The Tools of Neuroscience

Experiment

This volume establishes the conceptual foundation for sustained inves-

tigation into tool development in neuroscience. Neuroscience relies on
­diverse and sophisticated experimental tools, and its ultimate explanatory
target—our brains and hence the organ driving our behaviors—catapults
the investigation of these research tools into a philosophical spotlight.
The chapters in this volume integrate the currently scattered work on
tool development in neuroscience into the broader philosophy of science
community. They also present an accessible compendium for neurosci-
entists interested in the broader theoretical dimensions of their experi-
mental practices. The chapters are divided into five thematic sections.
Section 1 discusses the development of revolutionary research tools
across neuroscience’s history and argues to various conclusions concern-
ing the relationship between new research tools and theory progress in
neuroscience. Section 2 shows how a focus on research tools and their
development in neuroscience transforms some traditional epistemolog-
ical issues and questions about knowledge production in philosophy of
science. Section 3 speaks to the most general questions about the way we
characterize the nature of the portion of the world that this science ad-
dresses. Section 4 discusses hybrid research tools that integrate labora-
tory and computational methods in exciting new ways. Finally, Section 5
extends research on tool development to the related science of genetics.
The Tools of Neuroscience Experiment will be of interest to philoso-
phers and philosophically minded scientists working at the intersection
of philosophy and neuroscience.

John Bickle is Professor of Philosophy and Shackouls Honors College

Faculty at Mississippi State University and Affiliated Faculty in the
­Department of Neurobiology and Anatomical Sciences at the Univer-
sity of Mississippi Medical Center. He is author of four academic books
and editor of The Oxford Handbook of Philosophy and Neuroscience
(2009).

Carl F. Craver is a Professor in the Philosophy Department and the

­Philosophy-Neuroscience-Psychology Program at Washington University
in St. Louis. He specializes in the Philosophy of Science and has con-
tinuing research activity in the neuropsychology of memory. He is the
author of Explaining the Brain: Mechanisms and the Mosaic Unity of
Neuroscience and (with Lindley Darden) In Search of Mechanisms: Dis-
coveries across the Life Sciences.

Ann-Sophie Barwich is Assistant Professor at Indiana University Bloom-

ington (Department of History and Philosophy of Science and Medicine;
Cognitive Science). She specializes in olfaction as a model for theories of
mind and brain. Barwich is the author of Smellosophy: What the Nose
Tells the Mind (2020).
Routledge Studies in the Philosophy of Science

Understanding Perspectivism
Scientific Challenges and Methodological Prospects
Edited by Michela Massimi and Casey D. McCoy

The Aesthetics of Science

Beauty, Imagination and Understanding
Edited by Steven French and Milena Ivanova

Structure, Evidence, and Heuristic

Evolutionary Biology, Economics, and the Philosophy of Their
Relationship
Armin W. Schulz

Scientific Challenges to Common Sense Philosophy

Edited by Rik Peels, Jeroen de Ridder, and René van Woudenberg

Science, Freedom, Democracy

Edited by Péter Hartl and Adam Tamas Tuboly

Logical Empiricism and the Physical Sciences

From Philosophy of Nature to Philosophy of Physics
Edited by Sebastian Lutz and Adam Tamas Tuboly

The Explanatory Autonomy of the Biological Sciences

Wei Fang

The Tools of Neuroscience Experiment

Philosophical and Scientific Perspectives
Edited by John Bickle, Carl F. Craver, and Ann-Sophie Barwich

For more information about this series, please visit: https://www.routledge.

com/Routledge-Studies-in-the-Philosophy-of-Science/book-series/POS
The Tools of Neuroscience
Experiment
Philosophical and Scientific
Perspectives

Edited by
John Bickle, Carl F. Craver,
and Ann-Sophie Barwich
First published 2022
by Routledge
605 Third Avenue, New York, NY 10158
and by Routledge
2 Park Square, Milton Park, Abingdon, Oxon, OX14 4RN
Routledge is an imprint of the Taylor & Francis Group, an informa business
© 2022 selection and editorial matter, John Bickle, Carl F.
Craver, and Ann-Sophie Barwich; individual chapters, the
contributors
The right of John Bickle, Carl F. Craver, and Ann-Sophie
Barwich to be identified as the authors of the editorial
material, and of the authors for their individual chapters, has
been asserted in accordance with sections 77 and 78 of the
Copyright, Designs and Patents Act 1988.
All rights reserved. No part of this book may be reprinted
or reproduced or utilised in any form or by any electronic,
mechanical, or other means, now known or hereafter invented,
including photocopying and recording, or in any information
storage or retrieval system, without permission in writing from
the publishers.
Trademark notice: Product or corporate names may be
trademarks or registered trademarks, and are used only for
identification and explanation without intent to infringe.
Library of Congress Cataloging-in-Publication Data
A catalog record for this title has been requested

ISBN: 978-1-032-12799-6 (hbk)

ISBN: 978-1-032-17011-4 (pbk)
ISBN: 978-1-003-25139-2 (ebk)

DOI: 10.4324/9781003251392
Typeset in Sabon
by codeMantra
Contents

Foreword xi

Editors’ Introduction 1
J O H N B I C K L E , C A R L C R AV E R A N D A N N - S O P H I E B A RW I C H

SECTION 1
Research Tools in Relation to Theories 11

1 Tinkering in the Lab 13

JOH N BICK LE

2 Tools, Experiments, and Theories: An Examination of

the Role of Experiment Tools 37
G R E G O RY J O H N S O N

3 Science in Practice in Neuroscience: Cincinnati Water

Maze in the Making 56
N I N A A . ATA N A S OVA , M I C H A E L T. W I L L I A M S A N D
C H A R L E S V. VO R H E E S

4 Where Molecular Science Meets Perfumery: A Behind-

the-Scenes Look at SCAPE Microscopy and Its
Theoretical Impact on Current Olfaction 83
A N N - S O P H I E B A RW I C H A N D L U X U

5 A Different Role for Tinkering: Brain Fog, COVID-19,

and the Accidental Nature of Neurobiological Theory
Development 117
VA L E R I E G R AY H A R D C A S T L E A N D C . M AT T H E W S T E WA RT
viii Contents
SECTION 2
Research Tools and Epistemology 135

6 Dissemination and Adaptiveness as Key Variables in

Tools That Fuel Scientific Revolutions 137
A L C I N O J . S I LVA

7 Toward an Epistemology of Intervention: Optogenetics

and Maker’s Knowledge 152
C A R L F. C R AV E R

8 Triangulating Tools in the Messiness of Cognitive

Neuroscience 176
A N T O N E L L A T R A M AC E R E

9 Prediction, Explanation, and the “Toolbox” Problem 195

M A RC O J . N AT H A N

SECTION 3
Research Tools, Integration, Circuits, and Ontology 219

10 How Do Tools Obstruct (and Facilitate)

Integration in Neuroscience? 221
DAV I D J . C O L AÇ O

11 Understanding Brain Circuits: Do New Experimental

Tools Need to Address New Concepts? 239
DAV I D PA R K E R

12 Cognitive Ontologies, Task Ontologies, and Explanation

in Cognitive Neuroscience 259
DA N I E L C . B U R N S T O N

SECTION 4
Tools and Integrative Pluralism 285

13 “It Takes Two to Make a Thing Go Right”:

The Coevolution of Technological and Mathematical
Tools in Neuroscience 287
L U I S H . FAV E L A
 Contents ix
14 Hybrid Brains: Interfacing Living Neurons and Circuits
with Computational Models 305
AST R I D A. PR I NZ

SECTION 5
Tool Use and Development Beyond Neuroscience 319

15 Beyond Actual Difference Making: Causal Selections in

Genetics 321
JA N ELLA BA XT ER

List of Contributors 339

Index 345
Foreword

Stuart Firestein

The comic Emo Phillips has a very wry comment about the brain. He
muses that he always thought his brain was the most wonderful organ in
his body. Until one day he asked himself, “Wait a minute, who’s telling me
that?” Who indeed. This quip points out the hopelessness of understand-
ing the brain by introspection. Indeed, having a brain may be the biggest
obstacle to understanding it. Not that the organ is not smart enough to
understand how it works, but that the first-person experience of having a
brain, what it feels like it’s doing, is rarely a very good indication of what
is actually happening up there. In fact, it is usually dead wrong.
Thus, there is the need for instruments to visualize the structures of
the brain, to measure the activity of the brain and to probe its work-
ing at every level from single molecules to the whole buzzing, whirring
3.5 pounds of electrified pate. These instruments allow us to have a kind
of third-person disinterested view of the brain that is not as likely to be
swayed by its pernicious self-promoting propaganda. Or such would be
the ideal.
Of course, there always lurks the danger that the choice of measuring
device will reflect the answer we expect to get. As the geneticists warn,
you always get what you screen for. As neuroscientists, if we use elec-
trodes and amplifiers we will see the brain as a fundamentally electric
organ and believe we can understand it by recording voltages and ion
currents. The chemists and pharmacologists on the other hand will be
measuring neurotransmitters and binding kinetics revealing a picture
of a massive chemical factory. Then the anatomists with their golden
rule – function follows structure – will convince you that it’s all in the
connections and the vast parts list. And so it goes.
A hopeful solution to this is to take a pluralistic approach – not
­wholistic – but pluralistic. Recognizing that the brain is a subject (or is it
an object?) that can be examined at many different levels and from many
different perspectives. It is important to recognize that these multiple
perspectives are not likely to add up to a complete picture. Indeed, they
are more likely to maintain a very fragmented view of the brain. I for one
xii Foreword
believe this is the correct view of the brain. It will not be summed up in
a neat formula or model. Its very interest lies in its multiplicity.
When I first began in Neuroscience there were no departments of neu-
roscience at any major university. There were loose confederations of
faculty whose research was focused on the brain, but they all had their
primary appointments in various departments – anatomy, molecular
biology, physiology, evolution, psychology, computer science, electrical
engineering – all over the campus. Not to strike a nostalgic note here,
but there was something particularly correct about having brain research
distributed in that way. I’m still a bit uneasy about the proliferation of in-
stitutes dedicated to BRAIN RESEARCH, as if that was some monolithic
endeavor like building pyramids or cathedrals. It does seem like some-
thing a brain would be very pleased about – and that may not be good.
An alternative and more positive view of this concentration of brain
scientists into single departments or institutes is that neuroscience may
now find outreach to other fields even more beneficial. And those other
fields could even be outside the traditional scientific silos. They might in-
clude philosophy, literature, law, the arts, economics, sociology – indeed
many of these sorts of overtures are already in existence. Writers or art-
ists in residence at Neuroscience institutes have become more common;
books on music and the brain have proliferated; there are conferences
on ethics and law in the light of neuroscience findings. Perhaps counter-
intuitively neuroscience has become more pluralistic by becoming more
monolithic. After all, what we do with this brain is just as important to
understand as to how the parts fit together. The importance of instru-
ments and technology in driving the evolution of neuroscience both as
science and cultural phenomenon is the main subject of this collection,
so I will not go any further with this argument and allow the authors
herein to make it more thoroughly.
The pleasure of this volume is the inclusion of so many and varied
ideas about the brain, not only from practicing neuroscientists but also
philosophers and historians of neuroscience, all grounded in a wealth, an
abundant wealth, of technical, intellectual and instrumental approaches
that provide a wonderfully complicated view of this miraculous organ.
But who’s telling you that?
Editors’ Introduction
John Bickle, Carl Craver and
Ann-Sophie Barwich

The brain is science’s most photogenic organ. Glossy covers of modern

science journals routinely feature mesmerizing fluorescent neurons and
intricately woven cell structures with multicolored connections. And
increasingly, these figures are matched by the ability to intervene into
brain systems with drugs, electrical currents, magnetic fields and light,
to change them and bring them into line with our projected ideals. The
brain’s visual appeal mirrors the substantial insight that research on the
brain has advanced throughout the last century and beyond, as well as
our hopes and dreams for a future in which this central organ of control
is itself responsive to direct human commands.
The research tools of contemporary neuroscience have provided vec-
tors of access to brain structure and function that would have flabber-
gasted the pioneers of the field just 70 years ago. Tools are the most
conspicuous feature of neuroscience research, visible the moment you
walk into the lab: the scanner, the stereotactic surgery rig, the micro-
scope, the laser pointed at fixed tissue, the magnetic coils positioned
over localized regions of the brain, are themselves objects of our wonder
and vehicles of our imagination. These technologies drive progress in
neuroscience by opening new avenues by which we can “peer into” and
act upon brain systems. Through these technologies, we image active
brain areas, listen in on neurons talking to each other, and make mice
stop or go, feed or go wild, all with the flip of a switch.
Some have criticized neuroscience for its overreliance on technology
and fancy gadgets at the expense of theory, that neuroscience is “data
rich but theory poor.” Is neuroscience overwhelmed by the data gener-
ated by its research tools? And does this fixation on technology and the
data-collection it inspires explain why neuroscientists still lack a general
theory of how the brain makes the mind? This is a genuine concern.
Contrasted with physics, chemistry and evolutionary biology, neurosci-
ence does not yet offer a global theory of brain function. And a fasci-
nation with tools can, in fact, become a distraction from the hard work
of experimental design and theory building. Indeed, in his classic Ad-
vice for a Young Investigator, first published in 1897,1 Santiago Ramón

DOI: 10.4324/9781003251392-1
2 John Bickle et al.
y Cajal described “instrument addiction” as one ever-present form of
malaise threatening to derail the career of the young investigator.
Yet, even if we acknowledge Ramón y Cajal’s concern, there can be
no scientific theory of the brain or mind without the existence of instru-
ments that hone and refine our ability to intervene in and measure brains
and behavior. Good theory requires good data, and good data requires
instruments that work. Ramón y Cajal, himself a pioneer and devotee of
his own instruments (in his case, stains and light microscopes), would
surely acknowledge that nobody can understand contemporary neuro-
science or its history without attending to the arsenal of tools neurosci-
entists use to probe nature — to, in words often attributed to Francis
Bacon, “twist the lion’s tail.”
Specific research tools inform each of neuroscience’s many “levels” of
investigation, from the designer pharmacology used to study molecular
activities in intra- and inter-neuron signaling to the electrophysiology
rigs used to probe individual neuronal activities, to the functional imag-
ing modalities used to study neural systems, to the tasks and protocols
used to study behaving organisms and their cognitive processes. Indeed,
the often ambiguous notion of “levels” in neuroscience is closely tied to
the observational affordances of experimental tools and their heuristics.
Sometimes the localization of a particular event or item to a given level
is merely shorthand for describing the research tools and techniques em-
ployed to study it.
Some contend that every landmark discovery of contemporary neu-
roscience can be tied directly to the development and ingenious use of
one or more research tools (see Bickle this volume). Perhaps this claim is
exaggerated, but if so, only slightly. This centrality of research tools in
current neuroscience practice even motivates one of today’s most well-
funded and influential government-private enterprise partnerships, the
BRAIN Initiative. Its name is the acronym for Brain Research through
Advancing Innovative Neurotechnologies®, and it aims explicitly at “rev-
olutionizing our understanding of the human brain… [b]y accelerating
the development and application of innovative technologies,” (https://
braininitiative.nih.gov). The aim is to revolutionize neuroscience by
funding research into the development of new research tools and instru-
ments. The sheer dollar amounts associated with this initiative speak
volumes about neuroscientists’ recognition of the importance of new re-
search tools in their professional activities. Through 2019, BRAIN Ini-
tiative research awards already exceeded 1.3 billion US dollars (https://
braininitiative.nih.gov/about/overview).
Neuroscience is not unique in its reliance on experimental tools.
Many landmark discoveries in other sciences can likewise be tied di-
rectly to their innovations and developments. Telescopes guided Galil-
eo’s heliocentrism, the microscope was crucial to the development of the
germ theory of disease, and Rosalind Franklin’s X-ray crystallography
 Editors’ Introduction 3
led directly to the double helix model of DNA, to name just a few of the
most obvious examples. However, there is one crucial difference between
neuroscience and other tool-reliant sciences: the target of inquiry. Neu-
roscience’s ultimate explanatory target is our brains, the organ driving
our behaviors. This target catapults the investigation of research tools in
neuroscience into a brighter philosophical spotlight.
Philosophers are well suited to join neuroscientists in reflecting on the
development and use of research tools. Increasingly, young philosophers
are acquiring backgrounds in neuroscience as part of their undergrad-
uate, graduate, and postgraduate training. Acquiring this background
often involves using the tools involved in neuroscience research. Still,
sustained philosophical attention to the development and uses of exper-
iment tools in neuroscience remains rare. When new functional neuro-
imaging technologies such as positron emission tomography (PET) and
functional magnetic resonance imaging (fMRI) developed toward the
end of the twentieth century, there was a flurry of philosophical interest
in how these techniques worked and were being used in experiments in
the then-new cognitive neurosciences (e.g., Bogen, Roskies, cited in these
chapters). Unfortunately, those earlier discussions remained relatively
self-contained, as philosophers often linked their analysis of these new
techniques primarily to orthodox philosophical questions about the na-
ture of mind. Around that same time, the science-in-practice movement
across all of philosophy of science was successfully directing more philo-
sophical attention to aspects of day-to-day scientific activities and away
from philosophy of science’s traditional focus on the justification of sci-
entific theories and the relationship between theory and world. Many
philosophers of neuroscience came to count their research as part of the
broader science-in-practice movement, or the “practice turn.” Philoso-
phers of neuroscience have also been at the forefront of recent broader
interest in the philosophy of experiment. Nevertheless, the published
literature from this important refocusing on practices and experimenta-
tion still has one glaring lacuna. It has not yet addressed systematically
how the tools of scientific experiment come to be, how they are justified,
adopted, and proliferated, and finally how our theories change through
the uses of these tools, leading to cycles of change in the ability to inter-
vene into and detect brain function. In addition to epistemic and prag-
matic considerations about their application, another emerging question
concerns the explicitly cognitive role of tools in this context. Just how do
tools facilitate new thinking in scientific practice?
This gap in the philosophical literature presents philosophers with a
rare opportunity: to contribute fresh and novel ideas by reflecting on
how these research tools provide access to the brain and nervous system.
Furthermore, such philosophical contribution can play a notably active
role in the ongoing neuroscientific debate, given the dynamic nature of
the field and the relative recency with which many experimental tools
4 John Bickle et al.
have entered the brain sciences. (After all, fMRI is a child of the 1990s.)
Making such an active contribution will require detailed knowledge of
tool use in neuroscience practice. The project calls for collaborations
between philosophers and neuroscientists. Moreover, the work of the
handful of philosophers who have wandered onto this topic must even-
tually be combined into a coherent research program. This is what we
are trying to facilitate with this volume. As reflected in the chapters of
this volume, the work of a cadre of philosophers and philosophically in-
terested neuroscientists has begun to coalesce. We showcase their work
here in an effort to spawn more effective collaboration, to encourage
more consolidation into a recognized research area, and to instigate dis-
cussions and arguments that, we believe, will contribute meaningfully
and uniquely to our understanding of science as practiced.
The idea for this volume developed over the past half-decade. When
a handful of us first noticed our shared interest in a new tool in 21st-­
century neurobiology, two of the editors of this volume (Bickle and
Craver) joined up with two other philosophers of neuroscience (Sarah
Robins and Jacquiline Sullivan) to organize a submitted session on “op-
togenetics” at the 2016 Philosophy of Science Association meeting. Co-
incidentally, neuroscientist Stuart Firestein, who wrote the Foreward to
this volume, and the third editor, Barwich, were in the sessions’ audi-
ence. Attention to those discussions on the nature of new technologies in
the brain sciences seemed to persist, so Bickle and Craver co-organized
two recent professional workshops on tool development in neuroscience;
Barwich presented her research at both. The first was held in Pensacola
Beach, Florida, in September 2019. We expected 10–12 submissions to
our Call for Abstracts; by the submission deadline, we had received 25.
We held two ten-hour days of presentations and discussions. The second
conference was an Early Career Workshop on Neurotechnology, hosted
by the Center for Philosophy of Science and the Center for the Neural
Basis of Cognition at the University of Pittsburgh in January 2020 (for-
tunately, just before the Covid-19 pandemic shut down in-person con-
ferences). Bickle and Craver co-organized this second workshop with
Colin Allen and two History and Philosophy of Science Ph.D. candidates
at the time, Mahi Hardalupas and Morgan Thompson. Although the
general topic of this second workshop was broader than just research
tool development, that theme was prominent in our call for submissions.
In the end, we accepted eight submitted abstracts for presentation, each
was assigned a senior commentator; we invited four keynote speakers
and held a poster session. This workshop likewise comprised two full
days of presentations and discussions. And here our focus explicitly was
to promote work by junior investigators, the philosophers and scientists
who will be working on these topics for the next three to four decades.
These workshops, just like this volume, reflect the youth and diversity of
the philosophers and scientists who’ve taken interest in this topic. The
 Editors’ Introduction 5
philosophy of neuroscience is fortunate to have attracted not only highly
talented scholars but an impressively diverse group as well.
Those events led to this volume. Our fundamental goal is to promote
sustained investigations into tool development in neuroscience by phi-
losophers and neuroscientists. We hope this volume encourages more re-
searchers to reflect and perhaps extend this discussion to other sciences.
In these essays, broad epistemic considerations juxtapose seamlessly
with practical concerns. New insights emerge about the special place of
research tools in science, the norms of experimental practices, and the
strategies experimentalists use for teasing out nature’s secrets about that
wonderfully complex organ of thought, affect, and consciousness, the
human brain.
We’ve divided the 15 chapters in this volume into five sections. The
first section, Research Tools in Relation to Theory, has chapters orga-
nized around several case studies of tool development, some historical,
some recent. Each chapter discusses the development of revolutionary
(or potentially revolutionary) research tools across neuroscience and
supports different conclusions concerning the relationship between new
research tools and theory progress. Bickle describes the late-1950s de-
velopment of the metal microelectrode for single-cell neurophysiology
and the late-1970s development of the patch-clamp. He argues that the
historical details of these cases show that theory in neurobiology is dou-
bly dependent, not only on new research tools but also on atheoretical
“laboratory tinkering” through which these tools developed. Johnson
looks at examples of tools Bickle has stressed previously, especially gene
targeting and optogenetics. He finds a much less tidy picture than Bick-
le’s “tools-first” account proposes, and one that reflects an important
ambiguity in the claim that “theories depend on tools.” He argues that
sometimes theories and tools work in tandem, while in other cases, tools
are “first” and theory has little or no role in the investigation. Ata-
nasova, Williams, and Vorhees recount the process of the designing and
continuous refinement of the Cincinnati water maze, a tool utilized in
the study of rat egocentric learning and memory. This case study exem-
plifies the interplay between toolmaking, experiment, and theory where
tool availability often determines the kinds of experiments performed
in a certain laboratory. The authors introduce the notion of theories as
interpretative devices to capture cases in the practice of neuroscience
in which multiple experiments precede, rather than succeed, the the-
ory for which they provide empirical support. Barwich and Xu present
a first-hand account of Xu’s recent application of a new tool, SCAPE
microscopy, which uncovered a new mechanism of mixture coding at
the olfactory periphery. SCAPE (Swept, Confocally-Aligned Planar
Excitation) microscopy allows for fast three-dimensional, high-­
resolution imaging of entire small organisms (e.g., larvae) or large
intact tissue sections (e.g., in mice). Barwich and Xu illustrate how
6 John Bickle et al.
technologies facilitate a broader and different theoretical perspective by
being an integral part of the thinking process itself and by co-creating
mental structures. Hardcastle and Stewart show how “tinkering” with
19th-century manual “handheld” autopsy techniques led to discoveries
about otherwise puzzling “brain fog” reported in some advanced cases
of COVID-19 infection. This revamped technology enabled research-
ers to get around CDC (and common sense) restrictions on using aero-
sol sprays in brain autopsies of COVID patients and to find in infected
brains a class of large ­platelet-producing cells normally found in the
bone marrow. In this case, the “tinkering” with tools has led to new
implications about the blood-brain barrier, viral activity and the rela-
tionships between nervous and other bodily tissues.
We title the second section Research Tools and Epistemology. It in-
cludes chapters that show how a focus on neuroscience’s research tools
and their development transforms some traditional, much-discussed
epistemological issues in the philosophy of science, including the dis-
semination of knowledge, the integration of knowledge across fields of
neuroscience, and the relationship between explanation and prediction.
Silva argues that the power of a new research tool to capture scientists’
interests depends not only on its ability to reveal new and transforma-
tive phenomena, but also on its ease of use and transferability across
labs. He illustrates these dual(ing) concerns by discussing recent devel-
opments of “miniscopes,” miniaturized head-mounted fluorescent mi-
croscopes capable of recording from hundreds of cells for days or weeks
in freely-moving, behaving animals. Besides their potential impact on
science, miniscopes are also inexpensive, require no previous experience
with in vivo recordings, and can even be manufactured in the lab at
very low costs. How exactly do these epistemic and practical concerns
interact in daily laboratory science? Craver uses the early history of op-
togenetics technologies to study the norms governing the evaluation of
new interventionist technologies. Craver integrates this discussion with
James Woodward’s much-discussed manipulationist view of causal rele-
vance and lays out some foundational assumptions for an epistemology
of intervention. Craver calls attention to crucial normative distinctions
between those (the modelers) who use new tools for the purposes of ex-
ploring the brain and those (the makers) who use those tools to engineer
brain systems for our own ends, arguing for some significant epistemic
differences between these two groups of scientists. Tramacere analyzes
various cases of triangulation, where multiple tools have provided both
discordance and concordance of evidence. She argues that the use of
two or more measurement techniques (such as electroencephalogra-
phy (EEG), magnetic resonance imaging (MRI), and functional MRI
(fMRI)), has proven useful to minimize the limitations of each tool and
to refine experimental hypotheses about the role of the brain in cogni-
tion. Finally, she argues that triangulation, especially within the same
 Editors’ Introduction 7
experimental setting, is helpful with the integration of data and find-
ings, contributing to the understanding of how the mind works. Nathan
examines neuroimaging techniques in cognitive neuroscience and their
use in “reverse inferencing” psychological states from locations or pat-
terns of brain activity. Using these strategies the measured neural states
predict psychological features, but provide no causal-mechanistic ex-
planations of the psychological processes. This aspect of contemporary
cognitive neuroscience investigations, which Nathan also finds in other
human brain-mapping endeavors, does not sit easily with traditional
philosophy of science’s related treatments of explanation and prediction.
He suggests a “black-box” solution, according to which prediction and
explanation are grounded in the same underlying causal network, but
require different degrees of mechanistic detail.
Section 3 concerns Research Tools, Integration, Circuits and Ontol-
ogy. New research tools in neuroscience speak to the most general of
philosophical questions about how we characterize the nature of the
portion of the world that neuroscience addresses. The chapters in this
section cover three such questions: do research tools help or hinder at-
tempts to integrate the many fields and “levels” of neuroscientific inves-
tigation (or both)? The great increase in the contents of neuroscience’s
“toolkit” over the recent decades seems not yet to have yielded compar-
atively great explanatory increases in uncovering the mechanisms of be-
havior in even our simpler animal models. Is this because new tools need
the coincidental development of new neuroscientific concepts? How do
our research tools help us characterize the correct set of basic categories,
the “ontology,” of the brain and its activities? Colaço discusses CLAR-
ITY, a tissue clarifying tool that extends the scale at which optical mi-
croscopy can be used and so has successfully contributed what he calls
“local” integrations. However, his analysis of CLARITY also addresses
why the constraints and productivity of a research tool can contribute to
the difficulty, acknowledged by philosophers and neuroscientists alike,
of “systematic integration” of methods, data and explanatory schema
across neuroscience’s numerous fields. Parker considers the use of ex-
perimental tools to understand neural circuits. In the last 25 years, the
range of tools available for these analyses has increased markedly, but
he argues that neural circuit understanding has not increased signifi-
cantly despite techniques that allow us to examine more components
with greater precision and in more detail than before. He considers
whether we will have to develop tools that will allow us to address as-
pects of neurobiology that have not been considered in traditional circuit
approaches. Burnston investigates recent novel data-analysis tools that
some have proposed to offer illuminating explanations of how mental
functions are realized in the brain. He rejects this conclusion and instead
insists that results with these new tools recommend switching to a “task-
based cognitive ontology” of neuroscience’s basic explanatory target.
8 John Bickle et al.
Section 4, Tools and Integrative Pluralism explores how new com-
puting technologies have combined with more standard neuroscientific
wet-lab and behavioral laboratory tools and protocols to yield powerful
new approaches that integrated laboratory and computational meth-
ods in exciting new ways. We organize these two chapters historically.
Favela presents Hodgkin and Huxley’s Nobel Prize-winning work as
an instance of research constrained by both technological limitations
and mathematical assumptions. He then argues how the discovery of
scale-­invariant neuronal dynamics required both technological and
mathematical advances, and points out how these cases conflict with
the priority on technological dependence that some philosophers of
neuroscience have urged. Prinz describes the dynamic clamp technique,
which interfaces living neurons and circuits with computational mod-
els in real-time and at multiple levels, ranging from models of cellular
components and synapses to models of individual neurons to entire cir-
cuits. This technique creates hybrid in vivo-in silico systems where living
brains and computer models directly “talk to each other,” to take advan-
tage of combining experimental investigation of living neural systems
with the precise, theoretically-guided control over neural, synaptic, and
circuit parameters that computational models provide.
We end the volume with a section on Tool Use and Development Be-
yond Neuroscience. Neuroscience is hardly the only scientific endeavor
driven by its research tools. And yet none of the philosophies of specific
sciences has emphasized explicitly the place of tools in that science’s ac-
tivities or the implications tool use and development hold for philosoph-
ical reflection on its practices and products. Our hope is that this volume
spurs philosophical interest in research tools across all of science. To
illustrate that potential, in our final chapter Baxter investigates biolo-
gists conducting loss of function experiments, where they often single
out the genes they manipulate to explain the observable differences these
manipulations produce. She notices that the logical structure of the re-
sulting explanations vary in a regular fashion and argues that the types
of experimental tools that biologists use to intervene on genes account
for this variety.
Our broader goal of introducing a new concern into the philosophy
of neuroscience predates our specific interests in neuroscience’s research
tools. In mid-February 2008, two of us (Bickle and Craver) were in-
vited speakers at a one-day “Symposium on Neurophilosophy” hosted
by the then-new Ludwig-Maximilians-Universität Munich Center for
Neuroscience. Rather than immediately fly back to the States, we spent
the weekend at the Tegelberg in southern Bavaria. On the train back
to Munich late Sunday evening, before a long day of air travel back
to the States the next day, we decided that the ruthless reductionism-­
mechanism disagreement we had both been so committed to fighting
for nearly a decade was growing stale. (That was just as the issue of
 Editors’ Introduction 9
mechanisms was becoming the clearly dominant view in the philosophy
of neuroscience.) We wanted to find a new topic, perhaps even one on
which we could agree. It took us a while to find tool development; it
was another seven years after that first discussion until the idea for the
optogenetics symposium at PSA 2016 took shape. But it was fun to see
a new topic attract the attention of two longtime philosophical antag-
onists (and personal friends), especially one that also captured the at-
tention of Barwich and this outstanding collection of contributors from
philosophy and neuroscience.
By integrating these pluralistic approaches and interdisciplinary works
on tool use and development in neuroscience, we hope with this volume
to give this emerging research further momentum and to facilitate grow-
ing interest in this area.
We thank Andrew Weckenmann, editor of the Routledge Studies in
Philosophy of Science series, for his interest in pursuing a volume of
essays that seeks to introduce a new topic for further investigation, and
for the speed with which he and Routledge got our project into press. We
also are thankful for the helpful and positive comments of two anony-
mous reviewers on the book’s proposal.

Note
1 And recently made available in paperback by MIT Press (Cambridge: MIT
Press/Bradford Book, 2004).
Section 1

Research Tools in Relation

to Theories
1 Tinkering in the Lab
John Bickle

1 Putting Theory in Its Place

Progress in neuroscience is driven by its research tools. Nervous systems
are complicated and delicate and don’t readily reveal their machinations.
Available research tools are and have always been our only means of re-
vealing them. Philosophers have increasingly come to acknowledge this
fact, starting with a spate of interest in the workings of whole-brain
functional neuroimaging tools. More recently some have focused on re-
search tools common in neurobiology labs, including optogenetics and
DREADDs (Bickle 2016, 2018; Robins 2016, 2018; Sullivan 2018), gene
targeting techniques (Bickle 2003, 2015, 2016, 2019, 2020; Sullivan
2009), and various behavioral paradigms for rodent models (Sullivan
2010; Atanasova 2015).
The general philosophical lesson I have stressed in my contributions
to this growing literature is that in “wetlab” neurobiology, new tool
development drives everything else. It is where progress in this science
inevitably starts, and also where the bulk of attention in leading labs
is focused. This lesson contrasts sharply with the theory-centrism that
has characterized mainstream philosophy of science since its inception
as a recognized field in the mid-20th century. While contemporary
mainstream philosophy of science has certainly come to recognize the
existence and importance of scientific practices other than theorizing—
prediction, modes of observation, experimentation, to name a few—it
still too often subjugates these extra-theoretical practices to the goal and
service of theory progress and the confirmation of theory-world rela-
tions. Countering this mainstream theory-centrism was one explicit mo-
tive behind Ian Hacking’s (1983) push for a “back-to-Bacon” movement.
The still-advancing science-in-practice movement, which dedicates more
attention to extra-theory aspects of science, has been a promising recent
development. Still, a lot more needs to be done to square philosophy of
science with science’s actual day-to-day practices. More research on tool
development and use in laboratory neurobiology is a useful next focus.
A developer of one of laboratory neurobiology’s most influential re-
search tools, the high-impedance metal microelectrode (which you will

DOI: 10.4324/9781003251392-3
14 John Bickle
read about in the next section) once remarked that “our science seemed
not to conform to the science that we are taught in high school, with its
laws, hypotheses, experimental verification, generalizations, and so on”
(Hubel 1996, 312). Okay! So what did their science “conform to”? Our
philosophical view of what laboratory science is all about may hang in
the balance of these investigations.
It is important to clarify a number of points at the outset. First, re-
jecting theory-centrism about science is not tantamount to claiming that
theory has no place in wetlab sciences like neurobiology. Such a claim is
not only preposterous but also conflicts with the simple observation that
neuroscience’s best-confirmed current theories are from wetlab neurobi-
ology. The structure and function of ion channels and active transport-
ers that underlie neural conductance of action potentials; the detailed
molecular mechanisms of chemical neurotransmission in both pre- and
post-synaptic neurons; the intra- and inter-neuronal signaling pathways
leading to new gene expression, protein synthesis, and ultimately plas-
ticity of structure and responses in individual synapses, neurons, cir-
cuits, and ultimately organism behavior; these are current neuroscience’s
experimentally best-confirmed theories. The rejection of mainstream
philosophy of science’s theory-centrism is instead an attempt to put
theory in its proper place. Theory in wetlab neurobiology is dependent
upon the development and ingenious use of new research tools. The gen-
esis of every piece of theory just mentioned can be traced back directly
to the development of some new research tools. Theoretical progress in
this paradigmatic science of our times is secondary to and dependent
upon new tool development, both temporally and epistemically. Not the
other way around.
In this chapter, I will urge one more step away from theory-centrism
through further reflection on research tool development in neurobiology.
Theory progress therein is also dependent on what I will here call athe-
oretical laboratory tinkering. Theory progress in neurobiology is thus
doubly dependent, and hence tertiary in both epistemic and temporal
priority:

atheoretical laboratory tinkering → new experiment tool develop-

ment → theory progress.

My goal in this chapter is to argue for this additional step toward “put-
ting theory in its place,” in laboratory sciences like neurobiology.
Second important clarification: A common challenge to my attacks
on theory-centrism has been: “Surely the development of these research
tools presupposed theory! So how can theory be so dependent on them?”
Of course, theory informed the work that produced these research tools.
But as the detailed history of each tool shows, the theory progress that re-
sults didn’t guide the tool’s development. Here a distinction that emerged
 Tinkering in the Lab 15
in the recent philosophical literature on “exploratory experimentation”
is helpful. Franklin (2005) distinguishes experiments directed by both
theoretical background and local theory about the specific objects be-
ing investigated from experiments directed only by the theoretical back-
ground. Burian (2007) distinguishes theory-driven experiments from
theory-informed experiments depending upon whether or not the the-
ory provides expectations about what experimenters will find. Like ex-
ploratory experiments, tool development experiments appear to be these
latter types.1
A final clarification: Everyone concerned about the place of theory
in science should embrace philosopher Wilfrid Sellars’ quip, “the term
‘theory’ is one of those accordion words which, by their expansion
and contraction, generate so much philosophical music” (1965, 172).
I won’t “provide an analysis” of “theory” here. Sellars’ quip explains
why any “analysis” will simply be an explication of one of any number
of concepts that ambiguous word signifies. Instead, I’ll pursue my usual
metascientific approach (Bickle 2003; Silva et al. 2014) and present some
little-discussed history of laboratory neurobiology, namely, the devel-
opment of two of its most impactful research tools: the metal micro-
electrode for extended sessions of single-cell neurophysiology; and the
patch clamp, the tool that opened the door to the molecular revolution
that has guided mainstream neurobiology for the last four decades. The
mainstream philosophical sense(s) of “theory” that I’m trying to “put in
its (their) place” will hopefully become clearer in contrast with some his-
torical details about the kinds of atheoretical solutions that developers of
these laboratory research tools hit upon to get them to work.

2 Atheoretical Tinkering in David Hubel’s Lab circa 1957

For a quarter century, from the late-1950s well into the 1980s, electro-
physiology involving single-neuron recordings throughout the mamma-
lian cortex was a mainstay of basic neuroscience research. Not just in
vitro, in tissue slices, but in vivo, inside the brains of living animals. A
then-new research tool, the high-impedance conducting metal micro-
electrode, made some of that research possible. Armed with this new
tool, David Hubel and Torsten Wiesel (1962, 1965) found orientation-­
selective neurons and the hierarchy of simple and complex neurons in cat
primary visual cortex. Each won a one-quarter share of the 1981 Nobel
Prize for Physiology or Medicine for this work. A decade later Charles
Gross and colleagues (Gross et al. 1972) and two decades later Robert
Desimone and colleagues (Desimone et al. 1984) found face- and hand-­
selective neurons in the primate inferotemporal cortex. These neurons
were ­real-world approximations of Jerry Lettvin’s “grandmother” neu-
ron, the fanciful neuron that fires a high rate of action potentials when
and only when Granny is in its receptive field. Patricia Goldman-Rakic’s
16 John Bickle
group (Funahashi et al. 1993) found “working memory neurons” in the
primate dorsolateral prefrontal cortex, whose activity coded for the spa-
tial locations of stimuli that had been presented a few seconds earlier.
And William Newsome’s (Salzmann et al. 1992) and Ranulfo Romo’s
(Romo et al. 1998) labs transformed these recording metal microelec-
trodes into a microstimulation tool, using variants small enough to be
inserted directly into cortical columns to activate tiny clusters of sim-
ilarly-tuned cortical sensory neurons. These microstimulations altered
monkeys’ behavioral judgments about the character of the presented
sensory stimuli. This shortlist of groundbreaking results just scratches
the surface. (I’ll have more examples of landmark results from metal
microelectrodes later in this section.)
Less known is that David Hubel (1957) himself first perfected tung-
sten microelectrodes for cortical use, just a few years before he and
Wiesel embarked on their Nobel Prize-winning research. As Hubel re-
marks in his Nobel acceptance speech, recording from single cells in
the visual cortex was “potentially a gold mine” but “methods were
needed that would permit the recording of single cells for many hours,
and with eyes immobilized” (1981, 24). The standard glass micropi-
pette electrodes of the mid-1950s, which Vernon Mountcastle and Ste-
phen Kuffler had used to map the receptive fields of single neurons
in primate primary sensory and motor cortices and retina, tended to
break or bend at their tips during the much longer recording sessions
required for exploring visual neurons’ responses. Insulated metals were
the obvious choice for developing durable electrodes, but tip dimen-
sions necessary to resolve individual action potentials in single neu-
rons ranged from 5 microns for extracellular recordings to 1 micron
for intracellular recordings. 2 Achieving these tip dimensions in metal
faced serious technological difficulties. As Hubel remarks in the Sci-
ence publication introducing his technique, steel wire “becomes too
fragile near the tip when thus sharpened, and also requires too thick
a shaft” to resist breakage while being inserted through brain tissue
(Hubel 1957, 549). Following a suggestion by physicist Leon Levin,
Hubel chose tungsten as “by far the stiffest, easily available metal”
(1957, 549). Notice the nature of these initial problems and solutions,
as reflected in Hubel’s word choices, “becomes too fragile,” “requires
too thick a shaft,” “stiffest, easily available.” He confronted engineer-
ing problems that required practical solutions.
Hubel’s shift to tungsten only exacerbated his practical challenges.
He quickly discovered that one of the then-standard techniques for steel
electropolishing worked to generate tips of the required dimensions in
tungsten wire.3 Hubel consulted or appealed to no theory; he discovered
this strictly by trial and error, working with then-standard electropo-
lishing techniques. He also quickly discovered, by trial-and-error tin-
kering, that he could obtain almost any degree of taper in tungsten tips
 Tinkering in the Lab 17
by raising and lowering the wire during all but the final stages of elec-
tropolishing. However, insulating the tungsten microelectrodes proved
trickier. Hubel’s autobiographical remembrances illuminate the kinds of
problems this posed and the solution he stumbled upon:

I tried every coating I could find but nothing seemed adherent

enough or viscous enough. Formvar did not adhere and in any case
was available only in tank-car amounts. A solution of Lucite in
chloroform came close to working. One day while I was playing
around with this my neighbor in the next lab walked in with a can
of something called “Insulex” and said, “Why not try this?” I soon
found that when Insulex was thickened by evaporation it became
viscous enough to adhere to the wire, and suddenly I had an elec-
trode that was recording sensational single units.
(Hubel 1996, 303; my emphases)

The italics I’ve added emphasize the many practical considerations that
drove Hubel’s discovery. These are the phrasings of a skilled tinker, not
a physical chemist.
Hubel’s remembrances of inventing the hydraulic microelectrode ad-
vancer, required to drive the metal microelectrode to precise locations
in brain tissue in vivo, are equally revealing of the practical problems he
confronted and the tinkering solutions he found. “The problem was not
entirely simple … a hydraulic system seemed to be the best bet, but one
had to make the piston-and cylinder compatible with a chamber closed to
the atmosphere” (1996, 304). A closed chamber was required to prevent
movements of the cortex caused by pulsations in an open environment,
which would alter the electrode penetration path. Hubel found himself
“having to continually mollify machinists” when he brought back their
“skillfully built … latest model” and explained to them “why it couldn’t
work” (1996, 304). His solution? “Finally I decided I must learn how
to operate a lathe, and went to night school in downtown Washington,
D.C.,” (1996, 304). When all else fails, do it yourself (in Hubel’s case,
on your own time and dime). Note that this attitude about tool develop-
ment was from a future Nobel laureate.
Hubel’s tinkering attitude likewise drove his (and Wiesel’s) addition
to the histological “reconstruction technique” pioneered by Mountcas-
tle for confirming the exact placement of the recording electrode where
single-neuron recordings were taken. “Our addition … was the strat-
egy of making multiple small (roughly 100 μm diameter) electrolytic
lesions along each track by passing small currents through the tungsten
electrodes” (1981, 37). Did “theory” inform Hubel’s discovery of this
“addition”? Not hardly: “I worked out this method at Walter Reed by
watching the coagulation produced at the electrode tip on passing cur-
rents through egg white” (1981, 37). Tinkering redux.
18 John Bickle
And the rest really was history, in terms of progress in the neurobiol-
ogy of vision. Hubel and Wiesel’s work with Hubel’s tinkering-­generated
metal microelectrode and hydraulic microdriver on single neuron re-
sponses in cat primary visual cortex (striatal, V1), then on into extras-
triatal cortex (V2, V3), won them their shares of the 1981 Nobel Prize
for Physiology or Medicine “for their discoveries concerning informa-
tion processing in the visual cortex” (https://www.nobelprize.org/prizes/
medicine/1981/summary/). The Press Release announcing the award
ends by noting the principle of cortical organization based on individual
neuron function that directed subsequent neuroscience research for the
next quarter-century:

By following the visual impulses along their path to the various cell
layers of the optical cortex, Hubel and Wiesel have been able to
demonstrate that the message about the image falling on the retina
undergoes a step-wise analysis in a system of nerve cells stored in
columns. In this system each cell has its specific function and is
responsible for a specific detail in the pattern of the retinal image.
(https://www.nobelprize.org/prizes/medicine/1981/press-release/;
emphases in original)

The last sentence expresses the key theoretical concept of a sensory neu-
ron’s “receptive field,” which soon found application in motor and even
associational “cognitive” cortices. This experiment tool and approach
not only informed neuroscientific research through the 1980s that I sur-
veyed briefly at the beginning of this section, but its use also generated
the “subway map” models of visual cortical processing, and the distinc-
tion between the “dorsal visual stream” through the posterior parietal
cortex and the “ventral visual stream” through the inferior temporal cor-
tex (a distinction first found in nonhuman primates using Hubel’s metal
microelectrodes). All of this has been textbook neuroscience for three
decades. In his Nobel Award Ceremony Speech introducing Hubel and
Wiesel, Karolinska Institutet physiologist David Ottoman compared the
pair’s “translation” of the “symbolic calligraphy of the brain cortex” to

the deciphering of the hieroglyphic characters of the ancient Egyp-

tians … denoted as one of the greatest advances in the history of
philology. By breaking the code of the enigmatic signals of the visual
system you have made an achievement which for all time will stand
out as one of the most important in the history of brain research.
(https://www.nobelprize.org/prizes/medicine/1981/
ceremony-speech/)

All of this ground-breaking and enduring theory ties directly to

Hubel’s catch-as-catch-can, make-it-work, trial-and-error-tinkering
 Tinkering in the Lab 19
invention of a new experiment tool, and the new kinds of experiments
it enabled.

3 More on the Theory and Its Influence Driven by

Hubel’s Metal Microelectrode
Hubel and Wiesel’s shares of the 1981 Nobel Prize for Physiology or
Medicine stem in large part from three joint publications (1959, 1962,
1965).4 In these publications, the pair report their initial discoveries of
the receptive field properties of single neurons in the cat visual cortex,
both striate (Brodmann’s area 17) and extra-striate (Brodmann’s areas 18
and 19). They report their unexpected initial findings that bar stimuli at
specific spatial locations and orientations, rather than the expected small
circles drove activations in single neurons in area 17. This initial discov-
ery itself was by accident, resulting from Hubel and Wiesel’s attention as
experienced bench scientists, not appliers of systematic theory. One of the
only slides that initially evoked a consistently strong neuronal response in
a single cortical visual neuron in a single cat happened to have a vertical
crack near its bottom (see Hubel 1981, 27). Hubel and Wiesel noticed
the anomaly in that slide and immediately tried presenting slides with
bars oriented in various orientations instead of circles. Neurons across
V1 suddenly began responding vigorously to these new stimuli.
Extended recordings from individual cortical neurons in vivo during
the presentation of multiple stimuli, which Hubel’s metal microelectrode
permitted, led Hubel and Wiesel to their discovery that activity in these
neurons takes a Gaussian shape. A given primary visual cortex neuron
responds most actively (highest number of action potentials per time
unit) to bar stimuli at one specific orientation, e.g., vertical; its response
rate decreases systematically as bar orientations fall away from its most
preferred orientation in either direction. The careful penetrations and
follow-up histological techniques they refined and used with their new
tool also demonstrated that area 17 visual neurons in a given cortical
column shared receptive field properties (preferred spatial location and
bar stimuli orientations), but that a neuron’s exact rate of action poten-
tials depended on its location in a specific cortical layer.
These three publications also introduced another of Hubel and Wi-
esel’s contributions to subsequent neuroscience theory. They distin-
guished among “simple,” “complex,” and “hypercomplex” neurons,
strictly on measured receptive field properties and anatomical locations.
They speculated on circuitries obtaining between visual cortical neurons
through which the measured receptive field properties might be gener-
ated. Their (1965) paper demonstrated the feasibility of using the metal
microelectrode to explore receptive fields of individual neurons outside
of the primary visual cortex, and as Hubel states in his Nobel Prize ad-
dress, reported “the tendency toward increased complexity as one moves
20 John Bickle
centrally along the visual path, and the possibility of accounting for a
cell’s behavior in terms of its inputs” (1981, 36). Single-cell work on the
visual system also provided the experimental model for more detailed
work on other sensory systems, motor systems, and various cognitive sys-
tems throughout the association cortex in all cortical lobes. The influence
of those three Hubel and Wiesel publications has been enormous across
all of the neurosciences; this influence is well reflected by these papers’
citation numbers. Hubel and Wiesel (1965) has been cited 3,164 times;
Hubel and Wiesel (1959) 5,447 times; and their most-cited joint publica-
tion, Hubel and Wiesel (1962) has 15,303 citations.5 This number puts
their (1962) among the most cited papers in the history of neuroscience.

4 Laboratory Tinkering by Erwin Neher and Bert

Sakmann, circa 1974 … to 1980!
In the opening paragraph of his Nobel Prize address from December
1991, German biophysicist Erwin Neher describes the state of the search
for the mechanisms of neural conductance circa 1970. Alan Hodgkin
and Alex Huxley had “provided the basis” of the neuron action potential
nearly two decades before with their extensive experimental manipula-
tions and quantitative modeling of the fast sodium ion (Na+) influx, then
the slower potassium ion (K+) efflux across the axon membrane. But
“the question of the molecular mechanisms underlying these signals was
still open” (Neher 1991, 10). Hodgkin and Huxley (1952) speculated
about a voltage-operated “gate” in their formal model, and by 1970
“the terms Na-channel and K-channel were used frequently … although
no direct evidence for the existence of channels was available from bi-
ological preparations” (Neher 1991, 10). Experiments throughout the
1960s using artificial “black-lipid” membranes as models for biological
bimolecular lipid membranes had suggested that conductance changes
were from “single pore-like structures,” but in the biological membranes
“similar measurements were not possible at that time” (1991, 10). Us-
ing then-current recording equipment, background noise levels greatly
swamped currents of the sizes that were measurable in biological mem-
branes by more than two orders of magnitude. However, “indirect ev-
idence” suggested that channels with current conductances similar to
those measurable in artificial membranes should be operative in both
neurons and muscle cells. So:

It was very tempting to think about better methods for recording

currents from biological preparations. There was good reason to
hope that an improved technology would reveal a whole ‘microcos-
mos’ of electrical signals in a multitude of electrically and chemi-
cally excitable cell types.
(Neher 1991, 11; my emphases)
 Tinkering in the Lab 21
Neher’s words and attitude should be very familiar by this point in this
paper. A new experiment tool seemed the key next step to revealing the
mechanisms. And solving a host of practical problems in building such a
tool seemed to be the key first step: “an improved technology.” Was this
another job for laboratory tinkers?
The first task for resolving these tiny currents amidst extensive back-
ground noise was to isolate far smaller signal sources than that of the
membrane itself: to isolate small “patches” of the cell membrane. Neher
brought to this task his background in the spinning and use of glass
suction micropipettes from his Ph.D. work in Munich. Bert Sakmann,
fortuitously the scientist in the lab next door to Neher’s at the Max-
Planck Institut für Biophysikalische Chemie in Göttingen, had trained
in Bernard Katz’s lab and was experienced in treating cell surfaces en-
zymatically for precision current measurements. In the 1960s, working
with modifications of the voltage clamp, Katz’s lab had resolved macro-
currents from acetylcholine receptors from the end-plates of frog mus-
cle cells, which form the muscle component at neuromuscular junctions
(where motor neurons synapse onto muscle fibers). That work had pro-
vided the first direct experimental evidence of the “quantal” (step-wise)
nature of chemical neurotransmission. With their scientific backgrounds
Neher and Sakmann set to work to construct a “patch-clamp rig,” a min-
iaturized variation of Kenneth Cole’s quarter century-old voltage-clamp.
The basic idea of Neher and Sakmann’s patch clamp is simple. A tiny
hollow glass micropipette, filled with both electrolyte solution and a re-
cording electrode connected to an amplifier, is attached to the isolated
cell membrane being kept alive in a nutrient bath. “Tiny” is an under-
statement. The diameter of the pipette tip is in the single micron range
(see footnote 1 above), with a smoothened rather than sharp tip surface.
(At that time, sharp-tipped glass micropipette electrodes had been in
use for nearly three decades, to puncture membranes for intracellular
recordings.) A second electrode serves as a ground, completing a circuit.
For such “cell-attached” recordings the electrolyte solution inside the
pipette matches the composition of the nutrient bath solution. Often a
standard voltage-clamp rig is attached to the cell or membrane, to main-
tain a fixed membrane potential or to change the membrane potential
to precise values at specific times. Basic electrodynamics predicted that
achieving seals between micropipette tip and membrane surface with
resistances in the 10–100 gigaohms range should permit relatively noise-
free recordings of currents across the membrane in the single picoam-
pere range (1 ampere = 1,000,000,000,000 [1 trillion] picoamperes).
The ­pipette-membrane attachment is viewed under a microscope. The
entire rig is affixed to an anti-vibration table since the slightest vibration
can weaken the pipette-to-membrane seal.
As is typical with many ideas that seem simple, actually building a
functioning patch clamp rig was complicated. Neher and Sakmann had
22 John Bickle
a prototype rig in operation by 1974. They used it first to measure cur-
rents from a variety of cell preparations that contained acetylcholine
(Ach) receptors, as these were “probably the best-described channel in
biological membranes” at that time due to Katz’s ground-breaking work
from the previous decade (Neher and Sakmann 1976, 800). However, as
Neher later remarked, the pair “soon realized that it was not so easy to
obtain a satisfactory ‘seal’,” and despite their complementary expertise
and backgrounds on micropipette design and cell surface polishing, “our
initial attempts failed” (1991, 13). Trial-and-error work on firepolishing
micropipettes achieved the required smooth tips and reduced tip sizes
to the 3- to 5-micron range. This continually decreased the size of the
“patch” of the membrane the tool recorded from. Further refining var-
ious enzymatic treatments of membranes, especially to extra-synaptic
portions of denervated hypersensitive frog muscle fibers known from
pharmacological studies to be highly responsive to acetylcholine manip-
ulations, eventually generated seals sufficient to block background noise
and to measure square pulse waves of inward current of roughly 2–4
picoamperes from patches of tissue of roughly 10 square microns. A va-
riety of properties of Ach receptors they observed in their measurements,
predicted from Katz’s and others’ previous macroscopic recordings,
led Neher and Sakmann to conclude “that the observed conductance
changes are indeed recordings of single-channel currents” (1976, 801).
Their first paper reporting the patch clamp method and these initial
landmark but limited recordings was published in Nature in April 1976.
By the time their 1976 Nature publication appeared, Neher had spent
parts of 1975 and 1976 in Charles Stevens’ lab at Yale. He constructed
a patch clamp rig there and noted, in the phrasings of a true experimen-
talist, that “the fact that similar records could be obtained both in our
Göttingen laboratory and … at Yale … gave us confidence that they were
not the result of some local demon, but rather signals of biological sig-
nificance” (Neher 1991, 13). The square-wave nature of the measured
currents was particularly important because it was “proof” of Katz’s pre-
vious ground-breaking indirect macroscopic experimental evidence that
channels in biological membranes “open and close stochastically in an all-
or-none manner” (Neher 1991, 13; see also Neher and Sakmann 1976,
801 and Figure 3). The development of this new tool might have been
laborious and piecemeal, but the excitement generated by its early results
was palpable. “For the first time one could watch conformational changes
of biological macromolecules in situ and in real time” (Neher 1991).
Despite their excitement, Neher and Sakmann knew their initial
measurements were far from optimal. The big problem remained the
excessive background noise, traceable to the weakness of the pipette tip-
to-membrane seal relative to the tiny size of the currents. Too-weak seals
also led to currents from channels located under the pipette tip to be
recorded only partially. To strengthen the seal, preferably with internal
 Tinkering in the Lab 23
resistance in the gigaohm range, the pair and their collaborators “made
many systematic attempts to solve the seal problem” (Neher 1991, 15;
my emphasis). These included what you would expect from scientists
with Neher’s and Sakmann’s specific expertise and backgrounds: addi-
tional manipulation and cleaning of membrane surfaces, further minia-
turizing of pipette tip sizes and formation of even more steeply tapered
final tip lengths, coating pipette surfaces with different materials, re-
versing charges on glass surfaces, even developing a technique for mass-­
producing pipettes so that a new one could be used for each experiment,
solving the problem of miniscule pieces of membrane left on the tips
from previous experiments and weakening the seal (Neher et al. 1978;
Neher 1991). In other words, the pair did everything that then-current
theory suggested would strengthen the seal, yet all of this “with little
success” (Neher 1991, 15). Despite this failure to achieve the desired
seal, the new tool still proved useful for studying the properties of single
channels, especially cholinergic channels in hypersensitive muscle tissue.
But the patch clamp’s promise still seemed so much greater, for resolving
even smaller currents from other channels, if the gigaseal could some-
how be reliably and sustainably achieved.
Yet seemingly that achievement was not to be. As Neher notes, “by
about 1980 we had almost given up on our attempts to improve the seal”
(Neher 1991, 15). Notice the date Neher mentions. That was after more
than four years of working on the seal problem, virtually daily, since
their first Nature publication appeared in April 1976. Then one day:

we noticed by chance, that the seal suddenly increased by more than

two orders of magnitude when slight suction was applied to the
­pipette. The resulting seal was in the gigaohm range, the so-called
‘Gigaseal.’ It turned out that a gigaseal could be obtained repro-
ducibly when suction was combined with some simple measures to
provide for clean surfaces, such as using a fresh pipette for each
approach and using filtered solutions. The improved seal resulted in
much improved background noise …
(Neher 1991, 15)

Later in 1980, the group published their initial recordings of Na+ cur-
rents through the gigaseal-achieving patch clamp, through single ace-
tylcholine receptors in rat muscle fibers (Sigworth and Neher 1980).
By 1981 they published a full report of the improved new technique,
including how to mass-produce patch micropipettes, procedures for re-
liably achieving gigaseals, recording circuits and designs for improving
frequency responses, and improved procedures for preparing cell mem-
branes and physically isolating patches (Hamill et al. 1981). That paper
reported recordings using the improved technique from small-diameter
cells, including Na+ currents from single channels in bovine chromaffin
24 John Bickle
cells. Resolving currents from these channels had been impossible due
to too-weak seals, because of the currents’ tiny size and short duration.
The trial-and-error atheoretical tinkering attitude I am stressing is ap-
parent in Neher’s remarks quoted above (“by chance,” “it turned out”).
It is also explicit in his remark on the group’s capacity to amplify these
tiny measured currents. “Fortunately Fred Sigworth had just joined the
laboratory. With his experience in engineering, he improved the elec-
tronic amplifiers to match the advances in recording conditions” (1991,
15). Science writer (and biochemistry Ph.D.) Karen Hopkin has docu-
mented further remarks by Neher, Sakmann, and collaborators on their
achievement of the gigaseal (Hopkin 2010). Their “tinkering-first” at-
titude is unmistakable in those remarks. For example, Hopkins quotes
Neher suggesting that experimental know-how led to his breakthrough:
“you had to apply a little bit of suction in order to pull some mem-
brane into the orifice of the pipette … if you did it the right way it
worked” (Hopkin 2010; my emphases). She quotes lab member Owen
Hamill recalling “a weird period where we could no longer get gigaseals
… Then Bert [Sakmann] suggested that you have to blow before you
suck” (Hopkin 2010). Blowing solution through the clean pipette tip as
it approached the membrane apparently ejected any minute particles in
the tip that weakened the seal. No theory guided that discovery; it had
nothing to do with membrane polishing, Sakmann’s expertise. Rather,
this discovery was made “on the bench” by an experienced (and first-
rate) laboratory tinker.
Armed with their now-reliable new tool, Neher, Sakmann, and collab-
orators quickly demonstrated experimentally that the macroscopic ionic
Na+ and K+ currents in neurons measured in voltage clamp experiments
going back to Hodgkin and Huxley’s work were the sum of a huge num-
ber of microcurrents through individual channels specific to each ion.
The latter were now reliably measurable by the gigaseal patch clamp.
Either current could be induced through the voltage clamp stimulating
electrode by changing macroscopic membrane current. During multiple
current inductions, Na+ or K+ currents could be measured through the
patch clamp, with the other type of channels blocked pharmacologi-
cally by antagonists. An average of all of those (thousands of) patch
clamp recordings could then be computed and graphed. The graph of
time with respect to ion microcurrents exactly matched that for mem-
brane macrocurrents. The activities of Hodgkin and Huxley’s specula-
tive “voltage gates” in neuron membranes, and their role in generating
action potentials, were now observable on the lab bench. Coupling these
patch clamp discoveries with ones using an even older experimental tool,
X-ray crystallography (which had been around for nearly a century by
that time), quickly revealed channel structure. They were configured
proteins with differently charged amino acids at specific locations. And
the theory of ion channels and active transporters in neural conductance
 Tinkering in the Lab 25
that we possess today was quickly fleshed out experimentally. There is
an often-traced direct line of theory progress in neurobiology from Hod-
gkin and Huxley’s Nobel Prize to Roderick MacKinnon’s half-share of
the Nobel Prize in Chemistry in 2003 for his work on the structure and
mechanistic functioning of K+ channels. That path runs directly through
Neher, Sakmann and results obtained using their patch clamp.
And that was just the beginning of the patch clamp’s contributions to
neurobiology. As Salk Institute neurobiologist Charles Stevens, whose
Yale lab had hosted Neher’s 1975–1976 visit, remarked a decade ago,
“everything that we’re learned about the nervous system over the past
25 years we could not have done without patch camping. Erwin’s dis-
covery of the gigaseal made all the difference in the world” (quoted in
Hopkin 2010). The next few years following the reliable establishment
of the gigaseal bear out Stevens’ assessment. Ingenious experimenters
quickly discovered numerous ways the isolated patches of neuron mem-
brane could be manipulated. In his Nobel Prize address, Neher refers
to these further discoveries as “Unexpected Benefits” (1991, 15). They
were “unexpected” because they were not predicted by any theory; they
were discovered purely by chance or trial-and-error experimentation;
nobody saw these coming. In the words of this paper, they were the re-
sult of extended laboratory tinkering. First, numerous variations on cell
surface patching were found. Hamill et al. (1981) reported that a strong
pulse of suction through the pipette would not only establish the gigas-
eal but would also lesion the sealed membrane. This made the cytoplasm
of the cell continuous with the interior of the pipette and afforded ex-
perimenters a means to manipulate interior neuron function molecularly
in real time. These so-called “whole cell recordings” via the patch clamp
remain a laboratory standard in electrophysiology to this day.
Hamill et al. (1981) also reported that by applying suction to es-
tablish the gigaseal, then slowly lifting the pipette, the patch of mem-
brane sealed inside the pipette could be detached from the cell. This
“inside-out” patch-clamp variation exposed the cytoplasmic surface of
the membrane and its channels or receptors to molecular manipulation
via the solution in the nutrient bath. Finally, and most remarkably, the
gigaseal could be established with the cell membrane, strong suction
applied to lesion the membrane, as with the whole-cell recording tech-
nique, then the pipette lifted to rupture the cell membrane, only lift-
ing slightly more slowly than with the inside-out technique to remove
slightly more cell membrane on either side of the seal. The two detached
strands of the membrane would anneal together. This “outside-out”
patch clamp variation made the extracellular domain and its channel or
receptor components accessible to molecular manipulations for the first
time via the nutrient bath. All of the experimentally-confirmed theory
we now possess about receptors, including metabotropic receptors and
the variety of intra- and intercellular signaling their activation leads to,
26 John Bickle
traces back to these innovations with basic patch clamp techniques. All
these variations had been developed, refined, and put to use in numerous
laboratories worldwide by the time Sakmann and Neher published their
first comprehensive scientific review paper of the patch clamp technique
and its principal results, in 1984.
Neher himself, in his Nobel Prize address one decade later, remarked
on how transformative the patch clamp had already been in electrophys-
iology. He notes that the whole-cell recording variant quickly became
“the method of choice for recording from most cell culture prepara-
tions” because of its range of experimental applications: “many cell
types, particularly small cells of mammalian origin, became accessible
to biophysical analysis for the first time” (1991, 17). These cells simply
“would not tolerate multiple conventional impalements” that previous
tools required (1991, 17). Patch clamp whole-cell recording, inside-out,
and outside-out variants could also be used in combinations in detailed
multiple-experiment designs, and suddenly “individual current types
could be separated through control of solution composition on both
sides of the membrane” (1991, 17). The patch clamp radically shifted
experimental practices in laboratory neurobiology. It made the cells of
real interest, including mammalian neurons, experimentally accessible,
which technical limitations had precluded previously. And experiments
changed drastically, almost overnight. Neher reports some details:

This development shifted the emphasis of electrophysiological stud-

ies away from large-celled preparations, which usually were of in-
vertebrate origin, towards mammalian and human cell types. In the
first half of 1981, just before we first published a whole-cell char-
acterization of a small mammalian cell (bovine adrenal chromaffin
cells), only 5 out of 14 voltage-champ studies in the Journal of Phys-
iology were performed on cells of mammalian origin. The first 1991
issue of the same journal alone contained 10 voltage-clamp studies
on mammalian cells, none on invertebrates, and all using either the
whole-cell or single-channel recording techniques.
(1991, 17)

This shift constituted a real revolution in neurobiology, one reflected in

the actions, not just the words of bench neurobiologists. As I emphasized
in my (2016) concerning gene targeting techniques and optogenetics, the
revolution in neurobiology stemming from the patch clamp was likewise
built strictly on the development and ingenious use of this experiment
tool, itself the product of laboratory tinkering. The theory-centric com-
ponents of Kuhn’s (1962) famous model—paradigms, anomalies, crisis
science—are likewise not found in this history.
The most eye-opening and long-range influence of experiments with
the patch clamp were ones through which experimenters began to unravel
 Tinkering in the Lab 27
the vast networks of intracellular signaling pathways in active neurons.
These investigations constituted the leading edge of the “molecular rev-
olution” in neurobiology. The role of the patch clamp in this work be-
gan with early reports of “rundown” or “wash-out” when investigating
calcium (Ca2+) channels. We now know these channels to be subject to
modulation by second messengers, G-proteins, and phosphorylation; but
nearly 40 years ago none of this was known. Fenwick et al. first reported
“a slow run-down” of Ca2+ currents observed when using whole-cell
recordings, with “the speed of rundown … dependent on intracellular
Ca2+ concentration” (1982, 595). Channel activity measured through
the patch clamp disappeared rapidly, as a result of the perturbations
initiated by rupturing the cell membranes with whole-cell recording and
excised patch (inside-out, outside-out) techniques. Fenwick et al. spec-
ulated that the rundown “was apparently due to a progressive elimina-
tion of the channels available for activation” (1982, 595). Less than one
decade later, in his Nobel Prize address, Neher could already assert that
these puzzling results were “due to the loss of regulators, which at that
time were unknown,” through the ruptured membranes (1991, 19). But
these initially puzzling results only spurred more work using these new
experiment tools. As Neher claims, “subsequently, ingenious use of these
tools by many laboratories has revealed a whole network of interactions
between channels, second messengers, G-proteins, and other regulatory
proteins” (1991, 19). But in order to reveal this network, electrophysio-
logical recordings from these special channel-containing cells were not
enough. Experimenters also needed “to control or change systematically
the concentrations of second messengers” (1991, 19). And the patch
clamp and its standard variants, with their capacities to expose different
areas of the cell membrane and the cytoplasm to differing solutions in
the pipette and the surrounding nutrient bath, made these experimental
interventions possible.
The floodgates to increased knowledge of intracellular signaling had
opened. In his Nobel address, Neher cites three publications that quickly
followed Fenwick et al.’s (1982) initial report of rundown in Ca2+
­channel-containing cells. These three papers show the patch clamp’s ex-
panding role in both molecular neurobiology and physiology generally.
Fesenko et al. (1985) used the patch clamp to investigate “intracellu-
lar messengers” which couple activation of vertebrate retinal photore-
ceptive rod cell outer segments to single-channel activity. They found
that the activity of second messenger cyclic guanine monophosphate
(cGMP) “inside the inner membrane” was matched by increased cation
conductance through the individual channels (1985, 310). Kameyama
et al. (1986) found modulation of second messenger Ca2+ ion currents
during the phosphorylation cycle in patch-clamped guinea pig heart
cells. This work extended the use of patch clamping to investigate second
messenger activities in a wider range of tissue types. And Penner et al.
28 John Bickle
(1988) used the patch clamp to investigate multiple second messenger
pathways, including those involving Ca2+ and cyclic adenosine mono-
phosphate (cAMP) in rat peritoneal mast cells, producing elevated intra-
cellular Ca2+ concentrations released from intracellular storage sites. In
the three-decades span since Neher cited those three papers in his Nobel
Prize address, what we now know about these and other intracellular
signaling pathways and their significance for neuronal function has in-
creased by orders of magnitude. What is now presented in textbooks for
undergraduates (see, e.g., Purves et al. 2017, chapters 7 and 8) would
have astonished Neher and his fellow ground-breakers back in 1991.
The patch clamp and its standard variants made all this progress possi-
ble. This tool continues to be used by electrophysiologists and molecular
neurobiologists today, and sometimes still in the very forms that Neher
and others developed more than 40 years ago. All this is impressive his-
tory for a tool whose development resisted four years of “systematic”
attempts to achieve the gigaseal but yielded its promises to one lucky
accident by an observant experimentalist, which was followed quickly
by improvements found by a host of laboratory tinkers worldwide.

5 Answering the Call for “More Theory in

Neuroscience”
I have used some historical details about the development of the metal
microelectrode and the patch clamp to illuminate the double dependency
of theory progress in neurobiology, not only on the development of new
research tools but also the latter’s dependence on atheoretical laboratory
tinkering:

laboratory tinkering → new experiment tool development → theory

progress

I end this chapter by developing two further points of interest to both

philosophers and neuroscientists.
Nowadays we routinely hear calls for “more theory” in neuroscience,
especially from cognitive, computational, and systems neuroscientists and
their philosopher friends. Francis Crick was among the earliest callers for
this. As far back as 1979, while still an admitted “novice” in neurosci-
ence, Crick wrote an invited commentary on a series of essays published
in Scientific American on then-state-of-the-art neuroscience, all written
by leading neuroscientists. He remarked that while the articles “give a
good general idea of the progress that has been made” in understanding
the functioning, sensing, cognizing, emoting, and action-guiding brain,

what is conspicuously lacking is a broad framework of ideas within

which to interpret all these different approaches. Biochemistry and
 Tinkering in the Lab 29
genetics were in such a state until the revolution in molecular biol-
ogy. It is not that most neurobiologists do not have some general
concept of what is going on. The trouble is that the concept is not
precisely formulated …How then should a general theory of the
brain be constructed?
(1979, 133; my emphasis)6

Crick’s request for “more theory in neuroscience” was not unique

to him. Neurophilosopher Patricia Churchland, in her book that
initiated that field, likewise insisted that, at least concerning how
ensembles of neurons work to generate complex behaviors and cog-
nition, “there is no widely accepted theoretical framework, nor even
a well-defined conception of what a theory to explain such things
as sensorimotor control or perception or memory should look like”
(1986, 403). And while she admitted “some sympathy” for neurosci-
entists who judged that “theorizing about brain function is … slightly
disreputable and anyhow a waste of time,” she nevertheless offered a
number of reasons for “the value of theory” (1986, 403–407). Chap-
ter 10 of that book, nearly 80 of the book’s 482 pages, presents three
nascent “theories of brain function,” each one “illustrat[ing] some
important aspect of the problem of theory in neuroscience” (1986,
411). This same attitude carried over into her computational neuro-
science primer published six years later and co-authored with prom-
inent neuroscientist Terrence Sejnowski (Churchland and Sejnowski
1992). “‘Data rich but theory poor’ is a description frequently applied
to neuroscience,” they write; “in one obvious respect, this remains
true, inasmuch as we do not yet know how to explain how brains see,
learn, and take action” (1992, 16). The hope they express throughout
that book is that computational modeling, coupled with continued
experimentation into real brains by biochemists through neuropsy-
chologists, will provide the most promising strategy for generating
missing but needed theory.
Philosophers of neuroscience Ian Gold and Adina Roskies sounded
this familiar lament in their survey chapter from more than a decade
ago:

Neuroscience … has very few broad theories. It might be said that

the field is governed by a few global frameworks—a crude physical-
ism and perhaps computationalism—but these serve as fundamen-
tal or guiding assumptions rather than theories: They don’t provide
neuroscientists with predictive power in the way that physical theo-
ries do… Given its lack of theoretical richness, and the rather local
character of the theories that do exist, neuroscience looks quite dif-
ferent from both physics and evolutionary biology.
(2008, 351)
30 John Bickle
And as recently as 2016 Churchland and Sejnowski still lament: “neuro-
science is theory-poor” (2016, 667).
If it is a more and better theory that neuroscience needs or neuroscien-
tists want, then the lesson from a metascience of tool development rec-
ommends how to get that. Build new and better research tools! Without
obvious exception, the best, most well-confirmed theory in neuroscience
has resulted from the development of new research tools and their in-
genious uses on lab benches. As we saw in the previous section, the de-
velopment of one specific research tool, the patch clamp, generated the
“molecular wave,” not just a theory but a world view that has dominated
mainstream neurobiology for three decades.
A metascience of tool development in neurobiology yields an addi-
tional recommendation for those who want “more theory” in neuro-
science. My basic framework for tool development experiments (Bickle
2016), coupled with the new emphasis stressed in this chapter on the
place of laboratory tinkering, have all been derived retrospectively,
from historical case studies. But its basic concepts and lessons can also
be applied prospectively, to new tools under active initial development,
or even to ones simply being envisioned or imagined now. Furthermore,
the four case studies I have now analyzed provide exemplars for devel-
oping revolutionary new tools. What features does a new tool currently
under development share with, e.g., those that drove the development
of gene targeting techniques in behavioral neuroscience? Or those that
drove the patch clamp toward achieving the gigaseal reliably? In what
ways might the practical problems confronting developers of some envi-
sioned new tool, to make it work, resemble those Hubel confronted and
solved in his tinkerings that built the metal microelectrode and hydrau-
lic microdriver? Might new tool developers gain insights from details of
these historically successful cases and a metascientific analysis of their
shared dynamics? Might grant proposal reviewers, convinced of the
primary place of laboratory tinkering and new research tools for the
best-confirmed instances of theory progress across all of neuroscience,
look more favorably at proposals that emphasize the development of
new research tools?7 The abstract nature of my (2016) initial metasci-
entific account of components shared across different instances of tool
development grow increasingly detailed with each new case study to
which they get applied.
Interestingly, some neuroscientists seem to be groping toward conclu-
sions like these. Churchland and Sejnowski (2016), for example, couple
their claim (cited above) about neuroscience still being “­theory-poor”
with a surprising conjunct: “at the level of neuronal networks, neu-
roscience is also decidedly data poor” (2016, 667). Very much in
keeping with the lessons I am urging, they advocate explicitly for the
BRAIN initiative—Brain Research through Advancing Innovative
­Neurotechnologies—as a promising solution to this dual theory-data
 Tinkering in the Lab 31
poverty. Launched in the U.S. in 2013, the BRAIN Initiative is a
multi-federal agency public-private collaboration “aimed at revolution-
izing our understanding of the human brain … by accelerating the de-
velopment and application of innovative technologies” (https://www.
braininitiative.nih.gov/). As Churchland and Sejnowski insist, the hope
is that this Initiative will generate

new tools to obtain the data … new methods to analyze them …

inventing new technologies that will, it is hoped, foster the discovery
of a unified set of principles that link all levels of nervous systems
and thorough which brains perform their jobs.
(2016, 667)

The direction of influence that the BRAIN Initiative stresses should be

familiar to readers of this paper: new experiment tools → theory prog-
ress. Will BRAIN funders take heed of the lessons urged in this chap-
ter, that ‘atheoretical laboratory tinkering’ drives new tool development,
and put money toward proposals to do that?

6 Coda: Do Scientists Less Beholden to Theory Make

the Best Laboratory Tinkers?
I close this chapter by raising a related question. Are the scientists who
wind up inventing the research tools that revolutionize ­experiment-driven
fields typically the ones less beholden to or enamored with theory in
their day-to-day scientific activities? Does this attitude make them more
amenable toward the laboratory tinkering that drives new tool devel-
opment? And might this aspect of laboratory scientific practice become
institutionalized in the ways we train scientists?
I return to the attitudes of one of this chapter’s heroes. David Hubel
was a laboratory junkie. After he had perfected the process for making
and insulating tungsten microelectrodes, he recalls that he “spent the
next few months recording everything in the anesthetized cat’s nervous
system, from spinal cord to cochlear nucleus to olfactory bulb, almost
forgetting the original plan to record from awake, behaving animals”
(1996, 303–304). This laboratory tinkering resulted in his getting
“scooped” in publishing results using his new tool. Herbert Jasper of
McGill University, who ironically had been Hubel’s laboratory mentor
when Hubel was a medical student, traveled to Walter Reed in 1957
to learn firsthand Hubel’s tungsten electropolishing technique. He and
his post-docs returned to McGill, started producing the electrodes, and
were first to publish successful recordings from single neurons in awake
animals (primates) engaged in a cognitive (learning) task. In his scien-
tific autobiography, Hubel recalls his initial regret at his own dalliance:
“I wished I hadn’t taken so much time recording from so many parts of
32 John Bickle
cats’ brains” (1996, 304). But he also recalls soon coming to appreciate
the time he spent tinkering with his new tool:

The chance to play around at an early stage of one’s training is a

luxury denied to most beginning graduate students, who often start
in on a specialized problem assigned by an advisor, before having a
chance to try a few things for themselves.
(1996, 304; my emphasis)8

In the same paragraph, Hubel also remarks, “It has always surprised
me how few attempts are made to devise new methods—perhaps it is
because one is generally rewarded not for inventing new methods but
for the research which results from their use” (1996, 304). The lament
of a tinker!
This attitude pervaded Hubel’s entire approach to science. Speaking
for himself and Wiesel, he remarks that “we felt like 15th century ex-
plorers, like Columbus sailing West to see what he might find” (1996,
312; my emphasis). The impact of theory on their landmark research
was minimal:

If we had any “hypothesis” it was the simple-minded idea that the

brain, in particular the cerebral cortex, with all its ordered com-
plexity, must be doing something biologically meaningful with the
information that comes into it … So we recorded cells to see what
we might find.
(1996, 312–313)

And he generalizes this last point beyond himself and Wiesel: “I suspect
that much of science, especially biological science, is primarily explor-
atory in this sense” (1996, 313).
Hubel’s account of science’s “main pleasures” will be unfamiliar to
theorists: “To me the main pleasures of doing science are in getting ideas
for experiments, doing surgery, designing and making equipment, and
above all the rare moments in which some apparently isolated facts click
into place like a Chinese puzzle” (1981, 48). The first joy on Hubel’s list
is that of experimentalists, the second that of lab technicians, and the
third and fourth those of the tinkers, both mechanical and intellectual.
The task of writing Hubel’s obituary for Neuron, published a few
days after his death in October 2013, fell to Margaret Livingstone, his
second long-time research collaborator. She notes that he was “a giant in
our field,” but also “the guy in the lab who did the work that made him
great, and there is surely some connection between the way he daily went
about doing science and how successful he was” (2013, 735). She notes
the “two characteristics” that were most important to his scientific suc-
cess: “his mechanical inventiveness and his perseverance” (2013, 735).
 Tinkering in the Lab 33
She closes his obituary by remarking on the teaching Hubel undertook
after ending his research career:

When David did start teaching, he taught a Freshman Seminar at

Harvard College that was extraordinary popular, with ten times as
many students signing up each year as could be accommodated. He
taught them things that he thought were important but were missing
from most young people’s upbringing today: how to solder, how to
use power tools, how to suture skin, and how to wire up a simple
circuit.
(2013, 737)

I trust that Hubel’s tinkering-first attitude about science-in-general is

readily apparent in Livingstone’s remarks!
Intuitively, Hubel’s attitude about science seems fitting for some-
one destined to develop useful new experiment tools. The dedicated
­explorer-experimentalist will be attuned to the motivating problems for
new tools (Bickle 2016), and the dedicated tinker will be someone with
the patience and skills needed to build these new tools through ongoing
trial-and-error. But is there evidence for a general lesson here, between
those who’ve successfully built neuroscience’s most revolutionary, in-
fluential experiment tools and the general approach to science Hubel
articulated? If so, might this attitude be tracked, encouraged, and maybe
even taught as part of basic neurobiological training? Only further meta-
scientific work like that presented here can answer this question. But if
the lessons I’m stressing about the importance and dynamics of tool de-
velopment in neurobiology are correct, then neurobiology might stand
to gain enormously from these investigations.

Notes
1 A detailed investigation of this connection must await future work.
2 Microns, μm, millionths of one meter. The necessary tip dimensions thus
ranged from less than 4 one-hundred thousandths of an inch to less than 2
ten-thousandths of an inch.
3 By passing a 2- to 6-volt alternating current between the tungsten wire and
a nearby carbon rod, with the wire inserted into a 27-gauge hypodermic
needle, and its tip exposed and inserted into a saturated aqueous potassium
nitrite solution. I draw all details in this paragraph from Hubel’s (1957)
paper.
4 In the terminology of my metascientific model of tool development exper-
iments (Bickle 2016), these three publications present the metal microelec-
trode’s second-phase hook experiments, which bring the tool and its usage
potential to a wider range of scientists, and sometimes to the broader general
public, beyond specialists in the field in which the tool is developed.
5 These numbers were obtained from PubMed Central (https://www.ncbi.
nlm.nih.gov/pmc/), queried May 12, 2021.
34 John Bickle
6 Crick maintained this judgment to the end of his life. See one of his last pub-
lications before his death, co-authored with Christof Koch, where the pair
propose “a framework” for a neurobiology of consciousness, emphasizing
that a framework is not a theory, but rather just a “suggested point of view
for an attack on a scientific problem” (2003, 119).
7 See, e.g., my discussion (Bickle 2019) of Mario Capecchi’s remark about
his experiences with an NIH panel on his initial funding request for devel-
oping techniques of gene targeting by homologous recombination, followed
by subsequent remarks from the same panel a couple of years later after
he had persisted in his research despite that panel’s initial mistaken—and
­theory-grounded!—pessimistic judgment about its potential.
8 Hubel’s wording here should remind the reader of one of Ian Hacking’s
(1983) most quoted remarks about the development of microscopes: “We
should not underestimate … the pretheoretical role of … fooling around.”

References
Atanasova, N. (2015). “Validating animal models.” Theoria: An International
Journal for Theory, History and Foundations of Science 30: 163–181.
Bickle, J. (2003). Philosophy and Neuroscience: A Ruthlessly Reductive
Account. Dordrecht: Springer.
Bickle, J. (2015). “Marr and reductionism.” Topics in Cognitive Sciences (TOP-
ICS) 7: 299–311.
Bickle, J. (2016). “Revolutions in neuroscience: Tool development.” Frontiers in
Systems Neuroscience 10: 24. doi: 10.3389/fnsys.2016.00024.
Bickle, J. (2018). “From microscopes to optogenetics: Ian Hacking vindicated.”
Philosophy of Science 85: 1065–1077.
Bickle, J. (2019). “Linking mind to molecular pathways: The role of experiment
tools.” Axiomathes: Where Science Meets Philosophy 29(6): 577–597.
Bickle, J. (2020). “Laser lights and designer drugs: New techniques for descend-
ing levels of mechanisms in a ‘single bound’?” Topics in Cognitive Sciences
(TopiCS) 12: 1241–1256.
Burian, D. (2007). “On microRNA and need for explanatory experimentation
in post-genomic molecular biology.” History and Philosophy of the Life
Sciences 29(3): 283–310.
Churchland, P.S. (1986). Neurophilosophy: Toward a Unified Science of the
Mind-Brain. Cambridge: MIT Press.
Churchland, P.S. and Sejnowski, T.J. (1992). The Computational Brain. Cam-
bridge: MIT Press.
Churchland, P.S. and Sejnowski, T.J. (2016). “Blending computational and
experimental neuroscience.” Nature Reviews Neuroscience 17(11): 567–568.
Crick, F.H. (1979). “Thinking about the brain.” Scientific American 241(3)
(September): 219–232.
Crick, F.H. and Koch, C. (2003). “A framework for consciousness.” Nature
Neuroscience 6: 119–126.
Desimone, R., Albright, T.D., Gross, C.G., and Bruce, C. (1984). “Stimulus-­
selective properties of inferior temporal neurons in the macaque.” Journal of
Neuroscience 4(8): 2051–2062.
Fenwick, E.M., Marty, A., and Neher, E. (1982). “Sodium and calcium currents
in bovine chromaffin cells.” Journal of Physiology 331: 595–635.
 Tinkering in the Lab 35
Fesenko, E.E., Kolesnikov, S.S., and Lyubarsky, A.L. (1985). “Induction of cy-
clic GMP in cationic conductance in plasma membrane of retinal rod outer
segment.” Nature 313: 310–313.
Franklin, L. (2005). “Explanatory experiments.” Philosophy of Science (Pro-
ceedings) 72: 888–889.
Funahashi, S., Chaffe, M.V. and Goldman-Rakic, P.S. (1993). “Prefrontal neu-
ronal activity in rhesus monkeys performing a delayed anti-saccade task.”
Nature 365: 753–756.
Gold, I. and Roskies, A.L. (2008). “Philosophy of neuroscience.” In M. Ruse
(ed.), The Oxford Handbook of Philosophy of Biology. New York: Oxford
University Press, 349–380.
Gross, C.G., Rocha-Miranda, C.E., and Bender, D.B. (1972). “Visual proper-
ties of neurons in inferotemporal cortex of the macaque.” Journal of Physiol-
ogy 35(1): 96–111.
Hacking, I. (1983). Representing and Intervening. Cambridge: Cambridge Uni-
versity Press.
Hamill, O.P, Marty, A., Neher, E., Sakmann, B., and Sigworth, FJ. (1981). “Im-
proved patch clamp techniques for high-resolution current recording from
cells and cell-free membrane patches.” Pflügers Archiv (European Journal of
Physiology) 391: 85–100.
Hodgkin, A.L and Huxley, A.F. (1952). “A quantitative description of mem-
brane current and its application to conduction and excitation in nerve.”
Journal of Physiology 117: 500–544.
Hopkin, K. (2010). “It’s electric.” The Scientist. https://www.the-scientist.com/
uncategorized/its-electric-43471.
Hubel, D.H. (1957). “Tungsten microelectrode for recording from single units.”
Science 125(3247): 349–550.
Hubel, D.H. (1981). “Evolution of ideas of the primary visual cortex, 1955–
1978. A biased personal history.” Nobel Lecture. https://www.nobelprize.
org/nobel_prizes/medicine/laureates/1981/hubel-lecture.pdf.
Hubel, D.H. (1996). “David H. Hubel.” In L. Squire (ed.), The History of Neu-
roscience in Autobiography, vol. 1. Washington, DC: Society for Neurosci-
ence, 294–317.
Hubel, D.H. and Wiesel, T.N. (1959). “Receptive fields of single neurons in the
cat’s striate cortex.” Journal of Physiology 148: 574–591.
Hubel, D.H. and Wiesel, T.N. (1962). “Receptive fields, binocular interaction,
and functional architecture in the cat’s striate cortex.” Journal of Physiology
160: 106–154.
Hubel. D.H. and Wiesel, T.N. (1965). “Receptive fields and functional archi-
tecture in two non-striate visual areas (18 and 19) of the cat.” Journal of
Neurophysiology 28: 2290289.
Kameyama, M., Herscheler, J., Hofmann, F., and Trautwein, W. (1986). “Mod-
ulation of Ca current during the phosphorylation cycle of the guinea pig
heart.” Pflügers Archiv (European Journal of Physiology) 407: 123–128.
Kuhn, T. (1962). The Structure of Scientific Revolutions. Chicago, IL: Univer-
sity of Chicago Press.
Livingstone, M. (2013). “David H. Hubel, 1926–2013.” Neuron 80: 735–737.
Mishkin, M., Ungerleider, L.G., and Macko, K.A. (1983). “Object vision and
spatial vision: Two cortical pathways.” Trends in Neurosciences 6: 414–417.
36 John Bickle
Neher, E. (1991). “Ion channels for communication between and within
cells.” Nobel Lecture. https://www.nobelprize.org/nobel_prizes/medicine/
laureates/1991/neher-lecture.pdf.
Neher, E. and Sakmann, B. (1976). “Single-channel currents recorded from
membrane of denervated frog muscle fibers.” Nature 260 (April 29): 799–802.
Neher, E., Sakmann, B., and Steinback, J.H. (1978). “The extracellular parch
clamp: A method for resolving currents through individual open channels in
biological membranes.” Pflügers Archiv (European Journal of Physiology)
375: 219–228.
Penner, R., Matthews, G., and Neher, E., (1988). “Regulation of calcium influx
by second messengers in rat mast cells.” Nature 334: 499–504.
Purves, D., Augustine, G.J., Fitzpatrick, D., Hall, W.C., LaMantia, A.-S.,
Mooney, R.D., Platt, M.L., and White, L.E. (2017). Neuroscience, 6th Ed.
New York: Oxford University Press.
Robins, S.K. (2016). “Optogenetics and the mechanism of false memory.” Syn-
these 193: 1561–1583.
Robins, S.K. (2018). “Memory and optogenetic intervention: Separating the en-
gram from the ecphory.” Philosophy of Science 85(5): 1078–1089.
Romo, R., Hernández, A., Zainos, A., and Salinas, E. (1998). “Somatosensory
discrimination based on cortical microstimulation.” Nature 392: 387–390.
Sakmann, B. and Neher, E. (1984). “Patch clamp techniques for studying
ionic channels in excitable membranes.” Annual Review of Physiology 46:
455–472.
Salzmann, C.D., Murasugi, C.M., Britten, K.H., and Newsome, W.D. (1992).
“Microstimulation in visual area MT: Effects on direction discrimination
performance.” Journal of Neuroscience 12(6): 2331–2355.
Sellars, W. (1965). “Scientific realism or irenic instrumentalism.” In R.S. Cohen
and Marx W. Wartofsky (eds.), Boston Studies in the Philosophy of Science,
vol. II. New York: Humanities Press, 171–204.
Sigworth, F.J. and Neher, E. (1980). “Single Na+ channel currents observed in
cultured rat muscle cells.” Nature 287: 447–449.
Silva, A.J., Landreth, A., and Bickle, J. (2014). Engineering the Next Revolu-
tion in Neuroscience. New York: Oxford University Press.
Sullivan, J.A. (2009). “The multiplicity of experimental protocols: A challenge
to reductionist and non-reductionist models of the unity of neuroscience.”
Synthese 167: 511–539.
Sullivan, J.A. (2010). “Reconsidering ‘spatial memory’ and the Morris water
maze.” Synthese 177: 261–283.
Sullivan, J.A. (2018). “Optogenetics, pluralism and progress.” Philosophy of
Science 85(5): 1090–1101.
Van Essen, D.C., Anderson, C.H., and Felleman, D.J. (1992). “Information
professing in the primate visual system: An integrated systems perspective.”
Science 255(5043): 419–423.
2 Tools, Experiments, and
Theories
An Examination of the Role
of Experiment Tools
Gregory Johnson

1 Introduction
John Bickle (2016, 2018, 2019) offers two frameworks for thinking
about the role of experiment tools in neurobiology. First, he argues that
“revolutions in neuroscience” do not proceed in a Kuhnian manner such
that a dominant paradigm is replaced in response to an accumulation
of anomalies. Rather, revolutions in this science begin with motivating
problems that spur the development of new experiment tools, the impor-
tance of which is revealed in initial- and second-phase hook experiments
(2016). The initial-phase hook experiments demonstrate the feasibility
of the new tool. The second-phase hook experiments demonstrate its
usefulness in a wider range of experimental contexts and bring it to the
attention of a much larger audience—both scientific and more general.
Bickle extends this to a second framework, which is grounded in a
critique of “theory-centrism” in the philosophy of neuroscience (2018,
2019). As Bickle has it, theory-centrism is the view that neuroscience
needs theories—on the model of physics, early modern astronomy, or
evolutionary biology—that will drive successful research efforts. One
advocate of this view is Patricia Churchland. She writes,

If neuroscience is to have a shot at explaining—really explaining—

how the brain works, then it cannot be theory-shy. It must construct
theories. It must have more than anatomy and pharmacology, more
than physiology of individual neurons. It must have more than pat-
terns of connectivity between neurons. What we need are small-
scale models of subsystems and, above all, grand-scale theories of
whole brain function.
(1986, p. 406)

And then, more than two decades later, Ian Gold and Adina Roskies say,

The question remains whether neuroscience is the sort of science

that is doomed to be theory-poor, or whether this poverty is due to
its relative immaturity as a science. … The brain is an exceedingly

DOI: 10.4324/9781003251392-4
38 Gregory Johnson
complex biological organ which has evolved to perform a variety
of sophisticated tasks. It yields its secrets grudgingly. Nonetheless,
there is no principled reason why we cannot expect that, in time,
we will be able to formulate more general theories about the neural
processing that underlies these diverse functions.
(2008, p. 353)

Bickle aims to dispel this view that theory should have a central po-
sition in our understanding of neuroscience. Rather, he argues that we
should appreciate the central role of tool development and use. To the
extent that theory has a role in contemporary neuroscience, it is “ter-
tiary” in importance:

Rather than being the crux point on which everything else depends,
… theory turns out to be doubly dependent, and hence of tertiary,
not primary, importance. Our best confirmed theory is totally de-
pendent on what our experiment tools allow us to manipulate. And
those tools developed by way of solving engineering problems, not
by applying theory.
(2019, p. 578)

This can be interpreted in two ways. On the one hand, when a theory is
proposed, the theory depends on experiments and the tools used in those
experiments for its confirmation. Hence, without those tools, the theory
would fail to be confirmed. Arguably, though, dependence in this sense
doesn’t detract from a central role for theory in the scientific process.
This sense of dependence, however, doesn’t seem to be what Bickle has
in mind. For instance, in one place, he writes,

The molecular mechanisms of cognitive functions rank among con-

temporary neuroscience’s greatest theoretical achievements. And
yet this theory is tertiary in dependence. It comes directly from the
development and ingenious experimental use of some novel experi-
ment tools, to intervene into specific molecular processes in behav-
ing mammals. And those tools come from a catch-as-catch-can,
make-it-work, engineering-first attitude of the sort famously alluded
to by Hacking (1983), in his “microscope” argument for the relative
independence of “the life of experiment” from theory.
(2019, p. 594)

Here it seems that we are meant to understand that the temporal order,
as well as the order of importance, is, as Bickle later lays it out, “en-
gineering solutions → new experiment tools → better theory” (2019,
p. 595). I will call this the tools first (or anti-theory-centric) method with
the idea that, as Bickle stresses, the application of an experiment tool is,
 Tools, Experiments, and Theories 39
along with other factors, central to the investigation—but one of those
other factors is not the testing of a theory.
Surprisingly, given the role that it has in his analysis, Bickle is stu-
diously coy about what he means by theory. For the most part, rather
than use this term, I will use hypothesis, defined here as an explana-
tion for a process or phenomenon that still requires confirmation. By
focusing on hypotheses, I am deliberately setting aside theory used in
the sense of understanding, knowledge of the discipline, or completed
­explanation—or, as Churchland says, “this conglomeration of back-
ground assumptions, intuitions, and assorted preconceptions” (1986,
p. 405). I will take it for granted that theory in this latter sense is perva-
sive at all stages of neurobiological investigations.1
Bickle’s assertion that the tools first method is always used in con-
temporary neurobiology is a strong claim, and it will be our focus. In
Sections 2 and 3, I will look at two cases. The first, gene targeting and
investigations of the relationship between memory and long-term poten-
tiation, is extensively discussed by Bickle (2016, 2019). I find, however,
that a well-defined hypothesis does have a prominent role in these in-
vestigations. In short, a hypothesis was developed and then confirmed
by experiments using gene targeting. The second case, however—an op-
togenetic investigation of neurons in the extended amygdala that were
found to drive both anxiety and anxiety-reduction—illustrates the ap-
plication of Bickle’s tools first method.
The takeaway, then, is twofold. First, scientific method in contempo-
rary neurobiology is more varied than Bickle suggests, and sometimes
theory does have a central role. But, second, there are important investi-
gations in neurobiology that proceed without a hypothesis or theory as
the starting point (and without either coming into play at any point, for
that matter). This is, in part, a consequence of, as Bickle argues, experi-
ment tools that allow for ever more precise investigations of cellular and
molecular processes. It is also a consequence of the explanatory goals in
neurobiology, namely, the description of mechanisms. When these two
consequences come together, there is no longer an apparent need for the-
ories of the sort encountered in physics or evolutionary biology.

2 Gene Targeting and LTP

The first case involves two research tracks that intersected with pro-
ductive results in the 1990s. Long-term potentiation (LTP) is a lasting
increase in the efficacy of synaptic transmission following a sufficiently
strong stimulation from the pre-synaptic neuron. This phenomenon
was first reported by Terje Lømo in 1966 and then in more detail by
Lømo and Tim Bliss in 1973 (Lømo 1966; Bliss & Lømo 1973). The
idea that changes in the efficacy of synapses would be the neural basis
for learning and memory had been proposed before Bliss and Lømo’s
40 Gregory Johnson
work (Hebb 1949; Gardner-Medwin 1969), but even in the 1990s, the
idea that LTP—in particular, LTP at the CA3-CA1 synapses in the
­hippocampus—was the basis for at least some types of learning and
memory was still a hypothesis.
Richard Morris added some confidence to the hypothesis by demon-
strating that a pharmacological intervention targeted at NMDA re-
ceptors impaired spatial learning (Morris et al. 1986; Morris 1989).
DL-2-amino-5-phosphonopentanoic acid (AP5) prevents NMDA recep-
tors from allowing Ca2+ ions into the post-synaptic neuron, and, from
in vitro studies of hippocampal slices, AP5 was known to prevent LTP
(Harris et al. 1984). Morris infused AP5 into the ventricular system of
rats and then tested their performance on the Morris water maze. In
this test of spatial memory, rats or mice, swimming in a pool of opaque
liquid, explore until they find a submerged platform. After training, they
remember the location of the platform quite well using cues from out-
side the pool to orient themselves to its location—e.g., objects naturally
occupying the room (filing cabinets, posters on the wall, etc.) or images
placed around the room so that they can be seen from the pool. Mor-
ris found, however, that rats that had received the AP5 infusions were
impaired at learning the location of the platform, although they were
not impaired on a test of visual discrimination learning, which does not
depend on hippocampal functioning. These results support “the hypoth-
esis that the underlying neural mechanisms of LTP are activated during
and required for certain kinds of learning” (Morris & Kennedy 1992, p.
511). But, as Morris noted, the AP5’s disruption of the NMDA receptors’
function could have effects elsewhere in the hippocampus that are unre-
lated to LTP, and those other effects could be responsible for the spatial
memory impairment (Morris 1989, p. 3053). 2
The second research track was the development of targeted gene re-
placement as a tool for investigating the effects that specific genes and
the proteins that they encode have on behavior. When this technique is
used, as the name suggests, the gene for a single protein is replaced with
a non-functional copy. This was done, first, to fruit flies by Seymour
Benzer and his colleagues in the 1970s, and in the mid-1980s, Mario
Capecchi successfully knocked out developmental genes in mice. (See
Bickle 2016, 2019 for a fuller description of the development of this
tool in the 1980s.) This brings us to the intersection of targeted gene re-
placement and the relationship between LTP and spatial memory. Bickle
writes,

Taken together, the predicted general applicability of Capecchi’s

gene targeting techniques to mammalian nervous tissue, and the ex-
perimental demand to block LTP without disrupting other aspects
of synaptic function in order to test the alluring LTP → (rodent spa-
tial) learning and memory hypothesis, constituted the motivating
 Tools, Experiments, and Theories 41
problem for the application of gene targeting techniques in mamma-
lian behavioral neuroscience.
(2016, p. 4)

In the early 1990s, in response to this motivating problem, Alcino Silva

and Susumu Tonegawa used gene targeting for the first time in neuro-
biology to prevent LTP in mice and to investigate the LTP → spatial
memory hypothesis (Silva et al. 1992a, 1992b). They replaced the gene
for the α-form of calcium-calmodulin-dependent kinase II (α-CaMKII),
a protein that was believed at the time to have a role in LTP and is
now known to be needed for the insertion of new AMPA receptors in
the post-­synaptic cell’s membrane, one of the early steps that increases
the efficacy of transmission at the synapse. They found that this pre-
vented LTP in hippocampal neurons in vitro (1992b) and impaired spa-
tial learning in the Morris water maze (1992a). The α-CaMKII deficient
mice were not impaired, however, on a test of non-spatial memory.
Silva et al. report that their results “considerably strengthen the con-
tention that the synaptic changes exhibited in LTP are the basis for spa-
tial memory” (1992a, p. 206). But elsewhere, when discussing this use of
targeted gene replacement, Tonegawa points out,

In 1992, this work led to the report of the first knockout mice in neu-
roscience. It was shown that deletion of the α-CaMKII gene causes
a deficit in LTP at Schaffer collateral-CA1 synapses and an impair-
ment of spatial learning. Even before publication, however, we were
aware of the limitations of this approach in conventional knockout
mice. These limitations were primarily because the gene of interest
is deleted in the entire animal throughout the animal’s life. Although
no obvious developmental defects were observed in the α-CaMKII
knockout mice, more subtle defects could not be excluded. Further-
more, the universal absence of the protein in question (in this case,
α-CaMKII) certainly did not permit the establishment of a causal
relationship between CA1 LTP and spatial learning.
(2001; see also Tsien et al. 1996, p. 1327)

So, while Silva et al. showed that gene targeting could be used in neuro-
biological investigations and their study added some degree of confirma-
tion to the LTP → spatial memory hypothesis, their results encountered
much the same problem as did Morris’s. Like Morris, Silva and his col-
leagues disrupted one component in the process that gives rise to LTP
and, at the same time, this intervention impaired spatial learning. But,
in both cases, the intervention was broad enough to raise questions
about whether it was the absence of LTP or the impairment of some
other process that was responsible for the poor performance on the spa-
tial memory task. Four years later, however, further work on the gene
42 Gregory Johnson
targeting technique by Joe Tsien—who was also working in Tonegawa’s
lab—made it possible to draw a stronger conclusion.
Silva et al. had introduced the non-functional copy of the α-CaMKII
gene into embryonic stem cells that were then injected into fertilized
mouse embryos. The embryos were implanted in the uteruses of surro-
gate females, and some of the mice that were born contained the mutant
gene. Mating those mice with wild types and then mating mice that
were heterozygous for the mutant gene eventually yielded mice that were
homozygous for the mutant gene. This ensured that these mice lacked
α-CaMKII in CA1 neurons in the hippocampus. But these mice also
lacked this protein at all stages of development and in all brain areas
where it would normally be expressed: throughout the hippocampus, in
the cortex, septum, striatum, and amygdala.
Shortly after Silva et al.’s study was published, Tsien used the same em-
bryonic stem cell method to create two lines of mice. One line expressed
an enzyme derived from the bacteriophage P1 only in pyramidal cells in
the hippocampus’s CA1 region. This enzyme, Cre recombinase, targets
specific sequences of DNA, 34 base pair “loxP” sites, and it will excise
a sequence of DNA that is flanked by these 34 base pair sequences. (Or,
in a procedure that was, at this point, not yet developed, Cre recombi-
nase can flip the orientation of a gene so that, depending on its original
orientation, it either will be transcribed or won’t be transcribed.) In the
other mouse line, loxP sites flanked the NMDAR1 gene, a gene that
encodes an essential subunit of the NMDA receptor. But besides the in-
clusion of the loxP sites, these mice had normal genomes and developed
normally with a functioning NMDR1 gene. Mating mice from these two
lines produced offspring in which the Cre recombinase would excise the
NMDR1 gene in CA1 neurons, but nowhere else (since Cre recombinase
wasn’t expressed in any other neurons). And, since Cre recombinase only
begins to be expressed several weeks after birth, the loss of the NMDR1
gene did not interfere with prenatal or perinatal development.
Tsien and his colleagues found that the mice that lacked the NMDAR1
gene exhibited impaired spatial learning on the Morris water maze with-
out an impairment to non-spatial learning, and, in vitro, stimulating
CA3 axons did not induce LTP in CA1 neurons. Tsien et al. conclude:
“these results provide strong support for the hypothesis that NMDAR-­
mediated LTP in the CA1 region is crucially involved in the formation
of certain types of memory” (1996, p. 1328). And a little later they add:

Previous studies examined the correlation between spatial memory

and the site of hippocampal LTP (i.e., CA1, CA3, and dentate gy-
rus) by using a variety of conventional knockout mice. … Our new
evidence, while still correlational, is much stronger than that in the
earlier reports because we have singled out the CA1 synapses as a
site of plasticity impairment.
(1996, p. 1335)
 Tools, Experiments, and Theories 43
That said, the issue here is not whether this hypothesis has been suffi-
ciently confirmed or by whom. Rather it is that, in these studies, the re-
searchers are clearly seeking to confirm a hypothesis. Morris even lays it
out in textbook fashion when he says that “the hypothesis makes the im-
portant and testable prediction that AP5 should impair a subset of those
learning tasks that are impaired by hippocampal lesions but should be
without effect on tasks unaffected by such damage. The present series
of experiments tests this prediction” (1989, p. 3041). Each team of in-
vestigators, then, is judicious about the degree of confirmation provided
by their own results and even more so about the degree of confirmation
provided by the investigation that preceded theirs.
The LTP → spatial memory hypothesis was dependent on gene tar-
geting for its eventual confirmation, but, temporally, the hypothesis pre-
ceded the application of the tool, and verifying the hypothesis was the
goal of these investigations. Hence, the investigations did not employ the
tools first method, which is based on the absence of a well-defined hy-
pothesis.3 This, however, takes nothing away from gene targeting being
a revolutionary new tool, which, at least on Bickle’s terms, it has been. It
changed the types of investigations that could be undertaken, and its use
has become pervasive across cellular and molecular neuroscience (2016,
p. 2). Arguably, it has a greater claim to triggering a revolution for the
field than the second tool Bickle (2016) discusses, optogenetics, since
optogenetics depends on the manipulation of genes.

3 Optogenetics
The second “neuroscientific revolution” that Bickle (2016) describes
is the development and subsequent use of optogenetics. This tool,
which was initially developed in the early 2000s by Karl Deisseroth,
Edward Boyden, and colleagues working at Stanford University, uses
light to control the activation or inactivation of neurons (Boyden et al.
2005). This is done by taking genes for light-activated proteins from
algae, archaebacteria, and other microbial organisms and expressing
them in the neurons of interest. For instance, the frequently used ion
channel channelrhodopsin-2 occurs naturally in the green alga C. re-
inhardtii. When it is expressed in the neural membrane and exposed
to blue light, channelrhodopsin-2 allows positively charged ions to
enter the cell, which causes an excitatory response. Conversely, halor-
hodopsin, which is found in the single-celled archaeon N. pharaonis,
pumps negatively charged chloride ions into the cell when it is ex-
posed to yellow or orange light. This hyperpolarizing current inhibits
the cell. Once the light-sensitive protein is expressed in neurons, an
optical fiber that has been implanted in the animal’s brain over the
area of interest is used to deliver a specific wavelength of light. Thus,
neurons can be activated or inhibited while still allowing the animal
to move freely.
44 Gregory Johnson
Bickle describes the motivating problem that spurred the development
of optogenetics this way:

The motivating problem for optogenetics stemmed from the

­causal-mechanistic explanatory goals of mainstream neurobiology.
To test hypotheses about the causal role played by specific neurons
to produce behaviors routinely taken as indicators of particular
cognitive functions in mammalian models, experimenters need the
capacity reliably to intervene into the hypothesized mechanisms in
vivo—to activate or inhibit them in the behaving organism, as di-
rectly and as efficiently as possible. … If successful, neurobiology
would then have a great new experimental interventionist tool to
explore causal-mechanistic hypotheses relating activity in specific
neurons to behaviors—including behaviors routinely taken to indi-
cate the occurrence of particular cognitive functions. This was op-
togenetics’ motivating problem.
(2016, pp. 10–11)

This captures the central issue, but we can give a less hypothesis-centric
characterization of the problem. Here is how Boyden puts it:

As I interviewed for PhD programs in neuroscience in spring 1999,

I asked the scientists I met on each stop the same question: what
tools should a physical sciences-trained investigator develop, to help
understand the brain?
One fact that emerged from those conversations was that there
are many different kinds of neuron in the brain, which possess dif-
ferent morphologies, molecular compositions, and wiring patterns,
and which undergo different changes in disease states. New neuron
types, and new properties of existing neuron types, are being dis-
covered all the time. … In order to determine how different kinds
of neuron in the brain work together to implement brain functions,
and to assess the roles that specific sets of neurons play within neu-
ral circuits, it would ideally be possible to drive or quiet the activity
of defined neurons embedded within an intact neural network. By
driving the activity of a specific set of neurons, it would be possible
to determine what behaviors, neural computations, or pathologies
those neurons were able to cause. And by silencing the activity of a
specific set of neurons, it would be possible to determine what brain
functions, or pathologies, those neurons were necessary for.
(2011, p. 2)

This is, of course, related to Bickle’s characterization of the motivat-

ing problem, but here we begin, not with “causal-mechanistic hypoth-
eses,” but with the nature of the neuronal milieu. When distinguished
 Tools, Experiments, and Theories 45
by structure, function, the genes they express, or how they are arranged
and project to other neurons, the types of neurons packed into each brain
area are extremely diverse. Lesioning, pharmacological interventions,
and electrical stimulation and recording are useful tools, but they lack
the specificity to intervene on only a narrowly defined set of neurons.
Thus, the need for optogenetics, which allows for control of neurons
that are identified by genetic markers (which is related to their function
and structural features), location, and where their axons project.
Whereas before we saw gene targeting deployed to confirm the hy-
pothesis that LTP underlies spatial memory, we will now turn to a study
that is not investigating a hypothesis. This is not how all investigations
using optogenetics proceed, but the study is a fair exemplar of the tools
first method in neurobiology.

3.1 Optogenetics and Divergent Motivational States

The dopamine pathways in the brain are composed of dopamine-­
releasing neurons in the midbrain’s ventral tegmental area (VTA) and
substantia nigra that project to the prefrontal cortex, nucleus accumbens,
and striatum. Electrical stimulation, pharmacological interventions, and
lesioning have shown that these neurons have a role in reward-seeking
and motivated behaviors (e.g., Olds & Milner 1954; Ungerstedt, 1971;
Crow 1972; Wise et al. 1978). Not surprisingly, however, given that
dopaminergic and non-dopaminergic neurons are intermingled in the
VTA and have these diverse projections, the increased specificity of cell-
type targeting afforded by optogenetics has been embraced. Some of this
work has been done by Garret Stuber, who, not long after the very first
in vivo studies using optogenetics were performed, began investigating
the effects that dopaminergic neurons have on neurons outside of the
VTA and on behavior (Tsai et al. 2009; Stuber et al. 2010; Adamantidis
et al. 2011). He then investigated, again using optogenetics along with
other tools, the function of non-dopaminergic neurons in the VTA that
synapse on dopamine-releasing neurons (van Zessen et al. 2012). Then,
stepping further back from the dopaminergic neurons, he began investi-
gating the functions of neurons in other parts of the brain—the lateral
habenula, the bed nucleus of the stria terminalis, and the lateral hypo-
thalamus—that project directly or indirectly to the VTA (Stamatakis
& Stuber 2012; Jennings et al. 2013; Stamatakis et al. 2016). This has
yielded an increasingly detailed picture of the functional and structural
features of the neurons that provide inputs to the VTA-based dopamine
system. (For reviews of some of this work see Deisseroth 2014; Stamat-
akis et al. 2014.)
In the study that investigated the functional properties of neurons in
the bed nucleus of the stria terminalis that project to the VTA, Joshua
Jennings, Stuber, and colleagues found that a small number of neurons
46 Gregory Johnson
can drive and prevent behavior that reflects anxiety in mice (Jennings
et al. 2013). They introduce the study this way:

The extended amygdala, including the bed nucleus of the stria ter-
minalis (BNST), modulates fear and anxiety, but also projects to the
ventral tegmental area (VTA), a region implicated in reward and
aversion, thus providing a candidate neural substrate for integrating
diverse emotional states. However, the precise functional connectiv-
ity between distinct BNST projection neurons and their postsynap-
tic targets in the VTA, as well as the role of this circuit in controlling
motivational states, have not been described.
(2013, p. 224)

There is no question that they are starting with a thorough

­understanding—such as was available—of these brain areas. Their goal,
however, is not hypothesis-testing but rather to describe “the precise
functional connectivity” of these neurons and how these circuits control
motivation and behavior.
To carry out this investigation, they began with two genetically modi-
fied mouse lines. In the first, Cre recombinase was expressed only in neu-
rons that release the neurotransmitter glutamate. The adeno-­associated
virus was then used to deliver the channelrhodopsin-2 gene to neurons
in the ventral BNST (vBNST). Most of the viral DNA had been removed
from each copy of the virus and replaced with the gene for channel-
rhodopsin-2 inverted between loxP sites. When this viral construct was
injected into the vBNST, it infected, harmlessly, all neurons there. Ex-
pression of the channelrhodopsin-2 gene, however, was dependent on
Cre recombinase flipping the gene to the correct orientation so that it
could be transcribed. Thus, only the excitatory glutamate-releasing neu-
rons ended up with the light-sensitive ion channels. And since the viral
construct was only injected into the vBNST, the expression of channel-
rhodopsin-2 was limited to that area of the brain. In the other mouse
line, the mice expressed Cre recombinase only in neurons that release
the neurotransmitter γ-aminobutyric acid (GABA). Using the same viral
construct, channelrhodopsin-2 was expressed only in inhibitory GABA-­
releasing neurons in the vBNST in these mice.
After injecting the virus, Jennings et al. let four to six weeks elapse
so that channelrhodopsin-2 would be expressed, not only in and around
the neurons’ cell bodies but also in the presynaptic terminals that reach
into the VTA. The optical fiber for photostimulating the vBNST neurons
with blue light was then implanted in the VTA, thus ensuring that, while
vBNST → VTA neurons would be photostimulated at their distal axons,
neurons in the vBNST that project elsewhere would not be.4 The result
was that the neurons that could be optogenetically manipulated were se-
lected on the basis of the neurotransmitter that they release—glutamate
 Tools, Experiments, and Theories 47
or GABA—the location of their cell bodies, and where their axons proj-
ect, all without any apparent adverse effects on the mice.
Once the mice were prepared, Jennings et al. used several tests to in-
vestigate the functions of these neurons in freely moving mice. First,
they combined optogenetic control of the neurons with recording from
a microelectrode array implanted in the vBNST. So that they could dis-
tinguish the neurons in the vBNST that project to the VTA from other
neurons in the vBNST, they photostimulated the presynaptic terminals
in the VTA to generate action potentials that propagated up the neurons’
axons (i.e., from the pre-synaptic terminals to the cell bodies). Detec-
tion of a spike less than 20 ms after the first light pulse was delivered
indicated that the neuron being recorded was either a glutamatergic or
GABAergic vBNST → VTA neuron. With the ability to record the ac-
tivity of these neurons, Jennings et al. submitted the mice to an unpre-
dictable foot-shock procedure: mice received about 20 foot-shocks over
a 20-minute period while also being exposed to two constant contextual
cues, light from a house lamp and white noise. During this period—and
excluding the moments when the mice received the foot-shocks—they
found that, compared to an earlier baseline period, there was increased
activation of the glutamatergic vBNST → VTA neurons. Conversely,
the GABAergic vBNST → VTA neurons showed a reduction in activ-
ity during the 20-minute period. This established that the glutamater-
gic neurons respond during an aversive event while the activity of the
GABA­-ergic neurons is suppressed during such an event.
In the remaining parts of the study, photostimulation was delivered
during the experiments. In the first two, mice could choose to receive
or avoid photostimulation of the vBNST → VTA neurons either by
staying on one side of an arena or by nose poking. Mice in which the
glutamatergic neurons would be stimulated avoided photostimulation,
while the mice in which the GABAergic neurons would be stimulated
sought it out. Next, and interestingly, in an experiment with mice that
had been food-deprived, constant photostimulation of the vBNST →
VTA glutamatergic neurons caused a reduction in nose poking that
would deliver a food reward. In the final experiments, each group of
mice was submitted to one of two tests that are frequently used to as-
sess anxiety in rodents, the open field test, and the elevated plus maze.
In the first, following 20 minutes of constant photostimulation of the
glutamatergic vBNST → VTA neurons, mice spent significantly more
time in the corners of the open field arena—a 25 × 25 cm c­ hamber—
and less time in the center compared to controls, behavior that in-
dicates a heightened level of anxiety. Meanwhile, mice that received
constant photostimulation of the GABAergic neurons spent more time,
compared to control mice, exploring the open arms (as opposed to the
enclosed arms) of an elevated plus-shaped maze, indicating reduced
levels of anxiety.
48 Gregory Johnson
The takeaway is straightforward. Activation of the glutamatergic
vBNST → VTA neurons enhances anxiety, while activation of the GABA-­
ergic vBNST → VTA neurons suppresses it. There is more to say about
the neurons that are targeted by these vBNST → VTA neurons, but, even
without that detail, we can see that these different types of neurons—
from the same brain area and projecting to the same brain area—cause
very different types of motivation and behavior.5 Jennings et al. conclude:
“Although the canonical view of BNST function proposes a dominant
role of this structure in promoting anxiety states, the cellular and func-
tional complexity described here illustrates that particular BNST circuit
elements orchestrate divergent aspects of emotional and motivational
processing” (2013, p. 228). Granting that “the canonical view” is theory
is some loose sense, their goal was to investigate and describe these neural
circuits and that is what they accomplished. Moreover, a few years after
the study was completed, Stuber said this about their aims:

One thing that we began to think about more as we were finishing

this study was that these are ways that we can manipulate different
cell populations, but these are still very broad-spectrum cell target-
ing strategies. We are still, essentially, targeting all glutamatergic
neurons or all GABAergic neurons within this structure. And so, af-
ter our 2013 paper, we began to search for other genes—which may
be enriched in these structures, such as the BNST—that could serve,
potentially, as genetic handles or entry points [and] would allow us
to tap in and modulate and study more precise cell types.
(2016)

This led them to investigate the functional properties and connectivity

of the subpopulation of GABAergic neurons in the BNST that express
the gene for the neuropeptide nociceptin (Rodriguez-Romaguera et al.
2020). While it may be that Stuber and his team have expectations or
thoughts about what their experiments will reveal, their interest appears
to sidestep theory or hypotheses. Instead, they are seeking out more and
more details of these neural mechanisms.

4 Conclusion
Scientific method in contemporary neurobiology is varied. Hypotheses
may be formulated and tested, but they need not be. Ian Hacking (1983)
discusses this point, and it is instructive to see how far he is willing to
take it. He begins by quoting the 19th-century chemist Justus von Liebig
who wrote,

Experiment is only an aid to thought, like a calculation: the thought

must always and necessarily precede it if it is to have any meaning.
 Tools, Experiments, and Theories 49
An empirical mode of research, in the usual sense of the term, does
not exist. An experiment not preceded by theory, i.e., by an idea,
bears the same relation to scientific research as a child’s rattle does
to music.
(Über Francis Bacon von Verulam und die Methode der
Naturforschung, 1863; in Hacking 1983, p. 153)

Hacking, then, explains that there is a weak and a strong version of

Liebig’s statement about the role of theory. The weak version, which
Hacking finds uncontroversial, is simply that

you must have some ideas about nature and your apparatus before
you conduct an experiment. A completely mindless tampering with
nature, with no understanding or ability to interpret the result,
would teach almost nothing.
(1983, p. 153)6

The strong version of Liebig’s statement, meanwhile, has it that “your

experiment is significant only if you are testing a theory about the phe-
nomena under scrutiny” (1983, p. 154), and this Hacking believes to
be false. Of course, sometimes it is true that we precede that way, but
“one can conduct an experiment simply out of curiosity to see what
will happen” (1983, p. 154). The strong versus the weak reading of Li-
ebig’s statement is the distinction that we have encountered. Morris,
Silva, Tonegawa, Tsien, and their colleagues were testing a theory (or
a hypothesis, at least) about LTP and spatial memory. Jennings, Stuber,
and their colleagues—while having “some ideas about nature and [their]
apparatus”—were experimenting to see what they would find.
But while Hacking allows that experiments can be productively car-
ried out without being tests of a theory, he then adds that theory will
eventually be needed to make sense of the results (1983, pp. 158, 164–
165). One of his examples is Brownian motion. Although Robert Brown
was not the first to notice the phenomenon, he observed and carefully
recorded the motions of pollen in water in 1827. But, Hacking reports,
“for long it came to nothing” (1983 p. 158). It took Einstein’s molecular
theory of heat—which was itself confirmed in a series of experiments by
Jean Perrin—to explain Brown’s observations.
This is where we break from Hacking. He is not concerned to make
every experiment look like a test of a theory, but in the end, he maintains
that “plenty of phenomena attract great excitement but then have to lie
fallow because no one can see what they mean, how they connect with
anything else, or how they can be put to some use” (1983, p. 158). He
then concludes, “Thus I make no claim that experimental work could ex-
ist independently of theory” (1983, p. 158). Similarly, Churchland wor-
ries that “by shunning theory, one runs the risk that the data-gathering
50 Gregory Johnson
may be random and the data gathered, trivial” (1986, p. 404). Hacking
may be right when the examples come from physics, and Churchland is
writing in the 1980s, before the development of the tools discussed here
and before the mechanistic turn in the philosophy of biology, a point to
which we will turn in a moment. Their worries, in any case, appear not
to apply to Jennings et al.’s results. We may wish to know more about
the projections of dopaminergic neurons in the VTA, their effects on
behavior, and the inputs that they receive, but there is no real difficulty
with making sense of the results.
Part of the reason for this is because explanations in cellular and mo-
lecular neurobiology are mechanistic. Carl Craver, for one, is clear on
this point. Referring to “early twenty-first-century physiological sciences,
such as neuroscience,” he writes, “In such fields, the language of mecha-
nism is literally ubiquitous, and most scientists continue to demand that
adequate explanations reveal the hidden mechanisms by which things
work” (2013, p. 134). As is well known, this involves identifying the
parts that compose the system, both their structures and functions, and
how those parts are organized. William Bechtel puts it this way:

On my analysis, a mechanism is an organized system of component

parts and component operations. The mechanism’s components
and their organization produce its behavior, thereby instantiating a
phenomenon (Bechtel & Abrahamsen, 2005; …). According to this
view, a mechanism is a system operating in nature, and a mechanis-
tic explanation is an epistemic product. To arrive at a mechanistic
explanation, scientists must represent (sometimes verbally, but often
visually in diagrams) the component parts and their operations and
the ways in which they are organized.
(2005, pp. 314–315, italics in the original)

Seeking mechanistic explanations does not, however, by itself lead to

the tools first method. To the contrary, as Craver points out, these are
“hidden mechanisms,” and if we had little or no access to them, then
theorizing would be a prominent part of trying to understand them.
But, even though the components and operations of the brain are hetero-
geneous, fragile, and relatively inaccessible, experiment tools that have
been developed for investigating the brain are, as we have seen, provid-
ing access to cellular and sub-cellular mechanisms.7 Gene targeting (as
well as the viral delivery of genes) and optogenetics are two prominent
examples, but there are many others. One that bears mentioning because
it is, in a way, the complement to optogenetics is calcium imaging, which
uses proteins that can absorb and emit light (delivered to neurons by the
same methods that are used for light-sensitive ion channels) to image
neurons firing individual action potentials.
 Tools, Experiments, and Theories 51
While this is only a subset of the experiment tools used in contem-
porary neurobiology, it illustrates their scope nicely. Proteins can be
removed or added to specific types of neurons with gene targeting or
with viral gene delivery. At the cellular level, precise types of neurons in
specific locations can be activated or inhibited with optogenetics. How
these interventions effect behavior can then be tracked. Calcium imag-
ing, meanwhile, allows the activity of individual neurons to be tracked
when animals are given a behavioral task. Consequently, when the goal
is descriptions of ‘organized systems of component parts and component
operations’ and when the tools exist to identify those parts, their oper-
ations, and how they are organized, then, rather than expecting the de-
velopment of theories that must be submitted to experimental tests, we
should expect neuroscientific practice to be a process of exploring and,
piece by piece, adding detail to mechanistic explanations.

Acknowledgments
This chapter has benefited from many discussions of John Bickle’s work
on tool development at Mississippi State University and at the 2018 and
2019 Philosophy and Neuroscience at the Gulf annual meetings. I would
also like to thank John Bickle for his comments on a draft of this chapter.

Notes
1 I am not claiming that theory in the sense of a hypothesis (or a collection
of hypotheses) can be cleanly distinguished from any of the other senses of
theory or that we should always want to make such a distinction. I am only
claiming that the distinction will be useful here.
Bickle, for his part, uses theory in different ways. When he invokes and
rejects “theory-centrism,” he is adopting Churchland, Gold, and Roskies’
call for theory in the sense of proposed explanations that direct experimen-
tal investigations and are confirmed or disconfirmed by those investigations.
But he also uses theory in the sense of well-confirmed and mainly completed
explanations—for instance, when he says,
neuroscience, I claim (Bickle 2015, 2016, 2018), has lots of good,
well-confirmed theory; … the mechanisms of neuronal conductance
and transmission at chemical synapses, receptor and ion channel func-
tion, mechanisms of synaptic plasticity, including its molecular-genetic
and epigenetic mechanisms, details of anatomical circuitries linking
neurons to other neurons, and ultimately to sensory receptors and mus-
cle tissue.
(2019, p. 580).
Then, in his description of the development of gene targeting and optogenet-
ics, he is at pains to show that these tools were developed by “catch-as-catch
can laboratory tinkering” (2019, p. 594) not by the considered application
of theory—that is, not by the application of well-confirmed scientific expla-
nations (2018, pp. 1071–1074; 2019, pp. 591–594).
52 Gregory Johnson
2 Silva et al., in the introduction to their own study, make the point plainly:
Although LTP has been studied as a mechanism responsible for some
types of learning and memory, the actual evidence for this hypothesis is
not extensive. The main support for LTP as a memory mechanism is the
observation that pharmacological agents that block hippocampal gluta-
mate receptors of the N-methyl-D-aspartate (NMDA) class and thus pre-
vent the induction of LTP also impair spatial learning in rodents (Morris
et al. 1991). The problem with this evidence is that blocking NMDA
receptors disrupts synaptic function and thus potentially interferes with
the in vivo computational ability of hippocampal circuits. Perhaps the
failure of learning results not from the deficit in LTP but simply from
some other incorrect operation of hippocampal circuits that lack NMDA
receptor function.
(1992b, p. 201)
3 Bickle’s analysis is different. He focuses on the theory, or lack thereof, that
was used by Capecchi to develop gene targeting in mice (2019, pp. 591–594).
That, however, switches the focus from the theory that is stated in the mo-
tivating problem, and which clearly concerned these investigators, to theory
in a different context.
4 Control animals were from the same two lines of mice. Each received the injec-
tion of the viral construct and the optical fiber was implanted in its VTA. But, for
these mice, the viral construct did not contain the gene for channelrhodopsin-2,
and so their vBNST → VTA neurons could not be activated by light.
5 For the curious, using immunohistochemical staining, photostimulation,
and patch-clamp recording in brain slices, Jennings et al. found that the
vBNST → VTA neurons “formed functional synapses primarily onto
non-dopaminergic and medially located dopaminergic neurons.” They then
add, “these data provide a circuit blueprint by which vBNST subcircuits
interact with VTA-reward circuitry” (2013, pp. 225–226).
6 Although, he later adds,
Indeed even the weak version is not beyond doubt. The physicist George
Darwin used to say that every once in a while one should do a completely
crazy experiment, like blowing the trumpet to the tulips every morning
for a month. Probably nothing will happen, but if something did happen,
that would be a stupendous discovery.
(1983, p. 154)
7 Bickle, in another context, has made this same connection between new
experiment tools that began to be developed in the 1980s and 1990s and
mechanistic explanations (2015; see especially Section 4).

References
Adamantidis, A. R., Tsai, H.-C., Boutrel, B., Zhang, F., Stuber, G. D., Budygin,
E. A., Touriño, C., Bonci, A., Deisseroth, K., & de Lecea, L. (2011). Opto-
genetic interrogation of dopaminergic modulation of the multiple phases of
reward-seeking behavior. Journal of Neuroscience, 31(30), 10829–10835.
Bechtel, W. (2005). The challenge of characterizing operations in the mecha-
nisms underlying behavior. Journal of the Experimental Analysis of Behav-
ior, 84(3), 313–325.
 Tools, Experiments, and Theories 53
Bechtel, W., & Abrahamsen, A. (2005). Explanation: A mechanist alternative.
Studies in History and Philosophy of Biological and Biomedical Sciences,
36, 421–441.
Bickle, J. (2015). Marr and reductionism. Topics in Cognitive Science, 7(2),
299–311.
Bickle, J. (2016). Revolutions in neuroscience: Tool development. Frontiers in
Systems Neuroscience, 10, 1–13.
Bickle, J. (2018). From microscopes to optogenetics: Ian Hacking vindicated.
Philosophy of Science, 85(5), 1065–1077.
Bickle, J. (2019). Linking mind to molecular pathways: The role of experiment
tools. Axiomathes, 29(6), 577–597.
Bliss, T. V. P., & Lømo, T. (1973). Long-lasting potentiation of synaptic trans-
mission in the dentate area of the anaesthetized rabbit following stimulation
of the perforant path. The Journal of Physiology, 232(2), 331–356.
Boyden, E. S. (2011). A history of optogenetics: The development of tools for
controlling brain circuits with light. F1000 Biology Reports, 3(11), 1–12.
Boyden, E. S., Zhang, F., Bamberg, E., Nagel, G., & Deisseroth, K. (2005).
Millisecond-timescale, genetically targeted optical control of neural activity.
Nature Neuroscience, 8(9), 1263–1268.
Churchland, P. S. (1986). Neurophilosophy: Toward a unified science of the
mind-brain. Cambridge: MIT Press.
Craver, C. F. (2013). Functions and mechanisms: A perspectivalist view. In
P. Huneman (Ed.), Functions: Selection and mechanisms (pp. 133–158). Dor-
drecht: Springer.
Crow, T. J. (1972). A map of the rat mesencephalon for electrical self-­stimulation.
Brain Research, 36(2), 265–273.
Deisseroth, K. (2014). Circuit dynamics of adaptive and maladaptive behaviour.
Nature, 505(7483), 309–317.
Gardner-Medwin, A. R. (1969). Modifiable synapses necessary for learning.
Nature, 223(5209), 916–919.
Gold, I., & Roskies, A. L. (2008). Philosophy of neuroscience. In M. Ruse (Ed.),
Oxford handbook of philosophy of biology (pp. 349–380). Oxford: Oxford
University Press.
Hacking, I. (1983). Representing and intervening: Introductory topics in the
philosophy of natural science. Cambridge: Cambridge University Press.
Harris, E. W., Ganong, A. H., & Cotman, C. W. (1984). Long-term potentia-
tion in the hippocampus involves activation of N-methyl-D-aspartate recep-
tors. Brain Research, 323(1), 132–137.
Hebb, D. O. (1949). The organization of behavior: A neuropsychological the-
ory. New York: Wiley.
Jennings, J. H., Sparta, D. R., Stamatakis, A. M., Ung, R. L., Pleil, K. E., Kash,
T. L., & Stuber, G. D. (2013). Distinct extended amygdala circuits for diver-
gent motivational states. Nature, 496(7444), 224–228.
Lømo, T. (1966). Frequency potentiation of excitatory synaptic activity in the
dentate area of the hippocampal formation. Acta Physiologica Scandinavica,
68(S277), 128.
Morris, R. G. M. (1989). Synaptic plasticity and learning: Selective impair-
ment of learning rats and blockade of long-term potentiation in vivo by the
54 Gregory Johnson
N-methyl-D-aspartate receptor antagonist AP5. Journal of Neuroscience,
9(9), 3040–3057.
Morris, R. G. M., Anderson, E., Lynch, G. S., & Baudry, M. (1986). Selective
impairment of learning and blockade of long-term potentiation by an N-meth-
yl-D-aspartate receptor antagonist, AP5. Nature, 319(6056), 774–776.
Morris, R. G. M., Davis, S., & Butcher, S. P. (1991). Hippocampal synaptic
plasticity and N-methyl-D-aspartate receptors: A role in information storage?
In M. Baudry & J. L. Davis (Eds.), Long-term potentiation: A debate of cur-
rent issues (pp. 267–300). Cambridge: MIT Press.
Morris, R. G. M., & Kennedy, M. B. (1992). The Pierian Spring. Current Biol-
ogy, 2(10), 511–514.Olds, J., & Milner, P. (1954). Positive reinforcement pro-
duced by electrical stimulation of septal area and other regions of rat brain.
Journal of Comparative and Physiological Psychology, 47(6), 419–427.
Rodriguez-Romaguera, J., Ung, R. L., Nomura, H., Otis, J. M., Basiri, M. L.,
Namboodiri, V. M. K., Zhu, X., Robinson, J. E., Munkhof, H. E. van den,
McHenry, J. A., Eckman, L. E. H., Kosyk, O., Jhou, T. C., Kash, T. L., Bru-
chas, M. R., & Stuber, G. D. (2020). Prepronociceptin-expressing neurons
in the extended amygdala encode and promote rapid arousal responses to
motivationally salient stimuli. Cell Reports, 33(6), 108362.
Silva, A. J., Paylor, R., Wehner, J. M., & Tonegawa, S. (1992a). Impaired spa-
tial learning in alpha-calcium-calmodulin kinase II mutant mice. Science,
257(5067), 206–211.
Silva, A. J., Stevens, C. F., Tonegawa, S., & Wang, Y. (1992b). Deficient hippo-
campal long-term potentiation in alpha-calcium-calmodulin kinase II mutant
mice. Science, 257(5067), 201–206.
Stamatakis, A. M., & Stuber, G. D. (2012). Activation of lateral habenula in-
puts to the ventral midbrain promotes behavioral avoidance. Nature Neuro-
science, 15(8), 1105–1107.
Stamatakis, A. M., Sparta, D. R., Jennings, J. H., McElligott, Z. A., Decot,
H., & Stuber, G. D. (2014). Amygdala and bed nucleus of the stria terminalis
circuitry: Implications for addiction-related behaviors. Neuropharmacology,
76, 320–328.
Stamatakis, A. M., Van Swieten, M., Basiri, M. L., Blair, G. A., Kantak, P., &
Stuber, G. D. (2016). Lateral hypothalamic area glutamatergic neurons and
their projections to the lateral habenula regulate feeding and reward. Journal
of Neuroscience, 36(2), 302–311.
Stuber, G. D. (2016, September 28). Functional and molecular dissection of a
neural circuit for anxiety [Conference presentation]. DECODE Summit, Palo
Alto, CA. https://youtu.be/lj4_54w91wI.
Stuber, G. D., Hnasko, T. S., Britt, J. P., Edwards, R. H., & Bonci, A. (2010).
Dopaminergic terminals in the nucleus accumbens but not the dorsal striatum
corelease glutamate. Journal of Neuroscience, 30(24), 8229–8233.
Tonegawa, S. (2001). Roles of hippocampal NMDA receptors in learning and
memory. Riken BSI News, 12. https://bsi.riken.jp/bsi-news/bsinews12/no12/
issue1e.html.
Tsai, H.-C., Zhang, F., Adamantidis, A., Stuber, G. D., Bonci, A., de Lecea, L.,
& Deisseroth, K. (2009). Phasic firing in dopaminergic neurons is sufficient
for behavioral conditioning. Science, 324(5930), 1080–1084.
 Tools, Experiments, and Theories 55
Tsien, J. Z., Huerta, P. T., & Tonegawa, S. (1996). The essential role of hippo-
campal CA1 NMDA receptor–dependent synaptic plasticity in spatial mem-
ory. Cell, 87(7), 1327–1338.
Ungerstedt, U. (1971). Adipsia and aphagia after 6-hydroxydopamine induced
degeneration of the nigro-striatal dopamine system. Acta Physiologica Scan-
dinavica, 82(S367), 95–122.
van Zessen, R., Phillips, J. L., Budygin, E. A., & Stuber, G. D. (2012). Activa-
tion of VTA GABA neurons disrupts reward consumption. Neuron, 73(6),
1184–1194.
Wise, R. A., Spindler, J., deWit, H., & Gerberg, G. J. (1978). Neuroleptic-­
induced “anhedonia” in rats: Pimozide blocks reward quality of food. Sci-
ence, 201(4352), 262–264.
3 Science in Practice in
Neuroscience
Cincinnati Water Maze in the
Making
Nina A. Atanasova, Michael T.
Williams and Charles V. Vorhees

1 Introduction
Historically, philosophy of science was largely preoccupied with ques-
tions regarding the nature of scientific theories, what makes one theory
better than another, and what implications scientific theories have for
traditional philosophical questions. Philosophers conceptualized scien-
tific change in terms of theory change. Theory was considered a finished
product independent from its social and institutional context. However,
the process of theory in the making has yet to be carefully studied in
philosophy of science.
In our view, an integrated approach viewing science as a human prac-
tice embedded in a historical and institutional context should be ad-
opted for the study of multiple relevant factors involved in the making
of theories. This approach requires zooming out of the narrow view of
science which reduces it to its theories and looking at the bigger picture
of science as a practice, which generates theories along with technolog-
ical, medical, and social innovations motivated by societal needs and
available funding. Here we present a case study from contemporary ex-
perimental neuroscience which focuses on the routine practices of grant-
funded research and tool development which drive the research process
in biomedical science. We propose an account of science in which the-
ories serve as interpretative devices of experimental data. This result
contradicts some philosophical accounts of science that tend to reduce
it to its theories. By theory, we mean abstract models and narratives
which postulate comprehensive ontologies for the explanation of target
phenomena and the mechanisms that produce them.
Our approach recognizes the multiple functions of theories such as
prediction, explanation, and modeling, which have been extensively an-
alyzed in philosophy of science. While we recognize the importance of
all these functions, we show that the role theory plays in contempo-
rary neuroscience is as a device for interpretation of data generated by
exploring focused hypotheses about specific interventions, for example,
lesions of the hippocampus and their corresponding deficits in spatial

DOI: 10.4324/9781003251392-5
 Science in Practice in Neuroscience 57
navigation. Contemporary neuroscience is data-driven. Scientific inno-
vation consists in the development and consolidation of ensembles of
experimental tools and techniques. It depends largely on the availability
of funding and relevant technology.
Our goal is to present a case study which utilizes an integrated meth-
odology in which philosophers and scientists reflect together on the prac-
tice of science. The authors are a philosopher and two neuroscientists in
whose lab the philosopher worked while a graduate student. Although
the majority of the narrative is a first-person recount of the two scien-
tists, we use a third-person structure in order to ease the flow of the text.
In what follows, we provide a brief characterization of two traditional
philosophical views of science which we challenge. We then turn to the
philosophy of science-in-practice. Consistent with the turn to the philos-
ophy of science-in-practice, philosophers of neuroscience have focused
their attention on experimentation and tool development (Bickle 2006;
Sullivan 2009; Silva et al. 2014; Atanasova 2015). This motivated our
choice to trace the invention, development, and the continuous refine-
ments of an experimental tool in neuroscience, the Cincinnati water
maze (CWM). The case demonstrates the significance of the turn to the
philosophy of science-in-practice by exemplifying the notion of research
repertoires advanced by Leonelli and Ankeny (2015) and Ankeny and
Leonelli (2016). We show how factors such as availability of funding,
technology, and collaborations determine many rational choices in re-
search practice. Finally, we introduce the notion of theories as interpreta-
tive devices to capture the role of theory in contemporary neuroscience.

1.1 Postpositivist Views of Science and Scientific Change

Early-20th-century philosophy of science was dominated by logical
positivism which viewed empirical science as the most rigorous source
of knowledge. Scientific knowledge, according to logical positivism,
is based on empirical observations of the natural world that result in
law-like generalizations. Scientific explanations consist of identifying
generalizations from which the observed phenomena could be logically
derived. Theories, in this view, do not serve explanatory functions, they
are merely useful prediction devices. The success of science, according to
logical positivism, is measured by the success of its predictions.
By mid-20th century, many of the core beliefs of logical positivism
were challenged. Postpositivism took two forms: scientific realism and
scientific relativism. Scientific realists (e.g., Putnam 1976; Boyd 1980)
challenged the claim of positivism that scientific theories are merely pre-
diction devices. If they are so successful in predicting events, scientific
realists noted, theories must be getting something right about the un-
observable world. Scientific theories tell us truths about the world even
58 Nina A. Atanasova et al.
if we cannot directly observe all entities and processes they postulate.
This inspired many to treat scientific theories as explanatory devices
with comparable significance for both science and philosophy. Thus,
the borders between theoretical science and metaphysics were blurred
(Chakravartty 2007).
Scientific relativists (e.g., Kuhn 1970; Feyerabend 1988/1975), on the
other hand, criticized both positivists and realists for their conceptu-
alization of science as a rational activity detached from the lives of the
human beings who practice it. They focused on science as a social ac-
tivity and studied human interactions among scientists instead of the
relationship between theories and the world. Nevertheless, theories play
a central role in this account of science.
Inspired by the notion of the theory-ladenness of experiment and obser-
vation (Hanson 1965), Kuhn and Feyerabend claimed that the practice of
science is defined by the theories adopted by different scientific commu-
nities. Theories, from this view, function as local ontologies which define
the worldviews of the communities which adopt them. Thus, scientific dis-
agreement is attributed to incompatible worldviews characteristic of rival
communities. Since the proponents of this view believe that there is no
objective way to determine that one ontology is better than another, all
ontologies and corresponding theories are equally legitimate. Therefore,
scientific change, according to relativists, consists in switching from one
worldview to another. It is a radical replacement of one theory with another.
Both kinds of postpositivist approaches have rendered valuable in-
sights. However, taken to an extreme, both present a distorted image of
science. On one end of the spectrum, naïve scientific realists have a dog-
matic attitude toward science as a flawless source of ultimate knowledge
about the world. On the other end, radical relativists leave no room for
science as a specialized source of knowledge. This divide culminated in
the infamous Science Wars of the 1990s which drove a wedge between
intellectuals from the humanities and social sciences on one side and
natural scientists on the other (Gross and Levitt 1994; Gross et al. 1996;
Ross 1996). Philosophy of science has since carved space for a middle
ground. A number of philosophers of science now look both at the the-
ories generated in science and at the interactions between scientists. Phi-
losophy of science has turned to studying science-in-practice.

1.2 The Philosophy of Science-in-Practice

Postpositivist philosophy of science was preoccupied with scientific the-
ories. While we take issue with naïve realism which assumes, often un-
critically, the epistemic authority of science, we nevertheless assume that
there are facts of the matter regarding the way the world is. However,
judging which theory is better than another, in our view, depends on
judgments about the validity of the methods used to generate that theory.
 Science in Practice in Neuroscience 59
It matters, for example, how scientists come to know the mechanism of
long-term potentiation in order for discussions about its role in memo-
ry-formation to be justified. If the phenomenon of long-term potentiation
is not reproducible across laboratories and experimental conditions, if it
is not reliably associated with behaviors indicative of memory-­formation,
we would not be able to make sense of the discussions about its causal
role in the process (Dringenberg 2020). This is why it is important to
look behind the scenes of knowledge production to understand the prac-
tice that makes it possible. Studying the practice of science is thus as
important as studying its products, the theories. This realization is at the
core of the turn to the philosophy of science-in-practice.

1.2.1 The Society for Philosophy of Science in Practice

The philosophy of science-in-practice traces its roots to 20th-century phi-
losophy of experiment (Hacking 1983). However, it was not until the mid-
2000s that the approach matured as a distinct methodology of philosophy
of science. It was institutionalized with the initiation of the Society for Phi-
losophy of Science in Practice (SPSP), founded in 2006 by Rachel Ankeny,
Mieke Boon, Marcel Boumans, Hasok Chang, and Henk de Regt. SPSP
held its founding meeting at the biennial meeting of the Philosophy of
Science Association in Vancouver. The idea for SPSP was conceived of the
previous year at the conference “Philosophical Perspectives on Scientific
Understanding” in Amsterdam. According to SPSP’s mission statement:

Philosophy of science has traditionally focused on the relation be-

tween scientific theories and the world, at the risk of disregarding
scientific practice. In social studies and technology, the predominant
tendency has been to pay attention to scientific practice and its rela-
tion to theories, sometimes willfully disregarding the world except
as a product of social construction. Both approaches have their mer-
its, but they each offer only a limited view, neglecting some essential
aspects of science. We advocate a philosophy of scientific practice,
based on an analytic framework that takes into consideration the-
ory, practice, and the world simultaneously.
(SPSP Mission Statement)

The society has grown and given rise to a trend surpassing the frame-
work of postpositivism. Prominent leaders of the society, Rachel Ankeny
and Sabina Leonelli, introduced the notion of research repertoire to ac-
count for the collaborative practices in the contemporary life sciences.
They define research repertoire as follows:

…a distinctive and shared ensemble of elements that make it prac-

tically possible for individuals to cooperate, including norms for
60 Nina A. Atanasova et al.
what counts as acceptable behaviours and practices together with
infrastructures, procedures and resources that make it possible to
implement such norms.
(Leonelli and Ankeny 2015, p. 701)

…well-aligned assemblages of skills, behaviors, and material,

social, and epistemic components that groups may use to practice
certain kinds of science, and whose enactment affects the methods
and results of research, including how groups practice and manage
research and train newcomers.
(Ankeny and Leonelli 2016, p. 20)

Our case study exemplifies the enactment of one such repertoire in be-
havioral neuroscience. It shows the superiority of this account of science
over the, still influential, notion of scientific paradigm first advanced by
Thomas Kuhn and Paul Feyerabend in the 1960s–1970s.

1.2.2 Post-Kuhnian Philosophy of Science

The notion of repertoire was inspired by the realization of the limita-
tions of postpositivist philosophy of science. It was introduced as an al-
ternative to Kuhn’s (1970 [1962]) notion of a paradigm. Acknowledging
Kuhn’s equivocal use, we take paradigm to refer to the idea that a scien-
tific community is defined in terms of a shared ontology, shared ways of
formulating hypotheses, and standardized experimental methodologies.
The knowledge claims formulated by such a community are context-de-
pendent, bound by the corresponding ontological assumptions. This
often precludes members from one community to have meaningful ex-
changes with members of other scientific communities. This dissonance
ultimately precludes the integration of scientific knowledge generated by
different communities. Recent recounts of this concern include Sullivan
(2009), Longino (2014), and Hochstein (2016).
However, the philosophy of science-in-practice has brought about
post-Kuhnian alternatives to the stage. Ankeny and Leonelli’s notion of
scientific repertoire captures the interactions of local epistemic commu-
nities belonging to a single institution as well as large consortia which
may be spread out into a global network of collaborations.

1.2.3 Philosophy of Neuroscience in Practice

In this new line of philosophical studies of science, a number of phi-
losophers of science have engaged with analysis of current experimen-
tal practices in neuroscience (Atanasova 2015; Bickle 2016; Robins
2016; Sullivan 2016). With this research came the recognition of the
central role that tool development and innovation play in neuroscience.
 Science in Practice in Neuroscience 61
Neuroscientific practice is experiment-driven. Theory often catches up
after critical data are generated and are in need of synthesis and inter-
pretation. Experiments are often inspired by the available technology
rather than inspiring the design of novel technologies themselves. Many
decisions regarding what experiments to design and run are based on
pragmatic considerations (personnel, facilities, and funding).
Tool development and improvement thus need to take center stage
in the philosophy of neuroscience. The case study we offer is meant to
do justice to the epistemic and pragmatic problems through which tool
development progresses. The case study exemplifies the fundamental
role of the calibration strategy identified by Atanasova (2015). The story
about the development and refinement of the CWM we tell below is a
narrative about its calibration as a tool for the study of cognitive func-
tions and deficits in rats.
This case study exemplifies the dynamic interplay between tool de-
velopment, experiment, and theory. We aim to show that even though
theoretical considerations and background assumptions are implicit in
experimental design, this is not to say that comprehensive theories nec-
essarily guide experiments. Theoretical constructs such as “learning”
and “memory” are often very loosely defined and are rather fluid con-
cepts. They serve as placeholders in experimental hypotheses and are
under continuous reinterpretation and refinement in light of experimen-
tal results. As such they function like experimental tools (Feest 2010).
We thus maintain that while comprehensive theories about learning and
memory do get articulated, they often follow rather than precede the
experiments which support them.

2 The Cincinnati Water Maze in the Making

The CWM is an apparatus used for testing rat egocentric (based on in-
ternal bodily cues) learning and memory when performed in darkness
and requiring learning a sequence of turns to navigate to a goal (see
Figures 1a and b). In the beginning, the CWM was used to study a mix-
ture of egocentric and allocentric (based on external cues) learning and
navigation when rats were run under standard room lighting but has
since been modified as a test of egocentric memory complementary to
the widely utilized MWM which tests allocentric learning. The CWM
was introduced by Vorhees (1987) and has evolved continuously in the
Vorhees/Williams Lab at Cincinnati Children’s Research Foundation
ever since. The maze consists of a series of interconnected T-shaped cul-
de-sacs through which a rat navigates from a fixed start to an invariant
goal. The maze is partly submerged in water and the task is for rats to
find by trial and error the escape platform. The motivational factor in
this task is the natural aversion of rats to water. The CWM was inspired
by a simpler maze, the Biel water maze (BWM).
62 Nina A. Atanasova et al.

Figure 1a D
 rawing of Cincinnati water maze featuring the training path which
begins at S (start) and ends at G (goal). The two points are reversed
during reversal tests.

Figure 1b Photograph of Cincinnati water maze.

 Science in Practice in Neuroscience 63
2.1 The Biel Water Maze
The BWM was introduced by William Biel (1940). The apparatus was
used to explore developmental changes in the ability of young rats
to learn a complex task. Rats were trained to navigate by swimming
through a maze to reach an escape ladder. The maze consisted of a long
straight channel along one side used for preliminary training and a series
of five shorter T-shaped cul-de-sacs. Rats were placed in the starting
position and tested until they located the goal at the end of the long arm
where there was a ladder which afforded escape from the water. Rats
were given training trials for 2.5 days. Rats received one trial in the
straight arm in the morning and four maze training trials in the after-
noon. The next day they received five more training trials. There were 24
maze trials divided among two test sessions each day, one in the morning
and one in the afternoon. The first maze trial was in the morning after
the last training day and consisted of one trial and two trials in the af-
ternoon, followed by two morning and two afternoon trials for the next
five days ending with a single trial on the last day. Trials ended when the
rat found the escape ladder, but if it did not find it within ten minutes,
it was removed. Biel recorded the number of errors (not specified what
counted as an error), trials to reach criterion (criterion not specified), and
latency to reach the goal.
The maze was later used by Polidora and colleagues in a series of ex-
periments testing rat cognition. Polidora et al. (1963) aimed to “devise
a learning task, simple enough for rapid mastery by rats, but difficult
enough for reliable detection of relatively small differences in learning
ability” (pp. 817–818). They tested rats for five days. Rats were given
five pre-test trials in the straight channel on the first day. For the next
three days, rats were given seven trials per day in the maze. On the fifth
day, they were given five post-test trials in the straight channel. Rats
were given ten minutes of rest between trials. Polidora and colleagues
recorded transit time and errors (whole-body entries into a blind alley).
A few years later, the apparatus was utilized in a series of experiments
exploring behavioral and biochemical consequences of exposure to sev-
eral neurotoxins, after either pre- or postnatal exposure, by Richard
Butcher’s Lab at Cincinnati Children’s Research Foundation throughout
the 1970s. Butcher et al. (1970) used different procedures in which, in
addition to the training in the straight channel and the original start-
to-goal path in the maze, tested rats in a path which reversed the start
and the goal points of the maze after a number of trials in the original,
forward, path. Rats were given five trials in the straight channel on the
first day of the experiment. For the next two days, rats were given five
trials a day. Over the last two days, rats were tested on the reverse path
64 Nina A. Atanasova et al.
for five trials a day. Rats were removed from the maze if they did not find
the goal within ten minutes. Time and errors (whole body entries into
blind alleys) were recorded.
In 1976, Dr. Vorhees joined Dr. Butcher’s lab. While running exper-
iments using the BWM, Vorhees noticed that rats, treated in various
ways (experimental groups), tended to perform worse on the reverse
path compared with control groups, even though there were no differ-
ences on the forward path. This prompted him to investigate why this
happened. After numerous observations of how rats went through the
maze and made decisions, he noticed that if a rat was in the forward
path, it was forced to make a decision only at the end of a channel when
it encountered a perpendicular wall, where it had to turn right or left.
This was an inherent feature of the design of the maze. However, if the
rat was in the reverse path and swam until encountering a wall, it was
always a dead-end cul-de-sac, such that no matter whether it turned
right or left, it reached another dead-end. To advance through the maze,
the rat had to backtrack and take a side channel or anticipate the dead-
end and turn before it reached the end of a channel. Vorhees conjectured
that learning this anticipatory turn might be the reason why it was more
difficult for rats to solve the task in the reverse path. He hypothesized
that this more difficult aspect of the reverse path might be more sensitive
to subtle differences between treatments. If so, treatments might show
more pronounced effects if the maze was made more complicated and
had more anticipatory turns.

2.2 Introduction of the Cincinnati Water Maze

By 1987, Vorhees was the principal investigator of the lab. He set out
to explore his hypothesis that a more complex maze would increase
the sensitivity of the test. This is when he designed the CWM (Vorhees
1987). The new maze was made of nine T-shaped channels, as opposed
to five plus the long arm in the BWM. The channels were wider than the
BWM to accommodate adult rats and the long arm was eliminated. Vor-
hees compared the sensitivity of the two maze designs to cognitive defi-
cits occurring in rats after prenatal exposure to phenytoin. The choice
of drug was made on the basis of the well-established neurotoxicity of
the compound. He tested four groups of rats, one control and three with
exposure to different doses of the drug. The results showed dose-depen-
dent cognitive impairments in both mazes. However, the differences in
performance between groups were significantly larger in the CWM.
This approach to establishing the validity of a behavioral test in
neurobiology is analyzed in Atanasova (2015) where she refers to it as
calibration. Behavioral tests such as the CWM are typically used in com-
bination with other tests such as locomotor ability, anxiety, and vision.
These tests are known as test batteries. For each test in the battery, the
 Science in Practice in Neuroscience 65
rest serve as calibration devices. Some of them are meant to establish
prerequisite abilities for the test of main interest and others are used
for detection and elimination of artifacts of the procedure. When a test
is established within this larger context, it gets validated as an animal
model of some human ability, disorder, or other research-worthy condi-
tion. Further, when a new experimental design, or tool, is introduced,
neuroscientists calibrate it by testing it against factors that are known
to produce certain effects under circumstances similar to the test con-
ditions. For example, when a new animal model of depression is intro-
duced, it is run with antidepressants that are known to be effective in the
treatment of the symptoms of depression. Once it is established that anti-
depressants have effects that correspond to the modeled mood disorder,
the animal model is used to assess novel compounds for effectiveness as
antidepressants.
Similarly, Vorhees took a compound, phenytoin, known to produce
cognitive impairments and a tool, the BWM, which was established as
a valid test of cognitive impairments and redesigned it to make it more
useful. The validity of the test refers to the capacity of the tool to accu-
rately detect positive control agents known to be detrimental, but also
to accurately identify negative control agents as not toxic (Vorhees 1987,
p. 239). The BWM was successful at detecting the effects of known neu-
rotoxins. Vorhees took the new tool, the CWM, and tested it against an
established positive control compound. The new tool detected the same
effects better than the old one, and in later experiments, it detected ef-
fects which the old apparatus did not detect. This allowed Vorhees to
conclude that the new tool was more sensitive than the old one.

2.3 Trials and Errors

Michael Williams joined the lab in 1997. The CWM was continuously
used to study the cognitive deficits resulting from exposure to various
drugs and neurotoxins. Consistent with the calibration strategy, Vor-
hees and Williams used test batteries. Among the tests they commonly
employed was the MWM introduced by Richard Morris (1981). The test
is the most widely utilized spatial learning and memory test in neurobi-
ology. The MWM consists of a large circular tank filled with water with
a submerged escape platform (see Figures 2a–c). Animals are trained to
navigate to the platform using distal cues on the walls of the testing room
and away from the edge of the tank. There are numerous versions of the
installation and control conditions under which the test is run (Vorhees
and Williams 2006). In the Vorhees/Williams lab, the MWM is used as
a test complementary to the CWM in the sense that the MWM is used to
test allocentric navigation and the CWM – egocentric navigation.
The two instruments share a number of similarities. For example,
both detect cognitive deficits. However, while the CWM detects deficits
66 Nina A. Atanasova et al.

Figure 2a D
 rawing of Morris water maze featuring different positions of the
escape platform during different training and testing phases.

Figure 2b Photograph of Morris water maze.

in sequential/egocentric learning, the MWM detects deficits in spatial

navigation. Running multiple experiments in both mazes, Vorhees and
Williams noticed that after some treatments the performance in both
tasks was disrupted (Morford et al. 2002; Williams et al. 2003), but
 Science in Practice in Neuroscience 67

Figure 2c S creenshot of a bird’s-eye view of Morris water maze through a

recording camera.

after other treatments, the performance in one but not the other was
impaired (Williams et al. 2002, 2003). Vorhees and Williams realized
that such data pointed to a corresponding difference in the brain areas
and the neurotransmitters that were involved in solving the two tasks.
This inspired theoretical questions, but the maze would undergo sev-
eral modifications in response to calibration issues before they could be
addressed.
Vorhees and Williams discussed what makes the CWM unique and
what brain regions were involved in solving it. They knew that distal
cues are important for spatial memory, but the room that housed the
CWM had no prominent cues but had many standard room features
(such as air intake ducts and electrical outlets on the wall). Williams
wondered, “Are the rats using background spatial cues as well as prox-
imal cues in the CWM?” The task had presumably been solved on the
68 Nina A. Atanasova et al.
basis of egocentric route-based proximal cues. So, they explored ways to
eliminate spatial cues and the first thing they tried was hoods to cover
the rats’ eyes so that they could not see distal cues. Whishaw and Maas-
winkel (1998) made hoods to obstruct rats’ vision using a large circular
platform maze where they first found food visually, then had to re-find
it blindfolded.
Inspired by this publication, a student in the Vorhees/Williams lab,
Tracy Blankemeyer, made rat hoods. However, when the rats were
tested in the maze and the hoods got wet, they pried the hoods off, then
searched for the goal. Efforts to get the rats accustomed to the hoods
also failed by having the rats wear them for a week prior to maze testing.
The rats did adjust to the hoods and left them on but when put in the
maze, they again pried them off.
Therefore, Vorhees and Williams considered surgically blinding the
rats but found this objectionable. Neither thought they could sew the
eyelids shut even though this is a published procedure. They settled on a
more humane solution: they turned off the lights but provided a red light
because rat vision is reported to be deficient at the red end of the visible
spectrum. The maze was harder than under white light, but the rats still
solved the task rapidly. This suggested that there was enough light that
the rats were still able to see room cues. At that time, it was known that a
single landmark can be sufficient to solve an MWM. Therefore, Vorhees
and Williams thought rats might be using even minimal red light as an
orientation beacon to navigate to the goal.
This led to the decision to use infrared light. In this setup, rats per-
formed worse than under red light, but they still appeared to orient to
some unidentified light source. It turned out that the light leaking un-
derneath and around the door was sufficient for the rats to orient. A hint
that it was the door came out of experiments a graduate student, Nicole
Herring, did. She found that rats treated with methamphetamine had
CWM deficits (Herring et al. 2008). However, when she attempted to
replicate the results, the effect was smaller. Something had changed. Af-
ter inspection, it was determined that in the second experiment the door
was not closing fully, providing more light than in the first experiment.
Therefore, trim was installed around the door and the closer was fixed
so no light could get in the room. The principal investigators stood in
the room making sure after dark adaptation that nothing could be seen.
When experimenters entered the room to test a rat, they carried a red-
light flashlight to see enough to place rats in and out of the maze without
interfering with the dark adaptation of rats waiting to be tested. The dif-
ficulty rats had in solving the task under complete darkness meant that
they were no longer using visual cues for navigation. Performance was
monitored using an infrared-sensitive camera mounted above the maze
connected to a monitor in an adjacent room where the experimenter
scored performance. The tool was now calibrated to detect the effects
 Science in Practice in Neuroscience 69
which the investigators studied. In this process, there was little theoret-
ical speculation about what the effects of the experimental procedures
would imply for the nature of egocentric learning and memory, although
tacit knowledge about the behaviors of rats under experimental condi-
tions informed procedural changes. Some of the findings were accidental
others were predicted.
Another inadvertent discovery was the importance of training. This
came about when Vorhees and Williams helped another lab to set up
a CWM by providing them with drawings for the maze. Later, the lab
called Vorhees stating that the CWM was not working and that all rats
failed to learn even when tested under normal light. The other lab had
only asked for drawings, not test procedures. Vorhees had not provided
information on the straight channel trials under the assumption they
would later ask for procedures, but they never did. Hence, their rats
did not learn that escape was possible and because the CWM even in
the light is complex, and rats gave up searching without ever finding
the escape. The tool was not properly calibrated, not all procedures for
its proper use were followed. This is when Vorhees appreciated that the
straight channel was essential to teach rats about escape. Without this
knowledge, the rats became frustrated and would give up and tread wa-
ter until removed. Hence, it became clear that minimal training was
required for rats to solve a maze as difficult as the CWM. Until then,
running the rats through the straight channel, just like Biel did, was
considered an assessment of the rats’ swimming capacities before they
went into the maze, not a learning experience with positive transfer of
training to the maze. It was meant to establish that rats were proficient
swimmers and motivated to escape the water, and ensure they had no
motoric impairment if they were in the treated group. This procedure
accounted for the possibility that a treatment affecting swim speed could
confound the interpretation of group differences obtained in the maze.
Because the straight channel was used in the lab to obtain performance
(swim speed) information, it was not initially appreciated that it was
an essential part of the procedure. Employing the test for this purpose
is part of the calibration strategy, but it became clear that it was a nec-
essary component of the experimental setup. This shows that tacit pro-
cedural knowledge can make a difference in the way research develops
without the employment of a theory about the process which underlies
the studied behaviors. The causal connection between this element of
the procedure and the observed behavior was articulated after the fact.
It was not a hypothesis that was tested purposefully.
An even more recent change in the maze design was the introduc-
tion of a tenth T cul-de-sac. Williams noticed something that Vorhees
overlooked. In Vorhees’ drawing, one wall was slightly out of alignment
limiting the maze to nine cul-de-sacs. However, once the alignment was
corrected, there was room for an additional T. Recent data from the new
70 Nina A. Atanasova et al.
design show that the change made a difference in maze difficulty and it
now requires more trials for rats to become proficient. The rats’ optimal
performance in the ten-T version for the equivalent number of testing
days is not as good as in the nine-T version.
More changes were introduced in response to calibration problems.
For example, initially, only the errors in the arms of a T at the end of a
cul-de-sac were counted. Now stem and T-errors are counted in recog-
nition of the fact that all turns not leading to the goal are errors when-
ever a rat deviates from a direct path to the end regardless of whether it
goes the full distance to the end of a right or left arm of a T. Moreover,
counting only the end of cul-de-sac errors was not capturing valuable
information about mistakes rats made. As rats learn, one of the interme-
diate steps they go through is recognizing when they make a wrong turn
and swim down the stem of a T. Once the rats have visited that dead-
end T several times, thereafter when they turn in down that stem, they
recognize that they have been there before. The rat then stops short and
reverses course. Counting only T-arm errors missed these intermediate
errors. Adding stem errors further increased the sensitivity of the test.
The change in error counting occurred at a lab meeting when it became
clear that group differences were larger on some occasions depending on
how different experimenters counted errors. They needed to calibrate
experimenters as well as definitions of errors more precisely. This is not
trivial because when rats begin to learn the maze, they progressively in-
hibit errors. At first, rats go all the way to the end of a T and turn right
and left, and sometimes go back and forth before exiting. Later, rats
hesitate partway down the stem before turning around. These ‘inter-
mediate’ mistakes were being missed. The lab considered whether arm
errors were in a sense more severe than stem errors. To find out, the next
experiment counted both types of errors. Analysis of the data provided
no support for the idea that arm errors represent a more severe deficit
than stem or arm + stem errors in revealing group differences. Therefore,
all errors are now treated the same.
In addition to errors, experimenters track the time it takes for a rat
to find the platform (latency). However, there is a per-trial time limit
of five minutes designed to prevent fatigue. This aspect of the test pro-
cedure also evolved. Initially, the time limit was 12 minutes. Rats were
given two trials per day as before. However, rats that did not find the
goal after 12 minutes would get exhausted. This caused rats to become
tired faster on trial-2. The first approach to this problem was to give rats
reaching the 12-minute limit, a one-hour rest. Later, the trial limit was
reduced to ten minutes and the rest after a trial-1 failure was shortened
to 30 minutes. Over a period of several years, the limit was gradually
reduced from 10 to 6, and finally five minutes.
Some of the adjustments were also made in response to peer criticism.
Critics pointed out that the resting time between trials might not be
 Science in Practice in Neuroscience 71
enough and, if insufficient, would compromise performance on trial-2
of each day. In response, Vorhees and Williams tested performance with
rest intervals up to 30 minutes after a trial-1 failure to reach the goal
and found it made little difference unless the trial time limit was shorter
than five minutes. Therefore, five minutes was empirically determined
and became standard practice. The data from these refinement exper-
iments were not published separately because they were not sufficient
by themselves to be worthy of a paper even though they were essential
in the evolution of the method and are found within papers published
across many years.

2.4 The CWM in Practice

The apparatus has become a standard tool for testing egocentric route-
based learning and memory. It is an integral part of the test battery used
in the Vorhees/Williams lab to measure behavioral responses to drugs,
environmental agents, or gene mutations that cause neurological and
cognitive dysfunction. The use of the CWM has followed three paths in
its development, all of which are independent in their goals but intersect
and reinforce the ultimate success of the task.

2.4.1 Grant-Funded Research

A recent project in the Vorhees/Williams lab is a study of the cognitive
effects of proton exposure. It is a research question which arises in the
context of cancer treatment. The motivation is to find what treatments
minimize lasting cognitive side-effects from the treatment of brain tu-
mors. Oncologists know that people treated for brain tumors by X-ray
or proton radiation often have long-term cognitive problems. The ques-
tions they are addressing in normal rats include: “Do protons produce
fewer or different long-term cognitive effects than X-rays, and do dif-
ferent dose, dose-fractions, and dose-rates produce fewer deficits than
others?” The hypothesis is, since protons kill tumors more effectively,
maybe they also produce less collateral damage to surrounding tissue
than X-rays. Vorhees and Williams, working with oncologists, radiation
biologists, and medical physicists, developed an experimental plan to
test this hypothesis.
The way this works in practice is that the lab has tests of different
cognitive functions. They then use this toolset in different settings to
answer different experimental questions. In this case, by exposing rats to
protons and testing them in the CWM, MWM, and other tests in their
armamentarium, they can provide more comprehensive answers to the
questions of interest. The experiments are meant to test a focused hy-
pothesis about the cognitive effects of proton exposure. The researchers
do not evoke an overarching theory of cognition. They operationalize
72 Nina A. Atanasova et al.
cognitive deficits as deficits in behaviors observed under the experimen-
tal setup in response to a specific intervention. The toolkit is the crucial
factor in this process.
When the lab sets out to study a drug, environmental agent, or gene
mutation, they start by looking at the effect of the variable with the tools
that they know since they know the brain regions that mediate each test.
This provides first-pass information not only on what brain regions are
affected but also provides clues to which neurotransmitters and recep-
tors are likely affected. Once they generate preliminary data, they write
grant proposals to supply the means for further study of the effects of
the treatments.

2.4.2 Basic Research

Despite the fact that the majority of the choices made for the projects
carried on in the lab depend on grant funding, Vorhees and Williams
are driven by the quest for knowledge. For example, once Vorhees and
Williams eliminated distal cues that rats could use for navigation in the
CWM, they inferred that the rats’ ability to use the hippocampus for
spatial navigation in that task was discounted. This assumption was jus-
tified because the role of the hippocampus in spatial learning was well
established. The next issue became to find out what part of the brain was
used in solving the CWM. There were data published showing that the
striatum is involved in path integration which is an overlapping ability
with egocentric navigation. The next step, thus, was to lesion the stria-
tum without affecting motor behavior and see if this created deficits in
CWM learning.
When Amanda Braun entered the lab as a graduate student, she tested
the role of striatal dopamine on CWM learning. Her experiments pro-
duced data showing that rats with dopaminergic striatal lesions per-
formed much worse in the CWM than sham controls. However, to
test the specificity of the effect, the rats were also tested in the MWM.
Learning in the MWM was also impaired but less than in the CWM
(Braun et al. 2012). Next, Braun lesioned subregions of the striatum to
get a more specific idea of what regions were more important than oth-
ers. She did four experiments, one for each regional lesion. She lesioned
the dorsal striatum, the dorsal medial and dorsal lateral striatum, ven-
tral striatum (n. accumbens), or prefrontal cortex.
The data supported the idea Vorhees and Williams had hypothesized.
In these experiments, Braun again ran the rats in the CWM and MWM.
Entire dorsal striatal dopaminergic lesions produced effects on learning
in both mazes (Braun et al. 2012). This supported the notion that the two
regions are interconnected. So, it was not an absolute difference between
the two but the effect in the CWM was greater. When she lesioned the
medial vs. lateral dorsal striatum, she was able to differentiate between
 Science in Practice in Neuroscience 73
effects in the CWM vs. MWM (Braun et al. 2015). Hence, with the
entire dorsal striatum lesioned there were effects in both mazes whereas
lesions in subregions produced smaller or no effects on the MWM. They
also found that n. accumbens lesions impaired CWM learning, but pre-
frontal cortex dopamine lesions did not. This research was not extramu-
rally funded and yet it made a difference in how they thought about the
CWM compared with the MWM and other tests of learning and mem-
ory. It changed their approach to searching for other structure-function
cognitive relationships.
Braun’s experiments addressed concerns of other researchers about
the CWM as a unique tool. In the beginning, Vorhees and Williams’
peers would ask: “What does it mean? What are you measuring? Yes,
you get a difference among groups but what is the basis of the differ-
ence?” The initial lack of answers to these questions shows that the early
CWM experiments were not inspired by a preconceived theory about
what unobservable processes were producing the observed behaviors.
Braun’s results identified the principal brain regions involved in CWM
learning. Since the lesions were specific for dopamine, these studies im-
plicated dopamine as a key neurotransmitter in mediating CWM learn-
ing. Now a claim about the causal connection between affected brain
regions and behaviors was articulated. This eliminated many criticisms
because the results were empirically derived.
One of the projects Vorhees and Williams are considering is testing
the hippocampus’ involvement in CWM learning by lesioning the hip-
pocampus and running rats through both mazes since the MWM is
known to depend on the hippocampus and NMDA receptors but the
involvement of the latter for the CWM is not known. There are NMDA
receptors in the striatum too, but whether glutamate mediates egocen-
tric navigation circuits between the striatum and hippocampus is poorly
understood. These experiments would help elucidate the relationship be-
tween these forms of learning and memory. However, time and money
limit what can be done. Graduate students and postdocs have the luxury
to run experiments to test hypotheses based on curiosity and are less
constrained by the specific aims done for a funded grant. Principal in-
vestigators need to fulfill the aims of the grant and explore other ideas,
time and resources permitting.

2.4.3 Making Tools Better

The development of the maze was empirical, incremental, and based on
focused hypotheses at each step. It was motivated by trying to make
improvements in the utility of the apparatus to detect differences hy-
pothesized to be caused by different interventions. In other words, this
was a process of progressive calibration of the CWM. Method improve-
ments came along because experimenters needed something more than
74 Nina A. Atanasova et al.
the method provided at any given time; a tool that would do a better
job than in the past. Most changes resulted in the CWM working better
based on experimenter reasoning that was tested experimentally. When
a new idea suggested an improvement and the improved procedure
worked, the question became: Why does it work better? At this stage,
the apparatus inspired theoretical questions.
For example, Williams came up with a way to test the role of distal
cues (Vorhees and Williams 2014). The idea was to test one group of rats
in the dark and switch them to the light and another group in the light
and switch them to the dark. The results showed that when rats started
in the light and were switched to the dark, they made large numbers
of errors even though they knew the path to the goal. But if the rats
started in the dark, then were switched to the light, they had no dis-
ruption in learning. Having learned the maze in the dark the rats knew
the way without distal cues and adding these cues had little effect. The
rats may have made a cognitive map or formed a rule of how to solve
the maze. These are theoretical hypotheses. Rats learning under light
learned faster and made fewer errors, but rats used distal and proximal
cues to find their way. Once distal cues were eliminated, they did not
have enough proximal cue information to find the goal and had to do ex-
tensive relearning that proved highly disruptive. This was informative. It
showed that once rats developed an internal map, they were able to use it
regardless of vision-based cues. But for rats that used vision, when it was
eliminated, they had too few internal cues to fall back on because their
default solution relied more on the use of distal rather than internal cues.
When put in the dark, the rats were initially lost and made many errors.
The hippocampal spatial maps the rats developed no longer worked and
they were forced to switch to striatum-based learning. Hence, rats have
two navigational systems and when encountering a new situation, rely
on the system that is most efficient under the conditions they confront.
Generally, this is spatial rather than egocentric navigation. This is an
example of how the CWM started to generate theoretical insights. How-
ever, exploring theoretical hypotheses is still constrained by the dynamic
interplay between tool building and experimental design.
Most tests in the Vorhees/Williams lab are improved incrementally as
experimental work progresses. During a renovation, they were able to
build a larger, eight-foot diameter, MWM. The reasoning was derived
from what they had learned with the CWM: if the task is made more
difficult, perhaps it will be more sensitive to treatment effects. If their
hypothesis is correct, they would be able to detect treatment effects that
were previously not detected in a smaller diameter MWM, and they did.
Treatments that had not shown differences in six and seven-foot MWMs
showed significant effects in the eight-foot version. This was another
improvement that grew out of their experience in developing the CWM.
Note that the hypothesis they tested was not derived from a theory. It
was a hypothesis about how the experimental tool would interact with
 Science in Practice in Neuroscience 75
animal behavior. This shows how new theoretical hypotheses are in-
spired by the new possibilities afforded by improved tools.
The CWM is sensitive to perturbations of the striatum. And so, if a
treatment, whether pharmacological, lesion, or genetic, affects the stri-
atum, the CWM generally detects it. The procedures are still being re-
fined. One idea Vorhees and Williams are exploring is reversal learning;
a well-established phenomenon in learning theory. In classic T-mazes, it
is well known that after initial learning to one reinforced arm, switching
to the other (reversal) is easier or harder depending on the strength of
the initial learning. Over-training tends to make reversal learning eas-
ier (the overlearning reversal effect). In the MWM, reversal learning is
approximated by moving the hidden platform to the opposite quadrant
once the first position is learned. What is reversal learning in the CWM?
Williams and Vorhees reasoned it would be a mirror image of the maze.
Would it be harder for rats to learn the mirror image after reaching as-
ymptotic performance on the original version because they would have
retroactive interference? Or would it be easier because of positive trans-
fer of training and the prior learning would help them solve the mirror
image faster as in the overlearning reversal effect? They built a mirror
image maze and ran the experiments. As it turned out, the transfer was
positive. The negative interference the rats experience is outweighed by
the positive transference they get from knowing the general features of
the maze. This is an example of how, at this stage of the tool development
and utilization, theoretical questions inspire technological innovation.
The experimental goal of learning how a novel cognitive task could
be solved went hand in hand with the goal of making the apparatus
more sensitive. A lot of the refinements never make it into the scientific
literature. At the most, they are referenced as ‘we have shown previ-
ously that …, say 5 min between trials was sufficient to not compromise
trial-2 performance.’ Changes are described the first time they are pub-
lished, but after that, it is treated as standard. It disappears into the past.
These incremental steps are found periodically in review articles, such as
Vorhees and Williams (2006, 2014, 2015, 2016), but can be hard to
detect in data-based articles.

3 Reflections on Research Practices in Neuroscience

The case of the invention and development of the CWM exemplifies
general trends in neuroscience which generalize to most biomedical sci-
ences. We have identified three fundamental drivers of scientific innova-
tion: resources, collaborations, and theory-building.

3.1 Resources: Funding and Tool Availability

The CWM case exemplifies the significance of factors such as fund-
ing and experimental equipment. The development of tools such as the
76 Nina A. Atanasova et al.
CWM is a lengthy process. It took the Vorhees/Williams lab decades to
calibrate their equipment to work in ways that improved utility. Because
the test has been continuously refined, they are now in a position to look
at the data in a new experiment testing a new variable to determine if
the data do or do not fit predictions for a given test article. In such cases,
they can hypothesize what will be found and if the data are not consis-
tent with prediction, they can revise their hypothesis confident in the
knowledge that the CWM provided accurate data. This is the know-how
they acquired through the continuous use of the apparatus over time.
It is not irrational to choose to perform some experiments and not
others because of available funding. For example, an insufficient num-
ber of experimental subjects (be they humans or nonhuman animals)
because of insufficient funding, would jeopardize the statistical signifi-
cance of the results articulated in a study. Avoiding mistakes is another
reason for limiting the experimental design choices to a core set of tools
that the experimenter knows well. This is not to say that one should
never introduce new tools in the toolkit. It is, however, important to
recognize when a certain tool-making endeavor needs to be abandoned.
Sometimes a lab cannot get some equipment to generate useful data. In
such cases, experimenters tend to “cut their losses,” abandon a method
that does not seem productive, and try another one given the limits on
money, time, and personnel.

3.2 The Role of Collaboration

The specialization of labs using different tools and experimental tech-
niques would leave unbridgeable gaps and lead to incommensurable
theoretical commitments across the research communities trained in a
certain lab if Kuhn were correct about the rigidity of the worldviews he
ascribed to scientists. Contemporary neuroscience proves this ascription
unjustified. Ankeny and Leonelli’s scientific repertoires capture what
happens much better. Some labs have solved problems others have en-
countered. A single lab cannot install all equipment needed in the field.
To calibrate their results, scientists often reach out to colleagues in
other labs that have different tools in their toolboxes. This is why collab-
oration, rather than competition, is crucial. The venues for establishing
collaborations are, not surprisingly, the professional conferences where
neuroscientists “trade experience.”
On the large scale of present-day grant-funded research, collabora-
tions between multiple laboratories are crucial for innovation. Differ-
ent labs have different equipment and expertise, some of which can be
expensive and difficult for every lab to set up and maintain. The CWM
is one such tool. It is mostly used by the Vorhees/Williams lab because
of fabrication costs and space requirements. The MWM is widely used
because all one needs is a room, a circular water tank, and a platform.
 Science in Practice in Neuroscience 77
Moreover, many MWM tanks are drained and put in storage until
needed again. For many labs, it is impractical to build a large, complex,
permanently installed maze such as the CWM. However, having the
CWM in their toolbox, the Vorhees/Williams lab can answer questions
that labs without such a striatally based test cannot. There are no other
tests that measure egocentric learning and memory as well as the CWM.
Conversely, when other questions need to be addressed and Vorhees
and Williams do not have the equipment for it, they send their rats to
other laboratories. The decisions regarding tool utilization are con-
strained by the availability of resources and the stage of the research.
If the studies in the collaborating laboratories “were to find something
new,” Vorhees and Williams typically try to get the equipment set up in
their lab so they can continue. Alternatively, they may collaborate on
an ongoing basis with the other lab and apply jointly for grant support.
What hypotheses are explored depends on the available funding,
equipment/expertise, and collaboration between laboratories. For exam-
ple, the Vorhees/Williams lab developed a genetically modified rat using
CRISPR/Cas9 genome editing to delete exon 3 of the latrophilin-3 gene
associated with some cases of Attention Deficit Hyperactivity Disorder
(ADHD) (Regan et al. 2019). The LPHN-3 knock-out rat is hyperac-
tive, hyper-reactive to sudden stimuli, and has cognitive deficits, but they
have no direct way of testing attention and impulsivity. Therefore, they
sent these rats to Dr. Helen Sable at the University of Memphis who uses
operant conditioning to test attributes Vorhees and Williams could not.
Her data affirm that these rats are less attentive and more impulsive than
wildtype littermates.
Ankeny and Leonelli’s notion of research repertoires captures the col-
laborative practice of contemporary neuroscience. What the CWM case
shows is how laboratories develop specialized skills and techniques that
they contribute to larger-scale research in collaborative projects. The
research practice of neuroscience is driven by the availability of funding,
institutional infrastructure, and interactions between researchers.

3.3 Theory-Building: Theories as Interpretative Devices

Where did theory go? The influence of postpositivism is still strong in
the contemporary philosophy of science. Some versions of the Kuhnian
notion of paradigm lurk behind the concerns about the equivocal use of
terminology and theoretical constructs by different research teams in
neuroscience (Sullivan 2016), the claims about irreconcilable pluralism
in the study of behavior (Longino 2013), and irreconcilable codepen-
dence of psychology and neuroscience (Hochstein 2016). However, when
analyzed in detail, the practice of neuroscience displays the incremental
change of experimental practices and theoretical narrations rather than
“big paradigm shifts.” Neuroscience is characterized by an interplay of
78 Nina A. Atanasova et al.
tool development, experiment, and theory articulation. A single exper-
iment often leaves many unanswered questions. If a given hypothesis
seems to explain a certain effect, then there is the question of whether it
should be tested further to establish that it is indicative of a general phe-
nomenon. Unfortunate as it is, in neuroscientific practice, this question
is often left unaddressed, many times due to pragmatic considerations of
tool availability or funding. Therefore, opportunities to articulate com-
prehensive theories of the phenomena of interest are often unexplored.
This shows that the end goal of experimental neuroscience is not always
the articulation of theories.
Twentieth-century philosophy of science developed largely as a reflec-
tion of theoretical physics. It led to the view that the rest of the scientific
disciplines would eventually achieve the maturity of physics, if not turn
into physics themselves. This presupposed that the goal of science was
to produce theories about the natural world. On this view, experiment
plays only a supporting role to theory. Experiments are of interest only
as long as they produce evidence for one theory over another. Tool de-
velopment was not on the radar of philosophy of science until recently,
particularly with the boom of biomedical technology which enabled ex-
periments unthinkable 20 years ago.
In contemporary neuroscience, it is uncommon for researchers to be
guided by abstract theory in experimental practice. If neuroscientists
mention theories or when they announce that they are going to give a
“theoretical talk,” what they mean is something along the lines of: “We
need to think more broadly. We need to figure out how to fit all these
pieces of information together in a coherent hypothesis.” Unlike theo-
ry-driven physics, neuroscience typically articulates theories after a sig-
nificant number of diverse experimental data have accumulated and the
data need to be synthesized and made sense of within a theoretical frame-
work. In other words, the role of a theory is to interpret conglomerations
of data resulting from the testing of multiple focused hypotheses. For
example, the theory of cognitive maps in the hippocampus grew out of
electrophysiological experiments with recording electrodes put into the
awake rat hippocampus. It was combined with later experiments where
the relationship with the hippocampus was found using hippocampal
lesions to disrupt radial-arm maze learning. Then came Morris’ inven-
tion of the MWM that made assessing spatial learning clearer. Many
more electrophysiological experiments mapping “place cells” in the hip-
pocampus were performed, followed by many experiments showing how
hippocampal place cells map onto entorhinal grid cells and other ex-
periments that found head direction cells. These and other experiments
eventually culminated in a theory of the hippocampus-entorhinal cortex
of memory formation, consolidation, and retrieval. This is the theory
for which John O’Keefe, Edvard Moser, and May-Britt Moser won the
2014 Nobel Prize for the discovery of “cells that constitute a positioning
 Science in Practice in Neuroscience 79
system in the brain” which is the basis of memory. In this case, theory
followed rather than preceded the experiment.
In a sense, neuroscience is still working within an area of black boxes.
A lot is known about the brain and about how it functions. Neuro-
scientists often ask: “ What is the mechanism underlying a particular
observed change?” In reality it could be many different things. Neu-
roscience proceeds through experimental testing of focused hypotheses
about the effects of specific interventions under specific conditions until
a bigger picture of what happens is gradually assembled. This is when
articulating a comprehensive theory of the studied phenomena comes
into play.
This is how our case study unfolds. The CWM was originally used
to study spatial memory and navigation similarly to the MWM. Allo-
centric and egocentric navigation were recognized as distinct behaviors
but there was no assumption that they had distinct neural mechanisms.
When Vorhees and Williams started to observe systematic differences in
the performance of rats in the two mazes under identical interventions,
they started to ask questions about the underlying causes of the differ-
ences. They published their results, but peer reviews asked the question
about what CWM results meant. Vorhees and Williams sought an expla-
nation. This meant asking theoretical questions about how to interpret
data. Initially, they guessed they were studying path integration. How-
ever, a colleague pointed out that this could not be since path integration
is manifested in taking shortcuts in navigating a familiar terrain and this
was not possible in the CWM. Hence, they hypothesized that they were
studying route-based egocentric navigation. This was what prompted
the modifications of the maze that eventually led to using an infrared
camera to test the rats in darkness. Once the maze was standardized,
they were able to use it in the search for the underlying mechanism of
egocentric navigation.
Vorhees and Williams did not have a preconceived model of that
mechanism. There was not a theory to guide their efforts. To the extent
that they were operating under the assumption that observable behaviors
(e.g., navigation) attributed to a cognitive capacity (e.g., memory) have
underlying neural patterns, there was a theoretical assumption. This as-
sumption, however, characterizes neuroscience as a research endeavor.
It is not a theory that the CWM could test. Vorhees and Williams ar-
ticulated a theory to explain route-based egocentric navigation, which
had not been thoroughly investigated at the time. Their experiments
gradually identified the striatum as the predominant region involved in
egocentric navigation in the CWM. The next step was to identify the
neurotransmitters involved. Experiments showed that dopamine is one
of them but there may be others. All along Vorhees and Williams have
been testing focused hypotheses about the effects of specific interven-
tions on specific brain areas during egocentric navigation. These steps
80 Nina A. Atanasova et al.
are taking them closer to a comprehensive theory of the mechanism of
egocentric navigation, but this is still a theory in the making. Such is
the case with many research domains in neuroscience. This is why we
contend that the philosophy of neuroscience is best served by studying
experiment and tool development along with the theories they generate.

4 Conclusion
We presented a case study of the invention and calibration of a tool in
experimental neuroscience, the Cincinnati water maze. Our results show
that making choices regarding what experiments to pursue more often
than not relies on pragmatic considerations of funding, equipment, and
collaborations. Our analysis is consistent with the notion of scientific
repertoire and challenges approaches to philosophy of science which fo-
cus on the study of scientific theories rather than scientific practice. The
case we presented is exemplary of the tendency in contemporary neuro-
science for theories to be articulated for the purposes of interpretation
of disparate experimental data that need to be integrated into a cohesive
framework. Theory in this case functions as an interpretative device.

References
Ankeny, R. A. and S. Leonelli (2016). “Repertoires: A post-Kuhnian perspective
on scientific change and collaborative research”. Studies in History and Phi-
losophy of Science, 60: 18–28.
Atanasova, N. (2015). “Validating animal Models”. THEORIA: An Inter-
national Journal for Theory, History and Foundations of Science, 30(2):
163–181.
Bickle, J. (2006). “Reducing mind to molecular pathways: Explicating the re-
ductionism implicit in current cellular and molecular neuroscience”. Syn-
these, 151(3): 411–434.
Bickle, J. (2016). “From microscopes to optogenetics: Ian Hacking vindicated.”
Philosophy of Science, 85(5): 1065–1077.
Biel, W. (1940). “Early age differences in maze performance in the albino rat.”
The Journal of Genetic Psychology, 56: 439–453.
Boyd, R. (1980). “Scientific realism and naturalistic epistemology.” PSA: Pro-
ceedings of the Biennial Meeting of the Philosophy of Science Association,
2: 613–662.
Braun, A. A., D. L. Graham, T. L. Schaefer, C. V. Vorhees and M. T. Wil-
liams (2012). “Dorsal striatal dopamine depletion impairs both allocentric
and egocentric navigation in rats.” Neurobiology of Learning and Memory,
97: 402–408.
Braun, A. A., R. M. Amos-Kroohs, A. Guetierez, K. H. Lundgren, K. B. Se-
roogy, M. R. Skelton, C. V. Vorhees and M. T. Williams (2015). “Dopamine
depletion in either the dorsomedial or dorsolateral striatum impairs egocen-
tric Cincinnati water maze performance while sparing allocentric Morris wa-
ter maze learning.” Neurobiology of Learning and Memory, 0: 55–63.
 Science in Practice in Neuroscience 81
Butcher, R., C. Vorhees and H. Berry (1970). “A learning impairment associated
with induced phenylketonuria.” Life Sciences, 9, Part I: 1261–1268.
Chakravartty, A. (2007). A metaphysics for scientific realism. Knowing the un-
observable. Cambridge: Cambridge University Press.
Dringenberg, H. C. (2020). The history of long-term potentiation as a mem-
ory mechanism: Controversies, confirmation, and some lessons to remember.
Hippocampus, 30: 987–1012.
Feest, U. (2010). “Concepts as tools in the experimental generation of knowl-
edge in cognitive neuropsychology.” Spontaneous Generations: A Journal for
the History and Philosophy of Science, 4(1): 173–190.
Feyerabend, P. (1988/1975). Against method. London and New York: Verso.
Gross, P. R. and N. Levitt (1994). Higher superstition: The academic left and
its quarrels with science. Baltimore, MD: Johns Hopkins University Press.
Gross, P. R., N. Levitt, and M. W. Lewis (eds.) (1996). The flight from science
and reason. New York: New York Academy of Sciences.
Hacking, I. (1983). Representing and intervening: Introductory topics in the
philosophy of natural science. Cambridge and New York: Cambridge Uni-
versity Press.
Hanson, N. R. (1965). Patterns of discovery. Cambridge: Cambridge University
Press.
Herring, N. R., T. L. Schaefer, G. A. Gudelsky, C. V. Vorhees and M. T. Wil-
liams (2008). “Effects of (+)-methamphetamine on path integration learning,
novel object recognition, and neurotoxicity in rats.” Psychopharmacology,
199(4): 637–650.
Hochstein, E. (2016). Giving up on convergence and autonomy: Why the theo-
ries of psychology and neuroscience are codependent as well as irreconcilable.
Studies in History and Philosophy of Science, 56, 135–144.
Kuhn, T. (1970). The structure of scientific revolutions. Chicago, IL: University
of Chicago Press. First published 1962.
Leonelli, S. and Ankeny, R. A. (2015). “Repertoires: How to transform a project
into a research community”. Bioscience, 65(7): 701–708.
Longino, H. E. (2013). Studying human behavior: How scientists investigate
aggression and sexuality. Chicago, IL: University of Chicago Press.
Morford, L. L., S. L. Wood, G. A. Gudelsky, M. T. Williams and C. V. Vorhees
(2002). “Impaired spatial and sequential learning in rats treated neonatally
with d-fenfluramine.” European Journal of Neuroscience, 16 (3), 491–500.
Morris, R. G. M. (1981). “Spatial localization does not require the presence of
local cues”. Learning and Motivation, 12: 239–260.
Polidora, V. J., D. E. Boggs and H. A. Waisman (1963). “A behavioral deficit as-
sociated with phenylketonuria in rats.” Proceedings of the Society for Biology
and Medicine, 113: 817–820.
Putnam, H. (1976). “X* – What is ‘realism’?” Proceedings of the Aristotelian
Society, 76(1): 177–194.
Regan, S. L., J. R Hufgard, E. M. Pitzer, C. Sugimoto, Y. Hu, M. T. Williams
and C. V. Vorhees (2019). “Knockout of latrophilin-3 in Sprague-Dawley rats
causes hyperactivity, hyper-reactivity, under-response to amphetamine, and
disrupted dopamine markers.” Neurobiology of Disease, 130: 104494.
Robins, S. K. (2016). “Memory and optogenetic intervention: Separating the
engram from the ecphory”. Philosophy of Science, 85(5): 1078–1089.
82 Nina A. Atanasova et al.
Ross, A. (ed.) (1996). Science wars. Durham, NC: Duke University Press.
Silva, A. J. and J. Bickle (2009). The science of research and the search for
molecular mechanisms of cognitive functions. In J. Bickle (Ed.), The Oxford
handbook of philosophy and neuroscience (pp. 91–126). Oxford: Oxford
University Press.
Silva, A., A. Landreth and J. Bickle (2014). Engineering the next revolution in
neuroscience. New York: Oxford University Press.
Sullivan, J. A. (2009). “The multiplicity of experimental protocols: A challenge
to reductionist and non-reductionist models of the unity of neuroscience.”
Synthese, 167: 511–539.
Sullivan, J. A. (2010). “Reconsidering ‘spatial memory’ and the Morris water
maze.” Synthese, 177: 261–283.
Sullivan, J. A. (2016). “Optogenetics, pluralism, and progress.” Philosophy of
Science, 85(5): 1090–1101.
Vorhees, C. V. (1987). “Maze learning in rats: A comparison of performance
in two mazes in progeny prenatally exposed to different doses of phenytoin.”
Neurotoxicology and Teratology, 9: 235–241.
Vorhees, C. V. and M. T. Williams. (2006). “Morris water maze: Procedures for
assessing spatial and related forms of learning and memory.” Nature Proto-
cols, 1(2): 848–858.
Vorhees, C. V. and M. T. Williams (2014). “Assessing spatial learning and mem-
ory in rodents.” ILAR Journal, 55(2): 310–332.
Vorhees, C. V. and M. T. Williams (2015). “Reprint of ‘Value of water mazes for
assessing spatial and egocentric learning and memory in rodent basic research
and regulatory studies.’” Neurotoxicology and Teratology, 52: 93–108.
Vorhees, C. V. and M. T. Williams (2016). “Cincinnati water maze: A review of
the development, methods, and evidence as a test of egocentric learning and
memory.” Neurotoxicology and Teratology, 57: 1–19.
Whishaw, I. Q. and H. Maaswinkel (1998). “Rats with fimbria-fornix lesions
are impaired in path integration: A role for the hippocampus in ‘sense of di-
rection’”. The Journal of Neuroscience, 18(8): 3050–3058.
Williams, M. T., L. L. Morford, A. E. McCrea, S. L. Wood and C. V. Vor-
hees (2002). “Administration of D, L-fenfluramine to rats produces learning
deficits in the Cincinnati water maze but not in the Morris water maze: re-
lationship to adrenal cortical output.” Neurotoxicology and Teratology, 24:
783–796.
Williams, M. T., L. L. Morford, S. L. Wood, S. L. Rock, A. E. McCrea, M.
Fukumura, T. L. Wallace, H. W. Broening, M. S. Moran and C. V. Vorhees
(2003). “Developmental 3,4-methylenedioxymethamphetamine (MDMA)-in-
duced learning deficits are not related to undernutrition or litter effects: Novel
use of litter size to control for MDMA-induced growth decrements.” Brain
Research, 968(1), 89–101.
4 Where Molecular Science
Meets Perfumery
A Behind-the-Scenes Look
at SCAPE Microscopy and
Its Theoretical Impact on
Current Olfaction
Ann-Sophie Barwich and Lu Xu

1 Introduction: A Cognitive Approach to Tool Use

Imagine studying an entire brain in action, live stream in all its complex
interactive signaling, down to the single-cell level. You are able to im-
age an entire nervous system or large intact tissue sections reacting to
a stimulus. So, you may spritz an odor at a fruit fly or look at muscle
contractions in larvae to see what cells react and when. This must sound
like a dream to any neuroscientist studying structure function interac-
tions in the central nervous system and beyond. This dream turned into
reality with the invention of a new tool by the Hillman lab at Colum-
bia: SCAPE (Swept, Confocally-Aligned Planar Excitation microscopy),
a tool for three-dimensional, rapid live-stream imaging of small living,
freely moving organisms (e.g., C. elegans, drosophila) and large intact
brain tissues of bigger animals (as: in mice). Despite the relative recency
of its invention, SCAPE has already provided groundbreaking insights
into the workings of the mammalian and insect nervous and motor sys-
tem (Bouchard et al. 2015; Vaadia et al. 2019; Xu et al. 2020), one of
which will be the topic of this chapter.
Science in general, and neuroscience in particular, is driven by new
tools and technologies. Philosophical analysis of science has engaged
with various historical episodes, meticulously reconstructing the impact
of technologies on the trajectory of a field and its models. Yet, a clear
understanding of the epistemic role of tools in the research process still
remains at large. Of course, the general significance of tools in exper-
imental science is indisputable: research tools support observation and
inference-making, enlarge data production and accelerate data analysis,
and shape the communal standards with which materials and models are
selected as paradigmatic cases to explain phenomena of interest. Mean-
while, another—and perhaps less obvious—function of tools in science
is their explicitly cognitive dimension.

DOI: 10.4324/9781003251392-6
84 Ann-Sophie Barwich and Lu Xu
Previous cognitive theories of science, meaning cognitive approaches
that target the mental mechanisms and environmental affordances mak-
ing scientific reasoning possible, have been principally applied to scien-
tific models and theories (Giere 1988; Churchland 1996; Thagard 2012),
not the specific nature of tools. In recent years, Bickle (2016, 2019, 2020,
this volume) has shown with historical examples that engineering and
tinkering with tools has been a central driver of theories throughout the
history of neuroscience. Such an account pointing toward the cognitive
role of tools in neuroscience further invites a contemporary perspective
to complement and expand upon our understanding of scientific tool use
in action.
This chapter offers the rare opportunity of going behind the scenes at
the frontiers of science to get a better look at the pragmatic and cognitive
dimensions of tool use and development. Specifically, it provides a first-
hand account of how SCAPE arguably revolutionizes current research
on the brain by revealing how this tool was made to fit its experimental
potential, and notably one with tremendous theoretical implications.
The authors are well positioned to undertake this challenge. Between
2015 and 2018, Barwich was the resident philosopher in the Firestein lab
at Columbia University (Barwich 2020a), where Xu was the lead scien-
tist in the process of adopting the new tool of SCAPE to tackle mixture
coding in the nose (Xu et al. 2020). The results of these experiments
would turn out to be remarkable: Xu et al.’s study, published in Science,
uncovered a previously unknown molecular mechanism in the sensory
periphery of olfaction.
In what follows, we present a combined narrative of tool development
and its broader philosophical implications to better understand scientific
reasoning at the laboratory bench. This chapter integrates philosophical
analysis with contemporary scientific history to engage with the funda-
mental question of how scientists use new experimental tools to access
yet unknown features of research materials. Specifically, we want to un-
derstand how tools can perform a cognitive function that opens new the-
oretical perspectives for scientists in their experimental investigations.
What characterizes the cognitive function of scientific tool use in action?
Specifically, how do tools structure scientific reasoning by providing the
observational and conceptual scaffolds that create or stimulate the abil-
ity to conceive new perspectives on research objects or processes? How
does tool use afford the conceptual reconfiguration of a research topic?
In short, we want to understand how tools unlock conceptual possibili-
ties that were previously difficult to imagine.
SCAPE is an excellent example to show how tool innovations can pro-
vide scientists with new cognitive scaffolds in the advancement of scien-
tific theorizing. Traditionally, the cognitive structure of tools in research
practice has been considered mostly in terms of their theory-ladenness,
meaning their construction and application embody existing theoretical
 Molecular Science Meets Perfumery 85
assumptions that shape what we see and learn to understand with their
use (e.g., van Fraassen 1980, 2008; Barwich 2017). According to this
view, tools primarily assist and supplement scientific thinking. Here, we
want to highlight that research tools also embody a crucially constitu-
tive part of that thinking process, so much so that tools occupy an essen-
tial causal role in the mental mechanisms of scientific reasoning. In other
words, tools do not merely embody existing theoretical knowledge. In-
stead, tools are active drivers of new models and theories by accommo-
dating and, moreover, extending the cognitive process of scientists.
We present our analysis in three steps. First, we introduce the scientific
challenges and developments in recent research on odor coding, involv-
ing the receptors and the stimulus (Section 2). This will help to situate
the significance of the SCAPE study for general readers and highlight
the theoretical impact of its results. Then the SCAPE study takes center
stage (Section 3). Here, we blend the scientific study and its published
results with Xu’s own experience on getting the experiments to work
in the first place. In other words, we get a glimpse of how the sausage
was made! These details present the backdrop against which we clarify
the theoretical impact of the SCAPE study for theories of odor coding
(Section 4). We conclude with a (brief) philosophical reflection on the
cognitive dimensions of tool use (Section 5). Specifically, we draw on
theories of distributed cognition in recent cognitive science to suggest
that tools are a constitutive and active part of the reasoning process by
which scientists generate new exploratory and explanatory concepts.

2 The Scientific Challenge: A Code in the Nose

Crucial to any new technology’s success is identifying the right kind of
problem, specifically a situation more difficult to probe with previous
tools. SCAPE had found its perfect match in the mechanisms of odor
coding at the sensory periphery.

2.1 At the Periphery: The Olfactory Receptors

Why olfaction? Odor coding at the receptor level has been notoriously
difficult to investigate. Notably, the olfactory receptors were discovered
only 30 years ago (Buck and Axel 1991). This discovery transformed the
field of olfaction rapidly by catapulting it into mainstream neurobiology
and genetics, almost overnight (Firestein 2005; Firestein, Mombaerts,
and Greer 2014; Barwich 2020a, 2020b). The recognition of significance
reached far beyond olfaction and led to the 2004 Nobel Prize in Physiol-
ogy or Medicine for Buck (2005) and Axel (2005).
The reason for the great impact of the receptor discovery is twofold.
On the one hand, the olfactory receptors provided the key causal ele-
ments connecting the distal stimulus of odorants (airborne odoriferous
86 Ann-Sophie Barwich and Lu Xu
chemicals) with activity patterns in the brain. On the other hand, these
receptors were identified as the largest and genetically most diverse
members of the superfamily of G-protein coupled receptors (GPCRs).
GPCRs are a fundamental entity in the inventory of biochemical re-
search concerning a wide variety of biological processes, such as vision,
the regulation of immune responses, the detection of neurotransmitters
in the brain, and now also the encoding of chemical information as
smells (Snogerup-Linse 2012; Barwich and Bschir 2017). Indeed, olfac-
tory GPCRs proved a promising group of genes for future studies on
general structure-function behavior in the GPCR superfamily (Barwich
2015a).
What makes the study of odor coding at the periphery so challenging?
One central difficulty early on was the expression of odor receptors in
tissues other than olfactory sensory neurons.1 This hindered the abil-
ity of functional studies to test whether these receptors really constitute
olfactory receptors, that is, that they responded to odorants instead of
some other stimuli. Prior to the successful heterologous expression of
odor receptors in other tissues (Dey et al., 2011; Mainland et al. 2015;
Matsunami 2016), a study led by Haiqing Zhao in the Firestein lab
(Zhao et al. 1998) established the functional role of this newly discov-
ered GPCR family as unmistakably olfactory in nature. Briefly, this study
introduced a selected olfactory receptor gene (the rat receptor OR-I7)
via a viral vector into nasal epithelium tissue, resulting in a significant
overexpression of that receptor in epithelium tissue, and thus allowing
for the study of its foundational response and stimulus-response range.
However, this study could not yield insight into the principles of the
molecular detection mechanism. The problem here is that GPCRs are
generally quite difficult to stabilize. Besides, to this day there is no suc-
cessful X-ray crystallography structure of the mammalian olfactory
­receptors—further exacerbating structure-function studies. (Although
an X-ray crystallography structure has recently been obtained for an
olfactory receptor of the genetically notably different insect receptors;
see Butterwick et al. 2018). 2
The breakthrough arrived with a functional study by Bettina Malnic
in the Buck lab, which revealed that odorants are recognized combi-
natorially by the receptors (Malnic et al. 1999). “Combinatorial cod-
ing” means that one odorant is detected by different receptors, and one
receptor can interact with different odorants. Notwithstanding, under-
standing of receptor-odorant interactions has experienced only limited
progress since (Barwich 2020a). The reason for this stagnation resided
in the ongoing experimental inaccessibility of odor receptors. But the sit-
uation started to change in recent years (detailed review in Kurian et al.
2021). Central to this change was the application of SCAPE to odor cod-
ing in the mammalian epithelium, specifically to the coding mechanism
of mixtures (i.e., odor blends involving two or more odorants).
 Molecular Science Meets Perfumery 87
Mixture coding in olfaction had hitherto posed seemingly insur-
mountable experimental hurdles, so much so that the investigation of
mixtures was considered impracticable and unfeasible. Frankly, it re-
mains experimentally very tricky to examine the responses of individual
receptors—or sensory neurons (with one neuron expressing one receptor
gene; Mombaerts et al. 1996)3 —even to single odorants. Indeed, most
olfactory receptors have not yet been de-orphanized (meaning: their re-
sponse range to odorants and repertoire remains unknown).4 So how
could you possibly measure and compare thousands and thousands of
cells responding to multiple odors at once? You’d need a tool that can
document the population behavior of these numerous cells while also
allowing for detailed single-cell analysis.
This is what the invention of SCAPE microscopy afforded, and with
surprising experimental results. Now, to explain why the results of the
SCAPE study in olfaction were so surprising and remarkable, we first
must do a quick detour to give some details about the nature of the stim-
ulus and its perception.

2.2 In the Air: The Olfactory Table of Elements

If you tried counting odors, how many could you name? One peculiar
feature of your sense of smell is that, once you start paying attention to
it, the number of different scents seems almost endless. There are numer-
ous flowers with different fragrances, an array of food odors, and every
human or even every object has its own unique smell. We may not always
be able to give names to these various olfactory impressions, but we can
discriminate between them fairly well. At the same time, we clearly are
able to group many of these odors into general qualities, such as being a
green, a floral, a meaty, or a woody scent. How is that possible?
Imagine you want to invent a machine that can detect the slightest
change of features or concentration in an ever-changing and chem-
ically complex environment. This machine would need to be fast,
adaptive, sensitive, and widely tuned. Your nose is precisely such an
instrument.
Odors are numerous indeed, and their molecular basis is intriguingly
complex. Odorants come from a variety of sources and are of immense
structural diversity. Calculating the number of chemicals that can cause
different odor impressions will not necessarily lead to a final estimate,
though. An optimistic number was that Humans Can Discriminate More
than 1 Trillion Olfactory Stimuli after a study of the same title by the team
headed by Andreas Keller and Leslie Vosshall at Rockefeller University,
published in Science in 2014 (Bushdid et al. 2014). While this staggering
number has been disputed (Gerkin and Crasto 2015; Meister 2015), it re-
mains unquestioned that the range of olfactory stimuli is mindbogglingly
vast and covers at least several hundred thousands of different molecules.
88 Ann-Sophie Barwich and Lu Xu
(Besides, this further illustrates why the de-orphanization of olfactory
receptors presents quite the laborious challenge.)
Compared to the (still growing) periodic table of elements in chem-
istry, the biological table of olfactory elements appears limited at first
glance. Only molecules containing atoms of hydrogen, carbon, nitrogen,
oxygen, and sulfur elicit odor sensations. Your nose does not seem to
care much about other elements. Additionally, which of these molecules
really are odorous depends on the receptors, not the chemistry of the
stimulus in isolation because:

it is nearly impossible to predict whether a given molecule will be

odorous and what its odor quality might be from the chemical struc-
ture alone. Although all odor molecules are typically organic com-
pounds of low molecular weight, they may be aliphatic or aromatic,
may be saturated or unsaturated, and may have any of several polar
functional groups. However, there are many molecules that conform
to those characteristics, which are nonetheless odorless, to humans
and other animals.
(Poivet et al. 2018, 1)

Even within this seemingly restricted palette, the structural varieties

and combinatorial possibilities are countless. While being constantly
surrounded by hundreds of such odorants, your nose allows you to fil-
ter specific odor notes and detect differences resonating with minimal
variances of elements in your surroundings. It does so with remarkable
precision. A difference in one atom of two otherwise perfectly similar
molecules can make your perception of their odor quality vary entirely.
Consider nonanoic acid, a molecule smelling of cheese that has the
chemical formula CH3(CH 2)7COOH. By adding only one methyl group,
we end up with decanoic acid, CH3(CH 2)8COOH, which you will per-
ceive distinctly different as rancid smelling. Similar cases where minor
structural variations, even of just a single atom, result in drastic odor
differences are abundant.
As early as 1895, olfactory scientists like the French scholar Frédéric
Passy had noted that the sensitivity of the human nose is far greater than
that for spectral analysis.5 Or, in the more contemporary words of the
neuroscientist Stuart Firestein: “Your nose is arguably the most accurate
and sensitive chemosensor on the face of the planet!”6
The central scientific challenge in olfaction remains to explain just how
the nose, in tandem with the brain, ‘makes scents’ of this multitude of
volatile molecules. Scientists today still ask: is there any physical order to
perceptual odor space? Is it even meaningful to assume a physical order?
If it existed, how would such an order correspond to perceptual qualities?
Unlike the visual or auditory system, the physical stimulus of smell has
not yet been reduced to a small number of measurable causal attributes.
 Molecular Science Meets Perfumery 89
This is not a result of the seemingly subjective nature of smells but is
due to the molecular complexity of the olfactory stimulus. Consider the
stimulus in vision or audition. As wavelengths, colors, or sounds come
spatially ordered in a continuous linear scale. By contrast, the molecular
basis of smells does not comply with such a linear arrangement. There
is not one but multiple features that determine the quality of an odorant
(Keller and Vosshall 2016). The concept of the chemical similarity be-
tween odorants encompasses a multitude of parameters, including, ste-
reochemistry (a molecule’s geometrical arrangement), molecular weight,
functional groups, polarity, acidity/basicity, and many more (Ohloff,
Pickenhagen, and Kraft 2012). This multidimensionality of physical
features, encompassing approximately 5,000 molecular parameters,
hinders a straightforward answer to our initial question of how percep-
tual odor space is connected to its physical stimulus space. To date, no
general rule has been discovered that links a definite set of molecular
features to a specific smell (Rossiter 1996; Sell 2006; Barwich 2015b,
2020a; Poivet et al. 2016, 2018).
Perhaps, the sensible way around this problem is not to focus primar-
ily on the principles of organic chemistry but on the sensory system that
translates these features into smells. To get to the link between mole-
cules and odor, one might need to start with the system that extracts
sensory information from the stimulus. It turns out the scientific answer
to this challenge is far from obvious or even close to being settled.

3 Behind the Scenes: Notes of a Scientist

Investigations into the mechanism of mixture perception were consid-
ered to pose a challenge of experimental complexity primarily. However,
the SCAPE study revealed that the challenge of understanding mixture
coding at the sensory periphery is one of showing previously uncon-
ceived causal principles.

3.1 Introducing SCAPE to the Mix

Xu et al. (2020) found a molecular explanation for a peculiar and wide-
spread phenomenon about mixture perception in olfaction. Perfumers
have long known that odors do not behave additively in mixtures. A
comparison with the visual system helps to clarify this issue. Color vi-
sion acts additively, which means that we can calculate colors by adding
and subtracting ranges of electromagnetic wavelengths, translating into
colors in the visible spectrum of light. Take the color pink. Pink has no
associated wavelength but constitutes a brain’s computational feature,
namely: “pink is white light minus green.” In contrast, olfactory mix-
tures as perceptual wholes are not the sum of their (chemical) parts.
Additively in olfaction would imply we should be able to calculate what
90 Ann-Sophie Barwich and Lu Xu
the sum of components will smell like by putting together its individual
compounds. However, that is not the case.
Smell mixtures behave irregularly, often unpredictably. Perfumers
working on a formula routinely note that, at times, they may put a tiny
amount of a compound into a complex chemical blend, which then in-
stantly alters the quality of that entire mix (e.g., Laudamiel in Barwich
2020a). Conversely, the opposite case is also known to occur. Perfumers
can add significant amounts of a compound into a blend without result-
ing in any observable change in the overall scent quality of the mix. This
phenomenon has also been well documented in several psychophysical
studies (e.g., Cain 1974; Laing et al. 1984; Kay et al. 2005), referred to
as enhancement and suppression effects. Odor notes appear enhanced
or suppressed when blended into a mixture.
Such irregularity in complex mixture perception was assumed to con-
stitute either an effect of blending different chemicals into a mixture or
result from a higher-level mechanism in the neural computation of odor
images. Meanwhile, there may be a third, hitherto unconceived alterna-
tive suggesting an unknown mechanism at the receptor level in the sen-
sory periphery (instead of “higher-level” downstream computation). This
mechanism came to the fore with the experiments in Xu et al. (2020).
This is where the technological affordances of SCAPE microscopy
come into play.
SCAPE microscopy is a novel imaging technique developed by the
Hillman lab. SCAPE is a form of light-sheet microscopy which uses a
single, stationary objective lens at the sample. It can capture a sizeable
3D field of view (>1.2 mm × 1.2 mm × 400 μm) with cellular resolution
at over 10 volumes per second (Figure 1). SCAPE forms a 3D image by
illuminating the sample with an oblique light sheet, capturing fluores-
cence light excited by this sheet back through the same high numerical
aperture (NA) objective. A single-axis galvanometer in the system both
sweeps the light sheet from side to side at the sample and de-scans the
returning light to form a stationary oblique intermediate image plane.
This plane is then rotated using a set of two additional objective lenses
(O2 and O3) and focused onto a fast sCMOS camera. Compared with
other point-scanning based volumetric imaging methods, such as confo-
cal or spinning disk confocal microscopy, the light-sheet based SCAPE
approach dramatically increases the volumetric imaging speed and FOV
[field of view] with lower photobleaching and phototoxicity while keep-
ing similar spatial resolution in biological samples. (Further details in
Bouchard et al. 2015; Hillmann et al. 2019; Voleti et al. 2019.)
In its current installment (SCAPE is being constantly improved),
SCAPE microscopy allows studying the behavior of intact cells in times-
cales less than 0.1 s with a resolution high enough to analyze single-cell
activation of the full brain or large-scale neuronal circuits with different
animal models. This procedure creates several terabytes of data.
 Molecular Science Meets Perfumery 91

Figure 1 T
 he SCAPE Microscopy System. (a) Schematic of the SCAPE layout.
(b) Close-up of the O1 objective. An oblique light sheet images through
the O1 stationary objective lens and scans the sample laterally in the
x-direction to form a 3D image (Adapted from Voleti et al. 2019).

The application of SCAPE microscopy to the study of mixtures in

olfaction sounded intuitive from the start, Firestein remarked (quoted
in Barwich 2020a, 192): “The obvious thing to do with this would be
blends or mixtures to see the code.” SCAPE allowed us to observe two
things simultaneously in intact tissue and in real-time: the reaction of
individual cells to stimuli on the one hand and the behavior of entire
cell populations on the other. In other words, Xu was able to expose a
whole and intact section of epithelium tissue to one or more odors to
observe which and how many different cells were activated in response
to a specific stimulus.
So far reads the official story. Meanwhile, the idea to use SCAPE
microscopy to explore the molecular nature of mixture coding in ol-
faction had some pragmatic, and also some romantic, origins. Hillman
and Firestein were colleagues in the Biology Department at Columbia
University. (Hillman’s lab was originally on the fourth floor of the Sha-
piro building before moving to the new Zimmermann Institute (ZI); the
Firestein lab remains on the tenth floor of the Fairchild building on the
main Manhattan Campus to this day.) Meanwhile, there also existed
another connection: Xu, back then doing her graduate research on smell
receptors in the Firestein lab, is married to Wenze Li, a graduate student
of Hillman and working on the further development of SCAPE:

It was 2015 when I first learned about SCAPE microscopy (the year
that the Hillman lab published their first SCAPE paper). Wenze and
I were walking our dog Muffin when he decided to show off his re-
cent progress in the lab: a video showing a 3D Zebrafish heart beat-
ing. The contour of the heart was clearly outlined by the heart cells
92 Ann-Sophie Barwich and Lu Xu
(which were labeled with some sort of fluorescence); as the heart
contracts and relaxes, I could also see individual blood cells travel
through the ventricle and atriums. Wenze told me that this was im-
aged with a 3D imaging technique called SCAPE.
At the moment, I was frustrated by the difficulty that I have en-
countered in my own research: the olfactory sensory neuron dissoci-
ation protocol (which I have been using for the past three years with
no problem) no longer works, and this was the main method of my
study at the moment. Briefly, what the protocol does is to first strip
off the olfactory epithelium from the nasal turbinates of the mouse,
digest with enzymes, then collect olfactory sensory neurons from the
supernatant and place them on a piece of coverslip pre-coated with
concavcalin A – this coverslip will later be used for calcium imaging.
Normally I would get several hundred of neurons on each coverslip.
But during that time, I was only getting ~50 cells per coverslip – a pa-
thetic cell density, and I couldn’t figure out why because my colleagues
(Erwan Poivet and Narmin Tahirova) were using the same protocol
and same reagents and they were getting a normal number of cells.
This was the original momentum of me seeking an alternative
solution of imaging the olfactory sensory neurons. The cell disso-
ciation protocol was not perfect anyway: first, during the dissocia-
tion process, the axons are severed from the olfactory bulb and the
cilia are also damaged by the enzymes. Second, the efficacy is low.
Even with a good cell density, we could only get ~200–400 neu-
rons per field of view, among which only 5%–10% would respond
to a particular monomolecular odor stimulus. The population will
be further narrowed down if we want to study a receptor/neuron
responding to multiple odors. Finally, the dissociation process is ex-
tremely time-consuming. If we start at 9:30 am, the cells would be
ready for imaging at ~4 pm, during which we only have a 40min
break for lunch, and there is no way to know if the dissociation is
successful until the last step. Also, the cells can only be imaged the
same day – if you leave them in a cell incubator, many of them will
still be alive the following day but no longer functional.
As a lazy person,7 I always wanted to take a shortcut. The easiest
thing to come up with is to simply peel off the thin layer of the olfac-
tory epithelium and image it directly. It seems to be straightforward,
but in practice, it’s difficult to mount the epithelium on a coverslip
securely. And even though the epithelium is only ~100 um thick, its
surface is curved, making it difficult to visualize many individual
neurons in a single focal plane (imagine the difference of a flat piece
of paper and a piece of crumpled paper).
The invention of SCAPE has brought new hope to this (almost)
dead end. It’s capability of imaging non-transparent tissue in 3D
 Molecular Science Meets Perfumery 93
was exactly what I needed. How to harness this novel technique is
yet another story.
(ps. Several months later, I finally figured out why the dissociation
protocol was not working: we made a new batch of concavcalin A
from a different manufacturer, and somehow it takes longer to dry
out on the coverslip completely. If it’s not dry, it will be washed off
by the culture media along with all the neurons attached to it. I
used to prepare the coverslip in the morning of the experiment day,
which was fine for the old stock but not enough time for the new
stock. Meanwhile, Erwan and Narmin had prepared the coverslip
one night before, and that’s why they never had the problem!)

These were the roots of the SCAPE study. But back to its experimental
details: Xu used genetically engineered mice to track active cells with
a fluorescent glow. In consultation with the master perfumer Christo-
phe Laudamiel, she decided on two mixtures, each mixture consisting
of three different compounds: (a) acetophenone, benzyl acetate, and ci-
tral, and (b) dorisyl, dartanol, and isoraldeine. The central criterion for
choosing these components was that they were chemically and perceptu-
ally dissimilar. Additionally, the hypothesis was that the combination of
these odorants would yield a configural odor image—configural mean-
ing that the mixtures produced a quality different from their elemental
components. Furthermore, odor set 1 was tested in equal concentration;
odor set 2 in unequal concentration.

3.2 C
 hallenges: Choosing the Chemical Stimulus and Testing
Cell Responses
Scientific papers tend to present their findings without hiccups. Behind
the scenes, the story reads differently. Indeed, Xu dealt with several chal-
lenges already at the stage of choosing the chemical stimulus, specifically
the stimulus sets:

3.2.1 Choice of Odor Set 1

Stuart [Firestein] originally wanted me to test if “one odor can in-
hibit another odor,” which sounds extremely counter-intuitive. I ex-
pected this kind of event to be extremely rare if it happens at all.
Therefore, I want to use “potent” odors that activate a broad reper-
toire of neurons for me to analyze. It’s easy to think of acetophenone
and citral, as they are odors frequently used in olfactory studies and
are known to activate a lot of cells. Fortunately, they also fit the
criteria of being chemically and perceptually dissimilar. This is im-
portant because I was not looking for special cases of antagonism
94 Ann-Sophie Barwich and Lu Xu
that would normally happen between a chemical and its derivative,
but rather interactions that happen at a more general level.
To increase the complexity of odor interaction, I decided to intro-
duce a third component, but not a fourth, because it increases the
possible combinations from 7 (A, B, C, AB, AC, BC, ABC) to 15,
which will be too complicated to test. It is yet difficult to find a third
odor that is as potent as acetophenone and citral. Isoamyl acetate
is a close one, and we finally replaced it with benzyl acetate, which
has similar odor quality but is more frequently used in perfumery,
according to Christophe [Laudamiel].

3.2.2 Choice of Odor Set 2

While I aimed to test chemically and perceptually unrelated odors

with odor set 1, odor set 2 was selected to test the interaction be-
tween chemically unrelated but perceptually related odors. This re-
flects some cases in naturally occurring odors where perceptually
similar odors produce harmonic odor perception, such as rose oxide
and beta-damascone in rose. In fact, odor set 2 (named the “woody
accord”) is designed by perfumers to produce such harmonic smell
at a particular concentration ratio. Interestingly, we have observed
more inhibition in odor set 2 than odor set 1, indicating an under-
lying logic of doing subtraction (eliminating unwanted notes) rather
than addition (adding extra notes) in perfumery.

With the stimulus selection in place, and over the course of several trying
years, Xu repeatedly tested how the cells responded to each compound
when administered individually, in pairs, and then in the tripartite mix
(Figure 2):

The biggest challenge to me was to confirm that the modulation ef-

fects that I saw were real, not artifacts. It’s relatively easy to answer
qualitative questions (“Is this neuron activated by this odorant”) as
all you have to answer is yes or no. However, when it comes to quan-
titative questions (“how much a response is enhanced/inhibited”),
the difficulty of answering increases exponentially (I’ll explain why
in the following paragraphs). I was constantly stuck in the “doubt –
verification – confidence – doubt” cycle during the entire process of
this project, and each round of self-questioning resulted in the opti-
mization of some details.

A key technical challenge in Xu’s experiments concerned sample

mounting:
 Molecular Science Meets Perfumery 95

Figure 2 L
 u Xu working with the SCAPE 1.0 system during the pilot
experiments.

Rather than creating the SCAPE protocol from scratch, what I did
was to use the old cell dissociation protocol as a starting point and
gradually shaped it to the way I wanted (which ended up being quite
different). In our very first experiment, I prepared the OE [olfactory
epithelium] sample by cutting off the whole olfactory turbinates us-
ing a pair of fine scissors – this was actually an intermediate step
during cell dissociation. The next steps were to peel off epithelium
from the turbinates and cut it into pieces – I then placed it into a
perfusion chamber with the epithelium side facing up. Perfusion is a
standard thing to do in physiology in order to maintain the homeo-
stasis of the extracellular environment, and I simply adopted the rec-
ipe of Ringer’s solution we used to perfuse the dissociated neurons.
Next, I tried to deliver odor stimulus by injecting odor solutions us-
ing an automatic syringe injector, which didn’t work out because the
sudden change of flow rate would cause the sample to float around.
Given our field of view was only ~600 um × 400 um × 300 um at
the moment, a displacement less than 1mm would make the sample
disappear from the screen. Therefore, in the following experiments,
we had to transport the Agilent isocratic pump from our lab to the
Hillman lab using a small cart and used that for stable perfusion and
96 Ann-Sophie Barwich and Lu Xu
odor delivery. This lasted until the Hillman lab moved to the ZI, and
we purchased a new pump.
Nevertheless, the sample is still not stable enough in the chamber.
At first, I thought of using a smaller chamber such as the sample
is more confined in place, but then I realized it’s actually easier if I
don’t cut the turbinates at all and leave them attached to the mouse
head. This way all I have to do is to cut the mouse head sagitally and
remove the septum to expose the OE. What’s even better about this
prep is that if I cover the mouse hemi-head with a coverslip, a narrow
space will be naturally formed between the coverslip and the turbi-
nates, which allows a very small volume of liquid to flow in through
the nostril and out through the throat thanks to capillary action. Es-
sentially, this is a re-construction of the nasal cavity by replacing the
septum with a transparent coverslip, which allows us to observe the
OE from the medial side. To implement this idea, we have designed
our own perfusion chamber which resembles the shape of a mouse
head. In the final prep, the mouse hemi-head is placed facing down
to the coverslip on the bottom of the chamber, with an inverted ob-
jective acquiring the image from underneath (as shown in figure 1 of
our published paper, and figure 3 in this chapter). During subsequent
experiments, we further optimized sample mounting by applying a
blue light-cured dental gel to fix the sample to the chamber. As a
result, the displacement of the sample is now less than 10um at all
three dimensions during a one-hour imaging session.

Naturally, challenges did not end here, and Xu further focused on

the measures of response stability and the design of odor stimulation
sequence:

Several key factors can significantly affect the interpretation of re-

sults: rundown, variation, and photobleaching. Rundown means
when stimulated repetitively, the response magnitude of a cell will
gradually decrease. Variation means the response to the same stim-
ulus can fluctuate within a certain range. Photobleaching means the
fluorescent signal gets dimmer under constant excitation light. Due
to these factors, it is difficult to directly compare the response mag-
nitude of two stimuli even if they are given subsequently. However,
thanks to the pioneer work of Zita [Peterlin] et al. in our lab,8 we
already knew that the rundown of odor response is approximately
linear, and the natural variation of odor response is about 10%
(­Peterlin et al. 2008). This result has been confirmed with SCAPE
in our control experiment. In a simplified case [details of findings
in sections 3.4 and 3.5], to measure the effect of a potential modu-
lator B on odor A, the odor stimuli will be given as A – A + B – A,
and compare A + B with mean (A); if A + B is larger than A, the
 Molecular Science Meets Perfumery 97

Figure 3 S CAPE microscopy of intact mouse olfactory epithelium capturing

odor-evoked GCaMP activity with single OSN resolution. (a) The ex-
posed olfactory turbinates were mounted in a 3D-printed perfusion
chamber and imaged using SCAPE microscopy in an inverted configu-
ration. Perfusion solutions flow through the inlet at the nostril and the
outlet at the throat (bold arrows), the field of view covers the turbinate
IIb and III, and some of the neighboring turbinate. (b) 3D volumetric
rendering of SCAPE data acquired on the intact olfactory epithelium
showing baseline GCaMP6f signals in a field of view of 1,170 × 1,000
× 235 (y-x-z, μm). Enlarged side and top views of the acquired volume
are shown in the rectangle and square boxes accompanied with dotted
lines respectively, demonstrating data acquisition with single-neuron
resolution. GCaMP signals were extracted from 10,000’s cells to eval-
uate their individual responses to odor mixes revealing both enhance-
ment and suppression of odor responses by other odors at the sensory
level (Xu et al. 2020).

response to B alone will be subtracted to avoid false interpretation

of enhancement. In addition, we have set an inhibition/enhancement
threshold over 30% to account for the variation of odor responses.
These efforts may cause under-estimation of enhancement and inhi-
bition but are critical to ensure the overall interpretation of results
(Figure 4).
98 Ann-Sophie Barwich and Lu Xu

Figure 4 E
 volution of the SCAPE technique. A comparison is made between
data (3D images of mouse olfactory epithelium) acquired in 2015 and
2019, respectively. Note the improvement in both fields of view and
resolution of the image.

Each challenge was tackled by iterative adjustments, thus resulting in an

iterative improvement of the experimental procedure supporting data
validity, as well as the conceptual refinement of the experimental design
substantiating the possible scope of data interpretation.

Fine adjustments have also been made to minimize rundown by

maintaining the tissue in a healthier state, including carboxgenation
of the Ringer solution, adjustment of the Ringer recipe, improve-
ment of dissection procedure, and optimization of the time interval
between odor stimuli.
The control of photobleaching as well as further optimization of
image quality was achieved mostly by Wenze and his colleagues.
During the constant evolution of the SCAPE setup, the photobleach-
ing effect has been decreased to a minimal level while the resolution
and field of view has been greatly improved.

Obtaining stable measures and solid experimental results were not yet
the end point. A final challenge involved data analysis, to which we come
next.

3.3 Data Analysis

Another notable aspect in the making of the SCAPE study is its attention
to data analysis of cell behavior. Xu analyzed a vast sweep of ~10,000
 Molecular Science Meets Perfumery 99
olfactory receptor cells via genetic tools that target calcium-dependent
activity (GCaMP6f mice). That’s humongous data to analyze and a pleth-
ora of individual cells for which one must identify and interpret their ac-
tivity range and strength. Xu had to modify and work on the algorithm
to accomplish this task. Data analysis took place in several steps:

3.3.1 Preprocessing
The raw SCAPE data are spool files and need a specific script to load
into the workspace of Matlab. Since the specimen was scanned with an
oblique light-sheet, each 3D volume needs to be de-skewed to the orig-
inal scale. This part has been taken care of primarily by Wenze and his
lab-mates.

3.3.2 Image Registration

After optimization of sample mounting, the drift of the sample during
the experiment was minimal and easily corrected with various imaging
processing packages. However, the tissue also underwent deformation
at the scale of μm, which needed to be registered non-rigidly. Since the
diameter of each neuron is only 5–10 um, the goal was to limit displace-
ment at the submicron scale. Due to the lack of apparent “landmarks”
at this scale, registration with commonly used methods (such as Matlab
build-in functions or FIJI packages) did not yield satisfactory results.

To address this issue, I adopted the Non-Rigid Motion Correction

(NoRMCorre) algorithm developed by Pnevmatikakis et al. (2017)
and custom-modified it to perform deformable registration specif-
ically for our own dataset. Since tissue drift and motion artifact
within each 75 s trial was negligible, I took a single 3D volume (av-
eraged from ten volumes during the resting state) from each trial
and registered them to the template, which was also an averaged
3D volume from the reference trial. The displacement field was then
applied to all the volumes in each trial. This modification has sig-
nificantly reduced the calculation amount from the original frame-
by-frame update scheme. One drawback of this algorithm is that
it sometimes results in “outlier” output; to address this problem, I
created a graphical use interface (GUI) to manually correct the reg-
istration when necessary.
The overall registration results generated by the custom NoRM-
Corre scripts were satisfactory. However, manually validating im-
age sub-patches was labor-intensive. To overcome this bottleneck,
I switched to the Advanced Normalization Tools (ANTs) (Avants,
Tustison, and Song 2009), an Insight Toolkit (ITK)-based open-
source toolkit for image registration as well as segmentation,
100 Ann-Sophie Barwich and Lu Xu
template-building, and many other functionalities. ANTs utilize a
strategy termed Symmetric image Normalization (SyN) for deform-
able image registration. After fine-tuning, the performance of SyN
was comparable with that of NormCorre, sometimes slightly better
at resolving localized deformation. Although the SyN-based regis-
tration was still time-consuming (30–60 min per image pair), it did
not require supervision and, therefore, can be automatically pro-
cessed in batch through a custom bash script, which significantly
accelerated the registration process.

3.3.3 Denoising and Demixing of Calcium Signals

OSNs [olfactory sensory neurons] are densely packed in the intact olfac-
tory epithelium. Therefore, it is critical to eliminate the noise from neigh-
boring cells when analyzing a given OSN. This was achieved through the
application of Constrained Nonnegative Matrix Factorization (CNMF),
a framework that implements denoising, deconvolution, and demixing of
calcium imaging data simultaneously (Pnevmatikakis et al. 2016). CNMF
is capable of processing both 2D and 3D data, yet the 3D version is com-
putationally much more expensive. For this reason, Xu split the 3D image
series into multiple non-overlapping 2D stacks and processed them with
2D CNMF. Now, the CNMF algorithm has undergone a series of updates
since the publication of our paper, as well as many other computational
tools, and it’s possible to analyze our data with 3D CNMF in the future.

3.3.4 Data Sorting

The CNMF algorithm (as well as other demixing algorithms) inevitably
picks up numerous non-neuronal components and spontaneously acti-
vated neurons during analysis. For about 10,000 extracted time courses
in each experiment, only ~3,000 are from odor-responsive neurons
suitable for further processing, the rest being motion-induced baseline
fluctuation, spontaneous activity, or unhealthy neurons that stopped re-
sponding during the experiment.

Conventionally, time courses from dissociated neurons were

screened manually. It’s easier in practice than it sounds to screen
several hundreds of neurons, as odor response curves share a clear
temporal pattern that can be easily captured by trained eyes. How-
ever, when the number increases to 10,000, it becomes a different
story. To accelerate the data sorting process, I designed a 2D con-
volutional neural network to perform the initial screening of the
CNMF-­extracted time-courses and spatial loci. Essentially, I trained
a network to learn the way I would pick out the “good” (acceptable)
or ‘bad’ (rejected) neurons with ~90% of accuracy. The screening
 Molecular Science Meets Perfumery 101
results were then manually validated with a custom-designed GUI
to ensure data fidelity.

To recap, the specific requirements in the application of SCAPE required

several types of adjustment, including tissue preparation (e.g., sample
mounting) and data analysis (e.g., material deformation). These adjust-
ments likewise influenced the empirical grounding and validity of the
resulting model interpretation (i.e., greater ecological validity due to
perfusion chamber). Processing vast amounts of data, it soon became
undeniable that the experimental results were striking indeed.

4 The Findings
The patterns of cell populations showed remarkable and widespread
differences in their responses to the mixtures and their components in
isolation. Specifically, up to 38% of cells responding to a mix differed
compared to scans of the collection of cells responding to the individual
odorants in such mixture. More importantly, Xu found two different
and striking effects across receptor cell populations.

4.1 Effect One: Inhibitory Modulation

The first effect may have mirrored some expectations of olfactory re-
searchers concerning the nature of mixture coding. Yet it presented a
vital observation by confirming inhibition effects in odor coding at the
sensory periphery. Exposure to mixtures showed reduced activity across
epithelium cells (Figure 5, left), meaning that, overall, the number of
cells responding to the tripartite mixture was not equal to the sum of the
cells responding to its individual components. At first, this effect seemed
less surprising because of the combinatorial nature of odor coding. Some
odorants in a mixture may simply overlap in their activation of receptor
cells. Still, this explanation was hypothetical. Also, what was not ex-
pected was the sheer amount of inhibition.
Was this observed effect just a sign of combinatorial overlap, or is there
more to the story? Here, the ability to observe and compare the activity
of individual cells with SCAPE cast a new light on what was really going
on. Notably, some cells were responding to individual compounds and
also to binary mixtures. However, these cells remained inactive when
exposed to the tripartite blend. A closer look at cell responses to the
tripartite and binary mix indicated that Isoraldeine (asterisk; Figure 5,
left) served as the primary inhibitor instead of Dartanol (triangle; Figure
5, left). Moreover, Xu saw that some individual cells in the tripartite mix
remained silent but showed responses to at least one of the two binary
combinations of odorants. Consequently, such modulation constituted
not an outcome of combinatorial coding but inhibition.
102 Ann-Sophie Barwich and Lu Xu

Figure 5 A
 veraged heat maps of olfactory sensory neurons responding to two
sets of odor stimuli (in monomolecular solutions, binary and tripartite
combinations). Cells show significant suppression (left) and enhance-
ment (right) effects (Image modified from Xu et al. 2019).

These observations of inhibitory modulation at the periphery match

with behavioral effects mentioned in several psychophysical studies (e.g.,
Cain 1974; Laing et al. 1984; Kay et al. 2005). Now there was clear
evidence that these psychological effects were not necessarily produced
by “higher-level” mechanisms in the central nervous system but already
caused by interactions at the periphery. One key consequence of this
insight is that the expression of perceptual quality and variation in re-
sponse to an olfactory stimulus need not be an effect of cognitive bias
(e.g., familiarity with a stimulus or cross-modal cues), but can also point
to genetic divergence (e.g., in receptor expression).9
Even more remarkable was the realization that the functional role of
an odorant, meaning whether it served as an agonist and an inverse ago-
nist, depended on the mixture, instead of being a feature of the odorant
in isolation. In other words, an odorant could play multiple interactive
roles in varying ligand-binding contexts. Briefly, an odorant might act
as (i) an agonist to one olfactory receptor (i.e., an odorant binds to a
receptor to affect its basal activity), (ii) an inverse agonist to another
receptor (i.e., odorant binds to a receptor to decrease its basal activity),
and (iii) an antagonist (i.e., an odorant binds to a receptor to block other
 Molecular Science Meets Perfumery 103
odorants from affecting its basal activity). Such functional promiscuity
of odorants in mixture coding was quickly confirmed by other studies
(Inagaki et al. 2020; Pfister et al. 2020).

4.2 Effect Two: Enhancement

The second effect Xu noted during the SCAPE experiment would pose a
greater surprise. It would constitute a riddle for months to come. There
also were enhancement effects (Figure 5, right). Some cells showed activity
when exposed to the mixture but did not respond to any individual com-
ponents! These cells only reacted to these compounds when combined and
not in isolation! What did this effect mean? Firestein himself noted that
these results seemed baffling: “I can’t quite make sense of that part yet.”
The breakthrough occurred a couple of days before Barwich had to
submit her book manuscript on the science of olfaction to her publisher.
The book featured a chapter on the SCAPE study (Barwich 2020a,
Chapter 6). Firestein’s first message in his email to Barwich was fore-
shadowing (August 16, 2019): “I think we now know that enhancement
is real and that it makes sense. I will send you a copy of the manuscript
we are about to submit – Lu is putting some finishing touches on it to-
day.” He continued in a follow-up email the day after (August 17, 2019):

So, attached is a confidential draft of the inhibitor/enhancement

manuscript. We will submit by the middle of next week and then
put it up on bioarxiv. You should pay attention to the introduction
and the Discussion. Especially you might be interested in figure 7
(the last figure) [Figure 6 in this Chapter] which is a model of how
this may work in olfactory perception. It’s largely Lu’s idea and I
think it’s very simple and very clever. It is explained pretty well in
the discussion and in the figure legend but if you have any questions,
I’ll do my best to answer them.

Enhancement effects suggest a molecular mechanism at the receptor

level that had never been observed at the sensory periphery, but that was
known from pharmaceutical studies: allosteric modulation. The pres-
ence of an allosteric mechanism implies that a receptor has a potential
binding site in addition to the activity-inducing binding pocket for li-
gands. A ligand binding to the allosteric site does not lead to receptor
activation but results in a change of receptor conformation. In its new
conformation, that receptor is now receptive to components with which
it previously did not interact (Figure 7). By way of example, imagine a re-
ceptor {R1} does not interact with a specific odorant {O1} when O1 is ad-
ministered individually; R1 remains silent. But when O1 is presented as
part of a complex mixture with other odorants what happens is that an-
other odorant {O2} changes the configuration of R1. Here, O2 attaches
104 Ann-Sophie Barwich and Lu Xu

Figure 6 R
 eceptor modulation model in olfaction. Xu’s model of sparse coding
in mixture detection at the sensory periphery in olfaction. Comparison
between the received combinatorial model (left) and the new model of
receptor modulation (right) (Image 7 in Xu et al. 2019).

Figure 7 ( Quite) simplified model illustrating the principle of allosteric interac-

tion in receptor binding.

to the hypothesized allosteric site of R1 and modulates its activity, such

that R1 receptor now interacts with O1.
What was the reasoning process behind the new receptor mechanism
model? The challenge was to link cellular data with coding capacity:

The idea of figure 7 [Figure 6 in this chapter] came up during a dis-

cussion between Stuart and me in his office. We were going through
the manuscript. Everything seems to be there except for one piece
missing – what’s the significance of the modulation effect in the
periphery? With so many flavors of mechanism (partial agonism,
competitive antagonism, synergic enhancement, and probably also
allosteric enhancement), it seems to make the odor-receptor inter-
action more complicated and less predictable. “It’s ok if the modu-
lation effect doesn’t have any significance,” said Stuart, “maybe it’s
 Molecular Science Meets Perfumery 105
a flaw of the system.” Inspired by this comment, I started thinking
about what advantage this mechanism can possibly have. Reflect-
ing on the way we calculated the modulation effect, I realized the
existence of inhibition and enhancement is capable of flipping the
valence of odor response – something that would never happen in
a combinatorial coding scheme, where the only type of operation is
addition.
When I got back home, I told Wenze about this idea and asked
him what he thought about it. For half a minute, he didn’t say any-
thing. “This is so obvious,” he finally replied, “but people would
barely think of it in this way. Yet once you say it out loud everyone
would think it’s already in their mind.” We then shared this idea
with Elizabeth and all four of us gathered together to discuss it.
During this process, this idea has been refined and finally evolved to
the final model presented in our paper.
(It is a pity that we were not able to give strict mathematical proof
of how much it can improve the coding capacity and the associated
boundary conditions, such as the possibility and frequency of mod-
ulation effects.).10

In effect, these observations of various modulation effects explain how

mixtures can adopt odor qualities different from their components’
sensory qualities. Expertise in perfumery meets research in molecular
biology. By having entirely new cells reacting to a complex mixture,
they introduce distinct patterns of signals for the brain to compute into
odor images. The whole is indeed more than the sum of its (chemical)
parts.

4.3 There’s no eSCAPE from Paradigm Shift

Meanwhile, these SCAPE findings offer more than an explanation of the
emergence of some puzzling psychological effects in olfaction by way
of presenting a molecular mechanism. They point toward a revolution
in our understanding of the nature of odor coding and the conceptual
thinking required for its scientific study.11
Xu et al.’s (2020) study highlights that the scientific puzzle of odor
coding cannot be solved by an analytic chemical approach, comparing
chemical features of odorants in isolation and separated from their bind-
ing interactions with the receptors at the sensory periphery. This ana-
lytic chemistry approach had defined large parts of olfaction research
throughout the 20th century prior to the receptor discovery (Barwich
2015b). It still represents a popular, increasingly outdated understand-
ing of stimulus-response modeling in olfaction today (Barwich 2020a).
Xu’s results now presented another damning piece of evidence to the
analytic chemical approach. It suggested, in contrast, that olfaction must
106 Ann-Sophie Barwich and Lu Xu
be approached as a biological system, not a chemical formula, with the
complex forms of regulation that come with that shift in orientation.
Rather than chemistry, the data straightforwardly pointed to a biolog­
ical explanation. Modulation effects did not originate from a chemical
reaction since these organic chemicals are chemically inert under the
above experimental conditions.
Indeed, these modulation effects revolutionize our understanding of
odor coding. They reveal that combinatorial coding is not the only or
even the principal mechanism by which mixture coding in olfaction
operates. Think about it. The combinatorial coding of odor mixtures
would result in an indistinguishable smear of activation across epithe­
lium cells. Firestein remarked (in Barwich, 2020a):

Making conservative estimates that any given odor molecule can

activate 3–5 receptors at a medium level of concentration, then a
blend of just 10 odors could occupy as many as 50 receptors, more
than 10% of the family of human receptors. This will result in fewer
differences between two blends of 10 similar compounds.

Meanwhile, many smells consist of hundreds of aromatic compounds

(Maarse 1991). So how could the brain possibly know and distinguish
what chemicals hit the nose?

The SCAPE study presents a possible answer to the inherent neuro­

computational challenge arising from combinatorial coding at
the periphery: How does the brain discriminate different complex
mixtures from widespread and overlapping receptor activity? An­
tagonistic modulation at the receptor level would facilitate sparse
coding resulting in less ambiguous signal patterns. Aromatic
blends, such as coffee or roses, are composed of hundreds of differ­
ent components. The combinatorial code allows humans to detect
various odorant features in such mixtures and respond to a com­
plex and, in its constituents, unpredictable chemical environment.
However, with combinatorial activation alone, receptor activation
patterns quickly overlapped to form a broad and smudged signal,
which would lose its distinctiveness. Modulation, antagonistic and
allosteric, facilitates a unique receptor code for the discrimination
of complex mixtures.
(Kurian et al. 2021, 5)

This is why the significance of Xu et al. (2020) goes beyond mere techno­
logical finesse using a new tool to advance understanding of an old prob­
lem. The SCAPE study did not just provide some new details and more
depth of observation. It facilitated the development of an entirely new
theoretical approach to the study of odor coding, one that supplanted
 Molecular Science Meets Perfumery 107
the analytical approach and assured a more biologically realistic concep-
tual orientation to the study of the olfactory system.
Xu’s model, summarized in Figure B (and highlighted in Firestein’s
earlier email remarks) introduced a new model explaining how olfactory
signals at the periphery are sufficiently differentiated for the brain to
make sense of complex blends by showing either (i) a reduction of activ-
ity via inhibitory modulation that results in a sparser and thus differen-
tiated signal, or (ii) an enhancement of cell activity that allows the brain
to further distinguish complex mixture signals.
Crucially, the use of SCAPE in Xu et al. (2020) changed the chief tar-
get of explanation in theories of olfaction: explanations of odor coding
are less about the chemistry of the stimulus and more about what signals
actually reach the brain via the receptors.12

5 Discussion: Cognitive Scaffolding in Tool Use and

Development
The value of SCAPE microscopy and its successful application in olfac-
tion goes beyond mere technological affordances that probe unknown
receptor behavior by allowing a scientist to “see more” or “look deeper”
into layers and layers of cell tissue. It is the distinctly cognitive function
that SCAPE played in the research process.
We want to suggest that to understand how tools facilitate new ideas
in neuroscientific experiments, tools and technologies are best analyzed
as a constitutive and active part of the reasoning process of scientists.
This suggestion links to a range of modern theories in cognitive science
associated with the label of “distributed cognition” (e.g., McClelland
et al. 1986; Hutchins 1995). Theories of distributed cognition originated
during the 1980s and have garnered increasing attention in the philo-
sophical and scientific study of cognitive processes, especially over the
past couple of decades (Clark and Chalmers 1998; Clark 2008). Thus, it
is worthwhile implementing these theories further in philosophical ac-
counts of scientific practice (as also proposed, for example, in the works
Giere 2002; Giere and Moffatt 2003; Solomon 2007; Lui, Nersessian,
and Stasko 2008).
Distributed cognition understands the cognitive capacities of
­scientists—and, naturally, other cognitive agents—as embedded and
spread out in a broader system. It’s a systems theoretical approach that
is framing cognition primarily as a process, not in terms of its prod-
uct. According to this view, technologies (like tools or instruments),
scientific representations (such as formulas), social structures (such as
institutes and funding organizations), as well as other scientists (such
as collaborators and competitors) all are part of the cognitive process.
Critical to distributed cognition is the thought that these various ele-
ments are more than causal factors adding further information to the
108 Ann-Sophie Barwich and Lu Xu
content of an individual scientist’s knowledge of a research situation.
Instead, these elements are causally integral to the cognitive process in
which the individual scientist partakes as an epistemic agent. In other
words, these elements do not passively figure in an already well-defined
and independent mental process (e.g., by enhancing, accelerating, or bi-
asing the autonomous reasoning of the scientist). Instead, these elements
often enable and actively structure the cognitive mechanisms in ques-
tion in the first place.
A popular example to illustrate the workings of distributed cognition
is mathematics (McClelland et al. 1986, 44–8). Unless you are a genius
savant, complex calculations are often too complicated for most people
to process in their heads alone. Say, what is the product of 456 × 789?
Aiding the process of enacting such a calculation is an externalization of
the various steps involved, traditionally on paper. This externalization
structures the information so that, if one has learned the appropriate
steps, the resulting symbol manipulations can be correctly and swiftly
processed in the proper order. Other examples can be found in the his-
tory of science, for instance, with the manipulation of chemical formulas
as externalized symbol structures to investigate the elemental dynamics
in chemical reactions. The historian Klein (2003, 2–3) notably described
Berzelian formulas as “paper tools,” remarking:

(…) that chemists began applying chemical formulas not primarily

to represent and to illustrate preexisting knowledge, but rather as
productive tools on paper or ‘paper tools’ for creating order in the
jungle of organic chemistry. Berzelian chemical formulas… were
tools for experimentally investigating organic chemical reactions
and for constructing models of reactions and of the invisible consti-
tution of organic substances.

In the case of SCAPE, we likewise encounter a cognitively significant

structuring of steps in the reasoning process through the way in which
this tool, in parallel with the presentation of new empirical data, serves
as visualization and externalization of conceptual possibilities. What do
we mean by that?
Xu’s theorizing behind the novel sparse coding mechanism at the
olfactory periphery was not merely a consequence of the instrument’s
­theory-ladenness. The construction of SCAPE as a tool surely builds on
previous technological advances and modeling assumptions, including
genetically modified mice and fluorescent markers. These built-in mod-
eling assumptions in tools, of course, shape the way in which data is pro-
duced and what can be seen (e.g., analysis of this issue with the example
of protein synthesis in Rheinberger 1997). However, they cannot explain
how the newly produced data is viewed and interpreted by the scien-
tist. Recall Xu’s description of trying to explain the modulation effects
 Molecular Science Meets Perfumery 109
observed at the olfactory periphery. The model she came up with was
far from obvious or a direct derivation from the tool’s set-up. In other
words, focus on the existing theoretical assumptions that are embodied
and occasionally black-boxed by new technology remains insufficient to
understand how new theoretical models are developed via the use of new
gadgets and tools in scientific experimentation.
By way of example, some of the cognitively significant features struc-
turing scientific reasoning we can observe in the application of SCAPE,
without claims of exhaustiveness, are as follows:

1 Similar to the above example of mathematical calculation: the ap-

plication of SCAPE allowed for an organization and division of
otherwise incomprehensively complex data and data-points into
several steps and combinations—e.g., cells responding to monomo-
lecular, binary, and tertiary mixtures—that provided the basis for
­inference-making about the receptor activity patterns revealed.
2 Similar to the example of Berzelian formulas: the application of
SCAPE facilitated the external visualization of proportions to com-
pare elemental cell dynamics and interactive relationships. More-
over, these visualizations facilitated the possibility to abstract from
the present data set (meaning, the specific stimuli used) to cell be-
havior toward mixtures in general.
3 Additionally, the application of SCAPE coordinated the use of var-
ious other techniques in tandem with this new microscopic tech-
nique (e.g., the development of algorithms, the selection of tissues
and stimuli), as well as the coordination of collaborative activities
between researchers in two laboratories. A similar point is found
in Hutchins’ (1995) analysis of coordination on a naval vessel as
distributed cognition. Concretely, coordination adopts a cognitive
task in how experimental information is translated into proposi-
tional structures to be communicated. This propositionalizing then
scaffolds how the experimental data is integrated into the existing
literature on the topic and how follow-up strategies concerning ex-
perimental adjustments and revision are designed.
4 Further, the iterative tinkering,13 involving repeated adjustments
and modifications in the application of SCAPE, facilitated new op-
portunities not only for data recording but also for theorizing about
the data (and its scope). For example, initial challenges in sample
mounting (Section 3.2) resulted in an experimental set-up that was a
much closer match to the conditions involved in natural odor detec-
tion than previous tissue preparation.

Now, how did this re-organization, visual externalization, and coordi-

nation of data about the behavior of olfactory receptor populations pro-
mote fresh theorizing about odor coding?
110 Ann-Sophie Barwich and Lu Xu
Key to situating the theoretical impact of the SCAPE study in olfac-
tion is the ontological shift that has taken place in this study:

The SCAPE study is a demonstration [of] how new tools allow for a
targeted reconsideration of causal elements as defined by the limits of
previous technologies. Older technologies may have posited different
elements as central and as partaking in different levels of causal mech-
anisms – e.g., the behavior of individual cells versus cell populations.
The response of cells in the context of other cells as a population can
now be modeled on one causal plane and integrated into a causal mech-
anistic explanation with fewer levels of hypothetical mechanisms.
(Barwich 2020c, including a full analysis of the ontological
shift from single cells to cell populations in odor
coding. Also, see Figure 8)

Such ontological shift concerning the nature of causation at the olfactory

periphery does not simply arrive with new data. It stems from new think-
ing. We can compare Xu’s development of a sparse coding mechanism at
the olfactory periphery to the case of chemists fiddling with chemical for-
mulas or mathematicians playing with symbols combinations on paper.

Figure 8 C
 omparison of the ontologies in between the different models of cod-
ing at the periphery. SCAPE changed the ontological structure of odor
coding ad its fundamental components. Components (including their
relevant features) are not defined in isolation but as causal units via
their causal role within a mechanism. This means that the fundamen-
tal level in odor coding is not the {code = single odorant binding com-
binatorically to receptors} but {code = receptor populations responding
to multiple odorants in combination}.
 Molecular Science Meets Perfumery 111
Tools in experimental practice in such contexts, therefore, present an
extension of the thought process that allows for the creation of new pat-
terns, new pattern conceptualizations and combinations, and, in further
consequence, new inferences about the theoretical nature of this data.
The visual display of more comprehensive cell reactions to various
combinations of chemical stimuli carried a powerfully suggestive power
of introducing new significances, i.e., new causal relations and even new
causal constituents (i.e., population behavior as representing the overall
olfactory code, instead of single-cell calculations from individual stim-
uli). This display restructured the intellectual framework by introducing
new concepts, relations, and avenues of further research (i.e., the shift
from stimulus chemistry to receptor biology as a key determinant of
odor coding). Moreover, SCAPE produced new questions: what kind
of mechanism could cause enhancement effects, and to what potential
purpose from a system’s theoretical point of view?
The skeptic of distributed cognition may ask whether such a concep-
tual shift in thinking about odor coding could have also happened with-
out SCAPE. Perhaps someone could have considered this mechanism,
in principle. The point is, however, that it did not happen that way. To
understand the realities of scientific reasoning, we need to look at what
happens at the bench, instead of what is hypothetically conceivable in
the armchair. Besides, a distributed account of cognition does not posit
that those ideas are inconceivable without the appropriate technology.
It asserts that technologies facilitate a broader and different perspective
by being an integral part of the thinking process itself. In other words,
technologies do more than produce data to align with hypotheses—their
use co-creates mental structures in scientific reasoning. As the exam-
ple of SCAPE illustrated, these mental structures include iterative pro-
cedures of conceptual refinement, the direction of attention to specific
empirical observations (which may not always be transparent in their
interpretation), and the combination of otherwise separately occurring
phenomena on one observational place. (In a way, this can be compared
with the use of language, where syntax and semantics afford meaningful
but language-dependent creations, so much so that some works of po-
etry and especially humor are difficult to translate into another language
without sufficiently similar syntactic structure.)
In conclusion, our brief discussion of the cognitive function of tools in
neuroscientific experiments primarily serves as an invitation for philoso-
phers and researchers to further consider the ways in which tools struc-
ture and actively contribute to scientific theorizing. For the analytically
inclined philosopher, this discussion must remain wanting in conceptual
detail and argumentative explication. Yet, given the constraints of this
chapter—with its primary purpose (of providing a first-hand account of
a new tool and in its theoretical impact)—the concluding remarks are
best understood as a pointer toward the promise that targeted cognitive
112 Ann-Sophie Barwich and Lu Xu
studies of tools in experimental practice yields for the study of scientific
reasoning.

Acknowledgments
We are grateful to our colleagues in the Firestein and Hillman labs, es-
pecially Stuart Firestein and Wenze Li. Additional thanks belong to Carl
Craver, John Bickle, and Marco Nathan for their constructive sugges-
tions in the revision of this manuscript.

Notes
1 Review of the issues concerning the heterologous expression of olfactory
receptors in Peterlin, Firestein, and Rogers (2014).
2 Notably, it took almost ten years to find the insect receptors after the mam-
malian receptors had been discovered because they are substantially differ-
ent genetically (Clyne et al. 1999; Vosshall et al. 1999).
3 The “one neuron—one receptor gene” doctrine has been called into question
recently (Mombaerts 2004).
4 This laborious type of work is currently under way, though, for example in
the lab of Hiroaki Matsunami at Duke University.
5 See Passy (1895).
6 Personal communication.
7 The authors disagree on that verdict.
8 Peterlin was a postdoctoral researcher in Firestein’s lab before she moved to
work at Firmenich, Princeton.
9 Detailed analysis of the theoretical implications emerging from recent in-
sights into receptor coding in olfaction in Barwich (2020a).
10 Analysis of the mechanisms of odor coding at the periphery, including the
SCAPE study, and how it links to pattern recognition further downstream
in the central nervous system in Barwich (2020a).
11 Revolution not necessarily in the Kuhnian sense.
12 Further elaboration and details about the implications of this study in the
broader context of olfaction, and why smell needs to be modeled via the
biological processes encoding chemical features (rather than stimulus chem-
istry), in Barwich (2020a, 2020b).
13 The notion of “tinkering” is adopted from Bickle’s (this volume) analysis of
historical examples that look at tool engineering as drivers of theorizing in
neuroscience.

References
Avants, Brian B., Tustison, Nick and Song, Gang. 2009. “Advanced normaliza-
tion tools (ANTS).” Insight Journal 2(365):1–35.
Axel, Richard. 2005. “Scents and sensibility: A molecular logic of olfactory
perception (Nobel Lecture).” Angewandte Chemie International Edition
44:6110–27.
Barwich, Ann-Sophie. 2015a. “What is so special about smell? Olfaction as a
model system in neurobiology.” Postgraduate Medical Journal 92(1083):27–33.
 Molecular Science Meets Perfumery 113
Barwich, Ann-Sophie. 2015b. “Bending molecules or bending the rules? The
application of theoretical models in fragrance chemistry.” Perspectives on
Science 23(4):443–65.
Barwich, Ann-Sophie. 2017. “Is Captain Kirk a natural blonde? Do X-ray crys-
tallographers dream of electron clouds? Comparing model-based inferences
in science with fiction.” In: Thinking about Science, Reflecting on Art, ed.
by Otávio Bueno, George Darby, Steven French, Dean Rickle. London: Rout-
ledge, pp. 62–79.
Barwich, Ann-Sophie. 2018. “How to be rational about empirical success in
ongoing science: The case of the quantum nose and its critics.” Studies in
History and Philosophy of Science 69:40–51.
Barwich, Ann-Sophie. 2020a. Smellosophy: What the Nose Tells the Brain.
Cambridge, MA: Harvard University Press.
Barwich, Ann-Sophie. 2020b. “What Makes a Discovery Successful? The Story
of Linda Buck and the Olfactory Receptors.” Cell 181(4):749–53.
Barwich, Ann-Sophie. 2020c. “Imaging the living brain: An argument for ruth-
less reductionism from olfactory neurobiology.” Journal of Theoretical Biol-
ogy 512:110560.
Barwich, Ann-Sophie and Bschir, Karum. 2017. “The manipulability of
what? The history of G-protein coupled receptors.” Biology & Philosophy
32(6):1317–39.
Bickle, John. 2016. “Revolutions in neuroscience: Tool development.” Frontiers
in Systems Neuroscience 10:24.
Bickle, John. 2019. “Linking mind to molecular pathways: The role of experi-
ment tools.” Axiomathes 29(6):577–97.
Bickle, John. 2020. “Laser lights and designer drugs: New techniques for de-
scending levels of mechanisms ‘in a single bound’?.” Topics in Cognitive Sci-
ence 12(4):1241–56.
Bickle, John. Forthcoming. “ Research Tools in Relation to Theories.” In: The
Tools of Neuroscience Experiment: Philosophical and Scientific Perspectives,
ed. by John Bickle, Carl Craver, Ann-Sophie Barwich. London: Routledge.
Bouchard, Matthew B., Voleti, Venkatakaushik, Mendes, César S., Lacefield,
Clay, Grueber, Wesley B., Mann, Richard S., Bruno, Randy M. and Hillman,
Elizabeth M. C. 2015. “Swept confocally-aligned planar excitation (SCAPE)
microscopy for high-speed volumetric imaging of behaving organisms.” Na-
ture Photonics 9(2):113–9.
Buck, Linda B. 2005. “Unraveling the sense of smell (Nobel lecture).” Ange-
wandte Chemie International Edition 44:6128–40.
Buck, Lind B. and Axel, Richard. 1991. “A novel multigene family may encode
odorant receptors: A molecular basis for odor recognition.” Cell 65(1):175–87.
Bushdid, Caroline, Magnasco, Marcelo O., Vosshall, Leslie B. and Keller, An-
dreas. 2014. “Humans can discriminate more than 1 trillion olfactory stim-
uli.” Science 343(6177):1370–72.
Butterwick, Joel A., Del Mármol, Josefina, Kim, Kelly H., Kahlson, Martha
A., Rogow, Jackson A., Walz, Thomas and Ruta, Vanessa. 2018. “Cryo-EM
structure of the insect olfactory receptor Orco.” Nature 560(7719):447–52.
Cain, William S. 1974. “Odor intensity—Mixtures and masking.” Bulletin of
the Psychonomic Society 4:244.
114 Ann-Sophie Barwich and Lu Xu
Churchland, Paul M. 1996. The Engine of Reason, the Seat of the Soul: A Phil-
osophical Journey into the Brain. Cambridge: MIT Press.
Clark, Andy 2008. Supersizing the Mind: Embodiment, Action, and Cognitive
Extension. New York: Oxford University Press.
Clark, Andy and Chalmers, David. 1998. “The extended mind.” Analysis
58(1):7–19.
Clyne Peter J., Warr, Coral G., Freeman, Marc R., Lessing, Derek, Kim, Junhyong
and Carlson, John R. 1999. “A novel family of divergent seven-transmembrane
proteins: Candidate odorant receptors in Drosophila.” Neuron 22(2):327–38.
Dey, Sandeepa, Zhan, Senmiao and Matsunami, Hiroaki. 2011, “Assaying sur-
face expression of chemosensory receptors in heterologous cells.” Journal of
Visualized Experiments 48:e2405.
Firestein, Stuart. 2005. “A nobel nose: The 2004 Nobel Prize in physiology and
medicine.” Neuron 45:333–8.
Firestein, Stuart, Greer, Charles and Mombaerts, Peter. 2014. The molecular
basis for odor recognition. Cell Annotated Classic (accessed 09/29/2021).
https://els-jbs-prod-cdn.jbs.elsevierhealth.com/pb/assets/raw/journals/re-
search/cell/libraries/annotated-classics/ACBuck.pdf
Gerkin, Richard C. and Castro, Jason B., 2015. “The number of olfactory stim-
uli that humans can discriminate is still unknown.” eLife 4:e08127.
Giere, Ronald. 1988. Explaining Science: A Cognitive Approach. Chicago, IL:
Chicago University Press.
Giere, Ronald. 2002. “Scientific cognition as distributed cognition.” In: The
Cognitive Basis of Science, ed. by Peter Carruthers, Stephen Stich, Michael
Siegal. Cambridge: Cambridge University Press, p. 285.
Giere, Ronald N. and Moffatt, Barton. 2003. “Distributed cognition: Where
the cognitive and the social merge.” Social Studies of Science 33(2):301–10.
Hillman, Elizabeth M. C., Voleti, Venkatakaushik, Li, Wenze and Yu, Hang.
2019. “Light-sheet microscopy in neuroscience.” Annual Review of Neuro-
science 42:295–313.
Hutchins, Edwin. 1995. Cognition in the Wild. Cambridge: MIT Press.
Inagaki, Shigenori, Iwata, Ryo, Iwamoto, Masakazu and Imai, Takeshi 2020.
“Widespread inhibition, antagonism, and synergy in mouse olfactory sensory
neurons in vivo.” Cell Reports 31(13):107814.
Kay, Leslie M., Crk, Tanja and Thorngate, Jennifer. 2005. “A redefinition of
odor mixture quality. Behavioral Neuroscience 119:726–33.
Keller, Andreas and Vosshall, Leslie B. 2016. “Olfactory perception of chemi-
cally diverse molecules.” BMC Neuroscience 17(1):55.
Klein, Ursula. 2003. Experiments, Models, Paper Tools: Cultures of Organic
Chemistry in the Nineteenth Century. Stanford, CA: Stanford University Press.
Kurian, Smija M., Naressi, Rafaella G., Manoel, Diogo, Barwich, Ann-Sophie,
Malnic, Bettina and Saraiva, Luis R. 2021. “Odor coding in the mammalian
olfactory epithelium.” Cell and Tissue Research 338: 445–456.
Laing, David G., Panhuber, H., Willcox, M.E. and Pittman, E.A., 1984.
“Quality and intensity of binary odor mixtures.” Physiology & Behavior
33(2):309–319.
Liu, Zhicheng, Nersessian, Nancy and Stasko, John. 2008. “Distributed cogni-
tion as a theoretical framework for information visualization.” IEEE Trans-
actions on Visualization and Computer Graphics 14(6):1173–80.
 Molecular Science Meets Perfumery 115
Maarse, H., 1991. Volatile compounds in foods and beverages (Vol. 44). New
York: CRC Press.
Mainland, Joel D., Li, Yun R., Zhou, Ting, Liu, Wen LingL and Matsunami,
Hiroaki. 2015. “Human olfactory receptor responses to odorants.” Scientific
Data 2:150002.
Malnic, Bettina, Hirono, Junzo, Sato, Takaaki and Buck, Linda B. 1999. “Com-
binatorial receptor codes for odors.” Cell 96(5):713–23.
Matsunami, Hiroaki. 2016. “Mammalian odorant receptors: Heterologous ex-
pression and deorphanization” Chemical Senses 41(9):E123.
McClelland, James L., Rumelhart, David E. and PDP Research Group. 1986.
Parallel Distributed Processing. Vol. 2. Cambridge: MIT Press.
Meister, Markus. 2015. “On the dimensionality of odor space.” eLife 4:e07865.
Mombaerts, Peter. 2004. “Odorant Receptor Gene Choice in Olfactory Sensory
Neurons: The One Receptor–One Neuron Hypothesis Revis- ited,” Current
Opinion in Neurobiology 14(1):31–36.
Mombaerts, Peter, Wang, Fan, Dulac, Catherine, Chao, Steve K., Nemes, Adri-
ana, Mendelsohn, Monica, Edmondson, James and Axel, Richard. 1996.
“Visualizing an olfactory sensory map.” Cell 87(4):675–86.
Ohloff, Günther, Pickenhagen, Wilhelm and Kraft, Philip. 2012. Scent and
Chemistry: The Molecular World of Odors. Zürich: Wiley-VCH.
Passy, Frédéric. 1895. L’Année Psychologique, second year, p. 380.
Peterlin, Zita, Li, Yadi, Sun, Guangxing, Shah, Rohan, Firestein, Stuart and
Ryan, Kevin. 2008. “The importance of odorant conformation to the binding
and activation of a representative olfactory receptor.” Chemistry & Biology
15(12):1317–27.
Peterlin, Zita, Firestein, Stuart and Rogers, Matthew. 2014. “The state of the
art of odorant receptor deorphanization: A report from the orphanage.” Jour-
nal of General Physiology 143(5):527–42.
Pfister, Patrick, Smith, Benjamin C., Evans, Barry J., Brann, Jessica H., Trim-
mer, Casey, Sheikh, Mushhood, Arroyave, Randy, Reddy, Gautam, Jeong,
Hyo-Young, Raps, Daniel A., Peterlin, Zita, Vergassola, Massimo and Rog-
ers, Matthew E. 2020. “Odorant receptor inhibition is fundamental to odor
encoding.” Current Biology 30:2574–87.
Pnevmatikakis, Eftychios A., Soudry, Daniel, Gao, Yuanjun, Machado, Tim-
othy A., Merel, Josh, Pfau, David, Reardon, Thomas, Mu, Yu, Lacefield,
Clay, Yang, Weijian, Ahrens, Misha, Bruno, Randy, Jessel, Thomas M.,
Peterka, Darcy S., Yuste, Rafael and Paninski, Liam. 2016. “Simultaneous
denoising, deconvolution, and demixing of calcium imaging data.” Neuron
89(2):285–99.
Pnevmatikakis, Eftychios A. and Giovannucci, Andrea. 2017 “NoRMCorre:
An online algorithm for piecewise rigid motion correction of calcium imaging
data.” Journal of Neuroscience Methods 291:83–94.
Poivet, Erwan, Peterlin, Zita, Tahirova, Narmin, Xu, Lu, Altomare, Clara,
Paria, Anne, Zou, Dong-Jing and Firestein, Stuart. 2016. “Applying me-
dicinal chemistry strategies to understand odorant discrimination.” Nature
Communications 7:11157.
Poivet, Erwan, Tahirova, Narmin, Peterlin, Zita, Xu, Lu, Zou, Dong-Jing,
Acree, Terry and Firestein, Stuart. 2018. “Functional odor classification
through a medicinal chemistry approach.” Science Advances 4(2):eaao6086.
116 Ann-Sophie Barwich and Lu Xu
Rheinberger, Hans-Jôrg. 1997. Toward a History of Epistemic Things: Syn-
thesizing Proteins in the Test Tube. Stanford, CA: Stanford University Press.
Rossiter, Karen J. 1996. “Structure-odor relationships.” Chemical Reviews
96(8):3201–40.
Sell, Charles S. 2006. “On the unpredictability of odor.” Angewandte Chemie
International Edition 45(38): 6254–61.
Snogerup-Linse, Sara. 2012. “Studies of G-protein coupled receptors. The No-
bel Prize in chemistry 2012. Award ceremony speech.” The Royal Swedish
Academy of Sciences (accessed 09/25/2014). http://www.nobelprize.org/no-
bel_prizes/chemistry/laureates/2012/advanced-chemistryprize2012.pdf
Solomon, Miriam. 2007. “Situated cognition.” In: Philosophy of Psychology
and Cognitive Science , ed. by Paul Thagard. Amsterdam: North-Holland,
pp. 413–28.
Thagard, Paul. 2012. The Cognitive Science of Science Explanation, Discov-
ery, and Conceptual Change. Cambridge: MIT Press.
Vaadia, Rebecca D., Li, Wenze, Voleti, Venkatakaushik, Singhania, Aditi, Hill-
man, Elizabeth M. C. and Grueber, Wesley B. 2019. “Characterization of
proprioceptive system dynamics in behaving Drosophila larvae using high-
speed volumetric microscopy.” Current Biology 29(6):935–44.
van Fraassen, Bas. 1980. The Scientific Image. Oxford: Oxford University Press.
van Fraassen, Bas. 2008. Scientific Representation: Paradoxes of Perspective.
Oxford: Oxford University Press.
Voleti, Venkatakaushik, Patel, Kripa B., Li, Wenze, Campos, Citlali Perez,
­Bharadwaj, Srinidhi, Yu, Hang, Ford, Caitlin, Casper, Malte J., Yan, Rich-
ard Wenwei, Liang, Wenxuan, Wen, Chentao, Kimura, Koutarou D., Tar-
goff, Kimara L. and Hillman, Elizabeth M. C. 2019. “Real-time volumetric
microscopy of in vivo dynamics and large-scale samples with SCAPE 2.0.”
­Nature Methods 16(10):1054–62.
Vosshall, Leslie B., Amrein, Hubert, Morozov, Pavel S., Rzhetsky, Andrey and
Axel, Richard. 1999. “A spatial map of olfactory receptor expression in the
Drosophila antenna.” Cell 96(5):725–36.
Xu, Lu, Li, Wenze, Voleti, Venkatakaushik, Hillman, Elizabeth M. C. and
Firestein, Stuart. 2019. “Widespread receptor driven modulation in periph-
eral olfactory coding.” bioRxiv 760330.
Xu, Lu, Li, Wenze, Voleti, Venkatakaushik, Zou, Dong-Jing, Hillman, Eliza-
beth M. C. and Firestein, Stuart. 2020. “Widespread receptor driven modu-
lation in peripheral olfactory coding.” Science 368(6487):eaaz5390.
Zhao, Haiqing, Ivic, Lidija, Otaki, Joji M., Hashimoto, Mitsuhiro, Mikoshiba,
Katsuhiro and Stuart Firestein. 1998. “Functional expression of a mamma-
lian odorant receptor.” Science 279:237–42.
5 A Different Role for
Tinkering
Brain Fog, COVID-19, and
the Accidental Nature of
Neurobiological Theory
Development
Valerie Gray Hardcastle and
C. Matthew Stewart

John Bickle and his colleagues are current drivers of the science-in-­
practice movement and its focus on the extra-theoretical aspects of sci-
ence for philosophers (see, e.g., Ankeny et al. 2011). This movement grew
out of an intellectual frustration with the near divorce of the philosophy
of science from what scientists actually do in their day-to-day profes-
sional lives. Its adherents promote the detailed and systematic study of
the activities of science, while still maintaining the traditional philoso-
phy of science foci on rationality, justification, observation, evidence,
and theory (and in contrast to prevailing tendencies in the social studies
of science and technology, which focus almost exclusively on science as
a human construct). Philosophers of science-in-practice explore the his-
torical context and current practices that lead to a model or theory in an
effort to better understand the scientific process.
Bickle holds that revolutions in neurobiology-in-practice are driven
exclusively by new tool development, which he believes drives experi-
mental design (2016) – as opposed to, say, Thomas Kuhn’s (1962) model
of scientific theory development as a paradigm shift. He argues that
tool development is an analog of Ian Hacking’s (1983) famous “micro-
scopes” argument for the “relative independence” of “the life of exper-
iment” from theory, further downgrading theory’s purported centrality
in science (Bickle 2018). And following Hacking, Bickle holds that en-
gineering counts more than theory in the development of molecular and
cellular neurobiology research tools. Indeed, according to Bickle,

the genesis of every piece of theory [in neurobiology] … can be tied

directly to the development of new research tools. Theoretical prog-
ress in this paradigmatic science of our times is secondary to and
entirely dependent upon new tool development, both temporally
and epistemically.
(2021, p. 14, emphasis in the original)

DOI: 10.4324/9781003251392-7
118 Valerie Gray Hardcastle and C. Matthew Stewart
The larger, and perhaps more controversial, point that Bickle wants to
make is that tool development drives everything: not just theory develop-
ment, but resource allocation, experimental design, and researchers’ time
and attention. Bickle notes that Hubel and Weisel’s advancement in our
understanding of the receptive field properties in striate and extra-striate
cortex, which led to their theory of information processing in the visual
cortex, and which garnered them the 1981 Nobel Prize in Physiology
or Medicine, “rode on the back of Hubel’s catch-as-catch-can, make-
it-work, trial-and-error-tinkered invention of a new experiment tool,
the metal microelectrode and hydraulic microdriver, and the new kinds
of experiments it enabled” (2022, p. 18). If theories weren’t diminished
enough in their place in scientific practice, Bickle is further claiming that
tool development in turn relies on what he calls “laboratory tinkering,”
and what Hacking calls “fooling around” (1983, p. 199). So, actually,
“theory progress in neurobiology is thus doubly dependent, and hence
tertiary in both epistemic and temporal priority” (Bickle 2022, p. 14).
Tinkering in the lab leads to creating new tools for new experimental ap-
proaches, which, in turn leads to theoretical progress (see also Stix 2012).
Bickle goes on to surmise that perhaps this “double dependence” of
theory on tinkering and new tool development might be true across all
the bench scientific fields in the life sciences. These lessons taken from his
examination of neuroscience-in-practice stand in direct contrast to the
lament of many of the early authors in the philosophy of neuroscience:
that there was no theoretical framework for understanding the brain (e.g.,
Crick 1979; Churchland 1986; Churchland and Sejnowski 1992). Still
today, the lack of theory in neuroscience remains a concern among many
(Gold and Roskies 2008; Churchland and Sejnowski 2016). And still to-
day, there is no overarching theoretical framework in which to organize
the brain sciences (Hardcastle and Hardcastle 2015; Hardcastle 2017).
Bickle suggests that the lack of theory in neuroscience should not be
of great concern; instead, he follows Hubel’s explanation that the activi-
ties of neuroscientists “seemed not to conform to the science that we are
taught in high school, with its laws, hypotheses, experimental verifica-
tion, generalizations, and so on” (1996, 312). The actual impact of theory
on the day-to-day practice of neuroscience appears small. Instead, neuro-
scientists tinker around in their laboratories, intervening in the world as
they can with the tools they have (or create) and then seeing what comes
of it. As Hubel explains, they are akin to “15th century explorers, like
Columbus sailing West to see what he might find” (1996, p. 312).
But is this how bench neuroscience works across the board? This
chapter argues that the answer is yes and no. On the one hand, tool de-
velopment and tinkering are fundamental aspects of the neuroscientific
trade. But on the other hand, theories are just another tool that neuro-
scientists can utilize when poking about in the brain. Theories are not
tertiary outcomes of neuroscientific practice; rather, they are a utensil,
 A Different Role for Tinkering 119
just like a microelectrode, with which we use to probe the brain. We
illustrate this perspective by recounting two recent discoveries on the
impact of COVID-19 in the brain, both of which were supported by
the tinkering and practices of one scientist. In one case, theory clearly
drove tool innovation; in the other, as with Columbus, he was just sail-
ing West, hoping against hope to find something useful. But each case
resulted in significant theoretical advancement. Sometimes novel tools
drive theory, but other times, theory promotes novel tools. And both are
ways that neuroscience advances.

1 COVID-19
On 31 December 2019, China notified the World Health Organization
(WHO) about a cluster of cases of pneumonia in Wuhan City, home to
11 million. Less than two weeks later, there were 282 confirmed cases,
with six deaths. Four of these cases were outside of China in neighboring
countries. During this same timeframe, the virus responsible had been
isolated and its genome sequenced. The cause of the novel pneumonia
became known as COVID-19 (short for COronaVIrus Disease 2019),
and it was a new coronavirus,1 dubbed SARS-CoV-2 (Chaplin 2020;
Lago 2020). Less than eight weeks after that, there were over 118,000
cases and almost 5,000 deaths across 114 counties, and the WHO calls
it a pandemic. And now, at the time of this writing, a short 22 months
later, the WHO reports over 230 million cases worldwide, with almost
5 million dead (https://covid19.who.int). Forty percent of those cases and
half of the deaths have been in the Americas.
While it is not definitively known how or why the new coronavirus
appeared in humans, genomic analysis suggests that it originated in bats
and was transmitted to perhaps a pangolin, which was illegally sold
at a wet market in Wuhan, where it was then transmitted to humans
(Anderson et al 2020). SARS-CoV-2 is not the first coronavirus to have
been passed from other animals to humans – six have been identified
so far, four of which cause the common cold. A fifth is now known by
SARS (Severe Acute Respiratory Syndrome), which also originated in
China, likely also via bats (then perhaps through civets), in 2002. In the
two years in which it was active, just over 8,000 cases were identified
with roughly 10% of those infected dying. And the sixth, abbreviated
as MERS (Middle Eastern Respiratory Syndrome), originated in Saudi
Arabia in 2012, again likely via bats and then through camels. Unlike
SARS, MERS continues to infect humans to this day, with roughly
2,500 known infections and 860 deaths. Both SARS and MERS cause a
flu-like illness with symptoms ranging from asymptomatic infection to
the sniffles to severe pneumonia to acute respiratory distress syndrome,
septic shock, and multiorgan failure (Kahn and MacIntosh 2005; Chap-
lin 2020; Liu et al. 2020).
120 Valerie Gray Hardcastle and C. Matthew Stewart
MERS now seems to appear most often as a result of animal to human
transmission (perhaps as a function of tending to camels’ birthing their
young) and only secondarily through human spread. This coronavirus is
not terribly contagious and requires very close contact between humans
to be transmitted this way, such as occurs when one is caring for someone
who is ill (Durst et al. 2018). MERS is transmitted via respiratory aerosol
droplets and airborne particles, as is COVID-19. However, relative to
MERS, COVID-19 appears more contagious. This is probably largely
due to asymptomatic carriers and genetic mutations (Lago 2020). In the
four weeks after 55 index cases had been identified of SARS in East Asia,
an additional 3,000 cases had appeared (WHO 2003). But in the four
weeks after the first 59 cases of COVID-19 were identified, 25,000 cases
had been confirmed – an eight-fold increase in new cases (WHO 2020).
The COVID-19 pandemic is often compared to global flu outbreaks
as a way of putting the disease in context. However, seasonal flus kill
250,000–650,000 annually, depending on the outbreak. And the H1N1
swine flu epidemic in 2009–2010 caused 280,000 deaths worldwide (Kelly
et al. 2011). Obviously, these numbers do not compare with what we have
seen with COVID-19. The closest is perhaps the H1N1 flu pandemic in
1918–1919, in which 50 million worldwide died, with 650,000 of those
in the United States (Centers for Disease Control and Prevention 2019).
As with SARS and MERS, the clinical presentation of COVID-19 var-
ies from asymptomatic infection to mild illness to severe disease and
death. Sudden deterioration is common, most often in the second week of
the disease, and attributed to cytokine “storms2” or hyperinflammation.
Risk factors for severe disease include older age, immunosuppression,
hypertension, diabetes, and cardiovascular disease, chronic respiratory
disease, and kidney disease. Some of the risk factors could be explained
by the virus’s high affinity for binding to ACE2 (angiotensin-converting
enzyme 2), which is expressed by the epithelial cells in lungs, intestines,
kidneys, and blood vessels (Gouvea dos Santos, 2020).

2 COVID-19 in the Ear

Important to our discussion, the most common symptom of COVID-19
is a sudden loss of taste or smell, reported in 86% of those with mild
cases and over a third with more severe cases (Lechien et al. 2021). Newly
diagnosed COVID-19 patients are over 27 times more likely to complain
of loss of smell as compared to non-COVID-19 patients (Wagner et al.
2020). Often it is used as the primary identifying symptom of the disease.
Evidence suggests that this loss of smell happens in the nasal epithe-
lium in the upper nasal cavity, as neither the olfactory sensory neurons
nor the olfactory bulb, which receives input from the olfactory sensory
neurons, express the gene that encodes for the ACE2 receptor protein
which SARS-CoV-2 uses to enter host cells (Brann et al. 2020). The
 A Different Role for Tinkering 121
olfactory epithelium is a specialized tissue in the upper nasal cavity that
houses olfactory neurons and allows for odor detection. Both olfactory
sustentacular cells, which provide structural and metabolic support to
olfactory neurons, and olfactory basal cells, which act as stem cells that
regenerate the olfactory epithelium after damage, in the olfactory epithe-
lium express ACE2.
The hypothesis is that because olfactory mucosa is in such close prox-
imity to olfactory bulb projections, infected olfactory mucosa will af-
fect olfactory bulb processes (likely via an inflammatory response), even
though the virus itself will not bind to cells in that region (Meinhardt
et al. 2021). In any event, it quickly became obvious that otolaryngolo-
gists could not operate in the nose or mouth because of the potential for
aerosol spray from infected nasal and throat passages. Indeed, some oto-
laryngologists in China, Italy, and Iran died from operating on patients
early in the pandemic (Kulcsar et al. 2020).
COVID-19 has been billed as a respiratory illness (cf., WHO, 2021).
This description is not entirely off the mark, as SARS-CoV-2 is primarily
transmitted by breathing in the virus from the air, and it does indeed bind
with the epithelial tissue found in the lungs. However, and the import
of this point was originally missed in the first few months of the SARS-
CoV-2 outbreak, epithelial respiratory tissue is found outside the lungs.
Our respiratory tract does more than support breathing; it also removes
waste brought into the body along with the atmospheric gases we need to
function. Part of respiration includes a neurodigestive system, as it were.
For example, the epithelium in the upper nasal cavity comprises part of
our respiratory tract; our sinuses are a component of this neurodigestive
system. In addition to smelling the world around us, they move detritus
into our gut to be digested. Sinuses make mucus, which traps dust, pol-
len, dirt, and other unwanted particles breathed up from the nose, and
then sweeps this mucus to the back of the throat to swallow down as our
own internal garbage disposal system. A similar process occurs in our
lung epithelium. It consists of, among other cell types, goblet cells, which
produce and secrete mucus to trap pathogens and debris within the air-
way, and ciliated cells, which help sweep the trapped debris across the
airway tract to the back of the throat, where it too is swallowed down.
But the nasal passages and the lungs are not the only epithelial re-
spiratory tissue in our bodies. Our middle ear and mastoid also have
respiratory functions. The mastoid bone consists of a cell structure that
contains air pockets and is lined by modified respiratory epithelium. The
question of how the human tympanic cavity came to be lined by a con-
tinuous epithelial layer while the cavity is separated by ossicles, muscles,
and nerves, has been debated for years. The respiratory-like epithelium
was originally considered an extension of the upper respiratory tract,
thought to have developed because the pharyngeal pouch was turned
inside out in development, and the most recent evidence suggests that
122 Valerie Gray Hardcastle and C. Matthew Stewart
this hypothesis is still a good one (van Waegeningh et al. 2019). Re-
gardless of how the epithelium came to be in the inner ear, the mastoid
produces mucus and, in conjunction with the middle ear, sweeps ear
detritus trapped in the mucus through the Eustachian tube to the back
of the throat for swallowing, just as our other sinus cavities do.
We know that SARS-CoV-2 can infect the epithelial respiratory cells
in the lungs, which causes shortness of breath. We know it can infect
the epithelial respiratory cells in the sinuses, which causes the loss of
smell.3 And we know that SARS-CoV-2 can affect the olfactory system
of the brain via the mucosa in the upper nasal cavity. It therefore stands
to reason that SARS-CoV-2 could affect the epithelial respiratory cells
in the ear too. In other words, the virus could migrate to the middle
ear and infect the mastoid in COVID patients. And if it can infect the
mastoid, then perhaps it could also impact the central nervous system in
an analogous fashion to what happens between the nasal cavities and
olfactory bulb projections. These are the hypothesis that Matt Stewart
wanted to test.
There were some additional hints that these hypotheses might be true.
For example, viral infections can cause hearing loss, vertigo, and oc-
casionally tinnitus. Significant percentages of adults who have tested
positive for COVID-19 do report sudden hearing loss (8% of all cases),
vertigo (7%), or tinnitus (15%) among their symptoms, regardless of
the level of severity of the disease (Almuffarrij and Munro 2021). We
are starting to see longer-term effects on the auditory system for pa-
tients with “long-hauler” syndrome,4 including earaches, vertigo, loss of
hearing, and tinnitus, all persisting a month or more (Assaf et al. 2020).
Problems with hearing or dizziness generally reflect a challenge in the
inner ear, but tinnitus is not a peripheral phenomenon; it is a central
one. Its sufferers perceive sounds when they are not present, and sound
perception requires the involvement of auditory neurons.
Moreover, mouse data indicate that ACE2 is diffusely present in the
Eustachian tube, middle ear spaces, and cochlea of the mouse ear. Intra-
cellular entry of SARS-CoV-2 via ACE2 depends on spike protein cleav-
age mediated by Furin and spike protein priming by transmembrane
protease serine 2 (TMPRSS2). Both Furin and TMPRSS2 are present in
mouse ear tissue as well (Uranaka et al. 2020), suggesting that the mouse
ear could harbor SARS-CoV-2, which in turn suggests that human ears
could too.
In sum, there was evidence for both peripheral and central effects of
COVID-19 on the auditory system, as well as evidence that SARS-CoV-2
would be able to bind with the epithelial tissues found in the inner ear. If
these data were accurate, then the impact of SARS-CoV-2 could extend
to the auditory system as well.
To test the hypothesis that SARS-CoV-2 can infiltrate the auditory
system, one must examine the inner ears and brains of COVID-19
 A Different Role for Tinkering 123
patients. But doing so presents immediate technological challenges, for
there is no way to do this non-invasively in live patients. At the same
time, there was no good way to do this in deceased patients either, for
the normal protocol for accessing the brain in cadavers requires power
tools, which, of course, could easily broadcast the virus throughout the
autopsy theater. A new approach was needed – perhaps not new tools,
but at least a different way to access brains that did not utilize con-
temporary Western methods for doing so. This was the challenge that
confronted Stewart: If he wanted to test his hypotheses, then he would
need to devise a way to access human brains that would not risk infect-
ing himself or others with SARS-CoV-2. In this particular instance,
theory and hypothesis-testing are driving tool development, instead of
the other way around.
To cut to the end of the story, Stewart used essentially a hammer, a
chisel, and a tiny scoop to access the middle ear and auditory neurons.
These were not new tools, but rather old tools used in a novel way. These
comprised the surgery equipment used by head and neck surgeons in de-
veloping countries, who do not have access to modern mastoid surgery
drills or other power tools (Fagan and Jackler 2017). These tools in turn
reflect surgery techniques from the 1930s in the United States (Portman
et al. 1939) (See Figure 1 below).

Figure 1 M
 allet, rasp, chisels, curette, and gouge. From Fagan and Jackler
(2017), p. 2.
124 Valerie Gray Hardcastle and C. Matthew Stewart
Stewart partnered with the rapid autopsy program for inpatient
COVID-19 deaths at Johns Hopkins School of Medicine, pioneering the
use of hand tools as a safer method for autopsying the brains potentially
infected with SARS-CoV-2. To investigate whether the virus could be
found in the middle ear, he started with the usual method for detecting
SARS-CoV-2: the PCR (polymerase chain reaction) swab test to search
for genomic sequencing of SARS. He accessed the inner ear by trephin-
ing and curetting through the mastoid bone using a small hammer and
a chisel and then swizzled in the middle ear with a tiny scoop or gouge.
He and his team discovered that decedents who had died from compli-
cations of COVID-19 exhibited a very strong positive result to the PCR
test. Furthermore, this positive result continued to be evident for quite
long postmortem intervals – up to 96 hours after death (Frazier et al.
2020). SARS-CoV-2 could be found in human ears. (As an aside, this
discovery had immediate implications for medical care, because it was
now clear that providers should not peer into ruptured ears as part of
routine medical examinations because COVID-19 could potentially be
transmitted that way.)
But there is more. Stewart and others had good reason to suspect that
SARS-CoV-2 was infiltrating epithelial tissue in the inner ear. But to
confirm this suspicion, they would need to determine whether ACE2
receptors are indeed present in the human tympanic cavity. The natural
approach would be to use electron microscopy to examine which cells, if
any, in the inner ear have ACE2 receptors, or other markers like Furin or
TMPRSS2, which would promote binding with SARS-CoV-2.
This again is straightforward scientific hypothesis-testing. No ACE2
receptors, Furin, TMPRSS2, or similar would contraindicate the ability
of SARS-CoV-2 to infect the inner ear, which would further suggest
that the central effects that some patients with COVID-19 experienced
would have to be secondary implications of COVID-19. If SARS-CoV-2
cannot directly infect cells in the inner ear, then COVID-19-related tin-
nitus could not be due to SARS-CoV-2 itself but perhaps instead due to
inflammation or other responses by the immune system to fight off the
infection.
But using electron microscopy to examine which cells from the tym-
panic cavity express ACE2 is not an easy protocol to adopt because our
ear bones are among the hardest bones in our bodies. And, unfortu-
nately, the decalcification process required to soften them enough to be
able to slice them thin enough for microscopy examination normally
takes ten months – much too long in our race to understand the impacts
of SARS-CoV-2 on the human body. But once again, Stewart innovated
by modifying techniques from other eras used for other purposes. In
this instance, he reached back to his undergraduate studies and adapted
the procedures he had learned in the chemistry lab to thinly section and
decalcify rocks. Amazingly, using this novel method for decalcifying
 A Different Role for Tinkering 125
bones took only 13 days to soften the bones enough for microscopy
examination.
At the time of this writing, Stewart is now harvesting the entire ves-
tibular/cochlear part of the temporal bone for examination. While this
work is still ongoing, enough is known to be able to conclude definitively
that SARS-CoV-2 can infect the cochlear region. It is not just that there
were viral proteins present in the inner ear of COVID-19 cadavers in the
mucosa, but the bone tissue itself in the tympanic cavity can be infected.
Just as SARS-CoV-2 is ubiquitous in the lungs, nasal passages, and vas-
cular systems in patients with severe COVID-19, it is also present in the
ears.
Once again, we see theory driving tool innovation and not the other
way around. In this instance at least, instead of “the genesis …of the-
ory [in neurobiology] …[being] tied directly to the development of new
research tools” such that “theoretical progress …is secondary to and
entirely dependent upon new tool development, both temporally and
epistemically” (Bickle 2022, p. 14), we are seeing that the confirmation
of a neurobiological theory was tied directly to the development of new
research techniques such that while theoretical progress was indeed tem-
porally dependent upon new tool development, it was not epistemically.
However, we do not wish to push too hard on this conclusion or to try
to generalize it very far. In our next case study, we see that the same re-
searcher, still investigating the neurological implications of COVID-19,
this time does not have a hypothesis to drive his innovation. In this next
case, we see that theoretical progress was both temporally and epistem-
ically dependent on tool innovation.

3 COVID-19 and the Brain

As a reminder: the primary symptom of COVID-19, the loss of smell,
is neurological. Moreover, it is now very clear that many patients with
COVID-19 – both those with severe cases and those with mild – develop
additional neurological symptoms, often days or weeks after disease
onset. These symptoms can include headache, chronic pain, confusion,
memory loss, fatigue, insomnia, excessive drowsiness, alterations of
consciousness, anxiety, depression, and even psychosis. Quite common
is what is known as “brain fog,” or more formally as dysexecutive syn-
drome (Couzin-Frankel 2020, Helms et al. 2020). Alarmingly, over a
third of patients with COVID-19 receive a diagnosis for some psychi-
atric or neurological ailment within six months of getting infected with
the virus (Zubair et al. 2020).
Other viruses have had this impact on our brains, most notably a
variant of the measles virus and the strain of flu from the 2918 pandemic
(Hoffman and Vilensky 2017; Garg et al. 2019). Very recent analysis
suggests that the molecular neuropathologies found in decedents who
126 Valerie Gray Hardcastle and C. Matthew Stewart
had had COVID-19 roughly mimic those of Alzheimer’s disease, multi-
ple sclerosis, Huntington’s disease, autism spectrum disorder, and other
chronic CNS diseases. While the COVID-19 changes found in neurons
were not exactly the same as these other disorders, the overlap between
the two was striking (Yang et al. 2021).
These findings are worrisome, to put it mildly. A significant concern
among neurologists and clinical neuropsychologists is the possibility
that SARS-CoV-2 infection could lead to long-term cognitive dysfunc-
tion or decline (Pratt et al. 2021). While it is much too soon to be able to
know definitely whether humanity will be struggling with the long-term
effects of COVID-19 among a large proportion of our population for
decades to come, at the moment, this does appear to be a real and very
sobering possibility.
How might SARS-CoV-2 do this? One suggestion is that the virus
could pass through the blood-brain barrier and infect the brain directly,
as it apparently can in human’s close biological cousin, the rhesus mon-
key (Jiao et al. 2021). Another possibility is that the inflammation as-
sociated with the release of cytokines by our immune system as it fights
off SARS-CoV-2 could promote the accumulation of tau plaques in our
brain, which perhaps could eventually lead to dementia or Parkinson’s
disease (see, e.g., Heneka et al. 2020). And, of course, both hypotheses
could be true as well.
However, evidence for the viral RNA or proteins in the brain itself is
extremely sparse, including in the olfactory bulb (Mukerji and Solomon
2021), which is the neural system closest to the primary infection site.
Quantitative reverse transcriptase polymerase chain reaction tests (qRT-
PCR) did find very low levels of viral RNA in some of the brains (Puelles
et al. 2020, Solomon et al. 2020, Wichmann et al. 2020), but this was
much, much less than what was detected in the nasal epithelium – and
RNAscope and immunocytochemistry have not detected any viral RNA
or proteins in any brains studies thus far (Thakur et al. 2021).
At the same time, neuropathological examination of a variety of brain
areas from numerous patients who have died from SARS-CoV-2 shows
evidence of ischemic strokes, both large and small, with areas of dead
tissue, both globally and locally (Duarte-Neto et al. 2020, Reichard
et al. 2020; Remmelink et al. 2020; see also Lee et al. 2021; Thakur
et al., 2021). Given the minute levels of SARS-CoV-2 found in cadaver
brains, there is no obvious way for the virus to cause the stroke damage
found in the decedents. Clearly, something is happening in the brains of
persons with COVID-19, but what and how?
Unlike in our previous case, there was no theory to guide this inves-
tigation. Somehow, SARS-CoV-2 was instigating deleterious changes in
the brain. It was not doing this by direct infection; it likely was somehow
causing multiple small stokes. But what that mechanism might be and
how that mechanism was inducing ischemia was unknown. In this case,
 A Different Role for Tinkering 127
Stewart had to just start looking around in the brains of patients who
had died from COVID-19 to see if he could determine what might be
causing the changes. He had no real theory to guide his investigation,
just the tools he developed to autopsy COVID-19 cadavers and basic
histopathology. This was tinkering in its purest form.
As before, the autopsies would have to proceed without the usual
powered autopsy saw because of the potential for aerosol spray. Building
upon the older surgical techniques Stewart used to dissect the inner ear,
he attempted to access the temporal lobe without power tools. Needless
to say, it was quite physically challenging to chisel through the skull in
order to remove it and the dura while maintaining an adequate N95
facial seal!
The goal was to access the temporal lobe, which is not normally a tar-
get in autopsy, but the first attempts ended in a failed mess. With more
practice, and a new PAPR (Powered Air Purifying Respirator), Stewart
and his colleagues were able to successfully perform whole-brain ex-
traction and dissect randomly targeted cortical areas. There they discov-
ered an unexpected type of cell had clogged the capillaries in COVID-19
patient brains. Surprisingly, what they discovered were megakaryocytes,
the extremely large cells that normally reside in bone marrow and make
the platelets found in blood (see Figure 2) (Nauen et al. 2021).
It turns out that in times of inflammation, megakaryocytes can be
stimulated to move from bone marrow plates and to start circulating
in blood vessels. Because they are so large, they can directly block both
small capillaries as well as end-organ vessels. Autopsies of decedents with
COVID-19 had previously uncovered circulating megakaryocytes in the
liver, kidney, heart, and lungs. (This can occur with other diseases as
well.) These autopsies also showed that capillary blockage had occurred
in these same organs – regardless of when in the course of the disease the
patient died and sometimes despite full anticoagulation. The discovery
of megakaryocytes along with platelet-infused thrombi in the liver, kid-
neys, heart, and lungs suggests that these cells are playing a role in the
blockages. But megakaryocytes are not routinely found in the brain as
part of any normal process (Duarte-Neto et al. 2020, Rapkiewicz et al.
2020). (A few rare genetic diseases are correlated with megakaryocytes
in the brain, however.)
Nevertheless, a third of the brains that Stewart sampled contained
these megakaryocytes (see also Jensen et al. 2020). Because brain au-
topsy sections sample only a small portion of the entire cortex, finding
megakaryocytes at all indicates that the total number of them circulating
in brains could be quite large. By tracing the impact of these occlusions
through downstream tissue, Stewart and his colleagues determined that
the megakaryocytes spilled platelets in the brain, which then initiated
a thrombotic cascade. If large numbers of megakaryocytes block blood
flow in this manner in capillaries across the brain, then they might be
128 Valerie Gray Hardcastle and C. Matthew Stewart

Figure 2 S canning electron microscope image of a megakaryocyte in the mar-

row with red blood cells for comparison. BSIP SA.

creating a distinct pattern of ischemic destruction, which could poten-

tially lead to a distinct type of COVID-19 neurologic impairment.
One of the big mysteries about the long-term CNS sequelae after re-
covery is the ongoing brain fog that occurs in a disturbingly large num-
ber of patients, given that finding viral RNA or active infection in the
CNS is extremely rare and very limited. Recent comparison scans of al-
most 800 individuals pre- and post-SARS-CoV-2 infection indicate that
there is a consistent and significant loss of grey matter across cortex,
with the most profound effects being found a reduction of thickness
and volume in the gustatory-, olfactory-, and memory-related regions
of the brain (Douaud et al. 2021). Most disquieting is the limbic na-
ture of the gustatory and olfactory systems, and their proximity to the
hippocampus, which not only might account for the brain fog but also
might indicate that one long-term consequence of COVID-19 could very
well be dementia for some number of patients. And these changes could
 A Different Role for Tinkering 129
be wrought by multiple tiny strokes caused by megakaryocytes running
rampant in our brains.
Of course, we do not yet definitely know whether megakaryocytes
cause the neurological problems associated with long-hauler syndrome.
Nevertheless, this case presents a clear case of technological innova-
tion driving theory production: In this case, tinkering in the lab led to
novel uses for older (now unused) tools that opened a new experimen-
tal approach, which, in turn led to at least the potential for theoretical
progress.

4 Tool Innovation in Neurobiology-in-Practice

Increasingly, evidence suggests long-term brain involvement COVID-19.
Acutely ill patients are often confused and exhibit what appear to be
CNS disorders. Post-recovery, many patients complain of ongoing anx-
iety, depression, and difficulties with cognition. The challenge has been
how to reconcile patient experience with hitherto inconclusive brain
studies. Due to restrictions by the Centers for Disease Control and Pre-
vention (and common sense), routine autopsies cannot be performed on
the brains of COVID-19 patients because they involve instruments that
generate aerosol spray. But, after tinkering with a variety of different
approaches to get around this restriction, researchers at Johns Hopkins
School of Medicine returned to the lessons from previous generations
of practitioners and adapted old-fashioned manual surgery techniques.
With these new protocols for autopsy, their dissections revealed both
strong evidence for SARS-CoV-2 in the epithelial tissues of the inner ear
and the presence of megakaryocytes in the brain tissue of COVID-19
decedents. Both of these discoveries have huge implications for our un-
derstanding of how SARS-CoV-2 affects the body, both over short-term
and long-term periods.
Aside from being an ingenious solution, returning to hand-held tools
for autopsying brains is a textbook case of How Science Works: re-
searchers working to find the right tools that would allow them to peel
the lid off of a black box and figure out what is going on. In these cases,
the novel use of these tools is not informing theory per se; rather, the
desperate need to uncover the causal-mechanistic chain behind the neu-
rological sequelae of COVID-19 drove the innovative tool use. And then
the innovative use helped to confirm a hypothesis in one instance and
allowed for the development of a hypothesis in the other instance.
What do we learn from these case studies? In many respects, the les-
sons echo Barwich’s example of how a tool was exploited to develop
a new theoretical perspective in an otherwise rather hopeless circum-
stance. In the first case, contra Bickle’s examples, this theoretical ad-
vance was both predicted and expected; but in the second case, as in
Bickle’s examples, the advance was essentially an accidental finding.
130 Valerie Gray Hardcastle and C. Matthew Stewart
We agree that directed tinkering is indeed connected to significant
conceptual theoretical advancement more often than not in neurobi-
ology. But how it is connected can differ depending on the specifics
of the case. Sometimes tinkering and novel tool use drive hypothesis-­
testing, and sometimes they drive hypothesis generation. As adum-
brated by philosophers focused on science-in-practice, the processes
by which researchers come to their theories are as important, if not
more important, than the theories themselves. For, as many of the
papers in this volume illustrate, technological advancement is the
primary limiting factor in theory confirmation and advancement in
neurobiology.

Notes
1 The name coronavirus is derived from the Latin word corona, which means
crown or wreath. The name is due to the shape of virions on the surface of
the virus as visualized by electron microscopy creates an image reminiscent
of a crown (Chauhan 2020).
2 Cytokine literally means “cell mover,” which is exactly what these storms
do – they move cells from one part of the body to another.
3 We also know it can infect the epithelial cells in our blood vessels, but this
fact is not part of our current story.
4 “Long haulers” refer to patients who previously tested positive for COVID-19
but continue to experience adverse symptoms weeks or months after the vi-
rus has been cleared from the body.

References
Almufarrij, I., and Munro, K.J. (2021). One year on: An updated systematic
review of SARS-CoV-2, COVID-19, and audio-vestibular symptoms. Inter-
national Journal of Audiology. doi: 10.1080/14992027.2021.1896793.
Anderson, K.G., Rambaut, A., Lipkin, W.I., Holmes, E.C., and Garry, R.F.
(2020). The proximal origins of SARS-CoV-2. Nature Medicine 26: 450–452.
Ankeny, R., Chang, H., Boumans, M., and Boon, M. (2011). Introduction: Phi-
losophy of science in practice. European Journal of the Philosophy of Science
1: 303–307.
Assaf, G., David, H., McCorkell, L., Wei, H., Brooke, O, Akrami, A., Low, R.,
Mercier, J., and other member of the COVID-19 Body Politic Slack Group.
(2020). Report: What does COVID-19 recovery actually look like? An anal-
ysis of the prolonged COVID-19 symptoms survey by patient-led research
team. Patient Led Research Collaborative. https://patientresearchcovid19.
com/research/report-1/.
Bickle, J. (2016). Revolutions in neuroscience: Tool development. Frontiers in
Systems Neuroscience, 10: 24 doi: 10.3389/fnsys.2016.00024.
Bickle, J. (2018). From microscopes to optogenetics: Ian Hacking vindicated.
Philosophy of Science 85: 1065–1077.
Bickle, J. (2022). Tinkering in the lab. This volume, pp. 13–36.
Brann, D.H, Tsukahara, T., Weinreb, C. et al. (2020). Non-neuronal expres-
sion of SARS-CoV-2 entry genes in the olfactory system suggests mechanisms
 A Different Role for Tinkering 131
underlying COVID-19-anosmia. Science Advances 6: eabc5801. doi: 10.1126/
sciadv.abc580.
Centers for Disease Control and Prevention (2019). 1918 Pandemic (H1N1 vi-
rus). https://www.cdc.gov/flu/pandemic-resources/1918-pandemic-h1n1.html
Chaplin, S. (2020). COVID-19: A brief history and treatments in development.
Prescriber May: 23–28.
Chauhan, S. (2020). Comprehensive review of coronavirus 2019 (COVID-19).
Science Direct Biomedical Journal 43: 334–340.
Churchland, P.S. (1986). Neurophilosophy. Cambridge, MA: The MIT Press.
Churchland, P.S., and Sejnowski, T.J. (1992). The Computational Brain. Cam-
bridge: MIT Press.
Churchland, P.S., and Sejnowski, T.J. (2016). Blending computational and ex-
perimental neuroscience. Nature Reviews Neuroscience 17: 567–568.
Couzin-Frankel, J. (2020). From “brain fog” to heart damage, COVID-19’s
lingering problems alarm scientists. Science. Published July 31, 2020. doi:
10.1126/science.abe1147.
Crick, F.H. (1979). Thinking about the brain. Scientific American 241: 219–232.
Douaud, G., Lee, S., Alfaro-Almagro, F., Arthofer, C., Wang, C. Lange, F.,
Andersson, J.L.R., Griffanti, L., Duff, E., Jbabdi, S., Taschler, B., Windkler,
A., Nichols, T.E., Collins, R., Matthews, P.M., Allen, N., Miller, K.L., and
Smith, S.M. (2021). Brain imaging before and after COVID-19 in UK bio-
bank. Preprint posted at https://www.medrxiv.org/content/10.1101/2021.06
.11.21258690v1.
Duarte-Neto, A.N., Monteiro, R.A.A., da Silva, L.F.F., Malheiros, D.M.A.C.,
de Oliveira, E.P., Theordoro-Filho, T., Pinho, J.R.R., Gomes-Gouvea, M.S.,
Salles, A.P.M., de Oliveira, I.R.S., Mauad, T., Saldia, P.H.N., and Dolhnikoff,
M. (2020). Pulmonary and systemic involvement in COVID-19 patients as-
sessed with ultrasound-guided minimally invasive autopsy. Histopathology
77: 186–197. doi: 10.1111/his.14160.
Durst, G. Carvalho, L.M., Rambaut, A., and Bedford, T. (2018). MERS-
CoV spillover at the camel-human interface. eLife 7: e31257. doi: 10.7554/
eLife.31257.
Fagan, J., and Jackler, R. (2017). Hammer & gouge cortical mastoidectomy for
acute mastoiditis. In J. Fagan (Ed.) The Open Access Atlas of Otolaryngol-
ogy, Head & Neck Operative Surgery, pp. 1–14. https://vula.uct.ac.za/access/
content/group/ba5fb1bd-be95-48e5-81be-586fbaeba29d/Hammer%20%20
Gouge%20Mastoidectomy%20for%20acute%20mastoiditis-1.pdf.
Frazier, K.M., Hooper, J.E., Mostafa, H.H., and Stewart, C.M. (2020). SARS-
CoV-2 isolated from the mastoid and middle ear: Implications for COVID-19
precautions during ear surgery. JAMA Otolaryngology-Head & Neck Sur-
gery 146: 964–969.
Garg, R.K., Mahadevan, A., Malhotra, H.S., Rizvi, I. Kumar, N., and Uniyal,
R. (2019). Subacute sclerosing panencephalitis. Reviews of Medical Virology
29: e2058. doi: 10.1002/rmv.2058.
Gold, I., and Roskies, A.L. (2008). Philosophy of Neuroscience. In M. Ruse
(Ed.) The Oxford Handbook of Philosophy of Biology. New York: Oxford
University Press, pp. 349–380.
Gouvea dos Santos, W. (2020). Natural history of COVID-19 and current
knowledge on treatment and therapeutics. Biomedicine and Pharmacother-
apy 129: 110493. doi: 10.1016/j.biopha.2020.110493.
132 Valerie Gray Hardcastle and C. Matthew Stewart
Hacking, I. (1983). Representing and Intervening. Cambridge: Cambridge Uni-
versity Press.
Hardcastle, V.G. (2017). Thinking about the brain: What is the meaning of neu-
roscience knowledge and technologies for capital mitigation? In E.C. Monahan
and J.J. Clark (Eds.) Mitigation in Capital Cases: Understanding and Commu-
nicating the Life-Story. Washington, DC: American Bar Association, Chapter 5.
Hardcastle, V.G., and Hardcastle, K. (2015). A new appreciation of Marr’s lev-
els: Understanding how brains break. Topics in Cognitive Science 7: 259–273.
Heneka, M.T., Golenback, D., Latz, E., Morgan, D., and Brown, R. (2020).
Immediate and long-term consequences of COVID-19 infections for the de-
velopment of neurological disease. Alzheimer’s Research & Therapy 12: 69
doi: 10.1186/s13195-020-00640-3.
Helms, J., Kremer, S., Merdji, H., Clere-Jehl, R., Schenck, M., Kummerlen, C.,
Collange, O., Boulay, C., Fafi-Kremer, S., Ohana, M., Anheim, M., and Me-
ziani, F. (2020). Neurologic features in severe SARS-CoV-2 infection. New
England Journal of Medicine 382: 2268–2270. doi: 10.1056/NEJMc2008597.
Hoffman, L.A., and Vilensky, J.A. (2017). Encephalitis lethargica: 100 Years
after the epidemic. Brain 140: 2246–2251.
Hubel, D.H. (1996). David H. Hubel. In L. Squire (Ed.) The History of Neuro-
science in Autobiography, Volume 1. Washington, DC: Society for Neurosci-
ence, pp. 294–317.
Jensen, M.P., Le Quesne, L., Officer-Jones, J., Teodiósio, A., Traventhiran, J.
Ficken, C., Goddard, M., Smith, C., Menon, D., and Allinson, K.S.J. (2020).
Neuropathological findings in two patients with fatal COVID-19. Neuropa-
thology and Applied Neurobiology 47: 17–25 doi: 10.1111/nan.12662.
Jiao, L., Yang, Y., Yu, W., Zhou, Y., Long, H. Gao, J., Ding, K. Ma, C., Zhao,
S., Wang, J., Li, H., Yang, M., Xu, J., Want, J., Yang, J., Kuang, ., Luo, F.,
Qian, X., Xu, L., Yin, B., Liu, W., Lu, S., and Peng, X. (2021). The olfactory
route is a potential way for SARS-CoV-2 to invade the central nervous system
of rhesus monkeys. Signal Transduction and Targeted Therapy 6: 169. doi: 10.
1038/s41392-021-00591-7.
Kahn, J.S., and MacIntosh, K. (2005). History and recent advances in coronavi-
rus discovery. The Pediatric Infectious Disease Journal 24: S223–S227.
Kelly, H. Peck. H.A., Laurie, K.L., Wu, P., Nishiura, H., and Cowling, B.J.
(2011). The age-specific cumulative incidence of infection with pandemic
influenza H1N1 2009 was similar invarious countries prior to vaccination.
PLoS One 6(8): e21828. doi: 10.1371/journal.pone.0021828.
Kuhn, T. (1962). The Structure of Scientific Revolutions. Chicago, IL: Univer-
sity of Chicago Press.
Kulcsar, M.A., Montenego, F.L., Arap, S.S., Tavares, M. R., and Kowalski, L.P.
(2020). High risk of COVID-19 for head and neck surgeons. International
Archives of Otorhinolaryngology 24: e129–e130.
Lago, M.N. (2020). How did we get here? Short history of COVID-19 and other
coronavirus epidemics. Head & Neck 42:1535–1538.
Lechien, J.R., Chiesa-Estomba, C.M., Beckers, E., Mustin, V., M. Ducarme, M.
Journe, J., Marchant, A., Jouffe, L., Barillari, M.R., Cammaroto, G., Circiu,
M.P., Hans, S., and Saussez, S. (2021). Prevalence and 6-month recovery of
olfactory dysfunction: A multicentre study of 1363 COVID-19 patients. Jour-
nal of Internal Medicine. doi: 10.1111/joim.13209.
 A Different Role for Tinkering 133
Lee, M.H., Perl, D.P., Nair, G. Li, W., Maric, D., Murray, H., Dodd, S.J.,
Koretsky, A.P., Watts, J.A., Cheung, V., Masliah, E., Horkayne-Szakaly, I.,
Jones, R., Stram, M.N., Moncur, J., Hefti, M., Folkerth, R.D., and Nath, A.
(2021). Microvascular injury in the brains of patients with Covid-19. New
England Journal of Medicine 384: 481–483.
Liu, Y.-C., Kuo, R.-L., and Shih, S.-R. (2020). COVID-19: The first documented
coronavirus in history. Science Direct Biomedical Journal 43: 328–333.
Meinhardt, J., Radke, J., Dittmayer, C. et al. (2021). Olfactory transmucosal
SARS-CoV-2 invasion as a port of central nervous system entry in individuals
with COVID-19. Nature Neuroscience 24: 168–175.
Mukerji, S.S., and Solomon, I.H. (2021). What can we learn from brain autop-
sies in COVID-19? Neuroscience Letters 742:135528.
Nauen, D.W., Hooper, J.E., Stewart, C.M., and Solomon, I.H. (2021). Assessing
brain capillaries in coronavirus 2019. JAMA Neurology 78: 760–762. doi:
10.1001/jamaneurol.2021.0225.
Portmann, G., and Colleagues (1939). Translated by P. Voilé. A Treatise on the
Surgical Technique of Otorhinolaryngology. Baltimore, MD: William Wood
& Company, Medical Division, The Williams & Wilkens Co.
Pratt, J., Lester, E., and Parker, R. (2021). Could SARS-CoV-2 cause tauopathy?
The Lancet 20: P506. doi: 10.1016/S1474-4422(21)00168-X.
Puelles, V.G., Lütgehetmann, M., Lindenmeyer, M.T. et al. (2020). Multiorgan
and renal tropism of SARS-CoV-2. New England Journal of Medicine 38:
590–592.
Rapkiewicz, A.T., Mait, X., Carsons, S.E., Pittaluga, S., Kleiner, D. Berger, J.,
Thomas, S., Adler, N.M., Charytan, D.M., Gasmi, B., Hochman, J.S., and
Reynolds, H.R. (2020). Megakaryocytes and platelet-fibrin thrombi charac-
terize multi-organ thrombosis at autopsy in COVID-19: A case series. EClin-
icalMedicine. 24: 100434. doi: 10.1016/j.eclinm.2020.100434.
Reichard, R.R., Kashani, K.B., Boire, N.A., Constantopoulos, E., Guo, Y.,
and Lucchinetti, C.F. (2020). Neuropathology of COVID-19: A spectrum of
vascular and acute disseminated encephalomyelitis (ADEM)-like pathology.
Acta Neuropathologica 140: 1–6.
Remmelink, M., De Mendonca, R., D’Haene, N. De Clercq, S., Verocq, C.,
Lebrun, L., Lavis, P., Racu, M-L., Trepant, A-L, Maris, C., Rovive, S., Gof-
fard, J-C., De Witte, O., Peluso, L., Vincent, J-L., Decaestecker, C., Taccone,
F.S., and Salmon, I. (2020). Unspecific post-mortem findings despite multior-
gan viral spread in COVID-19 patients. Critical Care 24: 495.
Solomon, I.H., Normandin, E., Bhattacharyya, S. Mukerji, S.S., Ali, A.S., Ad-
ams, G., Hormick, J.L. Padera, R.F., and Sabeti, P. (2020). Neuropatholog-
ical features of Covid-19. New England Journal of Medicine 383: 989–992.
Stix, G. (2012). A Q&A with Ian Hacking on Thomas Kuhn’s legacy as “the
paradigm shift” turns 50. Scientific American (April). https://www.scientifi-
camerican.com/article/kuhn/.
Thakur, K.T., Miller, E.H., Glendinning, M.D., Al-Dalahman, A., Banu, M.A.,
Boehm, A.K. et al. (2021). COVID-19 neuropathology at Columbia Uni-
versity Irving Medical Center/New York Presbyterian Hospital. Brain. doi:
10.1093/brain/awab148.
Uranaka, T., Kashio, A., Ueha, R., Sato, T., Bing, H.Ying, G., Kinoshita, M.,
Kondo, K., and Yamasoba, T. (2020). Expression of ACE2, TMPRSS2, and
134 Valerie Gray Hardcastle and C. Matthew Stewart
furin in mouse ear tissue and the implications for SARS-CoV-2 infection. The
Laryngoscope 131: E2013–E217. doi: 10.1002/lary.29324.
Van Waegeningh, H.F., Ebbens, F.A., van Spronsen, E., and Oostra, R.-J.
(2019). Single origin of the epithelium of the human middle ear. Mechanisms
of Development 158: 103556. doi: 10.1016/j.mod.2019.103556.
Wagner, T., Schweta, F.N.U., Murugadoss, K. et al. (2020). Augmented cura-
tion of clinical notes from a massive EHR system reveals symptoms of im-
pending COVID-19 diagnosis. eLife 9: e58227. doi: 10.7554/eLife.58227.
Wichmann, D, Sperhake, J.P., Lutgehetmann, M. et al. (2020). Autopsy findings
and venous thromboembolism in patients with COVID-19: A prospective co-
hort study. Annals of Internal Medicine 173: 268–277.
World Health Organization (2003). Emergency preparedness, response. Update
83 – One hundred days into the [SARS] outbreak. https://www.who.int/csr/
don/20030618/en.
World Health Organization (2020). Coronavirus disease 2019 (COVID-19).
Situation report – 16. https://www.who.int/emergencies/diseases/novel-
coronavirus-2019/situationreports.
World Health Organization (2021). Coronavirus. https://www.who.int/
health-topics/coronavirus#tab=tab1.
Yang, A.C., Kern, F., Losada, P.M. et al. (2021). Dysregulation of brain
and choroid plexus cell types in severe COVID-19. Nature. doi: 10.1038/
s41586-021-03710-0.
Zubair, A.S., McAlpine, L.S., Gardin, T., Farhadian, S., Kuruvilla, D.E., and
Spudich, S. (2020). Neuropathogenesis and neurologic manifestations of the
coronaviruses in the age of coronavirus disease 2019: A review. JAMA Neu-
rology 77: 1018–1027. doi: 10.1001/jamaneurol.2020.2065.
Section 2

Research Tools and

Epistemology
6 Dissemination and
Adaptiveness as Key
Variables in Tools That Fuel
Scientific Revolutions
Alcino J. Silva

1 Introduction
New tools have often been catalysts of scientific revolutions in neurosci-
ence, from the role of the Golgi stain in the discovery of neurons to the
recent transformation of systems neuroscience by both optogenetics and
ground-breaking neuronal imaging methods. Sophisticated tools open
doors to novel observations and ideas, but not all new tools are either
feasibly adaptable to multiple related uses or equally accessible to neu-
roscientists. In this chapter, I will discuss the idea that the power of a
new tool is not only dependent on its ability to reveal new potentially
transformative phenomena but also on its adaptiveness, ease of use and
transferability. The impact of prohibitively complex or inaccessible tools
is often limited by the relatively small number of laboratories that even-
tually use them, while the power of accessible and nimble tools is am-
plified by their wide dissemination. The ability to carry out replication
and convergent follow up experiments is at the very heart of scientific
revolutions in biology, and therefore most of the tools that fuel innova-
tive leaps in neuroscience, and in other biological disciplines, are almost
without exception very flexible and easily accessible to the majority of
practitioners in their respective fields. I will illustrate these ideas by us-
ing examples from the recent history of neuroscience, and I will finish
the chapter by predicting that a new set of tools that has just become
available to neuroscientists, and with the properties mentioned above,
will have a significant impact on the field, including in bridging the cur-
rent gap between molecular and systems studies in neuroscience.

2 The Molecular Genetics Revolution in Neuroscience:

Some Personal Reflections
I still remember as a new graduate student in 1983 the sense of won-
der of opening Maniatis, as we then called Tom Maniatis’ highly influ-
ential Molecular Cloning Laboratory Manual (Maniatis, et al. 1982).
That incredible blue book included simple protocols for most of the
techniques I would use in the next five years of my Ph.D. studies in

DOI: 10.4324/9781003251392-9
138 Alcino J. Silva
the early days of the molecular biology and human genetics revolution
(McConkey 1993). Naively, I assumed that the accessibility and adap-
tiveness inherent in most tools that fueled the molecular biology and
genetics revolutions were common to all disciplines in biology. I quickly
discovered that matters were very different in neuroscience, the field I
decided to join. In the late eighties, there were deep divisions amongst
neuroscience fields that could be easily traced back to the tools that in-
dividual labs had access to. Strange that it may seem to neuroscientists
now, at that time neuroscience areas were mostly defined by the tools
used. For example, memory was studied by behavioral neuroscientists,
electrophysiologists and molecular biologists, but there was hardly any
overlap between the approaches they used and consequently the results
they published. For example, behavioral assays and brain lesion tools
were used in behavioral neuroscience to define the brain structures in-
volved in specific forms of memory (e.g., spatial memory), while electro-
physiologists used electrodes to probe the synaptic mechanisms thought
to underlie memory (e.g., long term potentiation or LTP), and molecu-
lar biologists were busy cloning the molecules (e.g., receptors, kinases)
that they thought may mediate mechanisms of memory. The barriers
between fields were so deep that there was even a deeply entrenched re-
luctance to cross them. I will never forget being reminded mockingly by
a leading neuroscientist that my “amateurish efforts” to use molecular
genetics, electrophysiology and behavioral neuroscience approaches in
my new laboratory were naïve since I would be a “jack of all trades and
master of none” (Nelson 2015)!
But all of this would eventually change with the wide use of transgenic
tools in neuroscience research (Grant and Silva 1994; Silva and Giese
1994; Tonegawa, et al. 1995; Silva 1996; Silva, et al. 1997). These tools
would break barriers between neuroscience fields and open the doors to
the integrative studies that now dominate the field (Silva, et al. 1997).
I still remember during my post-doctoral studies my first day in Jeanne
Wehner’s laboratory, and the excitement of starting behavioral studies
of the alpha calmodulin kinase II knockout mice that I had generated in
Susumu Tonegawa’s laboratory. With her post-­doctoral fellow Richard
Paylor, we showed that these mutant mice had deficits in hippocampal
learning (Silva, et al. 1992)! Then, I took the same knockout mice to
Charles Stevens laboratory, where along with Yanyan Wang we discovered
deficits in LTP induction(Silva, et al. 1992), thus tentatively connecting
the loss of that synaptic kinase (Bennett, et al. 1983) with hippocam-
pal LTP and h ­ ippocampal-dependent learning (Silva, et al. 1997). In my
own laboratory in 1992, we managed to have behavioral neuroscien-
tists, electrophysiologists and molecular biologists working side by side,
and the same was starting to happen in other neuroscience laboratories
(Silva, et al. 1997). Mutant mice served as bridges between previously
 Dissemination and Adaptiveness 139
separate fields in neuroscience, thus fueling a revolution that trans-
formed the field (­Mayford, et al. 1995; Silva, et al. 1997).
Transgenic mice, flies and worms are very flexible and easily dissem-
inated tools, since mutants can be generated for almost every gene of
interest, and then shared amongst laboratories. In the years that followed
those early heady days, countless mutants were used as integrative tools
by hundreds of laboratories to test or unravel molecular and cellular
mechanisms of nearly every behavior that had ever been studied (Bickle
2016). Within a few years, viral vectors were introduced to mutate spe-
cific regions of the mammalian brain, so as to confer molecular, cellular
and brain region specificity to studies of behavior (Chen, et al. 2019;
Nectow and Nestler 2020). At the same time, mutant mice were also used
in systems neuroscience studies (Rotenberg, et al. 1996; Cho, et al. 1998),
and they were instrumental in establishing connections between this field
and others in neuroscience, including molecular, cellular and behavioral
neuroscience. It is now difficult to find neuroscience papers in high-profile
journals that do not share the integrative character of these early mutant
mouse studies. A flexible and easy to share tool (i.e., mutant mice) has es-
sentially completely transformed neuroscience in a way that was difficult
to predict at the time (Morris and Kennedy 1992). I will never forget the
attacks that the field suffered in the early days of this integrative effort.
Many of our colleagues thought implicitly or explicitly that nothing of
consequence could be gained from studies that dared to bridge the gaps
between fields as far afield as molecular biology and behavioral neurosci-
ence (Nelson 2015). Their compelling arguments invoked the necessary
superficiality of such studies, the “inevitable” lack of deep expertise in-
volved, and the “unavoidable errors” that would come from such am-
ateurism. Their arguments were not without merit, but everything had
been changed by the irresistible prospect of making causal connections
across neuroscience fields, from molecular mechanisms, to the cellular
properties they regulated, to systems processes mediated by these cellular
mechanisms, all the way to behavior. The ruthless reductionism (Bickle
2007) inherent in work with molecular genetic tools had opened new
broad horizons in neuroscience and there was no going back. The power
of these new tools was inextricably linked to their easy dissemination,
adaptability and wide range of experimental applications (Bickle 2016).

3 The Optogenetics Revolution in Systems Neuroscience

The optogenetics revolution in systems neuroscience (Yizhar,
et al. 2011; Goshen 2014) is another example of the power that easily
disseminated and highly adaptive tools can have in a field (Bickle 2016).
Before these tools completely transformed the field, systems neurosci-
ence studies were heavily dependent on electrophysiological recordings
140 Alcino J. Silva
and computational modeling (Yizhar, et al. 2011). For example, electro-
physiological approaches record the profiles of action potentials of neu-
rons under different conditions and computational modeling attempts
to capture and account for these profiles in ways that are generative
(i.e., lead to further experiments that test the validity of the models).
These models are not necessarily explanatory causal statements (Buzsáki
2019), but instead mathematical formulations or computational simu-
lations that attempt to capture the complexity of the brain systems
studied. The discovery of proteins that can be used to either activate or
inhibit neurons with light (i.e., opsins) changed all of this (Boyden 2011).
With these methods, ideas and hypotheses could now be tested with a
number of experiments that directly manipulated the activity state of
specific neurons in targeted brain regions of behaving animals (Yizhar,
et al. 2011)! The initial optogenetic tools were followed within an as-
toundingly short time by a large number of other opsins with special
properties that allowed for previously unimaginable experiments (Rost,
et al. 2017), such as the specific activation and inhibition of nerve termi-
nals between two connected brain regions! The tsunami of papers that
followed was only possible because these tools were easily disseminated
and extremely flexible: opsin genes could be easily shared amongst lab-
oratories, and widespread genetic methods resulted in the engineering
of novel opsins with a wide range of different uses. The proliferation
of experiments where specific neurons could be precisely activated or
inhibited with a range of optogenetic tools changed fundamentally the
epistemological nature of systems neuroscience since these powerful
tools allowed the direct testing of causal explanations between neuronal
phenomena and brain states and behavior (Bickle 2016). This epistemo-
logical approach had fueled the molecular biology revolution, and now
it had made its way into systems neuroscience!

4 Tools for Defining and Connecting Phenomena in

Neuroscience
With this historical context set, I will describe next in detail key aspects
of this epistemology and how it could guide tool development. Tools
can be used to define the properties of phenomena (e.g., the sequence
of a gene, the shape of a synapse, the activation state of an ensemble of
neurons, the behavioral state of an animal), and to test possible causal
connections between these phenomena (e.g., whether a specific gene reg-
ulates changes in the shape of synapses, the activation of memory en-
sembles and memory formation itself) (Silva, et al. 2014; Bickle 2016).
Understanding the rules of how phenomena are defined and connected
can be useful in determining the tools that could have the biggest im-
pact at any one time in a particular field. Convergent evidence is crit-
ical for both defining and establishing causal connections amongst
 Dissemination and Adaptiveness 141
phenomena, and therefore it is important to define rules for measuring
the degree of convergence in any one set of experiments (see below).
Consistency is also critical in defining and connecting phenomena, and
it is a key consideration in evaluating any tool: the lower the variability
between the outcomes of identical experiments, the higher the useful-
ness of a given tool (Silva, et al. 2014).
To better define the role of tools in testing convergence in experi-
mental findings, I will start by introducing the different categories of
experiments to test or establish causality (i.e., connection experiments)
commonly carried out in neuroscience (Silva, et al. 2014; M ­ atiasz, et al.
2017; Matiasz, et al. 2018). In Positive Intervention experiments (↑), the
quantity or probability of one phenomenon (i.e., an Agent) is actively
increased, and the change or lack of change in another (i.e., the Target) is
measured. In contrast, in a Negative Intervention experiment (↓), either
the quantity or probability of an Agent is actively decreased, and the im-
pact on a Target is measured. In a Positive Non-Intervention experiment
(Ø↑), an increase in either the quantity or probability of the Agent is
observed without intervention and the effects on a Target are measured,
while in a Negative Non-Intervention experiment (Ø↓), a decrease in
either the quantity or probability of the Agent is observed without inter-
vention and the effects on a Target are measured. Convergence (or lack
of) between the results of experiments in the categories just outlined is
commonly used to evaluate the strength of evidence for any potential
causal connection in biology, including neuroscience. Neuroscientists
and other biologists also commonly perform a fifth type of experiment,
a Multi-Intervention (∇), in which more than one Agent is either posi-
tively or negatively manipulated simultaneously to determine whether
one or more secondary Agents mediate the effect of some primary Agent
on a given Target. This last type of experiment is critical in testing causal
paths between multiple phenomena (Bickle and Kostko 2018). Research-
Maps (Matiasz, et al. 2017, 2018), an online computational tool (re-
searchmaps.org), uses these experimental categories to generate causal
maps of biological phenomena.
The strength of evidence for causal connections depends on the
availability of tools for Intervention and Non-Intervention exper-
iments, and therefore it is easy to see how the categories outlined
above are used explicitly or implicitly to guide tool development. For
example, late in my graduate studies I became fascinated with the
emerging literature of molecular changes associated with brain states
and behavior (­Kennedy 1983). These Non-Intervention studies were
thought-provoking, and it became apparent that there was a great need
for approaches (i.e., Intervention tools) that could manipulate these
molecular processes to test whether they had meaningful effects on a
number of other phenomena, including behavioral phenomena, such
as learning and memory. Pharmacological manipulations were limited
142 Alcino J. Silva
by the difficulty of developing drugs with the molecular specificity
and toxicology profile required for neuroscience experiments. I had the
good fortune of doing my Ph.D. studies next door to Mario Capecchi,
one of the inventors of a set of tools that allow for the manipulation of
any gene in mice (i.e., gene targeting) (Capecchi 1989), and I decided
to use these gene Intervention tools in my post-doctoral studies in neu-
roscience. As described above, my Negative Intervention experiments
with the newly developed gene targeting approach (Capecchi 1989)
were critical in implicating the alpha calmodulin kinase II in synap-
tic plasticity and memory (Silva, et al. 1992; Silva, et al. 1992), and
they complemented Non-Intervention experiments that suggested a
role for this general class of molecules in plasticity and memory (Silva,
et al. 1992). Mouse transgenic tools, such as gene targeting, continue
to have a key role in neuroscience. Because of their wide application,
these techniques can be used for Positive and Negative Intervention
experiments to test the role of any gene in any other phenomena of
interest.
Beyond connecting phenomena, tools are also critical for defining the
identity or properties of phenomena (i.e., identity experiments) (Silva,
et al. 2014; Bickle 2016). Just as with connection experiments, conver-
gence and consistency are also critical for identity experiments: When
the same method gives reproducible results, and multiple methods give
convergent results concerning the properties of a given phenomenon,
then we can be assured that the results of the identity experiments in
question are reliable. For example, in defining spatial learning, neu-
roscientists use multiple methods including the Morris water maze
(Morris 1981), where animals swim in a pool of water looking for a
hidden platform, and the Barne’s Maze (Barnes 1979), where rodents
escape a brightly lit open field by learning to navigate to a hole with
an escape box. These two behavioral tools are very different, but they
have one thing in common: they measure spatial learning and mem-
ory performance. Therefore, finding that a given manipulation (genetic,
pharmacological, systems, etc.) affects both tasks in the same way (e.g.,
performance is disrupted or enhanced) is compelling evidence that it
affects spatial learning and memory. Similarly, hippocampal CA1 place
cells can be defined with electrophysiological (O’Keefe and Nadel 1978)
and imaging methods (Dombeck, et al. 2010): finding that manipulat-
ing phenomenon A in the hippocampal CA1region disrupts the ability
of these cells to fire in a location-specific manner, measured with these
two dramatically different methods, argues that phenomenon A has a
role in place specific firing in that region. Thus, convergence and con-
sistency are critical for tools used in both connection and identity ex-
periments. Convergence and consistency are also critical in experiments
designed for tool development and characterization (Silva, et al. 2014;
Bickle 2016).
 Dissemination and Adaptiveness 143
5 Tools for the Next Revolution in Neuroscience
Epistemological ideas, including those concerning the role of tools in
discovery, can not only be used to account for previous revolutions
in science, but more importantly, these ideas can hopefully be useful
in predicting and guiding future scientific developments. Although this
is by no means universally agreed to (see for example Chicharro and
Ledberg 2012; Buzsáki 2019) many neuroscientists believe that the
key goal of studies in Neuroscience is to account for brain states and
behavior with causal explanations that include multiple levels of anal-
yses (e.g., molecular, cellular, systems, brain regions, cognitive) (Silva,
et al. 2014; Bickle 2016; Sudhof 2017). However, currently, there is a
deep chiasm in neuroscience between molecular explanations of brain
states and behavior (i.e., molecular neuroscience) and those explana-
tions centered on brain systems (i.e., systems neuroscience). One of
the reasons for this chiasm is the lack of tools to test molecular ex-
planations of systems phenomena associated with brain states and be-
havior. Tools that address this problem could have a profound impact
in neuroscience research (Callaway 2005; Sudhof 2017). Ideally, these
tools would be able to track (i.e., for non-intervention experiments)
and manipulate (i.e., for intervention experiments) molecular function
with the appropriate temporal resolution, cellular and brain region
specificity required for the study of circuits, and their impact on brain
states and behavior. Remarkably, although these very tools (see be-
low) have been developed for other purposes, such as cancer research,
they have been rarely used in neuroscience. Below, I will review only
a couple of examples from each category of novel tools that could be
used to bridge the gap between molecular neuroscience and system
neuroscience studies. This is not meant as a comprehensive review of
this fast-growing field. Instead, the examples below were chosen to
illustrate the types of tools that I believe could have a transformative
impact in systems neuroscience. Again, the power of a new tool is
critically dependent on its adaptiveness, ease of use and transferability.
The classes of tools I will discuss next have these properties and this
is one of the reasons I believe they will become central to neuroscience
research.

5.1 Imaging Systems

Miniaturized head-mounted fluorescent microscopes, (miniscopes), ca-
pable of recording from hundreds of cells (including neurons) for days or
weeks in freely moving and behaving mice and other animals (Ghosh, et al.
2011; Ziv, et al. 2013; Cai, et al. 2016; Liberti, et al. 2017; Jacob, et al. 2018;
Aharoni, et al. 2019), are a powerful and easily disseminated tool. Single
unit recording systems, commonly used in systems neuroscience studies,
144 Alcino J. Silva
require significant expertise, considerable time investment, and can only
stably record neuronal activity for hours. In contrast, miniscopes require
no previous experience with in vivo recordings, laboratories can even
manufacture these miniscopes themselves at very low costs and stable
neuronal recordings can be routinely made for days or even weeks.
Although currently miniscopes have limited sensitivity and reso-
lution, upcoming technological improvements promise to overcome
these limitations. An important facet of miniscopes is the ongoing
­community-based improvement in their design, which continues to add
new features, such as for example, the ability to activate optogenetic
tools in the very fields of view (with closed-loop approaches [Potter,
et al. 2014]) that are being imaged (Stamatakis, et al. 2018). This opened
up the exciting possibility of simultaneously tracking and manipulating
molecular function in specific cells and circuits of behaving animals,
such as mice.
The inexpensive nature and considerable power of miniscope systems,
and consequently their expected widespread use in neuroscience labora-
tories world-wide, promises to be a significant catalyst for innovation in
systems neuroscience. Until recently, miniscopes have been almost solely
used for imaging changes in calcium concentration in neurons in vivo.
However, new molecular sensor developments (see below) promise to
dramatically expand the range of biological applications for this class of
imaging tools.
Recent advances in imaging go well beyond miniscopes, and include
other major advances which employ much larger, more technologically
complex and far more expensive microscopes used with head-fixed an-
imals, such as the two-photon mesoscope (Sofroniew, et al. 2016) and
holographic two-photon imaging (Yang, et al. 2018). However, these
systems’ cost and complexity preclude their easy dissemination, and
therefore most likely will limit the impact that they could have on neu-
roscience research.

5.2 Molecular Sensors and Reporters

In parallel with novel imaging devices, there has also been considerable
progress in molecular sensors and reporters that could be used to dra-
matically expand the scope of these imaging systems (Chen, et al. 2013;
Wang, et al. 2018). Fluorescence, autofluorescence and bioluminescence
have been used to engineer a plethora of different molecular reporters
and sensors that could potentially be coupled with imaging technolo-
gies (see above) to track the activation of specific molecular mechanisms
in a cell and circuit-specific manner in behaving mice, rats and other
animals. For example, Synapto-pHluorin is a novel pH imaging sen-
sor that includes the transmembrane synaptic vesicle protein VAMP2,
and a specific pHluorin called ecliptic pHluorin (Granseth, et al. 2006;
 Dissemination and Adaptiveness 145
Rose, et al. 2013). Within the low pH milieu inside of transmitter ves-
icles, synapto-­pHluorin is non-­fluorescent. When vesicles are released,
however, and synapto-pHluorin is exposed to the neutral extracellular
space, it is brightly fluorescent, thus allowing for visualization of synap-
tic release events in a circuit-­specific manner. Since they are essentially
molecular biology tools that can be delivered to specific cell populations
in the brain with widely used viral vectors, these molecular sensors will
be easily disseminated and easily adapted for the study of any behavior
or brain state.

5.3 Light Activated Manipulation of Molecular Mechanisms

Optogenetic tools have also been recently developed to manipulate most
intracellular signaling events, from ligand binding, receptor activation,
associated intracellular signaling, all the way to activation/suppression
of gene expression, and protein translation (Rost, et al. 2017; Kwon
and Heo 2020). For example, optical modulation approaches combined
with sophisticated genetic methods have been used to modulate recep-
tor function in specific cells of targeted systems and in defined brain
regions (e.g., Kim, et al. 2005; Airan, et al. 2009). Neurotransmit-
ter receptors, such as those for noradrenaline, serotonin, glutamate,
GABA, and dopamine, have had a key role in systems neuroscience,
and these novel optical manipulation tools that have been used to con-
trol the function of receptors in real time, provide an unprecedented
opportunity to modulate associated molecular signaling mechanisms
in a cell, circuit and brain region-specific manner, thus allowing for
experiments that could connect molecular properties associated with
these receptors with cellular and circuit processes modulating brain
states and behavior.
In addition to receptor function, there are also optogenetic approaches
to control protein interaction, phosphorylation, cleavage, aggregation/
disaggregation, allosteric changes, etc. (Kwon and Heo 2020). For ex-
ample, to control protein-protein interactions optogenetically, several
Light-Inducible Dimerization (LIDs) systems have been developed that
employ light-sensitive domains, including Phy-PIF, DrBphP, LOV2, and
CRY2-Cib (Zhang, et al. 2016; Klewer and Wu 2019; McCormick, et al.
2020). These domains change conformation after light activation, and
this has been used to regulate specific protein-protein interactions and
consequently their function (Kakumoto and Nakata 2013; Zhang, et al.
2014). In addition to controlling dimerization/aggregation, optogenetic
approaches have also been used to control enzyme activity by taking
advantage of light-triggered changes in protein conformation (Nguyen,
et al. 2016).
Other than controlling intracellular signaling, opto-tools have also
been engineered to modulate transcription. For example, site-specific
146 Alcino J. Silva
recombinase systems, such as Cre/Loxp, Flp/FRT, have been widely used
in viral vectors and in transgenic mice for modulating genes in a cell-
type specific manner (Wang, et al. 2011). To improve the temporal regu-
lation of these approaches, multiple variants were engineered that allow
for tight regulation by light (Schindler, et al. 2015; Kawano, et al. 2016;
Jung, et al. 2019; Yao, et al. 2020). Importantly, there are approaches
that depend on infrared deep brain stimulation and ultrasensitive opsins
that do not require cannulae implantation and associated damage of
brain tissue (Chen, et al. 2018; Jung, et al. 2019; Gong, et al. 2020).
Again, these genetic tools can be delivered to any cell population in the
brain with widely used viral vectors, and they can be adapted to test
their impact on any behavior or brain state.

6 Molecules, Systems and Behavior

The incredibly powerful imaging and manipulation tools mentioned
above (and many others that we did not have space to review) will
revolutionize systems neuroscience. It is important to acknowledge the
seminal contributions that systems neuroscience has made despite the
fact that until recently (see above) the majority of studies in the field
were exclusively focused on a small subset of cellular properties (elec-
trical properties such as action potentials) of a single brain cell type
(neurons), and that they were carried out mostly with a single class
(electrophysiology) of experimental approaches. This success is truly
remarkable, since we now have overwhelming evidence that the com-
plexity of brain states and behavior can be ascribed to nearly every cell
type in the brain (e.g., glia [Stevens 2003]), and can be traced back to
a wide range of biological mechanisms that go well beyond the electri-
cal properties of neurons. Despite the continuing success of traditional
neuroscience approaches, it is clear now that imaging and manipu-
lating the very molecular properties that both shaped brain systems
(not just neuronal circuits) during evolution, and that now account for
and constrain much of their functionality, will open countless win-
dows into the properties and mechanisms of the systems (neuronal,
glial, etc.) that underlie the wonderful complexity of brain states and
behavior.

7 Conclusion
In this chapter, I discussed how novel tools can be the catalysts of inno-
vation, and reviewed key properties, including ease of dissemination and
adaptability, that allow tools to open doors to ground-breaking obser-
vations and ideas. I argued that the potential usefulness of a new tool is
not only dependent on its ability to reveal new potentially transformative
phenomena, but also on its ease of use, adaptability and transferability.
 Dissemination and Adaptiveness 147
The ability to readily carry out replication and convergent follow up ex-
periments in many laboratories is at the very heart of scientific revolu-
tions in biology, and therefore most of the tools that fuel innovative leaps
in neuroscience, and in other biological disciplines, are frequently very
flexible and easily accessible to practitioners in the field. I propose that
molecular optogenetic tools, capable of optically tracking and manip-
ulating a wide range of molecular phenomena in diverse cell types and
systems in the brain of behaving animals, will be key catalysts for an
upcoming wave of innovation in neuroscience. These nimble and easily
disseminated tools will not only catalyze a much-needed synthesis be-
tween molecular and systems neuroscience, they will also move systems
neuroscience away from its neurocentric roots to a more biological realis-
tic inclusion of other cell types including for example astrocytes, glia and
oligodendrocytes.

References
Aharoni, D., B. S. Khakh, A. J. Silva and P. Golshani (2019). “All the light
that we can see: a new era in miniaturized microscopy.” Nat Methods 16(1):
11–13.
Airan, R. D., K. R. Thompson, L. E. Fenno, H. Bernstein and K. Deisseroth
(2009). “Temporally precise in vivo control of intracellular signalling.” Na-
ture 458(7241): 1025–1029.
Barnes, C. A. (1979). “Memory deficits associated with senescence: a neuro-
physiological and behavioral study in the rat.” J Comp Physiol Psychol 93(1):
74–104.
Bennett, M. K., N. E. Erondu and M. B. Kennedy (1983). “Purification and
characterization of a calmodulin-dependent protein kinase that is highly con-
centrated in brain.” J Biol Chem 258(20): 12735–12744.
Bickle, J. (2007). “Ruthless reductionism and social cognition.” J Physiol Paris
101(4–6): 230–235.
Bickle, J. (2016). “Revolutions in neuroscience: tool development.” Front Syst
Neurosci 10: 24.
Bickle, J. and A. Kostko (2018). “Connection experiments in neurobiology.”
Synthese 195: 5271–5295.
Boyden, E. S. (2011). “A history of optogenetics: the development of tools for
controlling brain circuits with light.” F1000 Biol Rep 3: 11.
Buzsáki, G. (2019). The Brain from Inside Out. New York, Oxford University
Press.
Cai, D. J., D. Aharoni, T. Shuman, J. Shobe, J. Biane, W. Song, B. Wei, M.
Veshkini, M. La-Vu, J. Lou, S. E. Flores, I. Kim, Y. Sano, M. Zhou, K. Baum-
gaertel, A. Lavi, M. Kamata, M. Tuszynski, M. Mayford, P. Golshani and A.
J. Silva (2016). “A shared neural ensemble links distinct contextual memories
encoded close in time.” Nature 534(7605): 115–118.
Callaway, E. M. (2005). “A molecular and genetic arsenal for systems neurosci-
ence.” Trends Neurosci 28(4): 196–201.
Capecchi, M. R. (1989). “The new mouse genetics: altering the genome by gene
targeting.” Trends Genet. 5(3): 70–76.
148 Alcino J. Silva
Chen, S., A. Z. Weitemier, X. Zeng, L. He, X. Wang, Y. Tao, A. J. Y. Huang,
Y. Hashimotodani, M. Kano, H. Iwasaki, L. K. Parajuli, S. Okabe, D. B. L.
Teh, A. H. All, I. Tsutsui-Kimura, K. F. Tanaka, X. Liu and T. J. McHugh
(2018). “Near-infrared deep brain stimulation via upconversion nanoparticle-­
mediated optogenetics.” Science 359(6376): 679–684.
Chen, S. H., J. Haam, M. Walker, E. Scappini, J. Naughton and N. P. Martin
(2019). “Recombinant Viral Vectors as Neuroscience Tools.” Curr Protoc
Neurosci 87(1): e67.
Chen, T. W., T. J. Wardill, Y. Sun, S. R. Pulver, S. L. Renninger, A. Baohan, E.
R. Schreiter, R. A. Kerr, M. B. Orger, V. Jayaraman, L. L. Looger, K. Svoboda
and D. S. Kim (2013). “Ultrasensitive fluorescent proteins for imaging neuro-
nal activity.” Nature 499(7458): 295–300.
Chicharro, D. and A. Ledberg (2012). “When two become one: the limits of
causality analysis of brain dynamics.” PLoS One 7(3): e32466.
Cho, Y. H., K. P. Giese, H. Tanila, A. J. Silva and H. Eichenbaum (1998). “Ab-
normal hippocampal spatial representations in alphaCaMKIIT286A and
CREBalphaDelta-mice.” Science 279(5352): 867–869.
Dombeck, D. A., C. D. Harvey, L. Tian, L. L. Looger and D. W. Tank (2010).
“Functional imaging of hippocampal place cells at cellular resolution during
virtual navigation.” Nat Neurosci 13(11): 1433–1440.
Ghosh, K. K., L. D. Burns, E. D. Cocker, A. Nimmerjahn, Y. Ziv, A. E. Gamal
and M. J. Schnitzer (2011). “Miniaturized integration of a fluorescence mi-
croscope.” Nat Methods 8(10): 871–878.
Gong, X., D. Mendoza-Halliday, J. T. Ting, T. Kaiser, X. Sun, A. M. Bastos, R. D.
Wimmer, B. Guo, Q. Chen, Y. Zhou, M. Pruner, C. W. Wu, D. Park, K. Deisse-
roth, B. Barak, E. S. Boyden, E. K. Miller, M. M. Halassa, Z. Fu, G. Bi, R. Des-
imone and G. Feng (2020). “An ultra-sensitive step-function opsin for minimally
invasive optogenetic stimulation in mice and macaques.” Neuron 107: 38–51.
Goshen, I. (2014). “The optogenetic revolution in memory research.” Trends
Neurosci 37(9): 511–522.
Granseth, B., B. Odermatt, S. J. Royle and L. Lagnado (2006). “Clathrin-­
mediated endocytosis is the dominant mechanism of vesicle retrieval at hip-
pocampal synapses.” Neuron 51(6): 773–786.
Grant, S. G. and A. J. Silva (1994). “Targeting learning.” Trends in Neurosci-
ences 17(2): 71–75.
Jacob, A. D., A. I. Ramsaran, A. J. Mocle, L. M. Tran, C. Yan, P. W. Frankland
and S. A. Josselyn (2018). “A compact head-mounted endoscope for in vivo
calcium imaging in freely behaving mice.” Curr Protoc Neurosci 84(1): e51.
Jung, H., S. W. Kim, M. Kim, J. Hong, D. Yu, J. H. Kim, Y. Lee, S. Kim, D.
Woo, H. S. Shin, B. O. Park and W. D. Heo (2019). “Noninvasive optical
activation of Flp recombinase for genetic manipulation in deep mouse brain
regions.” Nat Commun 10(1): 314.
Kakumoto, T. and T. Nakata (2013). “Optogenetic control of PIP3: PIP3 is suf-
ficient to induce the actin-based active part of growth cones and is regulated
via endocytosis.” PLoS One 8(8): e70861.
Kawano, F., R. Okazaki, M. Yazawa and M. Sato (2016). “A photoactivatable
Cre-loxP recombination system for optogenetic genome engineering.” Nat
Chem Biol 12(12): 1059–1064.
 Dissemination and Adaptiveness 149
Kennedy, M. B. (1983). “Experimental approaches to understanding the role of
protein phosphorylation in the regulation of neuronal function.” Annu Rev
Neurosci 6(493): 493–525.
Kim, J. M., J. Hwa, P. Garriga, P. J. Reeves, U. L. RajBhandary and H. G.
Khorana (2005). “Light-driven activation of beta 2-adrenergic receptor sig-
naling by a chimeric rhodopsin containing the beta 2-adrenergic receptor cy-
toplasmic loops.” Biochemistry 44(7): 2284–2292.
Klewer, L. and Y. W. Wu (2019). “Light-induced dimerization approaches to
control cellular processes.” Chemistry 25(54): 12452–12463.
Kwon, E. and W. D. Heo (2020). “Optogenetic tools for dissecting complex
intracellular signaling pathways.” Biochem Biophys Res Commun 527(2):
331–336.
Liberti, W. A., L. N. Perkins, D. P. Leman and T. J. Gardner (2017). “An open
source, wireless capable miniature microscope system.” J Neural Eng 14(4):
045001.
Maniatis, T., E. F. Fritsch and J. Sambrook (1982). Molecular Cloning: A Labo-
ratory Manual. Cold Spring Harbor, NY, Cold Spring Harbor Press.
Matiasz, N. J., J. Wood, P. Doshi, W. Speier, B. Beckemeyer, W. Wang, W. Hsu
and A. J. Silva (2018). “ResearchMaps.org for integrating and planning re-
search.” PLoS One 13(5): e0195271.
Matiasz, N. J., J. Wood, W. Wang, A. J. Silva and W. Hsu (2017). “­Computer-aided
experiment planning toward causal discovery in neuroscience.” Front Neu-
roinform 11: 12.
Mayford, M., T. Abel and E. R. Kandel (1995). “Transgenic approaches to cog-
nition.” Curr Opin Neurobiol 5(2): 141–148.
McConkey, E. H. (1993). Human Genetics: The Molecular Revolution. Boston,
MA, Jones and Bartlett Publishers.
McCormick, J. W., D. Pincus, O. Resnekov and K. A. Reynolds (2020). “Strat-
egies for engineering and rewiring kinase regulation.” Trends Biochem Sci
45(3): 259–271.
Morris, R. G. M. (1981). “Spatial localization does not require the presence of
local cues.” Learning Motivation 12: 239–260.
Morris, R. G. M. and M. B. Kennedy (1992). “The Pierian spring.” Current
Biology 2(10): 511–514.
Nectow, A. R. and E. J. Nestler (2020). “Viral tools for neuroscience.” Nat Rev
Neurosci 21(12): 669–681.
Nelson, N. C. (2015). “A knockout experiment: disciplinary divides and experi-
mental skill in animal behaviour genetics.” Med Hist 59(3): 465–485.
Nguyen, M. K., C. Y. Kim, J. M. Kim, B. O. Park, S. Lee, H. Park and W. D.
Heo (2016). “Optogenetic oligomerization of Rab GTPases regulates intracel-
lular membrane trafficking.” Nat Chem Biol 12(6): 431–436.
O’Keefe, J. and L. Nadel (1978). The Hippocampus as a Cognitive Map. Lon-
don, Oxford University Press.
Potter, S. M., A. El Hady and E. E. Fetz (2014). “Closed-loop neuroscience and
neuroengineering.” Front Neural Circuits 8: 115.
Rose, T., P. Schoenenberger, K. Jezek and T. G. Oertner (2013). “Developmental
refinement of vesicle cycling at Schaffer collateral synapses.” Neuron 77(6):
1109–1121.
150 Alcino J. Silva
Rost, B. R., F. Schneider-Warme, D. Schmitz and P. Hegemann (2017). “Op-
togenetic tools for subcellular applications in neuroscience.” Neuron 96(3):
572–603.
Rotenberg, A., M. Mayford, R. D. Hawkins, E. R. Kandel and R. U. Muller
(1996). “Mice expressing activated CaMKII lack low frequency LTP and do
not form stable place cells in the CA1 region of the hippocampus.” Cell 87(7):
1351–1361.
Schindler, S. E., J. G. McCall, P. Yan, K. L. Hyrc, M. Li, C. L. Tucker, J. M. Lee,
M. R. Bruchas and M. I. Diamond (2015). “Photo-activatable Cre recombi-
nase regulates gene expression in vivo.” Sci Rep 5: 13627.
Silva, A. J. (1996). Genetics and learning: misconceptions and criticisms. In
Gene Targeting and New Developments in Neurobiology, ed. S. Nakanishi,
A. J. Silva, S. Aizawa and M. Katsuki. Tokyo, Japan Scientific Societies Press,
3–15.
Silva, A. J. and K. P. Giese (1994). “Plastic genes are in!” Curr Opin Neurobiol
4(3): 413–420.
Silva, A. J., A. Landreth and J. Bickle (2014). Engineering the Next Revolution
in Neuroscience: The New Science of Experiment Planning. New York, Ox-
ford Press.
Silva, A. J., R. Paylor, J. M. Wehner and S. Tonegawa (1992). “Impaired spa-
tial learning in alpha-calcium-calmodulin kinase II mutant mice.” Science
257(5067): 206–211.
Silva, A. J., A. M. Smith and K. P. Giese (1997). Gene targeting and the biology
of learning and memory. Annu Rev Genet 31: 527–546.
Silva, A. J., C. F. Stevens, S. Tonegawa and Y. Wang (1992). “Deficient hippo-
campal long-term potentiation in alpha-calcium-calmodulin kinase II mutant
mice.” Science 257(5067): 201–206.
Silva, A. J., Y. Wang, R. Paylor, J. M. Wehner, C. F. Stevens and S. Tonegawa
(1992). “Alpha calcium/calmodulin kinase II mutant mice: deficient long-
term potentiation and impaired spatial learning.” Cold Spring Harb Symp
Quant Biol 57: 527–539.
Sofroniew, N. J., D. Flickinger, J. King and K. Svoboda (2016). “A large field of
view two-photon mesoscope with subcellular resolution for in vivo imaging.”
Elife 5: e14472.
Stamatakis, A. M., M. J. Schachter, S. Gulati, K. T. Zitelli, S. Malanowski, A.
Tajik, C. Fritz, M. Trulson and S. L. Otte (2018). “Simultaneous optogenetics
and cellular resolution calcium imaging during active behavior using a minia-
turized microscope.” Front Neurosci 12: 496.
Stevens, B. (2003). “Glia: much more than the neuron’s side-kick.” Curr Biol
13(12): R469–R472.
Sudhof, T. C. (2017). “Molecular neuroscience in the 21(st) century: a personal
perspective.” Neuron 96(3): 536–541.
Tonegawa, S., Y. Li, R. S. Erzurumlu, S. Jhaveri, C. Chen, Y. Goda, R. Pay-
lor, A. J. Silva, J. J. Kim, J. M. Wehner and C. F. Stevens (1995). “The gene
knockout technology for the analysis of learning and memory, and neural
development.” Prog Brain Res 105: 3–14.
Wang, H., M. Jing and Y. Li (2018). “Lighting up the brain: genetically encoded
fluorescent sensors for imaging neurotransmitters and neuromodulators.”
Curr Opin Neurobiol 50: 171–178.
 Dissemination and Adaptiveness 151
Wang, Y., Y. Y. Yau, D. Perkins-Balding and J. G. Thomson (2011). “Recom-
binase technology: applications and possibilities.” Plant Cell Rep 30(3):
267–285.
Yang, W., L. Carrillo-Reid, Y. Bando, D. S. Peterka and R. Yuste (2018). “Si-
multaneous two-photon imaging and two-photon optogenetics of cortical cir-
cuits in three dimensions.” Elife 7: e32671.
Yao, S., P. Yuan, B. Ouellette, T. Zhou, M. Mortrud, P. Balaram, S. Chatter-
jee, Y. Wang, T. L. Daigle, B. Tasic, X. Kuang, H. Gong, Q. Luo, S. Zeng,
A. Curtright, A. Dhaka, A. Kahan, V. Gradinaru, R. Chrapkiewicz, M.
Schnitzer, H. Zeng and A. Cetin (2020). “RecV recombinase system for in
vivo targeted optogenomic modifications of single cells or cell populations.”
Nat Methods 17(4): 422–429.
Yizhar, O., L. E. Fenno, T. J. Davidson, M. Mogri and K. Deisseroth (2011).
“Optogenetics in neural systems.” Neuron 71(1): 9–34.
Zhang, K., L. Duan, Q. Ong, Z. Lin, P. M. Varman, K. Sung and B. Cui (2014).
“Light-mediated kinetic control reveals the temporal effect of the Raf/MEK/
ERK pathway in PC12 cell neurite outgrowth.” PLoS One 9(3): e92917.
Zhang, Y., S. A. Sloan, L. E. Clarke, C. Caneda, C. A. Plaza, P. D. Blumen-
thal, H. Vogel, G. K. Steinberg, M. S. Edwards, G. Li, J. A. Duncan, 3rd, S.
H. Cheshier, L. M. Shuer, E. F. Chang, G. A. Grant, M. G. Gephart and B.
A. Barres (2016). “Purification and characterization of progenitor and ma-
ture human astrocytes reveals transcriptional and functional differences with
mouse.” Neuron 89(1): 37–53.
Ziv, Y., L. D. Burns, E. D. Cocker, E. O. Hamel, K. K. Ghosh, L. J. Kitch, A. El
Gamal and M. J. Schnitzer (2013). “Long-term dynamics of CA1 hippocam-
pal place codes.” Nat Neurosci 16(3): 264–266.
7 Toward an Epistemology of
Intervention
Optogenetics and Maker’s
Knowledge1, 2
Carl F. Craver

1 Introduction
The biological sciences, like other mechanistic sciences, comprise both a
modeler’s and a maker’s tradition. The aim of the modeler, in my narrow
sense, is to describe correctly the causal structures—the mechanisms—
that produce, underlie, maintain, or modulate a given phenomenon or
effect.3 These models are expected to save the phenomena tolerably well
(that is, to make accurate predictions about them) and, in many cases
at least, to represent the components and causal relationships compos-
ing their mechanisms. The aim of the maker, in contrast, is to build
machines that produce, underlie, maintain, or modulate the effects we
desire.4
The works of both maker and modeler depend fundamentally on the
ability to intervene into a system and make it work differently than it
would work on its own. My goal is to identify some dimensions of prog-
ress (or difference) in the ability to intervene in biological systems.
This project complements Alan Franklin’s (1986, 1990, 2012) pio-
neering work on the epistemology of experiment. Franklin focuses on
detection instruments and, specifically, on distinguishing “between a
valid observation or measurement and an artifact created by the ex-
perimental apparatus” (see 1986, 165, 192; 1990, 104). He argues that
scientists defend new detection techniques by showing that they detect
known magnitudes reliably, that their results conform to the expecta-
tions of a well-confirmed theory, and that their findings agree tolerably
well with those of other, more or less causally independent detection
techniques.5 Scientists sometimes defend their instruments directly by
appeal to their well-supported design principles and by showing their
results cannot be explained by known sources of error.6 Yet by focus-
ing on detection specifically, Franklin neglects a crucial aspect of causal
experiments: interventions. My goal is to take some preliminary steps
toward an epistemology of intervention by characterizing the norms by
which improvements in intervention methods are measured.7
In pursuit of this goal, I examine aspects of the early history of optoge-
netics. This technique matured and made its way into the neurosciences

DOI: 10.4324/9781003251392-10
 Toward an Epistemology of Intervention 153
around the turn of the Twenty-first century and subsequently has been
adopted widely and rapidly as an improved intervention method. Nu-
merous biologists have won awards for inventing and describing the key
components of this triumph of biological maker’s knowledge (including
Ernst Bamberg, Ed Boyden, Karl Diesseroth, Peter Hegemann, Georg
Nagel, and Gero Miesenböck, according to the latest Wiki update), and
the technique is currently featured in over 800 research articles per year
(see Kolar et al. 2018). By exploring why this technique was adopted so
readily and widely, we get a glimpse of the epistemic norms that make
one intervention technique better than another and into the arguments
by which such claims are defended.8
Optogenetics allows researchers to control the electrophysiological
properties of neurons with light.9 Researchers insert bacterial genes for
light-sensitive ion channels into target cells in a given brain region. The
virus that inserts this construct into cells commandeers the cell’s protein
synthesis and delivery mechanisms to assemble the light-sensitive chan-
nels and install them in the membrane. The light delivered through a
fiber-optic cable then activates or inactivates the channels, changing the
ionic current across the membrane and thereby modulating, producing,
or blocking neural signals.10
In what follows, I first present a schema for thinking about causal
experiments and complicate it to accommodate the complexity of op-
togenetics. I then discuss eleven dimensions of progress or difference
in the ability to intervene into brain function. These give us a sense of
the epistemic norms guiding the assessment of progress in intervention.
I close reflecting on how and why makers and modelers differ in their
assessment of the norms of intervention.

2 Causal Experiments
Figure 1 represents a simplified, standard causal experiment. A given
causal hypothesis or mechanism schema is instantiated in a target sys-
tem. The target system is the subject, organism, or system in which one
performs the experiment.11 One intervenes in the system to change one or
more target variables (T) and detects the resulting value of putative effect
variables (E).
The intervention technique is a means of changing one or more target
variables in the mechanism. In the simple case, one sets T to a value.
Sometimes, as in Kettlewell’s famed experiments on moth populations,
the researcher identifies natural circumstances that set the variable to a
value, but here I focus on interventions under researchers’ control.
Intervention checks detect the value of the target variable to see if the
intervention succeeded in setting T to the desired value. In lesion experi-
ments, for example, one confirms the location of the lesion using CT scans
154 Carl F. Craver

Intervention Check Detection

Target System

Intervention Eliciting Conditions

Figure 1 A standard form of causal experiment.

or post-mortem inspections. In pharmacological experiments, one might

measure the plasma concentration of the drug. Intervention checks can
also be used to compensate for unreliable intervention techniques; one
grounds inferences on the measured, rather than the intended, value of T.
On the other side of this schema, the detection technique is a device or
process that takes some feature or magnitude (here, E) in the experimen-
tal system as an input and returns a reading or measure as an output.
Detection techniques indicate a variable’s value: Litmus paper indicates
pH; an osmometer detects ionic concentrations; and functional MRI
(arguably) detects dendritic field potentials. Franklin focuses primarily
on the reliability of detection techniques (1990, 2009).
Eliciting conditions are interventions required to prepare the putative
effect variable for detection. In microscopic studies, experimenters stain
tissues. In classic PET studies, they inject radiolabeled substances. Elic-
iting conditions are causal interactions with the target system, but they
interact with the putative effect rather than the putative cause.12

3 Dimensions of Progress in Intervention

According to Woodward’s (2003) account of causal relevance, the claim
that T causes E amounts roughly to the claim that there exists an ideal
intervention on T that changes the value of E. Importantly, an inter-
vention need not be ideal in this sense to probe causal relationships.
Nonetheless, the ideal case is a good starting point for thinking about
progress in intervention.
Figure 2 summarizes Woodward’s criteria for ideal interventions.
Unidirectional arrows represent causal relations, the bidirectional dot-
ted arrow represents correlation, and bars across arrows indicate that
the relation must be absent. The depicted intervention is designed to
test whether T makes a difference to the effect variable, E. An ideal
 Toward an Epistemology of Intervention 155

U D
1

T S E
3 2

4
I C
5

Figure 2 Constraints on an ideal intervention (cf. Woodward 2003).

intervention, I, fixes the value of T, rendering it independent of all other

causal influences, U (1). This intervention changes E but only via the
change in T. That is, the effect is neither mediated directly (2) nor in-
duced by changing a variable, S, intermediate between T and E (3). Nor
does E change simply because I is correlated with some other variable,
C, that is causally relevant to E (4). Nor does I influence the detection
apparatus (D) used to measure E (5).13 Such conditions ensure, in short,
that any observed change in E can be attributed to the change I induces
in T and not to some other change wrought by or correlated with I.
The use of placebos in drug trials and sham surgeries in lesion studies
exemplify these ideals. Ideally, one contrives control conditions that
mimic the intervention in all respects except for the change to T. Such
controls allow one to assess whether the changes observed in E are due
to the changes in T or to some other unanticipated effect of I upon (or
some correlation of I with) some other variable causally relevant to E.14
Consider now how the schemas in Figures 1 and 2 might have to be
complicated to describe more adequately the complexity of experiments
as they occur in the “wild.” First, I is typically an extended causal se-
quence (as in Figure 3a). For example, I might involve an experimenter’s
choice (A), an action or procedure based on that choice (B), a device or
machine (C), which can also be decomposed into separate components,
that makes an alteration to the system (D) in which the target variable
(T) is located. Even Figure 3a is misleadingly simplified in representing
only a single intervention sequence. As optogenetics illustrates, the inter-
vention might involve a confluence of many tributary interventions (see
Figure 3b). Different sequences are involved, for example, in building the
genetic construct for the channels (determining what is delivered), in tar-
geting the intervention to specific cell populations (determining where
it is delivered), and in turning the channels on with light (determining
when it is activated). Some of these tributary interventions are what Fred
Dretske (1988) calls structuring causes; they build the mechanisms that
156 Carl F. Craver

I T

A B C D

A T

A2
where
A1 B C D E
what
F when
A3

Figure 3 (a) Interventions are themselves complex mechanisms. (b) Tribu-

tary interventions affecting what, where, and when optognetic influence is
delivered.

allow the experimenter to control neurons with light. Others, such as the
light, are triggering causes.
Many experiments involve more than one complex structural cause. The
target system, for example, may be surgically prepared or pre-treated to
make the experiment possible or convenient. The animal might have been
trained on a task or acclimated to an experimental setting. The target sys-
tem might be set into a baseline state against which the experimental effect
is evaluated. In an experiment discussed below, researchers first intervene
to stimulate the basolateral amygdala (BlA) with direct current; they do so
in order to test the effects of a second intervention, optogenetic inhibition
of BlA terminals in the central amygdala (CeA), on indicators of anxiety.
This brief example illustrates that to represent experimental interven-
tions as a single arrow into a single target T is often an extreme idealiza-
tion. Different ways of organizing multiple interventions give different
experiments their power to answer different kinds of questions about
the causal structure of the target system. Precisely because such causal
organization of experimental systems is key to their epistemic power,
an adequate philosophy of experiment must move beyond this idealized
simplification (See Craver and Darden 2013).
An adequate philosophy of experiment also must avoid assuming that
idealized constraints represented in Figure 2 must be satisfied for an
experiment to be epistemically valuable. Progress in intervention some-
times involves discovering previously unrecognized violations of such
principles and correcting them, but not always. In some experimental
setups, it is impossible or undesirable to change T in a way that makes
 Toward an Epistemology of Intervention 157
its value independent of T’s other causes (as required in 1). One might
wish, for example, to increase the amount of dopamine in a system while
allowing endogenous dopamine to vary under the influence of its typi-
cal, physiological causes. Interventions are also frequently ham-fisted,
changing many potentially relevant variables at once, if only because
technology does not permit a more surgical intervention. I discuss below
some situations in which ham-fisted interventions might be preferred.
And for the purposes of maker’s knowledge, a useful intervention might
be non-ideal and nonetheless uniquely useful.
These preliminaries in place, let us consider some dimensions15 along
which one evaluates the quality of an intervention method. Table 1 lists
11 epistemically relevant dimensions along which intervention tech-
niques might vary from one another. I discuss each in turn, using early
work on optogenetics as illustrative examples.

3.1 Number of Variables

Perhaps the most obvious dimension of progress in intervention con-
cerns the number and diversity of variables a researcher can control.
Many researchers believe the theta rhythm in the hippocampus is caus-
ally relevant to hippocampal function. Yet this claim will remain some-
what speculative in the absence of a means of intervening to change
the theta rhythm in a controlled experiment. The same can be said for
distributed cortical representations generally. For many areas of the cor-
tex, researchers presently have no idea how to intervene on the cortex to
produce physiologically relevant patterns of activation across widely dis-
tributed brain networks (perhaps involving millions of neurons). The de-
velopment of such a technique would be real progress and would help to

Table 7.1 Eleven dimensions of virtue in intervention

• Which variables?
• Number of (relevant) variables controlled
• Selectivity
• Physiological relevance
• Within variables
• Range
• Grain
• Nature of change
• Valence
• Reversibility
• Physiological relevance
• Level of control
• Efficacy
• Dominance
• Determinism
158 Carl F. Craver
put talk of “codes” and “representations” on firmer epistemic footing.16
So, beginning with the obvious: we make progress in intervention as we
expand our ability to manipulate more and more of the variables that
potentially make a difference to how things we care about work.

3.2 Selectivity among Variables

A second dimension, selectivity among variables, is more relevant to the
explosion of optogenetics in neuroscience. Deisseroth explains:

What excites neuroscientists about optogenetics is control over de-

fined events within defined cell types at defined times—a level of
precision that is most likely crucial to biological understanding be-
yond neuroscience.
(Deisseroth 2010)

Optogenetics achieves this selectivity using genetic mechanisms that

determine which cells express the channels. Boolean combinations of
promoters and inhibitors in gene regulatory constructs allow research-
ers to target the channels at specific cells, with specific chemical signa-
tures, specific morphologies, or specific topological connectivity. Only
infected cells with the desired signatures initiate expression and become
light-responsive.
This is a clear advantage over both electrophysiological and many
pharmacological techniques for intervening in electrophysiological be-
haviors of cells. Using a traditional electrode, the injected current spreads
through brain tissue, exciting all the cells in a region (e.g., excitatory and
inhibitory, neurons and glia) with no regard for different cell (or neu-
ronal) types. Given that different cell types plausibly act differently in
neural mechanisms, an intervention technique that fails to distinguish
them is impotent to evaluate a broad range of causal hypotheses.
Optogenetics is also more selective than population-level pharmaco-
logical interventions. Pharmacological interventions can be quite a bit
more selective than electrophysiological stimulation because agonists
and antagonists can be used to interfere with or stimulate specific ion
channels and intracellular molecular cascades. However, pharmaco-
logical techniques are often non-specific in the sense that a given drug
might work on many systems at once (e.g., on all cells with dopamine
receptors).
The more selective an intervention technique is among potentially
causally relevant variables, the more the technique allows one to test
independently whether those variables make a difference. For example,
the nucleus accumbens, a brain structure associated with cocaine re-
ward, contains two types of dopaminergic projection neurons: one pop-
ulated with D1 receptors, and the other populated with D2 receptors.
 Toward an Epistemology of Intervention 159

V1 V1 ?

I V2 ? E I V2 ? E

V3 V3 ?

Figure 4 Selectivity among variables.

Optogenetic stimulation of the D1 subtype enhances the reward value of

cocaine; stimulating the D2 subtype diminishes its reward value (Lobo
et al. 2010). Pharmacological intervention with dopamine alone cannot
reveal this mechanistically relevant difference.
More abstractly (as represented in Figure 4) suppose a modeler is test-
ing the hypothesis that variable T makes a causal difference to E. The
experimental goal is to intervene on T and to detect the consequences of
this intervention (if any) on the value of E. Now imagine a set of alter-
native hypotheses, each of which asserts a different putative cause vari-
able for E, V1, V2 , and V3, as in Figure 4. Clearly, an intervention that
changes V1, V2 , and V3 at the same time cannot alone discriminate their
causal contribution to E. Progress along this dimension would allow one
to change the putative cause without changing other putative causes at
the same time. In short, the more selective one’s intervention, the more
one can eliminate a certain class of artifacts and confounds (Craver and
Dan-Cohen 2021). Greater selectivity among variables is also valuable
for makers because more selective interventions can eliminate undesired
side effects of ham-fisted interventions.17

3.3 Grain and Range among Values of a Variable

Given selective control over a single variable, an intervention method
that affords more precise control over the value of that variable might be
preferred to one that affords less precise control. And a technique that
affords such precise control over a larger range of potentially relevant
values of that variable will likewise be preferred to a technique with a
restricted range.
First-generation optogenetic techniques gave researchers control
over the firing rates of action potentials with mean interspike intervals
around 100–200 ms. By 2012, scientists had modified the channel pro-
tein to produce action potentials reliably with a mean interspike inter-
val around 5 ms (see Gunaydin et al. 2010). As Boyden, et al., describe
it: “This technology thus brings optical control to the temporal regime
occupied by the fundamental building blocks of neural computation”
(Boyden et al., 2005; italics in original). The technique was intentionally
160 Carl F. Craver
modified so it could be used to explore the rates and patterns of electro-
physiological activity across the entire range of activity neurons exhibit
in their natural environments.
More abstractly, suppose two techniques target the same variable. A
technique is an improvement over its predecessor if it allows one to ex-
plore more fully the variable’s space of plausible switch-points (or tran-
sition zones). A switch-point in the value of a variable is a difference in
the value of the variable that makes a difference to the effect variable
(Craver 2007). Zero degrees Celsius is a switch-point (or, more accu-
rately, a transition zone, i.e., an imprecise switchpoint) in temperature
that makes a difference to whether water is liquid or solid. Switchpoints
might be analog, where each increment in the cause variable (no matter
how small) makes some difference to the effect variable (e.g., mass and
gravitational attraction), or they might be digital, with a minimally ef-
fective difference in the cause variable.
For this reason, it is valuable to be able to set the putative cause vari-
ables to the widest range of relevant values. In testing theories of heat,
for example, scientists sought ways to produce (and reliably detect) ex-
tremely high and low temperatures (see Chang 2004). Likewise, if one
is interested in exploring the effects of mean firing rate on the behavior
of a population of neurons, one would want control over at least the
observed spectrum of that firing rate, or perhaps even simply a plausible
range of known firing rates. The greater the range of the intervention
technique, the more modelers can use it to explore the space of possible
switch-points and transition zones. For similar reasons, the maker can
use such interventions to find the buttons and levers that might be ex-
ploited in the pursuit of maker’s knowledge.
An additional dimension of progress concerns the grain or precision
of the intervention: the smallest change the technique can reliably induce
in the putative cause variable. The ideal grain of an intervention varies
with the system in question and with the pragmatic uses to which the
intervention is to be put. Pharmacological interventions, for example,
give one the ability to influence the chemical environments of neurons
in ways that can directly affect neuronal activity. However, their effect
is imprecise compared to that induced by optogenetics. Pharmacological
antagonists increase or decrease the probability of neuronal activity, but
they do not precisely regulate mean firing rates, let alone precise tem-
poral patterns. Systems-level pharmacology thus affords comparatively
coarse control over electrophysiological activities. If one thinks precise
rates or temporal codes are relevant to neural mechanisms, as seems
likely, one would then have reason to prefer exploring those possibilities
with optogenetic interventions over pharmacological interventions.
These ideas are represented abstractly in Figure 5. At the top left is
an intervention into T that covers its entire range of values at a very fine
grain (i.e., very high precision). In the limit, I can be used to set T to any
 Toward an Epistemology of Intervention 161

t1 t2 t3 t4...
t10 t11 t12 t13...
I T
t100 t101 t103 t104
t500 t501 t502 tk
Fine Grain, t1 t2 t3 t4...
Large Range
t10 t11 t12 t13...
I
t100 t101 t103 t104...

T Fine Grain, t500 t501 t502... tk

Small Range

t1 t2 t3 t4...
t10 t11 t12 t13...
I
t100 t101 t103 t104...
t500 t501 t502... tk
Course Grain,
Large Range

Figure 5 Interventions with different grains and ranges.

value it can take. To the right is an intervention technique that allows the
researcher to set the putative cause variable to only part of the range of
possible values. The grain is the same, but its range is diminished relative
to the case just mentioned. On the bottom is a technique that covers the
same range as the first but at a coarser grain. The technique can set the
variable only to a value somewhere within a block of values.

3.4 Physiological/Ecological Relevance

Modelers and makers differ with respect to whether interventions must
be physiologically and ecologically relevant or, perhaps more accurately,
how physiologically and ecologically relevant they must be. The model-
er’s goal is to understand how a biological system works in its standard
operating conditions. She necessarily enforces a distinction between
how a system in fact works and how it might be made to work under ex-
treme or otherwise unusual conditions. Modelers therefore tend to prize
intervention techniques that can be used to manipulate physiologically
relevant variables in ways and within ranges that are similar to the ways
and ranges at which those variables operate in normal circumstances.
Modelers also often prize interventions that are ecologically relevant,
i.e., appropriate to the normal environment in which the organism lives
or in which the system functions. Of course, modelers sometimes are in-
terested in system behavior outside ecological and physiological ranges,
162 Carl F. Craver
but their interest extends to these conditions precisely because such be-
haviors reveal clues about underlying mechanisms in the normal or typ-
ical case.
To put this point abstractly, from the modeler’s perspective, one kind
of progress in intervention involves improving the ability to mimic the
types of change one normally sees in the target system. Some aspects of
this assessment include: (i) whether the intervention targets the variables
operative in the system as it typically works, (ii) whether the values to
which the target variable can be set are within the range those variables
take in the system’s normal or typical behavior, (iii) whether the stim-
ulus values change at rates or in manners comparable to the ways they
change during the system’s normal or typical functioning, (iv) whether
the changes in the stimulus values are induced via a mechanism similar
to the mechanism by which the changes are produced in the system’s
normal or typical behavior.
Many early papers on optogenetics emphasize the method’s physiologi-
cal relevance. They note that the technique is targeted at electrophysiolog-
ical properties known to be relevant to the function of the nervous system
(i). As discussed in Section 3.3, the technique targets the temporal range
of observed electrophysiological function (ii and iii). Given that optogenet-
ically induced spike trains can be timed specifically, and given that these
spike trains have a high degree of replicability trial to trial (see below),
researchers can produce electrical signals in neurons that resemble closely
the very signals neurons produce when they work. Although optogenetics
does not produce action potentials by precisely the same mechanisms by
which cells typically produce action potentials, the mechanism is more
like the physiological case, involving the flux of ions through a membrane-­
spanning channel (iv), than is direct electrophysiological stimulation.18
The terms “normal,” “physiological,” and “ecological” imply a teleo-
logical orientation, a focus on how things work when they are working
properly or as they typically do in the healthy life of the organism in the
environment it is supposed to inhabit.19 Makers care less than modelers
about how something typically or normally behaves; they focus instead
on how something might be made to serve our aims. Makers might care
about promoting health, life, and the good of the species, but they rec-
ognize other ends as well, even those that run counter to these teleolog-
ical goals. They might actively change the environment. They recognize
a set of buttons and levers inside and outside organisms available to
be pushed and pulled for purposes that might be irrelevant in (or even
contrary to) evolutionary or physiological contexts (see Craver 2010,
2014). 20 Though the maker must take practical constraints into consid-
eration, she is not generally constrained to intervene within the limits of
normal or typical functioning in this teleological sense. For the maker,
teleological biology offers only a blinkered view of the space of possible
causal structures that can be used for engineering.
 Toward an Epistemology of Intervention 163
3.5 Reversibility
A further dimension for evaluating intervention techniques is the capac-
ity to reverse the intervention. A clear example involves lesion techniques
in experimental neuropsychology, where one uses a scalpel, a vacuum,
or an electrical current to remove brain tissue. Because brain tissue can-
not be restored, the experimenter has to compare lesioned animals to
non-lesioned animals, or, alternatively, pre-lesioned animals to post-­
lesioned animals. In response to this epistemic limit, researchers have in-
vented methods, such as cooling and transcranial magnetic stimulation,
to produce reversible, functional lesions. These techniques allow one to
change the value of the target variable from “on” to “off” and back
again, allowing one to run experimental and control conditions in the
same animals and in whatever order or sequence one desires.
Pharmacological interventions are typically reversible, but only on
long timescales. In standard optogenetics applications, light opens or
closes the channel. In more advanced applications involving step func-
tion opsins, however, one wavelength of light activates the channel, al-
lowing ions to diffuse across the membrane, and another wavelength
turns it off. The reversal is, for practical purposes, instantaneous.
The value of reversibility is illustrated dramatically in a set of exper-
iments on the role of the BlA and the CeA in anxiety (Tye et al. 2011).
Reversible intervention allows one to test the effects of activating and
inactivating the BlA on anxiety (operationalized as the tendency to avoid
open spaces) in the same mice over time. One can make a mouse even
more anxious (i.e., less likely to enter open spaces) by stimulating its BlA.
The BlA sends axonal projections to the CeA (among other structures).
By inhibiting those projections to the CeA, one can nullify the effect of
the stimulation. Tye et al. report the effect of electrophysiological stim-
ulation in the BlA with or without optogenetic inhibition of terminals in
the CeA. The latter optogenetic stimulation can be turned on and off at
will, convincingly demonstrating a gating effect of the CeA on the BlA’s
contribution to anxious behavior.
Reversible intervention in such physiological systems allows one to
avoid confounds that result from comparing two groups of organisms
that might differ in innumerable and unmeasured ways. It also allows
one to test whether the observed effects of the intervention result from
order effects.

3.6 Bivalence
A related, but distinct, dimension of progress in intervention is to move
from interventions that can manipulate the value of a variable only in
one direction to interventions that can manipulate the value of the vari-
able in both directions. Lesion studies, for example, remove but cannot
164 Carl F. Craver

V1

C+

I CB E

C-

Figure 6 B
 ivalent interventions can increase and decrease the value of a variable
in the same target system.

stimulate brain tissue. Optogenetic methods, for reasons just described,

allow one to both stimulate and inhibit neuronal activity, raising it or
lowering it from baseline. Bivalence is represented in Figure 6.
Bivalence is a virtue because it facilitates the search for switch-points
and transition zones. A bivalent intervention can raise or lower the value
of the putative cause variable. Bivalent interventions can be used to ex-
plore whether the cause variable and the effect variable vary concomi-
tantly across the spectrum of their values and perhaps to describe that
concomitant variation mathematically. Bivalent interventions also allow
one to explore higher-order differences produced by different kinds of
change: to assess the effect of increases and decreases, different rates of
increase or decrease, and different patterns of increasing and decreasing.
Bivalent intervention techniques thus allow one to explore a wider range
of values, and a wider range of changes to the value of the relevant target
variables, than do univalent intervention techniques. If one combines
two univalent techniques, one for stimulation and one for inhibition,
one inevitably introduces experimental differences between conditions
that are avoided if bivalence is packaged in a single technique.

3.7 Efficacy
As noted in Section 2, interventions are typically complex causal se-
quences, starting with an action and ending with the change in the target
variable. Just as an experimenter’s efforts to detect a given variable can
be complicated by, for example, failures anywhere in the detection appa-
ratus or in the process by which the results are stored and manipulated,
the effort to intervene into a target system might be complicated by fail-
ures anywhere in the complex causal chain in the intervention method.
An experimenter might fail to take the appropriate action. Or she
could take the appropriate action, but her device might glitch and so fail
to initiate the causal sequence. Or the target system might somehow foil
the would-be intervention. In a drug trial, for example, the subject might
forget to take the pill, or the patient might have a stomach enzyme that
 Toward an Epistemology of Intervention 165
degrades the drug before it enters the blood. Bracketing experimenter
error, the efficacy of an intervention technique is the reliability with
which the intervention technique produces the desired change to the tar-
get variable.
Efficacy is in part a matter of objective frequency: given that an ex-
perimenter decides to set T to a target value (or target range), and given
that she has activated her instruments to initiate the chosen intervention,
what is the probability that T actually takes the target value (or falls
within the target range)? Or for non-binary values, what is the variance
in the value of T given the intervention to set T to a particular value?
In the case of optogenetics, two measures of efficacy have been partic-
ularly important. One is the extent to which the technique succeeds in
causing the expression of rhodopsin channels in the membranes of all and
only the target cells. Tsai et al (2009) describe one such result as follows:

Greater than 90% of the TH immunopositive cells were positive

for ChR2-EYFP near virus injection sites and more than 50% were
positive overall, demonstrating a highly efficacious transduction of
the TH cells.
(2009)

TH, tyrosine hydroxylase, is a marker for dopaminergic cells. ChR2-

EYFP is a marker for the expression of the Channel-Rhodopsin 2 gene.
Tsai et al. thus use intervention checks to establish the efficacy of their
intervention. This aspect of efficacy for optogenetics varies considerably
across different brain regions and in different experimental organisms.
Another measure of efficacy for optogenetics, focused on a distinct
tributary intervention, is the frequency with which pulses of light pro-
duce action potentials in the target cells. The first paper to describe op-
togenetic interventions (Boyden et al. 2005) is largely dedicated to this
measure. They show that light induces stable ionic currents with very
rapid activation kinetics. They explore different stimulus durations to
fine-tune the technique to deliver trains of action potentials reliably.
They characterize how many cells in a population produce action poten-
tials. They show that the same (or almost the same) patterns of action
potentials can be repeated time and again in the same neuron and across
different neurons by delivering the same light stimulus.
These distinct arguments about the efficacy of optogenetics address
different parts of the complex intervention mechanism. The first (ex-
pression rates) concerns the efficacy of the relationship between delivery
of the virus and channel expression in the target cell population. The
second concerns the efficacy of the relationship between light and elec-
trophysiological activity in the neurons. More generally, efficacy can be
understood in terms of the variance of the target variable given that one
initiates an intervention to fix its value. Given that one has decided to set
166 Carl F. Craver
a magnitude T to a particular value t, efficacy is the probability that T
takes a value within a range around t. One could determine the efficacy
of a technique by summing the squared distances between the intended
and the actual value on a number of trials and dividing by the number of
trials. Efficacy is improved by shrinking variance: bringing more of the
actual values closer to the intended value.
Efficacy is clearly desirable if one hopes efficiently to replicate one’s
experimental interventions. However, an intervention need not be per-
fectly efficacious to be useful, either for the modeler testing a causal
hypothesis or for the maker hoping to effect some practical outcome.
Whether an intervention is efficacious enough for a given epistemic or
practical objective depends on the variable in question and on the rele-
vant switch-points for the effect variable. For the modeler, even a very
inefficacious intervention can be useful so long as appropriate interven-
tion checks determine whether or not the intervention has produced the
desired outcome in T. For the maker, however, failures are failures, even
if one can tell that the failure can be blamed on the equipment. This is
another significant difference between modeling and making.

3.8 Dominance
The dominance of an intervention is the extent to which it satisfies con-
dition (4) in Woodward’s account of ideal interventions (see U in Figure
2). In an ideal intervention, the intervention technique sets T to a value
independent of T’s other causes. Such an intervention screens the value
of T off from all T’s other causes.
Woodward includes this requirement because T’s other causes might
change T’s value in unintended ways, foiling the causal inference. Sup-
pose, for example, one intervenes to reduce the value of T and fails to
notice a change in E. Perhaps this is because other causes of T compen-
sate for the intervention or coincidentally raise the value of T, erasing the
effect of the intervention. In that case, one should not conclude that the
induced change is causally irrelevant to the effect.21
But interventions need not screen T off from its other causes to be
useful, as Eberhardt (2009) demonstrates in his consideration of hard
and soft interventions. Hard interventions remove all causal arrows into
T besides the intervention. T’s value is influenced, if at all, only by the
intervention. Soft interventions do not fix T’s value; rather, they change
the conditional probability distribution over T’s values. In other words,
hard interventions are arrow-breaking, eliminating the influence of T’s
other causes; soft interventions are arrow-preserving, leaving at least
some of the parental influence on T intact.
Optogenetic interventions have been used in both hard and soft in-
terventions. A single neuron in a dish, for example, can be removed
from its cellular context leaving the light stimulus as the only exogenous
 Toward an Epistemology of Intervention 167
determinant of cellular activity. 22 That would be a hard intervention. In
contrast, researchers using step function opsins simply raise the mem-
brane potential to make it more probable that an incoming signal will in
fact lead to the generation of an action potential in the post-synaptic cell.
This is a soft intervention.
Eberhardt and Scheines (2007) discuss some of the advantages of hard
interventions for learning about a system’s causal structure. First, if one
uses a hard intervention, one can easily infer that the observed correla-
tion between T and the putative effect variable is not due to the action
of a common cause. By definition, a hard intervention severs the effect
of any such common cause on T. Second, hard interventions allow one
to assess the direction of causal influence between two correlated vari-
ables, i.e., to distinguish cases in which A causes B from cases in which
B causes A. Finally, because hard interventions allow the experimenter
to set the value of the cause variable, it gives the researcher informa-
tion that might be used, for example, to characterize the strength of the
causal relationship.
In other cases, hard interventions are either impossible or undesirable.
They are practically impossible when we don’t know how to break all
the causal arrows into T. They are undesirable in cases in which one
wants to leave the causal structure of the system intact, for experimen-
tal or practical reasons. The value of non-dominating interventions is
perhaps most apparent in the domain of maker’s knowledge. Many stan-
dard medical interventions are non-dominating. Insulin injections for
diabetes, for example, augment rather than replace endogenous insulin
production. L-Dopa treatments for Parkinson’s disease augment rather
than replace endogenous dopamine levels. In such cases, one intervenes
to boost residual capacities of the system and/or to encourage it to com-
pensate for its weaknesses, not to replace it with an artificial control
system (as a heart and lung machine replaces hearts and lungs while they
are off-line).
Modelers can also usefully deploy soft interventions. Eberhardt and
Scheines (2007) demonstrate, for example, that there are conditions in
which multiple simultaneous soft interventions on a system suffice to
determine the system’s causal structure in a single experiment (unlike
hard interventions). They show further that if only a single interven-
tion is allowed per experiment, then the choice between hard and soft
interventions makes no difference to the rate at which the data from
the series of experiments will converge on the correct causal structure
(Eberhardt and Scheines 2007). These results, however, are bracketed
by the assumption that one knows all the variables sufficient to deter-
mine the value of the effect variable. If there are possibly unmeasured
common-cause confounds, as is likely the case in most biological exper-
iments, then hard, dominating interventions have the clear advantage in
trying to learn the system’s causal structure.
168 Carl F. Craver
The move from soft interventions to hard interventions (or vice versa)
is not necessarily a form of progress in the ability to intervene in a system.
Hard and soft interventions should rather be seen as distinct interven-
tion strategies that can be used to solve different kinds of experimental
and practical problems. But in many discovery contexts, there is a clear
reason to prefer dominating interventions.

3.9 Determinism
The last dimension I consider is determinism. Some interventions are de-
terministic. Others are probabilistic. In a deterministic intervention, the
intervention technique is used to set the value of the target variable to one
and only one value. In indeterministic interventions, one allows a range
of possible interventions around a mean, for example. In different con-
texts, deterministic or indeterministic interventions might be warranted.
In several of the early optogenetic experiments, researchers intervened
by stimulating cells with pulses of light spaced around a mean interspike
interval (e.g., 100 ms). They allowed the precise timing of the spikes to
vary in a normal distribution around that mean.
Which sort of intervention is most appropriate depends on the hy-
pothesis under test. In these experiments, researchers wanted merely to
demonstrate that the method could generate action potentials at rates in
the approximately physiological range and, crucially, that they could do so
without presuming any particular pattern of stimulation. One might, how-
ever, hypothesize that mean interspike interval is the difference-­making
variable for the effect of interest (i.e., the system functions with a rate
code). In that case, the experimenter might choose an indeterministic in-
tervention to rule out the confounding possibility that observed effects are
due to a particular spike sequence (e.g., regularly spaced spikes at 20 Hz).
Interventions, it should be acknowledged, are typically de facto inde-
terministic because complex intervention mechanisms might always fail.
Most real-world interventions are probabilistic to some extent, if only
due to differences in efficacy and dominance. People make mistakes.
No intervention instrument is perfect. Experimental subjects might have
individual differences in how they respond to the intervention. Any for-
mal treatment of the epistemology of interventions must accommodate
the fact that interventions are often chancy affairs, and that chance can
enter into a complex intervention mechanism at many different stages.

3.11 Summary
Table 1 lists the epistemically relevant ways one intervention might differ
from another discussed above. In some cases, movement from one end
of that dimension to another is meaningfully interpreted as progress. In
other cases, any preference along that dimension depends fundamentally
 Toward an Epistemology of Intervention 169
on the empirical or practical problem the researcher faces. Such
­context-relativity, however, is underwritten by a deeper sense that there
are distinct kinds of causal question and that distinct forms of interven-
tion can be used most appropriately or efficiently to address them.

4 Conclusion: Makers and Modelers Revisited

Progress in the ability to intervene just is progress in the ability to con-
trol and produce phenomena with whatever reliability and precision we
desire. In medical contexts, for example, the dream is to deploy maker’s
knowledge to cure and prevent diseases, to build prosthetic devices (such
as artificial hearts and brain-machine interfaces), and to encourage re-
covery and reorganization in such systems after damage or disease. This
growth in human know-how is one of the primary fruits that Bacon
envisioned for science (1620; see Sargent 2001, 2012).
Maker’s knowledge is a potent stimulus to the search for model-
er’s knowledge (see Carrier 2004). In the laboratory, especially in sci-
ences such as neuroscience, the growth of knowledge about causal
systems depends fundamentally on inventing new ways to intervene
in those systems and control their behavior. Detection techniques,
such as functional imaging and single-unit recording, are indispens-
able in the search for causes. Yet observation and detection alone
are demonstrably incapable of disambiguating even relatively simple
causal structures (Eberhardt 2009). Advances in intervention, such
as optogenetics, make irreplaceable contributions to our ability to
search efficiently through the space of possible mechanisms for a given
phenomenon.
More than that, progress in intervention often brings with it new
conceptual tools, allowing researchers to envision previously occluded
parts of the space of possible mechanisms. Just as the telescope allowed
Galileo and Kepler to see new things, developments in intervention tech-
niques have the capacity to open new worlds and envision previously un-
dreamt of causal variables. Optogenetics will likely transform in subtle
and not-so-subtle ways how researchers think about the transmission of
information through neural circuits.
Although makers and modelers often overlap, as evidenced by the
invention of optogenetics itself, the fact that they have different goals
places different demands on their methods. The modeler, in my re-
stricted sense, starts with the task of understanding how a given biolog-
ical mechanism normally or typically works. They chose a phenomenon
of teleological interest and try to understand how the different parts of
a mechanism are organized together so the phenomenon occurs. The
model of the mechanism includes only the variables that make a dif-
ference to that phenomenon in the normal or typical conditions. The
modeler is paradigmatically the reverse engineer.
170 Carl F. Craver
Makers, in contrast, are engineers simpliciter. They begin with an engi-
neering problem: something they would like to do or make. That problem
might involve getting a biological system to do something utterly unnat-
ural or unusual as viewed from an evolutionary or physiological perspec-
tive (such as driving a robotic arm or playing PONG with EEG waves).
Makers are free to use any available components whatsoever, not merely
those evolution has managed to cobble together. The optogenetics mecha-
nism is a nice example: it involves infecting cells with artificially contrived
viruses bearing DNA that would not, under any typical ecological and
evolutionary conditions, ever make its way into the brains of mice and
rats. The contrivance is, in this sense, altogether unnatural and atypical.
Makers and modelers also have different epistemologies. The solution
to an engineering problem wears success on its sleeve: There is noth-
ing more to “getting it right,” for a maker, than producing the right
input-output relationship within design constraints. The engineered
solution need not be pretty (though elegance is an engineering value);
the solution need not mirror a natural system; it simply needs to work.
Modelers cannot rest comfortably with simply producing a phenome-
non. Their goal is to understand how some naturally occurring phenom-
enon is produced. They therefore have the additional epistemic burden
of describing a real phenomenon, building a model that “saves” that
phenomenon, and making sure the model describes the mechanism that
typically or normally produces that phenomenon in physiologically and
ecologically relevant conditions. Modelers, in other words, have com-
mitments to mirroring tolerably well the causal structure of the system.
Makers must respect the causal structure of the world, of course, but
they are not constrained to emulate it.
The science of genetics has always included both maker’s and modeler’s
traditions (see Waters 2008, 2014; Allen forthcoming). Today, the science
of genetics is arguably more about making than it is about modeling. In
contrast, much of the work in cognitive neuroscience over the last 30
years has been driven primarily by the modeler’s aspirations. The devel-
opment of optogenetics, however, marks a significant step in the coming
ascendancy of maker’s knowledge of the brain. Optogenetics is a potent
reminder that that brains will soon be under our control far more than we
could have dreamed even a few years ago. Optogenetics makes salient the
extent that science progresses not only by properly describing the world
but, perhaps even more importantly, by developing new ways to change it.

Notes
1 Thanks to the Minnesota Causality and Biology workshop for comments on
early drafts, especially John Beatty, David Danks, Christopher Hitchcock,
Sarah Roe, Roberta Millstein, Ken Waters, and Marcel Weber. Thanks also
to Andreas Hütteman’s Research Group on Causality and Explanation at
 Toward an Epistemology of Intervention 171
the Universität zu Köln for helpful discussion in Spring 2011. Paul Stein and
Mark Alford offered comments on a draft. Pamela Speh assisted with figure
production. Marshall M. Weinberg and the Molecular, Cellular, & Develop-
mental Biology Department at the University of Michigan provided funds and
a workshop that called my attention to this interesting topic in the first place.
2 This paper was initially produced in August 2012, in the early days of opto-
genetics. The technique has since become a standard tool in neuroscience. I
do not review all subsequent developments; but perhaps a focus on the early
days is justified as especially important for illustrating the epistemic norms
that determine whether a technique is accepted.
3 Not all models are mechanistic models (see Bogen 2005; Craver 2007;
Kaplan and Craver 2011). Here I focus on mechanistic modelers, reverse
engineers. Makers, in contrast, are engineers.
4 Martin Carrier (2004) discusses the relationship between applied and basic
science, which is related to the narrower distinction between modelers and
makers I have in mind. I take no stand on whether one kind of knowledge is
more fundamental than the other and am content to note only that they are
distinct when it comes to an epistemology of intervention. Sometimes, the
modeler and the maker combine, as in the use of hybrid models, to deliver
theoretically motivated interventions based on modeler’s knowledge (e.g.,
Prinz et al. 2004).
5 Chang (2004) discusses aspects of detection validity in history of the
thermometer.
6 Franklin (1990) grounds his account of these strategies in Bayesian statis-
tics. Mayo (1996) provides an alternative analysis grounded in error statis-
tics and severity of testing. Weber (2012) reviews some of the advantages
and disadvantages of these and other approaches to the epistemology of
detection.
7 Franklin (2009) describes on two biological experiments: Kettlewell’s ex-
periments on moths and the Meselsohn-Stahl experiment demonstrating the
semi-conservative nature of DNA replication. In Kettlewell’s experiments,
the intervention (industrial pollution) is out of the experimenter’s hands.
The primary interventions in the Meselsohn-Stahl experiment are (a) the ra-
diolabeling of Nitrogen, and (b) the replication of bacteria. The first of these
interventions is entirely in the service of detecting the difference between
parental DNA and offspring DNA (it is an eliciting condition). The second
is left to the bacteria themselves; the experimenter decides only when to stop
the process. His exemplars thus do not afford much opportunity to consider
interventions.
8 My focus on epistemic norms contrasts with other pragmatic and other con-
cerns such as cost, ease of use, learning curves, moral norms, technological
demands, public acceptance, which obviously also influence which tech-
niques thrive and which founder.
9 Some use the term optogenetics to describe both intervention and detection
techniques (e.g., Wiens and Campbell 2018). Optogenetic detection tech-
niques, for example, use gene constructs to make cells fluoresce when they
express a protein, for example. Others are used to activate intracellular sig-
naling molecules or to uncage biologically active molecules. Here I focus on
interventions involving channel-rhodopsins inserted into neural membranes.
10 For accessible reviews, see Deisseroth (2010, 2011), and Fenno et al. (2011).
An early paper presenting the technique is Zhang et al. (2006). For a discus-
sion of application in nonhuman primates, see Diester et al. (2011).
11 Rheinberger uses the term “experimental system” to refer to include for ex-
ample, the experimental subject, the intervention and detection techniques,
172 Carl F. Craver
lab protocols, preparatory procedures, storage devices, data analysis tech-
niques, and the like (see Weber 2005; Rheinberger 1997). Rheinberger de-
scribes experimental systems as the least unit of experimental analysis. I
focus exclusively on intervention techniques, leaving the rest of Rheinberg-
er’s experimental system as background.
12 They are not interventions on T with respect to E (see Woodward 2003).
Hacking (1983) uses staining as an example of the role of intervention in the
growth of scientific knowledge. But eliciting conditions do not play the same
epistemic role as interventions into putative causes. Kästner (2017) captures
this useful point by distinguishing interventions, in Woodward’s technical
sense, from mere interactions.
13 Woodward’s account of an ideal intervention does not include requirement
(5), as it is not irrelevant to the semantics of causal claims. It is, however,
relevant to how causal claims are tested and how artifacts are avoided (see
Craver and Dan-Cohen 2021). Likewise, in experiments crucially involving
an intervention check, one would want to know if the intervention influences
the intervention check independently of the change induced in the target
variable T.
14 Not all useful interventions are ideal in Woodward’s proprietary sense:
Woodward’s goal is to provide a semantics of causal claims (that is, an an-
swer to the question, “What do we mean when we assert that X causes Y?”).
He does not provide, nor does he intend to provide, an account of the con-
straints that an experiment must satisfy to reveal useful information about
the causal structure of a system. Nor does he intend his account to serve a
model of all causal experiments.
15 The term dimension is intended informally. Still, it conveys the key idea
that the usefulness of an intervention might depend on a number of inde-
pendently varying features of the intervention.
16 This idea underlies Hacking’s (1983) use of interventions as arguments for
realism.
17 One problem with current-generation optogenetics is that different cells ex-
press the opsin molecules to a different extent, yielding a differential re-
sponse to the same light stimulus and perhaps altering system-level behavior.
See Heitmann et al. (2017). This difference might explain the early failures
of optogenetic stimulation of primate motor cortex to initiate the required
movements (see Lu et al. 2015).
18 Optogenetic stimulation does differ from physiological mechanisms in ways
that might prove important. The temporal properties of the action potentials
produced by bacterial rhodopsins are not precisely the same as in standard
action potentials. As noted above, optogenetic stimulation drives the pop-
ulation of neurons all at once, more or less synchronously. Because popula-
tions of neurons interact in virtue of patterns of activation and inactivation
across populations, the inability to produce such distributed patterns is a
potential physiological limitation of the present-day method.
19 Judgments of typicality presuppose a choice of a reference class (think, for
example, of the normal cancer cell or the typical mammalian cell express-
ing bacterial rhodopsin). Judgments of normality are inherently tinged by a
preference for health, life, the good of the species, or some such valued end
(see Craver 2014).
20 Jacques Loeb revealed himself as a paradigmatic “maker” when he wrote to
Mach:
The idea is now hovering before me that man himself can act as a creator
even in living nature, forming it eventually according to his will. Man
 Toward an Epistemology of Intervention 173
can at least succeed in a technology of living substance. Biologists label
that the production of monstrosities; railroads, telegraphs, and the rest of
the achievements of the technology of inanimate nature are accordingly
monstrosities. In any case, they are not produced by nature; man has
never encountered them.
(in Pauly 1987, p. 51)
21 One can use intervention checks to confirm the intended value of the target
variable. Tye et al. (2011) use this approach. They record from the CeA
during two major interventions. The first stimulates the BlA. The second in-
hibits the cells of the CeA onto which the cells of the BlA project. The detec-
tion technique located in the CeA is designed to check that the interventions
(electrophysiological stimulation in the BlA and light inhibition of the CeA)
changed the activity of CeA cells as desired.
22 Even under such circumstances, one might still expect occasional “noise”
in the system. Ion channels are chancy machines; one might inhibit elec-
trical activity in a cell significantly and still see occasional spikes. In other
experiments, however, the light pulses are added in addition to the inputs at
dendrites that might generate spikes on their own.

References
Allen, G. (forthcoming). A History of Genetics, 1880–1980: Its Economic,
Social and Intellectual Context.
Bacon, F. (1620). The New Organon. In G. Rees (ed.). The Oxford Francis
Bacon (Vol. 11). Oxford: Clarendon Press. Revised 2004.
Bacon, F. (1626). The New Atlantis. From Ideal Commonwealths. New York:
P.F. Collier & Son. (c)1901 The Colonial Press, expired. Prepared by Kirk
Crady from scanner output provided by Internet Wiretap. This book is in
the public domain, released August 1993. http://oregonstate.edu/instruct/
phl302/texts/bacon/atlantis.html.
Bogen, J. (2005). Regularities and Causality; Generalizations and Causal Ex-
planations. In C. F. Craver and L. Darden (eds.), Mechanisms in Biology,
Studies in History and Philosophy of Biological and Biomedical Sciences,
36: 397–420.
Boyden, E.S., Zhang, F., Bamberg, E., Nagel, G., Deisseroth, K. (2005).
Millisecond-timescale, genetically targeted optical control of neural activity.
Nature Neuroscience, 8: 1263–1268.
Carrier, M. (2004). Knowledge and Control: On the Bearing of Epistemic Val-
ues in Applied Science. In P. Machamer and G. Wolters (eds.), Science, Values
and Objectivity. Pittsburgh, PA: University of Pittsburgh Press; Konstanz:
Universitätsverlag, 275–293.
Chang, H. (2004). Inventing Temperature: Measurement and Scientific Prog-
ress. New York: Oxford University Press.
Craver, C.F. (2007). Explaining the Brain: Mechanisms and the Mosaic Unity
of Neuroscience. Oxford: Clarendon Press.
Craver, C.F. (2010). Prosthetic Models. Philosophy of Science, 77: 840–851.
Craver, C.F. (2014). Functions and Mechanisms: A Perspectivalist Account. In
P. Hunneman (ed.), Functions: Selection and Mechanisms. Springer, Synthese
Press, 133–158.
174 Carl F. Craver
Craver, C.F. and Dan-Cohen, T. (2021). Experimental Artifacts. British Journal
for the Philosophy of Science. doi: 10.1086/715202.
Craver, C.F. and Darden, L. (2013). The Search for Mechanisms: Discoveries
across the Life Sciences. Chicago, IL: University of Chicago Press.
Deisseroth K. (2010). Controlling the Brain with Light. Scientific American,
303: 48–55.
Deisseroth, K. (2011). Optogenetics. Nature Methods, 8: 26–29.
Diester, I., Kaufman, M.T., Mogri, M., Pashaie, R., Goo, W., Yizhar, O., Ra-
makrishnan, C., Deisseroth, K., Shenoy, K.V. (2011). An Optogenetic Tool-
box Designed for Primates. Nature Neuroscience, 14: 387–397. Epub Jan 30.
Dretske, F. (1988). Explaining Behavior: Reasons in a World of Causes. Cam-
bridge: MIT Press. A Bradford Book.
Eberhardt, F. (2009). Introduction to the Epistemology of Causation. The Phi-
losophy Compass, 4(6): 913–925.
Eberhardt, F., Scheines, R. (2007). Interventions and Causal Inference. Philoso-
phy of Science, 74: 981–995.
Fenno, L.E., Yizhar, O., Deisseroth, K. (2011). The development and applica-
tion of optogenetics. Annual Review of Neuroscience, 34: 389–412.
Franklin, A. (1986). The Neglect of Experiment. Cambridge: Cambridge Uni-
versity Press.
Franklin, A. (1990). Experiment, Right or Wrong. Cambridge: Cambridge Uni-
versity Press.
Franklin, A. (2012). Experimentation in Physics. Stanford Encyclope-
dia of Philosophy. Available online at http://plato.stanford.edu/entries/
physics-experiment/.
Gunaydin, L.A., Yizhar, O., Berndt, A., Sohal, V.S., Deisseroth, K., Hegemann,
P. (2010). Ultrafast Optogenetic Control. Nature Neuroscience, 13: 387–392.
Hacking, I. (1983). Representing and Intervening. Cambridge: Cambridge Uni-
versity Press.
Heitmann, S., Rule, M., Truccolo, W., Ermentrout, B. (January 2017). “Op-
togenetic Stimulation Shifts the Excitability of Cerebral Cortex from Type I
to Type II: Oscillation Onset and Wave Propagation”. PLOS Computational
Biology, 13(1): e1005349.
Kaplan, D.M., Craver, C.F. (2011). The Explanatory Force of Dynamical Mod-
els. Philosophy of Science 78: 601–627.
Lobo, M.K., Covington, H.E., Chaudhury, D., Friedman, A.K., Sun, H.,
Damez-Werno, D., Dietz, D.M., Zaman, S., Koo, J.W., Kennedy, P.J., Mou-
zon, E. (2010). Cell Type–Specific Loss of BDNF Signaling Mimics Optoge-
netic Control of Cocaine Reward. Science 330(6002): 385–390.
Lu, Y., Truccolo, W., Wagner, F.B., Vargas-Irwin, C.E., Ozden, I., Zimmer-
mann, J.B., May, T., Agha, N.S., Wang, J., Nurmikko, A.V. (June 2015).
Optogenetically Induced Spatiotemporal Gamma Oscillations and Neuro-
nal Spiking Activity in Primate Motor Cortex. Journal of Neurophysiology,
113(10): 3574–3587.
Kästner, L. (2017). Philosophy of Cognitive Neuroscience: Causal Explana-
tions, Mechanisms & Empirical Manipulations. Berlin: Ontos/DeGruyter.
Kolar, K., Knobloch, C., Stork, H., Žnidarič, M., Weber, W. (2018). Opto-
Base: A Web Platform for Molecular Optogenetics. ACS Synthetic Biology, 7:
1825–1828. doi: 10.1021/acssynbio.8b00120.
 Toward an Epistemology of Intervention 175
Mayo, D.G. (1996). Error and the Growth of Experimental Knowledge. Chi-
cago, IL: University of Chicago Press.
Pauly, P.J. (1987). Controlling Life: Jacques Loeb and the Engineering Ideal in
Biology. New York: Oxford University Press.
Prinz, A., Abbott, L.F., Marder, E. (2004). The Dynamic Clamp Comes of Age.
Trends in Neurosciences 27: 218–24. doi: 10.1016/j.tins.2004.02.004.
Rheinberger, H.-J. (1997). Toward a History of Epistemic Things. Stanford,
CA: Stanford University Press.
Sargent, R.-M. (2001). Baconian Experimentalism: Comments on McMullin’s
History of the Philosophy of Science. Philosophy of Science 68: 311–317.
Sargent, R.-M. (2012). Bacon to Banks: The Vision and the Realities of Pur-
suing Science for the Common Good. Studies in History and Philosophy of
Science 43: 82–90.
Tsai, H.C., Zhang, F., Adamantidis, A., Stuber, G.D., Bonci, A., de Lecea, L.,
Deisseroth, K. (2009). Phasic Firing in Dopaminergic Neurons Is Sufficient
for Behavioral Conditioning. Science, 324: 1080–1084.
Tye, K.M., Prakash, R., Kim, S.Y., Fenno, L.E., Grosenick, L., Zarabi, H.,
Thompson, K.R., Gradinaru, V., Ramakrishnan, C., Deisseroth, K. (2011).
Amygdala Circuitry Mediating Reversible and Bidirectional Control of Anx-
iety. Nature, 471: 358–362.
Waters, K. (2008). How Practical Know-How About Experimentation Contex-
tualizes Theoretical Knowledge. Philosophy of Science, 75: 707–719.
Waters, K. (2014). Shifting Attention from Theory to Practice in Philosophy
of Biology. In M.C. Galavotti, D. Dieks, W.J. Gonzalez, S. Hartmann, T.
Uebel, M. Weber (eds.) New Directions in the Philosophy of Science. Berlin:
Springer International Publishing, 121–139.
Weber, M. (2005). Philosophy of Experimental Biology. Cambridge: Cam-
bridge University Press.
Weber, M. (2012). Experimentation in Biology. Stanford Encyclope-
dia of Philosophy. Available online at http://plato.stanford.edu/entries/
biology-experiment/.
Wiens, M.D., R. Campbell. (2018). Surveying the Landscape of Optogenetic
Methods for Detection of Protein–Protein Interactions. Wiley Interdisciplin-
ary Reviews. Systems Biology and Medicine. doi: 10.1002/wsbm.1415.
Woodward, J. (2003). Making Things Happen. New York: Oxford University
Press.
Zhang, F., Wang, L.P., Boyden, E.S., Deisseroth, K. (2006). Channelrhodopsin-
2 and Optical Control of Excitable Cells. Nature Methods, 3: 785–792.
8 Triangulating Tools in the
Messiness of Cognitive
Neuroscience
Antonella Tramacere

1 Why Triangulating Tools?

Imagine you are Gina, a four-year-old child playing in the kitchen and
finding a piece of brown stuff on the ground. At first glance, the stuff
does look attractive. It could be a piece of chocolate, which survived
from your last Easter egg. But it could also be a piece of soil coming
from an apartment plant, or a piece of dried excrement from Leon, the
house cat. Continuing to observe it is not going to give you any conclu-
sive information; therefore, you decide to touch the undefined object and
to inspect its texture. You are still unsure of its identity. You bring it to
the nose to smell it, but the dryness has wiped any significant odor out.
Finally, you decide to compare the touch and the smell of the object with
those of the soil of other plants present in the apartment, and of Leon’s
feces in the litter. Now, the suspicion becomes somewhat real; the object
is probably a piece of excrement from Leon.
This is not a classical textbook example, but it is still a case of trian-
gulation, namely the use of multiple information sources to identify an
object, supplying for the limitations of each source. Confronted with
the task of identifying an undefined object, Gina triangulates on it by
employing more than one sensory modality. The combination of vision,
touch, and smell has helped her to investigate the object from different
angles, and to supply for the limitation of each sensory modality. Gina
eventually refined the initial hypothesis, by minimizing the bias of each
modality, and through focused observations. Finally, she avoided eat-
ing a potentially harmful thing. The child has unwittingly employed an
epistemic strategy, which has been used by scientists and discussed by
philosophers for decades, perhaps centuries.
In the philosophy of science, triangulation is defined as the use of two
or more lines of determination to validate the existence or the properties
of objects (Wimsatt, 2012). Consider objects in a wide sense to be phe-
nomena, entities, regularities, research results and so on. According to
this definition, objects that have been validated through multiple lines of
determination are more likely to be robust, that is they are more likely to
be real, to be reliable and trustworthy, and to generalize across different

DOI: 10.4324/9781003251392-11
 Messiness of Cognitive Neuroscience 177
experimental situations (Culp, 1994; Wimsatt, 2012). Triangulation has
received considerable attention in the literature (Soler et al., 2012). Vari-
ous case studies have shown that scientists (at least sometimes) appeal to
triangulation to determine the degree of robustness about objects, and
thereby validate their hypotheses.
I will discuss triangulation in cognitive neuroscience. Identifying the
role of brain processes and mechanisms underlying cognition is the goal
of cognitive neuroscience. To make a parallel with Gina’s case, neurosci-
entists aim to identify some yet undefined object. But in the case of cog-
nitive neuroscience, the undefined stuff is constituted by brain processes,
operations, and functions in cognition. Additionally, like in Gina’s case,
pursuing this goal is made difficult by the limitations of each one of the
methods that can be used to investigate the object of interest.
In fact, in cognitive neuroscience, the role of brain mechanisms un-
derlying cognition can only be indirectly investigated, because of the
intrinsic limitations of experimental interventions, and of techniques
used to measure the variable (Weichwald & Peters, 2021). Neuroscien-
tists study brain activity and connectivity by using various tools (widely
understood as methods, techniques, or procedures), providing data that
are correlative in nature, and dependent on confounding factors that
may produce partial or contradictory results. Not only may the different
tools yield different outputs at the level of raw data, but they may not
even focus on the same aspects of the phenomenon, thus making it un-
clear whether they provide evidence for the same hypothesis.
Using different methods or tools for investigating the role of the brain
in psychological phenomena inevitably produces a considerable amount
of diverse data that await to be integrated and validated. Integration
and validation are needed because we want to know how data and find-
ings produced by different tools relate to each other and to what extent
they provide support for inferences bridging neural data to psychological
phenomena. Given the limitations of tools leading to diverse (types of)
data and potentially discordant results, could triangulation help with in-
tegrating findings and validating hypotheses about the role of the brain
in cognition?
This is not a trivial question. To date, much philosophical ink has
been spilled on the question of whether and to what extent triangulation
warrants belief in the triangulated-upon object, but less attention has
been paid to the ways that discordance is itself epistemically valuable.
This is the goal of this chapter. I will first present a case study, where
triangulating different tools (call this methodological triangulation) has
successfully increased the confidence in our understanding of a causal
relation. Specifically, I will analyze research on the causal role of high
values of systolic blood pressure in coronary heart disease. I will then
present a comparatively messy case from cognitive neuroscience to
178 Antonella Tramacere
analyze the epistemic force of triangulation when data and findings are
inconsistent and divergent. Specifically, I will focus on research about
the role of the mirror neuron system in autism spectrum disorder. By
comparing this case study to the previous one, I will show that triangu-
lating tools has comparable value in both cases.
Triangulation is useful when tools have limitations independently of
whether results are concordant or not because incoherence and diver-
gence are not insurmountable problems for validation of robustness. As
in Gina’s story, the utility of triangulation is not necessarily conditioned
on reaching the same conclusion with all information sources, but rather
on acquiring multidimensional information on objects and a modest
unification of phenomena.

2 Triangulation to Robustness
Triangulation refers to the use of multiple independent lines of deter-
mination to ascertain whether an object is robust, namely, to exclude
that it is an artifact of a particular method. The more a phenomenon
is confirmed under multiple independent lines of determination, the
more it is likely to be robust (Wimsatt, 2012). Triangulation as multiple
determination connects to an epistemic conception of robustness, as it
is used to deliver reliable connections between a claim and some fact
about the world. This conception of robustness is particularly relevant to
methodological triangulation because multiple tools for inquiring into a
phenomenon may or may not produce the same results or consistent in-
formation. Consequently, it is important to ascertain whether and how
various lines of evidence point to the same conclusion despite different
modalities or experimental procedures.
Robustness is not an all-or-nothing construct, because it comes in de-
grees and depends on the context of validation. This means that objects
or conclusions are evaluated against a set of contextual factors, depending
on the question asked by scientists, the formulated hypothesis, and the
specific methods used in the experimental setting. I will say more about
the context-dependent features of robustness throughout the article.
Many philosophers apply the term “robustness” strictly in the context
of debates about models and model outcomes (see Odenbaugh & Alex-
androva, 2011; Orzack & Sober, 1993; Weisberg, 2006). According to
these philosophers, the term ‘robustness’ indicates that model outcomes
do not change with a fixed body of data when assumptions are varied. I
do not use robustness in this sense. I am instead concerned with robust-
ness understood as a form of validation of phenomena through triangu-
lation of different tools (see Cartwright, 1991; Woodward, 2006) for
analyses of varieties of robustness).
Because independence is a condition of possibility for robustness (but
see Stegenga & Menon 2017 for a critical discussion), scholars have
 Messiness of Cognitive Neuroscience 179
largely discussed what makes one tool independent from another. Ac-
cording to a widely used notion, tools are independent when they have
unrelated sources of potential biases or confounding factors (Kuorikoski
& Marchionni, 2016; Munafò & Smith, 2018; Woodward, 2006).
When the possibility to incur an error using one tool is unrelated to
the possibility of errors with another tool, these tools can be defined as
independent. This implies that, to counterbalance the flaws or the weak-
ness of one tool with the strengths of another, various tools relying on
independent assumptions are needed.
Remember Gina’s case. She used different sensory modalities to tri-
angulate on the undefined object, because one single modality, such as
vision, could not provide conclusive evidence. On the contrary, the com-
bination of different modalities helped to counterbalance the weakness
of each one. In one sense, the sensory modalities are independent because
they are subjected to different sensory biases. Nevertheless, the sensory
modalities pertain to the same subject, that is Gina, and are therefore
dependent on potential cognitive biases that she may have. This sug-
gests that the independence of different tools is always contextual and
depends on the type of tool. Triangulation for different questions, with
different tools or with different subjects would require different criteria
of independence.
Triangulation also requires that evidence is commensurable, namely
tractable within a consistent set of background assumptions. Since
different methods often use “different languages” to generate data, it
is possible that scientists are not able to amalgamate them to validate
empirical claims about objects. Correctly individuating independence,
while avoiding incommensurability, has been considered a hard problem
for triangulation (Stegenga, 2009). Using different methods to inquire
into an object could increase the chance of errors, and since different
methods often rely on different sets of assumptions, any attempt to com-
bine them may lead to confusion.
Consequently, it seems that triangulation relies on integrating com-
mensurable and reliable data. We need to scrutinize data that depend
on limitations of tools or study design, and we need to know whether
and how different data pertains to the question under investigation. It
has been noted however that, if practitioners knew that the methods
used produce reliable and commensurable data about an investigated
phenomenon, they would not need triangulation (Hudson, 2013; Ste-
genga, 2009). They could simply use one single method for investigating
the phenomena, chosen for its precision and reliability.
These observations have produced skepticism about the value of trian-
gulation. If triangulation requires that evidence is reliable and commen-
surable, the use of multiple independent methods would be restricted
to specific situations. More specifically, triangulation would be valu-
able only when we know that the different methods produce reliable
180 Antonella Tramacere
evidence that can be amalgamated. But if scientists knew it in advance,
they would not need to triangulate methods. This is a serious criticism,
but I think that the skepticism for triangulation can be mitigated if we
consider it as neither a necessary nor a sufficient strategy for robustness.1
Triangulation is only one possible strategy to minimize confounders and
bias when we observe, measure, and thereby interpret and understand
objects, reaching a modest and local level of scientific unification.
In what follows, I will try to mitigate the criticisms addressed to trian-
gulation, by describing it as a way for minimizing the limitations of tools
and for contributing to the integration of data and findings. When we
don’t trust methods, evidence concordance works as a regulative ideal
to conduct research. Irrespective of the concordance of the evidence,
triangulation can be valuable to merging disparate sources of data ob-
tained through available methods and be instrumental to the validation
of empirical claims by assessing the robustness of data-to-phenomena
inferences.
I will discuss a case from etiological epidemiology to analyze how
triangulation can be used to increase the confidence in the causal role
of specific variables in the emergence of a disease. I will show how the
robustness of a causal relation can be assessed in ideal cases of evi-
dence concordance, and how multiple lines of determination can help
in determining the reliability of the methods, through analysis of the as-
sumptions and possible limitations of tools used in the experimental in-
terventions and manipulation of variables. Later, I will use this example
to analyze the case of evidence discordance in cognitive neuroscience.
I will leave aside the issue of what makes a causal claim causal, and
when causality can be inferred with reasonable certainty from data, as
this has been already the object of numerous investigations. Note that
I rely on interventionism (Cartwright, 2006; Hausman & Woodward,
1999), as an account of causation which illustrates (and problematize)
the ways scientists uncover the causal structure of the world through
focused interventions (Eberhardt & Scheines, 2007; Spirtes & Scheines,
2004). Questions asked by scientists are often causal and aimed at event
prediction, even when single experiments can only highlight correlations
among variables (Nathan, this volume). Because experiments and inter-
ventions are important to validate causal inference about phenomena,
analyzing the value of methodological triangulation is important to cor-
rectly identify the causal story behind the data.

3 A Case of Convergence from Epidemiology

In the International Journal of Epidemiology, Lawlor and colleagues
(2016) discuss whether triangulation of different statistical methods can
clarify the effect of lowering high blood pressure (BP) on the occurrence
of coronary heart disease (CHD).
 Messiness of Cognitive Neuroscience 181
One method is the good old prospective cohort study. This is an ob-
servational study which follows over time a group of similar individuals
differing with respect to certain factors (e.g., level of BP), to determine
how these factors affect certain outcomes (e.g., coronary accidents). By
conducting and meta-analyzing various observational studies, epidemi-
ologists have calculated the number of fatal and non-fatal CHD in indi-
viduals with different values of blood pressure across time (Lawlor et al.,
2016). The results showed that individuals with 10mmHg lower in BP
were less affected by CHDs.
Can we be confident that lowering high blood pressure decreases the
incidence of CHD? Not really. Observational studies have limitations.
It is possible that lowering high BP has no direct effects on CHD, be-
cause individuals with higher values of BP have typically more metabolic
comorbidities, and adiposity could for instance confound the observed
effect. This means that higher levels of BP per se might not directly in-
crease the risk of cardiovascular injuries; other factors such as obesity
could produce elevated values of BP, thus leading to a cycle of accumu-
lating vascular injury through time.
Epidemiologists also used another tool to investigate the relation be-
tween BP and CHD. The use of another technique to inquire into the
same causal question contradicts some criticisms that have been raised
against triangulation. Hudson (2013), for example, claimed that when
scientists compare one tool with the measurement of another tool, they
base this on the assumption that one tool is reliable, and the compari-
son serves to check the reliability of the second tool. On the contrary,
epidemiologists added another line of determination for checking the
reliability of the phenomenon, and not the tool.
Specifically, epidemiologists utilized randomized control trials (RCTs)
to inquire into the effect of a higher level of BP. Participants are ran-
domly assigned to the experimental group (treated with a drug) or to the
control group (treated with placebo). The experimental group has been
treated with an antihypertensive drug, to reduce BP by 10mmHg, while
the control group has only taken a saline solution. A decrease in cardio-
vascular accidents in the group treated with the drug would support the
hypothesis that lowering high values of BP causes a reduction in the risk
of CHD.
Interestingly, many RCT studies have been conducted on this ques-
tion. Later, the various results were integrated and meta-analyzed (Law
et al., 2009), supporting the hypothesis that lowering high BP of 10
mmHg effectively reduces risks of cardiovascular accidents. Is this result
robust enough to support the investigated causal hypothesis, and eventu-
ally guide preventive intervention in the population? Not yet.
Despite being considered the goal standard of epidemiologists, RCT
studies also present limitations. This is because the hypertensive drug
might not exclusively affect BP. The same drug could have several
182 Antonella Tramacere
additional effects, making it unclear whether the reduced incidence of
CHD in the experimental group is due to the target intervention or to
other independent factors. For instance, commonly used antihyperten-
sive drugs interfere with given regulative hormonal loops, which could
have cardioprotective effects independently from its known effect on BP.
Epidemiologists have inquired about the relation between BP and
CHD through a third line of determination (Ference et al., 2014). Specif-
ically, they have looked at genetic variability through a technique called
Mendelian randomization (MR). They were able to inquire into the en-
dogenous genetic variation associated with lower values of BP in large
samples of individuals while controlling for Body Mass Index and other
confounders. The results of MR largely converged with the other lines of
determinations, providing additional independent support that 10mmHg
lower BP significantly reduces the risk of cardiovascular accidents.
This seems to be successful, but there is still some problem. We now
know that lowering high values of BP decreases CHD in the population.
But how much should BP be lowered to save the highest number of in-
dividuals? Looking closer, the three lines of studies differed in respect
to the magnitude of the benefits of reducing high values of BP. Specifi-
cally, it seems that the benefit of lowering BP of 10mmHg is greatest for
genetic studies, intermediate for observational studies and least for the
RCTs. Which method should we trust more?
Contrary to the criticism developed by Stegenga (2009, 2012), for
which multiple determination is not valuable with discordance of ev-
idence, additional analyses propose a solution for the apparently dis-
cordant results. Scientists analyzed conditions of the study design, and
samples of individuals involved in each of the approaches. Specifically,
they have compared the duration of exposure to the risk factor (i.e.,
higher level of systolic blood pressure) in the target population, noting
that they differ among the various studies. When these differences are
taken into account, the three sets of results are broadly consistent with
each other and can provide more precise information for quantifying the
benefits of reducing BP of 10mmHg in the population.
Various aspects of this case study are relevant for answering the criti-
cisms addressed to the value of triangulation. Before analyzing these as-
pects, however, I will present a case from cognitive neuroscience, which
does not have such a happy ending, because the results of various tools
eventually are incongruent and do not converge. By comparing the two
cases, I will show that triangulation is beneficial also in situations of
evidence discordance.

4 A Case of Discordance from Cognitive Neuroscience

Considerable efforts have been devoted to the identification of the neural
bases causing the cognitive deficits observed in autism spectrum disorder
 Messiness of Cognitive Neuroscience 183
(ASD). Although a conclusive answer to this question has not been pro-
vided, progress has been made.
ASD is a multidimensional psychiatric disorder presenting a spectrum
of behavioral and cognitive impairments, such as difficulties with ver-
bal and non-verbal communication, repetitive or restrictive behaviors,
abnormal social interaction, and restrictive specialized interests (Myers
et al., 2007). ASD has risen in the last decades, with more diagnoses
across countries and categories of subjects, perhaps due to an increased
awareness and knowledge of the disease. Consequently, ASD has be-
come one of the most studied neuropsychiatric disorders.
During about the last 20 years, the discovery of mirror neurons has
boosted new lines of investigations on the specific neural mechanisms
underlying ASD (Williams et al., 2001). Mirror neurons are motor neu-
rons that activate during both the execution and perception of the same
or similar actions (Gallese et al., 1996), and that are organized in func-
tional spots in frontal-parietal areas. These neurons are embedded in
the so-called mirror neuron system (MNS), a network of interconnected
neural areas crucially implicated in social functions, such as mimicry,
empathy, imitation, non-verbal communication, and language (Ferrari
& Rizzolatti, 2015).
Interestingly, these functions overlap with the social cognitive symp-
toms in patients with ASD, motivating the formulation of the broken
mirror hypothesis (Ramachandran & Oberman, 2006), which states
that dysfunction of key nodes of the MNS is at the basis of the social
deficits observed in ASD. The broken mirror hypothesis has been tested
through multiple indirect tools because MNS dysfunction in patients
with autism cannot be investigated directly by putting electrodes in their
brains. Scientists have used a specific electroencephalogram (EEG) mea-
surement, called the mu wave, which is blocked anytime a person makes
a voluntary muscle movement and also when a person watches someone
else performing the same action.
Mu wave is insufficient to identify the role of mirror neurons in ASD
because EEG provides a good temporal resolution at the expense of the
spatial localization of neural activation. Functional magnetic resonance
(fMRI) has also been used. While offering a rough temporal character-
ization of neural activation through blood oxygenation level-dependent
measure in specific brain areas, fMRI allows better spatial localization
of the neural activation during specific tasks. But neither EEG nor fMRI
can provide a measurement of the motor potential to identify distur-
bance of mirror neurons’ activity. Using Transcranial Magnetic Stimu-
lation (TMS) during observation of others’ actions could do this, at the
expense of the localization of brain activation.
The use of the three techniques has produced inconsistent and dis-
cordant results. Some EEG studies did support (Oberman et al., 2005;
Palau-Baduell et al., 2011), while others did not support the relation
184 Antonella Tramacere
between MNS and ASD (Fan et al., 2010; Raymaekers et al., 2009). The
same has occurred with fMRI, insofar as some studies have observed
reduced activity in the MNS in ASD patients (Dapretto et al., 2006;
Fishman et al., 2014; Martineau et al., 2010), while other studies have
not (Dinstein et al., 2010). In a similar way, some studies conducted with
TMS did (Enticott et al., 2012), while others did not support the exis-
tence of the relation between mirroring mechanisms and ASD symptoms
(Enticott et al., 2013).
As a consequence, scientists asked whether the various techniques are
really investigating the same phenomenon; some questioned whether con-
sidering the mu wave in EEG experiments is suitable for identifying the
activity of the MNS (Hobson & Bishop, 2016) because mu suppression
could also be an index of differences in attentional engagements. Others
questions whether the experimental method (i.e., repetition suppression)
used with the fMRI can detect motor activation (Fuelscher et al., 2019).
Repetition suppression occurs when neural activity is suppressed within
a population of neurons following repeated activation by the successive
presentation of stimuli to which these neurons are sensitive. Since the
populations of mirror neurons activated during both observation and
execution of movements, repetition suppression effects are expected in
both conditions during repeated presentations of stimuli. The mixed re-
sults offered by fMRI investigations have led scientists to delve deeper
into the neurophysiological mechanisms of repetition suppression and to
refine the statistical methods used to analyze areas of activations.
Scientists also have analyzed the degree of independence of the ap-
proaches, the particular aspects of the experimental intervention, such
as the types of stimuli, and of the subject samples, such age and severity
of the symptoms, trying to determine whether and how all these factors
could play a role in the inconsistency and divergence of the observed
findings (Khalil et al., 2018; Yates & Hobson, 2020). For instance, it has
been noted that experimental conditions using stimuli with low or null
social relevance are incapable of detecting the activity of the MNS. Fur-
ther, disturbance of mirroring mechanisms is more consistent in younger
autistic subjects compared to adults.
Note that scientists utilized multiple tools within the same experi-
mental setting to investigate the role of MNS in some of the cognitive
deficits shown in ASD. This practice is progressively spreading in con-
temporary neuroscience (see Cichy & Oliva, 2020), and I will have more
to say about it later. For example, both TMS and EEG have been used
to inquire into the activation of the MNS in ASD, providing moderate
support for the existence of the relation between the neural network and
the symptoms of the disease (Cole et al., 2018). The combination of mul-
tiple tools within the same experimental setting has also been utilized
to inquire into the role of MNS activations in facial mimicry (Likowski
 Messiness of Cognitive Neuroscience 185
et al., 2012), a phenomenon in which the causal role of the MNS is now
considered relatively more robust.
Together, the analyses of limitations, confounders and contextual
conditions have led to the refinement of the initial hypothesis. The
novel hypothesis integrates diverse results obtained through previous
experiments, by stating that the observed inconsistencies could depend
on the MNS affecting social cognition by means of the interactions
with other brain regions, such as the prefrontal cortex and temporopa-
rietal junctions (Yates & Hobson, 2020). Consequently, scientists are
reasoning, only interventions using certain types of inputs (i.e., stim-
uli with social significance) may detect MNS dysfunction. Also, this
prediction seems to hold only for children with ASD, because adults
with autism do not consistently display both behavioral impairments
and atypical MS activity during tasks typically associated with MNS
functioning.
The research on this question is ongoing, and likely more experiments
will be conducted in the future. Regardless of the answer, it is important
to ask what we can learn about the value of triangulation by comparing
this messy case with the previous one, where triangulation has success-
fully helped to establish the existence of a causal relation. What is the
value of triangulation in cases of evidence discordance? Can the scien-
tific practices I have described help to mitigate the criticisms addressed
to triangulation?

5 Concordance as a Regulative Ideal

The rationale of triangulation is intuitive, at least in the case of evidence
concordance. Multiple determined evidence typically provides greater
support to a hypothesis (even though this is not always the case, see
Hudson, 2013; Rasmussen, 2001 for discussions). By triangulating var-
ious methods, epidemiologists have provided support for the claim that
higher values of BP cause a higher occurrence of CHD because the evi-
dence is concordant. But multiple tools have produced inconsistent and
divergent results in the neuroscientific case study, leaving us with the
uncertainty of whether disturbance of the MNS causes deficits in ASD.
What would triangulation prescribe in this case? How could it help us
with answering the research question?
Some philosophers have claimed that triangulation is far less signifi-
cant or frequent than commonly assumed (see Hudson, 2013), that scien-
tists typically resist examining alternative procedures, and use only one
method to inquire into phenomena, chosen because of its reliability. Oth-
ers (such as Stegenga, 2009, 2012) claimed that, even when triangulation
is used, findings are often inconsistent and/or divergent. In these situa-
tions, triangulation has no epistemic value, and can provide no guidance
186 Antonella Tramacere
for deciding which evidence is reliable, and which should therefore be
amalgamated to validate empirical claims. The only recommendation
that triangulation would provide, critics would continue, is to conduct
more experiments till convergence is reached, but this recommendation
is simply too general. No scientist would need such a vague maxim.
I think these criticisms do not take into consideration that triangula-
tion is used when tools are not trusted and that in this case, triangulation
allows to achieve a series of epistemic goals with both concordance and
discordance of evidence. In what follows, I will show that, irrespective
of the concordance of the evidence, multiple determination is used and
useful for error minimization, it requires (and thus promotes) integra-
tion of data and findings, it leads to refining hypotheses and eventually
explanations, and therefore contributes to achieve a better understand-
ing of objects and phenomena.
Let me explain this bold claim, by starting with the role of triangula-
tion for error minimization. Triangulation of tools requires controlling
for the reliability of the processes through which evidence is generated.
For example, inquiring about the relation between BP and CHD did not
stop with prima facie confirmation with one tool. Validating the strength
of the relation, and the magnitude of the estimated effect also required
controlling for bias, potential limitations, and residual confounders of
each of the methods. In a similar way, the investigation of the relation
between MNS and ASD did not stop with inconsistency and discordance
of results. The lack of convergence did not produce a stalemate or led
to abandoning the hypothesis, and rather motivated the analysis of why
concordance is not reached.
These practices suggest that multiple determination involves analyzing
background conditions, confounding variables and auxiliary hypotheses
of the methods and tools used to address the research question (Bogen
& Woodward, 1988). This is in line with Kuorikoski and Marchionni
(2016) analysis of triangulation, for which the latter involves reasoning
about the particular context of validation, about the processes that gen-
erate the data, and it is thereby useful for error minimization in both
cases of evidence concordance and discordance. This is also in line with
previous statements that robustness is always contextual because evi-
dence is evaluated against a set of contextual factors, depending on the
question asked by scientists, the formulated hypothesis, and the specific
methods used in the experimental setting.
Also note that, to conduct these analyses, epidemiologists and neuro-
scientists have scrutinized (i.e., eliminating statistical artifacts), selected
and processed raw data. They have analyzed the refined data and con-
trasted their value with results from different methods in the light of
the investigated question. Integration has occurred while pre-processing
and selecting raw data, while analyzing them, while interpreting the re-
sults of these elaborations, and while comparing these results with those
 Messiness of Cognitive Neuroscience 187
from different measures. Processing, merging, and processing data, un-
derstanding whether and how they produce results that may finally be
used to explain a phenomenon is a value in science because the more the
data (and then results) are integrated, the more the science is unified. 2
You may wonder how unification is even possible in case of inconsis-
tent and divergent evidence, for unification requires coherence. But if
you take unification to modestly mean interconnection between items
within and across sub-field (Grantham, 2015) all you need is to recipro-
cally connect findings produced by different experimental procedures.
Evidence is unified in the light of an updated hypothesis stating that
ASD in young subjects is due to dysfunctional activation in neural areas
beyond the MNS. The novel hypothesis connects and thus integrates a
sub-set of previous findings in a coherent way and leads to novel rounds
of experimentation that are thought to better address the original re-
search question. Comparing multiple evidence often relies on meta-anal-
yses and integrative reviews from multiple tools, on which scientist’s
base for conducting new experiments.
Critics of triangulation (such as Stegenga 2009) are right in saying that
we lack principled criteria to quantify in advance the weight of each line
of determination eventually converging on a conclusion; they are also
right in claiming that we lack ways to assess the degree of robustness of
a hypothesis when multiple determination is made and show inconsis-
tencies. It is also impossible to consider and compare all the background
conditions, potential biases, and auxiliary assumptions of tools, which
may eventually undermine the independence of the determination, and
consequently the robustness of claims. Robustness attributions to phe-
nomena are always open to revision.
But this is not a sufficient reason to dismiss the epistemic value of
triangulation, which is always contextual and dependent on the re-
search question. Accordingly, triangulation on tools has value, because
it requires mutual comparison and adjustment of results obtained by
different experimental groups using various experimental procedures.
Consequently, triangulation can contribute to (local and modest) inte-
gration and unification of findings because multiple evidence compari-
son involves checking whether the data themselves ‘make sense’, how it
relates to data from other tools, how they depend on common or inde-
pendent sources of error, and how well the data produced by a particular
technique pertain and support the initial hypothesis.
That triangulation can be understood as a valuable, context-depen-
dent practice for minimizing errors and biases is additionally confirmed
by the simultaneous use of multiple tools within the same experiment
described in the previous section. Call this simultaneous methodological
triangulation. This practice contradicts criticisms that, not only trian-
gulation is not common in science, but it is also not used to validate
phenomena in daily scientific practice (Hudson, 2013).
188 Antonella Tramacere
In contrast, I have shown that neuroscientists explicitly and simul-
taneously direct a range of methods at a particular cognitive system or
research problem about cognition, to gain a multidimensional under-
standing of how the system works. As such, simultaneous triangulation
is a confirmation that evidence concordance guides the comparison, the
mutual adjustments of independent evidence and eventually their inte-
gration. In other words, scientists conduct experiments with different
tools and expect (or better, they hope!) that the results of the experi-
ments will converge and be unified.
Simultaneous triangulation also provides support for the view that
error minimization depends on the nature of the research question. Con-
sider again the case of the causal role of SBP in CHD. In this case, using
different statistical methods with the same samples of participants was
considered a violation of independence, because it could increase the
chances of biases in recruiting and allocating participants. On the con-
trary, using multiple tools with the same subject’s sample and in the same
experimental setting in cognitive neuroscience is (at least in some cases)
desirable; it can clarify how different tools highlight (same or compat-
ible) aspects of the relation between neural components and cognition,
and for minimizing confounders due to interindividual brain variability.
In conclusion, while discordance of results may be a serious problem
for a notion of triangulation that aims to provide quantifiable criteria for
assessing robustness, it does not undermine the value of triangulation as
a regulative ideal for conducting research. I thus agree with those (Coko,
2020; Trizio, 2012) who suggest that multiple determinations can guide
scientists in evaluating the conditions for allocating confidence in em-
pirical claims, where these evaluations are contextual, depending on the
methods which have been used to address the research question, and
not dependent on a priori criteria. Concordance thus plays the role of a
“methodological attractor”—i.e., “an idealized logical scheme guiding
the efforts of experimenters and technicians, and at the same time, play-
ing a crucial role in the acceptance of a result by the scientific commu-
nity” (Trizio 2012, p. 120).

6 Discussion and Conclusion

Let us wrap up the journey we have done so far, before concluding with
general thoughts about the value of tool triangulation and the undefined
brown stuff that Gina found on the ground.
By taking into consideration a case of success and a case of (at least
temporary) failure, I tried to address criticisms stating that philosophers
have too often focused on scientific success, disregarding that “the lovely
paths of truth in the valleys of scientific struggle are often discordant”
(Stegenga, 2009, p. 656). The discussion of only two examples of tri-
angulation with different conclusions and different epistemic forces,
 Messiness of Cognitive Neuroscience 189
cannot make up entirely for this tendency. Nonetheless, I hope that I
have added reasons to claim that triangulation is effectively employed in
different sub-fields of science, and in different epistemic situations.
As Lawlor and colleagues have noted (2016), the term triangulation
(or any other synonym used in the philosophical literature) is not nec-
essarily mentioned, but the conducted analyses are often in line with
principles and requirements: multiply determine the existence or prop-
erty of a phenomenon to validate it, thereby assessing its degree of ro-
bustness. In both cases of methodological triangulation discussed here,
the rationale is similar: where methods and tools have limitations and
are not sufficiently reliable, multiple determination is a valuable strat-
egy, because it allows the researcher to control the process of evidence
production.
Scientists in different sub-fields pursue multiple lines of determination
and combine techniques to disambiguate unclear findings, conscious of
the fact that tools do produce results which are affected by confound-
ers. We are, especially with a novel and difficult discipline like cognitive
neuroscience, in a sort of “drunkard’s search.” We cannot directly reach
the entities that we need to measure, we often do not know whether and
how obtained measures are critical for our questions, we measure only
things that we can measure, trying to find some meaningful and reliable
patterns in there.
Bechtel stated this almost 20 years ago (2002)

In the cognitive neurosciences, where a major goal is to relate neural

structures to cognitive operations, no one technique can reveal what
cognitive operation is performed by a given brain area, but integrat-
ing the results from multiple techniques can provide a much better
understanding.

Neuroscientists will often find themselves in the situation of not know-

ing what to trust. In a situation like this one, it is not even possible to
be a purist of the method. Using multiple methods to address a causal
question in neuroscience is a rational strategy to analyze how results of
observations relate to each other through a process of selection, mutual
adjustment and integration of data and findings. This process is always
contextual and relative to the research question, tool, and context of
validation. But, despite its contextual and local aspect, it serves to high-
light reciprocal constraints between (psychological) phenomena and un-
derlying brain processes, therefore aiding our understanding of how the
mind works.
In the valley of neuroscientific struggle to understand how the mind
works, we are like Gina, a four-year-old child unsure of whether what
we are observing is chocolate, excrement, or something not properly ex-
citing, like dry soil, which after all can be molded to do something else.
190 Antonella Tramacere
We can proceed step-by-step; we can assume (or hope!) a consilience
of induction (Whewell & Butts, 1968), namely that what we discover
in one situation through the tools that we possess, will hold for novel
situations with same or different tools. Probably, we will not reach any
unified and orderly word where few principles hold for explaining things
within and beyond our mind, but we can still see how far we can go.

Acknowledgment
I would like to thank Francesco Bianchini, Marco Nathan, Adrian
Spohr and Alfredo Vernazzani for critical comments and discussions on
an early version of the chapter.

Notes
1 Triangulation is neither necessary nor sufficient for robustness. Triangula-
tion is only one among the many possible epistemic strategies for assessing
robustness about objects, based on the assumption that it is at best unlikely
that multiple seemingly independent methods provide the same conclusion.
This is known as the no miracle argument (Hacking, 1983). Triangulation
is also not necessarily connected to integration of data and findings, be-
cause integration can be reached through different strategies, and does not
even depend on the presence of a hypothesis to validate (O’Malley & Soyer,
2012).
2 This use of unification does not refer to classical ideals of intertheoretic and
hierarchical disciplinary reductions. Classic philosophical works on unifi-
cation have focused on interfield or disciplinary integration, aiming at a
reduction between neuroscience and psychology (Bickle, 1996; Churchland,
1986; Schaffner, 1993). I here acquire conceptions of modest unifications
(Grantham, 2015), understood as local interconnection between experimen-
tal results (other than concepts and theories) within and across sub-fields of
science.

References
Bechtel, W. (2002). Aligning multiple research techniques in cognitive neurosci-
ence: Why is it important? Philosophy of Science, 69(S3), S48–58. https://doi.
org/10.1086/341767
Bickle, J. (1996). New wave psychophysical reductionism and the method-
ological caveats. Philosophy and Phenomenological Research, 56(1), 57–78.
https://doi.org/ppr199656116
Bogen, J., & Woodward, J. (1988). Saving the phenomena. The Philosophical
Review, 97(3), 303–52. https://doi.org/10.2307/2185445
Cartwright, N. (1991). Replicability, reproducibility, and robustness: Comments
on Harry Collins. History of Political Economy, 23(1), 143–55. https://doi.
org/10.1215/00182702-23-1-143
Cartwright, N. (2006). From metaphysics to method: Comments on manipula-
bility and the causal Markov condition. The British Journal for the Philoso-
phy of Science, 57(1), 197–218.
 Messiness of Cognitive Neuroscience 191
Churchland, P. S. (1986). Neurophilosophy: Toward A Unified Science of the
Mind-Brain. Cambridge: MIT Press.
Cichy, R. M., & Oliva, A. (2020). A M/EEG-fMRI fusion primer: Resolving
human brain responses in space and time. Neuron, 107(5), 772–81. https://
doi.org/10.1016/j.neuron.2020.07.001
Coko, K. (2020). The multiple dimensions of multiple determination. Perspec-
tives on Science, 28(4), 505–541. https://doi.org/10.1162/posc_a_00349
Cole, E. J., Barraclough, N. E., & Enticott, P. G. (2018). Investigating mirror
system (MS) activity in adults with ASD when inferring others’ intentions
using both TMS and EEG. Journal of Autism and Developmental Disorders,
48(7), 2350–67. https://doi.org/10.1007/s10803-018-3492-2
Culp, S. (1994). Defending robustness: The bacterial mesosome as a test case. PSA:
Proceedings of the Biennial Meeting of the Philosophy of Science Association,
1994(1), 46–57. https://doi.org/10.1086/psaprocbienmeetp.1994.1.193010
Dapretto, M., Davies, M. S., Pfeifer, J. H., Scott, A. A., Sigman, M., Bookheimer,
S. Y., & Iacoboni, M. (2006). Understanding emotions in others: Mirror neu-
ron dysfunction in children with autism spectrum disorders. Nature Neuro-
science, 9(1), 28–30. https://doi.org/10.1038/nn1611
Dinstein, I., Thomas, C., Humphreys, K., Minshew, N., Behrmann, M., &
Heeger, D. J. (2010). Normal movement selectivity in autism. Neuron, 66(3),
461–9. https://doi.org/10.1016/j.neuron.2010.03.034
Eberhardt, F., & Scheines, R. (2007). Interventions and causal inference. Philos-
ophy of Science, 74(5), 981–995. https://doi.org/10.1086/525638
Enticott, P. G., Kennedy, H. A., Rinehart, N. J., Bradshaw, J. L., Tonge, B. J.,
Daskalakis, Z. J., & Fitzgerald, P. B. (2013). Interpersonal motor resonance
in autism spectrum disorder: Evidence against a global “mirror system”
deficit. Frontiers in Human Neuroscience, 7, 218. https://doi.org/10.3389/
fnhum.2013.00218
Enticott, P. G., Kennedy, H. A., Rinehart, N. J., Tonge, B. J., Bradshaw, J.
L., Taffe, J. R., Daskalakis, Z. J., & Fitzgerald, P. B. (2012). Mirror neuron
activity associated with social impairments but not age in autism spectrum
disorder. Biological Psychiatry, 71(5), 427–33. https://doi.org/10.1016/j.
biopsych.2011.09.001
Fan, Y.-T., Decety, J., Yang, C.-Y., Liu, J.-L., & Cheng, Y. (2010). Unbroken mir-
ror neurons in autism spectrum disorders. Journal of Child Psychology and
Psychiatry, 51(9), 981–8. https://doi.org/10.1111/j.1469-7610.2010.02269.x
Ference, B. A., Julius, S., Mahajan, N., Levy, P. D., Williams, K. A., & Flack, J. M.
(2014). Clinical effect of naturally random allocation to lower systolic blood
pressure beginning before the development of hypertension. Hypertension,
63(6), 1182–8. https://doi.org/10.1161/HYPERTENSIONAHA.113.02734
Ferrari, & Rizzolatti, G. (2015). New Frontiers in Mirror Neurons Research.
New York: Oxford University Press.
Fishman, I., Keown, C. L., Lincoln, A. J., Pineda, J. A., & Müller, R.-A.
(2014). Atypical cross talk between mentalizing and mirror neuron networks
in autism spectrum disorder. JAMA Psychiatry, 71(7), 751–60. https://doi.
org/10.1001/jamapsychiatry.2014.83
Fuelscher, I., Caeyenberghs, K., Enticott, P. G., Kirkovski, M., Farquharson, S.,
Lum, J., & Hyde, C. (2019). Does fMRI repetition suppression reveal mirror
neuron activity in the human brain? Insights from univariate and multivariate
192 Antonella Tramacere
analysis. The European Journal of Neuroscience, 50(5), 2877–92. https://doi.
org/10.1111/ejn.14370
Gallese, V., Fadiga, L., Fogassi, L., & Rizzolatti, G. (1996). Action recognition
in the premotor cortex. Brain: A Journal of Neurology, 119(Pt 2), 593–609.
Grantham, T. (2015). Conceptualizing the (dis)unity of science. Philosophy of
Science 71(2): 133–55. https://doi.org/10.1086/383008.
Hacking, I. (1983). Representing and Intervening: Introductory Topics in the
Philosophy of Natural Science. Cambridge: Cambridge University Press.
Hausman, D. M., & Woodward, J. (1999). Independence, invariance and the
causal Markov condition. British Journal for the Philosophy of Science,
50(4), 521–83. https://doi.org/10.1093/bjps/50.4.521
Hobson, H. M., & Bishop, D. V. M. (2016). Mu suppression—A good measure
of the human mirror neuron system? Cortex: A Journal Devoted to the Study
of the Nervous System and Behavior, 82, 290–310. https://doi.org/10.1016/j.
cortex.2016.03.019
Hudson, R. (2013). Seeing Things: The Philosophy of Reliable Observation.
New York: Oxford University Press.
Khalil, R., Tindle, R., Boraud, T., Moustafa, A. A., & Karim, A. A. (2018).
Social decision making in autism: On the impact of mirror neurons, motor
control, and imitative behaviors. CNS Neuroscience & Therapeutics, 24(8),
669–76. https://doi.org/10.1111/cns.13001
Kuorikoski, J., & Marchionni, C. (2016). Evidential diversity and the trian-
gulation of phenomena. Philosophy of Science, 83(2), 227–247. https://doi.
org/10.1086/684960
Law, M. R., Morris, J. K., & Wald, N. J. (2009). Use of blood pressure lowering
drugs in the prevention of cardiovascular disease: Meta-analysis of 147 ran-
domised trials in the context of expectations from prospective epidemiologi-
cal studies. BMJ, 338, b1665. https://doi.org/10.1136/bmj.b1665
Lawlor, D. A., Tilling, K., & Davey Smith, G. (2016). Triangulation in aetiolog-
ical epidemiology. International Journal of Epidemiology, 45(6), 1866–86.
https://doi.org/10.1093/ije/dyw314
Likowski, K. U., Mühlberger, A., Gerdes, A. B. M., Wieser, M. J., Pauli, P., &
Weyers, P. (2012). Facial mimicry and the mirror neuron system: Simultane-
ous acquisition of facial electromyography and functional magnetic resonance
imaging. Frontiers in Human Neuroscience, 6, 214. https://doi.org/10.3389/
fnhum.2012.00214
Martineau, J., Andersson, F., Barthélémy, C., Cottier, J.-P., & Destrieux,
C. (2010). Atypical activation of the mirror neuron system during percep-
tion of hand motion in autism. Brain Research, 1320, 168–75. https://doi.
org/10.1016/j.brainres.2010.01.035
Munafò, M. R., & Smith, G. D. (2018). Robust research needs many lines of evidence.
Nature, 553(7689), 399–401. https://doi.org/10.1038/d41586-018-01023-3
Myers, S. M., Johnson, C. P., & American Academy of Pediatrics Council on
Children with Disabilities. (2007). Management of children with autism
spectrum disorders. Pediatrics, 120(5), 1162–82. https://doi.org/10.1542/
peds.2007-2362
O’malley, M. A., & Soyer, O. S. (2012). The roles of integration in molecular
systems biology. Studies in History and Philosophy of Science Part C: Studies
 Messiness of Cognitive Neuroscience 193
in History and Philosophy of Biological and Biomedical Sciences, 43(1),
58–68. https://doi.org/10.1016/j.shpsc.2011.10.006
Oberman, L. M., Hubbard, E. M., McCleery, J. P., Altschuler, E. L., Ram-
achandran, V. S., & Pineda, J. A. (2005). EEG evidence for mirror neuron
dysfunction in autism spectrum disorders. Cognitive Brain Research, 24(2),
190–8. https://doi.org/10.1016/j.cogbrainres.2005.01.014
Odenbaugh, J., & Alexandrova, A. (2011). Buyer beware: Robustness analyses
in economics and biology. Biology and Philosophy, 26(5), 757–71. https://doi.
org/10.1007/s10539-011-9278-y
Orzack, S. H., & Sober, E. (1993). A critical assessment of Levins’s the strategy
of model building in population biology (1966). The Quarterly Review of
Biology, 68(4), 533–546. https://doi.org/10.1086/418301
Palau-Baduell, M., Valls-Santasusana, A., & Salvadó-Salvadó, B. (2011). Au-
tism spectrum disorders and mu rhythm. A new neurophysiological view. Re-
vista De Neurologia, 52(Suppl 1), S141–6.
Ramachandran, V. S., & Oberman, L. M. (2006). Broken mirrors. Scientific
American, 295(5), 62–9.
Rasmussen, N. (2001). Evolving scientific epistemologies and the artifacts of
empirical philosophy of science: A reply concerning mesosomes. Biology and
Philosophy, 16(5), 627–52. https://doi.org/10.1023/A:1012038815107
Raymaekers, R., Wiersema, J. R., & Roeyers, H. (2009). EEG study of the mir-
ror neuron system in children with high functioning autism. Brain Research,
1304, 113–21. https://doi.org/10.1016/j.brainres.2009.09.068
Schaffner, K. F. (1993). Theory structure, reduction, and disciplinary integration
in biology. Biology and Philosophy, 8(3), 319–47. https://doi.org/10.1007/
BF00860432
Soler, L., Trizio, E., Nickles, T., & Wimsatt, W. (Eds.). (2012). Characterizing
the Robustness of Science: After the Practice Turn in Philosophy of Science
(Vol. 292). Dordrecht, Netherlands: Springer Science & Business Media.
Spirtes, P., & Scheines, R. (2004). Causal inference of ambiguous manipula-
tions. Philosophy of Science, 71(5), 833–45. https://doi.org/10.1086/425058
Stegenga, J. (2009). Robustness, discordance, and relevance. Philosophy of Sci-
ence, 76(5), 650–61. https://doi.org/10.1086/605819
Stegenga, J. (2012). Rerum concordia discors: Robustness and discordant mul-
timodal evidence. Rerum Concordia Discors: Robustness and Discordant
Multimodal Evidence, 292, 207–26.
Stegenga, J., & Menon, T. (2017). Robustness and independent evidence. Phi-
losophy of Science, 84(3), 414–35. https://doi.org/10.1086/692141
Trizio, E. (2012). Achieving robustness to confirm controversial hypotheses: A
case study in cell biology. Achieving Robustness to Confirm Controversial
Hypotheses: A Case Study in Cell Biology, 292, 105–20.
Weichwald, S., & Peters, J. (2021). Causality in cognitive neuroscience: Con-
cepts, challenges, and distributional robustness. Journal of Cognitive Neuro-
science, 33(2), 226–47. https://doi.org/10.1162/jocn_a_01623
Weisberg, M. (2006). Robustness analysis. Philosophy of Science, 73(5), 730–
42. https://doi.org/10.1086/518628
Whewell, W., & Butts, R. E. (1968). William Whewell’s Theory of Scientific
Method. Pittsburgh, PA: University of Pittsburgh Press.
194 Antonella Tramacere
Williams, J. H. G., Whiten, A., Suddendorf, T., & Perrett, D. I. (2001). Imita-
tion, mirror neurons and autism. Neuroscience & Biobehavioral Reviews,
25(4), 287–95. https://doi.org/10.1016/S0149-7634(01)00014-8
Wimsatt, W. C. (2012). Robustness, reliability, and overdetermination (1981).
In L. Soler (ed.) Characterizing the Robustness of Science (pp. 61–78). Am-
sterdam: Springer.
Woodward, J. (2006). Some varieties of robustness. Journal of Economic Meth-
odology, 13(2), 219–40. https://doi.org/10.1080/13501780600733376
Yates, L., & Hobson, H. (2020). Continuing to look in the mirror: A review
of neuroscientific evidence for the broken mirror hypothesis, EP-M model
and STORM model of autism spectrum conditions. Autism, 24(8), 1945–59.
https://doi.org/10.1177/1362361320936945
9 Prediction, Explanation,
and the “Toolbox” Problem
Marco J. Nathan

1 The Once and Future Inference

Once upon a time, in a land far away—well, not all that spatially or tem-
porally distant, to tell you the truth, but that’s how fairy tales go—there
was a small flourishing country: the kingdom of Philosophy of Science.
Philosophy of Science was governed by a monarch, King Hempel, who
ruled by the covering-law model, a code affectionately referred to as
“CLM.” King Hempel was a fair, wise, and just leader. He viewed all his
subjects as equals. The country was split in two, oft bickering, counties,
Prediction and Explanation. But the law of the land, the CLM, treated
both as two sides of the same coin. The entire kingdom lived in unity,
simplicity, and harmony. Philosophy of Science, in a nutshell, was thriv-
ing and growing by the year.
Alas, nothing lasts forever, not even in fables. One gloomy day, King
Hempel was found mysteriously murdered as the rooster crowed the
crack of dawn. The assassin had lured the king into a dark alley and
mercilessly gored him with a flagpole, before dissolving like a shadow.
Things were never quite the same in good ole Philosophy of Science.
After a period of mourning, incredulity, and intellectual turmoil,
a new sovereign was finally found. The choice fell upon Sir Thomas
McKuhn, who had been coyly plotting a revolution to be appointed
Caesar. The paradigm shifted, as logic was no longer heralded as the
sole canon of rationality. There was a new sheriff in town, Dr. van der
Quine, who promoted a naturalistic attitude, drawing more attention
to the details of scientific practice. No more CLM. A new law of the
land—the causal-mechanistic model, the notorious “CMM”—was en-
forced by General DeSalmon. The new leaders brought a breath of fresh
air. Mechanism-driven “special” sciences, such as biology, neuroscience,
and psychology, were now viewed as inherently distinct from mathema-
tized, law-based, theory-driven physics. The old CLM was soon forgot-
ten and became a relic of a bygone past, confined to dusty textbooks.
Stories recounting the glory days of King Hempel were sung around the
campfire on chilly nights.

DOI: 10.4324/9781003251392-12
196 Marco J. Nathan
Years lazily went by. People eventually realized that not all that glit-
ters is gold. Incidentally, one also needs a baptismal act and the appro-
priate atomic essence. The new causal-mechanistic ideal, embodied in
the CMM, was far less ecumenical than its predecessor. With the demise
of CLM, explanation and prediction were no longer symmetric. War
soon broke out.
The playing field was hardly leveled. Causal explanation, led by De-
Salmon’s troops, swiftly vanquished the enemy. Prediction, outnumbered
and deemed dispensable, was banished from the kingdom. It searched
for a new home and eventually found one in a neighboring country, the
Republic of Natural Science. Planted in fertile soil, prediction quickly
grew into a strong, independent research field, with a host of new theo-
retical and experimental tools.
Meanwhile, in Philosophy of Science, the first cracks began to appear.
Galvanized by its initial success, van der Quine, DeSalmon, and their fol-
lowers were led to believe that the CMM could take care of all business
at once. And, at first, this appeared to be so. Explanation became the
once and future inference, relegating everything else to the background.
But, as it turns out, the mainstream causal-mechanistic approach, with
its focus on how things are brought about, is poorly equipped to account
for prediction. When forecasts finally started holding their own, this
became painfully clear.
The eventual passing of McKuhn and van der Quine left the kingdom
in a state of disarray. Civil war ran amok. Some citizens of Philosophy
of Science sought refuge in the Realm of Metaphysics, governed by Duke
Lewis Kellogg, an old foe of King Hempel. But the old duchy, once thriv-
ing, was also in a ruinous state and could barely take care of its own.
Others migrated to the larger neighbor, Natural Science, which, in the
meantime, had grown bigger and more powerful. Prediction and expla-
nation made amends. Most of them settled in Natural Science. Others
returned to Philosophy of Science. But there was finally peace and pros-
perity. And everyone lived happily ever after—isn’t this how any fairy
tale worth its salt is supposed to end?
Let’s snap out of fiction and back to reality. Our legend is inspired by
real events with momentous implications, the subject of this essay. Pre-
diction and explanation, once perceived as symmetrical, have been torn
apart and treated independently. The CMM, now dominant in philoso-
phy, captures forecasts indirectly, via causal-mechanistic relations. But
prediction, in many areas of the sciences, is rapidly gaining autonomy.
Purely predictive inferences, not backed up by causal explanation, are
increasingly becoming commonplace, fueled by conceptual and techno-
logical advancements. This raises the question of how predictive and
explanatory tools can work in unison while remaining distinct. This
puzzle, which I dub the “toolbox problem,” has a “black-box solu-
tion.” Prediction and explanation are ultimately grounded in the same
 The “Toolbox” Problem 197
causal network. Yet, predictive and explanatory tools require different
amounts of mechanistic detail, as well as varying degrees of abstraction
and idealization.
Here is the master plan. Section 2 outlines the status quo in contempo-
rary philosophy of science: the methodological hegemony of explanation.
Section 3 illustrates the ongoing predictive turn—the emancipation of
prediction from explanation—with examples from genomics, molecular
medicine, and cognitive neuropsychology. Section 4 raises a philosoph-
ical puzzle, namely, how to characterize the interdependence of predic-
tion and explanation while keeping these inferences distinct. Section 5
explores a “black box” solution to our toolbox problem. Section 6 wraps
up the discussion with general implications and concluding remarks.

2 The Hegemony of Explanation

There is a recent trend across the natural and social sciences, whereby
the epistemic focus is gradually shifting from explanation to prediction.
Meanwhile, philosophy of science has been going in the opposite direc-
tion, openly privileging explanation above any other kind of inference.
The hegemony of explanation is the subject of this section. We’ll shift to
prediction in the following one.
Traditionally, the three chief goals of science were taken to be descrip-
tion, explanation, and prediction. In the authoritative words of Herbert
Feigl (1949, p. 11),

The aims of the pure (empirical) sciences are then essentially the
same throughout the whole field. What the scientists are seeking are
descriptions, explanations, and predictions which are as adequate
and accurate as possible in the given context of research.

Elaborating on this insight, Feigl notes how “the first aim [description],
is basic and indispensable, the second and third [explanation and predic-
tion] (closely related to each other) arise as the most desirable fruits of
scientific labors whenever inquiry rises beyond the mere fact-­gathering
stage” (pp. 10–11). These remarks raise several related questions. Why
should science privilege these aims as paramount? What makes descrip-
tion “basic” and “indispensable”? In what sense are prediction and ex-
planation “closely related to each other”? To address these issues, we
must focus on how these three goals can be met.
Feigl could safely assert that prediction and explanation are closely
related because the covering-law model (CLM), the “official” theory
of explanation of logical positivism and logical empiricism—I will not
separate these movements here—took care of both endeavors simulta-
neously. As gestured in Section 1, Hempel and Oppenheim (1948), the
main proponents of the CLM, viewed prediction and explanation as two
198 Marco J. Nathan
sides of the same coin. Predicting an event is explaining something in
the future, yet to be observed. Conversely, explanation is retrodiction:
prediction of a happening that occurred in the past. Both prediction and
explanation are a matter of logically deriving an event from an expla-
nans consisting of background assumptions and laws of nature. This is
where description comes into the picture. A necessary presupposition
for predicting or explaining an event is an accurate characterization of
initial conditions and underlying nomological regularities. In short, all
three basic epistemic aims of science are brought together by the CLM.
This unified picture of scientific inquiry was bound to change with
the demise of the CLM, due to the discovery of asymmetries of explana-
tion, and the subsequent emergence of the causal-mechanistic paradigm.
Briefly rehashing this familiar story will be helpful to unravel the histor-
ical narrative.
Setting complications aside, the hiccup is straightforward. From the
standpoint of the CLM, prediction and explanation are structurally
analogous. This means that any explanatory inference is thereby pre-
dictive and, conversely, any prediction can be treated as an explana-
tion. Whether this symmetry holds water is questionable. Paraphrasing
the insights of Scriven (1958) and Bromberger (1966), among others,
the length of a shadow cast by a flagpole on a clear day, together
with the position of the sun and the basic laws of light refraction, can
be used to predict the height of the flagpole. But it is doubtful, to say
the least, whether a similar derivation constitutes an explanation of the
height of the pole. Along the same lines, reliable symptoms can be used
to predict a disease, but they don’t explain it.
Unsurprisingly, accounts which supplanted the CLM dropped the
alleged symmetry between explanation and prediction. Unification-
ism, pioneered by Friedman (1974) and Kitcher (1981, 1989), treated
explanation as a unifying endeavor, paying little to no attention to fore-
casting. As well-known as they are, unificationist analyses did not gain
much momentum. Hence, in what follows, I shall focus on the frame-
work that has received the lion’s share of attention and consensus. This
is the causal-mechanistic model, or CMM, for short.
What does it mean to approach explanation from a causal-­mechanistic
perspective? Several influential alternatives have been proposed (Salmon
1984, 1998; Woodward 2003; Strevens 2008). The overarching guid-
ing thought is that explaining an event involves specifying relevant de-
tails about what triggers it, that is, how the event in question is brought
about and how it would vary under different circumstances. In the
words of David Lewis (1986, p. 217, italics omitted), “to explain an
event is to provide some information about its causal history.” With this
core assumption firmly in place, it became clear that grasping complex
causal networks requires looking “under the hood” for mechanistic
underpinnings. A broad analysis of mechanisms and cognate notions
 The “Toolbox” Problem 199
was undertaken, with significant differences in nuance, by Bechtel and
Richardson (1993), Machamer et al. (2000), Glennan (2017), and fellow
advocates of the so-called “new wave of mechanistic philosophy,” which
gained full traction at the turn of the millennium.
As explanation became a matter of causes and mechanisms, what hap-
pened to prediction? It faded into the background. Allow me to elabo-
rate. Prediction was a hallowed topic in classic philosophy of science,
connected to induction and confirmation, long before the CLM. The no-
torious problem of induction—first posed by Hume (2000[1938], 1999
[1748]), rephrased in modern terms by Russell (1912), and revamped by
Goodman (1955)—challenges us to justify our belief that the future will
resemble the past. Expectations about the future are, at heart, a sort of
prediction. The issue of confirmation, as developed by logical empiri-
cists, is, in turn, a de-psychologized form of induction, which focuses
on purely logical, probabilistic relations between theory and evidence.
Two remarks. First, the class of inductive inferences, which encom-
passes all non-deductive forms of reasoning, is broader than the set of
predictions per se. Second, the problem of induction, conceived as a
yet-unresolved philosophical issue, is of marginal significance for prac-
ticing scientists who, pace Popper, routinely take for granted some prin-
ciple of induction. As such, the study of confirmation, traditionally part
of philosophy of science, was relocated and repackaged in the field of
formal epistemology.
Admittedly, a few philosophers of science have devoted attention
to prediction. Cartwright (1983, 2007, 2019) has emphasized the gap
between causal and predictive knowledge. Spirtes et al. (2000) have
sketched a mathematical framework for predictive inferences. In the
philosophy of medicine, Broadbent (2013) discusses prediction in epi-
demiology. And I, for one, have considered the role of predictions in di-
agnosis and prognosis (Nathan 2016). Nevertheless, to date, no one has
undertaken a systematic philosophical analysis of prediction comparable
to what has been provided for causal explanation.
Why did philosophers of science trade in prediction for explanation?
Could it be that forecasting has little to offer of philosophical value?
This seems unlikely. While a clear-cut answer is yet to be found, predic-
tion continues to live in the shadow of its counterpart, in theories of ex-
planation. To be sure, predictive accuracy remains a cardinal theoretical
virtue of many scientific models. Across various fields—from climatol-
ogy to economics, from precision medicine to epidemiology—a simu-
lation able to accurately anticipate the behavior of a complex system is
worth its weight in gold. Yet, I surmise that predictive accuracy is often
considered a more or less direct consequence of a deep understanding of
the system’s causal-mechanistic underpinnings. One predicts by dint of
explanatory power. Description and explanation remain the true gold
standard. This is my best attempt to rationalize the neglect of prediction
200 Marco J. Nathan
in philosophy of science. Otherwise, I would find the current situation
quite inexplicable.
In sum, with the demise of the CLM, the perceived symmetry of pre-
diction and explanation was lost. Despite a handful of valiant, if rela-
tively isolated attempts, philosophers of science swept prediction under
the rug, concentrating quasi-exclusively on description and explanation,
construed in causal-mechanistic fashion. Pointing the spotlight on expla-
nation left prediction without a template of its own. To be sure, formal
epistemologists, Bayesians in particular, focused on the revision of belief
based on new evidence. But this emphasis on confirmation falls short of
a full-blown model of prediction. Thus conceived, prediction could no
longer be viewed as a self-standing goal of science, but a by-product of
explanatory power, a welcome free lunch.
Was neglecting prediction the right move for philosophy of science?
For a while, this stance seemed justified by the attitude of the scientific
community, which also privileged explanation. Yet, things were bound
to change fast. Over the last few decades, many working scientists have
begun couching their epistemic aims in terms of predictive accuracy.
This is the “predictive turn.” It is a trend that has pushed researchers
away from explanation and toward purely predictive models. To drive
the point home, rather than providing an abstract description, the fol-
lowing section briefly introduces three case studies that illustrate this
predictive turn.

3 The Predictive Turn

Section 2 centered on how philosophy of science has come to privilege
explanation over prediction. We now shift emphasis to how the pre-
dictive turn is reshaping the scientific landscape. Preliminary question:
what’s the predictive turn? Following Weiskopf (2021), I employ this
expression to pinpoint a trend across the natural and social sciences,
where the main epistemic focus is gradually shifting from explanation
to prediction. I illustrate this movement with three case studies, where
tools play a fundamental role, at two distinct levels. First, making sense
of the predictive turn presupposes a conceptual “toolbox” that marks
off prediction from explanation. Second, and less metaphorically, these
inferences require a host of technological innovations that provide fine-
grained measurements to ground reliable predictions.
My first example centers on Genome-Wide Association Studies,
“GWAS” for short. Simply put, GWAS involve an examination of
­genome-wide sets of genetic variants across a population of reference
(Vrieze et al. 2012; van der Sijde et al. 2014; Visscher et al. 2017; Harden
2021). The goal is to employ standard genomic instruments and sta-
tistical analysis to determine whether DNA variation can be linked
to phenotypic, behavioral, or psychological traits. Is a specific allele
 The “Toolbox” Problem 201
correlated with, say, greater height or mathematical aptitude? And is this
link explanatory? While GWAS typically look for associations between
­single-nucleotide polymorphisms and human pathologies, the technique
can be applied to genetic properties of any kind, in humans, animals,
or plants. GWAS are now popular in neuroscience and neurology, for
instance, in autism research (Anney et al. 2017; Torrica et al. 2017;
­A lonso-Gonzalez et al. 2019).
What kind of knowledge does a GWAS provide? Ideally, by pinpoint-
ing markers associated with variation in traits within populations, we
should be able to establish a causal link between genetic and phenotypic
factors, thereby explaining the presence of the latter traits. Unfortunately,
this project may be overly optimistic. Starting with Lewontin’s (1974)
pioneering critique of Analysis of Variance (ANOVA)—a widespread
statistical tool—commentators have expressed skepticism regarding the
possibility of using GWAS to identify the causes of individual outcomes,
thereby causally explaining them, in full or in part. Nevertheless, GWAS
have not lost their intended purpose. They remain valuable instruments
of prediction. Allow me to elaborate.
Lewontin’s original insight has been developed along various lines.
Some modestly claim that the link between phenotypes and genes iden-
tified by GWAS markers is not a “paradigmatic” cause (Bourrat 2020).
Others advance the stronger point that the relation in question cannot be
truly causal because the association may be completely spurious (Rich-
ardson and Jones 2019), resulting from population structure (Sul et al.
2018), or relying on a pathway that operates via a genotype × environ-
ment covariance (Lynch and Bourrat 2017). Adjudicating between these
options lies beyond present concerns. The relevant point, for our own
purposes, is that a properly conducted GWAS unveils a correlation be-
tween genotype and phenotype. And a time-worn philosophical adage
reminds us that such associations don’t necessarily imply causation.
Importantly, the heart of the matter is not merely the truism that genes
alone do not causally explain any trait—they never do. The real concern
is that the causal relation between genes and phenotypes is often way
too indirect to provide an illuminating explanation of individual traits.
Hence, when GWAS provide little or no valuable causal-mechanistic in-
formation, genetic variation does not explain the trait—or, more mod-
estly, it does not explain it well. Still, genomic analysis comes in handy
to predict such characteristics. The GWAS framework, in short, is best
conceived not as an explanatory tool but, rather, as an instrument of
prediction. Witness the predictive turn in full force: the forecasting value
of a genomic tool is independent of its explanatory power.
Our second case study is an extension of our reflections on GWAS. As
noted, GWAS purport to identify genetic markers associated with phe-
notypic or behavioral profiles. But genetic markers are a special instance
of a much broader phenomenon: biomarkers. What is a biomarker?
202 Marco J. Nathan
The NIH Biomarkers Definitions Working Group (2001) concisely
defines a biomarker as “a characteristic that is objectively measured and
evaluated as an indicator of normal biological processes, pathogenic
processes, or pharmacological responses to a therapeutic intervention.”
To paraphrase, a biomarker is a measurable indicator, on any spatial or
temporal scale, that can be employed experimentally to evaluate charac-
teristics about its source, regardless of whether such source is a healthy
process or a pathological state.
A few examples should help drive the point home. Since 97.7°F–99.5°F
body temperature is a normal physiological state for a human being, an
approximate value within this range is a biomarker of good health. Sim-
ilarly, mutated MYH7 increases the probability of hypertrophic cardi-
opathy, and high prostate-specific antigen (PSA) levels are an indicator
of prostate cancer. All three states are biomarkers. From this standpoint,
GWAS set out to find genetic biomarkers. Unsurprisingly, the forecast-
ing value of GWAS extends, more generally, to biomarkers of all sorts.
A biomarker, simply put, grounds an informed prediction that a cer-
tain condition is or will be present. High PSA levels, under appropriate
circumstances, indicate that prostate cancer is the most likely diagnosis
for a patient. Yet, PSA levels do not causally explain the presence of the
tumorous mass. In general, a biomarker need not, and often does not,
causally explain the associated condition. How can a marker offer pre-
dictive value without any causal-explanatory insight?
I should make it very clear that one and the same entity can be both
a cause and a biomarker of a certain condition. Mutated MYH7 con-
tributes to hypertrophic cardiopathy and allows us to predict it too.
Similarly, the unstable expansion of a CAG repeat within the coding
region of the huntingtin gene is an indicator—and, unfortunately, a rare
instance of a sure-fire marker—of a devastating neurodegenerative dis-
order, Huntington’s Disease, that will develop when the subject reaches
their 30s or 40s. Now, it would be a mistake to pinpoint the genetic
mutation as the one and only cause of the pathology, as the network in
question turns out to be frustratingly complex. Still, the mutation is a
recognized cause of the neurological disorder. Having said this, such a
condition does not hold generally. A biomarker need not play any causal
role. To wit, as far as we know, there is no causal connection between
high PSA levels and prostate cancer. And, even when a biomarker does
play a causal function, its role as a cause is distinct from its role as a
biomarker. A biomarker need only predict, whereas a cause must also
contribute to an explanation. In short, biomarkers, qua indicators, need
not provide any substantial causal or mechanistic knowledge, without
affecting their status as biomarkers.
Recent philosophical research at the intersection of causation and
medicine has focused on the role of biomarkers in studying the etiology
of disease (Russo and Williamson 2012; Illari and Russo 2016; Ghiara
 The “Toolbox” Problem 203
and Russo 2019). While the tension between prediction and explanation
has not gone unnoticed, once again, it is explanation that has taken the
center of the stage.

On the one hand, biomarkers are not entities, things to which we

can attribute some causal power, in the same sense as HPV virus has
the power to initiate the onset of cervical cancer. Instead, biomark-
ers are clues, indicators, markers to detect in order to reconstruct
the missing link. On the other hand, and related to the previous
point, we need to say in which sense, if any, these continuous links,
or processes, between exposure and disease are causal. This is all
the more important because we seek to link heterogeneous levels as
the macro- and the micro-environment.
(Vineis et al. 2017, p. 4)

In sum, biomarkers provide no causal-mechanistic information about a

system. As such, they don’t suggest therapies or interventions. Biomark-
ers, together with all the technological and conceptual instruments re-
quired for their detection and analysis, are primarily tools of prediction,
not explanation.
My third and final case study illustrating the predictive turn in action
comes from the bourgeoning field of neuroimaging. Simply put, neuro-
imaging encompasses

a set of new technologies [that] has given us the ability to study how
the human brain works in greater detail than ever before. These
tools are known as neuroimaging methods, because they allow us
to create images of the human brain that show us what it is made of
(which we refer to as its structure) and what it is doing (which we
refer to as its function).
(Poldrack 2018, p. 1)

Arguably, the most revolutionary tool of all has been magnetic reso-
nance imaging (MRI). Now, MRI comes in various forms, and different
kinds of MRI scans measure specific aspects of the brain. Structural
MRI captures aspects of the makeup of neural tissue, such as how much
water or fat is present. To determine what the brain is doing, we need
functional MRI, or “fMRI,” which, roughly speaking, detects the
shadow of brain activity—engagement in a cognitive task—through its
effects on the amount of oxygen in the blood.
Let’s focus on so-called reverse inference, which basically allows one
to infer the engagement of a particular mental process from specific ac-
tivation patterns or locations in the brain, as opposed to a “forward
inference,” from mental states to brain processes. Here I provide an
admittedly simplified reconstruction of the underlying argument—for
204 Marco J. Nathan
a more detailed overview of reverse inference and its limitations, see
Nathan and Del Pinal (2017).
Consider a psychological generalization describing how humans typi-
cally behave while undertaking a certain cognitive task. For instance, sup-
pose that we perform better on a test when we listen to soothing music.
Presumably, there will be various cognitive processes that could explain
this tendency. For instance, music may increase our concentration or, al-
ternatively, listening to music may block negative emotions that interfere
with cognitive performance. How do we decide which option, if either, is
correct? Reverse inference shows how to employ neuroscientific evidence
to answer this question. The trick is to find some association between the
competing cognitive processes and some underlying areas or patterns of
neural activation. For instance, being focused may be associated with a
certain kind of brain activity x, whereas blocking emotions could be as-
sociated with a different neural pattern y. This being so, we can conduct
a neuroimaging scan while subjects are performing the test. If the scan
detects type-x activity and not type-y activity, this provides evidence that
what explains enhanced performance is focus, as opposed to lack of emo-
tions. To be sure, this constitutes an oversimplification, because most brain
regions and patterns of brain activity will be associated with several cogni-
tive tasks—the “lack of selectivity objection.” Various strategies have been
offered to rule out alternative explanations (Del Pinal and Nathan 2013;
Hutzler 2014; Machery 2014). We need not worry about details here.
Once again, the crucial point, for our present purposes, is that the as-
sociation of a cognitive subprocess with a neural region or pattern need
not be a causal one. Indeed, while it is possible that the patterns detected
via fMRI are what produce the mental state, this is the exception rather
than the rule. All that is required for the argument to go through is an
associative link, not a bona fide reduction. These associative bridge laws
may be conceived as probabilistic and context-sensitive relations that
do not identify their relata, either at the type-level or at the token-level
(Nathan and Del Pinal 2016). In short, the association between cog-
nitive state and neural pattern can be used to predict (“reverse-infer”)
the former from the latter or predict (“forward infer”) the latter from
the former. But neither inference is typically explanatory. To emphasize,
such inferences may well be explanatory, if the neural processes in ques-
tion are causally responsible for the mental state. Nevertheless—and this
is the take-home message—they need not be.
In conclusion, reverse and forward inference, alongside the required
fMRI technology, are, at heart, instruments of prediction, not explana-
tion. In this respect, they are like GWAS and biomarkers. The moral, of
course, isn’t that description and explanation are irrelevant in science.
What the predictive turn shows is that forecasting may become a goal in
and of itself, in the absence of explanation. What philosophical conse-
quences might this have?
 The “Toolbox” Problem 205
4 The Toolbox Problem
The previous section outlined three scientific tools, whose development
required a host of technological and conceptual advancements: GWAS,
biomarkers, and neuroimaging. All three can be conceived primarily as
instruments of prediction, as opposed to explanation. GWAS need not
uncover any causal link between genotype and phenotype. As such, they
frequently do not explain the presence of the trait under study, without
tainting their forecasting value. Biomarkers can increase our confidence
in the occurrence of a specific physiological process or reaction, while
not offering any insight as to what produces the state at hand. fMRI is
grounded in a series of associative bridge laws that reverse-infer the en-
gagement of a cognitive process based on neural activation or, vice versa,
forward-infer neural activity based on cognitive engagement. Again, all
of this can be effectively done without shedding any light on their causal
interplay and, thereby, not warranting any causal-mechanistic explana-
tion. These are clear illustrations of the predictive turn.
Some readers may find none of this strange, concerning, or remotely
controversial. As mentioned at the outset, prediction is a traditional goal
of science. Why should one be surprised by the emergence of predictive
tools?
The development of predictive tools per se is hardly shocking. The vex-
ing issue, from a philosophical perspective, is squaring the presence of
these instruments—characterized from a purely predictive, as opposed
to an explanatory standpoint—with the mainstream causal-mechanistic
outlook. I dub this the “toolbox problem” because it challenges us to
spell out a framework that captures how prediction and explanation can
mutually inform each other without reducing to one and the same. As
I will argue in due course, prediction is, indeed, a form of “settling for
less,” a cheaper version of explanation. But that does not undermine its
pivotal role.
In the heyday of logical positivism, the toolbox problem did not oc-
cur. The ruling theory of explanation, the CLM, treated prediction and
explanation as two sides of the same coin. “[The same logical] schema
underlies both explanation and prediction; only the knowledge situation
is different. In explanation, the fact Qa [the explanandum] is already
known. (…) In prediction, Qa is a fact not yet known” (Carnap 1966,
p. 17). Hence, a predictive inference is always explanatory and, vice
versa, an explanation also has predictive power. The statistical tech-
niques involved in a well-conducted GWAS provide an informed guess
as to whether an individual with genotype g will display phenotypic trait
t. If this prediction is borne out, we thereby have an explanation of t.
The same can be said about biomarkers and neuroimaging. If explana-
tion and prediction are logically equivalent, predictive and explanatory
oomph goes hand in hand.
206 Marco J. Nathan
When the CLM was displaced by the CMM, a wedge was drawn be-
tween prediction and explanation, which were no longer perceived as
symmetrical. So, how are these inferences related? Clearly, they support
each other. But how? Explanation is now a matter of causal-mechanistic
detail. If prediction is not logically equivalent to explanation, what is
the logical structure of a predictive inference? How does predictive value
enhance explanatory power and, vice versa, how does explanation con-
tribute to a prediction? This is the toolbox problem.
It is important not to conflate what I call the “toolbox problem” with
what has been dubbed the “causal fallacy.” The causal fallacy, as Broad-
bent (2013, p. 83) presents it, is the mistake of believing that all expla-
nations have predictive power.

Proving that X causes Y does not license a good prediction that

removing X will lead to a corresponding reduction in the incidence
of Y. To put it another way, just because X causes Y, that does not
mean that removing X is a sufficient means to removing Y. Just be-
cause the frying pan is making you hot does not mean that jumping
out of it is a good way to cool down.

Is the causal fallacy really a fallacy? Does all explanation enable pre-
diction? This remains an interesting open question. Still, this issue is
tangential to my toolbox problem, which is distinct. It involves explain-
ing how we can have prediction without explanation. How do we get
correlational structures underwritten by causal arrows that point in the
wrong direction?
As Pearl and Mackenzie (2018, p. 30) quip, “Good predictions need
not have good explanations. The owl can be a good hunter without
understanding why the rat always goes from point A to point B.” While
this seems unassailable, by focusing exclusively on causal-mechanistic
relations, one misses the point of the predictive turn. Attentive read-
ers will surely note that there are various ways of answering the tool-
box problem. The remainder of this section explores some intuitive
options and argues that none of them is ultimately successful. Then,
the following section will offer a different “black box” solution to our
conundrum.
A first reaction is the timeworn, ostrich-inspired tendency to stick
one’s head into the sand. From this standpoint, admittedly not especially
popular, the toolbox problem is not something we should lose sleep
over. Sure, the objection runs, prediction has a role to play in science.
And there may well be concepts that predict without explaining. Yet,
these are the exception rather than the rule. The issue is not widespread
enough to make it worth our while. Better to focus our attention on
what really matters for science, namely, mechanistic description and
causal explanation.
 The “Toolbox” Problem 207
This initial riposte strikes me as myopic. Our three case studies are
but the tip of the iceberg. Predictive inferences of this ilk are ubiquitous.
From computer science to climatology, from evolutionary theory to psy-
chology, prediction sans explanation is not the exception to the norm.
Furthermore, my examples are hardly marginal. GWAS are a central
tool in genomics. Biomarkers are the holy grail of precision medicine.
And neuroimaging, for better or for worse, has come to dominate cogni-
tive psychology. Hence, dismissing the issue is no more effective than the
cognate strategy of sticking your head in the sand in hope that a preda-
tor will chase itself away. Better to put our focus elsewhere.
A second, more promising response is to insist that the alleged struc-
tural discrepancy between explanation and prediction is merely illusory.
The CMM, properly understood and contextualized, is perfectly capa-
ble of taking care of both explanation and prediction. Both inferences
rely on causal-mechanistic knowledge.
Consider, first, biomarkers and associated conditions, such as the link
between high PSA levels and prostate cancer. To the best of my knowl-
edge, high PSA is not among the causes of cancer. Still, there must be
something that underlies and explains the reliable correlation between
biomarker and pathology. Obviously, it cannot be a cosmic coincidence
that high PSA is consistently associated with cancer. There must be some
causal connection.
A similar insight applies to neuroimaging. Take, for instance, a com-
puter classifier able to reliably reverse-infer the engagement of cogni-
tive state P from some location or pattern of neural activation N. Now,
presumably, P is not completely causally independent of the underly-
ing neural activity N. Clearly, it would be fallacious to infer from this
correlation, without further corroborating evidence, that P and N are
type-identical, or that N directly causes P. Still, for the association be-
tween P and N to be robust, and therefore useful in the context of a
reverse inference, there must be some causal connection between the
neural substrate and the mental superstrate.
In short, the argument runs, both examples point to the same con-
clusion. Prediction, just like explanation, is a thoroughly causal-­
mechanistic endeavor. The CMM can thus take care of both. Toolbox
problem solved.
There is something about this second rejoinder that strikes me as cor-
rect. Prediction, lest it becomes hocus pocus, must exploit mechanis-
tic links. My own solution, outlined in Section 5, will indeed posit a
causal connection between predictor and predicted state. Having said
this, as just presented, this reply is too crude. No one should dispute
the existence of a mechanistic pathway linking high PSA with cancer
or mental states with neural activity. Yet, such a connection does not
explain the pathology or mental state. In other words, the causal con-
nection between predictor and prediction is what ensures the robustness
208 Marco J. Nathan
of the inference (for a discussion of robustness in neuroscience, see
Tramacere, this volume). Still—and this is the crucial point—such
causal-­mechanistic connection need not explain the relevant outcome,
that is, the explanandum.
Moving on, a third reaction to the toolbox problem insists that pre-
diction and explanation are fundamentally distinct. Explanation follows
the CMM. Prediction does not. Thus, perhaps, the solution consists in
retaining the CMM for explanation, while developing an altogether dif-
ferent approach to prediction. From this perspective, two mistakes need
mending. First, the CLM wrongly assumed that prediction and explana-
tion are two sides of the same coin. The CMM rectified this by drawing
the two inferences apart. Second, the CMM fails to develop an indepen-
dent account of prediction. That’s the missing ingredient.
The predictable follow-up becomes: which model of prediction will
do the trick? Many alternatives are on the table, from statistical Bayes-
ian networks approaches, to a purely logical “covering-law model of
prediction” that drops any pretense of explaining, to hybrid approaches
that purport to combine both insights. All three avenues seem worth
pursuing, together with potentially others. Still, I’m skeptical about the
prospects of rendering prediction completely independently of explana-
tion. Why? For one thing, it offends the aesthetic taste of those of us who
long for some sort of unified scientific methodology. Second, and more
to the point, it leaves something out. Prediction and explanation do have
something in common, at least intuitively. Making a correct prediction
may open the door, even if a slight crack, to providing a corresponding
explanation. Conversely, explaining something is an effective strategy to
draw an effective prediction. To be sure, we are now back to square one.
We still need to address the toolbox problem by characterizing the na-
ture of this connection. True. But throwing in the towel and developing
an account of prediction completely severed from explanation throws
the baby out with the bath water. I’ll advance a better proposal in the
following section. First, however, we’ve got one final option to explore.
A fourth, and final rejoinder would be to return to the CLM. The
underlying idea is that we already have a theory that provides a uni-
fied treatment of prediction and explanation, namely, the covering-law
model. Perhaps the mistake was wandering away from Hempel and Op-
penheim’s hallowed approach.
This route seems to me to get us out of the frying pan and into the fire.
As noted all along, the CMM comes with costs and difficulties of its own.
Nevertheless, walking away from the deductive-nomological approach
was the right move. The notorious asymmetries are there to remind us
that predicting and explaining are not one and the same logical infer-
ence. This point was reinforced by our illustrations of the predictive turn.
Much explanation requires causation; prediction often does not, settling
for correlation. Rehashing the CLM, in other words, is a nostalgic way of
 The “Toolbox” Problem 209
reliving the glory days, a past that was abandoned for good reason. We’re
better off looking to the future, whatever that may be.
Let’s take a quick breather. This section introduced the toolbox prob-
lem, the issue of showing how prediction and explanation can, intui-
tively, mutually inform and support each other, without reducing one
to the other. There are many ways of addressing this challenge. We ex-
plored four possible rejoinders, none of which, I argued, hits the bull-
seye. Section 5 develops a different, more effective line of argument, a
“black-box” solution to our toolbox problem.

5 A Black-Box Solution
The previous section introduced the toolbox problem: the task of char-
acterizing the interplay between prediction and explanation. I explored
four rejoinders. First, one can deny that prediction is all that central to
science. Second, prediction, like explanation, could be accounted for at a
­causal-mechanistic level. Third, one may develop a self-standing model
that treats prediction independently of explanation. Fourth, the two infer-
ences might be reunited by returning to the old CLM. None of these strat-
egies, I maintained, is ultimately successful. At the same time, there is no
need to throw in the towel. This section sketches a “black-box” solution
to the toolbox problem and argues that it provides a viable path forward.
Let’s begin by emphasizing, once again, that prediction and expla-
nation are not structurally identical. Not all successful forecasts are
explanatory. Symptoms diagnose diseases without explaining them. Ex-
planation typically requires causation. Prediction, in contrast, only pre-
supposes correlation, that is, robust, reliable association.
Having noted this, the kind of association underlying predictive in-
ference calls for further elucidation. Even if marker B does not explain
condition C, there must be something responsible for the link between
B and C. Symptoms may not explain diseases, but what makes the cor-
relation between symptoms and diseases reliable, robust? In the case of
fMRI, specific patterns of neural activation N need not be type- or even
token-identical to the associated cognitive state M. (To be clear, M will be
token-identical to some neural state. But such neural state need not and,
typically, will not be the same neural state N that we use in our forward
or reverse inferences.) Still, there must be a causal mechanism which ex-
plains why engagement in C is reliably accompanied by activation in N.
In short, predicting C from B does rely on a causal-mechanistic story. The
point is that, as discussed at length, the causal-mechanistic relation in
question may well not correspond to a plausible explanation of C from B.
These considerations reveal that, while predictive and explanatory
inferences do not have the same structure, they are closely related and
inform each other. The question becomes what exactly makes them dif-
ferent. And the answer cannot simply lie in causation vs. correlation
210 Marco J. Nathan
since, as said, robust correlations are grounded in causal connections
of sorts. My proposal is to draw the relevant distinction in the amount
and kind of detail provided. To appreciate the point, it will be useful to
separate cases where predictions and explanations coincide from cases
where they do not.
Consider, first, a scenario where the same trait, property, or condition
is both predictive and explanatory of a certain state. Here, marker B is
a cause C of effect E. For instance, the modified huntingtin gene is both
a sure-fire predictor that Huntington’s Disease (HD) will occur when
the patient comes of age and is also one of the driving factors in neural
degeneration.

What is the difference between using the mutated gene to predict HD

and using it to explain it? When we explain HD, specifying some
causal-mechanistic details of how the gene induces the pathology will
normally deepen and enhance our explanation. To be sure, whether ad-
ditional information is invariably better, or we eventually reach a thresh-
old where all relevant data has been specified is a hotly debated topic in
the specialized literature (Strevens 2016; Craver and Kaplan 2020). Yet,
without getting bogged down in tangential complications, understanding
and describing how the genetic dysfunction triggers its devastating effects
will provide a firmer, more effective explanation of HD. In contrast, these
details will not enhance the prediction of HD. An appropriately chosen
genetic marker will tell us that the condition will appear, or how likely it
is to appear, without telling us why. This why-question, interesting as it
surely is, pertains to the purview of explanation, not prediction. Once we
establish that condition E has an x% probability of occurring, motivating
this likelihood is not going to improve the forecast in question.
So, for all intents and purposes, we can treat prediction as a mechanism
sketch (Craver and Darden 2013) or a black-boxed explanation (Nathan
2021), that is, a placeholder for the causal narrative linking B to E where
most, or even all the relevant causal-mechanistic details have been omitted.
Add these causal details back into the model—replace the black box with a
“grey” or “transparent” one—and the prediction will, ipso facto, turn into
an explanation of our explanandum event or data. In short, in this first type
of situation, a prediction is just a black-boxed explanation, exactly how
some advocates of the CMM suggested, more or less explicitly, all along.
Now, consider the alternative scenario, where a marker is predictive but
not explanatory. This could happen for various reasons. I consider two.
In a second possible case, effect E is solely or partially causally respon-
sible for biomarker B. In such circumstances, we can still use B to predict
E, as the causal link will bolster the association. Yet, as with the flagpole
 The “Toolbox” Problem 211
and the shadow, B does not explain E. If anything, it is E that explains
B. Scriven, Bromberger, and their followers were correct, after all, in
their diagnosis that prediction is symmetric while explanation is not.
Explanation tends to follow the arrow of causation; prediction does not.

A third possibility occurs when E1 and E 2 are both effects of a common

cause C. Here, E1 can be used to predict E 2 , just like E 2 could be used to
predict E1. However, E1 does not explain the occurrence of E 2 , just like
E 2 does not explain E1. Both E1 and E 2 are explained by their common
cause C. This third scenario is illustrated by the case of high PSA levels
and prostate cancer or, to exhume an old example, the relation between
stained teeth and lung cancer, both of which are the product of a com-
mon cause: smoking.

In these last two situations, the marker is predictive but not explanatory.
Note, once again, that there is an underlying causal-mechanistic connec-
tion that ensures the robustness of the association, enabling the prediction.
But, unlike our first scenario, adding etiological details will not turn the
prediction into an explanation—at least, not an explanation of our expla-
nandum. The shadow predicts the flagpole without explaining it. High
PSA levels predict prostate cancer but don’t explain it. Both effects are
explained by a common cause: genetic predisposition perhaps, although,
unfortunately, not much is currently known about this pathology.
Three brief remarks. First, why not use the common cause, as opposed
to the correlated marker to predict a condition such as cancer? More gen-
erally, could we not use causal explanations to predict effects? Of course,
we could. The problem is that causal-mechanistic stories are often quite
complex. To wit, not much is currently known about the network of causes
of, say, prostate cancer and Huntington’s Disease. Thus, having a simple
and easy to test marker comes in quite handy for diagnostic purposes.
Second, and relatedly, can a predictive link be explained? Could the
causal-mechanistic underpinnings of the relation between PSA levels
and cancer not be unveiled, transforming the forecast into a full-blown
explanation? Yes, they could. But should they? What exactly is there
to be gained? Allow me to elaborate. In some cases, we may want to
explain an association to make sure that it is robust enough. When the
212 Marco J. Nathan
link between PSA levels and cancer was first observed, researchers could
legitimately question the reliability of the marker in detecting a danger-
ous condition such as prostate cancer. And shedding light on why these
conditions tend to correlate may help mitigating legitimate doubts. At
the same time, once the robustness of the marker has been established
beyond reasonable doubt, an explanation of the marker may be superflu-
ous. What we want to explain is cancer, not PSA levels. When we have
substantive evidence that PSA predicts cancer, the explanation of PSA is
not all that important—a black box will suffice. It seems better to focus
on diagnosing cancer itself and to learn how to cure it.
Third, and finally, what is the relation between prediction and expla-
nation? The CLM treated predictive and explanatory inferences as two
sides of the same coin. The CMM drew them apart. Explanation requires
specifying the causes underlying a specific phenomenon; prediction does
not. These two perspectives can be reconciled by stressing that predic-
tion and explanation pertain to different levels of description, depending
on how much mechanistic information is provided. These considerations
shed light on ongoing debates concerning the nature of so-called “mathe-
matical” and “causal-dynamical” explanatory models, which include lim-
ited causal-mechanistic information (Ross 2015; Green and Jones 2016).
Now, whether these explanatory models should be understood as mecha-
nism sketches or as a different type of explanation, or even if they are best
conceived from a causal perspective (Woodward 2018; Ross 2021) are
questions that I shall not address here. However, our discussion stresses
that there is a spectrum of causal models, which vary depending on the
amount of abstraction and idealization. On one end of the spectrum, we
have detailed “glass box” explanations, where mechanistic components
are presented in detail, and the causal link should be as direct as possi-
ble. On the opposite side of the spectrum, we have black-boxed predic-
tions, where causal information can be omitted, and the causal link can
be much more indirect. In between, there is a continuum of causal models
of various sorts. To be sure, spelling out the epistemic norms from these
predictive and explanatory inferences, governing the amount of mecha-
nistic detail and the degree of abstraction is a worthwhile project. Yet,
this ambitious endeavor lies beyond the scope of the present work.

6 The Moral of the Story

Time to take stock. This essay began, in jest, with a fairy tale. While the
tone was humorous, the goal was serious: recounting some momentous
episodes in the last 50 years of philosophy of science. While prediction
has been neglected in this area of philosophy, some branches of science
have witnessed the emergence of a “predictive turn.” Admittedly, in
fields such as cellular and molecular neuroscience, mechanistic expla-
nation may continue to be the gold standard, at least for now, whereas
 The “Toolbox” Problem 213
prediction remains ancillary to understanding. However, GWAS, bio-
markers, and fMRI-based inferences show that the situation is quite
different in molecular medicine and cognitive neuropsychology, where
prediction is rapidly becoming just as important, if not more central
than explanation, description, and understanding. And these examples
are merely the tip of the iceberg. Hence, if philosophy wants to keep
up with methodological advances across the sciences, not merely in se-
lected areas, prediction can no longer be relegated to the sidelines. We
should devote to it the attention it deserves. This is the only viable strat-
egy for appreciating not just the success, but also the limits of scientific
explanation.
The second part of the essay discussed various lines of response to
what I dubbed the “toolbox problem,” that is, the task of providing a
unified framework for prediction and explanation, without reducing one
to the other. Many intuitive ways of addressing the challenge end up
missing the mark. My proposed solution boils down to a “black box”
approach. Both prediction and explanation, I suggested, describe the
same kinds of entities, laws, and mechanisms. But they do so at very
different levels of abstraction and idealization. The structural backbone
of these two inferences is the same. Yet, prediction requires not as much
causal-mechanistic structure as explanation, and these links may be
more or less direct. Effective prediction need not specify the nature of
the connection between marker and condition.
Epistemically speaking, explanation is much more demanding, in the
sense that it requires a much richer, more detailed characterization of the
underlying mechanistic substrate. Still, it would be a mistake to scorn-
fully dismiss the focus on prediction as “lazy science.” With toy exam-
ples, it is easy to brush off the role of correlation. Stained teeth won’t
be all that useful in the study of cancer, since we know that smoking is
the relevant causal variable. But one of the many lessons we learned in
the COVID-19 pandemic, is that, when frantically trying to block the
spreading of a virus, coughing and sneezing can be valuable predictors.
Once again, this is not to downplay the significance of explanation. The
point is that our best experimental practice requires a combination of
various kinds of inferences.
This black-box strategy has several virtues. In closing, I briefly men-
tion two. For one thing, it provides a joint treatment of prediction and
explanation, showcasing how these inferences are related albeit distinct.
The CLM’s mistake was assuming that to explain an event is thereby
to predict it. I’ve argued that there is a substantial discrepancy between
these two illocutionary acts, a difference which can be cashed out in
terms of the amount of detail required for success. Prediction requires
very minimal amounts of mechanistic detail. Explanation, in this re-
spect, is much more demanding. Yet, the two inferences are not com-
pletely independent. A successful prediction can be an invaluable guide
214 Marco J. Nathan
to an explanation and our model shows why: it pinpoints where addi-
tional causal-mechanistic structure may be required.
A second advantage of the view outlined here is that it rationalizes the
predictive turn, that is, the increasing prominence and growth of instru-
ments of prediction. From our present standpoint, it makes perfect sense
why forecasting is often considered the holy grail of science. Prediction is
much cheaper than explanation. If all we want to do is figure out whether
a specific state is present, a predictive marker is more than enough. Deeper
understanding is often desirable, but it is much more expensive and te-
dious to achieve. From this standpoint, prediction is the first step toward
explanation or the final step in some other pursuit. Whether or not we
want to take the additional steps demanded by explanation will depend
on several contextual features: cost, interest, difficulty, und so weiter.
Is the predictive turn something to cherish or bemoan? This remains
an open question. Some authors view the current focus on prediction as
a way of settling for less. Others see explanation as an expensive lux-
ury, distracting from more urgent and pressing issues—like curing dis-
eases and planning for climate shifts—where prediction is often enough.
While these considerations raise deep and important quandaries, this
is not the appropriate place to take an explicit position. The significant
conclusion drawn in this essay is that, regardless of our stance with re-
spect to prediction, it is high time philosophy gave rightful attention to
this prominent kind of inference.
My fairy tale had a happy ending. Prediction and explanation live
happily ever after, in peace and harmony. Strictly speaking, this is not a
caricatural portrait of the current situation, at least not yet. It’s up to us
to make sure that dreams turn to reality. But much remains to be done.

Acknowledgments
The author is grateful to Bill Anderson, Janella Baxter, John Bickle,
Carl Craver, Guie Del Pinal, Mallory Hrehor, and Antonella Tramacere
for constructive comments on various versions of this essay. An early
draft was presented on September 28, 2019, at the Tool Development in
Experimental Neuroscience workshop in Pensacola Beach, Florida. The
audience provided valuable feedback.

References
Alonso-Gonzalez, A. Calaza, C. Rodriguez-Fontenla, and A. Carracedo (2019).
“Novel Gene-Based Analysis of ASD GWAS: Insight into the Biological Role
of Associated Genes.” Frontiers in Genetics 10, 733.
Anney, R. J. et al. (2017). “Meta-Analysis of GWAS of Over 16,000 In-
dividuals with Autism Spectrum Disorder Highlights a Novel Locus at
10q24.32 and a Significant Overlap with Schizophrenia.” Molecular
 The “Toolbox” Problem 215
Autism 8(21). https://molecularautism.biomedcentral.com/articles/10.1186/
s13229-017-0137-9#author-information
Bechtel, W. and R. Richardson (1993). Discovering Complexity: Decompo-
sition and Localization as Strategies in Scientific Research. Princeton, NJ:
Princeton University Press.
Bourrat, P. (2020). “Causation and Single Nucleotide Polymorphism Heritabil-
ity.” Philosophy of Science 87(5), 1075–83.
Broadbent, A. (2013). Philosophy of Epidemiology. New York: Palgrave
MacMillan.
Bromberger, S. (1966). “Why-Questions.” In R. Colodny (Ed.), Mind and Cos-
mos: Essays in Contemporary Science and Philosophy, pp. 86–111. Pitts-
burgh, PA: University of Pittsburgh Press.
Carnap, R. (1966). An Introduction to the Philosophy of Science. Mineola, NY:
Dover.
Cartwright, N. (1983). How the Laws of Physics Lie. Oxford: Clarendon.
Cartwright, N. (2007). Hunting Causes and Using Them: Approaches in Phi-
losophy and Economics. Cambridge: Cambridge University Press.
Cartwright, N. (2019). Nature, the Artful Modeler. Chicago, IL: Open Court.
Craver, C. F. and L. Darden (2013). In Search of Mechanisms. Discoveries
Across the Life Sciences. Chicago, IL: University of Chicago Press.
Craver, C. F. and D. M. Kaplan (2020). “Are More Details Better? On the
Norms of Completeness for Mechanistic Explanation.” British Journal for
the Philosophy of Science 71(1), 287–319.
Del Pinal, G. and M. J. Nathan (2013). “There and Up Again: On the Uses and
Misuses of Neuroimaging in Psychology.” Cognitive Neuropsychology 30(4),
233–52.
Feigl, H. (1949). “The Scientific Outlook: Naturalism and Humanism.” Amer-
ican Quarterly 1(2), 135–48.
Friedman, M. (1974). “Explanation and Scientific Understanding.” The Journal
of Philosophy 71(1), 5–19.
Ghiara, V. and F. Russo (2019). “Reconstructing the Mixed Mechanisms of
Health: The Role of Bio- and Socio-Markers.” Longitudinal and Life Course
Studies 10, 7–25.
Glennan, S. (2017). The New Mechanical Philosophy. Oxford: Oxford Univer-
sity Press.
Goodman, N. (1955). Fact, Fiction, and Forecast. Cambridge, MA: Harvard
University Press.
Harden, K. P. (2021). “‘Reports of My Death Were Greatly Exaggerated’: Be-
havior Genetics in the Postgenomic Era.”’ Annual Review of Psychology 72,
37–60.
Hempel, C. G. and P. Oppenheim (1948). “Studies in the Logic of Explanation.”
Philosophy of Science 15, 135–75.
Hume, D. (1999 [1748]). An Enquiry Concerning Human Understanding. New
York: Oxford University Press.
Hume, D. (2000 [1738]). A Treatise of Human Nature. New York: Oxford Uni-
versity Press.
Hutzler, F. (2014). “Reverse Inference Is Not a Fallacy Per Se: Cognitive Pro-
cesses Can Be Inferred from Functional Imaging Data.” Neuroimage 84,
1061–69.
216 Marco J. Nathan
Illari, P. and F. Russo (2016). “Information Channels and Biomarkers of
Disease.” Topoi 35, 175–90.
Kitcher, P. (1981). “Explanatory Unification.” Philosophy of Science 48 (4),
507–31.
Kitcher, P. (1989). “Explanatory Unification and the Causal Structure of the
World.” In P. Kitcher and W. C. Salmon (Eds.), Scientific Explanation, pp.
410–505. Minneapolis: University of Minnesota Press.
Lewis, D. K. (1986). “Postscript E to ‘Causation’.” In Philosophical Papers,
Volume 2, pp. 193–212. New York: Oxford University Press.
Lewontin, R. (1974). The Genetic Basis of Evolutionary Change. New York:
Columbia University Press.
Lynch, K. E. and P. Bourrat (2017). “Interpreting Heritability Causally.” Philos-
ophy of Science 84(1), 14–34.
Machamer, P. K., L. Darden, and C. F. Craver (2000). “Thinking about Mech-
anisms.” Philosophy of Science 67, 1–15.
Machery, E. (2014). “In Defense of Reverse Inference.” British Journal for the
Philosophy of Science 65(2), 251–67.
Nathan, M. J. (2016). “Counterfactual Reasoning in Molecular Medicine.”
In G. Boniolo and M. J. Nathan (Eds.), Philosophy of Molecular Medicine:
Foundational Issues in Research and Practice, pp. 192–214, New York:
Routledge.
Nathan, M. J. (2021). Black Boxes: How Science Turns Ignorance into Knowl-
edge. New York: Oxford University Press.
Nathan, M. J. and G. Del Pinal (2016). “Mapping the Mind: Bridge Laws and
the Psycho-Neural Interface.” Synthese 193(2), 637–57.
Nathan, M. J. and G. Del Pinal (2017). “The Future of Cognitive Neuroscience?
Reverse Inference in Focus.” Philosophy Compass 12(7), e12427.
NIH Biomarkers Definitions Working Group. (2001). “Biomarkers and Surro-
gate Endpoints: Preferred Definitions and Conceptual Framework.” Clinical
Pharmacology and Therapeutics 69, 89–95.
Pearl, J. and D. Mackenzie (2018). The Book of Why: The New Science of
Cause and Effect. New York: Basic Books.
Poldrack, R. A. (2018). The New Mind Readers: What Neuroimaging Can and
Cannot Reveal about Our Thoughts. Princeton, NJ: Princeton University
Press.
Richardson, K. and M. C. Jones (2019). “Why Genome-Wide Associations with
Cognitive Ability Measures are Probably Spurious.” New Ideas in Psychology
55, 35–41.
Ross, L. N. (2015). “Dynamical Models and Explanation in Neuroscience.”
Philosophy of Science 82(1), 32–54.
Ross, L. N. (2021). “Distinguishing Topological and Causal Explanation.” Syn-
these 198, 9803–20.
Russell, B. (1912). The Problems of Philosophy. Indianapolis, IN: Hackett.
Russo, F. and J. Williamson (2012). “Envirogenomarkers. The Interplay be-
tween Difference-Making and Mechanisms.” Medical Studies 3, 249–262.
Salmon, W. C. (1984). Scientific Explanation and the Causal Structure of the
World. Princeton, NJ: Princeton University Press.
Salmon, W. C. (1998). Causality and Explanation. New York: Oxford Univer-
sity Press.
 The “Toolbox” Problem 217
Scriven, M. (1958). “Definitions, Explanations, and Theories.” In H. Feigl, M.
Scriven, and G. Maxwell (Eds.), Minnesota Studies in the Philosophy of Sci-
ence, Volume 2, pp. 99–195. Minneapolis: University of Minnesota Press.
Spirtes, P., C. Glymour, and R. Scheines (2000). Causation, Prediction, and
Search (2nd ed.). Cambridge and London: MIT Press.
Strevens, M. (2008). Depth. An Account of Scientific Explanation. Cambridge,
MA: Harvard University Press.
Strevens, M. (2016). “Special-Science Autonomy and the Division of Labor.” In
M. Couch and J. Pfeifer (Eds.), The Philosophy of Philip Kitcher, pp. 153–81.
New York: Oxford University Press.
Sul, J., L. Marti, P. Skidmore, A. Cassidy, S. Fairweather-Tait, L. Hooper, and
A. A. Roe (2018). “Population Structure in Genetic Studies: Confounding
Factors and Mixed Models.” PLoS Genetics 14(12), e1007309.
Torrica, B., A. G. Chiocchetti, E. Bacchelli, E. Trabetti, A. Hervás, B. Franke, J.
K. Buitelaar, N. Rommelse, A. Yousaf, E. Duketis, and C. M. Freitag. (2017).
“Lack of Replication of Previous Autism Spectrum Disorder GWAS Hits in
European Populations.” Autism Research 10(2), 202–11.
van der Sijde, M., A. Ng, and J. Fu (2014). “Systems Genetics: From GWAS to
Disease Pathways.” Biochimica et Biophysica Acta 1842(10), 1903–9.
Vineis, P., P. Illari, and F. Russo (2017). “Causality in Cancer Research: A Jour-
ney Through Models in Molecular Epidemiology and Their Philosophical In-
terpretation.” Emerging Themes in Epidemiology 14, 7.
Visscher, P., N. R. Wray, Q. Zhang, P. Sklar, M. I. McCarthy, J. Brown, A. Mat-
thew, and Yang (2017). “10 Years of GWAS Discovery: Biology, Function,
and Translation.” The American Journal of Human Genetics 101(1), 5–22.
Vrieze, S. I., W. G. Iacono, and M. McGue (2012). “Confluence of Genes, Envi-
ronment, Development, and Behavior in a Post-GWAS World.” Development
and Psychopathology 24(4), 1195–214.
Weiskopf, D. A. (2021). “Data Mining the Brain to Decode the Mind.” In F.
Calzavarini and M. Viola (Eds.), Neural Mechanisms: New Challenges in the
Philosophy of Neuroscience, pp. 85–110. Cham: Springer.
Woodward, J. (2003). Making Things Happen. A Theory of Causal Explana-
tion. New York: Oxford University Press.
Woodward, J. (2018). “Some Varieties of Non-Causal Explanation.” In A. Reut-
linger and J. Saatsi, Explanation Beyond Causation: Philosophical Perspec-
tives on Non-Causal Explanations, pp. 117–37. Oxford: Oxford University
Press.
Section 3

Research Tools,
Integration, Circuits, and
Ontology
10 How Do Tools Obstruct
(and Facilitate) Integration
in Neuroscience?
David J. Colaço

1 Introduction
For the past decade, philosophers have investigated how the methods,
data, and especially the explanatory frameworks of the subfields of neu-
roscience might be integrated (Craver 2007; Sullivan 2017). This interest
is unsurprising, as neuroscientists investigate diverse entities and phe-
nomena at different scales of the brain, which they aim to link via sci-
entific integration. Neuroscientists have developed integrative projects
(Shepherd et al. 1998; Markram 2012; Jorgenson et al. 2015; Amunts
et al. 2016) to both better understand the brain at different scales and
develop treatments for diseases (Markram 2012; Jorgenson et al. 2015).
Given that these integrative projects have been active for a decade,
it seems prudent to ask how successful they have been. Systematic in-
tegrative brain modeling projects, such as the Human Brain Project or
Blue Brain, have not fulfilled their ambitious goals (Yong 2019), and the
promised therapeutic interventions have yet to be delivered. At the same
time, neuroscientists have integrated locally, with piecemeal connections
made between subfields (Sullivan 2017). This disconnect between desires
and reality raises two questions. First, why has integrative neuroscience
not matched its expectations? Second, why does it succeed when it does?
In this chapter, I answer these questions by arguing that tools can both
obstruct and facilitate scientific integration, where tools are the mate-
rials and technologies that researchers use to study brain structure and
activity. My argument is built on two premises about popular tools in
neuroscience: (1) these tools have different constraints, (2) despite these
constraints, these tools productively generate knowledge. Together, these
premises entail that established tools contribute to knowledge production
often at the expense of the integration of methods, data, or explanatory
frameworks. This fact explains why we can find cases of local integration
across neuroscience, but systematic integration remains elusive.
In Section 2, I detail integration in neuroscience. I discuss how differ-
ent components – methods, data, and explanatory frameworks – might
be integrated, and I address the disconnect between desires for integra-
tion and its reality. In Section 3, I defend my premises for why tools

DOI: 10.4324/9781003251392-14
222 David J. Colaço
obstruct scientific integration. I illustrate them with examples from cog-
nitive neuroscience and molecular and cellular cognition. In Section 4,
I explain why some new projects, such as the BRAIN Initiative, prior-
itize the development of new tools to facilitate integration. I support
my claims with a case study of the tool known as CLARITY. This case
shows that tools can facilitate integration, explaining why local integra-
tion occurs despite as-of-yet unfulfilled desires of systematic integration,
given the constraints and productivity of tools.

2 W hat Kind of Scientific Integration Do Neuroscientists

Want?
Neuroscientists claim that integration will elucidate the brain as well
as foster treatments for psychiatric disorders. For instance, architects
of the tool-focused BRAIN Initiative (Brain Research through Advanc-
ing Innovative Neurotechnologies) suggest that integration will inform
interventions upon brain systems from cells to circuits, regions, and the
whole brain. Integration is needed, the architects argue, because ad-
dressing components in isolation has not resulted in viable interventions
(Jorgenson et al. 2015, p. 9; see also Bassett and Sporns 2017, p. 355).

2.1 What Is Scientific Integration?

Before evaluating these projects, I first detail what integration is. At its
broadest, scientific integration covers “various possible modes of rela-
tionship that can be developed between disciplines” (Bechtel 1993,
p. 33). This is a start, but it does not explain how subfields link via sci-
entific integration. To clarify the “how” question, O’Malley and Soyer
characterize scientific integration as “a multi-faceted dynamic, in which
methods, bodies of data and explanations are synthesized in order to un-
derstand and intervene more effectively” (2012, p. 59). This characteri-
zation captures that we can integrate methods (Mitchell and Gronenborn
2017), data (Leonelli 2016), or explanatory frameworks (Sullivan 2017).
The first mode is methodological integration: “directing a range of
methods at a particular biological system or research problem in order
to gain a multidimensional understanding of how the system works”
(O’Malley and Soyer 2012, p. 60). The BRAIN Initiative aims to in-
tegrate methods, given the desire to “facilitate the integration of many
multiple approaches into a single experiment” (Jorgenson et al. 2015,
p. 5). Thus, researchers want tools that can be used in tandem when they
investigate a neural system, so that they can acquire knowledge about
different aspects of this system (Kotchoubey et al. 2016).
The second mode is data integration: “the activity of making compa-
rable different data types from a huge variety of inconsistent sources,”
which “forms a new body of information that can be treated as a unified
 Integrating subfields of neuroscience 223
whole” (O’Malley and Soyer 2012, p. 61). The BRAIN Initiative aims
to integrate data: “datasets from different laboratories will cover spatial
scales ranging from micrometres to metres… and time scales ranging
from milliseconds to minutes, hours or even the lifetime of an organism”
(Jorgenson et al. 2015, p. 7). Data integration is an aim of the Human
Brain Project, where “unifying computer models” will “integrate all ex-
isting data” (Markram 2012, p. 52).
The third mode is explanatory integration: “when combinations of dif-
ferent methods invoke the fusion of explanatory frameworks, commonly
in the accommodation of multiple levels of phenomena” (­O’Malley and
Soyer 2012, p. 61). The BRAIN Initiative aims to integrate explanatory
models: “a unified view of the brain will cross spatial and temporal lev-
els, recognizing that the nervous system consists of interacting mole-
cules, cells and circuits” (Jorgenson et al. 2015, p. 9). This mode also is
the focal aim of the Human Brain Project (Markram 2012).
This three-part distinction is only one way we can account for scien-
tific integration, but it matches the interests of neuroscientists. Rather
than vaguely integrating “disciplines,” distinguishing methods, data,
and explanatory frameworks also characterizes how subfields of neuro-
science can be integrated. This is helpful, as integrating one mode need
not entail integrating the others. Further, it explains why integration
is conceptually distinct from epistemic aims like generalization. Gener-
alization is inducing or extrapolating about an experimental outcome;
integration is about linking methods, data, and frameworks.
We also can distinguish local and systematic integration. While there
is no hard separation, local integration is piecemeal both in terms of
modes and subfields: two subfields might link their methods, data, or ex-
planatory frameworks. For instance, neurobiology and behavioral neu-
roscience integrate methods and explanatory frameworks in the study of
spatial memory via variants of the Morris water maze (Sullivan 2016, p.
671). This is a case of local integration, which has generated knowledge,
predictions, and experimental protocols. By contrast, systematic integra-
tion is wholesale. The Human Brain Project and Blue Brain are attempts
at systematic data and, more importantly, explanatory integration, as
each is intended to develop a brain model that is informed by diverse
studies at different scales.

2.2 The Challenge

There is a consensus amongst philosophers and neuroscientists that in-
tegration is challenging. Sullivan notes that there are “conceptual and
methodological obstacles to integration” in neuroscience (2017, p. 129),
while Kotchoubey and colleagues note that there are “methodological
and epistemic problems that obstruct the development of human neu-
roscience” (2016, p. 1). Likewise, the architects of the BRAIN Initiative
224 David J. Colaço
note that integration is “a particularly daunting challenge” (Jorgenson
et al. 2015, p. 7).
The literature on integrative neuroscience suggests that there are
challenges for each mode. For methodological integration, there are
“methodological gaps” between subfields, which create “problems of
integration… when we are trying to connect several disciplinary tradi-
tions” (Kotchoubey et al. 2016, p. 5). For data integration, while there is
a deluge of data being produced about neural systems (Kotchoubey et al.
2016, p. 1), philosophers like Sullivan argue that “we should not expect
that amassing potentially discordant data from many different laborato-
ries under a single common label… will shed light on real divisions in the
causal structure of the world” (2017, p. 134). For explanatory integra-
tion, there is a worry about what experiments target, which undermines
the “connectability” of frameworks that are developed (Sullivan 2016, p.
665). Neuroscientists in different subfields may not investigate the same
targets, complicating the synthesis of their explanatory frameworks.
Integrating neuroscience is challenging but not impossible. As al-
ready mentioned, cases can be made for local integration. When look-
ing for more systematic integration, network neuroscience adopts an
“integrative perspective” when modeling neural systems (Bassett and
Sporns 2017, p. 353). Network models are designed to integrate data
from different studies performed at different scales, which can be used
by researchers to explanatorily integrate models of these scales. So far,
scientists have generated models that integrate observations performed
on hundreds of neurons with those at higher scales, as “network science
tools are perfectly suited to accomplish the important goal of crossing
levels or scales of organization, integrating diverse data sets, and bridg-
ing existing disparate analyses” (Bassett and Sporns 2017, p. 362).
At the same time, challenges abound in this nascent subfield of neu-
roscience. Beyond the complaint that these models fail to meet expec-
tations (Yong 2019), network models have yet to integrate explanatory
frameworks of most populations of neurons. These limitations reflect
that integrative neuroscience has yet to catch up with scientists’ desires.
More interesting is the fact that network modeling is only part of the
puzzle. Researchers need data that support these models (Sporns 2014,
p. 653). This fact indicates that there remain issues for integrating data
from existing neuroscience studies, which is due to how these studies
were performed. This implication, in turn, reveals that network neuro-
science (and modeling in neuroscience more generally) is but one part of
integrating neuroscience, which can benefit from new experiments.
Thus, there are successful (local) cases of neuroscience integration, but
there also is a disconnect between the desires for (especially systematic
explanatory) integration and its reality. Why is this the case? I explore
an answer that is hinted at by the network neuroscientists: understand-
ing integration requires us to reflect upon the tools of neuroscience. By
 Integrating subfields of neuroscience 225
‘tools,’ I mean the materials and technologies that researchers employ
when measuring or manipulating brain activity and structure (Bickle
2016). In this chapter, I focus on tools used in the production of data via
experimentation and observation.

3 How Do Tools Obstruct Scientific Integration?

The previous section addresses that there is a desire to integrate neuro-
science to a greater degree than what has been achieved. One obstacle,
it would seem, is the diversity of tools used by neuroscientists. However,
why might tools be obstacles?

3.1 Constraints and Productivity

First, all tools constrain empirical research. The choice of tools limits
what other tools can be used, what domains of systems can be inves-
tigated, what aspects of the system can be manipulated or measured,
and what data can be produced. Researchers use tools to achieve cer-
tain outcomes in an experiment, which requires that they be paired with
methods that complement a tool’s use. While exploration might lead to
discoveries about new ways to use a tool, using a tool beyond its con-
straints often results in a failure of its ability to produce knowledge.
Colaço describes the “constraining role” of tools by appeal to the
preparation and staining techniques used in microscopy: if “one chooses
to use a staining technique on a cell to investigate it with microscopy,
one will not be able to investigate any components of the cell that are
destroyed by the staining process” (2018, p. 38). Thus, tools whose con-
straints do not overlap are not usable in tandem due to requirements for
researchers to prepare or modify the system in distinct ways that accom-
modate the use of these respective tools. Further, the modifications each
tool makes to this target complicate the comparison of instances of the
use of tools that each modify but ostensibly measure or manipulate the
same target.
Not all constraints result from preparation or modification. Invasive
electrophysiological stimulation is constrained inasmuch as it cannot be
performed on humans unless cranial intervention is deemed necessary
for medical intervention. Other tools are constrained by issues of tem-
poral or spatial scale: magnetic resonance imaging (MRI) cannot image
individual neurons. Regardless of the reason for the constraint, if differ-
ent tools do not overlap in constraints, they cannot be used together to
investigate aspects of the same system.
Second, established tools are productive. Tools are developed and
standardized to be easy and efficient for researchers to reliably produce
useful data about the targets whose domain lies within the constraints
of the tool. Thus, despite the fact that tools have constraints, they are
226 David J. Colaço
effective when used within these constraints. These tools can be used to
generate knowledge about the systems that researchers want to study.
In this sense, tools are “productive” in an epistemic sense. Correspond-
ingly, these researchers also can publish research using these tools,
allowing them to be “productive” in the sense that they can achieve
prudential aims, such as publishing high-impact work, getting funding,
and building one’s professional reputation.
The productivity of tools used in existing experimental paradigms
is well-supported in the existing literature. For instance, Krohs claims
that “many experiments are… done in the way they are actually done,
because they are extraordinarily convenient to perform” (2012, p. 53).
Tools are easy and efficient means of achieving certain aims in an ex-
periment. However, a tool “strongly channels research,” and data that
are produced are determined by it (Krohs 2012, p. 53). Likewise, due to
their constraints, tools cannot be used in just any experimental design.
Thus, when researchers find tools to be productive for the aims that they
have, we should expect that researchers will develop protocols that use
them as opposed to those that do not.
A consequence of productivity in science rears its head in debates
about unification. Harp and Khalifa claim that researchers might sacri-
fice long-term gains in unifying research for them to achieve short-term
gains. Because there remain achievable goals in local research projects,
“the utility of an optimal local strategy is sometimes greater than that of
an optimal cosmopolitan strategy, even when the value of a local success
is lower than that of an unquestionable, and hence cosmopolitan, suc-
cess” (Harp and Khalifa 2015, p. 441). While these claims address unifi-
cation, a similar point can be made about integration in neuroscience. So
long as existing tools reliably produce useful data about neural systems
that researchers want to investigate, then we should expect researchers
to continue to use these tools even at the expense of methodological,
data, and explanatory integration.
Together, constraints and productivity suggest that researchers can
either continue to reliably produce useful data about the systems they
investigate, or they can use less productive tools to less reliably produce
data that may be less useful for investigating their targets. As long as the
established tools continue to produce knowledge about neural systems,
using these tools will be a reliable way for researchers to produce this
knowledge, regardless of their contribution (or lack thereof) to scientific
integration.

3.2 How Do Tools Obstruct Scientific Integration?

Methodological integration is most obviously obstructed by the diversity
of tools. Distinct tools have different constraints that are intimately tied
to their productivity. If these constraints require the neural system to
 Integrating subfields of neuroscience 227
be prepared or manipulated in incompatible ways, researchers will not
be able to direct “a range of methods at a particular biological system”
in tandem “in order to gain a multidimensional understanding” of this
system. For instance, microscopy often depends on staining and slicing
neural tissue in order for researchers to image it, while structural MRI
requires the system to be intact and free of metallic stains. These tools
cannot be used in tandem without undermining their respective produc-
tivity. While sequential uses of some tools, such as Logothetis and col-
leagues’ use of functional MRI (fMRI) and single-cell recording (2001),
are possible, this kind of integrative paradigm is challenging to perform
and cannot be performed on all brain structures that each tool can be
used on individually. These facts are illustrated by how infrequently
Logothetis and colleagues’ paradigm is reproduced.
By contrast, if tools with distinct constraints are used on different in-
stances of the same neural system, the modifications that must be made to
each system for researchers to use the particular tool within its constraints
can transform them in inconsistent ways. This inconsistent set of changes,
in turn, makes it difficult to relate the effects that each tool has on the
neural system under investigation. Along these lines, because tools limit
the ability for researchers to efficiently use other tools, the use of a range
of methods on a neural system can make a tool less productive. This not
only undermines the reliability of the data that any tool can produce when
used within its constraints; it also makes it harder to produce these data.
Perhaps readers think that methodological integration is not important
in the grand scheme of things. However, data integration is obstructed
by tools as well. The data produced from tools reflect their constraints:
these data reflect the assumptions, biases, and modifications that must
be made about the system in order for researchers to use the tool to pro-
duce reliable data about it. This is a problem if we accept that data inte-
gration “refers to the process of theorizing and modelling databases,” as
the theories and models we build of the databases will reflect constraints
of particular tools at the expense of those frameworks that are limited
by other tools that do not share the same constraints (O’Malley and
Soyer 2012, p. 61).
While this problem has its basis in the idea that experiments “can
produce highly variable results that are hard to model in any general-
izable way,” and “combining different data types remains a stubbornly
persistent problem” (O’Malley 2013, p. 552), this problem becomes
more serious when we consider the effect these procedures have on data
production. On one hand, if developing databases requires an appeal
to data produced through the use of tools, these datasets will reflect
the constraints of these tools. On the other hand, if researchers develop
databases based on a framework that is not indebted to the data pro-
duced from the use of tools, then these databases may not match the con-
straints of the diverse tools whose data will plug into these databases.
228 David J. Colaço
Either way, the tools complicate “the activity of making comparable dif-
ferent data types,” and they therefore can obstruct data integration.
Even with methodological and data integration obstructed, readers
might be most interested in explanatory integration. However, explana-
tory integration is obstructed as well. Explanatory frameworks develop
and advance based on data obtained from studies that use tools. As a
result, these frameworks reflect the constraints of the tools used to pro-
duce these data, and these frameworks also support the subsequent use
of these tools to produce data. While it is to be expected that any empir-
ical science’s explanatory frameworks will be influenced by the studies
that are performed in that discipline, neuroscience, having little in the
way of a systematic theory that guides all research, is built on the results
of experiment and observation.
The mere fact that different experiments share labels does not gener-
ate an integrated explanatory framework. These labels “do not reflect
the intra and interdisciplinary consensus as to how to generally define
or how to produce, detect and measure the phenomena designated by
those labels” (Sullivan 2017, p. 134). This fact, in turn, results from the
fact that constructs are differently operationalized in studies that de-
ploy different tools. Because operationalizations capture measurement
(Feest 2005), and the constraints of tools must match measures of target
phenomena, diverse tools can result in the need to measure targets in
distinct ways. At worst, operationalizations can be incompatible, but
the larger issue is that how they relate might be poorly understood by
researchers, thus undermining explanatory integration.
Neither the constraints nor productivity of tools entails that the three
modes of integration cannot occur. Rather, these features of tool use
show that the use of tools to generate scientific knowledge can be (and
often is) unsympathetic to the aims of scientific integration. In this sense,
tools can serve as obstacles to methodological, data, and explanatory
integration, leading to an undermining of systematic integration and sit-
uational cases of local integration.

3.3 T
 ools Obstructing Scientific Integration: MCC and
Cognitive Neuroscience
An example of how tools obstruct scientific integration between neu-
roscience subfields can be found when we compare different subfields
that study cognition. Take molecular and cellular cognition (MCC), a
subfield where molecular and cellular approaches are used to study be-
havioral and cognitive phenomena. MCC uses tools like “molecular ma-
nipulations (e.g., gene targeting, viral vectors, pharmacology), cellular
measures and manipulations (neuroanatomy, electrophysiology, optoge-
netics, cellular and circuit imaging), and a plethora of behavioral assays”
in the study of cognition (Silva et al. 2013, p. 8). Researchers use these
 Integrating subfields of neuroscience 229
tools to “intervene into molecular pathways in neurons and attempt to
develop explanations that bridge molecules, cells, circuits, and behav-
ior” (Silva et al. 2013, p. 12).
MCC does not readily integrate with other subfields that involve the
study of cognition, such as cognitive neuroscience. This lack of integra-
tion with cognitive neuroscience is a feature of MCC according to the
Molecular and Cellular Cognition Society:

Unlike Cognitive Neuroscience, which historically has focused on

the connection between human brain systems and behavior, the field
of Molecular and Cellular Cognition studies how molecular (i.e. re-
ceptor, kinase activation), intra-cellular (i.e. dendritic processes), and
­inter-cellular processes (i.e. synaptic plasticity; network representations
such as place fields) modulate animal models of cognitive function.
(Molecular and Cellular Cognition Society)

Thus, even though these two subfields ostensibly study the same cogni-
tive capacities, they study these capacities with distinct tools. In addi-
tion, this quote suggests that MCC researchers perceive their work to be
an alternative to cognitive neuroscience, rather than something that is
intended to integrate with it.
The differences between MCC and cognitive neuroscience can be il-
lustrated by how memory might be studied in each subfield. To study
memory deficits, MCC researchers use tools like electrophysiology and
transgenetics to manipulate synaptic and molecular activities in a mouse
model organism (Silva 2003). To use these tools on this model organ-
ism, the researchers must operationalize memory deficits in a way that
can be measured in mice. At this time, tools like electrophysiology and
transgenetics have not been used on human subjects due to ethical con-
straints, leaving open how the mechanistic schemata developed in MCC
relate to schemata that represent or explain memory phenomena in other
animals (let alone humans).
Cognitive neuroscientists use different tools than those that are used in
MCC. When studying human memory phenomena, they often use MRI
to correlate brain activity differences between healthy and impaired hu-
man subjects (Gabrieli 1998). This tool can be methodologically inte-
grated with other tools like EEG. However, methods like fMRI cannot
be concurrently used with electrophysiology in mouse model organisms,
due to electromagnetic constraints. This incompatibility obstructs the
integration of methods that employ tools with distinct constraints. Fur-
ther, how memory deficits are operationalized in the two cases is not
equivalent, making it difficult to compare the data from humans and
mice. This mismatch obstructs data integration from studies whose tar-
gets may not be equivalent, even if these targets share the same nomen-
clature. This fact ultimately obstructs the integration of explanatory
230 David J. Colaço
frameworks from these subfields. Because the relation between targets is
unclear, it also is unclear how the mechanistic schemata in mouse model
organisms relate to memory in humans.
This case does not exhaust the tools that are used in MCC or cogni-
tive neuroscience. For instance, cognitive neuroscience can employ more
invasive tools in model organisms like the rhesus macaque. Further, cog-
nitive neuroscience has integrated explanatory frameworks from neuro-
biology (though not MCC) in its history. For instance, neurobiological
work on long-term potentiation in the rabbit hippocampus (Colaço
2020) has greatly informed neuroimaging studies on the role of the hip-
pocampus in memory formation and consolidation. Thus, integration
between cognitive neuroscience and MCC can occur, but we must be
clear about exactly what of these subfields is integrated. This interaction
is just one example of a case where the reality fails to meet the desires of
the integration of neuroscience, but it highlights what integration is up
against. While local methodological, data, and explanatory integration
might occur – MCC and cognitive neuroscience have cases of all three –
each can be obstructed. Additionally, systematic integration remains elu-
sive and might not be desired by researchers in these respective subfields.
Nonetheless, we must ask how the methods, data, and explanatory
frameworks that are used in these subfields might be integrated. Each
mode of integration is obstructed. However, MCC and cognitive neu-
roscience each are individually successful, indebted to the fact that each
subfield has productive tools that, when used within their respective con-
straints, produce useful data about the systems the respective groups of
researchers target. These data, in turn, feed into the explanatory frame-
works of each subfield, including how the targets are operationalized and
how brain activity and structure are understood. Hence, the integration
of the methods, data, and explanatory frameworks of these subfields
counterbalance with the productivity of those that do not facilitate in-
tegration.1 Further, it is worth speculating whether researchers in either
camp want to integrate their methods, data, or explanatory frameworks,
even if there were no other obstacles to achieving more systematic inte-
gration of these two subfields.

4 What New Tools Change

The previous section explains how tools can obstruct methodological,
data, and explanatory integration due to their constraints and productiv-
ity. However, my cases show that tools are not an impassible roadblock
to scientific integration, despite their propensity in some circumstances
to obstruct it. Though systematic integration has failed to meet ex-
pectations, I have reviewed successful cases of local integration. What
does this fact tell us about integration in neuroscience? To answer this
question, I look to the role new tools in neuroscience play in facilitating
 Integrating subfields of neuroscience 231
integration. After all, the project known as the BRAIN Initiative aims
“to develop and apply new tools and technologies for revolutionizing
our understanding of the brain,” which is in contrast to the principally
modeling-focused projects of both the Human Brain Project and Blue
Brain (Jorgenson et al. 2015, p. 1).
Architects of the BRAIN Initiative note that integration is a “chal-
lenge,” and Section 2 highlights that the Initiative speaks to issues of in-
tegrating methods, data, and explanatory frameworks (Jorgenson et al.
2015, p. 7). Thus, the Initiative is designed to facilitate the integration
of methods, data, and ultimately explanatory frameworks regarding re-
search at different “scales” through the development of novel tools. The
microscale is the size of cells, molecules, and small circuits, while the
macroscale is the size of brain regions or the whole brain (Chang 2015).
The mesoscale is everything in-between, encompassing large circuits
and small populations of cells (Huang et al. 2018). It is also the least
investigated: “one of the greatest current challenges facing neuroscience
is that of bridging the gulf that exists between the molecular and the cel-
lular level on one hand and that of brain imaging, cognition and action
on the other” (Grillner et al. 2005, pp. 615–616).

4.1 CLARITY
A tool that drove the formation of the BRAIN Initiative is CLARITY,
which renders intact tissue transparent and amenable to optical and flu-
orescent microscopy paradigms (Chung and Deisseroth 2013). This tool
is indebted to the Deisseroth laboratory at Stanford. Prior to the devel-
opment of CLARITY, optical microscopy only could be performed on
the microscale, due to the need to slice tissue into thin segments so that
light can penetrate them. Indirect macroscale imaging tools – such as
MRI – cannot be used by researchers to investigate entities or phenom-
ena at the microscale or mesoscale, due to constraints of the coarseness
of indirect imaging. CLARITY solves the problem that some lipids in
cellular tissue scatter light. This technique removes these lipids and de-
velops a transparent support structure in the tissue.
The process begins when the tissue is placed in a monomer and linking
chemical solution. These chemicals diffuse into the tissue, and they bind
to proteins and nucleic acids in this tissue. However, these chemicals do
not bind to the lipids that scatter light. Following the binding, the im-
pregnated tissue is heat shocked, which causes the monomer and linking
chemicals to bind to one another, forming a “mesh.” The biomolecules
that bind to the chemicals embed in this mesh as it sets. Following the
development of the mesh, detergents are used by researchers to remove
the non-embedded molecules like the light-scattering lipids. Because the
mesh is in place, the tissue, minus the light-scattering lipids, remains in
the configuration it was in at the start of the process.
232 David J. Colaço
CLARITY has different constraints when compared to those of mi-
croscale slice microscopy or the indirect means of measuring or imag-
ing at the macroscale. Instead of researchers needing to slice tissue into
pieces, CLARITY allows researchers to prepare larger dimensions of tis-
sue. Because of its distinct constraints when compared to previous tools,
the use of CLARITY produces microscale data about the properties of
individual neurons and other cellular material as well as mesoscale data
about the structural relations of the population of neurons and other
material within the tissue.
Just like any other tool, CLARITY does not lack constraints. The
biological system in question is killed by CLARITY, and the hydrogel
and detergents used in the process can damage cellular structures in
predictable ways. What matters for my discussion of integration in neu-
roscience is that its constraints are novel when compared to other tools,
including the existing optical and indirect imaging methods mentioned
above, allowing researchers to produce data that are relevant to the data
produced using other tools. Thus, CLARITY overlaps other tools’ do-
mains: researchers can use it with existing light and fluorescent micros-
copy methods to investigate microscopic, mesoscopic, and macroscopic
structures.
Unlike many imaging tools of the past, CLARITY is a “multi-scale”
tool, with constraints covering the microscale, mesoscale, and macro-
scale. In this sense, CLARITY crosscuts previous research that used tools
with incompatible constraints. This is a direct consequence of CLAR-
ITY’s comparatively novel set of constraints: the change in constraints
affects how the tool informs existing frameworks, helping with method-
ological, data, and (perhaps most importantly) explanatory integration.
The creators of CLARITY have sought to make this tool productive
and easy to incorporate into existing paradigms that deploy other mea-
surement, imaging, and manipulation tools. Beyond publishing works
that communicate the theoretical basis and validity of the tool (Chung
et al. 2013), these creators developed a “wiki” to communicate the use
of the tool, the materials, and the kinds of projects for which CLAR-
ITY is useful (CLARITY Resource Center). This “CLARITY resource
center” also includes a forum for users of the tool to discuss their uses
of CLARITY as well as the issues that they face when they use it. This
forum is paired with “boot camps” that train researchers to use CLAR-
ITY. These resources provide researchers the opportunity to obtain the
first-hand experience of using the tool and second-hand experience from
other users. These strategies show that its creators aim to make the use
of CLARITY easy, efficient, and reliable.
The point of this section is not to suggest that CLARITY is some kind
of “silver bullet” tool that will resolve integration challenges in neuro-
science. Rather, it is one tool amongst several that can, and as I will
show does, contribute to integration. To put this tool into the broader
 Integrating subfields of neuroscience 233
context of neuroscience inquiry, we must consider some of the limita-
tions of CLARITY. For instance, it takes time and effort for researchers
to make a tool productive, and CLARITY is no exception. There remain
aspects of its use that make it less productive when compared to other
tools. It can take months for researchers to successfully perform CLAR-
ITY on neural systems, diminishing its efficiency compared to other
imaging tools.
There already have been attempts to reduce the time needed to prepare
the tissue and image it (Gradinaru et al. 2018), but the fact that the tool
is slow limits its productivity at this time. Likewise, the wiki dedicates
a section to troubleshooting the application of the tool, suggesting that
it can be difficult to master CLARITY. These troubleshooting tips will
likely help ease the use of CLARITY in the future. Nonetheless, their
existence reflects the fact that there is a degree of skill required for re-
searchers to use the tool correctly.

4.2 Does CLARITY Facilitate Scientific Integration?

CLARITY is an impressive tool. It can help to generate knowledge about
brain structure and integrate with some other methods of neuroscience,
and the data produced with CLARITY are designed to be integrated
with data resulting from the use of different tools (Gradinaru et al.
2018). Together, these facts support the use of CLARITY by researchers
to address brain structures that can be modeled, ultimately contributing
to explanatory integration. This is all well and good, but has this new
tool facilitated integration in neuroscience?
For methodological integration, CLARITY has been integrated into
MRI paradigms that allow researchers to advance their understanding
of how tissue types, such as lipids and proteins, differentially contrib-
ute to imaging results (Leuze et al. 2017). This local integration pro-
vides researchers with a better theoretical understanding of how the
biological material of the brain relates to their imaging results, with a
finding that lipids are the “dominant source of MRI contrast in brain
tissue” (Leuze et al. 2017, p. 412). This case shows that the integrative
use of multiple tools contributes to knowledge generation about brain
tissue and its relation to these tools that could not be achieved through
the use of either tool alone. Thus, methodological integration supports
both knowledge generation about neural systems and knowledge gener-
ation about the theoretical basis of MRI paradigms. The former fosters
growth in the theory of the system, while the latter fosters growth in the
theory of the technique (Colaço 2018).
For data integration, CLARITY has been paired with diffusion tensor
imaging for researchers to collect and model data about white matter
microarchitecture (Chang et al. 2017). Though non-invasive, diffusion
tensor imaging produces only indirect evidence of white matter structure
234 David J. Colaço
via the measurement of water diffusion in and around this structure.
When paired with CLARITY, researchers can correlate this water move-
ment data with structural data regarding the biological sources of this
diffusion. This local integration thus provides researchers a means of
combining and relating databases that result from distinct imaging tools,
or tools that measure different features of brain structures. Further, it
provides data about the relation between datasets, giving researchers an
understanding of how data from different tools relate and thus providing
them a useful form of “metadata” (Leonelli 2016, p. 97). By understand-
ing the means of integrating these databases, researchers aim to “inte-
grate 3D brain-wide molecular analyses with these large-scale efforts”
(Chang et al. 2017, p. 260).
For explanatory integration, CLARITY has been deployed to model
prefrontal cortex cell typology, which provides a means of understand-
ing how cell type and population structure correlate with behavior (Ye
et al. 2016). Because CLARITY allows for the modeling of intact struc-
tures, both phenotyping and cell wiring at the population level can be in-
tegrated into a model of the so-called “prefrontal cortex” of mice. This
local integration thus provides researchers a means of modeling the ro-
dent prefrontal cortex and its contribution to behavior while also acquir-
ing the cellular and molecular detail of the components of the prefrontal
cortex. This explanatory integration is possible because CLARITY can
be used by researchers to acquire both microscale and mesoscale data,
due to its novel constraints. The ability to generate a detailed, molecu-
larly satisficing, three-dimensional model of a brain area as complex as
the prefrontal cortex, even that of mice, also lays the groundwork for
producing population-level anatomical models of constructs like mem-
ory engrams, facilitated by the integrative use of clarifying tools like
CLARITY, labeling techniques, and manipulation tools like optogenet-
ics (Roy et al. 2019).
These cases of the use of CLARITY show how new tools can success-
fully facilitate all three modes of integration in neuroscience, with the
last case involving the facilitation of more than one mode concurrently.
What has been achieved in these cases is what the BRAIN Initiative was
designed to promote: create and hone a number of new tools that serve as
the seeds for integration to concurrently develop across several subfields
of neuroscience. By recent accounts, this strategy has been successful,
with numerous new tools facilitating local methodological, data, and
explanatory integration across a variety of neuroscience subfields (Koro-
shetz et al. 2018). These cases of integration address issues faced by re-
searchers who have specific aims in certain subfields of neuroscience.
Despite these cases being, by all accounts, successful instances of in-
tegration in neuroscience, all of these cases are nonetheless of local in-
tegration. Thus, CLARITY, like many other tools affiliated with the
BRAIN Initiative, has yet to contribute to the systematic explanatory
 Integrating subfields of neuroscience 235
integration that some researchers desire. For instance, nothing described
in this section matches the systematic brain modeling of the Human
Brain Project or Blue Brain. This fact puts CLARITY research in the
same camp as other cases of integration, such as Sullivan’s examples,
MCC, and cognitive neuroscience, that I discussed earlier in this chap-
ter. What does this fact tell us?
The success of these CLARITY cases results from the fact that adopt-
ing a tool with different constraints has allowed researchers to integrate
it with their current protocols and thus maintain a productive level of
research, both in terms of generating knowledge and in publishing excit-
ing, high-impact materials. In this sense, tools like CLARITY serve as
an evolution or amendment to current, productive practices, rather than
a radical revision to these practices. This ability to achieve local integra-
tion by changing constraints without sacrificing productivity is the back-
bone to the BRAIN Initiative’s successes. The same cannot be said for
attempts at systematic integration that focus on modeling, which have
not met their ambitious expectations of systematic data and explanatory
integration in the form of a unified model of brain structure and func-
tion. The changing of tools has allowed researchers to maintain their re-
search practices in a sympathetic form to their prior research, as opposed
to the need to sacrifice their current practices in an effort to achieve their
systematic desires. This lack of a comparative compromise mirrors the
tradeoff described by Harp and Khalifa: systematic integration does not
match with the utility of research practices, which is why it often is not
pursued. By contrast, local integration can be matched with the utility of
research practices when new productive tools are introduced. Hence, we
can find many successful cases of the three modes of local integration,
while successful systematic explanatory integration still is elusive.
While I have highlighted an apparent gulf between local and system-
atic integration, this distinction does not entail that local integration
cannot facilitate systematic integration (or vice versa). The BRAIN
Initiative’s results are compatible with the idea that we ultimately ulti-
mately gain a systematic model of brain structure and function, perhaps
through the linking of many cases of local integration. It also is compat-
ible with existing, modeling-focused initiatives for systematic data and
explanatory integration. What matters is that tool change can directly
and successfully facilitate local methodological, data, and explanatory
integration without any direct achievement of systematic integration.
Again, this explains why we can find cases of local integration in neuro-
science, while systematic integration remains elusive.
Ultimately, the case of CLARITY shows something important about
the role of tools in integrative neuroscience: tools can facilitate local in-
tegration for the same basic reasons that they can obstruct it. Tools have
constraints, which affect how they can be used to generate knowledge
in a reliable manner. Likewise, tools are productive, which affects how
236 David J. Colaço
they likely will be used to generate knowledge for epistemic and pruden-
tial reasons. If tools’ constraints do not overlap, it will be a challenge
to integrate research practices that respectively deploy them. If they do
overlap in a productive manner, which can be achieved by developing
new tools with constraints specifically designed to overlap in a produc-
tive manner, integration will be facilitated. CLARITY is one example of
a new tool that achieves this facilitation of scientific integration.

5 Conclusion
In this chapter, I have shown that resolving the challenges of integration
in neuroscience requires us to reflect upon both the constraints and the
productivity of the tools that are popular in the discipline. I have ex-
plained when tools are obstructions and facilitators of three modes of in-
tegration in neuroscience. The fact that productive but constrained tools
lead researchers to continue doing research at the expense of engaging
in integrative practices is a topic that has not been sufficiently addressed
in the philosophy of science, and it is worthwhile to appreciate that this
relation is challenging but not impossible for us to overcome.
As I have shown, local integration can be facilitated by new produc-
tive tools with novel constraints, even if this does not promote any sort
of systematic integration, especially in the short run. This is not a mere
possibility; it is actively achieved by projects like the BRAIN Initiative
and research with the tool CLARITY. Nonetheless, my premises of con-
straints and productivity are important to understanding the promise of
integration in neuroscience, as they explain why tools both obstruct and
facilitate scientific integration.

Note
1 This challenge reflects Bechtel’s claims that integration that creates a new
subfield such as MCC often results in a great deal of specialization. This
specialization, in turn, can result in disintegration between the new subfield
and previously established subfields (1993, p. 278).

References
Amunts, K., Ebell, C., Muller, J., Telefont, M., Knoll, A., & Lippert, T. (2016).
The human brain project: Creating a European research infrastructure to de-
code the human brain. Neuron, 92(3), 574–581.
Bassett, D. S., & Sporns, O. (2017). Network neuroscience. Nature Neurosci-
ence, 20(3), 353–364.
Bechtel, W. (1993). Integrating sciences by creating new disciplines: The case of
cell biology. Biology and Philosophy, 8(3), 277–299.
Bickle, J. (2016). Revolutions in neuroscience: Tool development. Frontiers in
Systems Neuroscience, 10, 24.
 Integrating subfields of neuroscience 237
Chang, E. F. (2015). Towards large-scale, human-based, mesoscopic neurotech-
nologies. Neuron, 86(1), 68–78.
Chang, E. H., Argyelan, M., Aggarwal, M., Chandon, T. S. S., Karlsgodt, K.
H., Mori, S., & Malhotra, A. K. (2017). The role of myelination in measures
of white matter integrity: Combination of diffusion tensor imaging and two-­
photon microscopy of CLARITY intact brains. Neuroimage, 147, 253–261.
Chung, K., & Deisseroth, K. (2013). CLARITY for mapping the nervous sys-
tem. Nature Methods, 10(6), 508.
Chung, K., Wallace, J., Kim, S. Y., Kalyanasundaram, S., Andalman, A. S.,
Davidson, T. J.,… & Pak, S. (2013). Structural and molecular interrogation
of intact biological systems. Nature, 497(7449), 332–337.
CLARITY Resource Center. (2020). http://clarityresourcecenter.org (Accessed
June 20th, 2021).
Colaço, D. (2018). Rethinking the role of theory in exploratory experimenta-
tion. Biology and Philosophy, 33(5–6), 38.
Colaço, D. (2020). Recharacterizing scientific phenomena. European Journal
for Philosophy of Science, 10(2), 1–19.
Craver, C. F. (2007). Explaining the brain: Mechanisms and the mosaic unity of
neuroscience. Oxford: Oxford University Press.
Feest, U. (2005). Operationism in psychology: What the debate is about, what
the debate should be about. Journal of the History of the Behavioral Sciences,
41(2), 131–149.
Gabrieli, J. D. (1998). Cognitive neuroscience of human memory. Annual Re-
view of Psychology, 49(1), 87–115.
Gradinaru, V., Treweek, J., Overton, K., & Deisseroth, K. (2018). Hydrogel-tis-
sue chemistry: principles and applications. Annual Review of Biophysics, 47,
355–376.
Grillner, S., Kozlov, A., & Kotaleski, J. H. (2005). Integrative neuroscience:
Linking levels of analyses. Current Opinion in Neurobiology, 15(5), 614–621.
Harp, R., & Khalifa, K. (2015). Why pursue unification?: A social-­
epistemological puzzle. THEORIA. Revista de Teoría, Historia y Funda-
mentos de la Ciencia, 30(3), 431–447.
Huang, L., Kebschull, J. M., Furth, D., Musall, S., Kaufman, M. T., ­Churchland,
A. K., & Zador, A. M. (2018). High-throughput mapping of mesoscale con-
nectomes in individual mice. bioRxiv, 422477. doi: 10.1101/422477
Jorgenson, L. A., Newsome, W. T., Anderson, D. J., Bargmann, C. I., Brown,
E. N., Deisseroth, K.,… & Marder, E. (2015). The BRAIN Initiative:
Developing technology to catalyse neuroscience discovery. Philosophi-
cal Transactions of the Royal Society B: Biological Sciences, 370(1668),
20140164.
Koroshetz, W., Gordon, J., Adams, A., Beckel-Mitchener, A., Churchill, J.,
­Farber, G.,… & White, S. (2018). The state of the NIH BRAIN initiative.
Journal of Neuroscience, 38(29), 6427–6438.
Kotchoubey, B., Tretter, F., Braun, H. A., Buchheim, T., Draguhn, A., Fuchs,
T.,… & Rentschler, I. (2016). Methodological problems on the way to inte-
grative human neuroscience. Frontiers in Integrative Neuroscience, 10, 41.
Krohs, U. (2012). Convenience experimentation. Studies in History and Philos-
ophy of Science Part C: Studies in History and Philosophy of Biological and
Biomedical Sciences, 43(1), 52–57.
238 David J. Colaço
Leonelli, S. (2016). Data-centric biology: A philosophical study. Chicago, IL:
University of Chicago Press.
Leuze, C., Aswendt, M., Ferenczi, E., Liu, C. W., Hsueh, B., Goubran, M.,…
& McNab, J. A. (2017). The separate effects of lipids and proteins on brain
MRI contrast revealed through tissue clearing. Neuroimage, 156, 412–422.
Logothetis, N. K., Pauls, J., Augath, M., Trinath, T., & Oeltermann, A. (2001).
Neurophysiological investigation of the basis of the fMRI signal. Nature,
412(6843), 150–157.
Markram, H. (2012). The human brain project. Scientific American, 306(6),
50–55.
Mitchell, S. D., & Gronenborn, A. M. (2017). After fifty years, why are protein
X-ray crystallographers still in business? The British Journal for the Philoso-
phy of Science, 68(3), 703–723.
Molecular and Cellular Cognition Society. (2021) https://molcellcog.org/about-
mccs (Accessed June 20th, 2021).
O’Malley, M. A. (2013). When integration fails: Prokaryote phylogeny and the
tree of life. Studies in History and Philosophy of Science Part C: Studies
in History and Philosophy of Biological and Biomedical Sciences, 44(4),
551–562.
O’Malley, M. A., & Soyer, O. S. (2012). The roles of integration in molecular
systems biology. Studies in History and Philosophy of Science Part C: Stud-
ies in History and Philosophy of Biological and Biomedical Sciences, 43(1),
58–68.
Roy, D. S., Park, Y. G., Ogawa, S. K., Cho, J. H., Choi, H., Kamensky, L.,…
& Tonegawa, S. (2019). Brain-wide mapping of contextual fear memory en-
gram ensembles supports the dispersed engram complex hypothesis. bioRxiv,
668483. doi: 10.1101/668483
Shepherd, G. M., Mirsky, J. S., Healy, M. D., Singer, M. S., Skoufos, E., Hines,
M. S.,… & Miller, P. L. (1998). The human brain project: Neuroinformatics
tools for integrating, searching and modeling multidisciplinary neuroscience
data. Trends in Neurosciences, 21(11), 460–468.
Silva, A. J. (2003). Molecular and cellular cognitive studies of the role of synap-
tic plasticity in memory. Journal of Neurobiology, 54(1), 224–237.
Silva, A. J., Landreth, A., & Bickle, J. (2013). Engineering the next revolution
in neuroscience: The new science of experiment planning. Oxford: Oxford
University Press.
Sporns, O. (2014). Contributions and challenges for network models in cogni-
tive neuroscience. Nature Neuroscience, 17(5), 652–660.
Sullivan, J. A. (2016). Construct stabilization and the unity of the mind-brain
sciences. Philosophy of Science, 83(5), 662–673.
Sullivan, J. A. (2017). Coordinated pluralism as a means to facilitate integrative
taxonomies of cognition. Philosophical Explorations, 20(2), 129–145.
Ye, L., Allen, W. E., Thompson, K. R., Tian, Q., Hsueh, B., Ramakrishnan,
C.,… & Deisseroth, K. (2016). Wiring and molecular features of prefrontal
ensembles representing distinct experiences. Cell, 165(7), 1776–1788.
Yong, E. (2019). The human brain project hasn’t lived up to its prom-
ise. The Atlantic. https://www.theatlantic.com/science/archive/2019/07/
ten-years-human-brain-project-simulation-markram-ted-talk/594493/
11 Understanding Brain
Circuits
Do New Experimental
Tools Need to Address New
Concepts?
David Parker

A primary motive of science is to explain and to translate this insight to

practical uses. Neuroscience aims to explain the neural basis of normal
and abnormal behaviour and cognition. This is considered one of the
biggest open questions in science: although often related to conscious-
ness there is much to do below this level. The human brain consists of
billions of interconnected neurons: assuming that each neuron is either
active or inactive (a gross simplification), the number of potential brain
configurations would exceed the number of elementary particles in the
universe (Sagan 1977, p. 66).
Despite this complexity, there are significant claims to understanding,
including the cellular mechanisms of memory (Morris 2003) and of the
cerebral cortex sufficient for a conscious human brain emulation (always
seemingly “within 10 years”; see Miller 2011). There are also claims for
translation. An obvious translation is to artificial intelligence and robotics,
for example, to address Moravec’s paradox, that we can easily simulate
higher-level functions (e.g. memory) but not supposedly simpler senso-
rimotor processes. A less obvious translation is to “improve” the healthy
brain (cognitive enhancement), an extension of traditional claims of treat-
ments for neurological and also for psychiatric conditions. The former
should face serious caution from the poor history of the latter (Middleton
and Moncrieff 2019), which presumably reflects the lack of understanding
of what the interventions are actually doing. To illustrate the limited sci-
ence behind these claims, a 2005 report by the UK Office of Science and
Innovation suggested that by 2017 cognitive enhancement may be “part
of the knowledge professionals toolkit”. This failed prediction reflects its
poor scientific basis: one drug, donepezil which reduces the breakdown of
acetylcholine, was claimed to “produce more acetylcholine, allowing the
brain to work at a higher level of efficiency” (see Howard-Jones 2007),
while transcranial current stimulation was claimed to “make areas of the
brain work even harder” (Thomson 2010; my italics).
These positive claims were contrasted by Torsten Wiesel who said to
understand the brain “we need a century, maybe a millennium…we are

DOI: 10.4324/9781003251392-15
240 David Parker
at a very early stage of brain science” (cited in Horgan 1999). Gunther
Stent (1969) suggested that analysing the physiological processes of be-
haviour was pointless as they will “degenerate into seemingly ordinary
reactions no more and no less fascinating than those occurring in the
liver” echoing Charles Sherrington’s claim of a remoteness “between the
field of neurology and that of mental health …physiology has not enough
to offer about the brain in relation to mind to lend the psychiatrist much
help” (Sherrington 1951). There is thus a lack of connection between
the components of the nervous system and the outputs they produce.
Examining components or behaviours alone is not sufficient: molecular
and cellular analyses require knowledge of the behaviour or goal of the
system, while top-down behavioural approaches need lower-level insight
to constrain potential explanations.
While it seems implicit in neuroscience claims, understanding lacks an
accepted definition (de Regt 2013). For example, Bassett and Gazzaniga
(2011, p. 6) wrote, “Understanding the brain depends significantly on
understanding its emergent properties”, while Gregory claimed under-
standing will remove any appeal to emergence (in Blackmore 2006, p.
105). Bassett and Gazzaniga (p. 8) also say “To understand mind-brain
mechanisms it is necessary to characterize relations between multiple
levels of the multiscale human brain system, including interactions be-
tween temporal scales”, while Dennett (1971), Newell (1982), Marr
(1982), and Glass et al. (1979), all following Gilbert Ryle’s (1949) claim
that claim understanding can be obtained at different levels, the compu-
tational or behavioural/cognitive, the programme or algorithm, or the
physical or implementational level.
Understanding and explanation in biology are claimed to appeal to mech-
anisms rather than laws (Railton 1981; Machamer et al. 2000). Functional
concepts and mechanisms explain how a system and its parts do what they
do, while understanding requires that the functional concepts and expla-
nations are both intelligible and correct (Grimm 2006). The phlogiston
theory made certain phenomena intelligible and had practical uses, but
obviously couldn’t be claimed as an understanding of combustion.
In neuroscience, it is easier to say what is not sufficient for understand-
ing. Description is not enough. Claims that more details will explain
are illustrated by connectomic approaches (Schroter et al. 2017; Morgan
and Lichtman 2013) and that an understanding will follow from record-
ing “from ever more cells over larger brain regions” (Mott et al. 2018).
These reflect an “illusion of depth” (Ylikoski 2009) by assuming the
more we know the more we will understand. Second, while prediction
or the ability to make targeted interventions is important (Woodward
2017), we can predict effects and reliably manipulate systems without
explaining how they happen. A classic neurophysiological example is
the Jendrassik manoeuvre used clinically for over a century to potentiate
 Understanding Brain Circuits 241
reflexes (e.g. the knee-jerk reflex) but which still lacks explanation (Nar-
done and Schieppati 2008). Lesions provide another example. Spinal
cord lesions evoke predictable sensory and motor disturbances that are
not explained by the lesion alone, but also reflect diverse functional
changes above and below the lesion site. Emphasis on the lesion has fo-
cused remedial approaches on regeneration to repair the lesion, but this
has failed to translate into a treatment (Steward et al. 2012), presumably
because it doesn’t properly explain the changes.
Even if we can explain and understand how an effect occurs, this still
leaves the question of “why”. Sherrington (1899) claimed that neuro-
physiology can only answer “how” questions: analysing the activity of
all the components involved we could understand how we run, but not
why (to exercise, compete, escape?). Why questions are teleological and
represent the goal or purpose, a final causality replaced in reduction-
ist accounts by the efficient cause, the mechanical account of how an
effect occurred. But explaining behaviour requires knowing how and
why it occurred. For example, a complete description of “how” may
not determine why a person exhibits certain psychopathology (a faulty
gene or faulty environment?), and thus won’t identify the most effective
treatment.
Neural circuit analyses, while far from a panacea (Parker 2010), offers
a middle-ground between bottom-up and top-down analyses, by con-
sidering how components interact to generate outputs. Minimal criteria
must for circuit understanding include (Selverston 1980):

1 Characterisation of the circuit output and the associated behaviour.

Reductionist analyses physically reduce or constrain systems to
varying extents to enable cellular analyses, making direct links be-
tween circuit behaviour and behaviour unlikely (Krakauer et al.
2017): spinal cord fictive locomotion provides an example (Parker
and Srivastava 2013).
2 The component cells that generate the output must be identified.
A circuit neuron is defined as being active when the circuit is ac-
tive, and its activity influences the circuit output. However, neither
criterion unequivocally identifies a neuron as a circuit component
(Parker 2010).
3 The connectivity between network neurons must be defined. It is
clearly impractical to try to determine all cell-to-cell connections
of larger circuits by piecemeal recordings, analyses instead describe
connections between cell populations.
4 Identification of circuit neurons and connectivity describes the net-
work structure, but the structure does not determine function. We
must also characterise the functional properties of circuit neurons
and synapses.
242 David Parker
The challenges of reductionist approaches in even simpler systems have
been raised, both in terms of their practical and conceptual basis (Selver-
ston 1980; Parker 2010; Krakauer et al. 2017). A practical complication
of reductionist explanations is the astronomical number of components
to characterise (akin to the “tyranny of numbers” faced by computer
engineers before integrated circuits; Warner 2001). Using Bell’s number,
Koch (2012) claimed that characterising a 1,000 neuron network at ex-
aggerated rates would take 2000 years. Koch’s argument is a strawman
as he uses non-biological conditions (e.g. all to all connectivity), but the
same point was made less theatrically using actual circuit features by
Selverston (1980) and Getting (1989).
A circuit approach moves above the level of component cells by an-
alysing how cell populations interact to generate an output. But the
output of a circuit is not behaviour (Krakauer et al. 2017). Analo-
gies are often used: Krakauer et al. correctly say we won’t under-
stand chess from knowing what the board and the pieces are made of
(Krakauer et al. 2017), but circuit analyses are not (or should not be)
about the properties of the components, but about how component
properties generate outputs. Lateral inhibition in the horseshoe crab
retina (Hartline et al. 1956) provides an example of a functional ex-
planation from a component analysis, a building block (Getting 1989)
or motif common to diverse sensory systems. In the chess analogy,
this would be equivalent to knowing what a rook did rather than what
it was made of.
Failure to explain behaviour when we have a wealth of lower-level
knowledge is also claimed to illustrate the insufficiency of circuit analy-
ses. C. elegans is the prime example (Krakauer et al. 2017), but this can
be claimed to reflect the need for more analyses. While we know a lot
about the genetics and structure of its nervous system, functional prop-
erties, a key criterion for circuit understanding (see above) are poorly
understood. Just as proper attention to behaviour is needed (Krakauer
et al. 2017), proper attention to cellular data is also important. This
requires attention to component and circuit aspects that go beyond a
focus on experimentally convenient molecular or anatomical details or
large cells or input/output components (Selverston 1980; Parker 2010)
to focus on the less tractable and poorly characterised elements (e.g.
interneurons). It also requires a proper critique of claims rather than the
appeal to authority that seemingly lets hypotheses or assumptions pass
as characterisations, explanations, or understanding. Aplysia offers an
example of a claimed circuit explanation of behaviour (Kandel 2001)
from a very detailed experimental focus limited to tractable components
that do not logically justify understanding (Glanzman 2010; see Parker
2019). While less notable, but still in receipt of significant personal
awards, Grillner’s understanding of the lamprey spinal cord locomotor
circuit reflects a mass of assumption and extrapolation to cover missing
 Understanding Brain Circuits 243
data and conceptual errors (Parker 2006, 2010). Both claims fail, at least
from a scientific perspective.

1 Experimental Reductionism
Mechanistic explanations of behaviour in neuroscience typically reduce
behaviours to their constituent molecules and cells (Selverston 1980; Get-
ting 1989; Ito 2006; Yuste 2015). Appeals to reductionism include its track
record; that by opening “black boxes” it allows greater potential for con-
trol and practical use; and that it offers greater generality and simplicity.
Reductive explanations relate to a “machine model” (Monod 1972), where
interlocking parts combine to move the system from an initial to a goal
state, an explanation showing how the component parts and their organ-
isation achieve the system goal (Darden and Craver 2009). Hanahan and
Weinberg (2000) claimed, “Two decades from now…it will be possible to
lay out the complete integrated circuit of the cell… we will then be able
to apply the tools of mathematical analysis to explain”. The microproces-
sor analogy is unfortunate as Jonas and Kording (2017) have shown that
current neuroscience approaches are insufficient to explain actual micro-
processors. Hanahan and Weinberg’s 20 years have passed, and features
have been identified that negate the circuit-board analogy for even a single
cell: a fluid cytoskeleton, “intrinsically disordered proteins”, enzymes that
affect numerous substrates or perform non-enzymatic functions, pleomor-
phic molecular assemblies with “probability clouds” of interactions, and
stochastic and probabilistic gene expression (see Nicholson 2019).
Apart from clinical case studies, most biological discoveries reflect the
use of model organisms with features that make them useful. This relates
to Krogh’s principle (originally outlined by Claude Bernard in the 19th
century; Jørgensen 2001), that for any problem there an animal on which
it can be most conveniently studied. Between the 1930s and 1960s a wide
range of invertebrate and lower-vertebrate model systems were introduced
to determine the general principles of nervous system function (strongly
associated with the field of neuroethology). These have relatively simple be-
haviours and accessible nervous systems containing relatively small num-
bers of often large cells that allowed circuit characterisations related to
behaviours. These natural advantages can be engineered to some extent in
more complex (mammalian) systems using early developmental stages or
reduced preparations (tissue slices or cultured cells). While we cannot sim-
ply extrapolate, the hope is that the conservation of function in simpler or
reduced preparations will help us explain more complex or intact systems.
The reductionist approach is illustrated by the field of molecular and
cellular cognition, a ‘ruthless’ reductionism that claims to explain the
molecular basis of cognition from genetic manipulations (Bickle 2003).
To some extent this link is obvious: if a change in behaviour reliably fol-
lows a manipulation, then the manipulation caused the change. But this
244 David Parker
is a correlation, and useful for that, not an explanation. Molecules don’t
connect directly like switches to behaviours but work through cells in
circuits that form systems that link to behaviour. While we could detail
the molecular intervention and the resulting behaviour, explaining the
links between them requires knowing how the molecular manipulation
worked through different levels.
Two aspects removed by the reductionist neural approach are glial cells
and non-wired effects. We now know that glia serve more than the tradi-
tional developmental or “housekeeping” role (clearing away released trans-
mitters and extracellular potassium caused by action potential signalling):
the tripartite synapse concept sees glia as functional components that re-
ceive and send signals to neurons (Araque and ­Navarrete 2010). Non-wired
signals are not passed along axons and synaptic contacts but through the
extracellular space and include volume transmission (the diffusion of trans-
mitters up to mm from their point of release; Svensson et al. 2019) and
extracellular electrical fields (“ephaptic” signals; Weiss and Faber 2010).
To consider the issue, assume that field effects are important (Figure
1). We know they occur as they are measured in EEGs, but their func-
tional role has received relatively little attention. Field effects are not a
component in the conventional sense but reflect the summed activity in
cell populations, an example of an emergent biological effect. This gen-
erates an extracellular signal that can affect the activity of other cells
depending on its magnitude and spread, which are not easy to predict and
depend on the specific features of the activity and the extracellular space
(Weiss and Faber 2010), which in turn depends on the arrangement of
other neurons and the geometry of the extracellular space. Extracellular
fields are thus influenced by neuronal activity and neuronal activity by
extracellular fields, a circular interaction. Neuronal activity also alters the
diameter, and thus resistivity of the extracellular space (Østby et al. 2009)
to alter the magnitude and spread of field effects, neuronal activity, and
the extracellular space… generating a circular interaction that influences
another circular interaction. We could, in principle, understand how this
influences an output if we knew the contribution of each cell to the extra-
cellular signal, and how it spreads through the extracellular space to alter
neuronal activity and the extracellular space, but this would at best pro-
vide a snapshot of constantly evolving dynamic activity. This argument
could be negated by begging the question and saying that field effects are
an unimportant epiphenomenon. But similar considerations apply to vol-
ume transmission and to circular interactions generated by conventional
“wired” feedback connections between neurons (Parker 2019; Svensson
et al. 2019) which cannot be easily dismissed to appeal to simple accounts.
It isn’t that we want or need knowledge of the minutiae of wired and
non-wired components and connections evolving in real time to explain
(Greenberg and Manor 2005, provide a sobering example of the failure to
explain when detail exceeds a small limit in a very simple system), but we
need to appreciate the role of these effects in any explanation.
 Understanding Brain Circuits 245

Figure 11.1 C
 ircular interactions along wired and non-wired pathways. (a) Neu-
ronal activity generates extracellular signals (shaded circle), that
can alter neuronal excitability. At the synaptic terminal, inputs are
evoked in postsynaptic cells and in glial cells (G), and glia can signal
back to neurons. Glial cells form a syncytium through gap junction
connections (dashed lines), which allows local activity to spread. (b)
Neuronal activity can cause swelling of glial cells to ‘shrink’ the
extracellular space and change the extracellular field, neuronal and
glial activity, transmitter release and glial cell activity, which alters
the extracellular space…, generating multiple circular interactions.
(c) These are known features of nervous systems, nothing is exag-
gerated, and if anything is too simplified. (d) Adding plasticity (both
­activity-dependent (ADP) and neuromodulator-evoked (NMod),
and the interactions between these effects) evokes additional circular
interactions. Even if complete (‘Laplacian’) detail was possible, we
would only have snapshots of constantly evolving dynamic activity.

An issue with constitutive reductionism is that each level of a biolog-

ical system establishes an order that is not necessarily reflected at other
levels. Sperry (1980, p. 201) suggested that

Once generated from neural events, the higher order mental patterns
and programs have their own subjective qualities and progress, op-
erate and interact by their own causal laws and principles which are
different from and cannot be reduced to those of neurophysiology.

We can stay far below the level of mental events to see this. A sodium
channel has atoms and molecules arranged to allow movement of so-
dium ions underlying an action potential; but an action potential is not
reducible to the sodium channel as it reflects multiple ion channels,
membrane pumps, and the membrane itself, that together generate the
voltage and electrochemical gradients needed for the sodium channel
to work.
246 David Parker
The reductionist focus on manipulating single system components
in isolation is really only possible in decomposable or “nearly decom-
posable” systems where interactions between parts are minimal, not in
systems where the interactions between parts are many and strong (“non-­
decomposable systems”; Simon 1962). Bassett et al. (2010) refer to Simon
(1962) in saying that hierarchical systems are nearly decomposable to claim
parts can be examined relatively independently. But Simon didn’t say that
hierarchical systems are nearly decomposable, just that “At least some
kinds of hierarchic systems can be approximated successfully as nearly
decomposable systems” (Simon 1969, p. 474). Near-­decomposability is
an assumption of reductionist approaches, a “fallible” heuristic (­Wimsatt
2006), and assumptions need to be considered. Non-decomposability
could reflect a temporary failure of experimental or analytical techniques
or concepts (e.g. Bechtel 2002) that disappears when we have the correct
concepts and details. But is this begging the question to maintain the
assumption of near-decomposability, a “promise of jam tomorrow”? It
is not enough to inductively claim that many major advances in the life
sciences have stemmed from the discovery of ways of decomposing a phe-
nomenon: that there are nearly-decomposable systems in biology does not
mean that this applies to all biological systems; we may need to appreciate
this to find ways of approaching these systems.
Non-decomposable systems have interactions between components
that are many and strong. This seems to define nervous systems where
feedforward, feedback and circular interactions are endemic, a heter-
archical rather than a hierarchical organisation of component parts
performing specific functions in fixed sequences. Heterarchical systems
offer challenges to reductionist approaches. Properties examined in re-
duced preparations or under quiescent conditions can change when the
system is active due to the recruitment of wired and non-wired feedback
loops or activity-dependent changes in component or system properties.
Heterarchical systems also show equifinality (many mechanisms gen-
erate the same output) and multifinality (single components influence
multiple outputs). Functional imaging has shown that instead of a one-
to-one mapping between a region and a function there is a high degree of
overlap among regions that are activated by tasks that share no cognitive
components (Schroeder and Foxe 2005). A given region can thus influ-
ence multiple functions. Price and Friston (2005) claim this will be de-
termined by its connectivity, but this ignores non-wired signals. The lack
of obligatory mechanistic sequences means the law of transitivity does
not hold in heterarchical systems, complicating the spatial and temporal
relations between levels, and meaning that explanations cannot simply
build up from lower-level properties. Noble (2012) provides an example
from the failure to explain the heart rhythm from lower-level properties.
In non-decomposable systems any manipulation, no matter how
“surgical”, will also necessarily affect other components through
 Understanding Brain Circuits 247
diaschisis (Carrera and Tononi 2014), an acknowledged neurological
term that seems less appreciated experimentally. Diaschisis literally
means “shocked throughout” to represent the widespread changes in the
brain caused by a focal lesion. Any manipulation, no matter how elegant
and refined, will necessarily affect downstream (and with feedback, up-
stream) components. Terms like “specific” or “targeted” imply precision
that may be justified from intention (a specific single component was
successfully targeted), but not from application (many components will
be affected). To offer more than a correlation, and more is often implied,
requires knowing how the manipulation alters the system. A manipula-
tion may cause an effect without explaining it.
Reductionist explanations also aim to define individual parameters.
For invertebrates, these can be single uniquely identifiable cells but in
vertebrates, they are cell populations, characterisation reflecting a popu-
lation average. Variability is endemic in these populations (Soltesz 2006).
Claude Bernard claimed that averages “confuse while aiming to unify,
and distort when aiming to simplify” (Bernard 1865, pp. 134–135). An
average value may not describe a parameter, especially if the sample size
is low, as the average value may not have been seen in any measure-
ment: would this be considered characterisation of a component? Aver-
aging also assumes that the variability is noise, when it may be a signal
or reflect the presence within a population or functional sub-groups
(Soltesz 2006). Variability does not rule out conventional reductionist
approaches but requires larger sample sizes to ensure that we properly
characterise components, which for small or inaccessible cells and syn-
aptic connections adds a significant burden.
Finally, reductionism assumes substantivalism, that components are
defined by their intrinsic characteristics (e.g. structure, location, or
transmitter content), rather than by their relationship to other compo-
nents. But a transmitter like GABA (or a neuron containing it) is not
intrinsically inhibitory, it is defined as such from the receptor it binds to
and the receptor’s inhibitory effect on the postsynaptic cell. Even then
the functional circuit effect could be excitation through disinhibition.
Highlighting these limits of reductionism does not invite the oppo-
site view prevalent with dichotomous thinking, that lower-level details
are irrelevant. Regions of the brain show specificity in the types of cells
they contain (from the 47 areas in Brodmann’s map of a century ago
to 98 now; Glasser et al. 2016), the specific anatomy, wiring patterns,
and functional properties suggesting that these details are important. If
everything could be reduced to Hopfield network representations then
the brain would be a collection of identical units that simply process dif-
ferent types of information and relay it to different areas (an argument
could be made for this arrangement in the cerebellar cortex). This par-
adoxically seems to be the view of the Human Brain Project, despite its
explicit focus on reductionist detail (Markram et al. 2015), that chaining
248 David Parker
together multiple simulated cortical columns will produce a fully con-
scious brain emulation. If we want to understand mechanisms, and for
effective and safe translations we should, we need these details.

2 New Tools
Neuroscience tools offer greater control and give more direct results at
lower-levels or scales, supporting a reductionist focus (compare the anal-
ysis of a protein on a gel with an fMRI scan). In practice, the development
of theories and instruments are mutually dependent: we develop tools
and techniques to perform specific analyses, and analyses are directed
by the available tools. The instruments and their use should be theory-­
neutral: a measurement obtained with a tool will be the same irrespective
of the theory under which it was used. However, how a tool is applied
and a measurement interpreted can be affected by a theory. The power
of newer technologies to focus on biological details can leave behaviour
as an “afterthought” (Krakauer et al. 2017), leading to explanations of
behaviour that fail to separate causation from correlation and thus gen-
erating erroneous assumptions of links between lower and higher levels.
New techniques are claimed to have caused “revolutions” in neuro-
science (“In short: understanding tool development is the key to under-
standing real revolutions in actual neuroscience”; Bickle 2016). Peter
Galison’s “Image and Logic” (1997) gives a view of scientific history
dominated by tools: steam-engine technology came before thermody-
namics and telegraphy and telephony came before information theory.
While Einstein, Heisenberg, Schrödinger and Dirac believed that prog-
ress in physics would continue through conceptual insight, experimen-
tal physics in the mid-20th century focused on new tools. Neuroscience
seems to have a similar dichotomy between the need for concepts or
techniques and tools.
It is easy to see new tools as revolutionary if a revolution is defined as a
fundamental change in the way things are done. This wouldn’t fit Thomas
Kuhn’s (1962) definition of a scientific revolution, “a noncumulative de-
velopmental episode in which an older paradigm is replaced in whole
or in part by an incompatible new one” (SSR, p. 92). A new technique
may revolutionise how or what experiments are done but would only
cause a scientific revolution if it led to a paradigm shift, not by increasing
precision or allowing novel analyses within a paradigm (this is “normal
science”; Kuhn 1962; Parker 2018). The terms scientific revolution and
paradigm shift are often used erroneously (e.g. Knafo and Wyart 2015),
presumably to emphasise something above “normal science” (seeing this
as a pejorative term also misunderstands its meaning; Parker 2018).
The technological aspect of neuroscience is expressed in the aims of
the US BRAIN (Brain Research Through Advancing Innovative Neu-
rotechnologies®) project, which promises technology “to produce a
 Understanding Brain Circuits 249
revolutionary new dynamic picture of the brain that, for the first time,
shows how individual cells and complex neural circuits interact in both
time and space”, and to find “new ways to treat, cure, and even prevent
brain disorders”. It doesn’t say how it will do this, seemingly beyond
examining more components in greater detail. Understanding the ner-
vous system comes from observation, both of brain and behaviour, and
the manipulation of its component parts either by accident (injuries or
neurological disorders) or experimental design.
Neuroscience research has traditionally been slow but high through-
put techniques are accelerating analyses (from automated “behavioural”
ethoscopes to high throughput anatomical, molecular and electrophysi-
ological analyses). But while this data gives knowledge, knowledge does
not necessarily explain.
In summarising neural circuit analyses, Selverston (1980) said that un-
derstanding would be limited to the simplest systems with the techniques
then available: these were intracellular and extracellular recording and
stimulation, various anatomical approaches combined with cell stains,
electron microscopy, the use of physical lesions, pharmacological agents,
the ability to kill labelled cells with UV light, and initial attempts at im-
aging. The range of tools has now increased markedly (see https://www.
biomedcentral.com/collections/ntfn). For example:

1 Single-cell transcriptomics allows the classification of cell types

beyond classical aspects like location, anatomy or some functional
property.
2 Molecular genetic approaches, for example using GAL4 driver lines
or the Cre-Lox allow specific cells to be labelled with fluorescent
markers under the control of specific promoters; promoter-driven
gene knock-outs or knock-ins can modify the function or develop-
ment of cells (e.g. using genetically encoded toxins like tetanus toxin
targeted to specific promoters, or KillerRed which generates condi-
tional toxic products when illuminated); and genetically engineered
expression of voltage or calcium reporters using specific promoters
to examine certain cell types or a pan-neuronal promoter for whole-
brain imaging (e.g. GCaMP, a fusion of calmodulin and GFP), with
latest developments offering a combination of both high sensitivity
and rapid responses (Zhang et al. 2020).
3 Optogenetics allows the rapid control of excitability by genetically
expressing light-sensitive proteins that depolarise or hyperpolar-
ise cells, offering the advantage of rapid loss or gain of function in
specific cells depending on the promoter used or region illuminated,
while chemogenetics genetically expresses receptors (DREADDS) in
cells to allow for their selective control by ligands novel to the system.
4 Connectomics uses high-throughput electron microscopy to charac-
terise cells, axonal processes, and synaptic connections, a “Google
250 David Parker
map” of the brain (Morgan and Lichtman 2013). The laborious
preparation and photographing of sections have been automated,
the bottleneck now being the reconstruction of the tissue. This can
be facilitated by expressing fluorescent proteins in cells (“Brainbow”
techniques, where Cre-Lox recombination results in a random com-
bination of different fluorescent proteins) to help trace processes
through tissue.
5 An aim of the BRAIN project is to examine multiple neurons at the
same time (Mott et al. 2018), to overcome the bottleneck of conven-
tional electrophysiology of examining, at best, a handful of neurons
at a time. Neuropixel probes allow thousands of neurons to be re-
corded in cortical and deeper brain structures in intact behaving
animals (Steinmetz et al. 2021).

These are all valuable additions to circuit research. But no tool, new
or traditional, is without caveats. The neural imaging field shows the
importance of highlighting caveats. Imaging was developed in the 1970s
(Selverston 1980), but highlighting dissatisfaction with its poor spatial
and temporal resolution led to successive improvements in reporters, mi-
croscopes, and analyses to the point where signals can now be imaged
with millisecond precision in multiple cells (Zhang et al. 2020; although
temporal precision falls as the number of cells imaged increases).
Many new tools offer significant control at the molecular and ­single-cell
level, making genetic tractability an advantage of a model system. Ge-
netically tractable models (Drosophila, C. elegans, zebrafish, mouse)
now dominate classical model systems (paradoxically, genetically tracta-
ble systems are less amenable to neurophysiological approaches). Instead
of the neuroethological approach of using diverse systems suited to ad-
dressing specific questions, we may be choosing questions to address in
genetically tractable systems, an example of the “law of the instrument”.
Before molecular genetic approaches, single-cell recording techniques
also promoted a reductionist explanatory and analytical focus of inver-
tebrate and vertebrate neural circuits (Selverston 1980; Getting 1989;
Ito 2006); the mechanisms of perception, learning and memory (Kandel
2001; Ito 2006; Bliss and Collingridge 2013).
New techniques often claim to overcome the limitations of previous
techniques, and they often do, but this doesn’t mean that they don’t
have caveats. Take molecular genetic approaches: positive or negative
effects of manipulations using these tools do not unequivocally identify
or eliminate a neuron as part of a circuit given the issues of heterarchical
systems discussed above (Parker 2019). Interpretation of effects, of both
molecular genetic and optogenetic approaches, is complicated by effects
on multiple cell types (“leaky” expression; Allen et al. 2015): ideally, a
single component at a time is affected so any effect can be related to that
component. Claims that molecular genetic approaches could “dissect”
 Understanding Brain Circuits 251
neural circuits (Kiehn and Kullander 2004), language that suggests sur-
gical precision, are negated by the promiscuity of supposedly selective
targeting: Gosgnach et al. (2006) claimed ‘selective ablation’ of one class
of spinal cord neurons but in the same paragraph say that two other
classes were increased and another reduced in number, a poor dissec-
tion. If multiple cells are affected then any explanation has to appeal to
all effects, not beg the question by dismissing unintended effects using
terms like “small increase” or “slight decrease” (Gosgnach et al. 2006).
Specific targeting of neuron classes is possible using combinatorial meth-
ods, for example using the regulatory elements of two neurally expressed
genes (Luan et al. 2020): INTERSECT (intron recombinase sites en-
abling combinatorial targeting) utilizes multiple recombinases expressed
in different cells type to more specifically target cell populations. Op-
togenetic probes can also be activated by activity-dependent immediate
early genes, increasing specificity by limiting targeting to recently active
cells (Guru et al. 2015).
There are issues even if a single population of cells is targeted. One
issue is that it doesn’t consider population variability. A knock-out or
optogenetic manipulation will affect all members of the population,
essentially averaging the effect. Whether this is acceptable or not de-
pends on the view of variability: if it is considered irrelevant or noise
then it doesn’t matter, but if variability is functionally relevant then it
does. In addition, genetic manipulations can result in adaptive changes
within minutes (i.e. not prevented by conditional knock-outs; Frank
et al. 2006). Homeostatic plasticity may lead to compensation that re-
sults in no apparent phenotype despite a key component being affected,
while diaschisis will add to the effect of the intended manipulation and
complicate interpretations of what caused a change (Parker 2019). Op-
togenetic stimulation can also evoke effects in the absence of optogenetic
proteins, possibly reflecting light-induced temperature changes (Allen
et al. 2015). These artefacts can presumably be controlled for (if they
are admitted as caveats; issues are not always highlighted as clearly as in
Allen et al. 2015); but diaschisis cannot be removed as it reflects normal
features of nervous systems, which means in this case that the potential
contribution of these factors to any explanation needs to be considered.
A connectomic map is a very desirable feature, and knowledge of
structure is important (ephaptic communication in the escape circuit of
goldfish is an example of a structural feature that explained a functional
phenomenon; Weiss and Faber 2010). But function must be determined,
it doesn’t drop out from structure despite claims that structure “may
enable predictions of circuit behaviour” (Lichtman and Sanes 2008,
p. 349) and that structure “will signify a physiological process without
the requirement of repeating the physiological analysis” (Morgan and
Lichtman 2013, p. 496). Evidence from many systems over many years
suggests a very limited ability to infer function from structure. This is
252 David Parker
evidenced by C. elegans, where a complete connectome has been avail-
able for over 30 years but we still lack insight into how the circuit works,
and the cerebellar cortex where the organisation has been known for de-
cades without insight into how the cerebellum does what it does, or even
what it is doing. This was very elegantly shown by the marked change
in output of a circuit of just two neurons depending on the functional
properties of the connections between them (Elson et al. 2002).
Although connectomic analyses can be automated, it still requires sig-
nificant effort. A recent connectome of a portion of the Drosophila brain
(around 25,000 neurons) took two years and hundreds of thousands of
hours on a paper with around 100 authors (Scheffer et al. 2020). Scal-
ing this up to larger systems may require a disproportionately greater
time. Connectomic detail is desirable but given the limited functional
insight is the time investment on detailed maps useful (1 mm3 of visual
cortex took six months to obtain the data (Landhuis 2020); considering
the volume of the visual cortex is 6 cm3 the time needed for a complete
connectomic map is biblical). A less detailed structure with functional
concepts may be a better approach.

3 Conclusion
Technological advances are needed to address the astronomical com-
plexity of even modestly sized networks. But conceptual advances are
needed to direct analyses and tool development; we need to know what
we need to know, and how to achieve it. Neuroscience has an enor-
mous amount of data on molecular, cellular, synaptic, and developmen-
tal mechanisms, and cognition and behaviour, but we lack a coherent
framework to link between these aspects. It isn’t that we need, or should
expect, a unified theory of brain function, just some idea of how to in-
tegrate between levels.
Complete reductionist detail of even modest circuits is not possible
with current techniques, it may never be, and may not be needed. Vari-
ous people have highlighted different explanatory levels (Ryle, Newell,
Marr, Dennett; see above), the semantic and computational, the pro-
gramme or algorithmic, and reductionist physical or implementational
approaches. Non-reductionist accounts of nervous systems include Lash-
ley’s equipotentiality hypothesis, Pribram’s holonomic brain theory, and
arguably most successfully in linking across levels cybernetic control
principles (Pickering 2010). For explanation appeal to a computational
or algorithmic level is reasonable, but is it enough for interventions? You
may know the ‘computations’ that a car engine performs, but you would
still want a mechanic who knows about the details of engines to fix
one. Neuroscience may need to introduce something similar to electro-
magnetic charge and electromagnetic forces as new fundamental phys-
ical properties to explain electromagnetism. While reductionism can
 Understanding Brain Circuits 253
appeal to successes, we shouldn’t assume its success. A key issue is that
although biological components operate according to physical laws, they
reflect organisation over various spatial and temporal scales rather than
independent parts (field effects may be the obvious example), not forget-
ting that the nervous system is in a body embedded in an environment.
Critiques of neuroscience reductionist approaches point to its failure
to explain, but current failure doesn’t necessarily mean that the ap-
proach is misguided. The Rosetta Stone was useful before it was decoded
(and it wasn’t decoded by having more of it or a reductionist analysis
of the stone). The same could apply to the relevance of lower-level (or
other) properties. For the nervous system, we are still in the early days
of understanding a complex system. We can continue to use statistical
approaches to explain higher-level phenomena from the mass action of
identical parts or abstractions that focus on representations or computa-
tions, while also considering the details in the diversity of brain regions,
neurons and synapses.
Even techniques that promise to examine multiple components (e.g.
the US BRAIN project; Mott et al. 2018) still examine components, and
there seems to be a drive to record from ever more cells. A recent review
states that we should now shift from considering individual neurons to
considering ensembles of neurons as the functional units of the nervous
system, and that new technologies will provide a greater understand-
ing of the link between the brain and behaviour (Yuste 2015), without
saying what links are needed and how to develop them. We are already
data rich but theory light, and assuming more components will give un-
derstanding is an illusion (Ylikoski 2009). Analyses at higher levels can
generate plausible inferences, but plausible is not necessarily correct.
Neuroscience papers increasingly use multiple methods, ‘a method-
ological decathlon’ seemingly to emphasise the importance of the work
(Krakauer et al. 2017). Using different methods can produce more robust
findings if the results of each are in agreement, and it avoids the dan-
ger of being biased towards certain factors and blind to others. But the
techniques should genuinely address questions rather than “throw” tech-
niques at a system. The prestige journals seem especially prone to multiple
techniques. There is a danger of multiple pieces of neuroscience informa-
tion seeming to make explanations more satisfying (Weisberg et al. 2008;
McCabe and Castel 2008), and thus interfering with proper critique.
Unless halted we may regret the move away from the traditional neu-
rophysiological/neuroethological approaches that explicitly address the
neurophysiology of natural behaviours in various model systems (ad-
dressing some of the concerns raised in Krakauer et al. 2017): we will
not simply lose the insight it can give, but also the insight it has brought.
This has always partly reflected a split between invertebrate and verte-
brate/mammalian communities (after moving to a vertebrate lab for a
post-doc after a PhD on insects I heard that neuroethology was “people
254 David Parker
doing weird things with insects”), and now reflects the appeal to geneti-
cally tractable systems.
Instead of philosophical and scientific competition over the merits of
lower-level reductionist or higher-level computational or representative
effects, they should be seen as different approaches to the same ques-
tion, not only as an epistemological diversity but also as an ontological
unity. As Simon (1962) said, “In the face of complexity, an in-principle
reductionist may be at the same time a pragmatic holist”. We can say the
same about tools and ideas. The history of neuroscience shows we need
both. We need to discuss concepts to identify what we actually need
to do and how to use and develop tools to do this. An important, and
seemingly relatively simple and non-technologial start is to stop allow-
ing trivial claims to characterisation and understanding by prominent
figures (Glanzman 2010; Parker 2006, 2019).

References
Allen, B. D., Singer, A. C., & Boyden, E. S. (2015). Principles of designing
interpretable optogenetic behavior experiments. Learning & Memory, 22,
232–238.
Araque, A., & Navarrete, M. (2010). Glial cells in neuronal network function.
Philosophical Transactions of the Royal Society London B: Biological Sci-
ences, 365, 2375–2381.
Bassett, D. S., & Gazzaniga, M. (2011). Understanding complexity in the hu-
man brain. Trends in Cognitive Science, 15, 200–209.
Bassett, D. S., Greenfield, D. L., Meyer-Lindenberg, A., Weinberger, D. R.,
Moore, S. W., & Bullmore, E. T. (2010). Efficient physical embedding of to-
pologically complex information processing networks in brains and computer
circuits. PLoS Computational Biology, 6(4), e1000748.
Bechtel, W. (2002). Decomposing the brain: a long-term pursuit. Brain and
Mind, 3, 229–242.
Bernard, C. (1865). An Introduction to the Study of Experimental Medicine
(H. C Greene (1949), Trans.). New York: Henry Schuman Inc.
Bickle, J. (2003). Philosophy and Neuroscience: A Ruthlessly Reductive Ac-
count. Dordrecht: Springer.
Bickle, J. (2016). Revolutions in neuroscience: tool development. Frontiers in
Systems Neuroscience, 10, 24.
Blackmore, S. (2006). Conversations on Consciousness: What the Best Minds
Think about the Brain, Free Will, and What It Means to Be Human. New
York: Oxford University Press.
Bliss, T. V., & Collingridge, G. L. (2013). Expression of NMDA receptor-­
dependent LTP in the hippocampus: bridging the divide. Molecular Brain, 6, 5.
Carrera, E., & Tononi, G. (2014). Diaschisis: past, present, future. Brain, 137,
2408–2422.
Darden, L., & Craver, C. (2009). Reductionism in biology, Encyclopedia of Life
Sciences. Chichester: J. Wiley and Sons Ltd.
 Understanding Brain Circuits 255
Dennett, D. C. (1971). Intentional systems. The Journal of Philosophy, 68,
87–106.
de Regt, H. W. (2013). Understanding and explanation: living apart together?
Studies in History and Philosophy of Science Part A, 44(3), 505–509.
Elson, R. C., Selverston, A. I., Abarbanel, H. D. I., & Rabinovich, M. I. (2002).
Inhibitory synchronization of bursting in biological neurons: dependence on
synaptic time constant. The Journal of Neurophysiology 88(3), 1166–1176.
Frank, C., Kennedy, M., Goold, C., Marek, K., & Davis, G. (2006). Mecha-
nisms underlying rapid induction and sustained expression of synaptic ho-
meostasis. Neuron, 52, 663–677.
Galison, P. (1997). Image and Logic: A Material Culture of Microphysics.
Chicago, IL: University of Chicago Press.
Getting, P. (1989). Emerging principles governing the operation of neural net-
works. Ann. Rev. Neurosci., 12, 185-204.
Glanzman, D. L. (2010). Common mechanisms of synaptic plasticity in verte-
brates and invertebrates. Current Biology, 20(1), R31–R36.
Glass, A. L., Holyoak, K. J., & Santa, J. L. (1979). Cognition. Reading, MA:
Addison-Wesley Publishing Company.
Glasser, M. F., Coalson, T. S., Robinson, E. C., Hacker, C. D., Harwell, J.,
Yacoub, E., Ugurbil, K., Andersson, J., Beckmann, C. F., Jenkinson, M.,
Smith, S. M., Van Essen, D. C. (2016). A multi-modal parcellation of human
cerebral cortex. Nature, 536, 171–178.
Gosgnach, S., Lanuza, G. M., Butt, S. J. B., Saueressig, H., Zhang, Y., Vel-
asquez, T., Riethmacher, D., Callaway, E. M., Kiehn, O., & Goulding, M.
(2006). V1 spinal neurons regulate the speed of vertebrate locomotor outputs.
Nature, 440, 215–219.
Greenberg, I., & Manor, Y. (2005). Synaptic depression in conjunction with
A-current channels promote phase constancy in a rhythmic network. The
Journal of Neurophysiology, 93, 656–677.
Grimm, S. R. (2006). Is understanding a species of knowledge? The British
Journal for the Philosophy of Science, 57, 515–535.
Guru, A., Post, R. J., Ho, Y.-Y., & Warden, M. R. (2015). Making sense of
optogenetics. The International Journal of Neuropsychopharmacology, 18,
pyv079.
Hanahan, D., & Weinberg R. A. (2000). The hallmarks of cancer. Cell, 100,
57–70.
Hartline, H. K., Wagner, H. G., & Ratliff, F. (1956). Inhibition in the eye of
limulus. The Journal of General Physiology, 39, 651–673.
Horgan, J. (1999). The Undiscovered Mind. London: Widenfeld and Nicholson.
Howard-Jones, P. A. (2007) Neuroscience and Education: Issues and Oppor-
tunities, TLRP Commentary. London: Teaching and Learning Research
Programme.
Ito, M. (2006). Cerebellar circuitry as a neuronal machine. Progress in Neuro-
biology, 78(3), 272–303.
Jonas, E., & Kording, K. P. (2017). Could a neuroscientist understand a micro-
processor? PLoS Computational Biology, 13(1), e1005268.
Jørgensen, C. B. (2001). August Krogh and Claude Bernard on basic principles
in experimental physiology. BioScience, 51(1), 59–61.
256 David Parker
Kandel, E. (2001). The molecular biology of memory storage: a dialogue be-
tween genes and synapses. Science, 294, 1030–1038.
Kiehn, O., & Kullander, K. (2004). Central pattern generators deciphered by
molecular genetics. Neuron, 41, 317–321.
Koch, C. (2012). Modular biological complexity. Science, 337, 531–532.
Knafo, S., & Wyart, C. (2015). Optogenetic neuromodulation: new tools for
monitoring and breaking neural circuits. Annals of Physical and Rehabilita-
tion Medicine, 58, 259–264.
Krakauer, J. W., Ghazanfar, A. A., Gomez-Marin, A., MacIver, M. A., & Poep-
pel, D. (2017). Neuroscience needs behavior: correcting a reductionist bias.
Neuron, 93(3), 480–490.
Kuhn, T. (1962). The Structure of Scientific Revolutions, 1st edn. Chicago, IL:
University of Chicago Press.
Landhuis, E. (2020). Probing fine-scale connections in the brain. Nature, 586,
631–633.
Lichtman, J., & Sanes, J. (2008). Ome sweet ome: what can the genome tell us
about the connectome? Current Opinion in Neurobiology, 18, 346–353.
Luan, H., Kuzin, A., Odenwald, W. F., & White, B. H. (2020). Cre-assisted
fine-mapping of neural circuits using orthogonal split inteins. eLife, 9, e53041.
Machamer, P., Darden, L., & Craver, C. F. (2000). Thinking about mecha-
nisms. Philosophy of Science, 67, 1–25.
Markram, H., Muller, E., Ramaswamy, S., Reimann, M. W., Abdellah, M.,
Sanchez, C. A., … & Schurmann, F. (2015). Reconstruction and simulation
of neocortical microcircuitry. Cell, 163, 456–492.
Marr, D. (1982). Vision: A Computational Investigation into the Human Rep-
resentation of Visual Information. New York: W.H. Freeman & Company.
McCabe, D. P., & Castel, A. D. (2008). Seeing is believing: the effect of brain
images on judgments of scientific reasoning. Cognition, 107(1), 343–352.
Middleton, H., & Moncrieff, J. (2019). Critical psychiatry: a brief overview. BJ
Psychiatry Advances, 25, 47–54.
Miller, G. (2011). Blue brain founder responds to critics, clarifies his goals. Sci-
ence, 334, 748–749.
Monod, J. (1972). Chance and Necessity: An Essay on the Natural Philosophy
of Modern Biology. New York: Vintage.
Morgan, J. L., & Lichtman, J. W. (2013). Why not connectomics? Nature Meth-
ods, 10, 494–500.
Morris, R. G. M. (2003). Long-term potentiation and memory. Philosophical
Transactions of the Royal Society of London. Series B: Biological Sciences,
358(1432), 643–647.
Mott, M. C., Gordon, J. A., & Koroshetz, W. J. (2018). The NIH BRAIN ini-
tiative: advancing neurotechnologies, integrating disciplines. PLoS Biology,
16(11), e3000066.
Nardone, A., & Schieppati, M. (2008). Inhibitory effect of the Jendrassik ma-
neuver on the stretch reflex. Neuroscience, 156(3), 607–617.
Newell, A. (1982). The knowledge level. Artificial Intelligence, 18, 87–127.
Nicholson, D. J. (2019). Is the cell really a machine? The Journal of Theoretical
Biology, 477, 108–126.
Noble, D. (2012). A theory of biological relativity: no privileged level of
causation. Interface Focus, 2, 55–64.
 Understanding Brain Circuits 257
Østby, I., Øehaug, L., Einevoll, G., Nagelhus, E., Plahte, E., Zeuthen, T., Lloyd,
C., Ottersen, O., & Omholt, S. (2009). Astrocytic mechanisms explaining
neural-activity-induced shrinkage of extraneuronal space. PLoS Computa-
tional Biology, 5, e1000272.
Parker, D. (2006). Complexities and uncertainties of neuronal network func-
tion. Philosophical Transactions of the Royal Society B: Biological Sciences,
361, 81–99.
Parker, D. (2010). Neuronal network analyses: premises, promises and uncer-
tainties. Philosophical Transactions of the Royal Society London B: Biologi-
cal Sciences, 365, 2315–2328.
Parker, D. (2018). Kuhnian revolutions in neuroscience: the role of tool develop-
ment. Biology and Philosophy, 33, 17.
Parker, D. (2019). Psychoneural reduction: a perspective from neural circuits.
Biology & Philosophy, 34(4), 44. doi:10.1007/s10539-019-9697-8.
Parker, D., & Srivastava, V. (2013). Dynamic systems approaches and levels of
analysis in the nervous system. Frontiers in Physiology, 4, 15.
Pickering, A. (2010). The Cybernetic Brain: Sketches of Another Future. Chi-
cago, IL: University of Chicago.
Price, C. J., & Friston, K. J. (2005). Functional ontologies for cognition: the
systematic definition of structure and function. Cognitive Neuropsychology,
22, 262–275.
Railton, P. (1981). Probability, explanation, and information. Synthese, 48,
233–256.
Ryle, G. (1949). The Concept of Mind. Chicago, IL: University of Chicago Press.
Sagan, C. (1977). Dragons of Eden: Speculations on the Evolution of Human
Intelligence. New York: Random House.
Scheffer, L. K., Xu, C. S., Januszewski, M., Lu, Z., Takemura, S.-Y., Hayworth,
K. J., Huang, G. B., … & Plaza, S. M. (2020). A connectome and analysis of
the adult Drosophila central brain. eLife, 9, e57443 C57441.
Schroeder, C., & Foxe, J. (2005). Multisensory contributions to low-level, 'uni-
sensory' processing. Curr Opin Neurobiol, 4, 454-458.
Schroter, M., Paulsen, O., & Bullmore, E. T. (2017). Micro-connectomics:
probing the organization of neuronal networks at the cellular scale. Nature
Reviews Neuroscience, 18, 131.
Selverston, A. (1980). Are central pattern generators understandable? Behav-
ioral Brain Sciences, 3, 535–571.
Sherrington, C. S. (1899). On the relation between structure and function as ex-
amined in the arm. Transactions of the Liverpool Biological Society, 13, 1–20.
Sherrington, C. S. (1951). Man on His Nature: The Gifford Lectures, Edin-
burgh, 1937–8. Cambridge: Cambridge University Press.
Simon, H. (1962). The architecture of complexity. Proceedings of the American
Philosophical Society, 106, 467–482.
Simon, H. (1969). The Sciences of the Artificial, Third Edition.: The MIT Press.
Soltesz, I. (2006). Diversity in the Neuronal Machine. New York: Oxford Uni-
versity Press.
Sperry, R. W. (1980). Mind-brain interaction: mentalism, yes; dualism, no.
Neuroscience. 5, 195–206.
Steinmetz, N. A., Aydin, C., Lebedeva, A., Okun, M., Pachitariu, M., Bauza,
M., Beau, M., … & Harris, T. D. (2021). Neuropixels 2.0: a miniaturized
258 David Parker
high-density probe for stable, long-term brain recordings. Science, 372,
eabf4588.
Stent, G. (1969). The Coming of the Golden Age. Garden City, NY: Natural
History Press.
Steward, O., Popovich, P. G., Dietrich, W. D., & Kleitman, N. (2012). Replica-
tion and reproducibility in spinal cord injury research. Experimental Neurol-
ogy, 233(2), 597–605.
Svensson, E., Aspergis-Schoute, J., Burnstock, G., Nusbaum, M., Parker, D., &
Schioth, H. (2019). General principles of neuronal co-transmission: insights
from multiple model systems. Frontiers in Neural Circuits, 12, 117.
Thomson, H. (2010). Mental muscle: six ways to boost your brain. https://
www.newscientist.com/article/mg20827801-300-mental-muscle-six-ways-
to-boost-your-brain/#ixzz67nP7rZgk.
Warner, R. (2001). Microelectronics: its unusual origin and personality. IEEE
Transactions on Electron Devices, 48, 2457–2467.
Weisberg, D., Keil, F., Goodstein, J., Rawson, E., & Gray, J. (2008). The seduc-
tive allure of neuroscience explanations. The Journal of Cognitive Neurosci-
ence, 20, 470–477.
Weiss, S., & Faber, D. (2010). Field effects in the CNS play a functional role.
Frontiers in Neural Circuits, 4, 1–10.
Wimsatt, W. C. (2006). Reductionism and its heuristics: making methodologi-
cal reductionism honest. Synthese, 151, 445–475.
Woodward, J. (2017). Explanation in Neurobiology: An Interventionist Per-
spective. In D. Kaplan (Ed.), Explanation and Integration in Mind and Brain
Science (pp. 70–100). Oxford: Oxford University Press.
Zhang, Y., Rózsa, M., Bushey, D., Zheng, J., Reep, D., Liang, Y., Brousaard,
G. J., … & Looger, L. L. (2020). jGCaMP8 Fast Genetically Encoded Cal-
cium Indicators. Janelia Research Campus. Online resource. doi:10.25378/
janelia.13148243.v4
Ylikoski, P. K. (2009). The Illusion of Depth of Understanding in Science. In H.
W. de Regt, S. Leonelli, & K. Eigner (Eds.), Scientific Understanding: Philo-
sophical Perspectives (pp. 100–119). Pittsburgh, PA: University of Pittsburgh
Press.
Yuste, R. (2015). From the neuron doctrine to neural networks. Nature Reviews
Neuroscience, 16, 487–497.
12 Cognitive Ontologies,
Task Ontologies, and
Explanation in Cognitive
Neuroscience
Daniel C. Burnston

1 Introduction
The development of new scientific tools provides opportunities for prog-
ress, but also gives scientists reason to reinvestigate, reconsider, and
maybe revise their assumptions about the domain under investigation. In
cognitive neuroscience, this has manifested in the debate over “cognitive
ontology” – that is, the set of mental functions or faculties investigated
by the neurosciences. Psychology comes equipped with a series of intu-
itive mental categories – perception, cognition, memory, imagination,
verbal reasoning, emotion, etc. Cognitive neuroscience has traditionally
proceeded under the assumption that these, or some suitably explicated
set of these, will be realized in the processes that neuroscientists investi-
gate. For better or worse, however, this assumption sits poorly with the
current evidence, which indicates the massive multifunctionality of indi-
vidual parts of the brain, the wide distribution of activity corresponding
to intuitive mental categories, and the importance of global network and
ecological context in determining what an individual part of the brain
does. These data stress, and perhaps break, the “new phrenological”
(Uttal, 2001) approach to cognitive and systems neuroscience, invali-
dating cherished means of analysis such as subtractive methodology and
reverse inference.
One powerful thought in the field is that part of the problem is our
intuitive conception of psychology. Indeed, Poldrack once said that “the
fundamental problem is our stone age psychological ontology” (Bunzl,
Hanson, & Poldrack, 2010, p. 54). Perhaps the standard mentalistic cat-
egories used in the psychological sciences are just too simple, too general,
and too crude to capture how the brain implements behavior. Perhaps
those categories need to be revised, refined, or even abandoned to under-
stand brain function. A host of questions immediately arises, however,
surrounding how committed we should remain to our standard list of
cognitive kinds. Do they successfully describe brain function, only at a
network level? Could we discover discrete implementations of kinds if
they are suitably amended, for instance by subdividing them into more
specific kinds? Or should they just be gotten rid of, resulting in a view

DOI: 10.4324/9781003251392-16
260 Daniel C. Burnston
of the brain on which it is “unanalyzable” (Uttal, 2001) into distinct
functions, where its function is “protean” and lacking generalizability
(Hutto, Peeters, & Segundo-Ortin, 2017)?
Theorists interested in cognitive ontology are thus facing a conun-
drum that is both methodological and ontological. What cognitive cat-
egories are realized in the brain cannot be determined independently of
our methods of investigation. But in cognitive neuroscience, those meth-
ods traditionally employ those categories as basic assumptions – i.e.,
they are what is being investigated, and thereby constrain the interpre-
tation of otherwise inscrutable brain data. Theorists have begun to use
formal tools from databasing, machine learning, and meta-analysis as a
way of addressing this problem. The hope is that the use of these tools
can turn the issue of cognitive ontology into a problem for data science,
rather than metaphysics. By analyzing large amounts of studies using an
agreed-upon, publicly shareable taxonomy of cognitive function, neuro-
scientists hope to be able to discover the ways in which cognitive catego-
ries relate to brain activation, and thereby provide a groundwork for the
substantiation, revision, or abandonment of those categories.
I refer to these projects collectively as “databasing and brain map-
ping” projects, and in this paper, I assess their status and prospects. Ul-
timately, I will argue that the problem is not so much with our intuitive
mental ontology per se, but with the standard explanatory framework
assumed by the cognitive neurosciences. The standard framework as-
sumes that categories of mental function are explanatory kinds, and
that cognitive neuroscience proceeds by showing how these explanatory
categories are instantiated in brain activity. As such, the standard frame-
work is committed to there being an ultimate taxonomy of distinct and
discretely realized cognitive kinds, whose instantiation in the brain caus-
ally explains behavior. I will argue that databasing and brain-mapping
projects, rather than substantiating this standard framework, should
inspire us to abandon it. Instead, I advocate an alternative view of neu-
roscientific explanation on which what explains are ways in which brain
systems organize to implement the informational demands of a particu-
lar task or context (Burnston, 2016, 2021). On this alternative, the best
reading of the role of psychological constructs is as heuristics for inves-
tigation, rather than as explanatory kinds (cf. Feest, 2010).
My aims are both descriptive and normative. I both believe that this
is what successful neuroscientific explanation does look like, and that
it is how we should think about it. This comes along with a variety of
methodological prescriptions, including a plea for increased focus on
task, rather than cognitive ontologies. I hope to clarify the potential
advantages and pitfalls of using formal analytical tools in the cognitive
ontology debate along the way.
I proceed as follows. In Section 2, I introduce the standard explana-
tory framework of cognitive neuroscience, articulate its commitments,
 Ontologies in Cognitive Neuroscience 261
and discuss methodological and empirical problems for the framework
present in the literature. Then, in Section 3, I outline some of the formal
tools that have been applied to the problem, and in Section 4 show that
no clear consensus has emerged on how results employing these methods
are supposed to relate to the standard framework. In Section 5 I out-
line my preferred approach to understanding the role of psychological
constructs in neuroscientific explanation, and in Section 6 show how
this approach offers distinct normative prescriptions than the standard
framework. Section 7 concludes.

2 The Standard Explanatory Framework of Cognitive

Neuroscience
As I understand it, the standard explanatory framework in cognitive
neuroscience is as follows. First, psychological kinds are explanatory.
Behavior is explained by citing mental functions such as memory, atten-
tion, language processing, action planning, etc. Second, explanation is
causal explanation. It is the realization of psychological kinds in neural
processes that explains behavior. To take one philosophical gloss on the
issue, consider Piccinini and Craver’s (2011) account of the role of psy-
chological theories in cognitive neuroscience. Psychology, on their view,
provides mechanism sketches of cognitive phenomena. They outline a
causal sequence of mental functions that can explain the phenomena of
interest. A full mechanism schema, on the other hand, will show how
this abstract functional organization is realized in lower-level causal in-
teractions in the brain, eventually bottoming out in the electrical and
chemical processes of individual cells. This can be read as a way of com-
bining Cummins’ (1983) functional decomposition approach to psycho-
logical explanation with the explanatory goals of cognitive neuroscience.
This kind of view is very influential (Boone & Piccinini, 2016), de-
scribes the traditional explanatory practices of cognitive neuroscientists
well, and comes with a number of advantages. First, it gives a metaphysi-
cally appealing picture of mental causation, wherein psychological states
cause behavior via their realization relation to the physical processes of
the brain. Second, it explains the importance of operationalization and
localization in cognitive neuroscience. ‘Memory’ is not something we
can study directly; we can only study behavior. On the standard story,
behavioral tasks are designed to dissociate the different components
of putative cognitive processes. The brain is then studied to show how
those functional differences are causally realized in distinct parts of the
brain. These changes can be measured either through differences in ac-
tivation – the traditional “subtraction” methodology in fMRI research –
or through intervention, by studying artificially induced or naturally
occurring brain injuries. By finding different localizations of distinct
functions, one explains the causal differences between, for example, a
262 Daniel C. Burnston

Figure 12.1 The standard explanatory framework.

memory process and an attentional process. A diagram of the standard

framework is provided above (Figure 12.1).
Despite its appeal, the standard framework faces a large class of prob-
lems, which we can generally refer to as individuation problems. In-
dividuation problems are the result of the realism about psychological
kinds inherent in the standard framework, along with the complexity
of causal processes in the brain. Basically, different explanantia need
to be distinct. The standard framework is committed, in each instance,
to establishing exactly what psychological functions are explaining a
behavior, how they interact, and how the behavior arises as a result of
that interaction. Individuation problems lead us to question whether this
commitment is in fact met in cognitive neuroscience – they suggest that,
very frequently, we have not, and even cannot, establish exactly which
psychological faculties are at work and the pattern of their interaction.
There are two kinds of individuation problems that have seen exten-
sive discussion in the literature. The first is an individuation problem
with operationalization and measurement. The question here is whether
tasks individuate particular mental faculties. Sullivan (2010), for in-
stance, questions whether the Morris water maze, a famous task in neu-
roscience, can be taken as a specific measurement of “spatial memory,”
the cognitive kind with which it is generally associated, rather than mea-
suring learning, the change of representational capacities, or just ability
to find a platform under the surface of some water. Sullivan (2014) also
applies the argument to tasks such as Stroop tasks. While standardly
thought of as measuring attention, Sullivan persuasively argues that el-
ements of attention, memory, language processing, and perceptual pro-
cessing are all indexed in the standard versions of Stroop tasks.
The second version of individuation problem applies to the kinds them-
selves. The worry is that psychological kinds are simply not distinct from
 Ontologies in Cognitive Neuroscience 263
one another. So, perhaps memory is not distinct, metaphysically speak-
ing, from action planning or imagination (De Brigard, 2014; Schacter,
Benoit, De Brigard, & Szpunar, 2015; but cf. Robins, 2016). Perhaps
“basic” emotions (Griffiths, 2002), or “concepts” (Machery, 2009), or
psychiatric categories such as “schizophrenia” (Tekin, 2016) do not cor-
respond to natural kinds. Recently, these worries have been extended to
core cognitive capacities like working memory (Gomez-Lavin, 2020).
I am moved by these considerations, but will not focus on them here.
What I will assess in detail is the last variety of individuation problem,
which occurs in the purported realization of cognitive kinds. The tradi-
tional approach to cognitive neuroscience hopes to specify the function
of each part of the brain. This “atomistic” (Burnston, 2021) approach
is motivated by the goal of causal decomposition in the brain. If we can
specify the function of each part, then we can understand any given be-
havior as the result of causal interaction between these functions. A one-
to-one mapping between the mental ontology and the neuroscientific
ontology would occasion a particularly powerful form of mechanistic
explanation.
Current data, however, sits uneasily with individuating atomic map-
pings between functions and brain activation. Widespread data from
many distinct parts of the brain suggest the multifunctionality of in-
dividual brain parts, overlap between instantiations of distinct mental
categories, and distribution of brain activation corresponding to distinct
kinds. Multifunctionality of particular parts of the brain undermines
the ability to say, given activation in a particular part, what function
that part is performing, and hence how it is causally interacting with
other parts. This is part of the problem with the traditional subtractive
methodology and reverse inference. Overlap and distribution of func-
tion undermine the ability to distinguish between the causal contribu-
tion of distinct kinds at the neural level.
It is worth considering further why this is. Suppose we are observing
a behavior, and activation in a number of parts of the brain. If one can
decompose the behavior into distinct psychological processes, and lo-
calize each of those processes, then one can theorize about the causal
interactions between them. But significant distribution and overlap of
instantiation muddy the division between localization and interaction.
This is because they allow for too much inferential freedom in how one
interprets the instantiation relation. Suppose two putatively distinct fac-
ulties overlap in their instantiation. Is this due to the fact that they share
some common functional core and other elements that differ? Is it be-
cause we have not explicated them sufficiently relative to each other? Or
is it because they are not, in fact, distinct after all? Similarly, given wide
distribution corresponding to a given function, is that distribution an
indication that the construct ranges over multiple distinct sub-functions
that interact? Or that our experiments in fact index multiple distinct
264 Daniel C. Burnston
functions? Or, again, that the construct does not describe the mechanis-
tic functioning of the brain?
In most cases of mental faculties, this is the situation that actually
obtains – the data suggests multifunctionality, distribution, and overlap.
But the standard framework requires that distinct explanantia be dis-
tinct, including in their instantiations. So, the current data conflicts with
the standard framework. Given this conflict, one can either attempt fur-
ther work to substantiate the standard framework, or one can abandon
it. To substantiate the framework, one would have to either try to fur-
ther differentiate the instantiation relations between distinct kinds, or
revise the ontology so that more specific mappings emerge. One could,
of course, pursue some combination. If abandoning the framework,
one would have to specify what kind of explanation results from that
abandonment.
The idea of revision of the cognitive ontology is appealing, and it is
often suggested in mechanistic contexts that higher-level kinds will have
to be split or revised in light of causal explanation at a lower level (Bech-
tel, 2008; Bickle, 2003). Similarly, it is often suggested that databasing
and brain-mapping techniques can help us revise our ontology. I will
consider these claims in detail in the next section. But it is worth noting
that the ontology revision proposal is less anodyne than is normally sup-
posed. Attempting to fine-grain our taxonomy does not guarantee that
distinct functions will be discovered – Feest (2010), for instance, nicely
explains how continuous attempts to distinguish implicit memory from
other forms of memory are what eventuated in the conclusion that im-
plicit memory may not be distinct from perceptual association. And even
a successful distinction may not be mechanistically useful – attempts to
distinguish face perception, body perception, and place perception, for
instance, do not show clearly distinct realizers but interdigitated “archi-
pelagoes” of voxels with a statistical preference for one kind of informa-
tion over another (Kanwisher, 2010). This is not a discovery of clearly
distinct parts with causal interactions between them.
In the next two sections, I introduce databasing and brain mapping
techniques in more detail, and argue that, while proponents of these
techniques are generally realist about psychological kinds, they do not
clearly opt for either substantiation, revision, or abandonment of the
standard framework, instead vacillating between these options. I also
raise doubts that the databasing and brain mapping techniques on offer
can perform any of these functions. This motivates my own take on the
issue, which I will pursue in Section 5.

3 Databasing and Brain Mapping

There are two main aspects to the databasing and brain-mapping proj-
ects I will discuss. The first is the collection of compendious amounts of
 Ontologies in Cognitive Neuroscience 265
neural data from across studies. The shareability of scientific data, as
well as the best means to collect it, disseminate it, and use it, are prob-
lems across the biological sciences (see, e.g., Bechtel, 2017; Darden et al.,
2018; Leonelli, 2012), and neuroscience is no different. One reaction to
the massive amount of research using, for instance, fMRI methodology,
is to try to systematize and understand this expansive dataset as a whole.
So, collection of the information is the first step. Several open-access
databases have been created in cognitive neuroscience to play this role,
including the Brain Map (Fox & Lancaster, 2002), Neurosynth (Yarkoni
et al., 2011), The Cognitive Atlas (Poldrack et al., 2011), the Experiment
Factory (Sochat et al., 2016), and The Cognitive Paradigm Ontology
(Turner & Laird, 2012).
While these projects differ in their precise focus, they all share a num-
ber of aspects. First, the idea is to collect activation data in a theoretically
unbiased way. What is archived is the raw activation data from a set
of fMRI studies. One can then ask questions about this data. Second,
one of the questions that everyone wants to ask about this data is how,
whether, and in what sense patterns in the data correspond to psycholog-
ical concepts. This is done in a number of ways. In the Cognitive Atlas,
each study, in addition to the data, is categorized according to the type of
tasks manipulated. Each task type is then defined, and the psychological
concepts that it is supposed to measure are listed. So, one can look for
the ways in which the same concepts are realized similarly or differently
across different tasks, or one can look at how different tasks/concepts
diverge or overlap. In Neurosynth, the text of papers is archived along
with the raw fMRI data, so one can look for ways in which usage of key
mental terms by scientists varies along with changes in brain activation,
and vice-versa. Finally, the hope across these projects is that the search
for patterns can be automated. Given the scale of the data set, automated
data analysis is used in the attempt to find meaningful patterns.
The analytical techniques applied to this collected data range from
traditional meta-analyses to statistical classifiers to generative, prob-
abilistic models, each with their associated benefits and detractions.
Meta-analytic techniques take already reported correlations between
cognitive concepts and activation patterns, and attempt to identify, gen-
eralize, and summarize the relationships discovered in the literature.
Statistical decoders train models to predict, given the presence of brain
activation, what cognitive concepts are being assessed in the range of
studies (or vice versa, see below). While there are a range of generative
models, one popular technique is Latent Dirichlet Allocation (LDA),
which is a Bayesian algorithm that models the text in a corpus of words
(in this case, the text from fMRI studies) as being generated by a group-
ing of topics, themselves construed as probabilistic groupings of indi-
vidual words. One can then attempt to correlate greater influence (or
“loadings”) of those topics with brain activation.
266 Daniel C. Burnston
There are a range of attitudes taken by brain mappers to their proj-
ects, which I will discuss in detail below. In general, however, I think
there are two fundamental assumptions they share. First, they are real-
ists about mental faculties. Second, and relatedly, they are committed
to the measurement relation between tasks and those faculties. For in-
stance, Hastings et al. (2014) describe the project as one on which “onto-
logical realism is a foundation” (p. 4). Lenartowicz et al. (2010) suggest
that “the elements of the mental ontology are not directly accessible but
rather must be accessed through experimental manipulations and mea-
surements (i.e., tasks)” (p. 680).
In these quotes, theorists are committing to the idea that mental func-
tions are real entities, and that tasks are measurements of them. This is
reflected in much of the databasing work. In the Cognitive Atlas and
Brain Map, for instance, tasks are explicitly categorized according to
the mental constructs they are supposed to measure. While Neurosynth
collects a range of textual data, the preprocessing of that data indi-
cates a realist commitment. Generally, LDA models using Neurosynth
focus on the abstracts of paper, and specifically on the cognitive terms
contained therein. In attempting to map these uses to the brain, then,
these projects assume that, at least at an abstract level, the concepts we
employ in thinking about the mind correspond to physical categories
within the brain.
As I will show below, this set of commitments interacts in complicated
ways with the standard framework. For now, I want to discuss a few
early results from these frameworks to show that, far from solving in-
dividuation problems, brain mapping projects tended to illustrate them.
The question will then be what attitudes brain mappers take to these
results.
In a meta-analysis of fMRI research, Anderson, Kinnison, and Pes-
soa (2013) compared different patterns of activation according to the
cognitive categories listed in Brain Map. They were interested in a num-
ber of properties, including the range of cognitive concepts associated
with each area’s activation, the distribution of activation corresponding
to those concepts, and the degree to which distinct brain areas were
likely to be active in studies measuring the same mental concepts. What
they showed was that individual parts of the brain exhibit a range of
“diversity profiles,” but that most areas’ activation corresponded with
significantly more than one cognitive concept. Moreover, areas within
previously identified functional-connectivity networks tended to be
highly “assortative,” meaning they tended to be active for similar cog-
nitive concepts, suggesting both the distribution of individual functions
and the overlap in neural activation between functions. Importantly,
taking functional networks such as the “fronto-parietal” network and
the “ventral attention” network, as basic units to correlate with cogni-
tive concepts showed a similar pattern of results.
 Ontologies in Cognitive Neuroscience 267
Poldrack, Halchenko, and Hanson (2009) performed a decoding anal-
ysis of the results from eight different fMRI studies investigating a range
of cognitive constructs. They began with statistical maps of the entire
brain – i.e., z-scored activation coordinates from every condition in the
eight studies. The question was then whether one could train a decoder to
predict which tasks and/or cognitive concepts were named in the studies,
such as “risk-taking”. They trained a support vector machine to predict,
on the basis of a given brain-wide activation, what task and what cogni-
tive concept was being measured in a case and showed that the classifier
could successfully classify both with 80%–90% accuracy. They further
trained a neural network with six hidden nodes to match the predic-
tive accuracy of the support vector machine. Importantly, however, the
nodes operated over a widely distributed set of voxels. When analyzed
as a six-dimensional system (one for each hidden node), each cognitive
concept was shown to be related to a combination of each dimension,
and each dimension was associated with a range of cognitive concepts.
Poldrack et al. (2012) performed a topic modeling analysis with the
following structure. First, they took the results and text from over
5,000 papers in the Neurosynth database. They began by exploring the
topic structure in the text. They then selected topics that corresponded
to mental concepts in the Cognitive Atlas and measured how the topic
loadings on these topics correlated with brain activity. Here, however,
they also show multifunctionality and distribution in the results. For
instance, they report:

topic 43 (with terms related to visual attention) was associated with

activity in the bilateral lateral occipital cortex, parietal cortex, and
frontal cortex. Topic 86 (with terms related to decision making and
choice) was associated with regions in the ventral striatum, medial,
orbital, and dorsolateral prefrontal cortex. Topic 93 (with terms re-
lated to emotion) was associated with bilateral activity in the amyg-
dala, orbitofrontal cortex, and medial prefrontal cortex.
(2012, p. e1002707)

These results are perfectly interesting in their own right, in that they
quantify the “specificity” with which our intuitive cognitive concepts
interact with brain activation. It is just that, on their face, they are in
conflict with the standard model because they show multifunctionality
and distribution rather than univocal relationships between psychologi-
cal constructs and activation. The question is what to do in response to
these results with regard to the standard framework.
Let me stress that I am reconstructing positions here – I think each of
the options I am about to articulate is present, to some degree, across
papers and theorists within the field. As far as I can tell, there are three
options with regard to the standard framework. First, one could attempt
268 Daniel C. Burnston
to substantiate the framework by pursuing more fine-grained analyses
in an attempt to discover more and more specific activation patterns for
particular cognitive concepts, perhaps further leading to decomposition
and causal explanation. Second, one could attempt to use the analyses to
revise our cognitive ontology. On this view, it might be the case that the
standard framework can be maintained, but only after the appropriate
revisions to the ontology. Third, one might use these results as motiva-
tion to abandon the standard framework altogether and opt for some
other kind of project. In the next section, I outline each of these perspec-
tives, along with examples from the literature which might suggest them,
and give reasons to question them. This will motivate my own proposal
about psychological constructs in Section 5.

4 Three Options with Regard to the Standard Framework

4.1 Substantiate?
One view one could take towards the standard framework is that it is
basically right, and our ontology is basically right, but that the results of
multifunctionality and distribution are due to insufficiently fine-grained
measurement. The solution, if this is one’s perspective, would be to re-
fine analyses so that the “true” and univocal associations between psy-
chological constructs and brain activation can be uncovered.
I think that this is the position that is least strongly considered in the
literature, but there are a few trends that suggest it. Indeed, one direction
in which the literature has gone over the last few years is in the direction
of more fine-grained analysis and the search for increased specificity.
Poldrack and Yarkoni (2016) thus describe the project as one of “quan-
tifying the true specificity of hypothesized structure-function associa-
tions” (p. 589). This, one assumes, means that they indeed take there
to be relations there to be discovered, further indicating realism about
psychological kinds. Moreover, recent projects take the goal of brain
mapping projects to be enabling both forward and reverse ­inference –
that is, the predictive ability of brain mapping models between con-
structs and activation patterns should be bidirectional. This, to me at
least, further suggests a belief in the importance of the instantiation
relation. Finally, there is how these projects are qualitatively described.
For instance, Varoquaux et al. (2018) suggest that one of the goals of
mapping projects is “precisely describing the function of any given brain
region” (2018, p. 1).
Varoquaux et al. performed a decoding analysis using a hierarchical
general linear model (GLM) framework. In particular, their reverse in-
ference required multiple layers of linear regressions on activity in the
brain. The first layer was tuned to individual oppositions between task
conditions. Then, a second layer used another regression that compared
 Ontologies in Cognitive Neuroscience 269
each cognitive term to all others, predicting which term was overall most
relevant. They compared the results of this decoder to other approaches,
showing that it resulted in sharper divisions between distinct functions.
There are a few things to be said here, however. First, this study
measured terms in the Cognitive Paradigm Ontology rather than the
Brain Map or the Cognitive Atlas, and these terms more directly de-
scribe task conditions (e.g. “response with left hand”) than psycholog-
ical constructs (e.g., “motor control”). Second, they focused primarily
on perceptual and motor areas for which there are already more-or-less
well-­understood general function ascriptions. Finally, even these results
showed distributed and interdigitated functional populations, with, for
instance, “face” and “place” areas being more or less separated, but
each involving multiple subpopulations distinct from each other.
Another recent approach to bidirectional decoding is from Rubin et al.
(2017), which employs LDA on over 11,000 articles from Neurosynth.
They start by noting that previous studies show mainly wide patterns of
activation for particular constructs, and thus are no help in finding “rel-
atively simple, well-defined functional-anatomical atoms.” To overcome
this, they performed an LDA analysis constrained both by the semantics
of the terms and by groupings in spatial coordinates. They report that,
not only were they able to uncover topics with relatively clear functional
upshot, (e.g., topics related to “emotion”), but that each topic “is as-
sociated with a single brain region.” At first, this sounds a lot like the
explanatory aim of the standard framework – i.e., to find a constrained
localization corresponding to each psychological function.
A closer reading questions this analysis, however. As the researchers
note, the probabilistic nature of the model suggests that the decoding
analysis uncovers the construct most likely associated with a given area,
but not the only one. This is further illustrated by the fact that individual
topics were allowed to spatially overlap in the model, and many mul-
tifunctional areas did indeed show significant overlap between related
topics. Further specifying to individual topics in many cases required
conditioning further on more spatial coordinates, hence suggesting,
again, distribution of function. So, while the results in this model are
predictive at a very specific construct-spatial level, it is not clear that
this reflects the reality of the system. And this is noted explicitly by the
researchers. It is worth quoting them in full:

While the topics produced by the model generally have parsimo-

nious interpretations that accord well with previous findings, they
should be treated as a useful, human-comprehensible approximation
of the true nomological network of neurocognition, and not as a
direct window into reality. For the sake of analytical tractability,
our model assumes a one-to-one mapping between semantic repre-
sentations and brain regions, whereas the underlying reality almost
270 Daniel C. Burnston
certainly involves enormously complex many-to-many mappings.
Similarly, rerunning the GC-LDA model on different input data,
with different spatial priors, a different number of topics, or with
different analysis parameters would necessarily produce somewhat
different results.
(Rubin et al., 2017, p. 14)

So, while one trend in the literature is to look for increasingly specific
relationships between extant psychological constructs and patterns of
activation, it is not clear that even successful results in this endeavor sub-
stantiate, or should be read as attempting to substantiate, the standard
framework.

4.2 Revise?
The idea that the databasing and brain mapping project can help us re-
vise our cognitive ontology is extremely common. For example, Poldrack
and Yarkoni (2016) suggest that “formal cognitive ontologies [are use-
ful] in helping to clarify, refine, and test theories of brain and cognitive
function” (p. 587), and that “biological discoveries can and should in-
form the continual revision of psychological theories” (p. 599).
These quotes suggest that, ultimately, the role of the databasing and
brain mapping project will be in helping us to explicate our mental on-
tology. Sometimes, this is pitched in terms of a discovery science – we
should let the brain tell us what its functional categories are, and revise
our ontology accordingly (Poldrack et al., 2012). In this section, I sug-
gest two related problems for this view. The first is the interpretability
problem, and the second is the seeding problem. In general, however, the
issue is this: without a rubric for how and when to revise our mental cat-
egories in light of brain mapping data, we lack the ability to use results
from brain mapping to revise the ontology in any specific way. This sug-
gests that metaphysical commitments about the nature of mental states
precede, rather than being compelled by, brain mapping data.
The interpretability problem is akin to a problem discussed by Carl-
son et al. (2018; cf. Ritchie, Kaplan, & Klein, 2016; Weiskopf, 2021)
for uncovering neural representations via machine learning techniques.
They argue that, given a particular ability to decode some stimulus from
neural activity, it is unclear how to interpret that result in terms of rep-
resentational content. The worry, I take it, is that the ability to decode
a stimulus from an activation does not mean that the activity represents
the stimulus under anything like the way we would describe it. The an-
alog problem here is that simply showing that a pattern of activity in
the brain is specific to, say, decision-making (or to a high topic loading
on a topic that happens to comprise words we associate with decision-­
making), doesn’t give us any indication of whether the pattern of activity
 Ontologies in Cognitive Neuroscience 271
is in fact performing something we would call “decision-making.” The
more distribution and overlap uncovered in the analysis, the more exac-
erbated this problem becomes, because of the inferential freedom dis-
cussed in Section 2.
So, given the association of a pattern of activity with a mental con-
struct, should we take that construct as substantiated, as in need of ex-
plication, or what? What degree of correlation/predictability, or what
degree of specificity, is required to count the kind as substantiated, and at
what point should we consider it in need of revision? The brain mapping
results themselves provide no rubric for how to make these decisions.
The seeding problem is related to the interpretation problem and is
based on the fact that even constructing the analyses requires adher-
ing, to an unspecified degree, to our extant cognitive constructs. In an
analysis based on Brain Map or the Cognitive Atlas, one only considers
concepts that are a current part of our mental ontology. This presumes
that the basic structure of the brain corresponds closely enough to those
categories in order for them to be useful in understanding the brain. But
what justifies this assumption? In principle, a specific-enough correla-
tion between mental constructs and brain activity might justify the as-
sumption, but it is precisely a lack of specificity of this type that prompts
the idea of ontology revision.
In topic-modeling analyses, the topics that are often focused upon are
the ones that one can intuitively or statistically pair with an a­ lready-known
mental construct. Varoquaux et al., for example, advertise that 100 of
the 200 topics in the model correspond to well-understood mental con-
structs. What about the other ones, however? Even given a substantia-
tion of some of our mental categories, what the analysis would suggest
is that our ontology is at least impoverished, and it does not come with
any prescriptions for what to say in these other cases.
Again, the point of this is not to discount the analysis. The point of
if it is just to deny that the analysis on its own offers us any principled
way of revising our cognitive ontology. Put differently, the principles for
ontology revision cannot be uncovered bottom-up from these analyses.
Metaphysical commitments must be undertaken in constructing and in-
terpreting the analyses themselves. Again, theorists in the field recognize
this problem. Poldrack and Yarkoni (2016), for instance, note that there
is “no algorithmic way” to approach ontology revision in light of specific
mapping results. They seem to suggest, however, that more analysis and
case-by-case thinking will allow for sufficient explication. The individ-
uation and seeding problems should raise concerns for that approach.

4.3 Abandonment?
One also finds more-or-less explicit discussion of the explanatory ide-
als of the standard framework in the literature. The clearest cases of
272 Daniel C. Burnston
these are Yarkoni and Westfall (2017) and Anderson (2014). Yarkoni
suggests explicitly that results from databasing and brain-mapping proj-
ects suggest abandoning explanation altogether, in favor of a purely pre-
dictive neuroscience. Anderson’s view does not cite prediction per se,
but does suggest that we need to change to a dispositional approach to
brain organization, wherein we do not understand a part of the brain as
contributing a specific causal influence at a specific time, but instead as
exhibiting dispositions to contribute to a range of functions.
I lack the space to assess these proposals in detail, but for my pur-
poses, it suffices to note that they both, more-or-less-explicitly, move
away from the mechanistic kind of explanation inherent to the standard
framework. Much has been said about the relative merits of mechanistic
explanation versus prediction in explanation (Craver, 2006), and now
is not the time to re-adjudicate these issues. What I want to argue for in
the remainder of the paper is that abandoning the standard framework
is not itself equivalent to abandoning mechanistic explanation. Instead,
we can abandon the standard framework by abandoning the central ex-
planatory role it affords to mentalistic constructs.

5 An Alternative View

5.1 Mental Constructs as Heuristics

My proposal is based on the following negative claim: posits of psycho-
logical faculties do not explain behavior. Individuation problems arise
from the notion that posits of psychological faculties are explanatory.
That is why they must be distinct from each other; that is why they must
be discretely realized; and that is why causal interactions between them
need to be established. Get rid of their explanatory status, and all of
those problems go away – it’s not particularly worrisome of psycholog-
ical kinds are not clearly distinct from each other, if specific behavioral
tasks don’t measure only one of them at the expense of others, or even
if they massively “crosscut” the causal patterns we measure in the brain
(Hochstein, 2016; Weskopf, 2011).
This leaves us with two questions. First, what does explain behavior?
And second, do psychological posits play any role in understanding it?
These questions can be asked within the context of brain mapping projects
as well. What sort of explanation should these projects be seen as working
towards? And should mental concepts play any role as they seek them?
My answers are as follows. First, information processing in the brain
explains behavior directly, and not in virtue of instantiating some par-
ticular mental function. As I will attempt to show, this proposal is com-
patible with each brain area being multifunctional, and with function
generally being distributed (Burnston, 2016, 2021). Second, mentalistic
concepts are best understood as playing a heuristic role (cf. Feest, 2010).
 Ontologies in Cognitive Neuroscience 273

Figure 12.2 The heuristic approach to mental constructs.

Rather than serving as explanantia, we should view mental categories as

helping us parse behaviors into rough, and revisable, similarity classes.
They provide traction on an otherwise impossibly complex space of
behavioral abilities and guide the search for important distinctions be-
tween types of behaviors, so that we can then investigate how the brain
implements those differences. Importantly, they can play this role even
if there is no fact of the matter about which neural processes implement
which psychological constructs. Hence, no individuation problems are
faced (Figure 12.2).

5.2 An Exemplar
The heuristic view makes a number of invocations about successful ex-
planations in neuroscience. First, overlap between mental constructs
across tasks and contexts should be just as important as separation be-
tween them. Second, understanding the differences in structure between
tasks is paramount for understanding neural function. Third, assuming
spatial decomposition between distinct purported mental faculties would
limit, rather than enabling, understanding of how the system works.
I will discuss one example in detail. Murray, Jaramillo, and Wang
(2017) pursued a modeling study of the interaction between the pre-
frontal cortex (PFC) and the posterior parietal cortex (PPC). The initi-
ating motivation for their study is that both the PFC and the PPC have
been shown physiologically to be involved across a wide range of both
working memory (WM) and decision-making (DM) tasks. The question,
then, is what their distinct contributions are.
Murray et al.’s approach was as follows. They modeled each area
as a fully recurrent neural network, and each area had distinct sub-­
populations selective for distinct perceptual stimuli. The PFC and PPC
networks were bi-directionally connected via long-range projections.
The main difference between the two populations was a difference in
274 Daniel C. Burnston
local structure. In particular, the PFC population was modeled as hav-
ing a higher degree of internal influence – both in self-excitation of each
subpopulation, and in inhibitory connections between them, than the
PPC. Given this network structure, the investigators could model the
dynamics of the system in a range of task types, and think about how
the network responded in each.
Murray et al. posited that one key factor involved in working memory
tasks is multi-stability. That is, the network can represent a range of
possible stimuli, but given that it has already represented one, it must
maintain that information across a delay, perhaps in the presence of
distractors. So, they modeled the presentation of a stimulus and whether
its representation could be maintained in the network even as other
modeled stimuli were presented. What they showed is that a particu-
lar dynamics occurred during “successful” working memory trials, in
which both PFC and PPC populations represented the stimulus during
presentation. During delay, presentation of a distractor would “switch”
the PPC representation to representing the distractor, but PFC would not
switch. After the distractor was removed, the PFC → PPC long-range
connections would enforce the PPC populations to “switch back” to
representing the remembered stimulus. This model predicted a range of
physiological results found in PFC and PPC during these kinds of tasks,
as well as predicting types and durations of distractor-presentation that
would cause errors. Importantly, this explanation relied on the degree of
internal connection in each area, and the feedback connections from the
PFC to the PPC. For instance, if the internal connections within the PFC
were not sufficiently strong, then it would not maintain the representa-
tion during distractor presentation.
For decision tasks, the investigators asked whether the network could
produce an evidence-accumulation-to-threshold kind of process. These
processes have been shown to be important for a range of decision-­
making processes including perceptual decisions and multi-attribute
choices (Teodorescu & Usher, 2013). Murray et al. modeled a perceptual
decision-making task, in which one out of a range of possible perceptual
outcomes must be decided on in the presence of a noisy signal. They
showed that the network could implement an evidence-­accumulation
process, in which buildup of evidence occurred primarily in the PPC
population and selection of outcome in the PFC population. Intriguingly,
these dynamics also were dependent on the degree of internal structure
in the populations. Specifically, if the PFC population had a lesser de-
gree of internal recurrent influence, the network would evolve towards
a decision more slowly, whereas if it was strongly internally connected
it would evolve very quickly. As predicted, at a higher degree of inter-
nal influence the network “decided” faster, which in turn contributed to
more errors when the stimulus was noisier, and more time to integrate
evidence would have been helpful.
 Ontologies in Cognitive Neuroscience 275
So, the very same network could implement both the kind of infor-
mation processing required in a WM task, and the kind required in a
perceptual DM task. One of the most intriguing results, however, is
that these two kinds of information processing trade-off in a network.
Greater resistance to distraction in the PFC network required a high
degree of internal influence in that module. But a high degree of inter-
nal influence also shortened the timeline over which perceptual evidence
could be accumulated. They posited that the particular structure of the
PFC-PPC circuit helps ameliorate this tradeoff. In particular, if one re-
moved the recurrent connections from the PFC to the PPC, then high
performance in the decision-making task would result in lower perfor-
mance in the working memory task.
I suggest that this kind of modeling project results in a mechanistic
understanding of the network, but only by exhibiting the three proper-
ties I discussed above. First, the understanding of the circuit developed
in the study starts out from the data point that both working memory
and decision employ overlapping circuits. Second, understanding the in-
formational requirements that are in common and differ across tasks is
central to the explanation. In particular, there is something in common
between working memory and decision-making tasks, namely that it is
useful to have both a population that is multiple in its responses paired
with a more categorically responding population. The difference in in-
ternal structure between the PFC and the PPC leads to the former hav-
ing more univocal responses, which lead to both its robustness in WM
contexts and its thresholding behavior in DM contexts. However, the
differences between the tasks are also vitally important, because they il-
lustrate the tradeoff in the network. WM contexts benefit from stronger
interconnection, since it increases resistance to distractors. But DM con-
texts benefit from weaker interconnection, since it allows for an increase
in evidence-gathering. This in turn leads to a mechanistic hypothesis,
namely that the distinction between the PFC and PPC circuits in their
degree of self-influence, and the feedback connection from the former to
the latter, help ameliorate this tradeoff.
Importantly, because the explanation takes this form, it would be a
mistake to attempt to spatially map WM and DM to distinct brain sys-
tems. There is not one part “doing” working memory and another part
“doing” decision-making, and therefore there is not a causal relationship
between so-individuated parts. There is one distributed circuit underlying
those intuitively distinct functions. Given this, I submit, there is no meta-
physically important distinction between working memory and percep-
tual decision-making. What there are are distinct task demands, and the
ways in which those demands are implemented by a distributed system.
The heuristic view, unlike the standard view, simply doesn’t assume
that there is a fact of the matter about (i) whether WM is really distinct
from DM, (ii) which tasks measure one versus the other, or (iii) whether
276 Daniel C. Burnston
a brain part really performs one rather than the other. What it suggests,
as seems to be the case, is that there are deep commonalities in the brain
systems performing these functions. The explanation also does not re-
quire that there be any firm division between the ultimate set of tasks
that are WM, versus those that are DM, tasks. There are simply tasks
with different informational requirements that are implemented differ-
ently in the network.
This is compatible with working memory and decision-making hav-
ing played important heuristic roles in the understanding of this sys-
tem. It was not, perhaps, initially obvious what the relationship between
WM and DM might be. The concepts were operationalized differently.
However, the persistent discovery of overlapping involvement in each
of these tasks by the distributed PFC/PPC circuit led to the question
of exactly how these functions are implemented. This led to a model-
ing project which uncovered both the commonalities and the differences
between the informational requirements of, and the neural processing
instantiating, tasks that correspond more-or-less closely to each of these
categories.
I have only discussed one example, but I take this to be an exemplar of
how multifunctional distributed circuits might be decomposed. I discuss
a variety of other examples in other venues (Burnston, 2020, 2021). If
this case is exemplary, however, then it stresses the normative bit of the
heuristic approach, as opposed to that of the standard framework.

6 Normative Upshot

6.1 For Ontologies

Whether each process instantiates attention, some form of memory, or
something else, I contend, is not as important for explaining how the
brain implements the informational demands of the particular task. The
normative upshot for databasing projects is that more explicit attention
needs to be paid to the kinds of behavioral paradigms at work and the
way that they tend to vary (Figdor, 2011; Sullivan et al., 2021).
This is not to say that databasing projects have not paid attention
to tasks. Projects such as the “Cognitive Paradigm Ontology” (Turner
& Laird, 2012) are specifically intended to index the different types of
behavioral experiments involved in studying cognition, and Sochat and
colleagues (2016) have recently argued that standardization and public-
ity of experimental design are vital for coordinating the study of cog-
nition between laboratories. But the particular way in which tasks are
approached in the field tends to strongly mirror the standard framework.
For instance, Sochat et al. (2016) argue that the Cognitive Atlas should
be “integrated” with a task ontology, by which they mean that each type
of behavioral experiment should be categorized according to the kind of
 Ontologies in Cognitive Neuroscience 277
psychological function it is used to study. Standardization is important
so that neuroscientists can “select paradigms based on the specific cog-
nitive functions that they are thought to measure” (2016, p. 7).
The heuristic approach views the situation differently. While an inves-
tigator may “base” their search for behaviors to study on associations
with cognitive functions that interest them, this basis is a heuristic one
and not an instance of measurement. That is, a neuroscientist interested
in “working memory,” broadly speaking, should not pursue particular
behavioral paradigms because they measure that explanatory construct,
but instead as a way of looking for behavioral distinctions that may
make a difference in how the brain processes information.
This has important upshot for how databases are constructed. In
Neurosynth and the Cognitive Atlas, behavioral tasks are given abstract
definitions, the assumption being that their role is to measure cognitive
functions rather than themselves serving as explananda. In the Cogni-
tive Paradigm Ontology and Brain Map, importantly, there are entries
for experimental manipulations such as epoch, stimulus, and response,
but as far as I can tell these categories have not been standardized, or-
ganized for comparative analysis, or as rigorously codified as the mental
constructs have been.
From the standpoint of the heuristic approach, this is generally insuf-
ficient. In the exemplars above, what makes a difference for explanation
is understanding how the brain responds differentially to variations in
task demands, from the presentation of the stimulus, often through a
delay or across learning, to an eventual behavior. Differential responses
to these variations are what explain how the brain implements the be-
havior. Other subfields of neuroscience employ different variations – for
instance, other areas of decision-neuroscience specifically vary reward
types and regimes in conjunction with changes in stimuli and behavioral
requirements. But entries in ontologies almost never contain detailed in-
formation of this sort, or at least that information is not standardized
and codified.
Now, given the potentially infinite ways that behavior can be varied,
codifying the kinds of behavioral and stimulus changes, the duration of
temporal epochs and how they relate, etc., will be both a significant and
a difficult task. But it is exactly the point of databasing projects to codify
large amounts of data and make it accessible across individuals trying to
explain things in a field. The normative bite of the heuristic view is that
this project is more important than trying to find the “true specificity”
between abstract mental constructs and brain activity.
This is not to deny that psychological concepts should be included in
databasing projects. What it does suggest though is that the current state
of the situation, on which most tasks are associated with a range of cog-
nitive concepts, is neither surprising nor problematic. If the goal of these
concepts is to serve as heuristics in the search for behavioral paradigms
278 Daniel C. Burnston
that index behavioral differences, then it is not surprising that these con-
structs should overlap both with each other and in the brain, and there
is no need to try to theorize that overlap away.
So, suppose a neuroscientist is interested in ‘working memory’ or ‘de-
cision’. They approach a databasing tool in trying to devise research
questions and develop experiments. What they find is that their con-
struct of interest overlaps with other related constructs, and they find
a huge number of extant behavioral distinctions that have been shown
to make a difference in how the brain operates in, broadly, those cog-
nitive contexts. This allows them to understand the behavioral distinc-
tions that have been done, where brain activity has been measured in
these conditions, etc. If they are interested in a particular part of the
brain, they may find related behavioral measures that might clarify its
function. If they are interested primarily in a construct, they may find a
range of areas and distinctions that they could investigate, or use in the
backdrop of forming new ones. The use of the psychological construct is
a heuristic for investigation, rather than an explanatory claim.
This is, of course, an ideal, but I think it is an ideal that is impor-
tantly distinct from the current focus in databasing and brain-mapping
projects.

6.2 For Functional Connectivity Analyses

An exciting series of recent functional connectivity studies have begun
the attempt to uncover dynamic principles for the entire brain during the
course of behavior. Functional connectivity measures the co-activation
of brain regions, with the assumption being that co-active regions are
coordinating in producing the relevant behavior. One can track changes
in functional connectivity with changes in the task or even across tem-
poral epochs of the same task. One can then ask a variety of questions
about the functional principles underlying these changes.
Here are just a few examples. Shine et al. (2016) hypothesized that
network interconnectivity scales with task complexity. So, they mea-
sured functional connectivity in a range of tasks and compared the over-
all degree of connectivity in the brain during each. They showed that
“language” tasks or “working memory” (in this case, N-back) tasks re-
sulted in higher levels of connectivity than more “simple” motor tasks.
Other studies have attempted to spatially decompose the brain into
parts that help organize its dynamic changes across tasks, versus those
that are in charge of performing those tasks. So, Shine et al. (2018) took
whole-brain data across a range of tasks and applied principal compo-
nents analyses to both the spatial and temporal data. They showed that
the first principal component in the temporal domain correlated with
activity in the spatial regions associated with the “rich club” network,
and hypothesized that these regions underlie an across-task organizing
 Ontologies in Cognitive Neuroscience 279
function, whereas regions more closely responding to the other compo-
nents were more task-specific.
This is cutting-edge work, and I do not wish to speculate too much
on its eventual upshot. What I want to suggest, however, is that the
heuristic approach offers distinct normative prescriptions than the stan-
dard framework for how to pursue further investigation. The standard
framework suggests that these dynamical processes must be divided up,
both spatially and temporally, according to distinct psychological facul-
ties, and the causal relationship between those faculties explained. The
heuristic view denies this necessity. Instead, it invocates the need to un-
derstand the structure of the tasks further – what distinctions in stimuli
or behavioral requirements drive the different dynamic shifts, and what
about the function of the brain networks involved enables them to enact
those requirements specifically?
The authors of these studies seem to view them, at least in part, as
stepping beyond the standard framework. As Shine and Poldrack note:
“these results shifted the focus from where in the brain a particular func-
tion resides to how the coordinated recruitment of segregated specialist
neural regions works together to accomplish the challenges associated
with complex behavioral tasks” (2017, p. 396).
But it is important to note that, at least as of now, the invocations of
the heuristic approach have not been followed. For instance, there is no
analysis of what “complexity” of tasks amounts to in the Shine et al.
(2016) paper. The notion is left intuitive. Nor is there any analysis of
what exactly makes tasks “language” tasks versus “memory” tasks in
the other studies. This is a lacuna in these projects, according to the
heuristic approach.

7 Conclusions
A number of years ago, it was common for textbooks in the philosophy
of mind to teach the following: either our intuitive conception of the
mind, with its commitments to intentional attitudes, etc., is true, or be-
haviorism is. I hope this strikes the modern reader as almost charmingly
anachronistic. One can find newer versions of the dichotomy, however.
Uttal (2001), in his famous criticism of fMRI research, argues that the
alternative to discovering discrete localizations for distinct cognitive fac-
ulties is to view the mind as “unanalyzable,” by which he means indivis-
ible into distinct parts. More recently, Hutto et al. (2017) have suggested
that the way to react to the “protean” – by which they mean dynamically
reconfigurable – functionality of the brain is to embrace enactivist views
of cognition, with their attendant rejection of mental representation and
computation (Anderson, 2014; Silberstein & Chemero, 2013).
I have tried to argue that one can abandon the standard explanatory
framework of cognitive neuroscience, and its attendant commitments about
280 Daniel C. Burnston
psychological constructs, without abandoning mechanistic explanation in
the brain. And, while I haven’t argued for it here, I claim elsewhere (Burn-
ston, 2020) that this general approach extends to representational explana-
tion as well. The heuristic approach to cognitive ontology is a very different
stance on explanation than is currently assumed in the literature, and I
believe it deserves to be taken as a realistic option in this emerging field.

References
Anderson, M. L. (2014). After phrenology: Neural reuse and the interactive
brain. Cambridge: MIT Press.
Anderson, M. L., Kinnison, J., & Pessoa, L. (2013). Describing functional di-
versity of brain regions and brain networks. Neuroimage, 73, 50–58.
Bechtel, W. (2008). Mental mechanisms: Philosophical perspectives on cogni-
tive neuroscience. New York: Routledge.
Bechtel, W. (2017). Using the hierarchy of biological ontologies to iden-
tify mechanisms in flat networks. Biology & Philosophy. doi:10.1007/
s10539-017-9579-x
Bickle, J. (2003). Philosophy and neuroscience: A ruthlessly reductive account
(Vol. 2). Dordrecht: Springer Science & Business Media.
Boone, W., & Piccinini, G. (2016). The cognitive neuroscience revolution. Syn-
these, 193(5), 1509–1534.
Bunzl, M., Hanson, S. J., & Poldrack, R. A. (2010). An exchange about local-
ism. In Hanson and Bunzl (Eds.), Foundational issues in human brain map-
ping (pp. 49–54). Cambridge: MIT Press.
Burnston, D. C. (2016). A contextualist approach to functional localization in
the brain. Biology & Philosophy, 31(4), 527-550.
Burnston, D. C. (2020). Contents, vehicles, and complex data analysis in neuro-
science. Synthese. doi:10.1007/s11229-020-02831-9
Burnston, D. C. (2021). Getting over Atomism: Functional Decomposition in
Complex Neural Systems. The British Journal for the Philosophy of Science,
72(3), 743-772. doi:10.1093/bjps/axz039
Carlson, T., Goddard, E., Kaplan, D. M., Klein, C., & Ritchie, J. B. (2018).
Ghosts in machine learning for cognitive neuroscience: Moving from data to
theory. Neuroimage, 180, 88–100.
Craver, C. F. (2006). When mechanistic models explain. Synthese, 153(3),
355–376. Retrieved from http://rd.springer.com/article/10.1007/s11229-
006-9097-x
Cummins, R. C. (1983). The nature of psychological explanation. Cambridge:
Bradford/MIT Press.
Darden, L., Pal, L. R., Kundu, K., & Moult, J. (2018). The product guides the
process: Discovering disease mechanisms. In E. Ippoliti & D. Danks (Eds.),
Building theories (pp. 101–117). Dordrecht: Springer.
De Brigard, F. (2014). Is memory for remembering? Recollection as a form of
episodic hypothetical thinking. Synthese, 191(2), 155–185.
Feest, U. (2010). Concepts as tools in the experimental generation of knowledge
in cognitive neuropsychology. Spontaneous Generations: A Journal for the
History and Philosophy of Science, 4(1), 173–190.
 Ontologies in Cognitive Neuroscience 281
Figdor, C. (2011). Semantics and metaphysics in informatics: Toward an ontol-
ogy of tasks. Topics in Cognitive Science, 3(2), 222–226.
Fox, P. T., & Lancaster, J. L. (2002). Mapping context and content: the Brain-
Map model. Nature Reviews Neuroscience, 3, 319. doi:10.1038/nrn789
Gomez-Lavin, J. (2020). Working memory is not a natural kind and cannot ex-
plain central cognition. Review of Philosophy and Psychology. doi:10.1007/
s13164-020-00507-4
Griffiths, P. (2002). Is emotion a natural kind? In R. C. Solomon (Eds.), Think-
ing about feeling (pp. 233–249). Oxford & New York: Oxford University
Press.
Hastings, J., Frishkoff, G. A., Smith, B., Jensen, M., Poldrack, R. A., Lomax, J., . . .
Martone, M. E. (2014). Interdisciplinary perspectives on the development,
integration, and application of cognitive ontologies. Frontiers in Neuroinfor-
matics, 8, 62.
Hutto, D. D., Peeters, A., & Segundo-Ortin, M. (2017). Cognitive ontology in
flux: The possibility of protean brains. Philosophical Explorations, 20(2),
209–223.
Kanwisher, N. (2010). Functional specificity in the human brain: a window into
the functional architecture of the mind. Proceedings of the National Acad-
emy of Sciences of the United States of America, 107(25), 11163–11170.
doi:10.1073/pnas.1005062107
Lenartowicz, A., Kalar, D. J., Congdon, E., & Poldrack, R. A. (2010). Towards
an ontology of cognitive control. Topics in Cognitive Science, 2(4), 678–692.
Leonelli, S. (2012). Classificatory theory in data-intensive science: The case of
open biomedical ontologies. International Studies in the Philosophy of Sci-
ence, 26(1), 47–65.
Machery, E. (2009). Doing without concepts: Oxford University Press.
Murray, J. D., Jaramillo, J., & Wang, X. J. (2017). Working Memory and
­Decision-Making in a Frontoparietal Circuit Model. Journal of Neurosci-
ence, 37(50), 12167–12186. doi:10.1523/JNEUROSCI.0343-17.2017
Piccinini, G., & Craver, C. (2011). Integrating psychology and neuroscience:
Functional analyses as mechanism sketches. Synthese, 183(3), 283–311.
doi:10.1007/s11229-011-9898-4
Poldrack, R. A., Kittur, A., Kalar, D., Miller, E., Seppa, C., Gil, Y., . . . Bilder,
R. M. (2011). The cognitive atlas: Toward a knowledge foundation for
cognitive neuroscience. Frontiers in Neuroinformatics, 5, 17. doi:10.3389/
fninf.2011.00017
Poldrack, R. A., Mumford, J. A., Schonberg, T., Kalar, D., Barman, B., & Yar-
koni, T. (2012). Discovering relations between mind, brain, and mental dis-
orders using topic mapping. PLoS Computational Biology, 8(10), e1002707.
doi:10.1371/journal.pcbi.1002707
Poldrack, R. A., & Yarkoni, T. (2016). From brain maps to cognitive ontologies:
Informatics and the search for mental structure. Annual Review of Psychol-
ogy, 67, 587–612.
Ritchie, J. B., Kaplan, D. M., & Klein, C. (2016). Decoding the brain: Neural
representation and the limits of multivariate pattern analysis in cognitive neu-
roscience. The British Journal for the Philosophy of Science, 70(2), 581-607.
Robins, S. K. (2016). Optogenetics and the mechanism of false memory. Syn-
these, 193(5), 1561–1583.
282 Daniel C. Burnston
Rubin, T. N., Koyejo, O., Gorgolewski, K. J., Jones, M. N., Poldrack, R. A., &
Yarkoni, T. (2017). Decoding brain activity using a large-scale probabilistic
functional-anatomical atlas of human cognition. PLoS Computational Biol-
ogy, 13(10), e1005649. doi:10.1371/journal.pcbi.1005649
Schacter, D. L., Benoit, R. G., De Brigard, F., & Szpunar, K. K. (2015). Episodic
future thinking and episodic counterfactual thinking: Intersections between
memory and decisions. Neurobiology of learning and memory, 117, 14–21.
doi:10.1016/j.nlm.2013.12.008
Shine, J. M., Bissett, P. G., Bell, P. T., Koyejo, O., Balsters, J. H., Gorgolewski,
K. J., . . . Poldrack, R. A. (2016). The dynamics of functional brain networks:
Integrated network states during cognitive task performance. Neuron, 92(2),
544–554. doi:10.1016/j.neuron.2016.09.018
Shine, J. M., Breakspear, M., Bell, P. T., Martens, K. E., Shine, R., Koyejo,
O., . . . Poldrack, R. A. (2018). The low dimensional dynamic and integrative
core of cognition in the human brain. bioRxiv, 266635. doi:10.1101/266635
Shine, J. M., & Poldrack, R. A. (2017). Principles of dynamic network re-
configuration across diverse brain states. Neuroimage. doi:10.1016/j.
neuroimage.2017.08.010
Silberstein, M., & Chemero, T. (2013). Constraints on localization and decom-
position as explanatory strategies in the biological sciences. Philosophy of
Science, 80(5), 958–970.
Sochat, V. V., Eisenberg, I. W., Enkavi, A. Z., Li, J., Bissett, P. G., & Poldrack,
R. A. (2016). The experiment factory: Standardizing behavioral experiments.
Frontiers in Psychology, 7, 610. doi:10.3389/fpsyg.2016.00610
Sullivan, J. A. (2010). Reconsidering ‘spatial memory’ and the Morris water
maze. Synthese, 177(2), 261–283. Retrieved from http://rd.springer.com/
article/10.1007/s11229-010-9849-5
Sullivan, J. A. (2014). Stabilizing mental disorders: prospects and problems.
In H. Kincaid & J. A. Sullivan (Eds.), Classifying Psychopathology: Mental
Kinds and Natural Kinds (pp. 257 - 281). Cambridge, MA: MIT Press.
Sullivan, J. A., Dumont, J. R., Memar, S., Skirzewski, M., Wan, J., Mofrad, M.
H., . . . Prado, V. F. (2021). New frontiers in translational research: Touch-
screens, open science, and the mouse translational research accelerator plat-
form. Genes, Brain and Behavior, 20(1), e12705.
Tekin, Ş. (2016). Are mental disorders natural kinds?: A plea for a new approach
to intervention in psychiatry. Philosophy, Psychiatry, & Psychology, 23(2),
147–163.
Teodorescu, A. R., & Usher, M. (2013). Disentangling decision models: From
independence to competition. Psychological review, 120(1), 1.
Turner, J. A., & Laird, A. R. (2012). The cognitive paradigm ontology: De-
sign and application. Neuroinformatics, 10(1), 57–66. doi:10.1007/
s12021-011-9126-x
Uttal, W. R. (2001). The new phrenology: The limits of localizing cognitive
processes in the brain. Cambridge: The MIT Press.
Varoquaux, G., Schwartz, Y., Poldrack, R. A., Gauthier, B., Bzdok, D., Poline,
J. B., & Thirion, B. (2018). Atlases of cognition with large-scale human brain
mapping. PLoS Computational Biology, 14(11), e1006565. doi:10.1371/
journal.pcbi.1006565
 Ontologies in Cognitive Neuroscience 283
Yarkoni, T., Poldrack, R. A., Nichols, T. E., Van Essen, D. C., & Wager, T. D.
(2011). Large-scale automated synthesis of human functional neuroimaging
data. Nature methods, 8(8), 665.
Yarkoni, T., & Westfall, J. (2017). Choosing prediction over explanation in
psychology: Lessons from machine learning. Perspectives on Psychological
Science, 12(6), 1100–1122. doi:10.1177/1745691617693393
Section 4

Tools and Integrative

Pluralism
13 “It Takes Two to Make a
Thing Go Right”
The Coevolution of
Technological and
Mathematical Tools in
Neuroscience
Luis H. Favela

It takes two to make a thing go right.

Rob Base and DJ E-Z Rock, It takes two (Ginyard, 1988)

1 Introduction
There should be no doubt that technological developments have played
significant roles throughout the history of scientific discoveries and prog-
ress. This is as true in the physical sciences (e.g., particle accelerators in
physics) as in the life sciences (e.g., microscopes in biology). What is less
apparent is the role mathematical developments have played in facilitat-
ing and supporting many of those discoveries. Mathematical tools for
analyzing data may not be at the forefront of discoveries centering on the
physical structure of investigative targets of interest (e.g., cells); but they
certainly are crucial in research focused on the dynamics of phenomena
(e.g., planetary motion). Consequently, for science to progress, research
on the movement and temporal aspects of phenomena often requires the
coevolution of technological and mathematical tools.
Recently, it has been increasingly argued by some philosophers of neu-
roscience that experimental tools are not just important but are funda-
mental to neuroscience research (e.g., Barwich, 2020; Bickle, 2016; Silva,
Landreth, & Bickle, 2014). Put in its sharpest terms, the line of thought
goes like this: From Golgi’s staining technique to functional magnetic
resonance imaging, and from deep brain stimulation to optogenetics,
the history of neuroscience is principally a history of tool development.
Moreover, it has been argued that this history is best characterized as
one that exhibits reductionist (Bickle, 2006, 2016) and mechanistic
explanations (Craver, 2002, 2005). Across these claims, little to no
mention of data analysis methods are mentioned nor the underlying
assumptions of those techniques. Here, I argue that the mathematical

DOI: 10.4324/9781003251392-18
288 Luis H. Favela
assumptions of applied data analyses have played crucial—though often
­underappreciated—roles in the history of neuroscience. First, I present
the Hodgkin and Huxley model of action potentials as an example of re-
search constrained by technological and mathematical limitations of its
time. Second, I draw attention to a feature of neurons that is overlooked
by the Hodgkin-Huxley model: scale-invariant dynamics. After describ-
ing scale-invariant dynamics, I then point out consequences scale-invari-
ant neuronal dynamics have for explanatory approaches in neuroscience
that rely on—what can be broadly described as—­decomposition strat-
egies of neuronal activity. I conclude by emphasizing the necessity of
mathematical developments in providing more appropriate and encom-
passing explanations of neural phenomena in toto.

2 Single Neuron Models

The canonical Hodgkin and Huxley (1952) model of action potentials
in the squid giant axon is considered not just “the single most successful
quantitative model in neuroscience” (Koch, 1999, p. 171), but also “one
of the great success stories in biology” (Häusser, 2000, p. 1165). Many of
the details of the model are not essential for my current aims (Figure 1a).
For detailed explanations of this model see Gerstner, Kistler, Naud, and
Paninski (2014), as well as Koch (1999) for discussion and further refer-
ences. For now, it is important to understand that this model treats the
action potential as an “all-or-none” event (e.g., Bear, Connors, & Para-
diso, 2016; Churchland & Sejnowski, 1992/2017; Eagleman & Downar,
2016). The action potential is treated as a binary event that occurs within
distinctly defined timescales (Figure 1b). Moreover, those timescales
have a lower boundary, specifically, 10 milliseconds (ms) in the canon-
ical Hodgkin-Huxley model (Hodgkin & Huxley, 1952, p. 528; Koch,
1999, p. 334; Marom, 2010, p. 23). What that means is that the action
potential of a neuron (i.e., its “spike” of activity) is treated within the
Hodgkin-Huxley model as occurring at least 10 ms from initiation to
termination of all involved processes (Marom, 2010, p. 22).
As was known then and now (e.g., Marom, 2010), although there
were empirically justifiable reasons at the time (e.g., Adrian & Zotter-
man, 1926), defining the action potential as a 10 ms event was due to
investigator observational preferences in combination with technologi-
cal limitations. Observational preferences were constrained by the lim-
its of the recording technology, namely, the voltage clamp. Although
the voltage clamp was instrumental in providing the data that lead to
the development of the Hodgkin-Huxley model, it was limited in its
ability to record the full range of ion channels, charged particles, and
other physiologically relevant features of neuronal activity (Schwiening,
2012). This resulted in the need to sum across molecular activity (Gerst-
ner et al., 2014)—certainly a necessity when calculating at the molecular
 “It Takes Two to Make a Thing Go Right” 289

Figure 1 H
 odgkin-Huxley model. (a) The canonical Hodgkin and Huxley
(1952) model of action potentials in the squid giant axon. Definitions
of key model variables (box). (b) The basic shape of an action potential
as produced by the Hodgkin-Huxley model (created with MATLAB
[MathWorks®, Natick, MA] script based on Kothawala, 2015). The
x-axis captures the entire range of time in which an action potential
occurs. According to the model, the lower temporal boundary of an
action potential is 10 ms. This means that the entire event, from start
to finish, occurs within that time frame.

scale—and collapse other physiological features into imprecise “leak”

terms, a sort of “catch all” variable used in models that have causally
relevant features that have not been precisely measured. Other limita-
tions involved the manner in which the data was calculated. Hodgkin
and Huxley calculated data from the voltage clamp via hand calcula-
tors (Koch, 1999, p. 160). Specifically, Hodgkin and Huxley utilized a
290 Luis H. Favela

Figure 2 T
 he Brunsviga 20, “one of the most popular mechanical calculators.
It was produced up to the early 1970s and marketed with the slo-
gan ‘Brains of Steel’” (Schwiening, 2012). (Reprinted with permission
from Wikipedia. CC BY-SA 2.0 DE.)

mechanical calculator, the Brunsviga 20 (Figure 2), which required them

to spend a few weeks and many thousands of rotations of the machine’s
crank in order to carry out the calculations (Schwiening, 2012).
Although the canonical Hodgkin-Huxley model is described by
some as being linear in nature (e.g., Gerstner et al., 2014; Hodgkin
& Huxley, 1952, pp. 538–540), there is debate about whether or not
it is able to capture the relevant types of nonlinearities exhibited by
feedback that are now established as occurring during action poten-
tials (e.g., Marom, 2010; Schwiening, 2012).1 Regardless of whether
or not the canonical Hodgkin-Huxley model is linear or nonlinear,
or can capture particular forms of feedback, it is clear now that even
single neurons are appropriately understood as nonlinear systems (e.g.,
Izhikevich, 2007).
Advancements in recording technologies have facilitated the ability of
neuroscientists to obtain more detailed data on neuronal activity (e.g.,
multielectrode arrays [MEA]; Gross, 2011), making it possible to re-
cord more detailed and accurate data from longer timescales of neuron
activity. As a result, it is becoming increasingly evident that the rele-
vant timescales for explaining even “basic” single-neuron activity can
require looking far below and above that 10 ms window presumed to
 “It Takes Two to Make a Thing Go Right” 291
characterize single neuron spikes. Action potentials do not appear to
have strictly defined windows of activity, specifically, nonlinearities
in the forms of feedback and hysteresis significantly contribute to the
event. 2 Instead of viewing action potentials as having clear startup and
finish conditions (Figure 1b), it is more accurate to view action poten-
tials as continuous, nonlinear cycles. This is clearly depicted in models
as early as the FitzHugh-Nagumo model (FitzHugh, 1961; Nagumo,
Arimoto, & Yoshizawa, 1962; Figure 3a) and more recent models, such
as the Izhikevich model (Izhikevich, 2007; Figure 3b).

3 Scale-Invariant Neuronal Dynamics

As mentioned in the previous section, there is debate as to the degree
or not that the canonical Hodgkin-Huxley model accounts for a wide
range of nonlinear features of action potentials, such as hysteresis. I am
not entering that debate here. The primary purpose for presenting the
Hodgkin-Huxley model was to provide an example of an achievement
in neuroscience that was constrained by the technology and related
mathematical techniques of its time (e.g., Brunsviga 20; Figure 2). Addi-
tionally, I also mentioned that some earlier models of single-neuron ac-
tivity, like the FitzHugh-Nagumo model, were able to explicitly account
for certain forms of nonlinearity (FitzHugh, 1961; Nagumo et al., 1962;
Figure 3a). In this section, I focus on scale invariance in neuronal ac-
tivity, which is a more recent finding in neuroscience. Scale invariance
was discovered in large part due to both improved technologies (e.g., re-
cording techniques) and mathematical techniques (e.g., fractal analysis;
more on that below).
At its most general, a phenomenon is “scale invariant” when its struc-
ture (i.e., spatial and/or temporal) is self-similar from various points of
observation (Bak, 1996; Gisiger, 2001). Scale invariance can be abstract
or statistical. Many illustrative examples of abstract scale-invariant spa-
tial structures are found in fractal geometry (Mandelbrot, 1983; Figure
4). The Koch curve (i.e., Koch snowflake, Koch triangle)—like the Can-
tor and Mandelbrot sets—exhibits a topology that is perfectly recre-
ated at each iteration or scale of observation. That is to say, the overall
“curved, triangle-like” shape appears at one spatial dimension (Figure
1a), is repeated at a higher dimension (Figure 1b), and again at another
dimension (Figure 1c). Scale invariance need not be perfectly precise as
in abstract fractals like the Koch curve. Scale invariance can be defined
in statistical terms. A phenomenon is statistically scale invariant when
certain properties are repeated at various scales. This is the type of scale
invariance found in natural structures such as bronchial tubes, coast-
lines, and unfurling ferns.
In recent years, an illustrative example of increased research on scale
invariance are the now well-known scale-invariant properties prevalent
292 Luis H. Favela

Figure 3 M
 odels of nonlinear single-neuron activity. (a) FitzHugh-Nagumo
model and phase space portrait. (Modified and reprinted with per-
mission from Scholarpedia. CC BY-NC-SA 3.0.). (b) Izhikevich model
and phase space portrait. Note that the phase space portrait depicts
a strange attractor as the Izhikevich model parameters are set to de-
pict chaotic time evolution. The FitzHugh-Nagumo model is not able
to depict chaotic neuronal behavior no matter how the parameters
are tuned. (Modified and reprinted with permission from Nobukawa,
Nishimura, Yamanishi, & Liu, 2015.)
 “It Takes Two to Make a Thing Go Right” 293

Figure 4 T
 he Koch curve is an example of an abstract spatial fractal. Here,
three iterations of self-similarity are depicted (a, b, and c). (Modified
and reprinted with permission from Wikipedia. CC BY-SA 3.0.)

in network structures. Scale-invariant networks have few nodes with

many connections and many nodes with few connections (Figure 5b).
Consequently, such networks have no specific or average number of
connections that characterize the entire system. Since such networks
resist being properly characterized by standard statistical means (e.g.,
averages), scale-invariant networks are more accurately described in
terms of power-law distributions (He, 2014). It has become commonly
accepted that many phenomena and systems of diverse composition are
scale invariant in this way, for example, cellular metabolism, Hollywood
actors that have worked together, sexual relationships, and the World
Wide Web (Barabási & Bonabeau, 2003). There is increasing evidence
that neural systems exhibit scale invariance (e.g., Boonstra, He, & Daf-
fertshofer, 2013; Di Ieva, 2016; He, 2014; Poli, Pastore, & Massobrio,
2015), such as neuron branching patterns and neuronal network con-
nections (Figure 5d). For current purposes, I focus on the scale invari-
ance exhibited by neuronal dynamics (for a wide range of examples see
Boonstra et al., 2013). In short, neuronal dynamics are considered scale
294 Luis H. Favela

Figure 5 R
 andom networks and scale-invariant networks. (a) Abstract random
network illustrating that most nodes have a comparable number of
connections. (b) Abstract scale-invariant network illustrating many
nodes with a small number of connections and few nodes with a high
number of connections. Dissociated neurons developing in vitro and
coupled to multielectrode array (MEA), exhibiting (c) random net-
work connectivity and (d) scale-invariant connectivity, with a highly
connected unit (center), or hub. (Modified and reprinted with permis-
sion from Josserand et al., 2021 and Poli et al., 2015. CC BY.)

invariant when there is no single or average time scale that properly

characterizes its activity, which includes attempting to define an event
as occurring within specific windows of time. There are several con-
sequences that result from the fact that many neural systems exhibit
scale-invariant spatial and/or temporal structures. In the next section, I
explore one such consequence, specifically, the inability of investigative
and explanatory approaches in neuroscience that rely on decomposition
strategies to properly account for scale-invariant neuronal dynamics.

4 Consequences of Scale-Invariant Dynamics for

Explanations in Neuroscience
In a recent paper, Bechtel (2015) argues against the claim that scale-
invariant biological phenomena cannot be explained mechanistically.3
He rejects the following argument, which I summarize as follows:

1 Mechanistic explanations require that the phenomena being ex-

plained have well-defined boundaries, such as a temporal boundary.
 “It Takes Two to Make a Thing Go Right” 295
2 Many biological phenomena exhibit scale-invariant features.
3 Scale-invariant phenomena have no well-defined temporal
boundaries.
4 Therefore, scale-invariant biological phenomena cannot be ex-
plained mechanistically.

Marom (2010) presents such an argument and serves as one of Bechtel’s

targets. Marom argues that there is empirical evidence suggesting that
neuronal activity is scale invariant and, thus, is just the type of biological
phenomenon that cannot be explained mechanistically. Marom’s argu-
ment includes discussion of the Hodgkin-Huxley model, which leads
him to conclude:

Indeed, the lesson from our journey across levels of organization, from
behavior through neural assemblies to single neurons and proteins,
suggests that dreams on all-encompassing microscopic timescale-based
descriptions, aimed at explaining the temporal richness of macroscopic
levels, should be abandoned. Other approaches are called for.
(2010, p. 23)

In short, Marom claims that there are no uniquely defined timescales

that could justify defining action potentials as events that have a lower
boundary of 10 ms. Consequently, macroscale neuronal activities that
appear as scale invariant are not merely the result of additive or linear
combinations of microscale contributions. Instead, they are truly scale
invariant: the micro timescales contribute to and constrain the macro
timescales, but so too does the macro contribute to and constrain the
micro, such that no single scale serves a more fundamental explanatory
role than the others.
Bechtel’s reply to Marom is that scale-invariant phenomena can still be
explained mechanistically. But to do so requires that we appreciate the
role of mechanisms in scientific practice. According to Bechtel, scientists
often posit “bounded mechanisms” for the purposes of testing hypothe-
ses (2015, pp. 84–85). A scientist can understand that a phenomenon is
interconnected (e.g., networks) and still pursue a mechanistic account of
that phenomenon by drawing boundaries around that organism. Those
bounded mechanisms are not abstractions, however. “Abstractions,” ac-
cording to Bechtel, leave information out. Instead, those bounded mech-
anisms are idealizations. Idealizations, according to Bechtel, are models
with simplifying falsehoods (2015, p. 85). For example, if phenomenon
X is understood to be highly interconnected, an explanation of X that
assumes that it is not affected by all its connections would be an abstrac-
tion. But to localize X to, for example, its nearest neighbors is to provide
a “first approximation” (2015, p. 85; italics in original) that appreci-
ates the practical challenges of accounting for all the actual connections.
Such an explanation would be both an idealization and a mechanism.
296 Luis H. Favela
Although he accepts that neuronal dynamics can be scale invariant,
Bechtel remains committed to providing mechanistic explanations of
those dynamics. Accordingly, Bechtel remains committed to mechanisms
being bounded, on the further stipulation that such bounded mecha-
nisms are idealizations and not abstractions. For example, the action
potential is a “bounded mechanism” that occurs within 10 ms windows.
Such an idealization is purported to be acceptable because it makes the
timescales of that phenomenon tractable to investigators’ cognitive lim-
itations (2015, p. 92)—and I would add technological limitations. Thus,
the Hodgkin-Huxley model can be understood as an idealization of ac-
tion potentials, with the 10 ms feature being a simplifying falsehood—
though not an abstraction that leaves out relevant features. This is a very
streamlined presentation of Bechtel’s argument, for example, he makes
a further claim that such idealized mechanisms can point out areas for
further investigation in a mechanistic explanation. What matters for my
current purposes is Bechtel’s attempt to make room within mechanistic
accounts to explain the scale-invariant activity.
There is a lot in Bechtel’s reply to Marom to agree with, for example,
the fact that scientists are epistemically limited creatures who need to
simplify some phenomena in order to get an intellectual grip on them.
However, I think Bechtel’s reply overlooks a central issue raised by
Marom. If mechanisms are, by definition, bounded, then scale-invariant
phenomena (e.g., flicker noise, power laws; Gisiger, 2001) are, by defi-
nition, not mechanisms. In the case of action potentials, the canonical
Hodgkin-Huxley model sets a lower boundary on the phenomenon at
10 ms. In other words, it treats action potentials as starting and finish-
ing within windows of time of at least 10 ms (Figure 1a). As discussed
above, such a claim was justified as being consistent with the best science
of the time (e.g., Adrian & Zotterman, 1926). With that said, it was con-
strained by technological (i.e., voltage clamp) and mathematical (i.e., the
type of calculations that could be conducted on a Brunsviga 20 calcu-
lator; Figure 2) limitations. Technological advancements have certainly
played a role in revealing scale-invariant dynamics, both within single
neurons and across neuronal populations (e.g., MEA; Gross, 2011).
However, data produced by advanced equipment alone does not provide
a substantial reason to justify the existence of scale-invariant dynam-
ics in neuronal systems. The other part needed for the right account—­
remember, “it takes two to make a thing go right”—is the pairing of
data from suitable technology with the appropriate mathematical tools.
In the case of single-neuron activity, the right mathematical tools are
those from nonlinear dynamical systems theory (NDST; e.g., Favela,
2020a, 2020b, 2020c; Izhikevich, 2007; Liebovitch & Toth, 1990).
NDST methods are crucial to assessing scale-invariant structures and
can contribute to establishing whether a phenomenon is justifiably
defined as scale-invariant or not and, if so, what kind of scale-invari-
ant characteristics it has. What is more, applying NDST methods to
 “It Takes Two to Make a Thing Go Right” 297
complex and nonlinear phenomena typically requires powerful com-
puters. For example, generating phase portraits of relatively simple
two-­dimensional dynamical systems was often not practical before com-
puters. Hodgkin and Huxley’s “Brains of Steel” mechanical calculator
was certainly not up to the task (Figure 2). In view of that, the Izhikevich
model of s­ ingle-neuron activity required both the appropriate processing
power (i.e., modern computers) and data analysis methods (i.e., NDST)
to provide qualitative and quantitative accounts of that phenomenon’s
nonlinear dynamics. Additionally, considerable computational process-
ing power is needed to analyze and model the more complex and nonlin-
ear features of neurons like chaotic dynamics (Figure 3b).
As mentioned above, nonlinear dynamics are not my central topic;
scale-invariant dynamics are. Scale-invariant properties are a particu-
larly unique set of phenomena with regard to the need for coevolving
technological and mathematical tools. Many aspects of mammalian bi-
ological phenomena exhibit scale-invariant structures, such as eye sac-
cades, heart beats, neuronal networks, and postural sway. Accordingly,
different mathematical tools are needed to properly determine the ways
they are scale invariant. For example, detrended fluctuation analysis
(Peng et al., 1994) can assess structural self-similarity in a signal; but it
will not necessarily make it clear if the structure results from linear or
nonlinear processes (Bryce & Sprague, 2012). In the case of appropriate
mathematical methods for assessing the scale-invariant dynamics of ac-
tion potentials, if such activity is, for example, fractal, then it would not
have been possible to accurately analyze such data, regardless of techno-
logical advancements, until the 1980s. The reason is because the concept
“fractals” and related data analytic methods were not introduced to the
broader scientific community until then (Mandelbrot, 1983).
In order to identify self-similar features of scale-invariant structures,
whether resulting from linear or nonlinear processes, the concept “frac-
tals” and their measurement must be part of an investigator’s toolbox.
Fractals, such as the Koch curve (Figure 4) are paradigmatic examples
of scale invariance: the overall structure of the system is maintained at
each level of observation. Such phenomena are thus not appropriately ex-
plained in terms that, for example, treat them as having an average value.
Instead, as Mandelbrot pointed out, such phenomena are appropriately
characterized via a fractal dimension. The fractal dimension provides a
quantitative means of characterizing scale invariance that accounts for
all of its scales. The equation for calculating the fractal dimension is:

n= 1
Sd (1)

It is helpful to explain this equation via the Koch curve. For demonstra-
tion purposes, we will look at a four-lined Koch curve (Figure 4a). Here
n is the number of line segments at a particular scale of observation;
298 Luis H. Favela
in this case, it is 4. Next, S is the scale factor, or the size reduction
at each iteration; here it is 1/3. Our equation is now: 4 = 1/(1/3)d, or
4 = 3d. We want to figure out d, or the fractal dimension. To do so, we
take the log of both sides: d = log 4/log 3, which gives us a fractal dimen-
sion d = 1.26. In English, this means that the fractal dimension of the
Koch curve is 1.26, which means it is not a straight line (d = 1) or a square
(d = 2), but closer to being a straight line than a square (d = 1.26). There
are various other methods for mathematically assessing fractals and
multifractals (Lopes & Betrouni, 2009).
The point of this example is to demonstrate that before Mandelbrot’s
invention—or, perhaps, discovery—of fractal geometry, it was not pos-
sible to appropriately account for scale-invariant phenomena, for exam-
ple, via more standard statistical tools such as calculating the arithmetic
mean. The consequence for neuronal activity is that it was not until the
1990s (e.g., Liebovitch & Toth, 1990) that scale-invariant dynamics
could be properly identified. Before then, such properties were misidenti-
fied via other statistical methods. Since scale-invariant structures have no
primary scale or average scale, they have no specific window to identify
as the start and finish boundary. Such a view of neuronal activity is fur-
ther evidenced by other NDST-based work, such as the Izhikevich model
(2007; Figure 3b), which treats action potentials as continuous cycles and
not binary, “all-or-none” events (cf. Figure 1b). If true, that is, if action
potentials are not bounded within discrete windows of time, then action
potentials cannot be accounted for via decomposition strategies, such as
those common to mechanistic approaches in actual scientific practice.4
In concluding this section, an important clarification needs to be made
in order to address a significant critique of the current line of argument.
The critique centers on the notion of “bounded” in regard to natural
phenomena. As discussed above, the currently relevant aspect of the
Bechtel/Marom debate centers on the idea that mechanistic explana-
tions treat targets of investigation as bounded, namely, as having delin-
eated borders, which can be spatial or temporal. The Hodgkin-Huxley
model of action potentials and its 10 ms event window were presented
as an example of such a bounded mechanism. Scale-invariant neuronal
dynamics was presented as an unbounded natural phenomenon, which
means it is not accessible to mechanistic explanation (i.e., if “mechanis-
tic explanations” include the stipulation of boundedness; see Bechtel,
2015; Marom, 2010). The critique of this line of argument centers on
the point that even scale-invariant neuronal dynamics are “bounded” in
a number of ways, for example, there is a window of time in which they
occur (e.g., they do not last for months, years, or centuries) and they
are spatially confined (e.g., they occur in an area of the brain, and not
across the whole brain, let alone body and world). This is an understand-
ably compelling critique. However, it does not concern the way in which
scale-invariant dynamics are “unbounded.”
 “It Takes Two to Make a Thing Go Right” 299
The way in which scale-invariant dynamics are unbounded involves
the inability of single, bounded values to characterize the phenomenon.
A time series (Figure 6) need not be infinite nor recorded from an event
that has no measurable spatial location in order to be scale invariant.
A scale-invariant time series exhibits the same pattern among windows
of various lengths of time. For example, if a heartbeat shows a pattern
of activity over 60 minutes, then, to be considered scale invariant, that
(statistically) same pattern should be shown in each of two 30-minute
windows of time, at each of four 15-minute windows, and so on. In that
way, the time series is not properly understood as “bounded” in that
there is no single length of time that characterizes the entire signal. That
is to say, it is not correct to treat the event as a bounded 60-minute event,
or a 30-minute event, and so on; but in terms of the structure of the
patterns across various scales. It is in that sense that Marom argues that
neuronal dynamics do not have timescales, and it is in that sense that
they are unbounded, and, thus, not properly explained mechanistically.

5 Conclusion
It is highly unlikely to find disagreement among the scientific research
community that technological advancements have paved the way
for some of the greatest advances and discoveries. What is less often
­acknowledged—especially in neuroscience—is the necessity of coevolv-
ing our mathematical tools with technological advances, and vice versa.
Consequently, technological advancements that produce more detailed
and accurate data recording will not alone necessarily provide proper ex-
planations of biological phenomena. Mathematical tools like those pro-
vided by NDST are needed as well in order to properly characterize data.
The Hodgkin-Huxley model was informed and constrained by the avail-
able technological (i.e., voltage clamp) and mathematical (i.e., Brunsviga
20 calculator) tools of the time. Since then, more advanced technology
(e.g., multielectrode arrays) and mathematics (e.g., fractal analysis)
have highlighted some of the limitations of the Hodgkin-­Huxley model
as a comprehensive model of action potentials across temporal scales.
Scale-invariant neuronal activity provides a rich example of this. In or-
der to identify scale-invariant activity, researchers needed more accurate
measurements (e.g., MEA), data analyses (e.g., detrended fluctuation
analysis), and—in this case—new concepts altogether. In order to prop-
erly account for scale-invariant activity, a new concept—namely, fractals
and the fractal dimension—was needed, as was accompanying innovative
mathematical analyses. One consequence of the existence of scale-invari-
ant neuronal activity discussed here involves the limitations of research
approaches centered on decomposition strategies—i.e., “mechanistic”
approaches understood in terms of actual scientific practice—to account
for phenomena that are without discrete (i.e., “bounded”) temporal
300 Luis H. Favela

Figure 6 T
 ime series exhibiting statistical scale invariance at multiple windows
of time. Data obtained from the spontaneous activity of single neurons
(i.e., mitral cells) in the rat main olfactory bulb are shown (Stakic,
Suchanek, Ziegler, & Griff, 2011). Synthetic time series reproduced
from results of detrended fluctuation analysis (Favela, Coey, Griff, &
Richardson, 2016). Statistical scale invariance demonstrated in win-
dows based on power of two: (a) 8,192 seconds, (b) 2,048 seconds, and
(c) 1,024 seconds. The overall structure of the time series is repeated
within each window of time. As a result, the time series is not properly
characterized via a single value (e.g., mean) or window of time.
 “It Takes Two to Make a Thing Go Right” 301
boundaries. In conclusion, an attempt has been made here to motivate
the claim that progress in neuroscience takes two things to make it go
right, namely, technological and mathematical advancements.

Acknowledgments
I would like to thank Ann-Sophie Barwich, John Bickle, and Carl Craver
for their interest in this project. Thanks to John Beggs for many conversa-
tions about several aspects of this material; I’ve learned a great deal from
him. Thanks to Mary Jean Amon for assistance with creating Figure 6.
Various stages and versions of this paper have benefited from constructive
feedback by audiences at the Tool Development in Experimental Neuro-
science: A Science-in-Practice Workshop & the 57th Annual Meeting
of the Alabama Philosophical Society (2019, September), the Center for
Philosophy of Science at the University of Pittsburgh (2019, October),
NeuroTech: An Interdisciplinary Early Career Workshop on Tools and
Technology in Neuroscience (2020, January), and the Neuroscience Alli-
ance at the University of Central Florida (2020, February).

Notes
1 “Linearity” can be understood as a mathematical relationship among data
(e.g., time series), in that each successive data point is additively related to its
previous point. Such phenomena exhibit outputs that are proportional to in-
puts (Lam, 1998). “Nonlinearity” can also be understood as a mathematical
relationship among data. But here, each successive data point can observe a
variety of relationships to previous points, such as exponential or multipli-
cative (Lam, 1998; May, 1976; Stam, 2005). Other, more interesting, forms
of nonlinearity found in biological systems include patterned dynamics
(e.g., fractals; Di Ieva, 2016) and phase transitions (e.g., catastrophe theory;
Isnard & Zeeman, 1976/2013).
2 Hysteresis is a nonlinear phenomenon. There are multiple forms of hystere-
sis that can be exhibited by both biological and nonbiological systems. All
forms are characterized by their being constrained by historical variation,
that is, when a system’s current state is strongly dependent on its history
(Haken, 1983). Ferromagnetic materials (e.g., common magnets that hold
up pictures on your refrigerator) provide clear illustrations of strong histori-
cal dependence characteristic of hysteresis effects (Dris, 2016). If a material
such as iron is magnetized, once it is demagnetized, then it will require dif-
ferent magnetic fields to re-magnetize it. The reason the same magnetic field
will not magnetize the piece of iron again is because the “system” (i.e., its
atomic composition) “remembers” its previous state, that is, is constrained
by states it has been in before. Thus, the current state of the system depends
on its previous state such that magnetization does not occur at absolute,
context-free values.
3 Bechtel’s (2015) use of “scale free” and the current of “scale invariant” refer
to the same phenomenon. The terms are commonly used interchangeably
in the relevant literatures (e.g., Barabási, 2016; Broido & Clauset, 2019;
Serafino et al., 2021). For that reason, when discussing Bechtel’s argument,
I will switch out “scale invariance” for “scale free” for consistency of termi-
nology in the current work.
302 Luis H. Favela
4 It is extremely challenging to define “mechanism” or “mechanistic explana-
tion” in a way satisfactory to most philosophers of science. In the current
work, I limit my application of the concept of “mechanism” to ways consis-
tent with actual scientific practice, that is, in the sense discussed by Bechtel
(2015). For that reason, I emphasize decomposition as common to the ways
scientists investigate and define “mechanisms” (e.g., Andersen, 2014; Illari
& Williamson, 2012; Khambhati, Mattar, Wymbs, Grafton, & Bassett,
2018). That is to say, scientists commonly aim to decompose—or “break
apart”—a target of investigation in order to understand it in terms of its
parts, namely, the constitution, contribution, and interaction of those parts.

References
Adrian, E. D., & Zotterman, Y. (1926). The impulses produced by sensory
nerve endings. Part 3. Impulses set up by touch and pressure. The Journal of
Physiology, 61(4), 465–483.
Andersen, H. (2014). A field guide to mechanisms: Part I. Philosophy Compass,
9(4), 274–283.
Bak, P. (1996). How nature works: The science of self-organized criticality.
New York, NY: Springer-Verlag.
Barabási, A.-L. (2016). Network science. Cambridge University Press. Retrieved
July 7, 2021 from http://networksciencebook.com
Barabási, A.-L., & Bonabeau, E. (2003). Scale-free networks. Scientific Amer-
ican, 288(5), 60–69.
Barwich, A. S. (2020). What makes a discovery successful? The story of Linda
Buck and the olfactory receptors. Cell, 181(4), 749–753. doi:10.1016/j.
cell.2020.04.040
Bear, M. F., Connors, B. W., & Paradiso, M. A. (2016). Neuroscience: Explor-
ing the brain (4th ed.). New York, NY: Wolters Kluwer.
Bechtel, W. (2015). Can mechanistic explanation be reconciled with scale-free
constitution and dynamics? Studies in History and Philosophy of Biological
and Biomedical Sciences, 53, 84–93.
Bickle, J. (2006). Reducing mind to molecular pathways: Explicating the re-
ductionism implicit in current cellular and molecular neuroscience. Synthese,
151, 411–434.
Bickle, J. (2016). Revolutions in neuroscience: Tool development. Frontiers in
Systems Neuroscience, 10(24). doi:10.3389/fnsys.2016.00024
Boonstra, T. W., He, B. J., & Daffertshofer, A. (2013). Scale-free dynamics and
critical phenomena in cortical activity. Frontiers in Physiology: Fractal and
Network Physiology, 4(79). doi:10.3389/fphys.2013.00079
Broido, A. D., & Clauset, A. (2019). Scale-free networks are rare. Nature Com-
munications, 10(1017), 1–10. doi:10.1038/s41467-019-08746-5
Bryce, R. M., & Sprague, K. B. (2012). Revisiting detrended fluctuation analy-
sis. Scientific Reports, 2(315), 1–6. doi:10.1038/srep00315
Churchland, P. S., & Sejnowski, T. J. (1992/2017). The computational brain
(25th anniversary ed.). Cambridge: The MIT Press.
Craver, C. F. (2002). Interlevel experiments and multilevel mechanisms in the
neuroscience of memory. Philosophy of Science, 69(S3), S83-S97.
Craver, C. F. (2005). Beyond reduction: Mechanisms, multifield integration and
the unity of neuroscience. Studies in History and Philosophy of Science Part
 “It Takes Two to Make a Thing Go Right” 303
C: Studies in History and Philosophy of Biological and Biomedical Sciences,
36(2), 373–395.
Di Ieva, A. (Ed.). (2016). The fractal geometry of the brain. New York: Springer.
Dris, S. (2016). Magnetic hysteresis. LibreTexts. Retrieved July 7, 2021, from
https://eng.libretexts.org/@go/page/333
Eagleman, D., & Downar, J. (2016). Brain and behavior: A cognitive neurosci-
ence perspective. New York, NY: Oxford University Press.
Favela, L. H. (2020a). Cognitive science as complexity science. Wiley Interdisci-
plinary Reviews: Cognitive Science, 11(4), e1525, 1–24. doi:10.1002/WCS.1525
Favela, L. H. (2020b). Dynamical systems theory in cognitive science and neuro-
science. Philosophy Compass, 15(8), e12695, 1–16. doi:10.1111/phc3.12695
Favela, L. H. (2020c). The dynamical renaissance in neuroscience. Synthese.
doi:10.1007/s11229-020-02874-y
Favela, L. H., Coey, C. A., Griff, E. R., & Richardson, M. J. (2016). Fractal analysis
reveals subclasses of neurons and suggests an explanation of their spontaneous
activity. Neuroscience Letters, 626, 54–58. doi:10.1016/j.neulet.2016.05.017
FitzHugh, R. (1961). Impulses and physiological states in theoretical models of
nerve membrane. Biophysical Journal, 1(6), 445–466.
Gerstner, W., Kistler, W. M., Naud, R., & Paninski, L. (2014). Neuronal dy-
namics: From single neurons to networks and models of cognition. Cam-
bridge: Cambridge University Press.
Ginyard, R. (1988). It takes two [Recorded by R. Ginyard (Rob Base) and R.
Bryce (DJ E-Z Rock)]. On It takes two [Vinyl]. Englewood, NJ: Profile Records.
Gisiger, T. (2001). Scale invariance in biology: Coincidence or footprint of a
universal mechanism? Biological Reviews, 76, 161–209.
Gross, G. W. (2011). Multielectrode arrays. Scholarpedia, 6(3), 5749.
doi:10.4249/scholarpedia.5749
Haken, H. (1983). Synergetics: An introduction (3rd ed.). Berlin: Springer-Verlag.
Häusser, M. (2000). The Hodgkin-Huxley theory of the action potential.
Nature Neuroscience, 3(11), 1165. doi:10.1038/81426
He, B. J. (2014). Scale-free brain activity: Past, present, and future. Trends in
Cognitive Sciences, 18(9), 480–487.
Hodgkin, A. L., & Huxley, A. F. (1952). A quantitative description of mem-
brane current and its application to conduction and excitation in nerve. The
Journal of Physiology, 117(4), 500–544.
Illari, P. M., & Williamson, J. (2012). What is a mechanism? Thinking about
mechanisms across the sciences. European Journal for Philosophy of Science,
2(1), 119–135.
Isnard, C. A., & Zeeman, E. C. (1976/2013). Some models from catastrophe
theory in the social sciences. In L. Collins (Ed.), The use of models in the
social sciences (pp. 44–100). Chicago, IL: Routledge.
Izhikevich, E. (2007). Dynamical systems in neuroscience: The geometry of
excitability and bursting. Cambridge: MIT Press.
Josserand, M., Allassonnière-Tang, M., Pellegrino, F., & Dediu, D. (2021). In-
terindividual variation refuses to go away: A Bayesian computer model of
language change in communicative networks. Frontiers in Psychology: Lan-
guage Science, 12(626118). doi:10.3389/fpsyg.2021.626118
Khambhati, A. N., Mattar, M. G., Wymbs, N. F., Grafton, S. T., & Bassett,
D. S. (2018). Beyond modularity: Fine-scale mechanisms and rules for brain
network reconfiguration. NeuroImage, 166, 385–399.
304 Luis H. Favela
Koch, C. (1999). Biophysics of computation: Information processing in single
neurons. New York, NY: Oxford University Press.
Kothawala, A. (2015). Simulation of nerve action potential using Hod-
gkin Huxley model [MATLAB script]. MATLAB Central File
Exchange. Retrieved July 7, 2021 from https://www.mathworks.com/mat-
labcentral/fileexchange/53659-simulation-of-nerve-action-potential-using-
hodgkin-huxley-model
Lam, L. (1998). Nonlinear physics for beginners: Fractals, chaos, solitons, pat-
tern formation, cellular automata and complex systems. Singapore: World
Scientific.
Liebovitch, L. S., & Toth, T. I. (1990). Using fractals to understand the opening
and closing of ion channels. Annals of Biomedical Engineering, 18, 177–194.
Lopes, R., & Betrouni, N. (2009). Fractal and multifractal analysis: A review.
Medical Image Analysis, 13(4), 634–649.
Mandelbrot, B. B. (1983). The fractal geometry of nature. New York, NY: W.
H. Freeman and Company.
Marom, S. (2010). Neural timescales or lack thereof. Progress in Neurobiology,
90, 16–28.
May, R. M. (1976). Simple mathematical models with very complicated dynam-
ics. Nature, 261, 459–467.
Nagumo, J., Arimoto, S., & Yoshizawa, S. (1962). An active pulse transmission
line simulating nerve axon. Proceedings of the IRE, 50(10), 2061–2070.
Nobukawa, S., Nishimura, H., Yamanishi, T., & Liu, J. Q. (2015). Analysis of
chaotic resonance in Izhikevich neuron model. PloS One, 10(9), e0138919.
doi:10.1371/journal.pone.0138919
Peng, C. K., Buldyrev, S. V., Havlin, S., Simons, M., Stanley, H. E., & Gold-
berger, A. L. (1994). Mosaic organization of DNA nucleotides. Physical Re-
view E, 49(2), 1685–1689.
Poli, D., Pastore, V. P., & Massobrio, P. (2015). Functional connectivity in in
vitro neuronal assemblies. Frontiers in Neural Circuits, 9(57). doi:10.3389/
fncir.2015.00057
Schwiening, C. J. (2012). A brief historical perspective: Hodgkin and Huxley.
The Journal of Physiology, 590(11), 2571–2575.
Serafino, M., Cimini, G., Maritan, A., Rinaldo, A., Suweis, S., Banavar, J. R.,
& Caldarelli, G. (2021). True scale-free networks hidden by finite size effects.
Proceedings of the National Academy of Sciences, 118(2), 1–11. doi:10.1073/
pnas.2013825118
Silva, A. J., Landreth, A., & Bickle, J. (2014). Engineering the next revolution
in neuroscience: The new science of experiment planning. New York, NY:
Oxford University Press.
Stakic, J., Suchanek, J. M., Ziegler, G. P., & Griff, E. R. (2011). The source of
spontaneous activity in the main olfactory bulb of the rat. PLoS One, 6(8),
e23990. doi:10.1371/journal.pone.0023990
Stam, C. J. (2005). Nonlinear dynamical analysis of EEG and MEG: Review of
an emerging field. Clinical Neurophysiology, 116(10), 2266–2301.
14 Hybrid Brains
Interfacing Living
Neurons and Circuits with
Computational Models
Astrid A. Prinz

1 Beyond the Connectome – Dynamics Matter

One of the core tenets of neuroscience is that in order to understand how
the brain processes information, we must understand the ‘connectome’,
meaning the structural layout of the nervous system – who connects to
whom, at scales ranging from individual neurons to entire brain regions
(Sporns 2010). For the unique model system that is Caenorhabditis ele-
gans, a nematode, the complete connectome of its 302 neurons has been
mapped (Cook et al. 2019), yet we are far from fully understanding how
its brain functions. Knowing the connectome of a brain may be a neces-
sary, but not a sufficient prerequisite to understanding nervous system
function. It is equally important to understand the temporal dynamics
of neuronal activity (Koppell et al. 2014).
Over the past decades, technological advances in several areas have
greatly increased our ability to monitor and manipulate neuronal activity.
These areas include multi-electrode neural interfaces, imaging of brain ac-
tivity with genetically encoded or externally applied optical probes sensitive
to neuronal membrane potential or calcium signals, and selective stimula-
tion of neuron populations with optogenetics. Nonetheless, many aspects of
and factors in neuronal signaling remain beyond the reach of experimental
monitoring and manipulation, especially on the functionally relevant time
scale of neuronal communication, which is in the range of milliseconds.
Computational modeling of neurons and neuronal circuits can help fill
this gap and provide an increasingly useful complement to experimental
investigation and manipulation of neuronal activity. In the following sec-
tions, I discuss strengths and potential pitfalls of modeling as a tool to
study neuronal circuit function and describe how ‘wet lab’ experiments
can be combined with computational modeling into hybrid circuits in
­real-time through the dynamic clamp, in a ‘best of both worlds’ scenario.

2 Computational Modeling – Strengths and Pitfalls

Computational models simulate the electrical activity and temporal
dynamics of neurons and neuronal circuits and can be constructed at

DOI: 10.4324/9781003251392-19
306 Astrid A. Prinz
different levels of nervous system organization, from the sub-cellular
level of molecular interactions to the level of entire brain regions com-
municating with each other, and even into the realm of social interac-
tions between brains (Prinz and Hooper 2017).
At all levels, computational modeling complements experimental
studies most beneficially if there is a feedback loop between the two:
models are constructed based on experimental data, models are used to
explore potential mechanisms of brain function and make testable pre-
dictions, those predictions are tested in experiments, the experimental
results inform refinement of the model – repeat the cycle. The benefits
of this iterative process are best captured by the German saying ‘Der
Weg ist das Ziel’ (loosely, ‘The journey is the reward’), indicating that
much can be learned from models that initially fail to match the ground
truth of experimental data, from how they fail, and from how they can
be refined.
For the purposes of discussing the dynamic clamp, I will focus now
on computational modeling at the scale of individual neurons and small
circuits of neurons. At this scale, temporal dynamics of neuronal and
synaptic activity can be represented computationally in the form of sys-
tems of coupled differential equations that are solved numerically on a
computer. Conductance-based models are a particular type of compu-
tational model at the cellular and small circuit scale and include dif-
ferential equations for each neuron’s membrane potential, activation of
each of the relevant ionic membrane conductances (for example, what
fraction of a neuron’s sodium channels are open at any given time), and
other dynamic variables, such as intracellular calcium concentration and
synaptic receptor binding and channel gating.
Conductance-based models can be powerful tools to investigate how
the temporal dynamics of a neuronal circuit’s activity depend on the
circuit connectivity and the intrinsic properties of the neurons and syn-
apses that constitute the circuit. Figure 1 provides an example of such a
small circuit model – the neurons and synapses that generate the motor
commands that allow the sea slug Tritonia diomedea to swim and es-
cape predation through a succession of dorsal and ventral body flexions.
Figure 1c shows the output of an early (1980s) computational model of
the circuit based on cellular and synaptic information available at the
time, which successfully reproduced activation and production of the
swim pattern (Getting 1983, 1989) and could be used to investigate how
circuit activity depends on cellular and synaptic parameters not accessi-
ble to experimental manipulation.
But Figure 1 also illustrates potential pitfalls of computational model-
ing. The original Getting model of the circuit shown in Figure 1c repro-
duced the experimentally observed motor pattern so well that researchers
were lured into thinking that this seemingly simple circuit was now ‘un-
derstood’. However, two decades later, newer information on synaptic
connectivity and neuromodulation in the circuit was incorporated into
 Hybrid Brains 307

Figure 14.1 M
 odels of the Tritonia escape swim circuit. (a) Schematic of the
core swim escape circuit, including neurons DRI – dorsal ramp
interneuron; DSI – dorsal swim interneuron; C2 – cerebral neu-
ron 2; VDI-B – ventral swim interneuron B. Triangles and circles
indicate excitatory and inhibitory synapses, respectively. 5HT in-
dicates circuit-wide serotonin modulation. (b) Escape swim motor
pattern recorded from living circuit. Bursts of action potentials in
the DSI and VDI-B interneurons correspond to dorsal and ventral
body flexions, which alternate to produce swimming. (c) Simulated
motor pattern from original Getting model. (d) The updated model
based on newer cellular and synaptic information fails to reproduce
the motor pattern. Scale bars in (b–d): 10 seconds, 20 mV. Adapted
from Calin-Jageman et al. (2007), with permission.
308 Astrid A. Prinz
the model and resulted in the failure of the updated model to produce
functional motor patterns (Calin-Jageman et al. 2007). This illustrates
the importance of constraining a computational model with experimen-
tal data as much as possible, and the vagaries of trusting results from
models with many cellular and synaptic parameters that are not well
constrained by experimental data.

3 Hybrid Living and Model Circuits – Best of Two Worlds

At the cellular and small circuit scale, the dynamic clamp technique pro-
vides a complementary approach that can overcome some of the poten-
tial pitfalls and uncertainties of computational modeling by combining
models with living neural systems in real-time, in a hybrid system con-
figuration (Robinson and Kawai 1993; Sharp et al. 1993a, 1993b; Prinz
et al. 2004; Goaillard and Marder 2006; Prinz and Cudmore 2011).
The basic operation of the dynamic clamp is illustrated in Figure 2 for
two example applications in which voltage-independent conductances g
are inserted into living neurons. A dedicated computer program hosts the
equation(s) describing the conductances that the experimenter intends to
mimic in a living neuron. The membrane potential (V) of the neuron or
neurons to be dynamically clamped is recorded with an electrode, trans-
formed into a digital signal, and fed into the computer in real-time. The
computer solves the equation(s) to calculate the ionic current (I) that would
be flowing through these modeled (virtual) conductances, and that calcu-
lated current I is injected into the living neuron in real-time. The neuron
thus receives a dynamically changing current, equivalent to the current it
would receive if it actually contained the membrane and/or synaptic con-
ductances that are modeled with the dynamic clamp. As the responses to
test pulses in Figure 2b indicate, the dynamic clamp does indeed change
the effective input conductance of the living neuron – the d ­ ynamic clamp
is therefore sometimes referred to as a conductance clamp.
While the example applications shown in Figure 2 are simple –
­dynamic clamp addition of an artificial inhibitory synapse to a living
neuron (Figure 2b), and subtraction from a living neuron of a leak con-
ductance introduced by electrode impalement to restore the cell’s nat-
ural excitability (Figure 2c) – the dynamic clamp can also be used to
introduce more complex, but physiologically realistic, conditions and
processes into one or several living neurons. For example, the dynamic
clamp can be employed to insert or remove one or more types of ion
channels with voltage-dependent gating dynamics, as shown in Figure
3. The figure summarizes a study in which dynamic clamp was used to
elucidate the cellular mechanisms of spike broadening in the R20 in-
terneuron in the sea hare Aplysia californica (Ma and Koester 1996).
Because this interneuron can initiate and modulate respiratory activity
in the animal depending on spike duration, this example illustrates how
the dynamic clamp can complement conventional electrophysiology and
 Hybrid Brains 309

Figure 14.2 B
 asic dynamic clamp applications: simulation of voltage-­independent
conductances. (a) Schematic of dynamic clamp configuration. The
membrane potential (V) of a living neuron is recorded, digitized,
and fed into a computer in real-time. Equations describing the mem-
brane or synaptic currents (I) to be modeled are solved by the com-
puter for every time step to calculate the current I that would flow
through these virtual conductances (g), given the momentary value
of V and the current’s reversal potential (E). The current I is injected
into the living neuron, also in real-time. (b) Membrane potential V
of a living neuron in response to application of the neurotransmitter
GABA (Gamma-aminobutyric acid, top) and in the same neuron in
response to application of a virtual synaptic conductance via dy-
namic clamp (bottom). In addition to the dynamic clamp current
I, brief current test pulses are injected throughout both traces; re-
duced amplitude of the resulting voltage deflections during GABA
application and during the virtual synaptic conductance clamp in-
dicates that both manipulations temporarily increase the membrane
conductance. Asterisks indicate times of application of GABA and
the beginning of dynamic clamp. From (Sharp et al. 1993b), with
permission. (c) Dynamic clamp subtraction of the leak conductance
introduced by impalement of a living neuron with a sharp electrode.
Leak subtraction virtually “repairs” the impalement damage to the
membrane, restoring the neuron’s intrinsic bursting activity pattern.
From Cymbalyuk et al. (2002), with permission.
310 Astrid A. Prinz
computational modeling to further our understanding of behaviorally
relevant cellular and circuit mechanisms. In the case of dynamic clamp
of voltage-dependent conductances as in Figure 3, the equations describ-
ing the conductances are systems of coupled differential equations for
the gating variables of the corresponding ion channels. In this – more
general and versatile – type of dynamic clamp application, the dynamic
clamp software and computer therefore must solve differential equations
numerically in real-time.
A further example of dynamic clamp versatility is provided in Figure
4, where dynamic clamp is applied to a pyramidal neuron in a slice of rat
somatosensory cortex (Chance et al. 2002). Electrical activity in brain
slices is often altered compared to activity in the intact brain, and is
difficult to control at a millisecond time scale and network-wide level
in slice experiments, making studies of the impact of realistic network
activity on individual neurons technically difficult. The example appli-
cation in Figure 4 shows how dynamic clamp can be used to expose
neurons in a slice to in vivo-like but fully controllable synaptic inputs
and conditions. How being embedded in a large network of other neu-
rons affects a neuron’s activity and information processing can thus be
systematically studied at a physiologically relevant time scale.
While Figure 4 shows how realistic synaptic input from a network
can be simulated with the dynamic clamp, the technique can also be
used to provide a living neuron with one or multiple synaptic partners in
the form of complete model neurons or small circuits simulated by the
dynamic clamp software and computer in real-time (Pinto et al. 2001),
allowing for the construction of truly hybrid networks.
As these example applications show, in contrast to the less physio-
logical manipulations provided by traditional current clamp or voltage
clamp, the dynamic clamp actually mimics in living neurons the presence
of ionic membrane channels, synaptic receptors and their associated dy-
namic conductances, or of an entire surrounding network and the inputs
it provides (Figures 2–4). In this sense, the dynamic clamp is akin to phar-
macological manipulation or overexpression or mutation of cellular and
synaptic conductances, or of embedding of a neuron in a network, but at
the flip of a switch, and with full experimental control over the virtual
conductance(s) by changing parameters in the dynamic clamp equations.
How, then, does the dynamic clamp approach combine ‘the best of both
worlds’? The computational component of dynamic-clamp constructed
hybrid circuits retains the full control that the experimenter has over
a purely computational model, because all parameters of the dynamic
clamp equations can be changed at will, easily, and over wide ranges,
often including parameters that are not – and in many cases will never
be – experimentally manipulatable. On the biological side of the hybrid
circuit, the living neurons in a sense serve as their own model, allowing
the experimenter to probe the effect of various manipulations introduced
 Hybrid Brains 311

Figure 14.3 D
 ynamic clamp addition and subtraction of voltage-dependent
conductances to study spike broadening. (a) Schematic of dynamic
clamp configuration. In contrast to Figure 2a, the equations describ-
ing the added/subtracted voltage-dependent conductances here are
numerically solved differential equations. The impaled neuron is
neuron R20 in the sea hare Aplysia, which can initiate and modu-
late respiratory pumping behavior depending on spike duration. (b)
Overlaid action potentials recorded from a repetitively stimulated
R20 neuron show spike broadening, with narrower spikes for early
stimuli and broader spikes for late stimuli (left). Pharmacologically
blocking two voltage-dependent potassium currents reveals that
they contribute to the initially narrow, then broadening spike shape
(middle). Spike broadening is restored by dynamic clamp addition
of the same K+ currents (right). (c) Subtraction of the same two
currents using dynamic clamp with negative conductance values g
mimics their blocking with pharmacology. Adapted from Ma and
Koester (1996), with permission.

via dynamic clamp in the context of the living system, without the need
for an accurate computational model of the entire neuron or circuit. This
eliminates some potential modeling pitfalls, such as the one described in
Figure 1. In other words, the dynamic clamp allows the researcher to use
computation selectively to control only the modeled aspects of a system,
312 Astrid A. Prinz

Figure 14.4 D
 ynamic clamp simulation of in vivo conditions in a brain slice. (a)
Schematic of dynamic clamp configuration. Simulated excitatory
(ge) and inhibitory (gi) synaptic conductance traces corresponding
to barrages of synaptic inputs from a surrounding network are sup-
plied to the dynamic clamp computer, which calculates the com-
bined excitatory and inhibitory synaptic current and injects it into
a pyramidal neuron in a slice of rat cortex in real-time depending
on the momentary membrane potential V. (b) Spike pattern of the
clamped neuron during simulated in vivo-like input barrage (bot-
tom) shows richer and more realistic temporal dynamics compared
to the same neuron’s spike pattern in response to constant current
injection (top). (c) Spike firing rate of the neuron as a function of
driving current when dynamic clamp is used to inject zero-fold (di-
amonds), one-fold (circles), two-fold (squares), or three-fold (tri-
angles) ge and gi. Simulated network synaptic input of increasing
strength affects the excitability of the neuron, with stronger net-
work input resulting in a smaller gain of action potential firing.
Adapted from Chance et al. (2002), with permission.

without having to trust a complex, full computational model that will

always only be an approximation of the living system.

4 Circuit Mechanisms of a Learned Behavior –

A Dynamic Clamp Study
Figure 5 illustrates how the dynamic clamp can be used to determine
cellular and synaptic mechanisms that underlie learning in a behavioral
conditioning paradigm. It exemplifies how the dynamic clamp can bridge
between different levels of nervous system organization, connecting cel-
lular, synaptic, and small circuit plasticity to whole-animal behavior.
The sea hare Aplysia californica, a marine mollusk, is a prime model
organism for the study of plasticity mechanisms underlying learning. In
particular, Aplysia exhibits associative learning in its feeding behavior,
which consists of cycles of protraction and retraction of the tongue-like
radula rasp to grasp food items (for example, seaweed). In naïve (con-
trol) animals in the absence of food, spontaneous radula cycles occur
with moderate frequency and irregularly. However, after an operant
conditioning regimen in which the animal receives a food reward con-
tingent on and immediately following each radula cycle, bite frequency
is increased and bites (radula cycles) occur at more regular intervals – the
 Hybrid Brains 313

Figure 14.5 D
 ynamic clamp reveals cellular and synaptic plasticity mechanisms
in learning-induced compulsive behavior. (a) Operant conditioning
protocol to induce reward learning in Aplysia feeding behavior. Im-
ages of the sea hare feeding apparatus viewed from the ventral side of
the animal, with the radula (tongue-like rasp) protracted (middle) or
retracted. Tick marks underneath each image show an example time
series of radula protractions during a 40 minute training period in
which a seaweed strip (top of frame) was touched to the animal’s lip
to incite feeding behavior, but not ingested. Control animals received
no food reward (left), contingently trained animals received a food
reward in the form of a calibrated injection of seaweed juice into the
mouth (symbolized by pipette), strictly timed to each radula bite cycle
(middle, reward timing indicated by arrows), and non-contingently
trained animals received an equal average number of food rewards,
but not timed in relation to the bites performed by the animal (right).
(b) One hour after the training, contingently trained animals show
both increased frequency and increased temporal regularity of rad-
ula bite cycles compared to control and non-contingently trained an-
imals, as indicated by the bite cycle interval histogram and fitted line
(middle). (a) and (b), adapted from (Nargeot et al. 2007), copyright
Society for Neuroscience. (c) In buccal neuronal circuits isolated from
contingently trained and non-contingently trained animals, dynamic
clamp was used to add/subtract a leak conductance Gleak to/from
three bite initiation neurons (B30, B63, B65) that control radula bit-
ing. Dynamic clamp was also used to increase/decrease the electrical
coupling strength Gcoupling between B30 and B63, and between
B63 and B65. The time of dynamic clamp operation is indicated by
the shaded area in voltage traces. Subtracting Gleak and increasing
coupling strength increased bite command frequency and regularity
in buccal circuits from non-contingently trained animals (thus mim-
icking reward learning, top traces). Conversely, adding Gleak and
decreasing coupling strength decreased frequency and regularity in
buccal circuits from contingently trained animals, effectively erasing
their behavioral learning (bottom traces). Scale bar, 30 seconds.
314 Astrid A. Prinz
animal has implicitly learned that radula bites can lead to a food reward,
and has adjusted the frequency and regularity (and thereby efficiency)
of its feeding motor program (Figure 5a). Animals that receive food re-
wards not timed to their bite cycles (non-contingent training) do not
increase the frequency and regularity of their radula cycles (Nargeot and
Simmers, 2011).
The small neuronal circuit controlling Aplysia’s buccal area, which
includes muscles that move the radula, can be isolated from the animal
and placed in a dish, where it continues to produce the neuronal bursts
of action potentials that in the intact animal would initiate and gov-
ern feeding behavior, called a ‘fictive biting’ motor pattern. Intriguingly,
buccal circuits isolated from contingently trained animals continue to
produce bite cycle motor patterns with higher frequency and regularity
than buccal circuits isolated from non-contingently trained animals or
control animals (Nargeot et al. 2007). The neuronal circuit in the dish
(in vitro) thus retains a circuit-level memory of what the animal learned
during training in vivo. At the same time, the in vitro circuit allows full
experimental access to the neurons that form the circuit and can be ma-
nipulated with the dynamic clamp.
Important components of the buccal circuit are three bite initiation neu-
rons, B30, B63, and B65, and the electrical synapses between them. In the
study summarized in Figure 5c, dynamic clamp was used to tease apart
how cellular excitability of B30, B63, and B65 and the strength of the
electrical synaptic coupling between them determines the frequency and
regularity of fictive biting (Sieling et al. 2014). The major findings are that

1 increasing neuronal excitability by using dynamic clamp to subtract

a leak conductance Gleak from all three neurons increases fictive
bite frequency,
2 increasing the strength Gcoupling of the electrical synapses between
B30 and B63 and between B63 and B65 with dynamic clamp in-
creases fictive bite pattern regularity,
3 applying both dynamic clamp manipulations simultaneously turns
the low frequency, irregular fictive pattern in a buccal circuit iso-
lated from a non-conditionally trained animal into a high frequency,
regular pattern (Fig. 5c top), and conversely,
4 applying the reverse dynamic clamp manipulations (i.e., decreasing
cellular excitability by adding a leak conductance and decreasing
electrical coupling) turns the high frequency, regular fictive pattern
in a buccal circuit isolated from a conditionally trained animal into
a low frequency, irregular pattern.

Notably, (3) and (4) suggest that learning in the intact animal likely oc-
curs through cellular and synaptic plasticity in the B30, B63, and B65
neurons and can be both mimicked and reversed by appropriately chosen
 Hybrid Brains 315
and simple dynamic clamp manipulations in the isolated circuit. This
approach and results demonstrate the potential explanatory power of
the dynamic clamp technique through bridging between levels of brain
organization and function.

5 Dynamic Clamp Constraints and Opportunities –

Historically, and Going Forward
Like any other technically cutting-edge tool, the dynamic clamp has lim-
itations and constraints that should not go unmentioned. One of these
constraints stems from the fact that in order to be dynamic clamped, neu-
rons must be electrically contacted to record their membrane potential
and inject current. This, by definition, limits the number of neurons that
can be simultaneously controlled with the dynamic clamp. However, as
Figures 4 and 5 illustrate, important questions about the mechanisms of
interaction of neurons with their surrounding circuits and networks can
nonetheless be studied with carefully crafted dynamic clamp experiments.
Another caveat to dynamic clamp experiments is that current injec-
tion usually occurs at the cell body, and that the ions that carry the
dynamic clamp current into the cell are not necessarily of the same type
as the ions that carry the membrane or synaptic current to be mimicked.
For example, while an outward flow of potassium ions might occur in
a living neuron far from the cell body in its dendritic tree, the dynamic
clamp current injection that is supposed to mimic it occurs at the cell
body and might be carried by the injection of chloride ions, dictated by
the composition of the electrode solution. It is therefore important for
experimenters to keep in mind that the spatio-temporal and chemical na-
ture of the dynamic clamp stimuli they provide to living neurons might
not exactly match the nature of the actual in vivo events to be simulated.
At the level of day-to-day neuroscience research practice, it should
further be noted that the installation, programming, and operation of a
dynamic clamp system – precisely because the technique is pushing the
envelope of what is technically feasible – requires some computer savvy.
However, ongoing efforts continue to make the dynamic clamp more
accessible and are lowering the threshold for dynamic clamp experimen-
tation in more and more labs (for example, see http://rtxi.org/).
The update rate of the dynamic clamp – what earns it the label
“­real-time” – depends on the hardware limitations of the dynamic
clamp equipment used, and on the speed with which signals can be
digitized and differential equations solved. Modern dynamic clamp
systems frequently achieve update rates of 25 kHz or higher, meaning
the feedback loop between the living neurons and the equations resid-
ing in the computer is completed multiple times per millisecond (Prinz
and Cudmore 2011, http://rtxi.org/). Such high update rates are espe-
cially important when the conductances to be modeled have very fast
316 Astrid A. Prinz
dynamics, as is the case for the voltage-gated ionic conductances that
underlie action potential generation and operate on sub-­millisecond
time scales. The dynamic clamp configurations used in the context of
such fast dynamics must be able to ‘keep up’ with the dynamics in the
living neurons, lest numerical inaccuracies in the clamped conductances
or – even worse – dynamic instabilities of the combined system arise
(Preyer and Butera 2009).
Historically, the then novel availability of sufficiently fast digitizing
equipment and increasingly powerful computers is likely the reason why
the dynamic clamp arose when it did – in 1993 – and that it arose at
that time simultaneously and independently in two separate research ar-
eas, neuroscience and cardiac physiology (Robinson and Kawai 1993;
Sharp et al. 1993a, 1993b). Since its inception, the dynamic clamp has
increasingly been adopted as a tool by researchers in both fields (Prinz
et al. 2004; Goaillard and Marder 2006; Wilders 2006), as constantly
improving hardware, dynamic clamp software, and computer speeds
continue to expand its capabilities.
More recently, dynamic clamp systems have been generalized beyond
the intracellularly recorded membrane potential as an input to, and the
calculated injection current as an output of the dynamic clamp. For ex-
ample, a recent technological advance uses optogenetics to genetically
encode current-generating opsin molecules in targeted cells and light-­
activate them to produce a dynamic clamp current through the cellu-
lar membrane directly, rather than by injecting the current through an
electrode into the cell (Quach et al. 2018). While in its current version
this novel system still requires an electrode to measure the membrane
potential, it is a big step toward contactless, non-invasive dynamic clamp
that could in future versions also simultaneously cover larger numbers of
neurons (Newman et al. 2015).
In conclusion, the dynamic clamp allows for the construction of hybrid
systems that combine and interface living neurons and circuits with com-
putational elements such as numerical models of membrane and synaptic
conductances, model neurons, and entire circuits and networks of neurons.
It is therefore an important component in the modern neuroscience toolkit.

Acknowledgments
I thank Fred H. Sieling and Romuald Nargeot for providing unpublished
data for Figure 5c.

References
Calin-Jageman RJ, Tunstall MJ, Mensh BD, Katz PS, Frost WN (2007). Param-
eter space analysis suggests multi-site plasticity contributes to motor pattern
initiation in Tritonia. Journal of Neurophysiology 98(4):2382–2398.
Chance FS, Abbott LF, Reyes AD (2002). Gain modulation from background
synaptic input. Neuron 35:773–782.
 Hybrid Brains 317
Cook SJ, Jarrell TA, Brittin CA, Wang Y, Bloniarz AE, Yakovlev MA, Nguyen
KCQ, Tang LTH, Bayer EA, Duerr JS, Buelow HE, Hobert O, Hall DH,
Emmons SW (2019). Whole-animal connectomes of both Caenorhabditis
elegans sexes. Nature 571:63–71.
Cymbalyuk GS, Gaudry Q, Masino MA, Calabrese RL (2002). Bursting in
leech heart interneurons: cell-autonomous and network-based mechanisms.
Journal of Neuroscience 22:10580–10592.
Getting PA (1983). Mechanisms of pattern generation underlying swim-
ming in Tritonia. II. Network reconstruction. Journal of Neurophysiology
49(4):1017–1035.
Getting PA (1989). Reconstruction of small neural networks. In: Koch C, Segev
I, eds. Methods in Neural Modeling. MIT Press, Cambridge, pp. 135–169.
Goaillard JM, Marder E (2006). Dynamic clamp analyses of cardiac, endocrine,
and neural function. Physiology 21:197–207.
Koppell NJ, Gritton HJ, Whittington MA, Kramer MA (2014). Beyond the con-
nectome: The dynome. Neuron 83(6):1319–1328.
Ma M, Koester J (1996). The role of potassium currents in frequency-dependent
spike broadening in Aplysia R20 neurons: a dynamic clamp analysis. Journal
of Neuroscience 16:4089–4101.
Nargeot R, Petrissans C, Simmers J (2007). Behavioral and in vitro correlates
of compulsive-like food seeking induced by operant conditioning in Aplysia.
Journal of Neuroscience 27(30):8059–8070.
Nargeot R, Simmers J (2011). Neural mechanisms of operant conditioning and
learning-induced behavioral plasticity in Aplysia. Cellular and Molecular
Life Sciences 68(5):803–816.
Newman JP, Fong MF, Millard DC, Whitmire CJ, Stanley GB, Potter SM
(2015). eLife 4: e07192.
Pinto RD, Elson RC, Szucs A, Rabinovich MI, Selverston AI, Abarbanel HDI
(2001). Extended dynamic clamp: controlling up to four neurons using a
single desktop computer and interface. Journal of Neuroscience Methods
108:39–48.
Preyer AJ, Butera RJ (2009). Causes of transient instabilities in the dynamic
clamp. IEEE Transactions on Neural Systems and Rehabilitation Engineer-
ing 17(2):190–198.
Prinz AA, Abbott LF, Marder E (2004). The dynamic clamp comes of age.
Trends in Neurosciences 27:218–224.
Prinz AA, Cudmore RH (2011). Dynamic clamp. Scholarpedia 6(5):1470.
Prinz AA, Hooper SL (2017). Computer simulation – power and peril. In:
Hooper SL, Buschges A, eds. The Neurobiology of Motor Control: Funda-
mental Concepts and New Directions. Wiley, Hoboken, NJ, 107–133.
Quach B, Krogh-Madsen T, Entcheva E, Christini DJ (2018). Light-activated
dynamic clamp using iPSC-derived cardiomyocytes. Biophysical Journal
115(11): 2206–2217.
Robinson HP, Kawai N (1993). Injection of digitally synthesized synaptic con-
ductance transients to measure the integrative properties of neurons. Journal
of Neuroscience Methods 49:157–165.
Sharp AA, O’Neil MB, Abbott LF, Marder E (1993a). The dynamic clamp: artifi-
cial conductances in biological neurons. Trends in Neuroscience 16:389–394.
Sharp AA, O’Neil MB, Abbott LF, Marder E (1993b). Dynamic clamp:
­computer-generated conductances in real neurons. Journal of Neurophysiol-
ogy 69:992–995.
318 Astrid A. Prinz
Sieling F, Bedecarrats A, Simmers J, Prinz AA, Nargeot R (2014). Differen-
tial roles of nonsynaptic and synaptic plasticity in operant reward learning-­
induced compulsive behavior. Current Biology 24(9):941–950.
Sporns O (2010). Connectome. Scholarpedia 5(2):5584.
Wilders R (2006). Dynamic clamp: a powerful tool in cardiac electrophysiology.
Journal of Physiology 576(2):349–359.
Section 5

Tool Use and

Development Beyond
Neuroscience
15 Beyond Actual Difference
Making
Causal Selections in Genetics*
Janella Baxter

1 Introduction
Loss of function studies in genetics are a central experimental approach
from which a substantial proportion of gene-centered explanations are
derived. Moreover, the gene-centered explanations that are derived from
the loss of function studies are crucial data for further research pro-
grams such as causal modeling of gene regulatory networks. While pre-
vious philosophical discussions have focused on what properties make
the gene uniquely significant in biological explanations, this paper has to
do with what makes the gene a useful tool for manipulation and control
of biological processes in loss of function studies. For, as we’ll see, the
causal properties that make genes central to loss of function studies are
not unique to genes alone but are likely to be shared by other related
biomolecules. Nevertheless, the centrality of loss of function studies to
contemporary biology and gene-centered explanations requires that our
philosophical analyses properly account for the causal and explanatory
status of genes in this area. A central aim in the history and philosophy
of biology literature has been to reconstruct the underlying justifications
for what makes genes explanatorily and experimentally significant. Sev-
eral, mutually compatible proposals have been made. The history of bi-
ology, it turns out, has had more than one operative gene concept. Thus,
a common view in the philosophy of biology has been that genes serve
a plurality of explanatory and experimental roles in different research
programs. Two proposals for the explanatory and experimental signif-
icance of genes have recently been articulated – the sequence specificity
view and the actual difference making approach (Waters 2007; Weber
2013, 2017).
In this paper, I argue that both the sequence specificity and actual
difference making views are inadequate for making sense of the subtle
reasoning at play in the loss of function studies. The argument against

* I owe special thanks to Alan Love, John Bickle, Antonella Tramacere, and Carl
Craver for their invaluable comments on this paper and to Gabriela Huelga-Morales
for welcoming me into her lab at the University of Minnesota, Twin Cities.

DOI: 10.4324/9781003251392-21
322 Janella Baxter
the sequence specificity view is easy. It simply involves observing that
the gene-centric explanations coming from loss of function studies do
not cite the sequence specificity of a gene as the explanatorily significant
property. Instead, the explanatorily significant property of genes in such
explanations is the pattern of gene expression in a biological system.
When it comes to patterns of gene expression, scientists conducting loss
of function studies distinguish between switch- and dial-like types of
gene expression. Switch- and dial-like gene expression represents two
types of causal control that are achievable thanks to recent technological
and experimental innovations.
Both switch- and dial-like causal control can be interpreted as more re-
fined concepts of Water’s actual difference making view. I embrace this in-
terpretation. However, in embracing this interpretation, I raise reasonable
concerns about why philosophers of science should understand the logic of
loss of function studies through the lens of switch- and dial-like causation
when we already have Waters’ actual difference making view. This is when
I incorporate a philosophy of technology to help justify preferring the
switch- and dial-like account over the actual difference making view.
I argue that the switch- and dial-like causal control represents a “spiral
of self-improvement” that is characteristic of scientific progress (Chang
2004, 44). On this view, technological and experimental progress isn’t
merely an advancement in practical know-how. I argue that this type of
progress can also usher forth conceptual change as novel technological
and experimental advancements help integrate novel concepts into con-
crete, empirical practices. In the case of loss of function studies, I argue
the novel techniques that make both knockout and knockdown exper-
imental techniques from which biologists may choose help switch- and
dial-like causal concepts make contact with the empirical world. Yet,
this recent development is not a replacement for the actual difference
making concept that characterized earlier stages of inquiry in genetics.
Rather, switch- and dial-like causal concepts build upon and refine the
older concept.
This account leaves philosophers of science with both the actual dif-
ference making and the switch- and dial-like accounts as legitimate ways
of understanding the logic of loss of function studies. Nevertheless, I
maintain that if our purpose is to understand the subtle reasoning char-
acteristic of loss of function studies, we should prefer the switch- and
dial-like account. For one thing, this account helps us track scientific
progress in genetics experimentation and theorization. For another, it
helps us understand the rationale behind a scientist’s choice between
switch- or dial-like control in a given experiment. For biologists are of-
ten principled in their choice between which type of control they employ
in an experiment. They will judge that one type of control is more likely
to be illuminating. The switch- and dial-like framework helps us under-
stand their reasoning for doing so.
 Beyond Actual Difference Making 323
The structure of my paper proceeds by (Section 2) laying out the se-
quence specificity and actual difference making accounts that have been
proposed in the philosophy of biology literature. Next (Section 3) I in-
troduce the logic of loss of function studies and I argue that dial- and
switch-like causal control is what justifies the causal selection of genes
in this experimental paradigm. And finally, (Section 4) I argue that for
the purpose of understanding loss of function studies we should prefer
the switch- and dial-like account over the actual difference making view.

2 Gene-Centered Explanations
It is widely acknowledged that contemporary and classical explanations
in the biological sciences often accord a special explanatory status to
genes. This is true despite the fact that many causal conditions are rel-
evant to any given effect. A central program in the philosophy of biol-
ogy literature has been to articulate the underlying rationale of singling
out genes in biological explanations that attempt to capture the sort
of causal control genes have over biological processes. One proposal,
extensively defended by C. Kenneth Waters (1994, 2004, 2006, 2007;
Rheinberger and Müller-Wille 2017), is that genes serve as actual differ-
ence makers in both contemporary and classical experimental areas of
biology. In what follows I distinguish between a narrow and broad sense
of actual difference making that is at play in Water’s work. Another
recent proposal is that genes have sequence specificity with respect to
the linear sequences of molecular products (Weber 2006, 2013, 2017;
Waters 2007; Woodward 2010; Griffiths and Stotz 2013). I lay out these
views in the present section.
The causal selection debate in the philosophy of biology literature
has primarily concerned the underlying rationale behind why biologists
highlight some causal variables in explanation and background others.
In particular, the debate has taken molecular genes and DNA as its cen-
tral case of causal variables that are frequently privileged in contempo-
rary biological explanations. The debate begins with the observation
that any given effect – biological or not – has a large set of relevant
causal conditions. For example, green fluorescence in jellyfish requires
the presence of the gene for the green fluorescent protein encoded in
the jellyfish’s genome, the transcription of the gene into messenger RNA
(mRNA), and the translation of mRNA into a protein. It also requires
the abundance of amino acids and biochemical energy in the form of
ATP, as well as viable temperature and pH levels, and so on. Yet, despite
this immense causal complexity, biologists often single out the gene for
green fluorescent protein when formulating their explanations of why
jellyfish fluoresce in the deep sea. Several proposals have been offered as
analyses of the underlying rationale for gene selection in many biological
explanations – sequence specificity (Weber 2006, 2013, 2017; Waters
324 Janella Baxter
2007, 2018; Woodward 2010; Griffiths and Stotz 2013; Griffiths et al.
2015) and actual difference making (Waters 1994, 2006, 2007).
Sequence specificity has to do with the causal control structural genes
have over the linear sequences of gene products, like RNA and proteins.
Structural genes are sequences of nucleic acid bases – adenine (A), thy-
mine (T)/uracil (U) (for RNA), guanine (G), and cytosine (C) – in DNA
that encode other molecular products. Structural genes are conceptually
distinct from regulatory genes (Gerstein et al. 2007). Regulatory genes are
sequences of nucleic acid bases in DNA that don’t have sequence specific-
ity over the linear sequences of other biomolecules. Instead, they enable
or inhibit the transcription of structural genes when regulatory factors –­
proteins and RNA molecules – bind to regulatory modules. Struc-
tural genes control the nucleic acid sequences of RNA transcripts by
Watson-Crick base pairing rules whereby adenine always specifies uracil,
thymine specifies adenine, guanine specifies cytosine, and cytosine spec-
ifies guanine. In this way, differences in the nucleic acid sequence of a
structural gene produce differences in the nucleic acid sequence of RNA
transcripts. In simple cases where there is no alternative splicing, differ-
ences in the nucleic acid sequences of structural genes also produce differ-
ences in the amino acid sequences of proteins. The nucleic acid sequence
of a structural gene determines the amino acid sequence of a protein in
units of three nucleic acids at a time or in units of codons. For the most
part (with the exception of stop codons – codons that instruct the protein
synthesis machinery to “halt” production), each nucleic acid triplet spec-
ifies one amino acid type. For example, the codon UGG “codes” only for
the amino acid tryptophan. There is some redundancy in the genetic code,
meaning that more than one codon specifies the same amino acid, as in the
case of ACU, ACC, ACA, and ACG all of which specify threonine. This
means that some alternative nucleic acid sequences of a protein coding
gene will make no difference to the same amino acid sequence of a protein.
However, many differences in the nucleic acid sequence of a protein cod-
ing gene will make a difference to the amino acid sequence of a protein.
For some authors, sequence specificity is the primary rationale for the
causal selection of genes in contemporary biological explanations. For
example, the sequence specificity of structural genes is at the heart of
C. Kenneth Waters’ account of what makes genetics successful (1994,
2006, 2007, 2018). In Waters’ view, the gene concept is flexible and can
refer to different entities depending on the purposes of a researcher. For
example, when a biologist wishes to explain the linear sequence of a
protein, they will appeal to the messenger RNA (mRNA) biomolecule
that is read and translated into a sequence of amino acids as the relevant
gene. By contrast, when a biologist wishes to explain the linear sequence
of an mRNA transcript, they will appeal to the relevant nucleic acid
sequence in DNA. In this way, there is no single entity that is a gene for
all research purposes. Yet, even on this flexible understanding of the
 Beyond Actual Difference Making 325
gene concept in contemporary biology, Waters’ analysis of contemporary
practice focuses entirely on structural genes. At times, Waters writes as
if structural genes are the only genes in the game when he writes: “Genes
are for linear sequences in products of gene expression” (Waters 1994,
177). This is further illustrated by Waters’ account of what has made
contemporary genetics successful.
Waters may feel no need to expand his account of the contemporary
gene concept because he has another analysis upon which he can fall.
This is his actual difference making view (Waters 2007). Like the struc-
tural gene concept, the actual difference making concept is also flexible.
It can be understood in broad or narrow terms. Broadly, a cause that
(1) actually varies relative to an otherwise uniform population and (2)
whose varying accounts for an actual difference in an effect variable is
an actual difference maker. The broad interpretation of actual difference
making makes no reference to any particular scientific paradigm – this is
in contrast to the narrow interpretation. All that’s required is that con-
ditions (1) and (2) be met. As we’ll see below although this can charac-
terize the loss of function studies in contemporary molecular biology, I
question the usefulness of this analysis below. The narrow interpretation
of actual difference making is specific to classical genetics of the early
20th century.1 Waters also argues that genes in classical genetics serve
as actual difference makers. On this construal, classical genes are ac-
tual difference makers with respect to actual differences in a phenotypic
trait when they are allowed to actually vary relative to an experimental
population (Waters 1994, 2004, 2007) In classical genetics, genes make
actual differences to genotype – the genetic composition of an organism –
which in turn (can) correlate with actual differences in a phenotypic
trait. Classical geneticists exploited both investigative and theoretical
principles of inheritance – e.g., independent assortment, segregation,
recombination, etc. – to generate lab-raised populations that were (rel-
atively) genetically identical with the exception of a few genes that were
permitted to actually vary. So, in a population of fruit flies, some indi-
viduals are homozygous – they carry two copies of the same gene – for
the wild type eye color (+/+), others are heterozygous for a wild type and
a mutant gene, say purple, (+/pr), and still others are homozygous for
purple (pr/pr). When phenotypic differences in eye color are obtained –
some individuals have red others have purple – classical geneticists in-
ferred that such differences must be due to the genetic differences they
were allowed to actually obtain (Bridges and Morgan 1919).2 Even
though classical geneticists knew that other genes – say, the vermillion
gene – could make a difference to eye color, Waters insists that it is the
wild type or purple genotypes that are explanatorily significant relative
to this experimental set-up. The vermillion gene, by contrast, is merely a
potential difference maker in this case and, thus, is not privileged in the
explanations of classical geneticists.
326 Janella Baxter
Actual difference making – in both the broad and narrow senses –
as well as sequence specificity are different types of causal control that
genes have with respect to biological processes. These accounts have
been proposed by various philosophers (notably, Waters 1994, 2004,
2006, 2007; Weber 2006, 2013, 2017) as descriptions of the reasoning
behind why biologists systematically privilege genes when formulating
their explanations. Waters (1994, 2004, 2006, 2007) and Weber (2004,
2018) both attempt to inform their philosophies by attending to the ex-
perimental practices of biologists. However, as I’ll argue below, their
views on the causal selection of genes have not been attentive to con-
temporary experimental approaches. For the causal selection of genes at
play in the loss of function study fits neither the sequence specificity nor
the actual difference making view very well. Experimental practices get
better over time as new technologies and experimental techniques are
developed. Moreover, explanatory practices often change alongside in-
novation in experimental approaches. I’ll argue below that technological
and experimental innovation can justify novel ways of conceptualizing
phenomena. As researchers are prompted to justify the inferences they
make from experiments, the explanations they formulate often need to
invoke novel concepts ushered forth by technological and experimental
innovations. Thus, it is likely that sequence specificity and actual differ-
ence making may not be up to the task of capturing the causal reasoning
at play in novel areas of biology. At least for the purpose of making
sense of the subtle logics implicit to loss of function studies – one of the
most central experimental paradigms used in contemporary biology –
our philosophical analysis should track the conceptual changes at work
in these sets of practices.

3 Taking Loss of Function Studies (More) Seriously

Loss of function studies are a central experimental approach to many
areas of biology. In what follows, I shall argue that the causal reason-
ing at play in this sort of experiment isn’t adequately captured by the
aforementioned literature. Loss of function studies systematically don’t
involve manipulation of a protein coding gene’s sequence specificity as
a means of manipulating biological processes. Nor do experimental-
ists carrying out these studies employ the theoretical tenets required of
classical geneticists for their explanations. At best, the broad sense of
Waters’ actual difference making approach best approximates the rea-
soning at play in this sort of experiment. Yet, as I’ll contend in the final
section, the broad sense of actual difference making is not a satisfying
analysis. In this section, I argue there are two kinds of causal concepts that
best characterizes the appeal to genes in many biological explanations –
namely, dial- and switch-like control over the concentration of gene
products in a living organism (Woodward 2010).
 Beyond Actual Difference Making 327
Loss of function studies are perhaps one of the most central experi-
mental techniques to contemporary biology. Not only do these experi-
ments aid crucially in the causal modeling of gene regulatory networks,
but – perhaps more modestly – they provide the necessary data for many
of the gene-centered explanations contemporary biologists provide.
Often the results of these studies inform and are informed by further
empirical and theoretical claims, such as causal modeling of gene reg-
ulatory networks. However, biologists needn’t be constructing causal
models when they formulate explanations based on a loss of function
study. At a minimum, the sorts of explanations biologists formulate
from loss of function experiments can characterize in qualitative terms a
causal relationship between a molecular gene (or set of genes) and some
phenotypic difference (Barbaric et al. 2007; Housden et al. 2016). The
name – “loss of function” – clearly indicates that gene function is at
the heart of this sort of experiment. But there can be multiple senses of
“function” used in biological vernacular, some of which are the central
target of loss of function studies.3 Perhaps the most appropriate way to
interpret functional claims in this sort of experimental program is from
the mechanistic tradition (Craver 2007). On this sort of analysis, an
entity’s activities contribute to the capacity of a more complex system
of which the entity is a part. For example, the protein Coronin aids in
regulating filament proteins (called actins) that are crucial for cellular
movement (Shen et al. 2014).
Loss of function studies exploit sequence specificity as a means to ma-
nipulating gene expression – but sequence specificity is not the relevant
explanans in this sort of experimental approach. Biologists have a host
of techniques by which to carry out loss of function studies – ranging
from less common methods like small molecule inhibitors and morpholi-
nos to much more pervasive approaches using RNA interference (RNAi)
and CRISPR-Cas9 (Housden et al. 2016). Often these techniques do rely
on sequence specificity. For example, CRISPR-Cas9 and RNAi have the
unique ability to identify and perturb the expression of coding sequences
embedded in DNA or RNA. Biologists “program” genes encoding
CRISPR-Cas9 or RNAi products so that when they are expressed in an
organism, these mechanisms act only on genes with a complementary
nucleic acid sequence. In this way, the molecular gene’s sequence speci-
ficity with respect to the linear sequences of products is both experimen-
tally relevant to loss of function studies and, indeed, explanatory. The
precise nucleic acid sequence of CRISPR or RNAi components makes
the difference between interference of one gene rather than another.
Differences in the CRISPR or RNAi sequences make a difference and
thereby explain which gene in an organism’s genome is perturbed. But
this is not the explanatory target of loss of function studies. The explan-
atory target of this sort of study is what accounts for some phenotypic
differences in the experimental population.4 The relevant explanans in
328 Janella Baxter
loss of function studies is gene expression. So, when the interruption
of a gene’s expression correlates with a phenotypic difference, it is the
targeted gene’s expression and not the gene sequence that is explanatory.
For example, loss of function studies of the Dpy gene in Caenorhabditis
elegans (or C. elegans) result in dumpy – short and fat – body types (Shen
et al. 2014). It is the perturbation of the Dpy gene’s expression that biolo-
gists single out in explanations of short/fat body types. The intervention
made by the loss of function tools, while reliant on the properties char-
acterized in the previous section, does not produce a change in the linear
sequence of a gene’s products. Instead, when successfully utilized, these
tools introduce changes that produce changes in the expression of a gene.
Loss of function studies proceed by manipulating the concentra-
tion of a gene’s product in an organism; however, this type of study
encompasses two related types of causal control – dial- and switch-
like. Dial-like control is a type of INF (short for influence)-specific-
ity (Woodward 2010). Causes with INF-specificity can take a range
of alternative values, each of which (ideally) determines one and only
one value from a range of alternatives that an effect variable can take.
INF-specificity is often analogized to the tuning dial on a radio. Tun-
ing dials often have many alternative values they can take. Tuning the
dial to one setting correlates with one radio station playing over the
speakers, while tuning to another setting correlates with a different
radio station playing, and so one for each possible setting of the dial.
This is very much like the dial-like control some kinds of loss of func-
tion interventions have over phenotypic traits. An important exception
is that the alternative settings of a radio station dial correlate with
discrete, alternative output values for the effect variable; whereas, di-
al-like control in loss of function studies needn’t. Instead, the alter-
native values a dial-like cause can take can correlate with continuous
values in the effect variable. In this way, a causal variable is analogous
to the modulation of brightness in a room by means of a light dimmer.
A light dimmer can be turned up or down to increase or decrease the
brightness of a room in a continuous way. This sort of control is often
achieved by means of gene knockdown approaches in loss of function
studies (O’Malley et al. 2010). For example, the amount of a fluores-
cent protein in a biological population can be increased/decreased by
administering more or less of the gene knockdown technology, RNAi
(Baggs et al. 2009). RNAi interferes with a gene’s expression by tar-
geting and degrading its mRNA transcripts. 5 The more RNAi reagent
added, ideally, the greater amount of a gene’s product is eliminated.
Sometimes this corresponds to a phenotypic difference. Experimental
investigation of how dial-like control of various genes make differences
to mammalian circadian rhythm cycles found that depleting the Bmal1
gene by about 40% corresponded to a slight increase in luminescence,
 Beyond Actual Difference Making 329
while a 60% decrease corresponded to an even greater increase. This
kind of causal control manipulates the concentration level of a gene
product in a matter of amounts – say, micromolars. Distinguishing be-
tween one causal value from another is, in this sort of case, determined
arbitrarily. The researchers could have just as easily tested a 38.76%
(rather than 40%) depletion in the Bmal1 gene product and discovered
that some change in the degree of luminescence obtains. What matters
is that some change in the causal variable corresponds to some change
in the effect variable. Any changes in the causal variable that fail to
do this don’t count as distinct values that the causal variable can take.
Distinct values correspond to changes that do make a difference to the
value of the effect variable.
A puzzling feature of knockdown experiments is that biologists use
the language of genes in their explanations of the phenotypic differences
that occur and not the language of mRNA. Baggs et al., for example,
write:

The depletion of only three genes…Clock, Bmal1, and Per1…dis-

rupted robust circadian oscillations in bioluminescence.
(Ibid, 0565)

The study from which this claim comes involves RNAi knockdowns of
several mRNA products involved in mammalian circadian rhythm cy-
cles. And yet, these authors speak of “the depletion of…genes…” This is
an illustration of how flexible the molecular gene concept is in contem-
porary biology. Waters (1994) has argued that a common gene concept
underlies the immensely diverse range of biochemicals to which biol-
ogists apply the term. Since nucleic acid sequences in DNA and RNA
often determine the linear sequences of other downstream products, the
word “gene” may properly apply to either type of biomolecule. Thus, the
molecular gene concept can be ambiguous when it is not specified what
product the gene is for and at what stage of gene expression. In this case,
the relevant gene refers to the mRNA transcripts that are the target of
RNAi. The nucleic acid sequences of the mRNA transcripts are genes for
the amino acid sequence of a protein during the stage of gene expression
called translation. Of course, there are genes for the mRNA transcripts
that are the target of the knockdown intervention. These are the Clock,
Bmal1, and Per1 nucleic acid sequences encoded in the organism’s DNA,
which determine the nucleic acid sequences of mRNA products during
the stage of gene expression called transcription. What this shows is that
sequence specificity plays a crucial role in the identification and individ-
uation of genes in biology; however, this should not be mistaken for the
causal significance of a gene. Loss of function studies intervene at var-
ious stages of gene expression – knockdown experiments intervene on
330 Janella Baxter
mRNA, and (as we’ll soon see) knockout experiments intervene directly
on DNA – and it is the various ways gene expression can be manipulated
that is conceptually relevant to the explanations biologists formulate.
Dial-like causal control differs from other types of loss of func-
tion interventions in that it leaves some amount of gene product in
the organism’s cellular environment. Gene knockout studies are an
increasingly pervasive experimental approach to the loss of function
studies. This sort of intervention involves switch-like causal control.
Instead of there being a range of many values a causal variable can take,
each of which associates with a different value in the effect variable,
switch-like causal control is all or nothing. On this sort of experimen-
tal approach, the contrast focus is between the presence or absence
of a gene product in a biological population. This sort of control is
best achieved by gene-­editing techniques like CRISPR-Cas9, which
break and rejoin DNA strands at a precise site where a protein cod-
ing gene is encoded. When the broken strands are rejoined, cellular
repair mechanisms can replace a codon that specifies an amino acid
with a stop codon. Ideally, the insertion of premature stop codons
in all redundant copies of a gene throughout an organism’s genome
will ensure that no protein is synthesized. When a stark phenotypic
difference between a knockout and control population is achieved
by means of a knockout intervention, the target gene often becomes
a standard test case by which to determine whether other knockout
interventions are successful. For example, knockout studies of the
Dpy gene in Caenorhabditis elegans (or C. elegans) result in dumpy –
short and fat – body types and are often used to test whether
­C RISPR-Cas9 reagents target a gene of interest (Shen et al. 2014). In
knockout studies, biologists privilege the absence of a gene whose DNA
sequence would determine the amino acid sequence of a protein in their
explanations of any phenotypic differences that might obtain. In the
case of Dpy knockouts, biologists highlight the absence of the Dpy nu-
cleic acid sequence in DNA to explain the phenotypic difference be-
tween the knockout and control population – in this case, short and fat
versus normal body types respectively.
One might object that either the narrow or broad construal of Waters’
actual difference making approach adequately captures the causal rea-
soning of loss of function experiments. On the narrow construal, loss
of function studies reason about the actual differences molecular genes
produce in ways that are similar to the reasoning strategies employed
by classical geneticists. For example, both strategies appeal to genes to
explain phenotypic differences when the genes are allowed to actually
vary relative to an otherwise genetically and environmentally uniform
population. Although it’s true that important conceptual and experimen-
tal parallels can be found between contemporary and classical genetics
(Vance 1996; Waters 2004, 2009), the causal reasoning strategies are
 Beyond Actual Difference Making 331
not the same. An important difference is the role dominant and recessive
traits played in the reasoning of classical geneticists (Bridges and Morgan
1919). In diploid organisms – or organisms carrying two copies of each
gene – dominant traits are produced by either homo- or heterozygosity
for a dominant allele. That is, either a dominant trait, such as the red eye
color, is caused by either a genotype with two copies of the red eye gene
(+/+) or a genotype with one copy of the red eye gene and one copy of a
mutant gene, like purple (+/pr). By contrast, recessive genotypes require
homozygosity of a mutant gene (pr/pr). For classical geneticists to explain
the difference between some phenotypic traits that emerged in their ex-
perimental populations, say red and purple eyes, they had to appeal to
the actual differences in genotypes with different dominant and reces-
sive alleles. Dominant and recessive traits are not a feature of the causal
reasoning in loss of function studies. What matters is that phenotypic
differences correlate with an experimenter’s intervention on a gene prod-
uct. Contemporary biologists do without the theory of dominant and
recessive traits to explain differences in their populations. An adequate
account of the experimental and explanatory strategies of classical and
contemporary geneticists ought to be sensitive to this subtle difference.
Perhaps a further objection is that Waters’ broad construal of actual
difference making adequately captures the causal reasoning of loss of
function studies. On this proposal, scientists attribute any actual dif-
ferences that might obtain in a (relatively) uniform population to causal
variables that they allow to actually vary. This reasoning appears to
capture the causal reasoning of contemporary biologists conducting loss
of function studies. Indeed, biologists attribute actual phenotypic differ-
ences that obtain in their experimental populations to the actual differ-
ences they induce in molecular genes. Yet I caution against embracing
this characterization. For it is too crude of an analysis to differentiate
between actual difference making practices that biologists recognize as
being illuminating (or, at least, have the promise of being illuminating)
and ones they don’t. As I’ll argue below, contemporary biologists can
make actual many types of differences when studying the functions of a
gene. Yet, biologists are systematic about the sorts of experiments they
conduct and the sorts they don’t. The broad construal of actual differ-
ence making cannot account for this. I’ll argue instead that we charac-
terize the causal reasoning at play in the loss of function studies in terms
of dial- and switch-like causal control.
Technical and experimental innovation can stimulate explanatory in-
novation. Loss of function techniques like RNAi and CRISPR-Cas9 with
their respective types of causal control exerted over genes give scientists a
way to make specific causal concepts operational (Chang 2014). That is,
these tools make the integration of dial- and switch-like causal concepts
with concrete empirical practices possible. While for some authors, tech-
nological innovation is itself “conceptually impoverished” as it is merely
332 Janella Baxter
the application of already developed conceptual frameworks from pure
science (Bunge 1966), technological innovation for loss of function stud-
ies (and perhaps many other experimental paradigms) suggests otherwise.
Rather, knockdown and knockout approaches provide scientists with jus-
tification for conceptualizing phenomena in one way rather than another.
As novel concepts are introduced by means of novel technological and
experimental innovations the logic of the explanations scientist formulate
changes as well. Scientists are often confronted with the need to explain
the reasoning behind the inferences they draw from the experiments they
conduct. In the cases I am discussing, the inferences drawn are often
rather modest – concerning things like what is causally responsible for
some observable effect in a particular population in a particular environ-
ment. Nonetheless, the techniques employed in a particular experiment
shape significantly the conceptual framework implicit to the explanations
biologists formulate. Depending on whether a scientist uses a knockdown
or knockout approach, this will determine whether they will conceptu-
alize the cause and effect variables in terms of a dial- or switch-like way.

4 Selecting the Right Intervention for the Job

So far I’ve argued that the kind of causal control exploited in loss of
function studies is not captured by either the sequence specificity account
or Waters’ narrow sense of actual difference making approach. But what
about the broad sense of actual difference making? By showing that ex-
perimentalists are principled about when they use dial-like and switch-
like causal control over genes, I’ll argue that the broad sense of actual
difference making cannot distinguish between actual difference making
strategies that they regard as illuminating and ones they don’t. I advocate
that we instead embrace dial- and switch-like causal control as a more
nuanced analysis of the causal reasoning biologists employ when singling
out genes in explanations that draw from loss of function experiments.
Richer and more nuanced concepts are what we should expect from tech-
nological and experimental innovation.
Experimentalists working in areas of genetics and molecular biology
have a host of interventions by which to make actual differences in liv-
ing systems as a means to studying how they work; however, a host of
constraints steer them towards particular tools for particular tasks. As
I’ve shown, a primary epistemic aim of many biologists is to produce
observable phenotypic differences in experimental populations by per-
turbing gene expression. For many biologists it is also crucial that they
can attribute phenotypic differences to functions that a gene (and its
products) have as a result of its evolutionary history – not as a result of
the biologist’s intervention. Furthermore, it is often important that any
phenotypic differences that may obtain as a result of an intervention help
illuminate how a gene’s functions contribute to the maintenance of an
 Beyond Actual Difference Making 333
organism’s life. These aims can constrain which experiments a biologist
selects. And finally, an intervention must hold the promise of illuminat-
ing a phenomenon within a time scale that is practical for the researcher.
As the techniques and methods of contemporary biologists con-
tinue to improve, biologists now have many ways to exploit the actual
­difference-making capabilities of genes – not all of which they employ.
One way has to do with a gene’s sequence specificity with respect to the
linear sequences of the products they encode. Modern gene-editing tools
have recently provided researchers with the power to intervene directly
on the nucleic acid sequence of a gene in such a way as to make actual
(almost) any possible gene sequence. This would count as an actual dif-
ference making approach to the study of gene function. In this sort of
experiment, researchers make an actual difference in the nucleic acid
sequence of a gene and associate any phenotypic differences that might
actually obtain in the population to their intervention. Yet, this is not
a common approach that illuminates the kind of phenomena many re-
searchers working in areas of molecular biology are studying. And for
good reason. Whether an alternative nucleic acid sequence of a molecu-
lar gene will produce an observable phenotypic difference is an empirical
matter that biologists are often not able to know prior to the experiment.
At the level of individual proteins, scientists have only a piecemeal un-
derstanding of how changes in the nucleic acid sequence of a protein cod-
ing gene will influence the protein’s structure and function. At a more
global systems level, the phenotypic characteristics of a living system
can remain undisturbed despite point mutations in genes (Kitano 2004).
Furthermore, the number of possible gene sequences an experimentalist
might have to test to determine which actually make a phenotypic differ-
ence can be quite large – for an average size gene of 900 base pairs, the
number of possibilities can be 4900 (since there are four alternative bases
for each position in the sequence). Experimentally testing each of these
can be quite inefficient for a researcher. Perhaps more problematic is the
possibility of creating a gene with novel functions – functions it would
not otherwise have were it not for human intervention. So, even if such
an intervention were to produce a phenotypic difference, a researcher
cannot know whether the phenotypic difference is due to an artificial
function they themselves created or to a function that the gene has ac-
quired over evolutionary history. The broad sense of actual difference
making would count among the illuminating actual difference making
practices, manipulation of a gene’s sequence specificity. Yet, this is only
an illuminating experimental approach for some types of investigative
questions. When it comes to uncovering a gene’s functions, the biological
community generally pursues a very different investigative path.
Achieving observable phenotypic differences that a biologist may at-
tribute to a gene’s evolved functionalities is no trivial matter. The pheno-
typic characteristics of living systems can display little to no differences
334 Janella Baxter
despite changes in the underlying processes that cause them (Kitano
2004; Barbaric et al. 2007). This can be due to a host of reasons. Pro-
tein functions are often undisturbed by point mutations in the structure
(Castagnoli et al. 2004). The presence of complex gene regulatory net-
works can (sometimes) mean that changes in the rate and duration of a
gene’s expression make no difference to the biological processes in which
the gene participates. It is also common for living systems to encode gene
products capable of similar functions, which can compensate for other
gene products in their absence (Mitchell 2009). While some changes in
the underlying processes of a living system can produce no phenotypic
differences, others can outright be unilluminating for other reasons. For
example, intervention on genes that are necessary for maintaining the
life of a biological organism can be lethal. Despite these challenges, biol-
ogists nevertheless conduct loss of function studies instead of other types
of experiments to uncover gene function.
While it is common for researchers to use both dial- and switch-like
interventions to study a gene’s expression, biologists are often principled
about which sort of intervention they conduct. If a researcher has reason
to believe a gene’s function will be compensated for by the functions of
other genes, they may avoid manipulating gene expression in a switch-
like way. This is the reasoning behind the dial-like control used in the
circadian rhythm study by Braggs et al. Instead of eliminating a gene’s
product(s) from a living system entirely, dial-like control allows for some
product to be present. Ideally, this sort of intervention will produce a
significant reduction in the amount of gene product to generate an ob-
servable phenotypic difference in the experimental population, but it is
an empirical matter as to whether the reduction in product will trigger
other compensatory mechanisms that prevent any phenotypic difference
from occurring. Complete elimination of a gene product from a biologi-
cal system that encodes compensatory mechanisms may not produce ob-
servable phenotypic differences. However, dial-like control over a gene’s
expression may. By contrast, switch-like causal control over a gene’s ex-
pression is often preferable to dial-like control when the gene of interest
has subtle phenotypic effects. For example, small amounts of the neu-
rofibromatosis type 2 gene are sufficient to maintain toxin sensitivity in
cancer cell growths (Shalem et al. 2014). Dial-like control on this gene’s
expression rarely shows any phenotypic difference. Switch-like control,
by contrast, shows a significant difference in toxin sensitivity. By turning
this gene’s expression from “on” to “off,” researchers are able to achieve
a clear phenotypic difference that is otherwise masked. Thus, in an ef-
fort to exaggerate a gene’s phenotypic differences, thereby making its
contribution more easily observable, biologists are more likely to prefer
interventions that eliminate a gene product from a living system entirely.
Switch- and dial-like causal concepts are recent refinements of Waters’
broad actual difference making concept. Technological and experimental
 Beyond Actual Difference Making 335
progress often involves the innovation of novel concrete, empirical prac-
tices that enable scientists to probe new and sometimes increasingly sub-
tle questions. This kind of progress doesn’t just represent innovation in
practical know-how (though I by no means with to diminish this kind
of knowledge). Often, with technological and experimental innovation
comes conceptual innovation as new practical innovations help put more
abstract concepts into contact with the empirical world. In the case of
loss of function studies, switch- and dial-like causal control achieved by
means of different experimental techniques have given biologists novel
and more subtle ways of probing the genotype’s relationship to the phe-
notype. With the ability to ask new and increasingly subtle questions
about the genotype’s relationship to the phenotype, techniques such as
RNAi and CRISPR-Cas9 have also given scientists new and increasingly
subtle ways to conceptualize the genotype’s relationship to the pheno-
type. So, while the broad interpretation of the actual difference making
view can absorb switch- and dial-like causal concepts, this doesn’t mean
doing so is illuminating for the purposes of philosophy of science. For
one thing, the broad interpretation of the actual difference making view
cannot account for the judgments scientists make about which sorts of
experimental interventions they deem likely to be illuminating in a par-
ticular context and which they deem likely to be unilluminating. For
another, progress in scientific inquiry often proceeds by means of what
Hasok Chang calls a “spiral of self-improvement” whereby later stages
of scientific inquiry build upon and refine previous ways of investigating
and thinking about phenomena (Chang 2004, 44). What this shows is
that philosophers of science have more than one way to make sense of
the underlying rationale biologists have for singling out genes for ex-
planatory purposes in loss of function experiments. We can think of the
rationale in terms of Waters’ broad sense of actual difference making
or in terms of the dial- and switch-like concepts I’ve articulated here.
In one sense, both ways of understanding loss of function studies are
appropriate. However, we should prefer one way over the other for spe-
cific kinds of purposes. I’ve argued that if our purpose is to understand
the subtle causal reasoning at work in one of the most widely conducted
experiments in contemporary biology, we should prefer the switch- and
dial-like account. Switch- and dial-like accounts help us track practical
and conceptual progress in biology. Moreover, this more refined set of
concepts helps us capture the logic behind the choices biologists make
when conducting knockdown or knockout experiments.

5 Conclusion
The gene is a central explanatory variable in much of classical and con-
temporary biology. I have argued that the rationale behind the wide-
spread causal selection of molecular genes varies depending on the type
336 Janella Baxter
of tools scientists use to manipulate and control genetic processes. This
is significant not just because it represents progress in practical know-
how. It’s significant because novel technologies and techniques can also
play a role in conceptual change as they make novel concepts opera-
tional. The story I am telling about the loss of function studies is a story
about how gene concepts have become increasingly sophisticated as bi-
ologist’s experimental tools have increased in number and power. This
means that the switch- and dial-like account defended in this paper is a
recent refinement of the broad interpretation of Waters’ actual difference
making view. Yet this doesn’t mean that the actual difference making
view is adequate for our purposes. If our purpose is to understand the
subtle reasoning at play in the loss of function studies – and I believe
philosophers of genetics should have this purpose – we should prefer the
switch- and dial-like account. This account better accommodates the
nature of technological and conceptual progress in the history of genet-
ics and it helps us appreciate the judgments biologists make about which
type of causal control is likely to be more illuminating.

Notes
1 By classical genetics, Waters really has in mind the system of practices and
theories of the American geneticist, Thomas Hunt Morgan and his lab. If by
“classical genetics,” one simply means the system of practices and theories
aimed at understanding genetics in the early 20th century, then Waters’ ac-
count only captures a subset of the science going on at the time. Different
laboratories around the world studying genetics operated by distinct styles
of thought at the time (see Harwood 1993).
2 Difference making alone cannot explain which genotype aligns with which
phenotypic difference. This required theoretical knowledge about dominant
and recessive traits as well as extensive empirical knowledge about which act
as dominant/recessive. Waters acknowledges this, however, it is noteworthy
his actual difference making approach cannot fully account for this aspect
of the reasoning at play in the experiments of classical genetics.
3 See Millikan (1989), Neander (1991a, 1991b), Wouters (2003), Walsh and
Ariew (2013) for a brief survey of the numerous function concepts in the life
sciences.
4 Some molecular tools do indeed produce phenotypic differences, however.
For example, genetic markers such as fluorescent proteins are an illustrative
example. Yet this sort of tool differs importantly from the loss of func-
tion tools presently discussed. Genetic markers are often used as an as-
say technique to help researchers detect the presence of a gene product;
whereas, loss of function techniques are methods for interrupting causal
relationships.
5 Dial-like control over the concentration of a gene product can be achieved
with a variety of methods including small molecules. Even gene-editing tech-
niques such as CRISPR-Cas9 can achieve this sort of control by targeting
only some copies of the same gene in a genome or by targeting all copies of a
gene at a particular stage of development. RNAi is by far the most popular
method.
 Beyond Actual Difference Making 337
References
Baggs J., T. S. Price, L. DiTacchio, S. Panda, G. FitzGerald, J. Hogenesch. 2009.
“Network Features of the Mammalian Circadian Clock.” PLoS Biology,
7(3), pp. 0563–75.
Barbaric I., G. Miller, T. N. Dear. 2007. “Appearances Can be Deceiving: Phe-
notypes of Knockout Mice.” Briefings in Functional Genomics and Proteom-
ics, 6(2), pp. 91–103.
Bridges C. B., Morgan T. H. 1919. The Second-Chromosome Group of Mutant
Characters. Carnegie Institution of Washington Publication 278. Washing-
ton, DC: Carnegie Institution, pp. 123–304.
Bunge M. 1966. “Technology as Applied Science.” Technology and Culture,
7(3), pp. 329–47.
Castagnoli L., A. Costantini, C. Dall’Armi, S. Gonfloni, L. Montecchi-Palazzi,
S. Panni, S. Paoluzi, E. Santonico, G. Cesareni. 2004. “Selective Promiscu-
ity in the Interaction Network Mediated by Protein Recognition Modules.”
FEBS Letters, 567, pp. 74–9.
Chang H. 2004. Inventing Temperature: Measurement and Scientific Progress.
New York: Oxford University Press.
Chang H. 2014. Is Water H 2O? Evidence, Realism and Pluralism. Boston Stud-
ies in the Philosophy of Science, Volume 293. New York: Springer Science &
Business Media.
Craver C.F. 2007. Explaining the Brain: Mechanisms and the Mosaic Unity of
Neuroscience. New York: Oxford University Press.
Gerstein M. B., C. Bruce, J. S. Rozowsky, D. Zheng, J. Du, J. O. Korbel, O.
Emanuelsson, Z. D. Zhang, S. Weissman, M. Snyder. 2007. What Is a Gene,
Post-ENCODE? History and Updated Definition. Cold Spring Harbor, NY:
Cold Spring Harbor Laboratory Press.
Griffiths P., K. Stotz. 2013. Genetics and Philosophy: An Introduction. New
York: Cambridge University Press.
Griffiths P., A. Pocheville, B. Calcott, K. Stotz, H. Kim, R. Knight. 2015. “Mea-
suring Causal Specificity.” Philosophy of Science, 82(4), 529–55.
Harwood J. 1993. Styles of Scientific Thought: The German Genetics Commu-
nity, 1900–1933. Chicago, IL: The University of Chicago Press.
Housden, B., M. Muhar, M. Gemberling, C. Gersbach, D. Stainier, G. Seydoux, S.
Mohr, J. Zuber, N. Perrimon. 2016. “Loss-of-Function Genetic Tools for Ani-
mal Models: Cross-Species and Cross-Platform Differences.” Nature, 18, 24–40.
Kitano H. 2004. “Biological Robustness.” Nature Reviews, 5, pp. 826–37.
Mitchell S. 2009. Unsimple Truths: Science, Complexity, and Policy. Chicago,
IL: University of Chicago Press.
Millikan R. G. 1989. “In Defense of Proper Functions.” Philosophy of Science,
56(2), pp. 288–302.
Neander K. 1991a. “Functions as Selected Effects: The Conceptual Analyst’s
Defense.” Philosophy of Science, 58(2), pp. 168–84.
Neander K. 1991b. “The Teleological Notion of Function.” Australasian Jour-
nal of Philosophy, 69(4), pp. 454–68.
O’Malley M. A., K. C. Elliot, R. M. Burian. 2010. “From Genetic to Genomic
Regulation: Iterativity in microRNA Research.” Studies in History and Phi-
losophy of Biological and Biomedical Sciences, 41, pp. 407–17.
338 Janella Baxter
Rheinberger H.-J., S. Müller-Wille. 2017. The Gene: From Genetics to Postge-
nomics. Chicago, IL: University of Chicago Press.
Shalem O., N. E. Sanjana, E. Hartenian, X. Shi, D. Scott, T. Mikkelsen, D.
Heckl, B. L. Ebert, D. E. Root, J. G. Doench, F. Zhang. 2014. “Genome-Scale
CRISPR-Cas9 Knockout Screening in Human Cells.” Science, 343, pp. 84–87.
Shen Z., X. Zhang, Y. Chai, Z. Zhu, P. Yi, G. Feng, W. Li, G. Ou. 2014. “Con-
ditional Knockouts Generated by Engineered CRISPR-Cas9 Endonuclease
Reveal the Roles of Coronin in C. elegans Neural Development.” Develop-
mental Cell, 30, pp. 615–36.
Vance R. 1996. “Heroic Antireductionism and Genetics: A Tale of One Sci-
ence.” Philosophy of Science, 63, pp. 36–45.
Walsh D. M., A. Ariew. 2013. “A Taxonomy of Functions.” Canadian Journal
of Philosophy, 26(4), pp. 493–514.
Waters C. K. 1994. “Genes Made Molecular.” Philosophy of Science, 61, pp.
163–185.
Waters C. K. 2004. “What was Classical Genetics?” Studies in History and
Philosophy of Science, 35, pp. 783–809.
Waters C. K. 2006. “A Pluralist Interpretation of Gene-Centered Biology.” In
Scientific Pluralism, Minnesota Studies in the Philosophy of Science, Volume
XIX, edited by S. H. Kellert, H. E. Longino, C. K. Waters. Minneapolis: Uni-
versity of Minnesota Press, pp. 190–214.
Waters C. K. 2007. “Causes That Make a Difference.” Journal of Philosophy,
104(11), pp. 551–579.
Waters C. K. 2009. “Beyond Theoretical Reduction and Layer-Cake Antireduc-
tion: How DNA Retooled Genetics and Transformed Biological Practice.” In
The Oxford Handbook of Biology, edited by M. Ruse. New York: Oxford
University Press, pp. 238–62.
Weber M. 2004. Philosophy of Experimental Biology. Cambridge: Cambridge
University Press.
Weber M. 2006. “The Central Dogma as a Thesis of Causal Specificity.” His-
tory and Philosophy of the Life Sciences, 28, pp. 595–610.
Weber M. 2013. “Causal Selection vs Causal Parity in Biology: Relevant Coun-
terfactuals and Biologically Normal Interventions.” In Causation in Biology
and Philosophy, edited by C. K. Waters, M. Travisano, J. Woodward. Minne-
apolis: University of Minnesota Press.
Weber M. 2017. “Discussion Note: Which Kind of Causal Specificity Matters
Biologically?” Philosophy of Science, 84(3), pp. 574–85.
Weber M. 2018. “Experiment in Biology.” The Stanford Encyclopedia of Phi-
losophy. https://plato.stanford.edu/entries/biology-experiment/.
Whittaker S., J.-P. Theurillat, E. Van Allen, N. Wagle, J. Hsaio, G. S. Cowley,
D. Schadendorf, D. Root, L. Garraway. 2013. “A Genome-Scale RNA Inter-
ference Screen Implicates NF1 Loss in Resistance to RAF Inhibition.” Cancer
Discovery, 3, pp. 351–62.
Woodward J. 2010. “Causation in Biology: Stability, Specificity, and the Choice
of Levels of Explanation.” Biology and Philosophy 25, pp. 287–318.
Wouters A. 2003. “Four Notions of Biological Function.” Studies in History
and Philosophy of Biological and Biomedical Sciences, 34, pp. 633–68.
Contributors

Nina A. Atanasova, Ph.D. is Lecturer of Philosophy at the Department

of Philosophy and Religious Studies at The University of Toledo. She
began her career in 2014 after receiving her Ph.D. in Philosophy (Phi-
losophy and the Life Sciences Track) from the University of Cincinnati
the same year. While a graduate student, she rotated in the Vorhees/
Williams Neurology Lab at Cincinnati Children’s Research Foundation.
She joined the International Association for Science and Cultural Diver-
sity as an Assistant Secretary General in 2015 and has been its Secretary
General since 2017. Her research centers on epistemology and method-
ology of neuroscience.
Ann-Sophie Barwich is Assistant Professor at Indiana University,
Bloomington, Department of History and Philosophy of Science and
Medicine and Cognitive Science Program. Her research specializes in
olfaction as a model for theories of cognition and the brain. She is the
author of Smellosophy: What the Nose Tells the Mind (Harvard Univer-
sity Press, 2020).
Janella Baxter is Lecturer in the Department of Philosophy at Wash-
ington University, St. Louis. Her area of research is philosophy of bi-
ology and technology. She is one of the co-founders of the Biological
Engineering Collaboratory, an interdisciplinary network for historians,
social scientists, and philosophers working at the intersection of biology
and engineering.
John Bickle is Professor of Philosophy and Shackouls Honors Col-
lege Faculty at Mississippi State University and Affiliate Faculty in the
Department of Neurobiology and Anatomical Sciences at the Univer-
sity of Mississippi Medical Center. His area of research is philosophy
of neuroscience, especially neuroscience-in-practice. He is the author of
four academic books, 100 articles and chapters in philosophy and neu-
roscience journals and volumes, and he edited The Oxford Handbook of
Philosophy and Neuroscience.
Daniel Burnston is Associate Professor of Philosophy at Tulane Univer-
sity, director of the Tulane Cognitive Studies Program, affiliate faculty
member in the Tulane Brain Institute, and co-editor of the Brains Blog.
He works on topics in philosophy of cognitive science and philosophy of
340 Contributors
science, focusing primarily on the organization of the brain and mind.
He has published in top philosophy of science journals, including Phi-
losophy of Science, BJPS, and Biology and Philosophy, as well as in
interdisciplinary venues in cognitive science and neuroscience.
David Colaço is Humboldt Postdoctoral Fellow at the Munich Center
for Mathematical Philosophy of LMU Munich. He is a historian and
philosopher of science who addresses reasoning and experimental prac-
tices in scientific research, with a focus on cognitive science and neuro-
science. He has published a number of papers, including Philosophy of
Science, Biology and Philosophy and Analysis.
Carl F. Craver is Professor in the Philosophy Department and the Phi-
losophy-Neuroscience-Psychology Program at Washington University.
He specializes in the philosophy of science, with particular expertise in
neuroscience and cognitive science. He is the author of Explaining the
Brain: Mechanisms and the Mosaic Unity of Neuroscience (Clarendon;
2007) and the co-author with Lindley Darden of In Search of Mecha-
nisms: Discoveries across the Life Sciences (Chicago; 2013).
Luis H. Favela is Associate Professor of Philosophy and Cognitive Sci-
ences at the University of Central Florida. He has been a Fellow at the
University of Pittsburgh’s Center for Philosophy of Science and Duke Uni-
versity’s Summer Seminars in Neuroscience and Philosophy, as well as
Visiting Researcher at Indiana University Bloomington’s Computational
Cognitive Neuroscience Laboratory in the Department of Psychological
and Brain Sciences. His interdisciplinary research at the intersection of
philosophy and the mind sciences applies complexity science, dynamical
systems theory, and ecological psychology to investigations of behavior,
cognition, and consciousness in diverse systems at various spatial and
temporal scales.
Stuart Firestein is Professor of Neuroscience in the Department of
Biological Sciences at Columbia University. His research focuses on the
vertebrate olfactory system, perhaps the best chemical detector on the
planet. Dedicated to promoting the accessibility of science to a public
audience Firestein serves as an advisor for the Alfred P. Sloan Founda-
tion’s program for the Public Understanding of Science. He is an AAAS
Fellow, a Guggenheim Fellow and has written two books on science for
a public audience Ignorance, How It Drives Science and Failure, Why
Science Is so Successful.
Valerie Gray Hardcastle is the St. Elizabeth Healthcare Executive
Director of the Institute for Health Innovation and the Vice President
for Health Innovation at Northern Kentucky University. An interna-
tionally recognized scholar, Hardcastle is the author of five books and
over 210 essays. She studies the nature and structure of interdisciplinary
theories in cognitive science and has focused primarily on developing a
philosophical framework for understanding conscious phenomena re-
sponsive to neuroscientific and psychological data. Most recently, she
 Contributors 341
is investigating how the nature of addiction can shed light on what it
means to be human.
Gregory Johnson is Instructor of Philosophy at Mississippi State
University. His area of research is philosophy of psychology and neu-
roscience. He has published in Philosophical Psychology, Minds and
Machines, and other journals. He is the author of Argument and Infer-
ence: An Introduction to Inductive Logic (MIT Press, 2016).
Marco J. Nathan is Associate Professor and Chair in the Department
of Philosophy at the University of Denver. His research focuses on the
philosophy of science, with particular emphasis on biology, neurosci-
ence, cognitive neuropsychology, and economics. He is the author of
Black Boxes: How Science Turns Ignorance Into Knowledge (OUP,
2021) as well as numerous articles in both philosophical and scientific
venues.
David Parker is University Senior Lecturer in the Department of Phys-
iology, Development and Neuroscience, University of Cambridge (UK).
His research focuses on neural circuits and neuromodulation, specifi-
cally on electrophysiological analyses of neuronal interactions and how
they contribute to the generation of spinal cord locomotor rhythms. He
has also examined how changes in spinal cord circuits could contribute
to functional recovery after spinal cord injury, research that has included
collaborative computational analyses. His work has been published in
the Journal of Neuroscience, the Journal of Neurophysiology, Neuro-
science, and Frontiers in Neural Circuits.
Astrid Prinz is Associate Professor in the Department of Biology at
Emory University. Coming from a background in Physics, she uses com-
putational modeling to study rhythmic pattern generation and homeo-
static regulation in small neural circuits, currently in the mouse and fruit
fly. She is the Past President of the International Organization for Com-
putational Neurosciences.
Alcino J. Silva is Director of The UCLA Integrative Center for Learn-
ing and Memory and is Distinguished Professor in the Departments of
Neurobiology, Psychiatry & Biobehavioral Sciences and Psychology at
UCLA. He pioneered the field of Molecular and Cellular Cognition, and
in 2002 founded and became the first President of the Molecular and
Cellular Cognition Society, an organization with more than 8,000 mem-
bers and chapters in North America, Asia, and Europe. His laboratory is
searching for the mechanisms that underlie the allocation, encoding, and
storage of information in the brain. Additionally, insights into mecha-
nisms of memory are being used by his lab to unravel the causes and de-
velop treatments for cognitive deficits associated with aging, intellectual
disabilities, and autism, as well as to enhance recovery after brain injury.
A key effort in his lab involves the development of computational tools
to integrate and summarize large amounts of causal information with a
searchable, graphical format (researchmaps.org).
342 Contributors
C. Matthew Stewart is Associate Professor of Otolaryngology, Head
and Neck Surgery at Johns Hopkins University, School of Medicine. His
research focuses on sensorimotor integration of the vestibular and visual
systems and his clinical practice expertise is in surgical restoration of
hearing as well as surgical management of tumors of the ear and skull
base. Most recently, he has been examining the human host cell mecha-
nisms of infection by the SARS-CoV-2 virus to explain the long-term re-
covery syndromes affecting the peripheral and central nervous systems.
He received an MD/PhD from the University of Texas Medical Branch
Galveston and is a Fellow of the American College of Surgeons.
Antonella Tramacere carries out research in the fields of philosophy of
comparative and evolutionary psychology, neuroscience, and cognition.
She has published in important neuroscientific journals, such as Trends
in Cognitive Science, Biological Reviews and Neuroscience, as well as in
philosophy journals, such as Topoi, Synthese and Biology and Philoso-
phy. She is engaged in international collaborations and projects with the
Max Planck for the Sciences of Human History in Leipzig (Germany),
with the Center for Mind and Cognition of the University of Bochum
(Germany), with the Laboratory of Symbolic Communication of Riken
in Tokyo (Japan), and with the Department of Philosophy and Religion
at Mississippi State University (USA). Currently at the Department of
Philosophy and Communication Study of the University of Bologna, she
is leading a theoretical and experimental project concerning emotions
and mentalizing in the visual perception of oneself.
Charles V. Vorhees, Ph.D. is Professor of Pediatrics at the University of
Cincinnati College of Medicine and Division of Neurology, Cincinnati
Children’s Hospital Medical Center. He is co-director of the Animal
Behavior Core and program director of the Teratology Training Pro-
gram. He is graduate faculty in the Neuroscience and Molecular and
Developmental Biology (MDB) graduate programs. He served as MDB
Program Director for six years and as an officer for another nine. He has
been a member of the Neuroscience Program Admission Committee for
ten years. He was Editor-in-Chief of Neurotoxicology and Teratology
for nine years. He is a founding member of the Developmental Neu-
rotoxicology Society and was president in 1984–1985 and 2012–2013.
He was an Eli Lilly Distinguished Lecturer in 1990, and a Society for
Neuroscience Grass Foundation Lecturer in 2002. He has served on
scientific advisory committees for the Food and Drug Administration,
the Environmental Protection Agency, the National Institutes of Health,
CalEPA, and the National Academy of Sciences.
Michael T. Williams is Associate Professor at the University of Cincin-
nati Department of Pediatrics and Cincinnati Children’s Research Foun-
dation Division of Neurology. Currently, he is co-director of the Animal
Behavioral Core at Cincinnati Children’s along with Charles Vorhees.
His research interests are broad within the realm of developmental
 Contributors 343
neurotoxicology and include the effects of manganese, proton particle
therapy, PCBs, pyrethroids, drugs of abuse, hydrocephalus, stress, as
well as the genetic models of developmental disorders. He has over 155
published manuscripts. He is a member of Society for Neuroscience,
International Behavioral Neuroscience Society (IBNS), Developmental
Neurotoxicology Society, and Society of Toxicology. He was made a fel-
low of IBNS in 2009 and won the Patricia Rodier Mid-career Award for
Research and Mentoring in 2020 (delayed until 2021). He is a current
member of the Neurotoxicology and Teratology editorial board.
Lu Xu is Postdoctoral Research Scientist in the Firestein lab at the
Department of Biological Sciences, Columbia University. She is currently
studying olfactory coding using a combination of physiological, molecu-
lar, and imaging techniques.
Index

action potential 14–16, 19–20, 24, calcium imaging 50–51, 92, 100, 115,
47, 50, 140, 146, 159, 162, 165, 148, 150
167–168, 172, 244–245, 288–291, calibration 61, 64–65, 67–70, 73, 76,
295–299, 303–304, 307, 311–312, 80, 313
314, 316 causal explanations 140, 143, 174,
adjustment 70, 98, 101, 109, 187–189 211
agent 52, 65, 71–72, 107–108, 141, causal-dynamical 212
249 causal-explanatory 202
algorithm 99–100, 109, 115, 240, causality 17, 41, 43–44, 47–48,
252, 265, 271 70–71, 73, 79, 81, 87, 95, 97, 102,
allosteric 103–104, 106, 145 108, 110–111, 119–120, 122,
Aplysia 242, 308, 311–314, 317 126–127, 129, 133, 141, 148,
artifacts 65, 94, 99, 152, 159, 172, 152, 154–161, 164–167, 169–170,
174, 178, 186, 193 172–174, 180–182, 185, 193, 199,
astrocytes 147, 151, 257 201–202, 204, 206–212, 215–217,
auditory system 88, 122–123 231, 241, 243–245, 247–248, 251,
256, 260–261, 263, 274, 289, 317,
Bacon, Francis 2, 49, 169, 173, 175 322, 325, 328, 331–332, 334, 338,
Barnes Maze 142, 147 341
Bechtel, William 50, 52–53, 189–190, causal-mechanistic 7, 44, 110, 129,
199, 215, 222, 236, 246, 254, 195–196, 198–201, 203, 205–214
264–265, 280, 294–296, 298, channel 14, 20, 22–25, 27, 36, 43,
301–302 46, 50–51, 63–64, 69, 153, 155,
Bell’s number 242 158–159, 162–163, 165, 171, 173,
Biel Water Maze 61, 63, 69, 80 175, 216, 226, 245, 255, 288, 304,
biomarker 201–205, 207, 210, 213, 306, 308, 310
216 Churchland(s, the) 29–31, 34, 37,
bivalence 163–164 39, 49–51, 53, 84, 114, 118, 131,
black box 7, 79, 109, 129, 196–197, 190–191, 237, 288, 302
206, 209–210, 212–213, 216, Cincinnati water maze 5, 56–57,
243, 341 61–62, 64, 80, 82
brain: and COVID-19 122–123, circuit 7–8, 14, 19, 21, 23, 33, 44,
125–129; “long hauler” syndrome 46, 48, 51–54, 73, 90, 143–147,
125, 128–129; and megakaryocytes 149, 151, 169, 175, 219, 222–223,
127–128; olfactory bulb 126; 228–229, 231, 239, 241–244, 247,
temporal lobe 127–128; viral 249–252, 254–258, 275–276, 281,
impacts 125–126 304–308, 310–316, 341
BRAIN initiative 2, 30–31, 222–223, clamp 8, 15, 21–28, 30, 35–36, 175,
231, 234–237, 256 288–289, 296, 299, 305–306,
Brownian motion 49 308–318
346 Index
closed-loop 144, 149 153, 155–160, 163, 167, 169–170,
coding 5, 84–87, 89, 91, 101, 103– 173–175, 181, 189, 192, 243,
112, 114, 116, 202, 324, 326–327, 248–250, 252, 269, 281, 310–314,
330, 333, 343 317, 321–324, 326, 328, 330–332,
cognition 4, 6, 29, 63, 71, 85, 107– 334–336
109, 111, 114, 116, 129, 147, 149, convergence 81, 137, 140–142, 147,
177, 185, 188, 222, 228–229, 231, 167, 180, 182, 186–188
238–239, 243, 252, 255–257, 259, covering-law model 195, 197, 208
276, 279, 281–282, 303, 339–342 COVID-19: autopsies 123–124,
cognitive 2–3, 7, 18, 20, 28, 31, 34, 127, 129; and the brain 122–123,
38, 44, 53, 61, 64–65, 71–75, 125–128, and the ear 122–123,
77–79, 81–85, 102, 107–109, 111, 125, 129; history of 119–120;
113–114, 116, 126, 132, 143, “long hauler” syndrome 122, 125,
149, 170, 174, 176–177, 179–180, 128–129; loss of smell 120, 122,
182–184, 188–190, 193, 197, 125; and megakaryocytes 127–129;
203–205, 207, 209, 213, 215–216, neurological impacts 125–126,
222, 228–230, 235, 237–240, 246, 128–129; risk factors for 120
254, 257–271, 276–282, 296, 303, Cre recombinase 42, 46, 150
339–342 cybernetic 252, 257
cognitive function 84, 107, 111, 229,
260, 270 data 1–2, 7, 29–31, 50, 52, 56, 61, 67,
cognitive impairments 64–65, 183 69–72, 76–80, 83, 90, 97–101, 104,
cognitive ontologies 259–260, 106, 108–111, 115, 122, 167, 172,
270, 281 177–180, 186–187, 189–190, 210,
cognitive ontology 7, 259–260, 264, 215, 217, 221–235, 238, 242–243,
268, 270–271, 280–281 249, 252–253, 259–260, 263–266,
cognitive science 53, 85, 107, 113, 270, 275, 277–278, 280, 283,
116, 132, 254, 281, 303, 287–290, 296–297, 299–301, 306,
339–340, 342 308, 316, 321, 327, 340
community 58, 60, 76, 81, 188, 200, data analysis 7, 83, 98–99, 101, 172,
253, 297, 299, 333, 337 265, 280, 287–288, 297, 299
computation 8, 24, 28–29, 34, 44, 52, data integration 222–224,
89–90, 100, 105, 131, 140–141, 227–229, 233
159, 174, 240, 252–257, 279, databases 75, 227, 234, 260, 264–
281–282, 297, 302, 304–306, 308, 267, 270, 272, 276–278
310–312, 316, 340–341 decomposition 155, 215, 246, 254,
computational modeling 8, 29, 140, 261, 263, 268, 273, 276, 278, 280,
305–306, 308, 310–312, 341 282, 288, 294, 298–299, 302
concentration 27–28, 87, 93–94, 106, determinism 157, 168
144, 154, 204, 306, 326, 328, 336 diaschisis 247, 251, 254
conductance 14, 20, 22, 24, 27, 35, differential equations 306,
51, 306, 308–317 310–311, 315
conductance-based models 306 discordance 6, 177, 180, 182–183,
conductance clamp 308–309 185–186, 188, 193, 224
connectomics 240, 249, 251–252, discovery 2, 6, 8, 16–19, 24–25, 31,
256–257, 305, 318 44, 52, 69, 78, 81, 85–86, 89, 105,
consistency 141–142, 301 112–113, 116, 119, 124, 127, 129,
constraints 7, 111, 155–156, 162, 132, 137–138, 140, 143, 149,
170, 172, 189, 221–222, 225–232, 156, 168, 174, 183, 190, 198,
234–236, 282, 315, 332 215, 217, 225, 237, 243, 246,
control 1, 8, 26–27, 29, 43, 45–47, 259–260, 264–265, 268, 270,
52–53, 64–65, 82, 96, 98, 120, 276, 279–281, 287, 291,
129, 131, 145, 147–149, 151, 298–299, 302, 329, 338, 340
 Index 347
distributed cognition 85, 107–109, gene expression 14, 145, 150, 243,
111, 114 322, 325, 327–330, 332, 334
dominance 157, 166, 168 gene knockdown 328
dopamine 45, 55, 72–73, 79–81, 145, gene knockout 150, 249, 330
157–159, 167 gene targeting 5, 13, 26, 30, 34,
dynamic clamp 8, 175, 305–306, 39–41, 43, 45, 50–52, 142, 147,
308–318 150, 228
dynamics 8, 30, 33, 53, 108–109, 116, genetics 29, 45, 48, 75, 80, 85, 99,
148, 151, 274, 282, 287–288, 291, 102, 120, 127, 137–140, 142,
293–294, 296–299, 301–306, 308, 145–150, 155, 158, 170, 173,
312, 316 182, 200–202, 210–211, 214–217,
242–243, 249–251, 256, 321–322,
ear 120–125, 127, 129, 131, 134, 342 324–326, 330–332, 336–338, 343
Einstein, Albert 49, 248 genome-wide association studies 200
electrical synapses 314 gigaseal 23–25, 28, 30
electrode 16–17, 21, 24, 31, 78, 138, glia 146–147, 150, 158, 244–245, 254
158, 183, 308–309, 315–316 glutamate 46–48, 52, 54, 73, 145
electroencephalography (EEG) 6, 170, Golgi (stain) 137, 287
183–184, 191, 193, 229, 304 GPCRs 86, 113, 116
electrophysiology 2, 15, 25–28, 78,
138–140, 142, 146, 153, 158, 160, Hacking, Ian 13, 34–35, 38, 48–50,
162–163, 165, 173, 225, 228–229, 53, 59, 80–81, 117–118, 130,
249–250, 308, 318, 341 132–133, 172, 174, 190, 192
engineering 6, 16, 24, 36, 38, 82, heterarchical systems 246, 250
84, 93, 112, 117, 140, 144–146, heuristics 2, 246, 258, 260, 272–273,
148–150, 162, 169–171, 175, 238, 275–280
242–243, 249, 304, 317, hierarchical 15, 190, 246, 268, 280
338–339 hippocampal learning 138
enzymes 21–22, 42, 92, 120, 145, 164, hippocampus 40–43, 52–54, 56,
243 72–74, 78, 81–82, 128, 138, 142,
ephaptic 244, 251 148–151, 157, 230, 254
epidemiology 180–182, 185–186, 192, Hodgkin-Huxley-model 8, 20,
199, 215, 217 24–25, 173, 288–291, 295–299,
epithelial tissue 120–122, 124, 303–304
129–130 homeostasis 95, 251, 255, 341
error 16, 61, 63–65, 70, 74, 139, Hubel, David 14–20, 30–33, 35, 118,
152, 165, 171, 175, 179, 186–188, 132
243, 274 hybrid brains 305
error minimization 186, 188 hybrid circuit 310
etiological 180, 202, 211 hybrid networks 310
hybrid system 308, 316
feedback 244, 246–247, 274–275, hypothesis 6, 14, 32, 39–45, 48–49,
290–291, 301, 306, 315 51–52, 56, 60–61, 64, 69, 71,
fluorescence 92, 144, 232 73–79, 93, 111, 115, 118, 121–123,
fMRI 3–4, 6, 183–184, 191, 203–205, 125–126, 129–130, 140, 153,
209, 227, 229, 238, 248, 261, 158–159, 166, 168, 176–178, 181,
265–267, 279 183, 185–187, 190, 193–194, 238,
fractals 291, 293, 297–299, 301–304 242, 252, 275, 295

GABA 46–48, 55, 145, 247, 309 identity experiments 142

GCaMP 97, 249 imaging 2–3, 5–6, 50–51, 83, 90, 92,
gene-centered explanations 321, 96, 99–100, 113, 115, 131, 137,
323, 327 142–144, 146, 148, 150–151, 169,
348 Index
192, 203, 215, 225, 228, 231–234, machine learning 260, 270, 280, 283
237, 246, 249–250, 287, 305, 343 maker’s knowledge 152–153, 157,
inflammation 120–121, 160, 167, 169–170
124, 127 mastoid 121–124, 131
inhibition 43–44, 46, 51, 70, 93–94, mathematical 8, 105, 108–110, 140,
97, 101–103, 105, 107, 114–115, 164, 195, 199, 201, 212, 243,
140, 156, 158, 163–164, 173, 242, 287–289, 291, 296–299, 301, 304,
247, 255–256, 274, 307–308, 312, 340
324, 327, 338 maze 5, 36, 40–42, 47, 56–57, 61–64,
injection 42, 46, 52, 95, 154, 158, 66–70, 72–75, 77–80, 82, 142, 223,
165, 167, 308–309, 312–313, 262, 282
315–317 mechanism 5, 7–9, 14, 20–21, 25, 34,
innovation 2, 26, 56–57, 60, 75–76, 36, 38–40, 44, 48, 50–53, 56, 59,
84, 119, 125, 129, 144, 146–147, 79–82, 84–86, 89–90, 102–106,
200, 239, 322, 326, 331–332, 335, 108, 110–113, 126, 130, 134,
340 138–139, 144–146, 148, 152–153,
integration 6–9, 36, 46, 56–57, 60, 155–156, 158–160, 162, 165,
72, 79–82, 109–110, 138–139, 168–174, 177, 183–184, 197–199,
148–149, 177, 179–181, 186–190, 202, 206–207, 209–210, 212–213,
192–193, 219, 221–238, 242–243, 215–217, 229–230, 237, 239–240,
252, 256, 258, 274, 276, 281–282, 243, 246, 248, 250, 252, 255–257,
285, 302, 317, 322, 331, 341–342 261, 263–264, 272, 275, 280–281,
interpretative device 5, 56–57, 287, 294–296, 298–299, 302–303,
77, 80 306, 308, 310, 312–313, 315, 317,
intervention 6, 27, 35–36, 40–41, 327, 330, 334, 337, 340–342
44–45, 51, 53, 56, 72–73, 79, 81, membrane 20–27, 35–36, 41, 43, 149,
118, 132, 141–143, 152–169, 153, 162–163, 167, 174, 245, 303,
171–174, 177, 180–182, 184–185, 305–306, 308–310, 312, 315–316
191–192, 202–203, 221–222, 225, membrane conductance 306, 309
239–240, 244, 252, 258, 261, 282, membrane current 24, 35, 174, 303
328–335, 338 membrane potential 21, 167, 305–
intervention experiments 141–143 306, 308–309, 312, 315–316
invariance 61, 192, 291, 293–297, mice 1, 5, 40–42, 46–47, 52, 54, 83,
299–301, 303 93, 99, 108, 138–139, 142–144,
iterative 98, 109, 111, 291, 293, 298, 146, 148, 150, 163, 170, 229, 234,
306, 337 237, 337
microelectrode 5, 13, 15–19, 28,
Kuhn, Thomas 26, 35, 37, 58, 60, 30–31, 33, 35, 47, 118–119; metal
76–77, 81, 112, 117, 132–133, 248, 5, 13, 15–19, 28, 30, 33, 118
256–257 microscope 1–2, 5–7, 21, 34, 38,
53, 80, 83, 87, 90–91, 97, 107,
lesion 17, 25, 43, 45, 56, 72–73, 75, 113–114, 116–117, 124–125, 128,
78, 82, 138, 153, 155, 163, 241, 130, 143–144, 147–150, 225, 227,
247, 249 231–232, 237, 249–250, 287
Liebig, Justus 48–49 miniscopes 6, 143–144
linearity 89, 96, 268, 290, 295, 297, modulation 27, 46, 48, 52, 94, 96,
301, 323–325, 327–329, 333 101–108, 116, 145–146, 152–153,
Long Term Potentiation (LTP) 39–43, 229, 307–308, 311, 316, 328
45, 49, 52–54, 59, 71, 81, 126, molecular and cellular cognition 222,
128–129, 132, 138, 150–151, 226, 228–229, 238, 243, 341
230, 254, 256, 258, 342 Molecular Cloning Laboratory
loss of function 8, 321–323, 325–332, Manual 137
334–337 Morris Water Maze 36, 40–42, 66–67,
luminescence 328–329 80, 82, 142, 223, 262, 282
 Index 349
Muffin 91 place cells 78, 142, 148, 150
multifunctionality 259, 263–264, population behavior 87, 111
267–269, 272, 276 postpositivism 57, 59, 77
prediction 6–7, 13, 21–22, 25, 29,
nasal 86, 92, 96, 120–122, 125–126 40, 43, 56–57, 69, 76, 88, 104,
network 7, 27, 30, 44, 60, 100, 157, 129, 137, 139, 143, 152, 180, 185,
183–184, 191, 197–198, 202, 208, 195–214, 217, 223, 232, 239–241,
211, 224, 229, 236, 238, 241–242, 244, 251, 265, 267–269, 271–272,
247, 252, 254–255, 257–259, 274, 283, 306
266–267, 269, 273–276, 278–280, productivity 7, 222, 225–228, 230,
282, 293–295, 297, 302–304, 310, 233, 235–236
312, 315–317, 321, 327, 334, 337, protein 14, 40–43, 108, 120, 122,
339 144–145, 147, 149, 153, 159,
nonlinearity 290–292, 296–297, 301, 171, 238, 248, 323–324, 326–330,
304 333–334, 337
noradrenaline 145 protocol 2, 8, 36, 82, 92–93, 95,
123–124, 129, 137, 172, 223, 226,
oligodendrocytes 147 235, 313
ontology 7, 56, 58, 60, 110, 219, 254, psychophysical 90, 102, 190
257, 259–260, 263–266, 268–271,
276–277, 280–282 randomization 181–182, 192
opsins 140, 146, 148, 163, 167, 172, realism 36, 57–58, 80–81, 172, 262,
316 264, 266, 268, 337
optogenetics 4–6, 9, 13, 26, 34, 36, real-time 8, 91, 116, 305, 308–310,
39, 43–47, 50–53, 80–82, 130, 137, 312, 315
139–140, 144–145, 147–153, 155– receptive field 15–16, 18–19, 35, 118
160, 162–166, 168–172, 174–175, reductionism 8, 34, 36, 53, 80, 82,
228, 234, 249–251, 254–256, 281, 113, 139, 147, 190, 241–248, 250,
287, 305, 316 252–254, 256, 258, 280, 287, 302
relativism 57–58
paradigm 37, 60, 77, 105, 117, 133, relevance 6, 154, 157, 161–162, 184,
195, 198, 227, 248, 265, 269, 193, 253
276–277, 282, 312, 323, 325 repertoire 57, 59–60, 76–77, 80–81,
patch clamp 15, 21–28, 30, 35–36 87, 93
pathways 14, 27–28, 34–35, 45, 53, respiration 119–122, 308, 311
80, 113, 149, 151, 201, 207, 217, reverse inference 7, 203–204, 207,
229, 245, 302 209, 215–216, 259, 263, 268
perfumery 83, 89–90, 93–94, 105 reversibility 157, 163, 175
periphery 5, 84–86, 89–90, 101–110, revolution 2, 5, 15, 26–27, 29–31,
112, 116, 122, 342 33–37, 43, 53, 81–82, 84, 105–106,
philosophy of biology 35, 50, 53, 131, 112–113, 117, 130, 132, 137–140,
175, 321, 323, 339 143, 146–150, 195, 203, 231, 236,
philosophy of neuroscience 5, 8–9, 35, 238, 248–249, 254, 256–257, 280,
37, 53, 60–61, 80, 118, 131, 217, 302, 304
339 robustness 176–178, 180–181, 185–
philosophy of science 3–4, 6, 9, 13, 194, 207–212, 253, 275, 329, 337
34–36, 53, 56–60, 77–78, 80–82, ruthless reductionism 8, 113, 139,
113, 117, 130, 173–176, 190–193, 147, 243
195–197, 199–200, 212, 215–217,
236–238, 255–256, 280–282, SARS-CoV-2: ACE2 receptor protein
301–303, 335, 337–341 120–122, 124; and the brain 122–
philosophy of science-in-practice 123, 125–126, 128–129; detection
57–60 of 124; origination 119–120; and
phosphorylation 27, 35, 145, 149 strokes 126–129
350 Index
scaffolds 84, 107, 109 surgery 1, 32, 68, 123, 127, 129,
scale 7, 76, 89, 99, 202, 221, 223– 131–133, 155–157, 246, 251, 342
225, 231, 240, 248, 252–253, 257, synaptic 8, 14, 21, 39–42, 45, 51–55,
265, 278, 289, 291, 293–301, 303, 138, 140, 142, 144–145, 148–149,
305–308, 310, 313, 316, 333, 340 229, 238, 241, 244–245, 247,
scale-invariant 8, 288, 291, 293–299 249, 252–253, 255–256, 306–310,
selectivity 157–159, 191, 204 312–318
sensors 144–145, 150
serotonin 145, 307 theory: as tool 118–119
signaling 2, 14, 25, 27–28, 83, 145, tinkering 5–6, 13–18, 20–21, 24–26,
149, 171, 174, 305 28, 30–33, 51, 84, 109, 112,
single-cell 5, 15–16, 20, 43, 83, 87, 117–119, 127, 129–130
90, 110–111, 151, 227, 249–250 tools first method 38–39, 43, 45, 50
single neuron 15–19, 31, 35, 97, transcranial magnetic stimulation 163,
166, 288, 290–292, 295–296, 300, 183, 239
303–304 transcriptomics 249
single unit 17, 35, 143 transgenic 138–139, 142, 146, 149
spatial learning 40–42, 52, 54, 65, 72, tributary intervention 155–156, 165
78, 82, 142, 150 triggering 43, 156, 334
spatial memory 36, 40–43, 45, 49, 55,
67, 79, 82, 138, 223, 262, 282 values 21, 159–162, 164–166, 173,
Sperry, Roger 245, 257 177, 181–182, 185, 299, 301, 311,
spike 47, 122, 162, 168, 173–174, 328–330
288, 291, 308, variability 141, 182, 188, 247, 251
311–312, 317 viral vectors 86, 139, 145–146, 148,
striatum 18–19, 35, 42, 45, 54, 72–75, 228
77, 79–80, 118, 267 virtual 148, 308–310
structure-function 1, 14, 45, 73, 83, voltage-dependent 308, 310–311
86, 235, 257, 268, 333
suppression 47–48, 90, 97, 102, 145, Wimsatt, William 176–178, 193–194,
184, 191–192 246, 258