PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

This page intentionally left blank

PHYLOGENETIC
ANALYSIS
OF DNA SEQUENCES

Edited by
Michael M. Miyamoto
University of Florida
Gainesville, Florida

Joel Cracraft
University of Illinois
Chicago, Illinois

New York Oxford

OXFORD UNIVERSITY PRESS
1991
Oxford University Press
Oxford New York Toronto
Delhi Bombay Calcutta Madras Karachi
Petaling Jaya Singapore Hong Kong Tokyo
Nairobi Dar es Salaam Cape Town
Melbourne Auckland

and associated companies in

Berlin Ibadan

Copyright © 1991 by Oxford University Press, Inc.

Published by Oxford University Press, Inc ,
200 Madison Avenue, New York, New York 10016

Oxford is a registered trademark of Oxford University Press

All rights reserved. No part of this publication may be reproduced,

stored in a retrieval system, or transmitted, in any form or by any means,
electronic, mechanical, photocopying, recording, or otherwise,
without the prior permission of Oxford University Press

Library of Congress Cataloging-m-Publication Data

Phylogenetic analysis of DNA sequences / edited by
Michael M. Miyamoto, Joel Cracraft
p. cm. Includes bibliographical references and index
ISBN 0-19-506698-7
1. DNA—Evolution. 2. Nucleotide sequence. 3. Molecular
evolution I Miyamoto, Michael M II Cracraft, Joel.
OH371.P47 1992 574.87'3282—dc20 90-28771

98765432
Printed in the United States of America
on acid-free paper
Contents

1. Phylogenetic Inference, DNA Sequence Analysis, and the

Future of Molecular Systematics 3
Michael M. Miyamoto and Joel Cracraft

2. DNA Sequencing: Strategy and Methods to Directly

Sequence Large DNA Molecules 18
Jerry L. Slightom, David R. Siemieniak, and
Leang C. Sieu

3. The Application of Automated DNA Sequence Analysis to

Phylogenetic Studies 45
Robert J. Ferl, C. Joseph Nairn, Jin-Xiong She,
Edward Wakeland, and Ernesto Almira

4. Computer Alignment of Sequences 59

Michael S. Waterman, Jana Joyce, and Mark Eggert

5. Aligning DNA Sequences: Homology and Phylogenetic

Weighting 73
David P. Mindell

6. Relative Efficiencies of Different Tree-Making Methods for

Molecular Data 90
Masatoshi Nei

7. Compositional Statistics Evaluated by Computer

Simulations 129
Arend Sidow and Allan C. Wilson

8. Weighted Parsimony: Does it Work? 147

Walter M. Fitch and Jia Ye

9. Testing the Theory of Descent 155

David Penny, Michael D. Hendy, and Michael A. Steel
vi CONTENTS

10. Parsimony and Phylogenetic Inference using DNA Sequences:

Some Methodological Strategies 184
Joel Cracraft and Kathleen Helm-Bychowski

11. Evolutionary Analysis of Length-Variable Sequences:

Divergent Domains of Ribosomal RNA 221
Allan Larson

12. Statistical Methods for Testing Molecular Phylogenies 249

Wen-Hsiung Li and Manolo Gouy

13. Discriminating Between Phylogenetic Signal and Random

Noise in DNA Sequences 278
David M. Hillis

14. When are Phylogeny Estimates from Molecular and

Morphological Data Incongruent? 295
David L. Swofford

15. Congruence Among Data Sets: A Bayesian Approach 334

Ward C. Wheeler

Index 347
Preface

The comparative analysis of DNA sequences is becoming increasingly im-

portant in systematic and evolutionary biology and will continue to do so
as faster and more efficient methods for collecting these data are developed.
Large amounts of comparative sequence data will be required to answer
most molecular systematic questions, but this labor-intensive effort will
only be the first of several problems faced by the systematist. Although
the use of DNA sequences in systematics is still in its infancy, already a
healthy mixture of opinion exists about the most appropriate methods for
reconstructing phylogenetic history from nucleotide data. Moreover, some
question whether DNA sequences will prove to be more informative in all
cases when compared to more traditional data-bases. Thus, in using DNA
sequences comparatively, the biologist is confronted by staggering com-
plexities that are often not appreciated even by the expert systematist or
molecular evolutionist.
This volume has assembled an internationally recognized group of in-
vestigators representing different theoretical viewpoints and disciplines to
address critically a diversity of questions about DNA systematics. The book
begins with an introduction by Miyamoto and Cracraft, followed by 14
additional chapters emphasizing data acquisition, sequence analysis, and
the broader systematic importance of nucleotide information. Contributors
on data acquisition have focused on improved techniques for obtaining
comparative sequence information by manual (Slightom et al.) and auto-
mated (Ferl et al.) approaches. With regard to data analysis, authors have
concentrated on methodological problems dealing with sequence alignment
(Waterman et al. and Mindell) and different tree-building algorithms (Nei,
Sidow and Wilson, Fitch and Ye, and Penny et al.). Finally, contributors
have focused on more general issues having broad implications within
systematics. Specifically, their chapters have concentrated on the evalua-
tion of phylogenetic reliability and information content of different se-
quences and data sets (Cracraft and Helm-Bychowski, Li and Gouy, and
Hillis), on the relationship between molecular evolutionary bias and phy-
logeny reconstruction (Larson), and on the application of consensus and
congruence approaches in systematics (Swofford and Wheeler).
This book has its roots in the symposium "Recent Advances in Phylo-
genetic Studies of DNA Sequences," which was part of the special cen-
viii PREFACE

tennial celebration of the American Society of Zoologists, held in con-

junction with the Society of Systematic Zoology, on December 26-30,
1989 in Boston, Massachusetts. Researchers from different disciplines and
approaches presented papers at the symposium, but rather than just de-
scribe their methods or dwell on particular groups, each participant con-
centrated on the strengths, limitations, and assumptions of their approaches
relative to others. The diversity of topics and viewpoints represented at
the symposium constituted its greatest strength, and in turn, has now be-
come the most important quality of this book.
The following people and organizations are profoundly acknowledged
for their assistance. The American Society of Zoologists and Society of
Systematic Zoology contributed administrative and financial assistance for
the symposium. J.S. Farris, J. Felsenstein, A.G. Kluge, T.D. Kocher, J.A.
Lake, and A. Meyer presented papers at the symposium, but chose not to
contribute chapters. External reviews of the individual chapters were pro-
vided by M.W. Allard, J.M. Carpenter, J. Felsenstein, D.H.A. Fitch,
W.M. Fitch, D.M. Hillis, R. Holmquist, R.L. Honeycutt, B.F. Koop,
W.-H. Li, D.R. Maddison, D.P. Mindell, C.J. Nairn, J.L. Patton, D.R.
Siemieniak, J.L. Slightom, D.L. Swofford, B.S. Weir, W.C. Wheeler, R.
Wilson, C.-I. Wu, and E.A. Zimmer. W.F. Curtis of Oxford University
Press is recognized for his encouragement and for his help in seeing this
volume through the initial stages of production. The book's index was
compiled by M.R. Tennant, and K. Lee, and G. Kiltie, and A. McClaughry
helped with all aspects of the secretarial work. Both editors were supported
by National Science Foundation awards during the organization of the
symposium and completion of the book. Financial assistance was also pro-
vided by our respective Departments and Universities. All of these indi-
viduals and institutions are gratefully thanked for their help and support.
Finally, Michele R. Tennant and Terry Root are especially thanked for
their continuous love and support.

M.M.M.
J.C.
Contributors

Ernesto Almira Kathleen Helm-Bychowski

Interdisciplinary Center for Department of Anatomy and Cell
Biotechnology Research Biology
University of Florida University of Illinois
Gainesville, Florida 32611 Chicago, Illinois 60680

Joel Cracraft Michael D. Hendy

Departments of Anatomy and Cell Department of Mathematics and
Biology and Biological Sciences Statistics
University of Illinois Massey University
Chicago, Illinois 60680 Palmerston North, New Zealand

David M. Hillis
Mark Eggert Department of Zoology
Departments of Mathematics and The University of Texas
Molecular Biology Austin, Texas 78712
University of Southern California
Los Angeles, California 90089 Jana Joyce
Departments of Mathematics and
Robert J. Ferl Molecular Biology
Interdisciplinary Center for University of Southern California
Biotechnology Research and Los Angeles, California 90089
Department of Vegetable Crops
University of Florida Allan Larson
Gainesville, Florida 32611 Department of Biology
Washington University
St. Louis, Missouri 63130
Walter M. Fitch
Department of Ecology and Wen-Hsiung Li
Evolutionary Biology Center for Demographic and
University of California Population Genetics
Irvine, California 92717 University of Texas
Houston, Texas 77225
Manolo Gouy
Laboratoire de Biometrie David P. Mindell
Universite Lyon I Department of Biological Sciences
69622 Villeurbanne University of Cincinnati
Cedex, France Cincinnati, Ohio 45221
X CONTRIBUTORS

Michael M. Miyamoto Jerry L. Slightom

Department of Zoology Molecular Biology—Unit 7242
University of Florida The Upjohn Company
Gainesville, Florida 32611 Kalamazoo, Michigan 49007

C. Joseph Nairn Michael A. Steel

Departments of Vegetable Crops and Fakultat fur Mathematik
Botany Universitiit Bielefeld
University of Florida Postfach B640
Gainesville, Florida 32611 D-4800 Bielefeld 1, Germany

Masatoshi Nei David L. Swofford

Department of Biology and Institute of Illinois Natural History
Molecular Evolutionary Genetics Survey, Urbana
Pennsylvania State University 607 East Peabody Drive
University Park, Pennsylvania 16802 Champaign, Illinois 61820

Edward Wakeland
David Penny Division of Basic Sciences
Department of Botany and Zoology Department of Pathology and
Massey University Laboratory Medicine
Palmerston North, New Zealand University of Florida
Gainesville, Florida 32611
Jin-Xiong She
Division of Basic Sciences Michael S. Waterman
Department of Pathology and Departments of Mathematics and
Laboratory Medicine Molecular Biology
University of Florida University of Southern California
Gainesville, Florida 32611 Los Angeles, California 90089

Arend Sidow Ward C. Wheeler

Department of Molecular and Cellular Department of Invertebrates
Biology American Museum of Natural History
University of California Central Park West at 79th St.
Berkeley, California 94720 New York, New York 10024

Allan C. Wilson
David R. Siemieniak Department of Molecular and Cellular
Molecular Biology—Unit 7242 Biology
The Upjohn Company University of California
Kalamazoo, Michigan 49007 Berkeley, California 94720

Leang C. Sieu Jia Ye

Molecular Biology—Unit 7242 Department of Biology
The Upjohn Company University of Southern California
Kalamazoo, Michigan 49007 Los Angeles, California 90089
PHYLOGENETIC ANALYSIS OF DNA SEQUENCES
This page intentionally left blank
 1
Phylogenetic Inference, DNA Sequence
Analysis, and the Future of Molecular
Systematics

MICHAEL M. MIYAMOTO AND JOEL CRACRAFT

The Emergence of Phylogenetic Systematics

Comparative biology has had a long and noble history (Rieppel, 1988). Its
central goal has been to understand the bewildering diversity of form ob-
served across the world's organisms. Concepts such as taxa (in particular,
species) and homology have been the core intellectual instruments for
facilitating this understanding. Thus, comparative biologists have labored
to sort the world's organisms into species (however construed), largely
based on their characteristics of form (but viewed broadly, including any
intrinsic attribute), and to describe their similarities and differences. For
several hundred years now, this effort to compare has led to the realization
that similarities and differences among species are best ordered in terms
of a hierarchy of relationships, which we now take to represent the primary
pattern of life's history as it has diversified over the last 4 billion years.
Systematics is the science of comparative biology and the primary goal
of systematists is to describe taxic diversity and to reconstruct the hierarchy,
or phylogenetic relationships, of those taxa (Hennig, 1966; Eldredge and
Cracraft, 1980; Nelson and Platnick, 1981; Wiley, 1981). The overall im-
portance of this goal is that corroborated hypotheses of relationship are
the prerequisites for all inferences regarding what needs to be explained
(i.e., various patterns) within the context of an historical perspective. This
might involve repeated patterns of geographic distribution from one group
to another (Nelson and Platnick, 1981), the distribution across lineages of
certain behavioral, ecological, or morphological attributes, or the sharing
among taxa, or perhaps among spatially segregated populations, of vari-
ation at the level of DNA. Just as importantly, it is only by having an
hypothesis for the hierarchical pattern of shared attributes—with its im-

3
4 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

plications for understanding character polarity—that one can begin to

formulate explanations for derived novelties that have arisen and become
fixed in populations.
Procedures for constructing phylogenetic hypotheses have been most
fully developed by the discipline of phylogenetic systematics, or cladistics
(Hennig, 1966), which presently dominates the field of systematics (Hull,
1989). The emergence of cladistics can be related to its direct connection
to genealogy in that only shared derived characters (synapomorphies) are
used as evidence to support hypotheses about phylogenetic relationships.
With respect to those hypotheses, similarities due to the retention of pri-
mitive features (symplesiomorphies) are ignored because they are unin-
formative regarding relationships. The cladistic approach is thus concep-
tually coherent in that only synapomorphies are considered to provide
evidence for monophyly, and only monophyletic groups have objective
reality as historical entities.
The increasing acceptance of cladistics can also be attributed to its place-
ment of systematic methodology within a broader scientific context. Thus,
one of the more important philosophical advances has been to require that
phylogenetic hypotheses conform to observations as closely as possible
(Farris, 1983). Stated differently, those hypotheses which make the fewest
ad hoc assumptions about shared patterns of character-state distributions
are to be preferred (Eldredge and Cracraft, 1980; Wiley, 1981), yet the
hypotheses themselves will always remain vulnerable to testing by addi-
tional observations. This philosophical and methodological approach, known
as parsimony, has become an essential component of modern systematic
methodology (Farris, 1983; Sober, 1983, 1988), and as we shall discuss
below, also provides the rationale for another important component of
phylogenetic inference: congruence analysis.

The Growth of Comparative Sequence Analysis

Prior to the 1960s, most systematic studies utilized morphological char-
acters as evidence for relationships. However, over the last 20 to 30 years,
the contributions of molecular approaches to phylogenetic research have
increased steadily (Hillis, 1987; Patterson, 1987; Doolittle, 1990; Hillis and
Moritz, 1990). At this time, DNA sequences are becoming the preferred
database for molecular systematics. Nuclear genomes, coupled with their
extranuclear counterparts, are exceedingly large and complex, thereby of-
fering the systematist an almost endless array of characters with different
structural/functional properties, mutational/selectional biases, and evolu-
tionary rates (Li et al., 1985; Nei, 1987; Li and Graur, 1991). Therefore,
the potential exists for using the evolutionary characteristics of particular
genomic regions to match those sequences to specific systematic questions.
This permits the use of sequences to address a broad range of systematic
questions. Thus, studies of population variation can be investigated with
mitochondrial DNA (mtDNA) sequences from the noncoding control re-
PHYLOGENETIC INFERENCE AND DNA SEQUENCE ANALYSIS 5

gion, a part of the molecule which often shows significant variation, even
among individual organisms (Greenberg et al., 1983). At the other ex-
treme, the phylogenetic relationships within and among kingdoms can be
evaluated with highly conserved stretches of coding DNA, for example,
those for ribosomal RNAs (rRNAs) (Field et al., 1988; Lake, 1988; Sogin
et al., 1989; Mindell and Honeycutt, 1990). The tremendous flexibility in
their resolving power ensures the future importance and widespread ac-
ceptance of DNA sequences in comparative research.
Several recent advances have now made it possible for systematists, even
those with little formal training in molecular biology, to obtain DNA se-
quence data on a routine basis. Although DNA sequencing has become a
simple laboratory procedure (Sanger et al., 1977; Sanger, 1981), the po-
lymerase chain reaction (PCR), without question, is most responsible for
the increased interest in using DNA sequences in comparative work (Saiki
et al., 1988; Erlich, 1989; Kocher et al., 1989). Pairs of oligonucleotide
primers are chosen to flank the region of DNA that is to be amplified by
PCR. Following cycles of DNA denaturation by heat, primer annealing by
cooling, and strand extension with a thermostabile enzyme such as Taq
polymerase, microgram quantities of double-stranded DNA can be syn-
thesized from nanogram amounts of template. By modifying the ratios of
the two primers, additional amplification cycles can produce single-stranded
DNA (ssDNA), which can then be directly sequenced to obtain nucleotide
data (Gyllensten and Erlich, 1988). As very little template is required
(theoretically, only one molecule), the original samples for PCR need not
be fresh or frozen tissue (Paabo et al., 1988; Kocher et al., 1989). Thus,
DNA can be amplified from alcoholic and formalin specimens, as well as
from skins, feathers, bones, single hairs, and mummified material (Higuchi
et al., 1984; Paabo et al., 1988, 1989; Paabo, 1989). PCR has therefore
greatly enhanced the value of samples which were previously not accessible
to the molecular systematist. The ability to use museum specimens is of
special importance, particularly since natural biological diversity continues
to be lost at an alarming rate. Furthermore, PCR studies of museum spec-
imens collected at different times promise to add a temporal component
to phylogenetic and evolutionary analysis (Thomas et al., 1990).

PHYLOGENETIC INFERENCE USING DNA SEQUENCES:

METHODOLOGICAL PROBLEMS

Despite the spectacular technical advances in obtaining comparative se-

quence data, many difficulties still exist for the investigator, not only during
the process of acquiring the sequence information itself, but also in sub-
sequent analyses undertaken to generate phylogenetic hypotheses for the
taxa of interest. Some of these difficulties and limitations are discussed
below.
6 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Acquisition of Comparative Sequence Data

Most comparative sequence data that have been used in molecular system-
atics were obtained by conventional cloning procedures (Sambrook et al.,
1990), although many investigators have now adopted asymmetrical PCR
as the method of choice (Gyllensten and Erlich, 1988; Allard et al., 1991).
While PCR is relatively simple in theory and practice, many laboratories
attempting to obtain comparative sequence data using asymmetrical am-
plification have often experienced variable success in producing high quality
ssDNA, especially when trying to amplify fragments over 800 to 1000 base
pairs (bp) in length. Yet, other factors may be as important as fragment
length in influencing the quality of single-stranded product from asym-
metrical PCR, and reaction conditions must be carefully optimized (Gyl-
lensten, 1989). These difficulties may be only transitory, however, since
new procedures for obtaining single-stranded template from double-stranded
PCR reactions have the potential for greatly facilitating large-scale com-
parative studies (Higuchi and Ochman, 1989; Mitchell and Merril, 1989).
Only after it becomes possible to obtain large amounts of comparative
sequence information rapidly, accurately, and at reasonable cost, will DNA
sequences fulfill their potential for systematics. In order to answer system-
atic questions, the most important objective is to obtain character-state
data, which can often be accomplished most effectively by aligning, in
tandem, those segments of sequence that are easiest to collect; for example,
by using several so-called universal PCR primer-pairs (Kocher et al., 1989).
Thus, it will sometimes be the case that systematist may not bother to
sequence both complementary strands of DNA, much less, complete genes.
In instances such as this, the goal of the systematist will frequently conflict
with that of the molecular evolutionist who may be interested in different
questions involving the evolutionary patterns seen across entire genes. Such
conflicts are regrettable, given the general importance of DNA sequences
to fields other than molecular systematics (Fickett and Burks, 1989).
Once DNA sequences have been determined, they are deposited in
permanent data banks that can be accessed by others for many different
purposes. Given their general utility, the accuracy of these sequences should
therefore be verified by checking both complements against one another.
For a similar reason, nucleotide sequencing using RNA templates should
be avoided in favor of direct DNA analysis (Hillis et al., 1990).
A major concern of molecular evolution is the relationship between the
higher-level structure of genes and their products (including, for example,
proteins and structural RNAs), and their primary sequences (Gerbi, 1985;
Gautheret et al., 1990; Johnson et al., 1990). Such information is also
relevant to molecular systematics. For example, knowledge about the sec-
ondary structure of rRNA may provide an important basis for assigning
weight to different regions of the molecule in a phylogenetic analysis (Wheeler
and Honeycutt, 1988; Smith, 1989). Consequently, obtaining complete
gene sequences across a spectrum of taxa becomes important, because data
PHYLOGENETIC INFERENCE AND DNA SEQUENCE ANALYSIS 7

of this kind will permit the detailed analyses of molecular structure, func-
tion, and evolution necessary to justify the weighting (Irwin et al., 1991).
Acquiring sequence data is labor intensive; therefore, molecular syste-
matists generally use only single individuals to represent each taxon in their
study. As a consequence, the assumption is made (usually implicitly) that
polymorphism does not affect the phylogenetic analysis because within-
group variation is negligible. However, this assumption may not be well-
founded, especially when the branch points under consideration are of a
relatively recent age, as would be true in the case of a recent radiation of
species within a clade (Nei, 1986). Under these circumstances, the se-
quences may not reflect the true species relationships of the group, even
if a single branching pattern is strongly supported by the data. Instead, the
strong support might reflect a gene phylogeny that may not be congruent
with the cladistic history of the taxa themselves. Such misleading situations
arise because of random sorting of ancestral polymorphic alleles by drift after
cladogenesis. The only way to test whether a gene phylogeny is congruent
with the species phylogeny is to compare the current results against data from
other individuals and for unlinked genes (Nei, 1986; Felsenstein, 1988).

Sequence Alignment
The first step in analyzing nucleotide data is to align the sequences against
one another (Waterman, 1984, 1989; see also papers in Doolittle, 1990). This
initial step is crucial, as all subsequent analyses are dependent on the final
alignment. In many instances (e.g., involving protein-coding sequences of
mtDNA), sequence alignment is easy and can be done by eye, without the
aid of an alignment algorithm. On the other hand, many kinds of sequences
vary so much across taxa that computer-assisted alignment is essential to
minimize the differences among them. Most computer procedures use some
measure of similarity (or dissimilarity) to search for the best alignment for a
given pair of sequences. Different pairwise comparisons are then combined
to produce the final overall result. Because of their indirect approach to the
problem, such strategies may not reveal the optimal alignment for multiple
sequences. Algorithms for the simultaneous comparison of multiple se-
quences related by a given tree topology have been developed, but are
limited by their dependence on the tree topology itself and by the need
for large amounts of computer time (Sankoff, 1975; Sankoff and Ceder-
gren, 1983; see also Waterman et al., and Mindell, this volume).
In molecular systematics, sequence alignment is essentially a problem
of homology1 of character-state data that are the individual nucleotides at
each base position. Morphologists have traditionally recognized homolo-

1
Here, sequence homology is equated with orthology (common ancestry by speciation) and
paralogy (ancestry by gene duplication) is not included in the definition (Fitch, 1970; Pat-
terson, 1988). In phylogenetic analysis, paralogous comparisons are avoided, because they
provide evidence on the order of gene duplications, but not of speciation. Thus, only or-
thologous sequences are of importance in reconstructing species relationships.
8 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

gies by the criteria of similarity and congruence (Patterson, 1988). Hom-

ologies themselves are hypotheses of synapomorphy that are first postu-
lated on the basis of observed similarity (Eldredge and Cracraft, 1980).
After phylogenetic reconstruction, these assessments are then evaluated
according to their agreement, or disagreement, with the most-parsimonious
solution (i.e., with respect to their congruence with other characters).
Those similarities which conflict with the most-parsimonious distribution
of the entire suite of characters are reinterpreted as homoplasies (conver-
gences, parallelisms, and reversals).
Homology has been largely equated with similarity by comparative mo-
lecular biologists (see Patterson, 1988). Unfortunately, this emphasis on
similarity has obscured the conceptual and methodological relationship
between homology and character support for a genealogical hypothesis.
Even though methods of sequence alignment may maximize overall simi-
larity prior to a phylogenetic analysis, hypotheses of base-position ho-
mology nevertheless remain falsifiable by character congruence on the tree
that most parsimoniously explains the distribution of all the data across
the taxa. The same tests of similarity and congruence that apply to mor-
phological characters, therefore, apply to sequence data as well.
The importance of sequence alignment to comparative analysis is high-
lighted by the use of large-subunit rRNA sequences to resolve relationships
among prokaryotes and eukaryotes. The original study of these groups by
Lake (1988) emphasized small-subunit rRNA sequences and concluded
that eocytes and eukaryotes are each other's closest living relatives. A
major implication of this arrangement was the possibility that eukaryotes
originated from a sulfur-metabolizing, thermophilic, anucleate ancestor.
However, both Lake (1990) and Gouy and Li (1990) have noted that the
results for large-subunit rRNA sequences are dependent on the particular
alignment used as well as the choice of species. The instability of the large-
subunit rRNA results reemphasizes the need for more rigorous and efficient
methods of aligning multiple sequences in which the tasks of alignment
and phylogenetic analysis are more fully integrated (Hein, 1989, 1990; see
also Waterman et al., and Mindell, this volume).

Reconstructing Phylogenetic History

That nucleotide sequences provide a rich source of data for resolving phy-
logenetic relationships is no longer disputed. What is being debated, rather,
is not so much how phylogenetic hypotheses can be constructed from those
data (many procedures are capable of generating such hypotheses), but
which method should be preferred and how we might objectively assess
the veracity, or reliability, of the result. These questions define a host of
critical problems within systematics in general, many of which bear on the
nature of comparative evidence and its application to evaluating alternative
hypotheses of relationship. Entangled with these problems are a series of
debates over methods of phylogenetic inference that take as their focus
PHYLOGENETIC INFERENCE AND DNA SEQUENCE ANALYSIS 9

what investigators either assume or infer about the evolutionary dynamics

of nucleotide sequences. None of these controversies, however, appear to
be unique to molecular data. For years, morphologists have debated whether
it is desirable or possible to use assumptions or inferences about the ev-
olutionary characteristics of morphological features and whether one can
apply these to evaluate the reliability of those characters as evidence for
phylogenetic inference.
A good deal is known about the evolutionary dynamics of DNA (e.g.,
see the summaries of Li et al., 1985; Nei, 1987; Li and Graur, 1991), and
assumptions about the nature of its evolution have been incorporated into
all methods of phylogenetic inference in the form of weighting schemes,
correction factors for multiple mutations, and so forth. Thus, it is not the
use of these assumptions per se that would lead one to prefer one method
over another inasmuch as it is generally possible to incorporate any given
assumption into most tree-building algorithms (except in those special cases
when certain assumptions, such as constant rates, are a logical corollary
of the method). Justification for a particular method will therefore have
to come from elsewhere, and various possibilities have been proposed.
Felsenstein (1988), for example, identifies two approaches to justification:
"hypothetico-deductive," which he equates with the application of parsi-
mony; and statistical. Other investigators have sought to justify the choice
of method by simulation analysis in which DNA sequences evolve under
various assumptions and according to "known" phylogenies. The method
of phylogenetic inference that recovers the "true" phylogeny, given the
particular a priori assumptions of the simulation, is preferred over those
which do not (Sourdis and Krimbas, 1987; Sourdis and Nei, 1988; Saitou
and Imanishi, 1989).
There are at least two general sets of methods of phylogenetic inference
that can be applied to sequence data (we appreciate the fact that others
may categorize these differently) (Felsenstein, 1981, 1988; Swofford and
Olsen, 1990). The first set utilizes discrete character data and includes two
well-known approaches; Hennigian cladistics (often called the parsimony
method, but this term can be applied more broadly to other procedures
as well; see below), and maximum likelihood. A second set of procedures
clusters intertaxon similarity/dissimilarity distance measures derived from
paired comparisons of the sequences. Categorizing methods in this way
mirrors some of the most intense debates seen in the recent systematic
literature, and also emphasizes certain theoretical and methodological
problems that should concern all workers undertaking phylogenetic infer-
ence.
It is probably the case that the majority of nucleotide sequence data sets
have been analyzed using distance methods. Surprisingly, however, few
investigators who apply distance algorithms have addressed the extensive
literature criticizing and defending the use of distances in phylogenetic
inference (Farris, 1981, 1983, 1985, 1986a, b; Felsenstein, 1984, 1986).
Distance analysis receives its justification from the argument that phylo-
10 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

genetic inference must be treated as a case of statistical inference (Felsen-

stein, 1984, 1986), yet as the debates have highlighted, this is not entirely
self evident (Farris, 1983). Indeed, the conflict between using methodo-
logical parsimony2 and statistical inference as the basis for justifying any
tree-building method will undoubtedly remain a controversial issue within
systematics for some years to come.
If one chooses to apply a statistical approach, then the data must possess
certain properties in order for the statistical procedures to have validity.
This is the basis for much of the disagreement: it is well known that sys-
tematic data, including DNA sequences, often do not satisfy the underlying
assumptions of statistical models (Sanderson, 1989; Swofford and Olsen,
1990). This appears to be a problem for the statistical viewpoint but not
for methodological parsimony. In principle, at least, methodological par-
simony may be applied no matter what the underlying characteristics of
the data. It is certainly the case that the statistical structure of data is very
much a concern for all methods of analysis—even in cladistics one assumes
independence of characters—but the exact nature of that structure is itself
transparent to methodological parsimony. This forms the basis for the
suggestion that methodological parsimony is a more general approach to
hypothesis evaluation than is statistical inference.
Methodological parsimony is a general criterion in science for ajudicating
the effectiveness of alternative hypotheses in accounting for data (Farris,
1983; Sober, 1983). It applies to all methods of phylogenetic inference
which rely on an optimality criterion (Swofford and Olsen, 1990). Some
quantity is being minimized or maximized in all of these methods, and the
decision to choose the minimum or maximum solution forms the basis for
the application of parsimony. It must be understood, however, that this
approach does not address the truth of the resulting hypotheses. The ap-
plication of statistical methods has been offered as one way to evaluate
the efficacy, or reliability, of hypotheses relative to others. If one chooses
not to apply a statistical approach, however, the most-parsimonious so-
lution remains the single best hypothesis for explaining the data, given the
criterion forming the basis for the parsimony decision. To claim otherwise
is to argue that our choice of hypothesis does not have to conform to
evidence, a claim that will lead scientific discovery nowhere.
As noted, it has become popular to use simulations to compare the
properties of various tree-building algorithms. The reliability of a method
is then judged on whether it accurately identifies the "true" tree used in
the simulation. No one would suggest that a given approach will always
find the true phylogenetic relationships for a set of real taxa, and few would
suggest that simulations are not informative regarding when a particular

2
We distinguish here between parsimony methods and methodological parsimony. The for-
mer is generally equated with cladistic algorithms in which the parsimony criterion is the
minimization of homoplasies. Methodological parsimony is a more general, philosophical
criterion, and keeping the distinction in mind may help clarify the debate over the role of
statistics versus parsimony in phylogenetic inference.
PHYLOGENETIC INFERENCE AND DNA SEQUENCE ANALYSIS 11

procedure might fail to find the correct tree. But there are some funda-
mental philosophical and empirical differences between simulations of fic-
titious taxa and their DNA sequences, on the one hand; and real-world
taxa and their sequence characteristics, on the other. These differences
place important limitations on the general significance of simulated results.
The results of simulation are logically dependent upon the model of
nucleotide evolution used to generate the sequences, along with the pa-
rameters of the model such as generation time, time to cladogenesis, and
sequence length, as well as the algorithm used in the specific analysis. The
evolutionary models used in many simulation studies are exceedingly sim-
ple, and even though they will surely become more sophisticated (e.g.,
more "realistic") in the future, such studies will still face a credibility gap.
There is no single evolutionary model, no matter how complicated, that
can mimic all of the historical complexities of sequence data. Even as
models increase in sophistication, no one who studies DNA would expect
that a particular model of sequence change will necessarily hold across
different clades of taxa. If the assumptions of these models lack veracity
with respect to real-world situations, their relevance to empirical cases will
continue to be questioned.
In short, because of the immense complexities of the real world, simu-
lation studies can only be expected to identify some of the limitations of
a given method of phylogenetic inference. It is unrealistic to expect any
computer model to capture all of the nuances of the molecular evolutionary
process. Thus, the successful reconstruction of a simulated phylogeny by
a particular method does not necessarily guarantee similar success when
the method is applied to actual data. Some important parameter of the
actual data may have been overlooked in designing the simulation, thereby
limiting the general significance of its conclusions. There is a need, there-
fore, for an approach to evaluate methods and phylogenetic results in the
absence of truth about the real world. In the next section, we suggest that
congruence analysis, an extension of methodological parsimony, offers one
such approach for comparing the reliability of competing methods and/or
empirical results.

CONGRUENCE ANALYSIS AS ARBITER

Congruence analysis is a time-honored tradition of science, in general, and

of systematics, in particular. Unfortunately, its importance has been over-
looked within molecular systematics, especially by those searching for a
means to evaluate the reliability of phylogenetic methods. As a principle,
the notion of congruence applies broadly, from instances of pattern rec-
ognition to expectations of theory. Thus, it is congruence of observations
that signifies the presence of a pattern and implies the need for a common
explanation; it is congruence of evidence that allows us to prefer one
hypothesis over another. In both instances, the use of concordance can be
viewed as an extension of the application of methodological parsimony.
12 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

In principle, the true phylogenetic relationships for a given set of taxa

will never be known with certainty. In the absence of such knowledge,
systematics has long relied on studies of congruence among data sets to
assess the reliability of its phylogenetic hypotheses. Well-corroborated pat-
terns of relationship are accepted as the best estimates of the true phytogeny
because it is more parsimonious to accept a common explanation—com-
mon descent, for example—than to conclude that two (or more) identical
results were obtained independently. For such reasons, studies of congru-
ence have been viewed by systematists as a powerful way to resolve difficult
phylogenetic questions (Penny et al., 1982; Kluge, 1989).
Congruence analysis provides an important mechanism with which to
evaluate the reliability of different tree-building methods (see also Mick-
evich, 1978). Given a common data set, those approaches which lead to
congruent results should be preferred over those that do not. Furthermore,
such comparisons can then serve as a basis for investigating why a particular
method has failed to converge on a congruent pattern that is supported by
different sets of data and/or other methods of phylogenetic inference. Stud-
ies of congruence, therefore, can provide insights into the limitations and
assumptions of different tree-making algorithms.
We suggest that empirical investigations of congruence among actual
data sets using different tree-building procedures, along with varying the
assumptions of those methods, will likely provide more insight into the
efficacy of these approaches than will simulation analysis, which is generally
far removed from the real world. One approach might rely on phylogenetic
hypotheses of major groups of organisms that are highly corroborated by
a variety of molecular and nonmolecular data, such as higher-level rela-
tionships within the primates (Miyamoto and Goodman, 1990). Those
methodological approaches which lead to congruence with well-corrobo-
rated hypotheses should be preferred over those that rarely do. In this
way, the reliability of different tree-building methods can be evaluated
against our best estimates of phylogenetic relationship without interference
from oversimplifications of the real world and from individual bias during
the selection of parameters in simulation analyses.
Congruence analysis also permits one to evaluate the reliability of dif-
ferent weighting schemes of character transformations for use in a phy-
logenetic study. In their phylogenetic investigation of cervid and other
artiodactyl mtDNA sequences, Kraus and Miyamoto (1991) found that
cladistic analysis of all mutations (transitions, transversions, and gap events)
resulted in a phylogeny supporting the monophyly of antlered deer in the
family Cervidae. In contrast, when only transversions were counted, the
same method (i.e., parsimony) led to a solution whereby antlered deer
were not monophyletic. The former arrangement favoring antlered deer
monophyly is corroborated by morphological data, unlike its alternative
solution. Thus, congruence suggests in this case that the use of all mutations
provides more reliable results than analysis of transversion differences
alone. The failure of transversion parsimony to obtain a reliable result can
PHYLOGENETIC INFERENCE AND DNA SEQUENCE ANALYSIS 13

be attributed to the loss of cladistic information offered by transitions at

these lower hierarchical levels.
The primary limitation on obtaining knowledge of phylogenetic history
has not been so much with difficulties encountered with any particular
method, but often with the actual data (Kraus and Miyamoto, 1991). We
either have too few data or the evidence is of such poor quality (because
of extensive homoplasy, intraspecific polymorphism, paralogy) that no
method can use them effectively. Conversely, when the data are highly
structured, the same groupings will very likely emerge no matter what
method of phylogenetic inference is employed. The real problem occurs
when different methods yield incongruent results for the same data. Is this
the result of the methods per se, the choice of underlying assumptions used
in the approach (e.g., correction factors, weighting schemes), or because
the data are ambiguous? By relying on studies of congruence, answers to
such questions will continue to accumulate for real-world situations.

THE FUTURE OF MOLECULAR SYSTEMATICS

The growing importance of DNA sequences for phylogenetic inference and

for analysis of evolutionary processes at the molecular level requires that
investigators strive for complete gene sequences whose accuracy has been
verified by sequencing both complements. Such efforts will enhance the
overall value of these comparative data to biology as a whole. With regard
to systematics, a more detailed understanding of molecular evolution will
facilitate the development of improved methods of phylogenetic recon-
struction. Because large amounts of sequence data will be needed to ad-
dress problems about the extent of intraspecific polymorphism, and to
resolve many vexing phylogenetic questions, we might expect that coop-
erative research projects involving different laboratories will be required.
Continued technological developments (with respect to collecting se-
quence data), coupled with the vast phylogenetic information contained
in nuclear and extranuclear genomes, virtually guarantees that nucleotide
sequences will become the primary source of systematic data in the near
future. Before the full potential of these data is realized, however, many
of the problems noted above will need to be resolved. We have proposed
that congruence analysis will play an important role in resolving contro-
versies over systematic methods and phylogenetic relationships. Congru-
ence provides the ultimate test of reliability in the absence of revealed
truth, and as such, adoption of this methodological criterion can be re-
garded as crucial to the continued development of the field. Despite the
growing importance of sequence data, it cannot be stressed strongly enough
that there remains a pressing need to enlarge our nonmolecular database
of systematic characters. If we are to gain meaningful insights into the
"whats," "hows," and "whys" of the history of life, phylogenetic studies
14 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

will have to rely on all available comparative information, both molecular

and nonmolecular.

ACKNOWLEDGMENTS

We thank M. W. Allard, D. P. Mindell, and M. R. Tennant for their

helpful suggestions about the manuscript. This research was supported by
grants from the National Science Foundation to MMM (BSR-8857264 and
BSR-8918606) and JC (BSR-8805957 and BSR-9007652).

REFERENCES

Allard, M.W., D.L. Ellsworth, and R.L. Honeycutt. (1991) The production of
single-stranded DNA suitable for sequencing using the polymerase chain
reaction. BioTechniques 10: 24-26.
Doolittle, R.F. (ed.). (1990) Molecular Evolution: Computer Analysis of Protein
and Nucleic Acid Sequences. Methods in Enzymology, vol. 183. Academic
Press, New York.
Eldredge, N., and J. Cracraft. (1980) Phylogenetic Patterns and the Evolutionary
Process. Columbia University Press, New York.
Erlich, H.A. (ed.). (1989) PCR Technology. Principles and Applications for DNA
Amplification. Stockton Press, New York.
Farris, J.S. (1981) Distance data in phylogenetic analysis. Pp. 3-23 in Advances
in Cladistics: Proceedings of the First Meeting of the Willi Hennig Society
(V.A. Funk and D.R. Brooks, eds.). New York Botanical Garden, Bronx.
Farris, J.S. (1983) The logical basis of phylogenetic analysis. Adv. Cladistics 2:1-
36.
Farris, J.S. (1985) Distance data revisited. Cladistics 1: 67-85.
Farris, J.S. (1986a) On the boundaries of phylogenetic systematics. Cladistics 2:
14-27.
Farris, J.S. (1986b) Distances and statistics. Cladistics 2: 144-157.
Felsenstein, J. (1981) Numerical methods for inferring evolutionary trees. Q. Rev.
Biol. 57: 379-404.
Felsenstein, J. (1984) Distance methods for inferring phylogenies: a justification.
Evolution 38: 16-24.
Felsenstein, J. (1986) Distance methods: a reply to Farris. Cladistics 2: 130-143.
Felsenstein, J. (1988) Phylogenies from molecular sequences: inference and reli-
ability. Ann. Rev. Genet. 22: 521-565.
Fickett, J.W., and C. Burks. (1989) Development of a database for nucleotide
sequences. Pp. 1-34 in Mathematical Methods for DNA Sequences (M.S.
Waterman, ed.). CRC Press, Boca Raton.
Field, K.G., G.J. Olsen, DJ. Lane, S.J. Giovannoni, M.T. Ghiselin, E.G. Raff,
N.R. Pace, and R.A. Raff. (1988) Molecular phylogeny of the animal king-
dom. Science 239: 748-753.
Gautheret, D., F. Major, and R. Cedergren. (1990) Computer modeling and display
of RNA secondary and tertiary structures. Pp. 318-330 in Molecular Ev-
olution: Computer Analysis of Protein and Nucleic Acid Sequences (R.F.
PHYLOGENETIC INFERENCE AND DNA SEQUENCE ANALYSIS 15

Doolittle, ed.)- Methods in Enzymology, vol. 183. Academic Press, New

York.
Gerbi, S.A. (1985) Evolution of ribosomal DNA. Pp. 419-517 in Molecular Ev-
olutionary Genetics (R.J. Maclntyre, ed.). Plenum Press, New York.
Gouy, M., and W.-H. Li. (1990) Archaebacterial or eocyte tree? Nature 343: 419.
Greenberg, B.D., I.E. Newbold, and A. Sugino. (1983) Intraspecific nucleotide
sequence variability surrounding the origin of replication in human mito-
chondrial DNA. Gene 21: 33-49.
Gyllensten, U. (1989) Direct sequencing of in vitro amplified DNA. Pp. 45-60 in
PCR Technology. Principles and Applications for DNA Amplification (H.A.
Erlich, ed.). Stockton Press, New York.
Gyllensten, U.B., and H. A. Erlich. (1988) Generation of single-stranded DNA by
the polymerase chain reaction and its application to direct sequencing of
the HLA-DQA locus. Proc. Natl. Acad. Sci. USA 85: 7652-7656.
Hein, J. (1989) A new method that simultaneously aligns and reconstructs ancestral
sequences for any number of homologous sequences, when the phylogeny
is given. Mol. Biol. Evol. 6: 649-668.
Hein, J. (1990) Unified approach to alignment and phylogenies. Pp. 626-645 in
Molecular Evolution: Computer Analysis of Protein and Nucleic Acid Se-
quences (R.F. Doolittle, ed.). Methods in Enzymology, vol. 183. Academic
Press, New York.
Hennig, W. (1966) Phylogenetic systematics. University of Illinois Press, Urbana.
Higuchi, R., B. Bowman, M. Freiberger, O.A. Ryder, and A.C. Wilson. (1984)
DNA sequences from the quagga, an extinct member of the horse family.
Nature 312: 282-284.
Higuchi, R.G., and H. Ochman. (1989) Production of single-stranded DNA tem-
plates by exonuclease digestion following the polymerase chain reaction.
Nucleic Acids Res. 17: 5865.
Hillis, D.M. (1987) Molecular versus morphological approaches to systematics.
Ann. Rev. Ecol. Syst. 18: 23-42.
Hillis, D.M., A. Larson, S.K. Davis, and E.A. Zimmer. (1990) Nucleic acids III:
sequencing. Pp. 318-370 in Molecular Systematics (D.M. Hillis and C. Mor-
itz, eds.). Sinauer Associates, Sunderland.
Hillis, D.M., and C. Moritz (eds.). (1990) Molecular Systematics. Sinauer Asso-
ciates, Sunderland.
Hull, D.L. (1989) The evolution of phylogenetic systematics. Pp. 3-15 in The
Hierarchy of Life. Molecules and Morphology in Phylogenetic Analysis (B.
Fernholm, K. Bremer, and H. Jornvall, eds.). Elsevier Science Publishers
B.V., Amsterdam.
Irwin, D.M., T.D. Kocher, and A.C. Wilson. (1991) Evolution of the cytochrome
b gene of mammals. J. Mol Evol. 32: 128-144.
Johnson, M.S., A. Sali, and T.L. Blundell. (1990) Phylogenetic relationships from
three-dimensional protein structures. Pp. 670-690 in Molecular Evolution:
Computer Analysis of Protein and Nucleic Acid Sequences (R.F. Doolittle,
ed.) Methods in Enzymology, vol. 183. Academic Press, New York.
Kluge, A.G. (1989) A concern for evidence and a phylogenetic hypothesis of
relationships among Epicrates (Boidae, Serpentes). Syst. Zool. 38: 7-25.
Kocher, T.D., W.K. Thomas, A. Meyer, S.V. Edwards, S. Paabo, F.X. Villa-
blanca, and A.C. Wilson. (1989) Dynamics of mitochondrial DNA evolution
in animals: amplification and sequencing with conserved primers. Proc. Natl.
Acad. Sci. USA 86: 6196-6200.
16 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Kraus, F., and M.M. Miyamoto. (1991) Rapid cladogenesis among the pecoran
ruminants: evidence from mitochondrial DNA sequences. Syst. Zool. 40:
117-130.
Lake, J.A. (1988) Origin of the eukaryotic nucleus determined by rate-invariant
analysis of rRNA sequences. Nature 331: 184-186.
Lake, J.A. (1990) Archaebacterial or eocyte tree? Nature 343: 418-419.
Li, W.-H., and D. Graur. (1991) Fundamentals of Molecular Evolution. Sinauer
Associates, Sunderland.
Li, W.-H., C.-C. Luo, and C.-I. Wu. (1985) Evolution of DNA sequences. Pp.
1-94 in Molecular Evolutionary Genetics (R.J. Maclntyre, ed.). Plenum
Press, New York.
Mickevich, M.F. (1978) Taxonomic congruence. Syst. Zool. 27: 143-158.
Mindell, D.P., and R.L. Honeycutt. (1990). Ribosomal RNA in vertebrates: ev-
olution and phylogenetic applications. Ann. Rev. Ecol. Syst. 21: 541-566.
Mitchell, L., and C.R. Merril. (1989) Affinity generation of single-stranded DNA
for dideoxy sequencing following the polymerase chain reaction. Anal.
Biochem. 178: 239-242.
Miyamoto, M.M., and M. Goodman. (1990) DNA systematics and evolution of
primates. Ann. Rev. Ecol. Syst. 21: 197-220.
Nei, M. (1986) Stochastic errors in DNA evolution and molecular phylogeny. Pp.
133-147 in Evolutionary Perspectives and the New Genetics (H. Gershowitz,
D.L. Rucknagel, and R.E. Tashian, eds.). Alan R. Liss, New York.
Nei, M. (1987) Molecular Evolutionary Genetics. Columbia University Press, New
York.
Nelson, G., and N. Platnick. (1981) Systematics and Biogeography: Cladistics and
Vicariance. Columbia University Press, New York.
Paabo, S. (1989) Ancient DNA: extraction, characterization, molecular cloning,
and enzymatic amplification. Proc. Natl. Acad. Sci. USA 86: 1939-1943.
Paabo, S., J.A. Gifford, and A.C. Wilson. (1988) Mitochondrial DNA sequences
from a 7000-year old brain. Nucleic Acids Res. 16: 9775-9787.
Paabo, S., R.G. Higuchi, and A.C. Wilson. (1989) Ancient DNA and the poly-
merase chain reaction. The emerging field of molecular archaeology. /. Biol.
Chem. 264: 9709-9712.
Patterson, C. (1987) Introduction. Pp. 1-22 in Molecules and Morphology in Ev-
olution: Conflict or Compromise! (C. Patterson, ed.). Cambridge University
Press, New York.
Patterson, C. (1988) Homology in classical and molecular biology. Mol. Biol. Evol.
5: 603-625.
Penny, D., L.R. Foulds, and M.D. Hendy. (1982) Testing the theory of evolution
by comparing phylogenetic trees constructed from five different protein
sequences. Nature 297: 197-200.
Rieppel, O.C. (1988) Fundamentals of Comparative Biology. Birkhauser Verlag,
Basel.
Saiki, R.K., D.H. Gelfand, S. Stoffel, S.J. Scharf, R. Higuchi, G.T. Horn, K.B.
Mullis, and H.A. Erlich. (1988) Primer-directed enzymatic amplification of
DNA with a thermostable DNA polymerase. Science 239: 487-491.
Saitou, N., and T. Imanishi. (1989) Relative efficiencies of the Fitch-Margoliash,
maximum-parsimony, maximum-likelihood, minimum-evolution, and neigh-
bor-joining methods of phylogenetic tree construction in obtaining the cor-
rect tree. Mol. Biol. Evol. 6: 514-525.
PHYLOGENETIC INFERENCE AND DNA SEQUENCE ANALYSIS 17

Sambrook, J., E. Fritsch, and T. Maniatis. (1990) Molecular Cloning. A Laboratory

Manual/second edition. Cold Spring Harbor Laboratory, Cold Spring Har-
bor.
Sanderson, M.J. (1989). Confidence limits on phylogenies: the bootstrap revisited.
Cladistics 5: 113-129.
Sanger, F. (1981) Determination of nucleotide sequences in DNA. Science 214:
1205-1210.
Sanger, F., S. Nicklen, and A.R. Coulson. (1977) DNA sequencing with chain-
terminating inhibitors. Proc. Nad. Acad. Sci. USA 74: 5463-5467.
Sankoff, D. (1975) Minimal mutation trees of sequences. SIAM J. Appl. Math.
28: 35-42.
Sankoff, D., and R.J. Cedergren. (1983) Simultaneous comparison of three or
more sequences related by a tree. Pp. 253-263 in Time Warps, String Edits,
and Macromolecules: The Theory and Practice of Sequence Comparison (D.
Sankoff and J.B. Kruskal, eds.). Addison-Wesley, Reading.
Smith, A.B. (1989) RNA sequence data in phylogenetic reconstruction: testing the
limits of its resolution. Cladistics 5: 321-344.
Sober, E.R. (1983) Parsimony methods in systematics. Adv. Cladistics 2: 37-47.
Sober, E.R. (1988) Reconstructing the Past: Parsimony, Evolution, and Inference.
MIT Press, Cambridge.
Sogin, M.L., U. Edman, and H. Elwood. (1989) A single kingdom of eukaryotes.
Pp. 133-143 in The Hierarchy of Life. Molecules and Morphology in Phy-
logenetic Analysis (B. Fernholm, K. Bremer, andH. Jornvall, eds.). Elsevier
Science Publishers B.V., Amsterdam.
Sourdis, J., and C. Krimbas. (1987) Accuracy of phylogenetic trees estimated from
DNA sequence data. Mol. Biol. Evol. 4: 159-166.
Sourdis, J., and M. Nei. (1988) Relative efficiencies of the maximum parsimony
and distance-matrix methods in obtaining the correct phylogenetic tree. Mol.
Biol. Evol. 5: 298-311.
Swofford, D.L., and G.J. Olsen. (1990) Phylogeny reconstruction. Pp. 411-501
in Molecular Systematics (D.M. Hillis and C. Moritz, eds.). Sinauer Asso-
ciates, Sunderland.
Thomas, W.K., S. Paabo, F.X. Villablanca, and A.C. Wilson. (1990) Spatial and
temporal continuity of kangaroo rat populations shown by sequencing mi-
tochondrial DNA from museum specimens. /. Mol. Evol. 31: 101-112.
Waterman, M.S. (1984) General methods of sequence comparison. Bull. Math.
Biol. 46: 473-500.
Waterman, M.S. (1989) Sequence alignments. Pp. 53-92 in Mathematical Methods
for DNA Sequences (M.S. Waterman, ed.). CRC Press, Boca Raton.
Wheeler, W.C., and R.L. Honeycutt. (1988) Paired sequence difference in ribo-
somal RNAs: evolutionary and phylogenetic implications. Mol. Biol. Evol.
5: 90-96.
Wiley, E.O. (1981) Phylogenetics. The Theory and Practice of Phylogenetic Sys-
tematics. J. Wiley Sons, New York.
2
DNA Sequencing: Strategy and Methods
to Directly Sequence Large DNA
Molecules

JERRY L. SLIGHTOM, DAVID R. SIEMIENIAK, AND

LEANG C. SIEU

The goal of the Human Genome Initiative, to produce the complete nu-
cleotide sequence of the genome of human and other organisms, will re-
quire the development of rapid and efficient methods for obtaining nu-
cleotide sequences. Thus, improvements in present sequencing methods
are of great interest and importance if we hope to sequence the 3 x 109
base pair (bp) human genome. Presently, two methods exist for obtaining
nucleotide sequences, the enzymatic method developed by Sanger et al.
(1977) which uses DNA polymerases and chain termination dideoxy nu-
cleotides for base determinations; and the chemical method developed by
Maxam and Gilbert (1977) which uses base-specific chemical reactions for
base determinations. Of these two methods, the enzymatic method has
been found to be readily adaptable to automation, and thus has received
more attention in the form of research on different polymerases and re-
agents for the utilization of these polymerases. The T7 polymerase de-
scribed by Tabor and Richardson (1987) and a polymerase from Thermus
aquaticus (Taq) (Saiki et al., 1988) have, for the most part, replaced the
use of Escherichia coli DNA polymerase I (Klenow fragment) and reverse
transcriptase as the enzymes of choice for enzymatic sequencing.
Automation of different aspects of the nucleotide sequencing reactions
(DNA purification, enzymatic reactions, and information transfer) has been
the subject of numerous papers, and these steps are continuously being
improved. Some of the laborious aspects of DNA sequencing have been
partially automated, such as the use of robotic workstations to perform
enzymatic reactions (Frank et al., 1988; Wilson et al., 1988; Zimmerman
et al., 1988) the use of film readers (Elder et al., 1986), and informational

18
DNA SEQUENCING: STRATEGY AND METHODS 19

transfer using fluorescent tags (Smith et al., 1986; Ansorge et al., 1987;
Prober et al., 1987).
Present methods using random sequencing strategies for automating DNA
sequencing suffer from several weaknesses: (1) the need to subclone short
DNA fragments; (2) the length of reliable nucleotide sequence reads are
short, in the range of 500 bp; (3) meaningful analysis cannot be done until
a substantial amount of data has been accumulated; and (4) the strategy
requires a considerable amount of "over-sequencing." For the random
sequencing strategy, it is generally accepted that an insert should be se-
quenced a minimum of five times (Edwards et al., 1990), after which the
individual sequences are assembled into a contiguous sequence with the
aid of computer programs known as random sequence handlers (Staden,
1982). These computer programs do have their limitations, especially when
repetitive sequences are encountered, such as the highly repetitive Alu
element. In many cases, even after excessive random sequencing of an
insert, gaps remain that must be closed using a directional sequencing
approach. It would be much more efficient if large DNA molecules could
be directly sequenced, in a directional manner, because coverage could be
done by sequencing the insert only 2.5 times, allowing for overlaps between
sequencing runs on both strands.
This report describes our attempt to develop a nucleotide sequencing
strategy that avoids many of the problems described above, yet remains
as simple as possible (relying on proven methods). Our strategy has two
major goals. First, to conveniently sequence large DNA molecules [50
kilobase (kb) range]; and second, to obtain the longest sequence reads as
possible. We describe here methods for obtaining the nucleotide sequence
of cosmid and 1 cloned inserts utilizing mostly double-stranded enzymatic
sequencing (Sanger et al., 1977; Chen and Seeburg, 1985; Zagursky et al.,
1985) using T7 polymerase, and to a smaller degree, the chemical sequenc-
ing method (Maxam and Gilbert, 1977).

EXPERIMENTAL AND DISCUSSION

Strategy for Directly Sequencing Cosmid and A Cloned DNA Inserts

The use of T7 polymerase to obtain nucleotide sequence information from
large DNA molecules such as cosmid and A clones has been described
previously (Zagursky et al., 1985; Manfioletti and Schneder, 1988; Leve-
dakou et al., 1989; Kesterson et al., 1989). However, these methods are
somewhat limited because sequences are obtained only from the vector
cloning junctions or from genetic elements (e.g., genes, repetitive DNA)
known to be located within the cloned insert. Having such a small number
of sequencing points for a "walk" across a large insert would be too time
consuming. This problem can be overcome by obtaining more sequencing
points (internal to the insert) from which to initiate double-stranded di-
20 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

deoxy sequencing; these points are referred to as initiation points (IPs).

These additional IPs should be obtained using a method that is independent
of the cloned DNA insert; that is, it could be used for any cloned insert.
Several methods for obtaining IPs are feasible: (1) the use of the chemical
sequencing procedure (Maxam and Gilbert, 1977); (2) the use of splinkers
to add a known primer sequence to a restriction enzyme site (Kalisch et
al., 1986); (3) the more recently developed use of exonuclease III (Sorge
and Blinderman, 1989); and (4) the use of short oligomer primers for
random priming (Studier, 1989; Siemieniak and Slightom, 1990). Any one
of these methods is appropriate for obtaining additional IPs, and extensive
sequence reads from IPs are not necessary, as the switch to oligomer
primer-directed sequencing method requires an accurate sequence reading
of only the length of the oligomer primer. However, after switching to the
enzymatic sequencing method, its efficiency can be greatly improved as
the length of the sequence reads is increased. Thus, our second goal, to
obtain the longest reads possible (see below), allows us to reduce the
following: the number of oligomer primers needed, the number of enzy-
matic reactions, the amount of template DNA, the amount of overlapping
sequence, and, most importantly, the number of sequencing gels needed.
Our sequencing strategy for large DNA molecules is outlined in Table
2-1. For this strategy we use the chemical sequencing method to obtain
IPs by obtaining the sequence of small (100 to 1,500 bp) fragments from
within the cloned insert. These small DNA fragments are obtained by using
two different restriction enzymes known to yield, upon sequential diges-

Table 2-1 Strategy for Sequencing Large DNA Inserts

1. Obtain "stable" cosmid or A DNA clones. The use of overlapping clones is desirable but
not necessary.
2. Obtain oligomers specific for the 5' and 3' regions flanking the vector cloning site.
3. Analyze insert with various restriction enzymes for the identification of enzyme pairs that
can be used to obtain single 32P-end-labeled fragments (100 to 1,500 bp in length). The
sequence of these 5'-end-labeled fragments provides internal IPs.
4. Sequence the insert 5' and 3' flanking regions using flanking oligomer primers and obtain
the sequence of four to five 32P-end-labeled internal fragments (see Fig. 2-1).
5. Design oligomer primers to continue the sequence of the flanking and IPs within the
cloned insert.
6. Use PCR amplification to map the orientation and relative positions of sequence contigs
with respect to each other and flanking contigs (see Fig. 2-2). Obtain additional IPs if
needed.
7. Extend sequence contigs using the enzymatic sequencing method; enter new sequence
information into the computer as soon as available, and order new sets of oligomer primers.
Alternate among clones (if available) to avoid waiting for the delivery of oligomer primers.
Each oligomer set should yield about 5 to 8 kb of new sequence information.
8. Order opposite-strand oligomer primers as sequence information becomes available, and
store oligomers until needed or until time allows for their use. This allows for effective
use of DNA synthesis instrument, and will accelerate obtaining the sequence of the op-
posite strand.
9. Continue the sequencing process described in steps 7 and 8 until all contigs merge.
DNA SEQUENCING: STRATEGY AND METHODS 21

tion, identifiable DNA fragments. The first restriction enzyme digestion

product is subjected to 5'-end-labeling with [32P-y]-adenosine triphosphate
(ATP) (see below) followed by digestion with the second enzyme which
releases a single 32P-end-labeled DNA fragment. Generally, the sequence
of a small number (four to eight) of these unmapped DNA fragments are
obtained from a cloned insert. Each IP is extended in both 5' and 3'
directions, resulting in a growing continuous DNA sequence region that
is referred to as a "sequencing contig."
The overall efficiency of this method can be greatly enhanced by se-
quencing several large cloned inserts at one time; this process helps to
eliminate "dead time" between the delivery of oligomer primers. Presently,
extensive regions of the genomes of human and many other species have
been cloned as sets of overlapping clones; for example, the T-cell receptor
variable gene families (Lai et al., 1988). The use of this strategy for se-
quencing three overlapping cosmid clones is schematically illustrated in
Figure 2-1. The use of overlapping cosmid clones is also useful because
the overlapping boundary regions provide additional IPs within the adja-
cent cosmid clones (Fig. 2-1).
The relative location of the randomly generated IPs can be mapped using
polymerase chain reaction (PCR) amplification. The 5' and 3' oligomer
primers from each IP are used in a series of PCR amplifications designed
to cross-challenge each primer, including the oligomers from regions flank-
ing the vector cloning site. An example of this type of PCR mapping is
shown in Figure 2-2, where the relative location of four contigs was de-
termined; two contigs are separated by about 4.1 kb (lane 1), and another
pair of contigs is separated by about 8 kb (lane 6). Mapping by PCR can
be used effectively for determining the location of contigs even if they are
separated by a considerable distance; knowledge of the relative locations
of four to eight contigs would be sufficient to obtain a partial map of a
complete 40 kb insert. The method used to amplify the DNA fragments
in the range of 8 kb is identical to that described by Saiki et al. (1988), except
that 5 mM DTT was used and the extension time was 10 minutes. This
mapping information is useful in determining if the IPs are sufficiently spread
throughout the insert to allow for efficient sequencing; if not, additional IPs
should be obtained before closure of existing (closely spaced) contigs.
With the synthesis of 10 to 16 oligomer primers (for five to eight contigs,
including flanking contigs) the first round of oligomer-directed sequencing
can be initiated. After sequencing and film exposure, the sequence infor-
mation near the ends of each read can be used for designing the next set
of oligomer primers. The remaining sequence information can be read
while waiting for delivery of the new oligomer primer set. This process is
continued until the ends of each sequencing contig merge. The primary
goal of this initial closure is to obtain a DNA sequence of the insert at a
confidence level of about 95% or better. While accumulating the first strand
sequence, the oligomer primers needed for sequencing the opposite strand
should be ordered.
Figure 2-1 Schematic representation of the sequencing strategy outlined in Table 2-1. Here we show how the
strategy could be used to obtain the sequence of three overlapping cosmid clones. The region of cloned genomic
DNA is represented by the top line and the positions of three overlapping cosmid clones are shown below this
line. Internal initiation points (IPs) are obtained by chemically sequencing 32P-end-labeled fragments that are
the result of sequential digestion with two restriction enzymes; the use of the restriction enzymes EcoRI (E) and
BamHI (B) is illustrated. Nucleotide sequences obtained by the chemical sequencing of 32P-end-labeled fragments
are shown by the arrows above the cosmid cloned inserts. Additional internal contigs are obtained at the 5'
and 3' overlapping regions, which in this illustration adds four sequencing contigs. Extensions of the flanking
and internal sequencing contigs, using the enzymatic sequencing method, are shown by arrows below the
cosmid cloned inserts.
DNA SEQUENCING: STRATEGY AND METHODS 23

Figure 2-2 PCR amplification mapping of sequencing contigs located within cos-
mid H7.1, which contains part of the human T-cell receptor/3-gene family (Lai et
al., 1988). Oligomer primers corresponding to DNA regions which flank (5' and
3') the cloning site of the cosmid vector pTL-5 (Lund et al., 1982) were used in
various combinations with oligomer primers from internal sequencing contigs.
The lanes marked M, contain the kb DNA size standard purchased from Bethesda
Research Laboratories (BRL). Lane 1 shows that two internal sequencing contigs
are separated by about 4.1 kb. Lanes 2 to 5 and 7 show no amplified DNA frag-
ment indicating that the oligomer primer parts used are either separated by large
distances, or that the primer pairs are not orientated in the direction needed for
amplification (Saiki et al., 1988). Lane 6 shows that two other sequencing contigs
are separated by a distance of about 8 kb.

The strategy described above and in Table 2-1 possesses several distinct
advantages, which include:
1. no subcloning of cosmid and A inserts is needed;
2. limited restriction enzyme site mapping data are needed;
3. the use of overlapping inserts increases efficiency;
4. meaningful analysis of the accumulated sequence information is
straightforward;
5. inserts can be completely sequenced by sequencing only 2.5-fold;
6. films provide hard-copy data storage.
24 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Of these advantages, the fourth is particularly interesting, especially to the

individual(s) involved in obtaining the sequence information because it
allows them the opportunity to immediately evaluate the accumulated data.
In fairness, to those who may want to try our strategy, we must also
describe some disadvantages that are associated with the direct sequencing
of large DNA molecules:
1. the need to isolate large amounts of cosmid or A DNAs (about 400
|xg for a cosmid and about 200 u,g for a /I insert);
2. the need for a large number of oligomer primers (that presently cost
about $100 each);
3. insert instability can be a problem (however, this is a problem for
any sequencing method);
4. lack of an accurate film reader for transferring sequence information
into a computer.
Of these disadvantages, the lack of an efficient method to transfer infor-
mation into the computer is the most severe. There are many DNA se-
quencing methods that have the ability to generate sequencing films faster
than they can be read. The need to obtain a large amount of DNA template
is not an extreme disadvantage, but does require a commitment to isolated
cosmid and A DNAs from several liters of E. coli cells. From the standpoint
of cost and the need for a large number of oligomer primers, this method
is quite expensive, and at our present price it would cost more than $10,000
for only the oligomers to sequence a 40 kb insert. However, because of
recent advances in the synthesis of oligomer primers (K. Beattie, personal
communication), their cost should be reduced considerably (about one-
fifth to one-tenth) in the near future, at which time this disadvantage will
be removed from the list.

The Chemical Sequencing Method

The use of specific chemicals for DNA strand breakage to generate a
breakage pattern which is consistent with the four DNA bases was devel-
oped by Maxam and Gilbert (1977). A more detailed description of these
procedures, including trouble-shooting information for the base-specific
reactions, has also been described by Maxam and Gilbert (1980). The
chemical reactions involve three consecutive steps: base modification, re-
moval of the modified base from its sugar, followed by strand breakage at
the exposed sugar molecule. The specific base reactions involve the use of
dimethyl sulfate (DMS) to methylate the N-7 of guanine and the use of
the nucleophile hydrazine to split thymine or cytosine ring structures. The
identification of either a thymine or cytosine ring structure is achieved by
performing the reaction in the presence of 1 M NaCl; only cytosine reacts
at an appreciable rate. The presence of an adenine can be determined by
treatment with formic acid, which weakens both adenine and guanine gly-
cosidic bonds by protonating the purine ring nitrogens. In all cases, pi-
DNA SEQUENCING: STRATEGY AND METHODS 25

peridine is used to catalyze /J-elimination of phosphates from the exposed

sugar, which results in breakage of the DNA strand. However, the base-
specific chemical reaction is the final reaction step in this DNA sequencing
method. Prior to performing these reactions, the DNA fragment must be
end-labeled with [32P-y]-ATP, separated on a preparation gel, extracted
from the gel, and then subjected to the base-specific chemical reactions.
Although the chemical sequencing procedure has been extensively used
for the past 14 years, its use has declined during the past few years due to
the use of modified DNA polymerases, such as T7 and Tag polymerases,
which are much more reliable than E. coll polymerase I (Klenow fragment).
In the past, sequences obtained using the chemical procedure were con-
sidered to be less subject to errors. However, the reactions required con-
siderable trouble shooting, quality control of the chemicals used, and very
pure DNAs. These drawbacks have probably been responsible for much
of its continued decline and disfavor. Despite these problems, we have
found the chemical sequencing method to be very useful for obtaining
sequence information from regions internal to large DNA inserts (see Table
2-1). We feel that the advantage of obtaining additional IPs compensates
for difficulties encountered in using the chemical sequencing procedure.
Because we have had extensive experience with the chemical sequencing
procedure, we do not consider its use to be difficult. However, we do
realize that this could be a problem for others who do not have such
experience. We have included this section to detail some of the procedures
and precautions that we use to eliminate many of the problems which
beginners have with chemical sequencing. The following descriptions high-
light several problem areas associated with chemical DNA sequencing.

DNA Purification
The use of impure plasmid, cosmid, or A DNAs can be responsible for
many of the problems common to both chemical and enzymatic sequencing
procedures. It is important to use a DNA purification method that con-
sistently yields DNA which can be sequenced. The use of crude lysate
procedures is not appropriate for chemical sequencing due to the presence
of large amounts of E. coll DNA and RNAs that interfere with the 5'-end-
labeling procedure (see below). For the purification of A phage we have
found the best method involves the use of a CsCl step gradient followed
by a CsCl equilibrium gradient (Slightom and Drong, 1988).
The isolation of plasmid and cosmid DNAs can be accomplished by any
number of methods, and many of these methods are modifications of the
alkaline-extraction procedure described by Birnboim and Doly (1979). One
particular derivation of this method involves the use of 7 M NH4OAc for
the neutralization step (Morelle, 1989). Table 2-2 describes a modification
of this plasmid isolation procedure for the large-scale preparation of plas-
mid or cosmid DNAs. The most important part of the procedure is the
use of two CsCl-ethidium bromide gradient bandings, using a fixed-angle
26 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Table 2-2 Preparation of Plasmid or Cosmid DNAs for Sequencing by Chemical

or Enzymatic Methods
1. Grow bacteria overnight culture (50 mL) in appropriate antibiotic.
2. Add supplements to M9 media (Maniatis et al., 1982): 1 mL of 0.1 M CaCl2; 1 mL of
1.0 M MgSO4; 10 mL of 40% glucose.
3. Add appropriate antibiotic and 10 mL of bacterial overnight growth to 1 L of M9 media.
Shake at 37°C, 150 rpm until O.D.650 reaches about 0.5.
4. If plasmid can be amplified, add 150 mg of chloramphenicol (3 mL of stock 50 mg/mL
in EtOH) per L of M9. Amplify overnight at 37°C shaking at 120 rpm.
5. After overnight growth, chill cells on ice followed by centrifugation in 1 L bottles at
4,200 rpm (Sorval RC-3B) for 15 minutes.
6. Resuspend cells in 25 mL of 0.9% NaCl, 10 mM Tris-HCl, 10 mM EDTA (pH 7 0).
Transfer into a 50 mL polypropylene Oak Ridge tube and centrifuge 4,200 rpm for 10
minutes.
7. Resuspend cells in 5 mL of 50 mM glucose, 25 mM Tris-HCl, 10 mM EDTA (pH 8.0)
and 8 mg/mL lysozyme. Incubate at room temperature for 10 minutes.
8. Add 10 mL of freshly prepared alkaline solution (0.2 N NaOH and 1% SDS). Mix by
gently inverting the tube 3 to 6 times. The suspension should become somewhat clear
and viscous. Place on ice for about 10 minutes.
9. Add 7.5 mL of 7.5 M ammonium acetate solution and mix the contents by gently mixing
the tube for about 10 seconds (the pH should be about 7.8 and thus not require any
adjustment). Maintain on ice for about 10 minutes to allow most of the protein, high
molecular weight RNA, and chromosomal DNA to precipitate.
10. Centrifuge tube for 10 minutes at 15,000 rpm, 4°C. Remove the clear supernate and
transfer to a clean tube Add an equal volume of phenol extraction solution (see Table
2-4) and gently mix. Separate phases by centrifuging at 5,000 rpm for 5 minutes. Remove
the lower phenol solution and repeat the extraction using a solution of chloroform and
isoamyl alcohol (1:0.04); repeat phase separation centrifugation.
11. Transfer upper DNA solution to a clean tube and add 0.6 volumes of isopropanol and
incubate at - 70°C for 10 minutes.
12 Centrifuge at 15,000 rpm for 10 minutes, then remove supernate. Wash the DNA pellet
by adding 5 mL of 100% EtOH, gently mixing, followed by incubation at -70°C for 10
minutes.
13. Centrifuge at 15,000 rpm for 10 minutes, then remove supernate and excess EtOH using
a cotton swab. Add 8 mL of DNA dialysis buffer [10 mM Tris-HCl (pH 8.0), 10 mM
NaCl, 1 mM EDTA] and gently resuspend the DNA pellet.
14. Measure volume and add 1.05 g of CsCl/mL. Gently mix until all of the CsCl is in
solution; then add 100 u.L of ethidium-bromide solution [50 mg/mL in dimethylsulfoxide
(DMSO)]. Avoid exposure to light.
15. Prespin the mixture at 19,000 rpm, 15°C, for 30 minutes to remove large RNAs (nbo-
somes) and denatured proteins.
16. Transfer DNA solution to either a Beckman quick-seal tube or clear Oak Ridge tube.
Centrifuge in a fixed-angle rotor according to the specifications necessary to establish a
CsCl gradient. Generally, quick-seal tubes can be spun at speeds in the range of 60,000
rpm for 24 hours, while Oak Ridge tubes can be spun at about 38,000 rpm for 48 hours.
17. After centrifugation, remove all DNA bands from the top and not the side (DNA removal
from the side will mix the small RNAs into the DNA band). First, remove the upper
bacterial chromosomal band if present. Then carefully remove the lower supercoil plasmid
or cosmid DNA band and add to a new centrifuge tube.
18. The supercoil plasmid or cosmid DNA band should be in a volume of about 2 mL. Fill
the remaining portion of the centrifuge tube with 1.58 Q CsCl (1.05 g CsCl/mL of DNA
dialysis buffer) and add an additional 25 (j,L of ethidium-bromide solution (50 mg/mL in
DMSO). Gently mix and centnfuge again as described in step 16.
DNA SEQUENCING: STRATEGY AND METHODS 27

Table 2-2 Continued

19. Remove supercoiled plasmid or cosmid DNA band from the top and transfer to a 15 mL
conical tube (Falcon). Extract ethidium-bromide by adding an equal volume of isopro-
panol (which has been saturated with a NaCl solution to ensure phase separation).
Isopropanol extract 2 to 3 times or until no "pink" ethidium-bromide color remains.
20. Transfer DNA solution into a dialysis tube and dialyze against 1 L of DNA dialysis buffer;
3 changes.
21. After dialysis, determine DNA concentration by measuring its O.D.26o (1 O.D. = 50
M-g/mL).

rotor such as the Beckman type Ti 70.1. Although it is tempting to reduce

or eliminate the time needed for these CsCl bandings by using only one
banding or by using vertical angle rotors, we have found that shortcuts can
be the cause of many problems encountered in DNA sequencing reactions.
We strictly avoid the use of vertical rotors for the preparation of plasmid
and cosmid DNAs which are to be 5'-end-labeled. The vertical rotors do
not offer an effective purification of the plasmid or cosmid DNAs from
contaminating small RNA molecules (in the range of 10 to 50 nucleotides).
It is essential that the concentration of these RNAs be reduced as much
as possible because their 5'-OH groups effectively compete with the tar-
geted DNA during the 32P-end-labeling reaction. In fact, because of their
small size and large numbers, these small RNA molecules are the preferred
substrate of the end-labeling reaction. Removal of these small RNAs is
very difficult, and the use of ribonuclease (RNase) A only results in in-
creasing their concentration because the products of RNase A treatment
are small ribonucleotides. These RNAs are generally not detected on ac-
rylamide gel electrophoreses because they are rapidly eluted.

5'-End-Labeling with [32P^y]-ATP

It is well known that restriction enzyme sites that yield protruding 5'-OH
groups, end-label much more efficiently than flush-ended 5'-OH groups,
and that recessed 5'-OH groups are the most difficult to 5'-end-label.
However, we have found that if the DNA is free of these small RNA
molecules, all restriction enzyme sites can be end-labeled regardless of the
location of its 5'-OH group (Slightom et al., 1980; Slightom et al., 1987;
Bosma et al., 1988; Siemieniak et al., 1991). We have 5'-end-labeled sites
such as PstI, Apal, Sad, Kpnl, and others, and have obtained 32P-specific
activities similar to that obtained for sites such as BamHI, Bglll, EcoRl,
and others that have protruding 5'-OH groups. Having the ability to end-
label most restriction enzyme sites greatly increases the usefulness of the
chemical sequencing procedure.
End-labeling reactions are generally done using about 10 (xg of plasmid,
20 (Jig of cosmid, or 30 (jig of A DNAs. These DNAs are first subjected to
digestion by the restriction enzyme of choice. Digestions are done with
five to tenfold enzyme excess and completeness of the restriction enzyme
28 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Table 2-3 5'-End-l_abeling of DNA Fragments

1. Obtain restriction enzyme site map of insert DNA to identify which restriction enzymes
can be sequentially used to obtain a single 32P-end-labeled DNA fragment.
2. Digest 10 u,g of plasmid, 20 ji,g of cosmid, or 30 |xg of A DNA with the first restriction
enzyme. Digest in a total volume of about 50 u,L using salts recommended by the enzyme
vendor. Digest the DNA using between five- to ten-fold excess enzyme for 1 hour. Digest
at the optimal temperature recommended for the enzyme.
3. Check completeness of digest by removing 250 to 500 ng of DNA from the digestion mix
and electrophoresing in a mini-agarose gel. During this time, allow the sample to continue
to digest.
4. After the DNA sample is completely digested with the first enzyme, add one-tenth volume
1 M Tris-HCl (pH 8.4) to raise the pH to between 8.0 and 8.4. Add 5 to 10 U of calf
intestinal alkaline phosphatase (CIP) (from Boehringer-Mannheim, see Table 2-4 for
preparation of CIP).
5. Incubate reaction at 55°C for 30 minutes, then remove all enzymes (CIP and restriction
enzyme) by 2 extractions with phenol:chloroform:isoamyl alcohol (1:1:0.04) followed by
1 extraction with chloroform:isoamyl alcohol (1:0.04).
6. Add one-tenth volume of 3 M NaOAc and 2.5 volumes of EtOH and incubate at - 70°C
for 10 minutes Spin in an Eppendorf centrifuge for 7 to 10 minutes; however, if A DNA
is being end-labeled, this spin should be reduced to about 1 minute so as not to pack the
DNA pellet. Wash the DNA pellet once with 100% EtOH and repeat the Eppendorf
spin.
7. Remove excess EtOH by briefly drying in a Savant Speedvac; be sure not to dry A samples
completely as they will become difficult to resuspend. Resuspend DNA in 15 u.L of H2O
and 15 u.L of denaturation buffer (see Table 2-4). Use 25 u,L of each liquid for A samples.
Carefully check samples to ensure that all of the DNA pellet has been resuspended.
8. Heat samples to 65°C for 5 minutes, cool on ice, and then add 4 ^.L polynucleotide kinase
buffer (see Table 2-4). Mix the sample briefly, and then add 100 |xCi of [32P-y]-ATP
followed by the addition of 10 U of polynucleotide kinase (US Biochemical). Briefly mix
the sample, spin down the solution in an Eppendorf centrifuge, and incubate at 37°C for
30 minutes.
9. Heat-kill the polynucleotide kinase by incubating at 65°C for 5 minutes then add one-
tenth volume 3 M NaOAc and 2.5 volumes of EtOH and precipitate the labeled DNA
away from most of the unreacted [32P-y]-ATP.
10. Dry the DNA pellet in a Savant Speedvac (do not completely dry A DNA samples) and
then resuspend DNA pellet in 30 jxL H2O followed by the addition of the appropriate
second restriction enzyme buffer Check to ensure that the DNA pellet has been resus-
pended.
11. Add the second restriction enzyme (five- to ten-fold excess) and digest at the appropriate
temperature for 1 hour Following digestion, remove the restriction enzyme by 1 ex-
traction with the phenol extraction solution (see Table 2-4), and 1 extraction with chlo-
roform:isoamyl alcohol (1:0.04) Removal of the restriction enzyme is done to ensure
that it [or added bovine serum albumin (BSA)] does not interfere with acrylamide gel
separation of the labeled DNA fragments.
12. Add 3 to 5 u,L of sucrose loading dye (see Table 2-4) and load DNA sample onto a
glycerol acrylamide gel. Electrophorese DNA sample to allow separation of the expected
DNA fragments.
13. Locate 32P-labeled DNA fragments by exposing to film (see text). The size of DNA
fragments can be determined by using a 32P-end-labeled-size standard mixture such as
the 1 kb ladder available from BRL (treat with CIP before 32P-end-labehng, see steps 4
and 5). Bands can be located by the attachment of side strips marked with chemilumi-
nescent paint, followed by a short exposure to film. The exposed film provides an effective
guide for localizing the end-labeled fragments.
14. Remove the 32P-end-labeled DNA fragments and place in a 1.5 mL Eppendorf tube.
Crush the acrylamide gel matrix using a pointed spatula and then wash the acrylamide
DNA SEQUENCING: STRATEGY AND METHODS 29

Table 2-3 Continued

fragments remaining on the spatula into the tube using 400 \t,L of elution solution (see
Table 2-4). Incubate the tube at 37°C overnight with shaking in an Eppendorf vortex.
15. Separate the DNA-containing elution solution from the acrylamide gel matrix by passing
through a 1 mL plastic Eppendorf pipet tip that has been fitted with a small plug of
siliconized glass wool. A 10 x 70 mm borosilicated glass culture tube is used to collect
the eluted sample.
16. The elution process can be enhanced by spinning the tip and tube at 3,000 rpm. After
the spin, add 100 nL more elution solution to the pipet tip and repeat the spin.
17. Transfer the sample to a fresh 1.5 mL Eppendorf tube, add 1 mL 100% EtOH, mix, and
incubate at -70°C for 10 minutes. Pellet the DNA by spinning for 7 to 10 minutes in
an Eppendorf centrifuge. Remove the supernatant and resuspend in 50 p,L H2O. Add
500 n.L EtOH, incubate at -70°C for 10 minutes.
18. Centrifuge sample for 5 minutes in an Eppendorf centrifuge, dry DNA pellet in a Savant
Speedvac, and then resuspend DNA in 30 JJLL H2O. Some of the excess acrylamide that
remains can be removed by freezing the sample at -70°C, then spinning 5 minutes in
an Eppendorf centrifuge, followed by transfer to a fresh 1.5 mL Eppendorf tube.
19. Add 3 JJL.L sonicated carrier DNA, 2.5 u.g/u.L (final concentration of 0.25 u.g/M'L) (see
Table 2-4). The 32P-end-labeled fragment is now ready to be aliquoted among the 4 base-
specific reactions used for the chemical sequencing method (Table 2-6).

digest can be checked by electrophoresis on a small agarose gel. The 5'-

terminal phosphates are removed by the addition of 5 to 10 U of calf
intestinal alkaline phosphatase (CIP) (from Boehringer-Mannheim) fol-
lowed by replacement of this phosphate with a radioactive (32P) phosphate.
The 5'-32P-end-labeling procedure is outlined in Table 2-3.
The 32P-end-labeled DNA fragments can be located by placing paper
strips, which have been marked with a unique pattern using chemilumi-
nescent paint, on both sides of the gel. Exposure of these strips to light,
followed by a short exposure to X-ray film, will leave the unique pattern
as a guide on the exposed film. Using this film guide, along with a 32P-
end-labeled internal size standard (we use the one kb ladder obtained from
BRL), the location of 32P-end-labeled fragments can be determined and
removed by cutting around them with a razor blade. The 32P-end-labeled
DNA fragments are removed from the acrylamide by first placing them
into a 1.5 mL Eppendorf tube followed by using a pointed spatula to mash
the gel matrix and the addition of 400 |j,L of elution solution (Table 2-4).
Removal of the 32P-end-labeled DNA fragment from the acrylamide gel
matrix is described in Table 2-3.

Base-Specific Chemical Cleavage Reactions

The chemicals needed for base-specific reactions can be obtained from the
vendors recommended by Maxam and Gilbert (1980) or as a DNA chemical
sequencing kit from Sigma (SEQ-1). If a small number of chemical se-
quencing reactions are planned, use of a chemical sequencing kit is rec-
ommended, leaving the chemical quality control problems to the kit's ven-
dor. If a large number of reactions are planned, stock chemicals can be
30 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Table 2-4 Solutions Needed for Chemical and Enzymatic Sequencing

General Solutions
Preparation of phenol extraction solution
Melt bottle of Mallinckrodt AR phenol at 65°C.
Saturate with DNA dialysis buffer; about 250 mL/500 g phenol.
Add 2 M Tris-HCl (pH 9.5) until pH of phenol solution is between 7 and 8 (about 65 mL).
Add 0.2% 8-OH quinoline.
Add an equal volume of ch)oroform:isoamyl alcohol (1 :0.04) ; yields about 850 mL of phenol
extraction solution.
Glycerol acrylamide gel (7.5% acrylamide, 20% glycerol; 100 mL)
40 mL — 50% solution of glycerol (use 50% for ease of pouring)
10 mL— 10X Tris-borate-EDTA (TBE) electrophoresis buffer
15 mL — acrylamide stock solution (48% acrylamide, 2% bis)
35 mL— H2O
10X TBE electrophoresis buffer (4 L)
Tris 484.4 g
Boric acid 205.4 g (crystals)
EDTA 14.9 g
Dissolve in a final volume of H2O, check pH, should be about 8.3.
Sucrose loading dye (100 mL)
Sucrose 60%
10X TBE 0.5 mL
Dyes 0.1% xylene cyanol, bromophenol blue, and orange G
Solutions Used for 5'-End Labeling
Preparation of CIP alkaline phosphatase:
Obtain CIP from Boehringer Mannheim, (lyophilized) 4,000 U
Resuspend in 1 mL glycerol buffer (10 mM Tris-HCl, pH 8.0; 1 mM MgCl2; 40% glycerol).
Carefully mix. Aliquot 100 (jiL into 1.5 mL Eppendorf tubes and store at -20°C.
Denaturation buffer
20 mM Tris-HCl pH 9.5
0.1 mM EDTA
1.0 mM spermidine
Polynucleotide kinase buffer
0.5 M Tris-HCl pH 9.5
0.1 M MgCl2
50 mM DTT
50% glycerol
Elution solution
0.5 M NH4OAc
10 mM MgOAc2
1.0 mM EDTA
0.1% SDS
Chemical Sequencing Solutions
DMS reaction buffer
50 mM sodium cacodylate (pH 8.0) Sigma # C-0250
1.0 mM EDTA
DMS stop buffer
1.5 M NaOAc
l.OM /3-mercaptoethanol
100 (j,g tRNA (carrier)
DNA SEQUENCING: STRATEGY AND METHODS 31

Table 2-4 Continued

Hydrazine stop buffer
0.3 M NaOAc
0.1 mM EDTA
25 (xg tRNA (carrier)
Preparation of carrier DNA
Obtain solution of calf thymus DNA, concentration of about 2.5 mg/mL. Sonicate exten-
sively using the large sonication tip. Average size of fragments should be about 1,000 bp
(analyze on acrylamide gel). Phenol extract and EtOH precipitate. Resuspend in DNA
dialysis buffer and measure concentration (O.D.260).
DNA sequencing gel loading solution
80% Formamide
lOmM NaOH
l.OmM EDTA
0.1% Xylene cyanol
Enzymatic Sequencing Solutions
Termination mixes (nucleotide concentrations, as supplied in kits)

G mix A mix T mix C mix

dGTP 80 M.M 80 u.M 80 MM 80 M.M
dATP 80 M-M 80 (jiM 80 M.M 80u.M
dCTP 80 M-M 80 M-M 80 M.M 80u.M
dTTP 80 M-M 80 MM 80 uM 80 U.M
ddGTP 8 M-M ddATP 8 M-M ddTTP 8 M.M ddCTP 8 M-M
NaCl 50 M-M 50 M-M 50 uM 50M-M
Nucleotide chase mix (for extensions, not part of kit)
dGTP 200 M-M
dATP 125 M-M
dCTP 50 M-M
dTTP 200 fiM
NaCl 50 M-M
Modified termination mix for sequence extension
G mix: 2.5 M-L G termination mix, 2.5 M-L chase mix
A mix: 3.0 M-L A termination mix, 3.0 (xL chase mix
T mix: 3.0 o.L T termination mix, 3.0 M-L chase mix
C mix: 2.5 uL C termination mix, 2.5 M-L chase mix

obtained; however, purity of these chemicals should not be taken for granted.
Table 2-5 lists some of these chemicals used and the vendor from which
they can be purchased. Table 2-5 also lists the conditions that we use for
storage of these chemicals. Before using chemicals such as DMS, hydrazine,
and formic acid, one should carefully read the instructions for safe handling,
storage, and disposal. Both DMS and hydrazine are poisonous and volatile;
and hydrazine is flammable. Formic acid is caustic and can cause severe
burns to exposed skin. All of these chemicals should be dispensed in a
fume hood, and plastic or latex gloves should be worn. Discard waste (even
supernatants from ethanol precipitations) and pipette tips into solutions of
inactivating reagents; 5 M sodium hydroxide for DMS; and 3 M ferric
chloride for hydrazine.
32 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Table 2-5 Chemical Vendors and Storage

CHEMICAL VENDOR STORAGE
Acrylamide BDH Chemicals Room temperature
Bis-acrylamide BRL 4°C
Dimethyl sulfate Aldrich (99%) 25 mL brown bottles, 4°C
Formic acid Aldrich (97%) 4°C
Hydrazine Eastman Kodak 1.5 mL tubes, -80°C
(#902)
Piperidine Fisher Scientific 4°C
Sodium cacodylate Sigma (#C-0250) Room temperature
SDS BDH Chemicals Room Temperature
y-methacryloxy- Sigma (M-6514) 4°C
propyltrimethoxy-
silane
Repelcote Atomergic Chemical Room temperature
Corporation

The quality of DMS and formic acid supplied by the listed vendors
appears adequate, thus, problems with these base-specific reactions are
generally not due to the chemicals. We have found the storage of small
aliquots (in 25 mL brown bottles) of DMS at 4°C to be helpful, and that
formic acid can also be stored for extended periods of time (several years)
at 4°C without noticeable changes in reaction specificities. Such stability
for hydrazine is not the case, however, as its base-specific reactivity is very
inconsistent from bottle to bottle—even if the bottles have the same chem-
ical lot number. This is probably because hydrazine is more susceptible to
variations in storage and shipping conditions. We have found it necessary
to test three to four different bottles of hydrazine before finding one that
yields clean base-specific reactions for both C and C + T reactions. Even
a sample of hydrazine of adequate purity cannot be stored at 4°C for long
without a noticeable loss of reaction specificity. For prolonged storage of
the hydrazine, it is aliquoted into 1.5 mL Eppendorf tubes (0.8 mL/tube)
and stored at -80°C, and once an aliquot is opened and partially used it
is discarded (using the approved procedure described above). We have
been able to store hydrazine aliquots at - 80°C for over 4 years with little
loss of base-specific reactivity. Additional information concerning storage,
analysis, and safe use of these base-specific chemicals has been described
by Maxam and Gilbert (1980).
All base-specific reactions are conveniently done in 1.5 mL Eppendorf
conical polypropylene tubes and, for added convenience, different color
tubes can be used to identify each of the four base-specific reactions.
Solutions used in various sequencing reaction steps are listed in Table
2-4. The DNA sequencing chemistry begins with the addition of the base-
specific chemical, and the extent of the reaction depends on the concen-
tration of these chemicals; DMS (G-specific), formic acid (A + G-specific),
DNA SEQUENCING: STRATEGY AND METHODS 33

hydrazine (C + T-specific), and hydrazine in 1 M NaCl (C-specific). These

reactions are also dependent on the concentration of each base represented
in the 32P-end-labeled DNA fragment. That is, on a weight basis, the
concentration of any particular base in a shorter fragment is higher than
that for a larger fragment, thus short DNA fragments will appear to react
slower than long DNA fragments. This is important to know, because the
chemical sequencing method will be done on fragments that range in size
from 100 to 1,500 bp. Reaction times for shorter DNA fragments (in the
range of 100 to 700 bp) and for larger DNA fragments (in the range of
700 to 1,500 bp) are included in Table 2-6. Because the reaction time listed
for the DMS reaction of larger DNA fragments is quite short, the number
of G-specific reactions done at one time should be kept small (two to four).
If the reaction is somewhat slow (measured in minutes), the number of
samples in each reaction set can be increased; it is not difficult to handle
six to eight A-, C + T-, or C-specific reactions at one time.
After performing the base-specific reactions listed in Table 2-6, the re-
acted DNA samples either contain methylated Gs, missing purine bases,
or open pyrimidine bases. Upon treatment of these samples with 1 M
piperidine (Table 2-6), the 7-methyl G residues are opened and all of these
ring-opened bases are displaced from the sugar backbone followed by /3-
elimination of the phosphates from the empty sugars. The piperidine re-
action is accomplished using high-temperature incubation (90°C) for 30
minutes; caution should be taken to ensure that the tops remain on the
Eppendorf tubes during this step. Following the piperidine cleavage step,
samples are spun briefly and then placed into a Savant Speedvac evaporator
for removal of the piperidine as a water-piperidine azeotrope. To ensure
complete removal of piperidine, this evaporation process is repeated twice
by adding 100 jxL of H2O for each repeat. These base-specific reacted
DNA samples are quite stable and can be left under vacuum overnight to
ensure complete removal of the piperidine. After the final drying step, the
relative amount of radioactivity in each sample is determined by counting
Cerenkov radiation; the 1.5 mL Eppendorf tubes are placed in scintillation
vials and counted in a scintillation counter using the counting efficiency
setting for tritium. Knowledge of the relative amounts of radioactivity in
each sample is needed to determine the volume of sequencing gel loading
solution (Table 2-4) to add. Samples are resuspended in as small a volume
as possible to maximize the amount of radioactivity loaded onto the se-
quencing gel. The volume of loading solution added is influenced by several
factors: the number of times each sample must be loaded; how consistent
the level of radioactivity is among the four base-specific reactions; and the
level of radioactivity among the different DNA fragments to be loaded on
the same sequencing gel. The final step in the sequencing reaction involves
heating the sample to 90°C for three minutes and loading them onto the
sequencing gel (described below) for electrophoretic separation of the re-
acted fragments.
34 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Table 2-6 Reaction Conditions for the Chemical DNA Sequencing Method
G G + A T + C C
200 u,L DMS buffer 18 |xL H2O 18 R.L H2O 15 |xL 5 0 M NaCl
1 (xL carrier DNA 1 (xL carrier DNA 1 fxL carrier DNA 1 |xL carrier DNA
5 (xL labeled DNA 9 (xL labeled 9 jxL labeled 5 ixL labeled
DNA DNA DNA
0.6 |xL DMS 20 jxL formic acid 20 u,L hydrazine 20 (xL hydrazine
Mix, spin Mix, spin Mix, spin Mix, spin
Incubate 20°C Incubate 20°C Incubate 20°C Incubate 20°C
30 sec (100-700 2.5 min 2.5 min 2.5 min
bp)
10 sec (700 bp-1.5 1.0 min 1.0 min 1.0 mm
kb)

Add 50 (xL DMS Add 350 (xL hy- Add 350 ixL hy- Add 350 jxL hy-
stop buffer drazine stop drazine stop drazine stop
Add 1 mL EtOH Mix, spin, then Mix, spin, then Mix, spin, then
add 1 mL EtOH add 1 mL EtOH add 1 mL EtOH
Mix, incubate Mix, incubate Mix, incubate Mix, incubate
-70°C lOmin -70°C 10 min -70°C 10 min - 70°C 10 min

Spin, 4°C, 7 mm Spin, 4°C, 7 min Spin, 4°C, 7 min Spin, 4°C, 7 mm
Discard super' into Discard super' Discard super' Discard super'
5 M NaOH into 5 M NaOH into 1 M FeCl3 into 1 M FeCl3

Add 350 u.L 0.3 M Add 350 U.L 0.3 Add 350 ixL 0.3 Add 350 uX 0.3
NaOAc, pH 6.0 M NaOAc, pH M NaOAc, pH M NaOAc, pH
60 6.0 6.0
Add 1 mL EtOH Add 1 mL EtOH Add 1 mL EtOH Add 1 mL EtoH
Mix, incubate Mix, incubate Mix, incubate Mix, incubate
-70°C10min - 70°C 10 min -70°C 10 mm -70°C 10 min

Spin, 4°C, 7 min Spin, 4°C, 7 mm Spin, 4°C, 7 min Spin, 4°C, 7 mm
Discard super' into Discard super' Discard super' Discard super'
5 M NaOH into 5 M NaOH into 1 M FeCl3 into 1 M FeCl3

Add 0.8 mL EtOH Add 0.8 mL Add 0.8 mL Add 0.8 mL

EtOH EtOH EtOH
Mix, incubate Mix, incubate Mix, incubate Mix, incubate
-70°C 10 min -70°C 10 mm -70°C 10 min -70°C 10 mm

Spin, 4°C, 7 mm Spin, 4°C, 7 mm Spin, 4°C, 7 mm Spin, 4°C, 7 min

Discard super' into Discard super' Discard super' Discard super'
5 M NaOH into 5 M NaOH into 1 M FeCl3 into 1 M FeCl3

Speedvac 5 min Speedvac 5 min Speedvac 5 mm Speedvac 5 min

Add 20 u.L H2O Add 20 |xL H2O Add 20 (xL H2O Add 20 u,L H2O
Speedvac dry Speedvac dry Speedvac dry Speedvac dry

Add 100 |xL 1 M Add 100 u.L 1 M Add 100 u.L 1 M Add 100 (xL 1 M
piperidine, mix piperidine, mix piperidine, mix piperidine, mix
Incubate 90°C 30 Incubate 90°C 30 Incubate 90°C 30 Incubate 90°C 30
min min mm min
Spin, Speedvac dry Spin, Speedvac Spin, Speedvac Spin, Speedvac
dry dry dry
DNA SEQUENCING: STRATEGY AND METHODS 35

Oligomer-Directed Enzymatic Sequencing of Cosmid and A Inserts

Oligomer-directed enzymatic sequencing of cosmid and A inserts provides
an opportunity to efficiently sequence large DNA molecules in a directed
manner. This method is much more efficient than using the chemical se-
quencing approach described above because it uses much less DNA, and
it is not subject to limitations due to the lack of or over-availability of
restriction enzyme sites. However, as we have discussed above and outlined
in Table 2-1, the combined use of both sequencing methods does provide
a strategy for efficient sequencing of large DNA inserts. We have found
both TV and Tag polymerase to be effective for enzymatic sequencing;
however, because most of our experience has been with T7 polymerase,
its use is described in the procedure outlined in Table 2-7. The procedure
described in Table 2-7 has been found to be simple, yet the most consistent
to yield excellent results with either plasmid, cosmid, or A. DNAs. However,
for this procedure to be consistent, we recommend that the DNA templates
be as pure as possible (see above and Table 2-3). We have recently used
this procedure to obtain over 80 kb of nucleotide sequence information
for a set of cosmid clones and we have experienced a very high rate of
success for primer-directed sequencing: in excess of 85% of the reactions
yield meaningful results (D. R. Siemieniak, unpublished results). Addi-
tional information concerning direct sequencing of cosmid DNA has been
described by Siemieniak et al. (1991). This high rate of success of our
cosmid sequencing procedure is directly attributed to the use of pure cosmid
DNAs and the accuracy of the sequence used for the design of the next
walking oligomer primer.
Enzymatic Sequencing Reactions
The procedure outlined in Table 2-7 can be used for the enzymatic se-
quencing of plasmid, cosmid, or A DNAs; these steps are essentially those
described by the T7 vendor (US Biochemical) for sequencing plasmid DNAs.
We have found that the commercially available T7 sequencing kits are
adequate for obtaining sequence reads in the range of 600 bp using double-
stranded template. However, to achieve even longer sequencing reads, in
the range of 800 bp or more, we have found it necessary to use a modified
termination mix, referred to as the extension mix. The termination mix
supplied in the T7 kit contains a dideoxy:deoxy nucleotide ratio of 1:10;
while a ratio of 1:30 is used in the extension mix (Table 2-4). Using the
procedure outlined in Table 2-7 and the gel system described below, we
have routinely obtained nucleotide sequence reads which extend 700 to
800 bp beyond the oligomer primer.
The T7 polymerase reaction is stopped by drying down in a Savant
Speedvac, followed by resuspending in 12 (xL of sequencing gel loading
dye (Table 2-4). We have found the loading solution supplied in the T7
kits to be inadequate for obtaining complete strand denaturation when
used as described in the protocol book, possibly because the final concen-
36 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Table 2-7 Procedures for Double-Stranded Sequencing

1. Use 2 (xg plasmid DNA, 3 jxg cosmid DNA, or 4 (xg A DNA, final concentration of 1 jxg/
u,L, adjusted concentration after denaturation.
2. DNA mix:
2 to 4 U.L of DNA (step 1)
2 (j,L 5 x reaction buffer (200 mM Tris-HCl, pH 7.5, 100 mM MgCl2, 200 mM NaCl)
1 (JL! oligomer primer (50 ng)
x H2O (to a total of 10 fxL)
10 (jiL total volume.
3. Denature DNA by placing tube in boiling H2O for 5 minutes, cool on ice for 5 minutes,
then add to DNA mix (step 2).*
4. Aliquot termination mixes (see Table 2-4); for normal sequencing dispense 2.5 ^.L of the
appropriate termination mix into tubes labeled G, A, T, C. Prewarm tubes in a 37°C
heating block. If sequencing reads beyond 600 bp are desired, the nucleotide concentrations
in the termination mixes are altered by the addition of a chase mix, changing the ratio of
dideoxy.deoxy nucleotides to 1:30 (see Table 2-4).
5. Dilute labeling mix (7.5 ^M dGTP, 7.5 |xM dCTP, 7.5 M-M dTTP, 50 mM NaCl) 1:5. This
will allow for reads close to the primer.
Labeling mix 2 |o,L*
H2O 8 (j,L
6. Dilute Sequenase (1:8) in cold enzyme dilution buffer. (10 mM Tns-HCl, pH 7 5, 5 mM
DTT, 0.5 mg/mL BSA)
Enzyme dilution buffer 7 ^L
Sequenase version 2.0 1 p,L
7. Reaction mixture is obtained by adding the following.
DNA mix 10 ^.L (from step 2)
0.1 M DTT 1 (xL
Dilute labeling mix 2 p,L (from step 5)
[32P-a]-dATP (3,000 Ci/mM) 2 u.L
Dilute Sequenase (3.25 U) 2 u,L (from step 6)
Total 17 (xL
Incubate at room temperature for 5 minutes
8. Transfer 3.5 jxL of reaction mixture (step 7) to each termination tube (G, A, T, C) (step
4), mix and incubate at 37°C for 5 minutes.
9. Stop the reaction by drying in a Savant Speedvac (5 minutes) and then add 12 |xL of
loading solution (see Table 2-4). Heat samples to 90°C for 3 minutes, cool on ice, and
then load gel.
'See US Biochemical Sequenase protocols.

tration of formamide is only 40%. Use of the sequencing gel loading so-
lution described in Table 2-4 (which also contains 10 mM NaOH) or even
the supplied loading solution (if added after the drying step) greatly reduced
the number of sequence pauses found in our long gels. This indicates that
many sequencing artifacts, attributed to the polymerase, may indeed be
the result of inefficient denaturation of the duplex. In addition, the use of
the thermostated gels (see below) which allow for a constant gel temper-
ature of 55°C is helpful in reducing the number of gel artifacts that many
researchers attribute to polymerase pauses. Thus, inefficient denaturation
of the duplex can cause serious problems in the accuracy of the sequence
obtained.
DNA SEQUENCING: STRATEGY AND METHODS 37

Sequencing Gel Technology: Reaching for 1 kb

The second goal of our sequencing strategy is to maximize the amount of
information obtained from each sequence reaction. Most commercially
available sequencing gel apparatuses allow for the use of gels in the range
of 40 to 60 cm in length, and a few apparatuses are capable of supporting
100 cm gels (C.B.S. Scientific Inc., and Fotodyne, Inc.). Sequencing gels
in the range of 60 cm can routinely separate DNA fragments in the range
of 600 bp and can be effectively used with our sequencing strategy. How-
ever, to increase efficiency, we routinely use 100 cm gels. The use of long,
thin, thermostated DNA sequencing gels has been previously investigated
(Garoff and Ansorge, 1981; Ansorge and Barker, 1984) and much of our
gel system is derived from the work of these investigators.
Our sequencing gels are 20 cm wide and 100 cm long and can be used
with a wedge spacer system (0.2 mm top and 0.4 or 0.6 mm bottom) (Olsson
et al., 1984) or can be used with a continuous 0.2 mm spacer. The wedge
gel system is referred to as a "bellbottom" gel (Slightom et al., 1987) as
it contains 60 cm long wedge spacers (C.B.S. Scientific Inc.) that start with
a 0.4 mm wedge at the bottom of the gel and has a thickness of 0.2 mm
at the top (60 cm); a 0.2 mm spacer is used for the remaining 40 cm of
the gel. The thermostated plates are made according to the diagram shown
in Figure 2-3. The material and measurements shown in Figure 2-3 are for
the assembly of a thermostated plate measuring 20 cm x 104 cm. Ther-
mostated plates can be made for any size gel, and our experience suggests
that their use will increase the resolving power and accuracy of any se-
quencing gel system. The thermostated plate has been extremely helpful
for sequencing DNAs that have a high G + C composition, regardless of
whether the chemical or enzymatic sequencing method is used. If needed,
gel temperatures can be raised (above 70°C if thermostable glass is used).
We recently found such high-temperature gels to be necessary for sequenc-
ing a transposable element (70% G + C) from Streptomyces fradiae (Sie-
mieniak et al., 1990).
Thermostated plates are treated with Repelcote (an organic based silane
solution available from Atomergic Chemical Corporation) to prevent gel
sticking, while the face plate is treated with y-methacryloxypropyltrimeth-
oxysilane that bonds the gel matrix directly to the glass. Procedures for
treating both the thermostated and face plates and for removing the gel
matrix from the face plate are given in Table 2-8. These large sequencing
gels are filled with acrylamide sequencing solution (4% acrylamide) while
they are lying flat (horizontal) on a bench. This is done by fitting a one-
eighth-inch hole, which has been drilled through the face plate 1 inch from
the bottom and 1 inch from one side of the plate, with a 50 mL syringe
(Fig. 2-4). The gel mold can be filled using hydrostatic pressure, but the
process can be accelerated by fitting the syringe with the plunger and forcing
the solution into the gel mold. After filling the gel mold, the slot-forming
comb is put into place and the bottom spacer is removed, followed by
38 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

1. Plate Glass, 2 Sheets, 41 inches (108 cm)

x 8 34 (22.2 cm)

2. Spacers and baffles

A. 1 spacer 41 inches (104 cm) x 3/4 inch (2 cm)
B. 1 spacer 36 inches (91.5 cm) x 3/4 inch (2 cm)
C. 1 spacer 71/4 inches (18.4 cm) x 3/4 inch (2 cm)
0. 1 spacer 63/4 inches (17.25 cm) x 3/< inch (2 cm)
E. 1 spacer 4 Vz inches (11.5 cm) x 3/4 inch (2 cm)
F. 9 baffles 6 inches (15.25cm) x V4 inch
Figure 2-3 Construction of thermostated gel plate. This diagram shows the design
of the thermostated gel plate used for the sequencing gels described in the text and
shown in Figure 2-4 and 2-5. Measurements are given for the glass parts needed to
construct a gel which is 20 cm wide and 104 cm long. The thermostated plates are
constructed by first cleaning glass surfaces with acetone (in a fume hood) and the
glass parts are glued together using 100% silicone sealer (clear, so as to allow visibility
through the fused glass). Apply plenty of silicone sealer and use pressure to remove
all air bubbles between glass surfaces. Material used for the inlet (bottom) and outlet
(top) connectors can be rigid plastic or metal tubing. Assemble glass parts on a clean,
level surface. After assembly, allow plates to set (under the weight of several face
plates) for at least 24 hours. Test thermostated plates for leaks prior to using.

excreting a small amount of the acrylamide gel solution from the bottom
to ensure that the gel makes uniform contact with the electrophoresis
buffer. The hole in the face plate is plugged with a small amount of grease.
This method of pouring large gels is rapid (these large gels can be poured
in less than a minute) and avoids the introduction of bubbles, which is one
of the major obstacles to using long gels.
After pouring, the gels are allowed to set for about one hour, and then
placed into the gel stand where the necessary tubing from the circulating
bath is attached to the thermostated plate (Fig. 2-5). The connection be-
tween the water bath and gel plates uses sealed fittings that, when discon-
nected, shut off water flow (these connectors are available from Cole-
Parmer). After filling the upper buffer chamber, the combs are removed,
the slots are washed free of any gel debris, and the circulating water is
used to heat the gel to 55°C prior to loading. The 4% bellbottom gel is
DNA SEQUENCING: STRATEGY AND METHODS 39

Figure 2-4 Horizontal pouring of sequencing gel. Sequencing gel face plates con-
tain a one-eighth-inch hole drilled 1 inch from the bottom and 1 inch from one side.
After assembling the gel mold, acrylamide sequencing gel solution is poured into a
50 mL syringe and, following removal of air bubbles, the barrel of the syringe is
fitted into the hole and the acrylamide solution is allowed to fill the gel mold. After
the solution clears the top region of the syringe, the plunger can be fitted into the
syringe and used to force the remaining solution into the gel mold. After the gel
mold is filled, the slot former is put in place, the bottom spacer is removed, and
the hole is plugged with a small amount of grease.

preelectrophoresed for about one hour at 2,000 volts, but preelectropho-

resis of the 4% straight gel is not necessary. Sequenced DNA samples
which have been resuspended in the loading solution (Table 2-4) are heated
to 90°C for three minutes, placed on ice, and then loaded onto the gels.
Both the wedge and straight gels are designed for maximum nucleotide
sequencing reads, and they need to be loaded only once. Maximum nu-
cleotide sequencing reads for shorter gels generally require the inconven-
ience of multiple sample loadings. The amount of sequence sample loaded
depends on the size of the slots used; for example, loading volumes of
about 0.5 to 1.0 (xL can be used with a 5 mm slot; while only about 0.2
|xL should be used in a 3 mm slot. Loading of these thin gels is done using
plastic micropipet tips that are available from Drummond Scientific Co.
A typical 4% acrylamide-7 M urea bellbottom gel is electrophoresed
until the xylene cyanol dye is about 70 cm from the top, while the 4%
acrylamide-7 M urea straight gel is electrophoresed for a total of 45,000
to 50,000 volt-hours (no dye markers are visible). After electrophoresis,
40 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Table 2-8 DNA Sequencing Gels: Solutions and Plate Treatments

DNA Sequencing Gel Solutions
Acrylamide sequencing stock solutions (100 mL)

Components Gel percentage (4%)

Urea 50 g
Bis-acrylamide 0.2 g
Acrylamide 3.8 g
10X TBE 12.5 mL

y-methacryloxypropyltrimethoxysilane solution

25 mL EtOH (100%)
0.75 mL 10% acetic acid
75 H.L y-methacryloxypropyltrimethoxysilane

Silane solution
2% dichlorodimethylsilane in chloroform or obtain Repelcote from Atomergic Chemi-
cals Corporation
Treatment of Plates
Thermostated plate
Wash plate with soap and water
Rinse with water until all the soap residue has been removed
Wash plate with 100% EtOH, use lint-free wipes (Kimwipes)
Place thermostated plate in a hood
Apply Repelcote and allow plate surface to dry
Buff plate surface with Kimwipes

Treatment of face plate

Wash plate with soap and water
Rinse with water until all the soap residue has been removed
Wash plate with 100% EtOH using Kimwipes
Apply y-methacryloxypropyltrimethoxysilane solution with Kimwipe
Allow plate to dry
Remove excess y-methacryloxypropyltrimethoxysilane by rinsing with 100% EtOH and
buffing with a Kimwipe
Repeat rinsing with 100% EtOH—generally, this buffing is repeated 2 to 3 times
If too much y-methacryloxypropyltrimethoxysilane remains on the face plate it could
cause the acrylamide gel to also adhere to the thermostated plate (even if it has been
silane treated). If this happens it could cause the acrylamide gel to rip upon plate sepa-
ration.

Removal of gel matrix from face plate

Soak free plates in a solution of 2% KOH for at least 2 hours, or until needed

gels are removed frorrt the stand, separated, and the DNA fixed and the
urea removed by soaking the face plate in a 10% acetic acid bath for ten
minutes. Excess acetic acid is removed by an extensive water rinse. Then
the gels are allowed to dry at room temperature, or, if available, in a warm
room (37°C). If a warm room is used, gels dry in about three hours, and
once dry, the gel matrix has a thickness of about 10 /A which increases the
DMA SEQUENCING: STRATEGY AND METHODS 41

Figure 2-5 Use of thermo-

stated DNA sequencing gel sys-
tem including safety cabinets.
The DNA gel strands and safety
cabinets were obtained from
Fotodyne, Inc. and fitted with
tubing that allows heated water
to circulate from the water bath
(baths, shown on the lower left,
contain about 6 L) into the ther-
mostated plate. Electrophoresis
power supplies are located on
top of the safety cabinets and
contain ground-default safety
switches. Safety cabinets also
contain shutoff switches that
disconnect the power from the
gel when open.

efficiency of the film exposure and band resolution. Gels are exposed to
X-ray roll film (Eastman Kodak XAR351 available from Fotodyne, Inc.)
by forming a sandwich with the film being placed between the gel plate
and a clean gel plate followed by wrapping with aluminum foil and enclosing
within a large (custom made) black plastic bag. Films are generally exposed
at room temperature, except for chemical sequencing reactions, which are
exposed at -70°C with the aid of an intensifying screen-(Quanta III, 35
cm x 1 m, available from DuPont). Nucleotide sequence reads from a
typical 4% bellbottom gel generally start about 50 bp from the end-labeled
restriction enzyme site (chemical method) or oligomer primer (enzymatic
method) and extend out to about 400 bp; while a typical 4% straight gel
can be read starting from about 350 bp and extend out to better than 800
bp.

CONCLUDING REMARKS

Recent advances in nucleotide sequencing technologies have lead us to

believe that alternative strategies should be investigated for obtaining nu-
cleotide sequence information. We believe that these new strategies will
involve the direct sequencing of larger and larger DNA molecules, and
that this information will be collected in a directional manner that will
42 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

allow for rapid analysis. Toward these goals we have been developing
strategies and methodologies for directly sequencing large DNA molecules.
The results of our test suggest that a directional sequence strategy can be
improved to a level where it competes favorably with the random sequenc-
ing strategy. We feel that the major contribution of this work is to show
that large DNA molecules can be conveniently sequenced directly and
directionally. We realize that additional technological improvements are
needed before the direct sequencing procedure and strategies described
here will be widely applied. Additional improvements such as automated
sequencing reactions, the use of nonradioactive labels, the use of automatic
-film readers, and reduction in the price of oligomer primers, would greatly
benefit our strategy. The use of short primers has recently been suggested
by Studier (1989), and the use of a small library of oligomers would elim-
inate most of the expenses and time involved in oligomer synthesis. The
effectiveness of using a limited library of oligomers (nonamers) for genome
sequencing was simulated using computer-aided techniques by Siemieniak
and Slightom (1990). Thus, because many of these technologies are already
in use, or are in the advanced stages of development, their incorporation
into a strategy for directly sequencing large DNA molecules is possible
and most likely inevitable.

ACKNOWLEDGMENTS

We thank Eric Lai and Lee Hood for the use of cosmid clone H7.1 (de-
scribed in the legend of Fig. 2-2) and Roger Drong for his help in critically
reading this manuscript.

REFERENCES

Ansorge, W., and R. Barker. (1984) System for DNA sequencing with resolution
of up to 600 base pairs. J. Biochem. Biophys. Meth. 9:33-47.
Ansorge, W., B. Sproat, J. Stegemann, C. Schwager, and M. Zenke. (1987) Au-
tomated DNA sequencing: ultrasensitive detection of fluorescent bands dur-
ing electrophoresis. Nucleic Acid Res. 11:4593-4602.
Birnboim, H.D., and J. Doly. (1979) A rapid alkaline extraction procedure for
screening recombinant plasmid DNA. Nucleic Acids Res. 7:1513-1523.
Bosma, P.J., E. Van den Berg, T. Kooistra, D.R. Siemieniak, and J.L. Slightom.
(1988) Human plasminogen activator inhibitor-1 gene: promoter and struc-
tural gene nucleotide sequences. /. Biol. Chem. 263:9129-9141.
Chen, E.Y., and P.H. Seeburg. (1985) Supercoil sequencing: a fast and simple
method for sequencing plasmid DNA. DNA 4:165-170.
Edwards, A., H. Voss, P. Rice, A. Civitello, J. Stegemann, C. Schwager, J.
Zimmerman, H. Erfle, C.T. Caskey, and W. Ansorge. (1990) Automated
DNA sequencing of the human HPRT locus. Genomics 6:593-608.
Elder, J.K., O.K. Green, and E.M. Southern. (1986) Automatic reading of DNA
sequencing gel autoradiographs using a large format digital scanner. Nucleic
Acids Res. 14:417-424.
Frank, R., A. Bosserhoff, C. Boulin, A. Epstein, H. Gausepohl, and K. Ashman.
DNA SEQUENCING: STRATEGY AND METHODS 43

(1988) Automation of DNA sequencing reactions and related techniques: a

workstation for micromanipulation of liquids. BioTechniques 6:1211-1213.
Garoff, H., and W. Ansorge. (1981) Improvements of DNA sequencing gels. Anal.
Biochem. 115:450-457.
Kalisch, B.W., S.A. Krawetz, K.-H. Schenwaelder, and J.H. Vande Sande. (1986)
Covalently linked sequencing primer linkers (splinkers) for sequence analysis
of restriction fragments. Gene 44:263-270.
Kesterson, R.A., S.A. Kerner, and J.W. Pike. (1989) Cosmid sequencing using
Sequenase: a novel approach for determining genomic structure. US Bio-
chemical Comments: 17-18.
Lai, E., P. Concannon, and L. Hood. (1988) Conserved organization of the human
and mouse T-cell receptor p-gene families. Nature 331:543-546.
Levedakou, E.N., U. Langegren, and L. Hood. (1989) A strategy to study gene
polymorphism by direct sequence analysis of cosmid clones and amplified
genomic DNA. BioTechniques 7:438-442.
Lund, T., F.G. Grosveld, and R.A. Flavell. (1982) Isolation of transforming DNA
by cosmid rescue. Proc. Natl. Acad. Sci. USA 79:520-524.
Manfioletti, G., and C. Schneider. (1988) A new and fast method for prepar-
ing high quality lambda DNA suitable for sequencing. Nucleic Acids Res.
16:2873-2884.
Maniatis, T., E. Fritsch, and J. Sambrook. (1982) Molecular Cloning. Cold Spring
Harbor Laboratory, Cold Spring Harbor.
Maxam, A., and W. Gilbert. (1977) A new method for sequencing DNA. Proc.
Natl. Acad. Sci. USA 74:560-564.
Maxam, A., and W. Gilbert. (1980) Sequencing end-labeled DNA with base specific
chemical cleavage. Methods Enzymol. 65:499-560.
Morelle, G. (1989) A plasmid extraction procedure on a miniprep scale. BRL
Focus 11:7-8.
Olssen, A., T. Moks, M. Uhlen, and A.B. Gaal. (1984) Uniformly spaced banding
patterns in DNA sequencing gels by the use of field-strength gradients. /.
Biochem. Biophys. Methods 10:83-90.
Prober, J.M., G.L. Trainor, R.J. Dam, F.W. Hobbs, C.W. Robertson, R.J. Za-
gursky, A.J. Cocuzza, M.A. Jensen, and K. Baumeister. (1987) A system
for rapid DNA sequencing with fluorescent chain-terminating dideoxynu-
cleotides. Science 238:336-341.
Saiki, R.K., D.H. Gelfand, S. Stoffel, S. Scharf, R. Higuchi, G.T. Horn, K.B.
Mullis, and H.A. Erlich. (1988) Primer-directed enzymatic amplification of
DNA with a thermostable DNA polymerase. Science 239:487-491.
Sanger, F., S. Nicklen, and A.R. Coulson. (1977) DNA sequencing with chain-
terminating inhibitors. Proc. Natl. Acad. Sci. USA 74:5463-5467.
Siemieniak, D.R., L.C. Sieu, and J.L. Slightom. (1991) Strategy and methods for
directly sequencing cosmid clones. Anal. Biochem. 192:441-448.
Siemieniak, D.R., and J.L. Slightom. (1990) A library of 3342 useful nonamer
primers for genome sequencing. Gene 96:121-124.
Siemieniak, D.R., J.L. Slightom, and S.-T. Chung. (1990) Nucleotide sequence of
Streptomyces fradiae transposable element Tn4556: a class-II transposon
related to TnJ. Gene 86:1-9.
Slightom, J.L., A.E. Blechl, and O. Smithies. (1980) Human fetal G^- and A-y-
globin genes: complete nucleotide sequences suggest that DNA can be ex-
changed between these duplicated genes. Cell 21:627-638.
44 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Slightom, J.L., T.W. Theisen, B.F. Koop, and M. Goodman. (1987) Orangutan
fetal globin genes: nucleotide sequences reveal multiple gene conversions
during hominid phylogeny. J. Biol. Chem. 262:7472-7483.
Slightom, J.L., and R. Drong. (1988) Procedures for constructing genomic clone
banks. Section A8, pp. 1-42 in Plant Molecular Biology Manual (S.B. Gelvin
and R.A. Schilperoort, eds.). Martinus Nijhoff, Dordecht.
Smith, L.M., J.Z. Sanders, R.J. Kaiser, P. Hughes, C. Dodd, C.R. Connell, C.
Heiner, S.B.H. Kent, and L. Hood. (1986) Fluorescence detection in au-
tomated DNA sequence analysis. Nature 321:674-679.
Sorge, J.A., and L. Blinderman. (1989) ExoMeth sequencing of DNA: eliminating
the need for subcloning and oligonucleotide primers. Proc. Natl. Acad. Sci.
USA 86:9208-9212.
Staden, R. (1982) Automation of the computer handling of gel reading data pro-
duced by the shotgun method of DNA sequencing. Nucleic Acids Res. 10:4731 -
4751.
Studier, F.W. (1989) A strategy for high-volume sequencing of cosmid DNAs:
random and directed priming with a library of oligonucleotides. Proc. Natl.
Acad. Sci. USA 86:6917-6921.
Tabor, S., and C.C. Richardson. (1987) DNA sequence analysis with a modified
bacteriophage T7 DNA polymerase. Proc. Natl. Acad. Sci. USA 84:4767-
4771.
Wilson, R.K., A.S. Yuen, S.M. Clark, C. Spence, P. Arakelian, and L.E. Hood.
(1988) Automation of dideoxynucleotide DNA sequencing reactions using
a robotic workstation. BioTechniques 6:776-787.
Zagursky, R.J., K. Baumeister, N. Lomax, and M.L. Berman. (1985) Rapid and
easy sequencing of large linear double-stranded DNA and supercoiled plas-
mid DNA. Gene Anal. Tech. 2:89-94.
Zimmerman, J., H. Voss, C. Schwager, J. Stegemann, and W. Ansorge. (1988)
Automated Sanger dideoxy sequencing reaction protocol. FEBS Lett. 233:432-
436.
3
The Application of Automated DNA
Sequence Analysis to Phylogenetic
Studies

ROBERT J. FERL, C. JOSEPH NAIRN, JIN-XIONG SHE, EDWARD

WAKELAND, AND ERNESTO ALMIRA

Automated DNA sequence analysis has developed to the point that many
individuals and research organizations have considered or will soon con-
sider investing in this technology. We discuss here some of the factors
involved in automated DNA sequence analysis, particularly as performed
by a service facility and as applied to phylogenetic studies. We also show,
by comparative studies involving ribosomal RNA (rRNA) genes and po-
lymerase chain reaction (PCR)-generated templates, that automated DNA
sequencing saves tremendous time and can be every bit as accurate or
inaccurate as manual methods.
The purpose of this chapter is to discuss the current state of automated
DNA sequencing and its application to the generation of sequence data
for use in phylogenetic studies. We address the subject from a purely end-
user point of view. That is, we are not on the cutting edge of automated
methods development, nor do we have available a wide array of different
automated sequencers and large amounts of comparative performance data.
Rather, as part of an active, readily accessible service facility (University
of Florida Interdisciplinary Center for Biotechnology Research, Gaines-
ville, FL), we have explored the practical aspects of automated DNA
sequence analysis as currently available to the general scientific public and
practiced by the average user.
We begin with a general introduction to the notion of DNA sequencing
by a core service facility. Does such a facility really work? Can service
sequencing, particularly automated service sequencing, provide quality data?
Can any investigator (the hard-core molecular biologists as well as those
with less molecular expertise) have access to DNA sequence information?

45
46 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

The answer to these questions is yes, especially if given the appropriate

funding and administrative atmosphere.
We then move on to a brief description of our sequencing facility and
the particular automated sequencing instrument upon which the facility is
based. After an abbreviated coverage of the methods involved and ex-
amples of actual DNA sequencing projects, we present a comparison of
automated versus manual examples of the kinds of sequencing projects
likely to be undertaken by a user interested in generating sequence infor-
mation from a large number of alleles or homologous genes.

AUTOMATED DNA SEQUENCING

Automated DNA Sequencing as a Service

The Interdisciplinary Center for Biotechnology Research at the University
of Florida, like the biotechnology centers at many universities, participates
in a wide range of activities, and maintains a set of core service labora-
tories.1 These service laboratories consist of high-end analytical equipment
that is purchased, maintained, and staffed from Center funds. Thus, faculty
members have access to equipment that would otherwise be beyond their
individual funding capabilities. The budget from the Center generally cov-
ers all expenses other than the expendables actually used, allowing these
service laboratories to provide state-of-the-art technology at a contract
price that encourages regular use.
The DNA Sequencing Core Laboratory, for example, is set up to provide
DNA sequencing data from any fragment of DNA. The cost (keeping in
mind that only expendables are recovered by this fee) is $50 per reaction,
and from that reaction the average user will receive approximately 300
bases of DNA sequence information.
Typically, a potential user will contact us for general information on
vector and primer choices, DNA cloning and purification procedures, and
other hints as to how to develop their sequencing project. For each user,
we provide a set of protocols, as well as advice on overall strategy for the
project. Then the sequencing and analysis cycle starts, with the user pro-
viding the DNA (and primers if necessary) and the Sequencing Core Lab-
oratory providing the data. The information from each sequencing run is
stored in computer files, and also supplied to the user as hard-copy output
of the raw data together with an interpretation of the data as a sequence
of bases. Users are encouraged to become aware of the nuances of data
interpretation and to make use of all of the raw data, not just the string
of machine-interpreted bases. Thus the user still maintains the primary

'For further information on the DNA Sequencing Core Laboratory at the University of
Florida, feel free to contact Dr. Ernesto Almira or Dr. Robert Ferl. Potential purchasers of
automated sequencers are encouraged to make full use of the demonstrations and literature
available from the manufacturers. The manufacturers will also provide a wealth of specific
technical literature on DNA purification and sequencing reactions.
AUTOMATED DNA SEQUENCE ANALYSIS IN PHYLOGENETIC STUDIES 47

responsibility for sequence accuracy, including final decisions on base call-

ing, number of repetitive runs to obtain the final sequence, and the pro-
duction of sequence from both complementary strands.
By using the Sequencing Core Laboratory, the user is freed from the
purchase of expensive sequencing apparatuses, radiolabeled nucleotides,
X-ray film, and all of the other minutiae associated with manual DNA
sequencing. Instead, funds and effort can be directed to producing the
clones or fragments for sequencing and analyzing the subsequent data.
Indeed, most of the faculty that currently use the Sequencing Core Lab-
oratory are new to DNA sequencing, though faculty already established
in DNA sequencing also use the Laboratory for occasional or particularly
difficult sequencing projects. We have had users come to us for very short-
term projects, such as to simply confirm the sequence of a probe or insert.
We have also had users generate sequences of tens of kilobases (kb) over
longer periods of time.
From our experience as a service sequencing facility, it has become clear
that automated DNA sequencing through such a facility is a viable alter-
native to manual sequencing. The cost, at least within the center concept
as developed at the University of Florida, is minimal, and the automated
nature of the process means a short turnaround time. Thus, many inves-
tigators who would not otherwise have done so are in the business of
generating and analyzing DNA sequence information.

A Brief History of Automated DNA Sequencing and Factors Involved in

Choosing an Automated Sequencer
While a full review of automated DNA sequencing is beyond the needs of
this discussion, familiarity with certain aspects of the technology will pro-
vide a useful perspective for the data comparisons presented in the next
sections. It should also be noted that automated DNA sequencer research
and production is a rapidly evolving field in its own right, and some of the
views presented here (and indeed some of the factors that influence the
decision of which sequencer to purchase) might well change in the near
future.
Automated DNA sequencing can be a misleading phrase. What is really
meant is automated analysis of DNA sequencing reactions. Other than
some modifications that we will discuss below, the enzymatic protocol to
produce the DNA sequence is currently the same, time-honored Sanger
dideoxy terminator system. What has become automated is the electro-
phoretic separation and detection of the products of the Sanger method
(Sanger et al., 1977).
Manual sequencing utilizes fairly large gels and electrophoretic separa-
tion for a specified time, followed by autoradiography. The sequence is
read from the bottom of the gel to the top, because in a given period of
time the smaller fragments will have migrated farther than the larger ones.
The autoradiogram presents a detailed view of the separation achieved at
a certain time-point.
48 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

All automated sequencers also utilize electrophoresis, but in a funda-

mentally different way. Automated sequencers, instead of looking at the
whole gel at one point in time (as in an autoradiogram), look at one point
of the gel over time. They measure the time it takes for a band to traverse
a specified distance in the gel. Smaller bands traverse the gel more rapidly
than larger ones, and arrive at the detection window in a shorter period
of time. Thus the output is very different from the "ladder" observed in
the manual sequencing autoradiogram. Instead, an electrophoretogram is
produced, presenting the detected bands as peaks on the Y axis, and time
of electrophoresis on the X axis. Each peak is then identified as an A, T,
G, or C depending upon the detection system (Fig. 3-1).
The main difference among automated sequencers is the mechanism by
which the bands are detected. One type of machine detects radioactively-
labeled DNA, while the other detects fluorescently-labeled DNA. As a
service facility, we did not consider the machines designed to read radio-
active bands as they elute from the bottom of the gel for two reasons. First,
we did not want to handle the large amounts of radioactive label that would
be involved in large-scale service sequencing. Second, all of these machines
require four reaction tubes and four gel lanes per sequence template. In
machines designed to detect fluorescently-labeled DNA, only one gel lane
is required to analyze the sequence from a template, and the number of
reaction tubes per template can be reduced as well.
The two companies currently marketing automated sequencers that op-
erate by detection of fluorescently-labeled DNA fragments are DuPont
and ABI (Applied Biosystems). Sequencers from both companies use a
laser to scan the gel to detect the DNA fragments. Both use different
fluorescent tags to label the DNA from the different reactions, so that a
single lane is all that is needed to produce sequence from a template. (The
slightly different color of each tag allows the machine to not only detect
when a band is passing the detector, but also whether it represents a
fragment terminated at A, T, G, or C.) Both employ essentially standard
sequencing gel technology. Both also provide a similar type of sequence
hard-copy output. The main difference between the machines is in the
chemistry of the flourescent tags, and this leads to differences in perform-
ance and overall approach.
Except for some very recent developments, the chemistry of the system
from ABI employs a set of four differently labeled primers, one primer
for each of the sequencing reactions, A, T, G, and C (Smith et al., 1986).
After the annealings and extension reactions in the presence of the standard
dideoxy chain terminators, the reactions are stopped, combined, and run
on a single lane of the sequencing gel.
DuPont employs a set of four differently labeled dideoxy terminators
(Prober et al., 1987). After annealing with a standard primer in a single
tube, the extension reaction takes place in the presence of the four labeled
dideoxy chain terminators. After the reaction is stopped, it is loaded in a
single lane on the sequencing gel. In the output shown in Figure 3-1, each
Figure 3-1 Sample output from the DuPont automated DNA sequencer. As DNA fragments pass the scanning
window of the laser detector, the two photomultipliers (PMT1 and PMT2) of the system record the fluorescence
of the band. Each photomultiplier reads a different wavelength, and the actual color of the band (and therefore
the sequence identity of the terminator) is determined by the ratio of signal derived from the two detectors. The
software compares the peak ratios to call the base sequence, which is printed below the peaks. However, since
the output from the photomultiphers is also graphed as a function of time, the raw data can also be evaluated
by the operator.
50 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

band (or peak) is then identified as an A, T, C, or G depending upon its

color, which is determined by the ratio of fluorescence detected by two
photomultiplier tubes (PMT1 and PMT2).
The choice between machines becomes a choice between chemistries and
the limitations and benefits therein. In 1987 we chose the DuPont machine
for the following three reasons. First, we (correctly) anticipated that a large
proportion of our sequencing would involve custom primers and the DuPont
system allows their simple and direct use. The ABI system did have the
capability to use custom primers, but at the cost of the extra steps to add
the fluorescent labels to the primers. From a service facility perspective,
we wanted to minimize our efforts for each step, while still maximizing
our ability to use any primer. (It should be noted that ABI now offers
labeled terminator chemistry on their latest system.) Second, the labeled
dideoxy terminator chemistry allows the extension reactions to be carried
out in a single tube. Thus, by using the DuPont system we could reduce
our manipulations per sequence template by 75%. Again, this was desirable
from a service facility perspective. Third, as we will demonstrate below,
the labeled terminator chemistry removes many types of artifacts that can
impact a sequencing analysis.
These reasons were our own, and firmly based on the availabile data at
the time. Any person or group currently considering the purchase of an
automated sequencer should carefully reexamine the performance char-
acteristics, chemistries, and flexibilities of all machines now available. This
discussion should not be taken as an endorsement of the DuPont automated
sequencer, but merely reflects our views at the time of our purchase.
In order to test the effectiveness of the automated sequencing approach
for phylogenetic analysis, we chose two examples for comparison with
manual methods. First, we chose a relatively large gene, the small-subunit
rRNA gene from Gingko biloba (Nairn et al., 1991). With a panel of 12
primers for conserved regions, we conducted a standard set of manual
sequencing experiments to produce a confident sequence, and we submitted
the same template DNA and primers to the DNA Sequencing Core Lab-
oratory for automated analysis. Second, we compared manual and auto-
mated sequencing of a PCR generated fragment produced from mouse
genomic DNA (exon 3 of MHC Class II A(3 gene).

RESULTS AND DISCUSSION

Automated Sequencing of a Cloned Gene: the 185 rRNA Gene from

Gingko biloba
The small-subunit rRNA gene from Gingko biloba was recovered from a
genomic library and subcloned into pUC 19 for sequencing. The pUC
subclone was used directly as a double-stranded template for both the
manual and automated runs. The coding region of the rRNA gene is ap-
AUTOMATED DNA SEQUENCE ANALYSIS IN PHYLOCENETIC STUDIES 51

proximately 1,800 base pairs (bp). The overall setup of the 12 primers is
shown in Figure 3-2. The average distance between adjacent primers is
approximately 300 bases. The locations for the primers were chosen from
within those regions exhibiting the most conservation over wide taxa. For
the purposes of this comparison, we have not included the price of the
primers, since that cost would be the same for both the automated and
manual methods.
The manual sequencing strategy and resulting read lengths are presented
in Figure 3-2. The method for sequencing was according to the KiloBase
Sequencing System instruction manual as delivered by Bethesda Research
Laboratories, Life Technologies Inc. (Chen and Seeburg, 1985). Initially,
three separate gel runs of up to 14 hours each were required to generate
the primary data set. This resulted in an average read length of over 500
bases of sequence per run. However, as is common with rRNA gene se-
quencing experiments, several areas of severe collapse or compression
(where multiple bands appear in all lanes) were encountered that required
additional reactions and gels for final resolution (see further discussion
below). We have found that each collapse requires a separate approach
for resolution, and no single modified reaction type was sufficient to remove
all collapse regions. In all, four additional gels, with several modified
reaction sets (such as ITP, deaza-7-GTP, and Sequenase) were required
to produce the final sequence.
Thus, after eight sequencing gels and over 60 reactions, a sequence was
generated that contained no ambiguities. Because of the length of the
readable sequence, each area was sequenced on both strands and dupli-
cated on at least one strand. All collapsed regions were fully resolved.
Automated sequencing utilizing the protocols suggested by the manu-
facturer required only 12 sequencing reaction tubes and one five-hour run
on the DuPont automated sequencer. The general results of the automated
sequencing run are shown in Figure 3-3. The distinguishing feature of the
automated analysis is that the effective reading lengths per primer were
about half that derived from manual methods. In fact, in several cases of
nonoverlap, sequence information was therefore obtained from only one
strand. (The consequences of these shorter reads will be considered in the
next section.)

Figure 3-2 Manual sequencing strategy and results for the Gingko biloba 185 rRNA
gene. Note that the read lengths of the individual sequence runs allow a complete
overlap of data from both strands.
52 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Figure 3-3 Automated sequencing strategy and results for the Cmgko biloba 185
rRNA gene. Note that the shorter read lengths resulted in several areas of the gene
that were not sequenced on both strands.

Table 3-1 Ginkgo biloba rRNA Gene Analysis: Automated

versus Manual Sequencing
Sequence length 1,811 base pairs
Base mismatches 10 positions
Percent sequence agreement 99.5%
Percent sequence agreement minus re- 100%
gions representing "overlap failure"
by automated sequencing

Table 3-1 presents the comparison of the data obtained from an ex-
haustive manual sequencing approach and a single automated analysis. In
this table, the data are strictly compared, and any mismatches in the align-
ment are counted as machine errors. From this data set, an error rate for
the DuPont automated sequencer of 0.5% would be calculated. However,
a very key point is that this comparison is based on the data as derived
from the single reads as shown in Figure 3-4. That is, it includes sequence
derived from some areas that were covered only on one strand.
Therein lies the most important lesson in automated sequence analysis.
The same rules for responsible sequencing must apply to both manual and
automated methods. In other words, only those data that are confirmed
with opposite-strand data should be given confidence.
Since the readable runs on the DuPont sequencer failed to extend beyond
300 bases from each primer, the overlap failures created gaps in the se-

Figure 3-4 Location of sequence mismatches by the automated method. Note that
all cases of miscalled bases in the automated analysis occurred in areas of the
sequence that were not covered by complementary data from both strands. N,
number of errors in single-stranded regions; O, overlap failures.
AUTOMATED DNA SEQUENCE ANALYSIS IN PHYLOGENETIC STUDIES 53

quence data where data were available from only one strand (see Fig. 3-4).
In every case of mismatch between the automated and the manual analyses,
the error in the automated run occurred in an area of-failed overlap. Indeed,
when the errors within the overlap gaps are eliminated from the analysis,
the agreement between automated and manual methods is 100% (Table
3-1).
At the time of this comparison, read length on the DuPont sequencer
was a significant limitation. However, as mentioned earlier, protocols and
equipment continue to evolve, and our average run length now exceeds
300 bases. Thus, it is likely that a repeat of the automated analysis would
achieve complete overlap.
There was one remarkable benefit from the automated analysis. The
regions of collapse that plagued the manual sequencing gels were totally
absent from the automated run. Figure 3-5 shows one example in which a
dramatic collapse region exists on the regular sequencing gel. The collapse
was subsequently resolved by modified reaction and gel conditions. How-
ever, it can be clearly seen that the DuPont sequencer read through the
region the first time and without modification of the standard reaction
conditions. Collapse regions in a manual gel are most likely the result of
premature chain termination in the sequencing reaction, or hairpins in the
fragments. In the DuPont system, any premature terminations are not
detected (because the label is on the dideoxy terminators) and thus have
no adverse influence on the read. In addition, because of the sensitive
computer-controlled electrophoresis, precise temperature control is always
realized and fold-back collapses are essentially eliminated.
Thus, using cloned DNA as template, automated sequence analysis is
extremely accurate and much faster than traditional manual methods. The
only drawback in the comparison was that the automated sequencer pro-
duced runs of shorter readable length. If the shorter read length prevents
overlap, then errors can occur because of the inability to cross-check with
the complementary strand. However, as automated protocols continue to
develop, this will become less and less of a problem.

Automated Sequencing of PCR-Generated Templates

Clearly, one of the main approaches in future phylogenetic analysis of

DNA sequences is in the application of PCR technology (Erlich, 1989).
Direct sequencing of PCR-generated products is the fastest way to gain
DNA sequence information from any DNA source that is not already
subcloned into a convenient vector.
In order to evaluate the efficacy of combining automated sequence anal-
ysis and PCR technology, we compared manual and automated sequences
derived from alleles of exon 3 of the mouse MHC Class II Ap gene. PCR
was performed according to the method described by She et al. (1991).
PCR amplification of 1 u-g of genomic DNA produced a DNA fragment
which was subsequently purified by agarose gel electrophoresis. The pur-
54
AUTOMATED DNA SEQUENCE ANALYSIS IN PHYLOCENETIC STUDIES 55

ified DNA fragment and one of the primers was then used to perform an
asymmetrical amplification of one strand. The amplified single-stranded
DNA product was divided and subjected to both manual and automated
sequence analyses using the opposite member of the primer pair to prime
the sequencing reactions.
The results clearly indicated that PCR-derived templates can be directly
sequenced by both manual and automated methods. In a comparison of
the manually derived sequence aligned with the sequence produced from
the automated analysis, over the effective region of the analysis (approx-
imately 220 bases), there is 100% agreement between the two methods.
Again, it must be remembered that confident sequence can only be arrived
at if sequences from both strands are compared and analyzed.

On-Screen Data Analysis

The comparison utilities present in the computer control system of the

automated sequencer allow direct, on-screen comparison of different data
sets (Amorese and Hochberg, 1989). Thus, in a search for sequence po-
lymorphisms, the actual data as well as the derived sequences for several
runs can be loaded into windows on the screen. Presumptive polymorph-
isms can be visually confirmed by this direct comparison method, rather
than by the tedious rereading of autoradiograms. Another example from
the mouse MHC II locus (Fig. 3-6) shows a comparison of the output data
generated from analysis of the sequences of PCR-generated templates for
two different alleles. Visual confirmation of a G/A transition is easy and
obvious.
In a similar fashion, the direct comparison utility of the automated se-
quencer also allows the direct, on-screen alignment of the two comple-
mentary reads for any given area of the sequence. If the sequences derived
from the complementary strands show any mismatches or ambiguities, the
raw sequence data can be reevaluated in paired windows from each se-
quence. From these raw data the operator can make an immediate informed
interpretation based on the quality of the peaks in each sequence, and can
correct one of the sequences or can decide to rerun the reaction(s).

Figure 3-5 Comparison of one dramatic collapse region affecting the manual but
not the automated analysis. (Top) The center four lanes of this sequencing gel dem-
onstrate a common sequencing artifact that is especially prevalent in rRNA gene
sequencing projects. This particular collapse region required several additional at-
tempts to finally resolve the collapse and derive the correct sequence (the comple-
ment to the automated run is shown here). (Bottom) The automated run through this
region, however, shows no artifact. The collapse region is indicated on the output
and the correct sequence is displayed.
56 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Figure 3-6 Demonstration of allelic polymorphism detected by automated analysis

of PCR-generated templates. These two panels represent output from automated
sequencer analysis of PCR templates derived from two distinct alleles of the mouse
MHC II locus. The G/A transition (highlighted in boxes) is clear as is the similarity
of the remainder of the sequence output.

CONCLUSIONS
We hope that this presentation has shown that automated sequencing, even
by a central sequencing facility, is a viable, perhaps even a desirable,
alternative to traditional DNA sequencing methods.
The concept of a central sequencing service facility that services a wide
array of researchers is appropriate for several reasons. First, the automated
AUTOMATED DMA SEQUENCE ANALYSIS IN PHYLOGENETIC STUDIES 57

systems can handle a fairly large throughput of sequencing reactions and

can thereby service a number of users simultaneously. Second, as a group
service facility, maximum use of the sequencer allows the cost of machine
purchase and/or maintenance to be amortized among the users (or, under
the Center concept, to be absorbed by the organization providing the
service). Finally, use of an automated sequencing facility relieves the
individual researcher from a large amount of the effort and other capi-
tal outlays involved in producing sequence information. This allows more
time to be spent in data analysis or production of the templates for
analysis.
We have also shown that automated sequence analysis can be every bit
as accurate as manual sequencing, but only if the same rules of responsible
sequencing are employed. That is, all sequences should be confirmed on
both DNA strands and any ambiguities should be resequenced. Given one
run on one template, the machines are capable of what might seem to be
high error rates, but a similar error-rate observation must also be admitted
for any single manual sequencing run as well. Responsible sequencing
protocol demands complementary data in each case. Our data indicate that
for both cloned genes and PCR-generated templates, automated sequen-
cers equal the accuracy of manual analyses.
The one limitation of automated sequencers is the length of readable
sequence derived from a single reaction. In our hands, the DuPont se-
quencer produces very accurate sequence for 250 to 400 bases, depending
on the quality of the template. While machines from other manufacturers
(such as ABI) claim higher run lengths, it seems that automated sequencers
are only beginning to be able to match the 600 to 900 bases that can be
read in many laboratories using manual techniques. This can seriously
effect experimental design, as we saw in our analysis of the rRNA gene,
from Gingko biloba. While the manual sequencing provided large amounts
of redundant overlap, the automated sequencer could not provide over-
lapping sequence in every case, leading to the potential loss of information.
However, even though run lengths continue to increase, the shortfall in
overlaps could be avoided by simply reducing the space between adjacent
primers.
The benefits of automated analysis include the ability to sequence a large
number of templates at one time, with a relatively short gel run that does
not require constant attention or numerous loadings. For the production
of sequence information from a large number of alleles or species, this
approach is ideal. This is especially true when combined with PCR to
generate the templates. In some cases, such as with the rRNA genes, gel
and enzymatic artifacts are greatly reduced or absent in automated runs,
thus adding to the overall accuracy. Finally, because all the data are com-
puter-resident, entry errors are eliminated and comparisons can be made
on-screen and directly from the raw data, thus enhancing the proofreading
of ambiguities in the sequence and eliminating the storage, shuffling, and
rereading of autoradiograms.
58 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

ACKNOWLEDGMENT

The DNA sequencer was purchased through a Biomedical Research Grant

through the Division of Sponsored Research at the University of Florida,
Gainesville, FL.

REFERENCES

Amorese, D.A., and A.M. Hochberg. (1989) A fluorescence-labeled DNA analysis

system for sequencing and mapping. Am. Biotech. Lab. 7:38-46.
Chen, E.J., and P.H. Seeburg. (1985) Supercoil sequencing: a fast and simple
method for sequencing plasmid DNA. DNA 4:165-170.
Erlich, H.A. (ed.). (1989) PCR Technology. Principles and Applications for DNA
Amplification. Stockton Press, New York.
Nairn, C.J., M. Miyamoto, and R.J. Ferl. (1991) Phylogenetic implication of the
DNA sequence of the 18s rRNA gene of Gingko biloba. Mol. Biol. Evol.,
submitted.
Prober, J.M., G.L. Trainor, R.J. Dam, F.W. Hobbs, C.W. Robertson, R.J. Za-
gursky, A.J. Cocuzza, M.A. Jensen, and K. Baumeister. (1987) A system
for rapid DNA sequencing with fluorescent chain-terminating dideoxynu-
cleotides. Science 238:336-341.
Sanger, F., S. Nicklen, and A.R. Coulson. (1977) DNA sequencing with chain
terminating inhibitors. Proc Natl. Acad. Sci. USA 74:5463-5467.
She, J.X.,F. Bonhomme,E. Desmarais,andE.K. Wakeland. (1991) Mitochondrial
DNA sequence evolution and phylogeny in the genus Mus. Mol. Biol. Evol.,
submitted.
Smith, L.M., J.Z. Sanders, R.J. Kaiser, P. Hughes, C. Dodd, C.R. Connell, C.
Heiner, S.B.H. Kent, and L. Hood. (1986) Fluorescence detection in au-
tomated DNA sequence analysis. Nature 321:674-679.
4
Computer Alignment of Sequences

MICHAEL S. WATERMAN, JANA JOYCE, AND MARK EGGERT

Sequencing of DNA has precipitated a revolution in biology. Rapid se-

quencing techniques were introduced a little more than a decade ago, and
the rate of sequencing continues to accelerate. As of spring 1991 the in-
ternational databanks (EMBL, GenBank, and DDBJ) contained approx-
imately 50 x 106 base pairs (bp) of DNA sequences with an average
sequence length of about 1,000 bp. Improvements in sequencing technology
continue to be made, and the associated discoveries in biology are stag-
gering.
It is instructive to look at volumes of Dayhoff s Atlas (1979) and consider
the data available in the early 1970s. Hemoglobin and cytochrome c se-
quences had been determined from a variety of organisms, and efforts were
being made to infer the evolutionary relationships from macromolecular
sequences. See Fitch and Margoliash (1967) for a pioneering paper in this
area. Today, in contrast, there are families of sequences for many genes.
Journals such as Journal of Molecular Evolution and Molecular Biology
and Evolution are largely devoted to the study of sequence evolution.
Recently, there has been an exciting new initiative in molecular biology,
the Human Genome Initiative. This project has as its objective the char-
acterization of an entire human genome of 3 x 109 bp. Even though the
basic goal of the genome project is very important and will take several
years to accomplish, the project has much broader implications and could
be the basis of biology for the next century. First of all, the human genome
cannot be considered in isolation from those of other organisms. For ex-
ample, the genomes of Escherichia coli, nematode, and yeast are being
characterized. With a fairly detailed restriction map of E. coli already in
existence (Kohara et al., 1987), both the experimental techniques and the
basic biology learned from work on these organisms are essential to the
Human Genome Initiative. From our point of view, genomic analysis is
an important new emphasis in biology. Genomes from different organisms
can be studied from a comparative viewpoint, and variation within a species

59
60 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

is becoming accessible at the molecular level. When large amounts of

sequence are produced, it will be an enormous task to detect functional
and evolutionary relationships within and between the DNAs of organisms.
In this chapter we will consider sequence alignment, one of the most
utilized computational tools for the study of molecular sequences. See Bell
and Marr (1989) for related discussions of these topics. In the section on
sequence alignment, we discuss the alignment problem, including a max-
imum-likelihood heuristic to motivate alignments that maximize a scoring
function. The computer algorithms used to align full sequences and to
locate maximum segment alignments are given. Alignments from 5S ri-
bosomal RNA (rRNA) gene sequences are presented to illustrate the al-
gorithms. The statistical significance of alignments is also discussed. Finally,
there is a brief survey of multiple sequence alignments.

Sequence Alignment

Several different questions might be asked about a sequence, one of which

concerns unexpected relationships with other sequences. These discoveries
are sometimes made by screening a database with the sequence. Wilbur
and Lipman (1983), and Lipman and Pearson (1985), devised a search
algorithm based on the computer science technique of hashing, and theirs
is the method of choice for such questions. Several groups have imple-
mented versions of their technique. Essentially, the search looks for exact
matching A:-mers (usually k = 4 to 6 for DNA, and k = 2 for protein)
and reports regions where the test sequence has a high density of matches
with the database. Then an optimized alignment region is presented. There
is no further discussion of database searches in this chapter, but they should
not be overlooked.
Sequence alignment is a popular computer activity. The alignments are
often based on some explicit optimization function, rewarding matches and
penalizing mismatches, insertions, and deletions. Sequence alignment can
give useful information about the evolutionary or functional relationships
between sequences.
The generic two-sequence alignment problem can be stated as follows:
a = fljflj • • • an and b = b^b2 . . . bm are two sequences over a finite
alphabet. For DNA and RNA, the alphabet has four letters, and for pro-
teins, the alphabet has 20 letters. Molecular evolution events include sub-
stitutions of one letter for another as well as insertions and deletions (indels)
of letters. What is the minimum number of these events required to trans-
form a to b? The relationships between a and b implied by the minimal
set of events are usually displayed as alignments where letters of a are
written directly over their corresponding letters in b. It is necessary to
weight the transformations in order to come closer to biology, so the
problem has a more general statement.
Let us look at a simple example where a = AAGTTC and b = AGCCC.
COMPUTER ALIGNMENT OF SEQUENCES 61

An alignment that might be suggested by the sequences is

AAGTTC
A- GCCC
where there are three matches, two mismatches, and one indel. This align-
ment represents a specific hypothesis about the evolution of the sequences;
three of the nucleotides have not changed since the common ancestor,
there have been (at least) two substitutions, and one nucleotide has been
either inserted or deleted. Sequences are aligned by computer to find good
alignments. A computer is necessary because of the exponential number
of possible alignments. For example, two sequences of length 1,000 have
over 10600 possible alignments.
There remains the question of how to score an alignment. We present
a simple heuristic derivation of "alignment score." If p is the probability
of an identity, q the probability of a substitution, and r the probability of
an indel, the above alignment has probability
Pr = p3q2r1.
Define score 5' by the log likelihood:
5' = log Pr = 3(logp) + 2(log q) + l(log r).
Define S = S' — n log s = S' — 6 log s, where s is a constant satisfying
\og(p/s) = 1. We have simply subtracted a constant from 5'. 5 becomes
5 = 3 - 2/i - 1(5,
where ju = log(q/s) and d = log(r/s). Fortunately, score defined by
S = #identities - /^substitutions - d#indels
can be efficiently computed as we will see in the next section. It also has
the simple maximum likelihood interpretation we have presented. We cau-
tion the reader that this simple reasoning is only a heuristic. See Tavar6
(1986) for a related mathematically rigorous discussion.
We distinguish two types of alignments: (1) alignments of full sequences;
and (2) finding segments of sequence that can be well aligned. Full sequence
alignment should only be attempted when it is believed that the sequences
are related, from one end to the other. If this is not the case, the sequences
can be forced into incorrect relationships due to the necessity of matching
the less similar segments. A more conservative approach is to run a max-
imum segments analysis that only finds those segments of the sequences
matching at or above some preset level.
Consensus sequence analysis is usually done by the biologists "eye" and
by experiment. Of course, we only believe a protein-binding site when it
is verified by experiment, but analysis by inspection can be biased. So it
is useful to have computer methods that can find consensus patterns best
fitting explicitly stated criteria. Computer algorithms for consensus can be
applied to align sets of multiple sequences.
62 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

DYNAMIC PROGRAMMING ALIGNMENT OF TWO SEQUENCES

In 1970 Needleman and Wunsch wrote a landmark paper that approached
sequence comparison (alignment) with an algorithm that is one of a class
of optimization techniques known as dynamic programming. Their algo-
rithm finds maximum similarity between two sequences, where matches
receive positive weight, and mismatches, insertions, and deletions receive
nonpositive weight. Some mathematicians began to attempt to define a
distance between sequences and so to construct a metric space. This search
culminated with Sellers (1974), where the desired results were obtained
for single insertions and deletions, and with later workers who extended
this work to multiple insertions and deletions (Waterman et al., 1976).
While it has been proven that similarity and distance are equivalent con-
cepts when matching full sequences (Smith et al., 1981), for certain other
problems similarity is superior, and we presented similarity here.
Modern DNA sequencing has revealed introns and exons. This suggests
the problem of finding the best matching pieces (segments) between two
sequences. While this problem was first approached with distance methods,
a much simpler approach through similarity was taken by Smith and Water-
man (1981), and an extension of that technique is presented below (Water-
man and Eggert, 1987).
Matching or aligning entire sequences should be attempted when the
sequences are known or suspected to be closely related. Even when this
is the case, an extraordinary number of best alignments can result; many
of these will differ only slightly from one another. We illustrate this below
with some recommendations on how to deal with the situation. Most se-
quence comparisons will, however, involve sequences only significantly
related in pieces, if at all. In those cases a full alignment is not informative,
and the maximum segments algorithm is the most appropriate. These al-
gorithms can produce segment matchings which are best, second best, and
so on; this is shown in the examples below.

Aligning Full Sequences

As explained above, what is to be presented is a similarity method for
sequence alignment. The function s(a,b) is to define similarity between
the letters a and b. In the examples below, matches receive weight 1 and
mismatches receive -1, so that s(a,b) = 1 if a = b and s(a,b) = -1 if
a =f= b. Deletions of length k receive weight —wk and below, only w± is
used where wk = kw^. The algorithm is based on recursively computing a
matrix 5.
First
So,o = 0, 5,,0 = - Wi, 1 ^ i < n SOJ = - w;> 1 < / < m.
Then
5j - = max {^.jj.! + 5(a,. fe;), max {5(i/-fc - wk}, max {5,-_fcy - wk}}.
k^l ksil
COMPUTER ALIGNMENT OF SEQUENCES 63

For single insertions and deletions only,

S,i} = max {£,_!,,_, + s(atb^, S,_ij - wl,Slii_l - Wj}.
The idea of these calculations is that Slit is the maximum similarity of
a^a2 . • .aiandb1b2. . . b,. This is why for example, S0j = wr The recursion
is based on the ways an alignment can end:
. . . a, corresponds to S,_lj_l + s(at,b^
...b,
and
. . .- corresponds to 5,_ y _ 1 - w^
...b,
and so on.
When « = m, the multiple insertions and deletions program runs in time
proportional to n3, while the single insertion and deletion program runs in
the much preferred time n2. Fortunately, when the function wk is linear in
k, wk = a + /3*k, the running time can still be made n2 (Gotoh, 1982).
Multiple insertions and deletions are important, as adjacent bases are de-
leted or inserted by one event and should be weighted accordingly.
Alignments can be produced from the matrix by two methods. One is
accomplished by saving information at each matrix entry that indicates
which events were necessary to calculate 5,r Then, beginning at.5n m , this
information is used to retrace our steps to 50j0, thus producing the optimal
alignments. The second technique is recomputing to see which events pro-
duced the value of each matrix entry beginning at Snm. Both methods take
little time in comparison with the matrix construction.
For the examples, we first align genes of 55 rRNA E. coli (rrnB operon)
with the closely related 55 rRNA Beneckia Harvey i. The 55 rRNA gene
sequences used in this chapter are taken from Erdmann et al. (1983). As
mentioned above, s(a,b) = +1, if a = b, and s(a,b) = -0.9 if a 4 b
and wl = 2. (wk = <x>, k s? 2, so that only single insertions and deletions
are allowed.) The two sequences have similarity of 78.1. There are two
optimal alignments, which result from simply changing which of the two
Ts is mismatched with the C. See Figure 4-1 for these two alignments.

Figure 4-1 The optimal full sequence alignments of the genes for 55 rRNA E. coli
and 5S rRNA 6. harveyi (top and bottom sequences, respectively). The maximum
segment alignments of the same sequences are identical to the full alignments.
64 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

For a second example, we align genes of 55 rRNA E. coli with the more
distantly related 55 Mycoplasma capricolum. Here the similarity is 18.3,
with 175 different but optimal alignments. The 175 alignments are repre-
sented in the four alignments of Figure 4-2 (Top). The representation is
similar to that of Figure 4-1, so that the first alignment actually represents
a total of 2 x 5 x 3 = 30 alignments, the second represents 2 x 5 x 2 =
20, the third 5 x 5 x 3 = 75, and the fourth 5 x 5 x 2 = 50. This

Figure 4-2 (Top) The four alignments represent the 175 optimal alignments of the
genes for 55 rRNA E. coli and 55 rRNA M. capricolum (top and bottom sequences,
respectively). (Bottom) The two alignments represent the two highest scoring segment
alignments of the same sequences.
COMPUTER ALIGNMENT OF SEQUENCES 65

makes a total of 30 + 20 + 75 + 50 = 175 alignments. These alignments

have a great deal in common. As will be seen in the next section, a max-
imum segments algorithm will produce almost all of the common alignment.

Maximum Segments
This is a preferred method for exploring sequence relationships if computer
time is available. See Smith et al. (1985) for database searches with the
algorithm. It consists of a simple alteration of the preceding method due
to Smith and Waterman (1981). The recursively defined matrix here is H
and recursion begins with H00 = H,0 = 0, 1 £ i s n H0j = 0, 1 <
/ < m and
H,t = max{H,_1J_1 + s(ai,b/),max{Hij_k - wk},max{H,_k>J - wk},0}.
*al Aral

For single insertions and deletions only,

Hit, = max {//,_!,,_! + s(anbj), Ht_^ - w^H^,^ - w1; 0}.
Notice that a simple addition of zeros in the boundary conditions and
recursion is the only change from the definition of S. These simple changes
have the pleasant effect of causing Htil to be the maximum similarity of all
possible segments ending in a, and br Thus max,;//,y is the score of the best
aligned segments in a and b.
After the 1981 Smith-Waterman algorithm, much work has gone into
finding second, third, . . . best segment matches. Fortunately, there is also
a simple, useful solution to this question. There is no problem with the
observation that the best segment matching is associated with the entries
where max,,//,, is achieved. Since large entries influence the entries nearby,
it is not clear whether or not the second-best matching is near the first or
not. One way to make certain of finding the second best is to recalculate
the matrix, only this time no match, mismatch, insertion, or deletion from
the maximum segment can be used. With single or linear insertion and
deletion functions, only a small part of the matrix need be recomputed.
To illustrate the algorithm, we compare the gene sequences of 55 rRNA
E. coll and 55 rRNA M. capricolum as above. There the full similarity
was with 175 different but optimal alignments. They are presented in the
four collections of alignments in Figure 4-2 (Top). Asking for the two best
nonintersecting segment alignments gives the alignments of Figure 4-2 (Bot-
tom). Notice that the regions of alignment common to all 175 alignments
in Figure 4-2 (Top) are almost entirely given in Figure 4-2 (Bottom). A
portion of the first matrix calculation is given in Table 4-1. No recalculation
has been done so that the reader can check understanding of the method.
STATISTICAL DISTRIBUTION OF ALIGNMENT SCORES

Spurious matches are a problem in interpretation of the sequence align-

ments. In this section, we present some results for the distribution of scores
max//,y from the maximum segments algorithm.
Table 4-1 Maximum Segments Matrix for the Genes of 55 rRNA E. coli and 55 rRNA M. capnco/um
T G C T T G G C G A C C A T A G C G A T T
T 10 0 0 10 10 0 0 0 0 0 0 0 0 10 0 0 0 0 0 10 10
G 0 20 0 0 1 20 10 0 10 0 0 0 0 0 1 10 0 10 0 0 1
C 0 0 30 10 0 0 11 20 0 1 10 10 0 0 0 0 20 0 1 0 0
C 0 0 10 21 1 0 0 21 11 0 11 20 1 0 0 0 10 11 0 0 0
T 10 0 0 20 31 11 0 1 12 2 0 2 11 11 0 0 0 1 2 10 10
G 0 20 0 0 11 41 21 1 11 3 0 0 0 2 2 10 0 10 0 0 1
G 0 10 11 0 0 21 51 31 11 2 0 0 0 0 0 12 1 10 1 0 0
C 0 0 20 2 0 1 31 61 41 21 12 10 0 0 0 0 22 2 1 0 0
G 0 10 0 11 0 10 11 41 71 51 31 11 1 0 0 10 2 32 12 0 0
G 0 10 1 0 2 10 20 21 51 62 42 22 2 0 0 10 1 12 23 3 0
C 0 0 20 0 0 0 1 30 31 42 72 52 32 12 0 0 20 0 3 14 0
A 0 0 0 11 0 0 0 10 21 41 52 63 62 42 22 2 0 11 10 0 5
G 0 10 0 0 2 10 10 0 20 21 32 43 54 53 33 32 12 10 2 1 0
T 10 0 1 10 10 0 1 1 0 11 12 23 34 64 44 24 23 3 1 12 11
A 0 1 0 0 1 1 0 0 0 10 2 3 33 44 74 54 34 14 13 0 3
G 0 10 0 0 0 11 11 0 10 0 1 0 13 24 54 84 64 44 24 4 0
C 0 0 20 0 0 0 2 21 1 1 10 11 0 4 34 64 94 74 54 34 14
G 0 10 0 11 0 10 10 1 31 11 0 1 2 0 14 44 74 104 84 64 44
C 0 0 20 0 2 0 1 20 11 22 21 10 0 0 0 24 54 84 95 75 55
G 0 10 0 11 0 12 10 0 30 10 13 12 1 0 0 10 34 64 75 86 66
G 0 10 1 0 2 10 22 2 10 21 1 4 3 0 0 10 14 44 55 66 77
T 10 0 1 11 10 0 2 13 0 1 12 0 0 13 0 0 1 24 35 65 76
G 0 20 0 0 2 20 10 0 23 3 0 3 0 0 4 10 0 11 15 45 56
G 0 10 11 0 0 12 30 10 10 14 0 0 0 0 0 14 1 10 2 25 36
T 10 0 1 21 10 0 10 21 1 1 5 0 0 10 0 0 5 0 1 12 35
C 0 1 10 1 12 1 0 20 12 0 11 15 0 0 1 0 10 0 0 0 15
C 0 0 11 1 0 3 0 10 11 3 10 21 6 0 0 0 10 1 0 0 0
C 0 0 10 2 0 0 0 10 1 2 13 20 12 0 0 0 10 1 0 0 0
COMPUTER ALIGNMENT OF SEQUENCES 67

The average sequence in GenBank or EMBL is 1,000 bases long. Figure

4-3 shows best segment alignments for independent simulations of ten
independent pairs of length 1,000 DNA sequences with P(A) = P(C) =
P(G) = P(T) = 0.25. The algorithm parameters are s(a,b) = 1 if a = b,
s(a,b) = -p for a + b, and w(k) = dk, where fj. = 1.5 and d = 2.1. It
is remarkable that these segmental matchings from random sequences are
so long and score so well. Simulations such as this suggest that understand-
ing the distribution of score (max//y) under the null hypothesis of inde-
pendence is an important goal. If the analysis of "interesting" alignments
proceeds on an ad hoc basis, it is easy to be misled by statistically insig-
nificant alignments. As the genome projects get underway and megabases
of sequence are produced, these statistical considerations will assume in-
creasing importance. The examples of this chapter are of DNA sequences,
but the general theory allows analysis of protein and other sequences.
When two random sequences of length n(a and b) are written in a fixed
alignment, the resulting sequence of matches and mismatches can be iden-

Figure 4-3 Simulation results from the maximum segments algorithm applied to
ten pairs of independent identically distributed DNA sequences of length n = 1,000
with equally likely letters. The algorithm parameters are^ = 1.5 and w(k) = 2.1/c.
68 PHYLOCENETIC ANALYSIS OF DNA SEQUENCES

tified as a sequence of coin tosses. The probability that the klh toss is a
head is P(x[ = y,). Our object in this section is to study long runs of
matches, which in this case are long head runs. The celebrated Erdos-
Renyi law (Erdos and Renyi, 1970) gave order magnitude behavior of the
longest head run containing (1 - a) x 100% tails where a > P(H) = p.
For length Rn of pure head runs (a = 1.0) their result is

with probability one,

while for general a > p their result is

with probability one,

where G(a,p) = a \og(alp) + (1 - a)log((l - a)/(l - p ) ) is relative

entropy. For a = 1, G(a,p) = (log lip) and the two results are consistent.
Other work (Arratia et al., 1986) gives more precise results for this law
and gives

and

where 0.577 . . . is the Euler-Mascheroni constant, 6 = ln(l/p), and the

remainders r^ri) and r2(ri) are of very small, but nonzero magnitude. For
512 fair coin tosses, the mean is approximately 9.33, while the standard
deviation is about 1.93.
The probability results quoted here are for independent and identically
distributed coin tosses. To carry the results over to sequence matching,
the two sequences of length n are assumed to have bases chosen inde-
pendently and identically with p = Pjtwo bases match} = p2A + p2c +
P2c + Pr- The formulae (1) and (2) hold with n replaced by n2.
To consider approximate matching, we allow a fixed number of mis-
matches k. The mean length of the longest match with k mismatches be-
comes, from Arratia et al. (1986), approximately

where q = 1 — p and all logs are taken to base 1/p. The variance remains
approximately jr2/(602) + Vu.
The Erdos-Renyi law for the length Rn of the longest 100% head run of
n coin tosses then extends to a law for the length Mn of the longest match
between two sequences. We have recently shown (Arratia and Waterman,
COMPUTER ALIGNMENT OF SEQUENCES 69

1989) that the length of the longest run of matches containing (1 - a) x

100% mismatches satisfies

with probability one,

which is the Erdos-Renyi law with n replaced by n2. This last theorem has
been obtained by use of the theory of large deviations.
Considering these results, it is not surprising that log(n) laws hold far
beyond the longest exact head run or match. The expected behavior of
max Hv is of importance in evaluating sequence comparisons. If a located
match is at or below that expected from random sequences of similar
composition, then the match should not be further considered without
additional biological information. These distributions have been shown to
fit biological sequences quite well for algorithm parameters not covered
by the theorems above. We have also proven that max HtJ undergoes a
phase transition (Waterman et al., 1987). For larger values of (fi,6), max
Htj grows proportional to log(n) and for smaller values max HtJ grows
linearly with the sequence length. There are only two modes of behavior
at this precision. The logarithmic and linear regions of this two-dimensional
parameter space have been determined numerically in a Monte Carlo study.

MULTIPLE SEQUENCE ALIGNMENT

Many problems of interest in biology involve more than two sequences.

For example, the analysis of E. coli promoter sequences for regions of the
sequences critical to polymerase binding to DNA involves all sequenced
promoters. See Galas et al. (1985) for such an analysis. While analysis of
a set of sequences for evolutionary relationships can be performed by
pairwise sequence analysis, it is often essential to simultaneously consider
the full set of sequences.
Alignment of r > 2 sequences by dynamic programming is possible in
principle, but serious computational difficulties must be overcome. In San-
koff (1975) a technique is presented to optimally align r sequences given
a tree that relates them. This method uses dynamic programming and
Fitch's (1971) parsimony algorithm. The method is very computationally
intensive. Another method is that of Waterman et al. (1976) in which a
tree is not explicitly assumed. All sequences are treated symmetrically so
that a so-called star phylogeny is implicitly assumed. These methods take
at least time proportional to nr where n is the common sequence length.
Recently, some significant progress has been made in the area of multiple
alignment by dynamic programming (see Pearson and Lipman, 1988). Lip-
man and Carrillo have made it feasible to find optimal alignments of up
to six sequences by greatly reducing the number of cells the dynamic pro-
gramming algorithm must consider. Carrillo and Lipman (1988) give the
basic algorithm. In Altschul and Lipman (1989), it is extended to an al-
70 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

gorithm that aligns the sequences given a tree that relates them. In Lipman
et al. (1989) up to eight sequences the length of an average protein can
be aligned.
It is often important to analyze many sequences. A consensus algorithm
for multiple DNA sequence alignment is given that matches words of length
and degree of mismatch chosen by the user (Waterman, 1986). As above,
all sequences are treated symmetrically. The alignment maximizes an align-
ment scoring function. This method can easily align 100 sequences of length
1,000. In Waterman and Jones (1990) the method is extended to protein
sequences.
A heuristic approach that has proven very useful is to infer the evolu-
tionary tree while making a multiple alignment. Usually, the various meth-
ods proceed by performing pairwise alignments. In Johnson and Doolittle
(1986), and Feng and Doolittle (1987), one such development is pursued.
Taylor (1986, 1987) has also written algorithms very useful for protein
analysis. Many of these methods of multiple sequence analysis are pre-
sented in chapters of a comprehensive volume edited by Russell Doolittle
(1990). The program of Hein (1989) combines and refines many of the
other techniques, and this would be our first choice to align and deduce
an evolutionary tree for a set of sequences.

PROGRAM AVAILABILITY

The maximum segments algorithm and consensus alignment algorithms for

DNA and protein sequences are all available from the authors. The im-
plementations are written in the C language.

ACKNOWLEDGMENTS
This work was supported by grants from the National Institutes of Health
and from the National Science Foundation.

REFERENCES

Altschul, S.F., and D.J. Lipman. (1989) Trees, stars, and multiple biological se-
quence alignment. SIAM J. Appl. Math. 49:197-209.
Arratia, R., L. Gordon, and M.S. Waterman. (1986) An extreme value theory for
sequence matching. Ann. Statist. 14:971-993.
Arratia, R., and M.S. Waterman. (1989) The Erdos-Renyi strong law for pattern
matching with a given proportion of mismatches. Annals Prob. 17:1152-
1169.
Bell, G.I., and T.G. Marr. (eds.). (1989) Computers and DNA: The Proceedings
of the Interface Between Computation Science and Nucleic Acid Sequencing
Workshop. Santa Fe Institute Studies in the Series of Complexity, vol. 7.
Addison-Wesley, Redwood City.
COMPUTER ALIGNMENT OF SEQUENCES 71

Carrillo, H., and D. Lipman. (1988) The multiple sequence alignment problem in
biology. SIAMJ. Appl. Math. 48:1073-1082.
Dayhoff, M.O. (ed.). (1979) Atlas of Protein Sequence and Structure, vol. 5, suppl.
3. National Biomedical Research Foundation, Washington, D.C.
Doolittle, R.F. (ed.). (1990) Molecular Evolution: Computer Analysis of Protein
and Nucleic Acid Sequences. Methods in Enzymology, vol. 183. Academic
Press, New York.
Erdmann, V.A., E. Huysman, A. Vandenberghe, and R. De Wachter. (1983)
Collection of published 5S and 5.8S ribosomal RNA sequences. Nucleic
Acids Res. Il:rl05-rl33.
Erdos, P., and A. Renyi. (1970) On a new law of large numbers. J. Anal. Math.
22:103-111; reprinted in Selected Papers of Alfred Renyi, vol. 3, 1962-1970.
Akademiai Kiado, Budapest, 1976.
Feng, D., and R.F. Doolittle. (1987) Progressive sequence alignment as a prereq-
uisite to correct phylogenetic trees. /. Mol. Evol. 25:351-360.
Fitch, W.M. (1971). Towards defining the course of evolution: minimum change
for a specific tree topology. Syst. Zool. 20:406-416.
Fitch, W.M., andE. Margoliash. (1967) Construction of phylogenetic trees. Science
155:279-284.
Galas, D.J., M. Eggert, and M.S. Waterman. (1985) Rigorous pattern recognition
methods for DNA sequences: analysis of promoter sequences from E. coli.
J. Mol. Biol. 186:117-128.
Gotoh, O. (1982) An improved algorithm for matching biological sequences. /.
Mol. Biol. 162:705-708.
Hein, J. (1989) A new method that simultaneously aligns and reconstructs ancestral
sequences for any number of homologous sequences, when the phylogeny
is given. Mol. Biol. Evol. 6:649-668.
Johnson, M.S., and R.F. Doolittle. (1986) A method for the simultaneous align-
ment of three or more amino acid sequences. J. Mol. Evol. 23:267-278.
Kohara, Y., K. Akiyama, and K. Isono. (1987) The physical map of the whole E.
coli chromosome: application of a new strategy for rapid analysis and sorting
of a large genomic library. Cell 50:495-508.
Lipman, D.J., S.F. Altschul, and J.D. Keceioglu. (1989) A tool for multiple se-
quence alignment. Proc. Natl. Acad. Sci. USA 86:4412-4415.
Lipman, D.J., and W.R. Pearson. (1985) Rapid and sensitive protein similarity
searches. Science 227:1435-1441.
Needleman, S.B., and C.D. Wunsch. (1970) A general method applicable to the
search for similarities in the amino acid sequences of two proteins. /. Mol.
Biol. 48:444-453.
Pearson, W.R, and D.J. Lipman. (1988) Improved tools for biological sequence
comparison. Proc. Natl. Acad. Sci. USA 85:2444-2448.
Sankoff, D. (1975) Minimal mutation trees of sequences. SIAM J. Appl. Math.
28:35-42.
Sellers, P. (1974) On the theory and computation of evolutionary distances. SIAM
J. Appl. Math. 26:787-793.
Smith, T.F., and M.S. Waterman. (1981) Identification of common molecular
subsequences. J. Mol. Biol. 147:195-197.
Smith, T.F., M.S. Waterman, and C. Burks. (1985) The statistical distribution of
nucleic acid similarities. Nucleic Acids Res. 13:645-656.
Smith, T.F., M.S. Waterman, and W.M. Fitch. (1981) Comparative biosequence
metrics. J. Mol. Evol. 18:38-46.
72 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Tavare, S. (1986) Some probabilistic and statistical problems in the analysis of

DNA sequences. Lee. Math. Life. Sci., Am. Math. Soc. 17:57-86.
Taylor, W.R. (1986) Identification of protein sequence homology by consensus
template alignment. /. Mol. Biol. 188:233-258.
Taylor, W.R. (1987) Multiple sequence alignment by a pairwise algorithm. CA-
BIOS 3:81-87.
Waterman, M.S. (1986) Multiple sequence alignment by consensus. Nucleic Acids
Res. 14:9095-9102.
Waterman, M.S., and M. Eggert. (1987) A new algorithm for best subsequence
alignments with application to tRNA-rRNA comparisons. J. Mol. Biol.
197:723-728.
Waterman, M.S., L. Gordon, and R. Arratia. (1987) Phase transitions in sequence
matches and nucleic acid structure. Proc. Natl. Acad. Sci. USA 84:1239-
1243.
Waterman, M.S., and R. Jones. (1990) Consensus methods for DNA and protein
sequence alignment. Pp. 221-237 in Molecular Evolution: Computer Anal-
ysis of Protein and Nucleic Acid Sequences (R.F. Doolittle, ed.). Methods
in Enzymology, vol. 183. Academic Press, New York.
Waterman, M.S., T.F. Smith, and W.A. Beyer. (1976) Some biological sequence
metrics. Advances in Math. 20:367-387.
Wilber, W.J., and D.J. Lipman. (1983) Rapid similarity searches of nucleic acid
and protein data banks. Proc. Natl. Acad. Sci. USA 80:726-730.
5
Aligning DNA Sequences: Homology
and Phylogenetic Weighting

DAVID P. MINDELL

Ever since the term homology was defined as "the same organ in different
animals under every variety of form and function" by Owen (1848), it has
frequently been both used and debated. Homology is a concept repre-
senting an hypothesis of correspondence between features, and is used at
different levels of biological organization (e.g., phenotypic, genotypic).
Most evolutionary biologists, particularly systematists, would now agree
that homology describes the relationship between features that are similar
due to common descent among organisms, as opposed to those that are
similar due to convergent or parallel evolution. Thus, identification of
homologous features is a crucial step in determining phylogenetic rela-
tionships. Demonstrating this, homology and synapomorphy have been
presented by some as equivalent relationships (Hennig, 1966; Platnick,
1979), both being used to identify monophyletic groups.
Phylogenetic analysis of DNA sequences involves their alignment in such
a way that homologous nucleotide base positions within homologous genes
are compared. Sequences of DNA offer the largest set of characters and,
hence, potential homologies for phylogenetic studies. By logical extension,
however, they also offer the largest set of potentially convergent and,
hence, phylogenetically misleading characters, indicating the importance
of homology assessment. Sequences differ from morphological characters
in their linear organization, and in their greater (although highly variable)
susceptibility to mutational change. These differences have required dif-
ferent methods in assessment of molecular sequence, as opposed to mor-
phological, homology relationships. The objective of this chapter is to
examine the practice of DNA sequence alignment in the context of ho-
mology assessment. I point out that species sequences should be aligned
in descending order of phylogenetic relationship (phylogenetic weighting
of alignments) to maintain the continuity of information which forms the

73
74 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

basis of relationships of homology. Using mitochondrial ribosomal RNA

(rRNA) sequences, I also show how shuffling the order of input for se-
quences in multiple alignments may be used to help determine phylogenetic
relationships among taxa whose divergences occurred relatively close to-
gether in time.

TYPES OF HOMOLOGY AND "RECOGNITION CRITERIA"

A useful distinction has been made between features that are homologous
between different individuals and those that are homologous within a single
individual organism. These two types, as recognized in a morphological
context, have been called phylogenetic and repetitive homology, respec-
tively (Van Valen, 1982; Roth, 1984). Repetitive homology includes both
serial homology (or repetition), as in the legs of a millipede; and nonserial
repetitions, as in a snake's scales or leaves of a tree. This distinction is
useful as it facilitates the practice of systematics; phylogenetic homologies
are appropriate for phylogenetic analysis of organisms, repetitive homo-
logies are not.
The same and additional homology distinctions have been made for
molecular features, in accord with increased understanding of mechanisms
of molecular change. Sequences of DNA or amino acids that are homol-
ogous between individuals have been termed orthologous, whereas dupli-
cate genes or proteins within an individual are paralogous (Fitch, 1970).
Chromosomal mutations mixing various pieces of genes in new combina-
tions (Gilbert, 1978, 1985) can cause "partial paralogy" among genes and
proteins, and horizontal transfer of DNA sequences between species can
mix "native" and "foreign" sequences (termed "xenology") within an in-
dividual (Gray and Fitch, 1983; see review by Patterson, 1987). For mo-
lecular data, orthologous sequences are appropriate for phylogenetic anal-
ysis of organisms, whereas paralogous sequences are appropriate for
phylogenetic analysis of genes.
In its original (morphological) context the main criteria for recognizing
homologous features have been: (1) an "essential sameness," as deter-
mined by properties of similarity in function, structural position, compo-
sition, or ontogenetic origin (Remane, 1956; Jardine, 1967; Bock, 1974;
Riedl, 1978); and/or (2) congruence with other postulated homologies, as
determined by their agreement in supporting the same phylogenetic rela-
tionships (Wiley, 1975; Eldredge and Cracraft, 1980). These same criteria
of similarity (excluding the ontogenetic origin property) and congruence
are applicable to DNA sequence characters; however, the similarity cri-
terion has required a "new" means of assessment, known as sequence
alignment.
ALIGNING DNA SEQUENCES 75

DNA SEQUENCE ALIGNMENTS

Homology For Sequences Is Not "A Statistical Concept"

The aligning of DNA sequences is the hypothesizing of a homology rela-
tionship for each nucleotide base position. Alignments are relatively easy
for protein-coding genes, due to the presence of structured reading frames
with predictable features of more frequent change at third base positions
within codons, and recognizable start and stop codons. Nonprotein-coding
sequences, such as rRNA genes and mitochondrial D-loop regions, are
difficult to align because they lack these readily distinguishable features.
This is due to the relative abundance of nucleotide base insertion and
deletion events that can occur when there are no selective constraints to
maintain reading frames coding for specific amino acids. Most of the fol-
lowing discussion will pertain to nonprotein-coding DNAs. In phylogenetic
analyses, nucleotide base positions are treated as characters, and each of
the four possible bases represent alternative states. Insertion and deletion
of nucleotide bases connote gain or loss of nucleotide positions; however,
they also may be considered as additional character-states.
A variety of approaches have been taken in aligning sequences, all of
which share a fundamental similarity criterion, in seeking to maximize the
"property" of identical nucleotide bases at aligned positions, and thereby
minimize differences between sequences. In computer-based alignments,
maximizing identical bases is accomplished by assigning various costs (or
penalties) for invoking insertion/deletion (gap) events and nucleotide base
substitutions. This is followed by scoring of different potential alignments,
and selection of the least costly one. The a priori assignment of costs is
crucial, as the number of potential unconstrained alignments for DNA
sequences from even two species is large, and, given the existence of only
four different nucleotide bases and unlimited insertions/deletions, gaps
could be used to achieve "perfect alignment" for any set of sequences.
For example, in the sequence alignment program by Myers and Miller
(1988) that I have used in analyses reported below, the solution involves
finding a set of operations that converts one sequence to another at a
minimum cost. The allowed operations are substitution, deletion, and in-
sertion of nucleotides. Costs for each type of substitution are specified in
a matrix, and the cost of inserting a gap n bases long is g + nh, where g
is the cost, assigned by the user, for opening a gap and h is the user-
assigned cost for increasing gap size by one base. Assignment of different
costs can significantly alter alignments (Fitch and Smith, 1983), and should
be carefully considered. Thus, homology assignments are based on a prob-
abilistic assessment, in that preference for a particular alignment is based
on preassigned costs and comparison of multiple, overall alignment scores.
This quantitative aspect of determining homologies contributed to an
earlier terminological confusion (i.e., use of the term "percent homolo-
gous" to describe the relationship between two sequences), which is now
resolved (Reeck et al., 1987). Choice among possible alignments is prob-
76 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

abilistic; however, once an alignment is adopted, those (hypothesized)

homologous nucleotide positions are qualitatively related. Recognition of
this distinction is important, as it enables us to state clearly that homology
hypotheses based on sequence alignments are no different than other (e.g.,
morphological) homology hypotheses, and that homology in the case of
DNA sequence data has not become a statistical concept as Patterson
(1987:9) has inferred.
Confusing the means of determining homology relationships with the
nature of that relationship may lead to discomfort among some systematists
when they hear that homologies are quantitatively determined in molecular
sequence alignments. However, a quantitative basis for assessing proposed
homologies is nothing new. It is inherent in the homology recognition
criterion (discussed above) of congruence with other postulated homolo-
gies, as determined by their agreement in supporting the same phylogeny.
Phylogenetic congruence among proposed homologies is quantitative, in
that if two or more sets of postulated homologies (synapomorphies) conflict
(support different relationships among taxa), the larger set is preferred.
This is the basis of a most-parsimonious phylogeny.
Alignment Sensitivity to Two Classes of Variation
The fact that we discriminate among possible alignments based on quan-
titative scores makes those alignments, and subsequent phylogenies, sen-
sitive to two classes of variation: (1) variation in rates of DNA sequence
change; and (2) variation in the inclusion or exclusion of taxa from different
alignment bouts. Alignment sensitivity is related to an inherent assumption
of rate constancy in many alignment procedures. In the absence of a priori
weighting schemes for different types of substitutions, or for order of input
of sequences into an alignment (of more than two sequences), alignment
algorithms seek maximal matching at all nucleotide positions. Any differ-
ences in rate of change between sequences can, potentially, introduce
differences in the amount of nonhomologous nucleotide matches accu-
mulating per time-unit since species' divergences, and such nonhomologous
nucleotide matches, or homoplasy, might then contribute to an "inaccur-
ate" phylogeny. Nonhomologous matches can arise from multiple substi-
tutions ("multiple hits") at a single nucleotide position occurring since two
taxa last shared a common ancestor.
Regarding the first class of variation, rates of DNA sequence change
are known to vary: (1) by type of base substitution, as seen in the bias
toward transitions in both mitochondrial (Brown et al., 1982) and nuclear
(Fitch, 1967; Li et al., 1984) DNA; (2) by gene region depending on local
functional constraints, as seen in faster rates of change in introns relative
to exons (Li et al., 1985); (3) by gene, as seen in greater sequence con-
servation in functionally important genes; and (4) across taxa, as seen in
faster rates of change in rodents compared to humans (Wu and Li, 1985;
Britten, 1986). Rate heterogeneity has been a primary concern among
phylogeneticists, hence the diversity of algorithms for phylogeny recon-
ALIGNING DNA SEQUENCES 77

struction with different underlying assumptions regarding rate equality or

the lack of it.
Accommodating rate heterogeneity at the level of DNA sequence align-
ments, though important, has received less attention. Initial procedures
that can be taken include matching phylogenetic questions with sequences
from appropriate (informative) genes or gene regions, and preferentially
maintaining alignment within conserved regions over known variable re-
gions, as done by Gray et al. (1984). Further alignment procedures that
could be taken correspond to character weighting procedures in phyloge-
netic analyses. Alignment, in determining nucleotide character homolo-
gies, is an integral step in phylogenetic analysis, and as such, weighting
based on variable rates of character change may be reasonably applied to
alignments as well. Although this presumes knowledge of evolutionary rate
heterogeneity across the taxa and sequences involved, examples of this
approach include assigning higher alignment costs to transversion over
transition substitutions, to substitutions and gaps in functionally con-
strained over less constrained regions, and in species with relatively slower
rates of change over those with faster rates.
The second class of variation to which alignments are sensitive, differ-
ences in specific taxa included in any given alignment, raises issues no less
complex than those raised by the first class; however, a prescriptive gen-
eralization is more easily made. Namely, DNA sequences from various
taxa should be included (input) in a multiple alignment bout in decreasing
order of phylogenetic relationship, an approach which may be called phy-
logenetic weighting of alignments. At the beginning of this chapter I dis-
cussed the widespread agreement that relationships of homology denote
similarity due to shared inheritance (of a feature) from a common ancestor.
Thus, homology may also be defined on the basis of continuity of infor-
mation (Van Valen, 1982). If so, that information is primarily genetic, and
the "continuity" is provided by genealogy (Roth, 1988). Alignment of
sequences from distant relatives, prior to alignment of more closely related
taxa, would misrepresent the continuity of information in those DNA se-
quences. Similarly, equal treatment (via simultaneous alignment) of se-
quences from all taxa would also remove the genealogical basis of the
homology assessment (unless all the taxa involved experienced simulta-
neous origins from a common ancestor). Thus, phylogenetic weighting of
sequence alignments avoids the problem of confounding similarity and
relationship.
Use of phylogeny to inform alignments may, at first, appear circular
because alignments themselves are used to infer phylogeny. However, this
is an artifact of viewing alignment and phylogenetic analysis as independent
procedures with different objectives when, in fact, they are mutually de-
pendent procedures in assessment of evolutionary history based on nu-
cleotide sequences. Few systematists would seek to determine character
homology (as in sequence alignment) without considering hypothesized
relationships among taxa.
78 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Figure 5-1 Two alternative alignments

(A) (B) for a string of nucleotide bases from each
of three species. (A) is preferred on the
12345 12345 basis of maximal nucleotide matching,
with one substitution and two gaps in-
voked. (B) requires two substitutions and
Species 1 GT-AC GT-AC
one gap, and would be preferred given a
Species 2 G-CAG GC-AG sister relationship between species 1 and
2, relative to species 3, to preserve conti-
Species 3 GTCAC GTCAC nuity of information.

In Figure 5-1, two alternative alignments for an hypothetical sequence

of nucleotide bases from each of three species are presented to demon-
strate: (1) that optimal matching can mask true homology and continuity
of information, in the absence of phylogenetic weighting of alignments;
and (2) that the relative weight, or cost, applied to gaps (insertion/deletion
events) can effect alignment preference. Given a decision to assign less
weight (or cost) to gaps, the first alignment (Fig. 5-1A) is preferred on the
basis of maximal nucleotide matching, because only one substitution and
two gaps are invoked, as opposed to two substitutions and one gap in the
second alignment (Fig. 5-1B). However, if loss of nucleotide position 3 is
uniquely shared by species 1 and 2, which are mammals, relative to species
3, which is a bird, the second alignment (Fig. 5-1B) is preferred on the
basis of maintaining continuity of information. Specifically, the homolo-
gous relationship of the second nucleotide position in species 1 and 2 is
retained. Of course, the first alignment is still feasible, even given the
"known" sister relationship of species 1 and 2. The priority of continuity
of information and genealogy over optimal matching, however, is based
on the fact that homologies are genealogical in origin, whereas similarity
measures, such as optimal matching, are used to assess homology rather
than to define it. The effect of differential weighting of gaps is seen in the
preference for the first alignment (Fig. 5-1A) if gaps are discounted and
the lack of a clear preference for one alignment over the other if gaps are
given the same weight as substitutions.
The early computerized techniques for aligning molecular sequences
either maximized matches or minimized the number of substitution, in-
sertion, or deletion events required to make two sequences equivalent
(Needleman and Wunsch, 1970; Sankoff, 1972; Sellers, 1974). In gener-
alizing this approach to more than two sequences, Sankoff and colleagues
have used "given" phylogenies as the basis for input order of sequence
data in multiple species alignments (Sankoff and Rousseau, 1975), clearly
recognizing the need for phylogenetic weighting of alignments (see Sankoff
et al., 1976:136). Multiple sequence alignment algorithms such as the con-
sensus method of Waterman and Jones (1990), which find sequence pat-
terns that are conserved across taxa but which make no attempt to use
phylogenetic information are inappropriate for the reconstruction of phy-
logeny. Such algorithms are well-suited, however, for similarity searches
ALIGNING DNA SEQUENCES 79

used in identifying unknown sequences (by scanning databanks such as

GenBank), RNA secondary structure, palindromes within sequences, and
so on.
Recently, several computer programs have been developed integrating
alignment and phylogeny analyses that are essentially phenetic in approach.
The "progressive alignment" programs of Feng and Doolittle (1987, 1990)
for amino acid sequences proceed by: (1) determining an initial phenogram
using modified Needleman and Wunsch (1970) alignment scores and the
method of Fitch and Margoliash (1967) to determine branching order; (2)
progressively aligning the sequences in order of their inferred relatedness
from the phenogram; (3) filling gaps with neutral elements; and (4) con-
structing a new tree based on distances calculated from the new overall
alignment. The same strategy used by Feng and Doolittle has been adapted
for nucleotide sequences and microcomputers by Higgins and Sharp (1988,
1989). Konings et al. (1987) (see Hogeweg and Hesper, 1984) also deter-
mine an initial tree from nucleotide sequence distance calculations, and
align sequences according to inferred relationships. However, if the re-
sulting tree differs from the initial one, they realign the sequences and
determine a new tree, repeating this last procedure until the alignment
order and inferred phylogenetic relationships match. The "unified ap-
proach" programs of Hein (1989a,b, 1990) for nucleic and amino acid
sequences also use distances from pairwise alignments to make an initial
tree. Rearrangements of the tree are performed to improve fit to the
distance data, and sequences are then aligned in order of their relationship
inferred from the phenogram. The alignments selected are those that sup-
port a parsimony tree having the same topology as the distance tree. In
this way, the parsimony criterion of minimizing evolutionary change is
applied; however, so-called parsimony analyses of alternative 'alignment
sets are constrained to match the topology of the distance tree.
These approaches have not yet received wide use by systematists, but
certainly will as more sequence data become available. It would be in-
formative to compare results of the above approaches with parsimony trees
(for the same and alternative sequence alignments) that have not been
constrained to fit a distance data set. Where available, independent data
sets could be used in parsimony analyses to provide the initial tree used
for determining order of sequence input to multiple alignments.

ALIGNMENT PROBLEMS ASSOCIATED WITH SHORT INTERNODES

The integrated nature of sequence alignment and phylogenetic analysis

results in a "catch-22" (as coined by novelist Joseph Heller in describing
unavailable prerequisites). Phylogenetic weighting of alignments (aligning
sequences in decreasing order of phylogenetic relatedness) requires knowl-
edge of phylogeny, which the alignment is intended to provide. An initial
tree based on pairwise distances (as in the alignment analyses discussed
above) is most likely to be phylogenetically "inaccurate" where species
80 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

sequences are saturated with change, and where species divergences are
close in time (as in radiations of taxa). In the former case, homoplasies
due to multiple substitutions may unduly affect the tree, and in the latter
case, brevity of uniquely shared ancestry results in few or no synapomor-
phies, and short internodes within a tree. Cavender (1978) and Felsenstein
(1978) have discussed the potential increase in phylogenetically false evi-
dence associated with short internodes and unequal rates of change among
taxa. Even given equal rates of change, stochastically evolving molecular
characters may be unable to support phylogeny reconstruction where time
of independent ancestry for sister taxa is short (see Lanyon, 1988).

Recognizing Short Internodes

The first step in addressing the problem of short internodes is recognizing
them. Lack of phylogenetic resolution from variable DNA sequences may
stem from: (1) "noise" associated with multiple substitutions outweighing
the "signal" from informative differences; and/or (2) brief periods of shared
ancestry yielding no "signal" (shared derived characters) at all, as discussed
above. In some instances, brief shared ancestry for taxa has been inferred
based on fossil or biogeographic evidence and independent phylogenetic
analyses failing to find any "signal." Examples of radiations of taxa with
relatively short periods of shared ancestry include avian and mammalian
orders (see Carroll, 1988 and references therein). Where there is less cer-
tainty of brief shared ancestry, the sequences themselves may provide
evidence. In theory, percent divergence scores for pairwise comparisons
among diverse taxa may be plotted against estimated time since divergence,
where divergence times are "known" from fossil evidence, and the resulting
line or curve may be used to estimate divergence times in instances where
other evidence is lacking. This will, of course, be complicated by hetero-
geneity in rates among taxa.
As an example, consider a plot of percent sequence divergence scores
for species comparisons of mitochondrial small-subunit rRNAs (125 in
vertebrates) (Table 5-1, Fig. 5-2). The relationship between "% Sequence
Divergence" and "Estimated Divergence Time" is roughly linear for spe-
cies divergences of 85 million years or less. Thereafter, divergence accu-
mulates more slowly over time, as multiple substitutions at individual nu-
cleotide positions lead to increasing homoplastic similarity between sequences.
If no fossil evidence was available for the comparisons among represent-
atives of mammalian orders, those sequence divergence values might be
secondarily overlain on the curve (Fig. 5-2), and estimated values for di-
vergence times taken from the horizontal axis. Clustering of percent di-
vergence scores suggests brief periods of shared independent ancestry for
those taxa, assuming similarity in rates of sequence change. Note, however,
the overlap in percent divergence scores for different "known" estimated
divergence times (Table 5-1), making this approach, for 12S rRNA at least,
highly approximate. Conversely, if percent divergence scores were not
Table 5-1 Pairwise Comparisons of Mitochondrial Small-Subunit rRNAs*
Species 1 2 3 4 5 6 7 8 9 10
1. Mus 57/20 43/85 41/85 52/85 52/85 52/85 47/300 46/350 44/600
2. Rattus 8.6 44/85 46/85 54/85 56/85 54/85 46/300 50/350 42/600
3. Bos 21.3 21.8 79/30 64/85 60/85 62/85 49/300 47/350 47/600
4. Giraffa 20.4 22.5 11.9 59/85 61/85 60/85 52/300 51/350 48/600
5. Homo 24.1 22.9 20.9 22.4 94/8 93/19 50/300 47/350 48/600
6. Pan 23.6 22.7 21.1 21.9 4.3 93/16 45/300 49/350 49/600
7. Pongo 26.4 24.8 23.1 23.7 9.1 9.3 47/300 48/350 49/600
8. Gallus 24.3 30.2 30.0 30.4 27.8 28.8 30.0 60/350 52/600
9. Xenopus 28.4 28.7 25.7 -27.2 28.2 28.2 28.1 27.3 48/600
10. Drosophila 34.7 35.3 37.7 36.2 37.4 38.4 38.8 39.0 40.9
'Values below the diagonal are percent divergence scores, calculated as: 100{l-[number identical nucleotides/(number shared base positions + total number gaps)]}. Values
above the diagonal are: percent of substitutions that are transitions/estimated divergence times in millions of years before present. Species names and data sources are: Mus
musculus (Van Etten et al., 1980), Rattus norvegicus (Kobayashi et al , 1981), Bos laurus (Anderson et al., 1982), Giraffa camelopardalis (Tanhauser, 1985), Homo sapiens
(Anderson et al., 1981), Pan troglodytes (Hixson and Brown, 1986), Pongo pygmaeus (Hixson and Brown, 1986), Gallus gallus (Desjardins and Morais, 1990), Xenopus
laevis (Dunon-Bluteau and Brun, 1986), and Drosophila yakuba (Clary and Wolstenholme, 1985). Estimates for divergence times are based on Schopf et al. (1983), Shosham
et al. (1985), Britten (1986), Hixson and Brown (1986), Carroll (1988), and references therein
82 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Figure 5-2 Percent sequence divergence for mitochondria! small-subunit rRNA,

and percent transitions versus estimated divergence time (open circles and filled
circles, respectively) for species pairwise comparisons. Data are from Table 5-1 and
Hixson and Brown (1986). Points represent averages for species comparisons having
the same estimated divergence times. Percent sequence divergence was calculated
as given in Table 5-1.

clustered, the taxa involved in those comparisons would not be inferred

as separated by short internodes within a tree. Thus, lack of resolution of
their relationships could be more reasonably attributed to overabundance
of "noise," rather than an absence of "signal."

Shuffling Input Order for Alignments

Demonstration
Once sets of taxa separated by short internodes have been identified, an
obvious procedure in seeking to resolve their relative branching order is
to shuffle the input order of their nucleotide sequences in subsequent
multiple alignment bouts, and to examine resulting phylogenetic trees. If
the resultant trees are the same, order of input of sequences has no major
effect, and the relationships may be considered unbiased by alignment
input order. I have taken this approach in alignment and phylogenetic
analysis of 12S and 165 rRNA sequences from representatives of three
mammalian orders: human (Homo sapiens), cow (Bos taurus), and mouse
(Mus musculus), using the chicken (Callus gallus) as an outgroup. The
three mammalian (ingroup) species sequences were aligned in all three
possible orders: (Homo, Mus)Bos; (Homo, Bos)Mus; (Bos, Mus)Homo.
In each case, sequences for the first two species enclosed by parentheses,
above, were aligned, and a consensus sequence was made from that align-
ment in which all variable nucleotide sites were overwritten as "N" (am-
ALIGNING DNA SEQUENCES 83

biguous). This consensus sequence was then aligned with the third ingroup
species. Another consensus sequence was made, as above, from these three
sequences combined, and then the outgroup was aligned to this second
consensus sequence. Pairwise alignments were done using the algorithm
of Myers and Miller (1988). Program parameters for "open gap cost" (cost
for opening a gap in sequence) and "unit gap cost" (cost for increasing
the size of the gap by one base) were both set to 10. Alignments done with
these settings could not readily be improved by visual inspection and ad-
justing alignments by hand, whereas experimental alignments with various
other settings could be improved after visual inspection. Phylogenetic anal-
yses were done using the Phylogenetic Analysis Using Parsimony (PAUP)
computer program (Swofford, 1989). This approach is reasonable for small
numbers of species; however, as the number of variable sites increases
with inclusion of more taxa, ambiguous ("N") designations in the consensus
also increase, and too many such ambiguous scores will eventually con-
found the alignments. Use of "N" scores is conservative, and analyses
could also be done in which variable nucleotide sites were overwritten with
the appropriate International Union of Pure and Applied Chemistry (IUPAC)
ambiguity scores combining the nucleotide variation observed. This would
facilitate alignment of more sequences from more species.
To investigate the effects of the order of alignment of species' sequences,
I constrained phylogenetic analyses to each of the three possible tree to-
pologies, for each of the three different "alignment order" data sets. I
have also employed two weighting strategies for the nucleotide characters,
equal weighting for all nucleotide substitutions, and giving zero weight to
(omitting) transitions (Fig. 5-3). Omitting transitions, however, is not meant
to imply that they contain no phylogenetic information.

Interpretation
The phylogenetic analyses with shuffling of input order of alignments dem-
onstrate that phylogenetic hypotheses can be quite sensitive to alignment
order of species sequences. In analyses of 165 rRNA sequences, two dif-
ferent most-parsimonious (designated by scores with an asterisk in Figure
5-3) tree topologies were found under conditions of equal weighting for
all characters. Considering transversions only, all three possible topologies
were found as most parsimonious, varying with alignment order. There is
an expected tendency for tree topology to reflect the order of sequence
alignments. In 165 rRNA analyses of transversions only, branching order
within most-parsimonious trees always corresponded with alignment order.
In contrast, 125 rRNA analysis under each weighting scheme consistently
identifed a single tree topology as most parsimonious, regardless of align-
ment order. This convergence toward a single topology suggests that hy-
pothesized relationships based on 125 rRNA are not an artifact of align-
ment order. 125 rRNA analyses with all nucleotide substitutions equally
weighted unite Bos and Mus as sisters, as found by Cedergren et al. (1988)
using a different alignment-phytogeny approach, whereas consideration of
 84 PHYLOCENETIC ANALYSIS OF DNA SEQUENCES

H M B H B M B M H
FORCED G G G
TREE
TOPOLOGY

(1) (2) (3)

ALIGNMENT
ORDER — *• (H,M)B (M,B)H (H,B)M (H,M)B (M,B)H (H,B)M (H,M)B (M,B)H (H,B)M

PARSIMONY SCORE
12S rRNA
All chars. 565 561 559 559 559 547 548* 538* 541*
Transv. only 273 275 272 264* 263* 256* 274 272 268

16S rRNA
All chars. 1021 1014 1007 1013* 1002 986* 1017 987* 1001
Transv. only 562* 563 571 564 553 550* 564 545* 564

Figure 5-3 Lengths for each of three possible phylogenetic trees (parsimony scores)
using each of the three possible alignment orders (input order of sequences for
multiple alignments) for Homo sapiens (H), Bos taurus (B), and Mus musculus (M)
with Callus gallus (C) as an outgroup. For each alignment order, mitochondrial
rRNA sequences for two species enclosed by parentheses were aligned (Myers and
Miller, 1988), and a consensus sequence was made from that alignment in which
all variable nucleotide sites were overwritten as "N" (ambiguous). This consensus
sequence was then aligned with the third ingroup species, another consensus se-
quence was made, and the outgroup sequence was aligned to that second consensus
sequence. Parsimony scores are given for analyses under two weighting schemes,
equal weight for all substitutions, and zero weight for transitions (transversionsonly).
Asterisks denote the score for the most-parsimonious tree topology under the given
alignment order and weighting scheme. Data sources for both small-subunit (12S)
and large-subunit (16S) rRNA analyses are given in Table 5-1.

transversions only place Homo and Bos as sisters. The bias toward tran-
sitions in mitochondrial sequences (Brown et al., 1982; Thomas and Beck-
enbach, 1989), including rRNAs (Hixson and Brown, 1986), is the basis
of the difference in tree topologies. The faster rate of accumulation for
transitions makes them more susceptible to multiple substitutions at a given
nucleotide position, and less informative for phylogeny as divergence time
between species pairs increases. Based on small-subunit mitochondrial rRNA
comparisons (Fig. 5-2), percent transitions begins leveling off at about 50%.
Transitions begin to saturate with change, and become phylogenetically
less informative for species divergences over about 20 million years old.
Evolutionary patterns among eutherian orders are not well documented
by fossils (Clemens et al., 1979; Novacek, 1982), and estimates for diver-
gence times among the three mammalian orders represented here (Ro-
dentia, Artiodactyla, Primates) vary from 55 to 90 million years ago based
on both fossil and biochemical (e.g., Shoshani et al., 1985) evidence, but
it is reasonable to suppose that, in the present analyses, transitions are less
reliable than transversions. In this light, the best mitochondrial rRNA
phylogenetic hypothesis is tree 2 in Figure 5-3, having Homo and Bos as
ALIGNING DNA SEQUENCES 85

sisters relative to Mus. This variously conflicts (Miyamoto and Goodman,

1986; Novacek, 1986) or agrees (McKenna, 1975; Easteal, 1990) with hy-
potheses based on other data sets.
Further support for the value of transversions in this case, and the phy-
logeny inferred from them, is found in the greater relative difference in
tree length between the most-parsimonious tree and the other two trees
for transversions alone, compared to that for trees based on all characters
combined. For analyses presented in Figure 5-3, the sum of additional steps
required to support the two less-parsimonious trees constituted 24 and 12%
of the mean tree length for transversions only in the 125 and 16S rRNA
genes, respectively, compared to 17 and 9%, in the 125 and 165 rRNA
genes respectively, for all characters combined.
Assigning different weights to different types of nucleotide character
change, as above, is unreasonable if it is done after phylogenetic analysis
in order to "save" a particular set of relationships. However, it is reasonable
where such weighting can be justified beforehand, based on knowledge of
relative rates of character change as done here. Weighting could also have
been done using a recursive, successive approximations approach (Car-
penter, 1988). Because DNA sequence homologies, and hence types of
character change (i.e., transitions or transversions), are based on quanti-
tative alignment scores, and because there are only four possible nucleotide
character-states, nonhomologous similarity will arise often. Differential
weighting of types of molecular character change may be needed in many
instances, lest homoplastic similarity overwhelm homologous similarity in
phylogenetic analyses.

ACKNOWLEDGMENTS

I am grateful to Jim Carpenter, Joel Cracraft, Brent Mishler, and Michael

M. Miyamoto for valuable criticisms of an earlier draft of this paper. Steven
Sawchuck kindly assisted in conducting species sequence comparisons.

REFERENCES

Anderson, S., A. T. Bankier, B. G. Barrell, M. H. L. de Bruijn, A. R. Coulson,

J. Drouin, I. C. Eperon, D. P. Nierlich, B. A. Roe, F. Sanger, P. H.
Schreier, A. J. H. Smith, R. Staden, and I. G. Young. (1981) Sequence
and organization of the human mitochondria! genome. Nature 290:457-465.
Anderson, S., M. H. L. de Bruijn, A. R. Coulson, E. C. Eperon, R. Sanger, and
I. G. Young. (1982) Complete sequence of bovine mitochondrial DNA:
conserved features of the mammalian mitochondrial genome. J. Mol. Biol.
156:683-717.
Bock, W. J. (1974) Philosophical foundations of classical evolutionary classification.
Syst. Zool. 22:375-392.
86 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Britten, R. J. (1986) Rates of DNA sequence evolution differ between taxonomic

groups. Science 231:1393-1398.
Brown, W. M., E. M. Prager, A. Wang, and A. C. Wilson. (1982) Mitochondrial
DNA sequences of primates: tempo and mode of evolution. J. Mol. Evol.
18:225-239.
Carpenter, J. M. (1988) Choosing among multiple equally parsimonious clado-
grams. Cladistics 4:291-296.
Carroll, R. L. (1988) Vertebrate Paleontology and Evolution. W. H. Freeman, New
York.
Cavender, J. A. (1978) Taxonomy with confidence. Math. Biosci. 40:271-280.
Cedergren, R., M. W. Gray, Y. Abel, and D. Sankoff. (1988) The evolutionary
relationships among known life forms. J. Mol. Evol. 28:98-112.
Clary, D. O., and D. R. Wolstenholme. (1985) The ribosomal RNA genes of
Drosophila mitochondrial DNA. Nucleic Acids Res. 13:4029-4045.
Clemens, W. A., J. A. Lillegraven, E. H. Lindsay, and G. G. Simpson. (1979)
Where, when and what—a survey of known Mesozoic mammal distribution.
Pp. 7-58 in Mesozoic Mammals (J. A. Lillegraven, Z. Kielan-Jaworowska,
and A. Clemens, eds.). University of California Press, Berkeley.
Desjardins, P., and R. Morais. (1990) Sequence and gene organization of the
chicken mitochondrial genome. /. Mol. Biol. 212:599-634.
Dunon-Bluteau, D., and G. Brun. (1986) The secondary structures of the Xenopus
laevis and human mitochondrial small ribosomal subunit RNA are similar.
FEES Lett. 198:333-338.
Easteal, S. (1990) The pattern of mammalian evolution and the relative rate of
molecular evolution. Genetics 124:165-173.
Eldredge, N., and J. Cracraft. (1980) Phylogenetic Patterns and the Evolutionary
Process. Columbia University Press, New York.
Felsenstein, J. (1978) Cases in which parsimony or compatibility methods will be
positively misleading. Syst. Zoo/. 27:401-410.
Feng, D., and R. F. Doolittle. (1987) Progressive sequence alignment as a pre-
requisite to correct phylogenetic trees. J. Mol. Evol. 25:351-360.
Feng, D., and R. F. Doolittle. (1990) Progressive alignment and phylogenetic tree
construction of protein sequences. Pp. 375-387 in Molecular Evolution:
Computer Analysis of Protein and Nucleic Acid Sequences (R. F. Doolittle,
ed.). Methods in Enzymology, vol. 183. Academic Press, New York.
Fitch, W. M. (1967) Evidence suggesting a non-random character to nucleotide
replacements in naturally occurring mutations. J. Mol. Biol. 26:499-507.
Fitch, W. M. (1970). Distinguishing homologous from analogous proteins. Syst.
Zoo/. 19:99-113.
Fitch, W. M.,andE. Margoliash. (1967) Construction of phylogenetic trees. Science
155:279-284.
Fitch, W. M., and T. F. Smith. (1983) Optimal sequence alignments. Proc. Natl.
Acad. Sci. USA 80:1382-1386.
Gilbert, W. (1978) Why genes in pieces? Nature 271:501.
Gilbert, W. (1985) Genes-in-pieces revisited. Science 228:823-824.
Gray, G. S., and W. M. Fitch. (1983) Evolution of antibiotic resistance genes: the
DNA sequence of a kanamycin resistance gene from Staphylococcus aureaus.
Mol. Biol. Evol. 1:57-66.
Gray, M. W., D. Sankoff, and R. J. Cedergren. (1984) On the evolutionary descent
of organisms and organelles: a global phylogeny based on a highly conserved
ALIGNING DNA SEQUENCES 87

structural core in small subunit ribosomal RNA. Nucleic Acids Res. 12:5837-
5852.
Hein, J. (1989a) A new method that simultaneously aligns and reconstructs an-
cestral sequences for any number of homologous sequences, when the phy-
logeny is given. Mol. Biol. Evol. 6:649-668.
Hein, J. (1989b) A tree reconstruction method that is economical in the number
of pairwise comparisons used. Mol. Biol. Evol. 6:669-684.
Hein, J. (1990) Unified approach to alignment and phylogenies. Pp. 626-644 in
Molecular Evolution: Computer Analysis of Protein and Nucleic Acid Se-
quences (R. F. Doolittle, ed.). Methods in Enzymology, vol 183. Academic
Press, New York.
Hennig, W. (1966) Phylogenetic Systematics. University of Illinois Press, Urbana.
Higgins, D. G., and P. M. Sharp. (1988) CLUSTAL: a package for performing
multiple sequence alignment on a microcomputer. Gene 73:237-244.
Higgins, D. G., and P. M. Sharp. (1989) Fast and sensitive multiple sequence
alignments on a microcomputer. CABIOS 5:151-153.
Hixson, J. E., and W. M. Brown. (1986) A comparison of the small ribosomal
RNA genes from the mitochondria! DNA of the great apes and humans:
sequence, structure, evolution, and phylogenetic implications. Mol. Biol.
Evol. 3:1-18.
Hogeweg, P., and B. Hesper. (1984) The alignment of sets of sequences and the con-
struction of phyletic trees: an integrated method. J. Mol. Evol. 20:175-186.
Jardine, N. (1967) The concept of homology in biology. Br. J. Philos. Sci. 18:125-
139.
Kobayashi, M., T. Seki, K. Yaginuma, and K. Koike. (1981) Nucleotide sequences
of small ribosomal RNA and adjacent transfer RNA genes in rat mito-
chondria! DNA. Gene 16:297-307.
Konings, D. A. M., P. Hogeweg, and B. Hesper. (1987). Evolution of the primary
and secondary structures of the Ela mRNAs of the adenovirus. Mol. Biol.
Evol. 4:300-314.
Lanyon, S. M. (1988) The stochastic mode of molecular evolution: what conse-
quences for systematic investigations? Auk 105:565-573.
Li, W.-H., C.-C. Luo, and C.-I. Wu. (1985) Evolution of DNA sequences. Pp.
1-94 in Molecular Evolutionary Genetics. (R. J. Maclntyre, ed.). Plenum
Press, New York.
Li, W.-H., C.-I. Wu, and C.-C. Luo. (1984) Nonrandomness of point mutation as
reflected in nucleotide substitutions in pseudogenes and its evolutionary
implications. J. Mol. Evol. 21:58-71.
McKenna, M. C. (1975) Toward a phylogenetic classification of the Mammalia.
Pp. 21-46 in Phytogeny of the Primates (W. P. Luckett and F. S. Szalay,
eds.). Plenum Press, New York.
Miyamoto, M. M., andM. Goodman. (1986) Biomolecular systematics of eutherian
mammals: phylogenetic patterns and classification. Syst. Zool. 35:230-240.
Myers, W. W., and W. Miller. (1988) Optimal alignments in linear space. CABIOS
4:11-17.
Needleman, S. B., and C. D. Wunsch. (1970) A general method applicable to the
search for similarities in the amino acid sequence of two proteins. /. Mol.
Biol. 48:443-453.
Novacek, M. J. (1982) Information for molecular studies from anatomical and fossil
evidence on higher eutherian phylogeny. Pp. 3-41 in Macromolecular Se-
88 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

quences in Systematics and Evolutionary Biology (M. Goodman, ed.). Plenum

Press, New York.
Novacek, M. J. (1986) The skull of leptictid insectivorans and the higher-level
classification of eutherian mammals. Bull. Am. Mus. Nat. Hist. 183:1-112.
Owen, R. (1848) On the Archetype and Homologies of the Vertebrate Skeleton. R.
and J. E. Taylor, London.
Patterson, C. (1987) Introduction. Pp. 1-22 in Molecules and Morphology in Ev-
olution: Conflict or Compromise! (C. Patterson, ed.). Cambridge University
Press, London.
Platnick, N. I. (1979) Gaps and prediction in classification. Syst. Zoo/. 27:472-
474.
Reeck, G. R., C. de Haen, D. C. Teller, R. F. Doolittle, W. M. Fitch, R. E.
Dickerson, P. Chambon, A. D. McLachlan, E. Margoliash, T. H. Jukes,
and E. Zuckerkandl. (1987) "Homology" in proteins and nucleic acids: a
terminological muddle and a way out of it. Cell 50:667.
Remane, A. (1956) Die Grundlagen des Naturlichen Systems der Vergleichenden
Anatomie und Phylogenetik. 2. Geest und Portik, Leipzig.
Riedl, R. (1978) Order in Living Organisms. A Systems Analysis of Evolution.
John Wiley and Sons, New York.
Roth, V. L. (1984) On homology. Biol. J. Linn. Soc. 22:13-29.
Roth, V. L. (1988) The biological basis of homology. Pp. 1-26 in Ontogeny and
Systematics (C. J. Humphries, ed.). Columbia University Press, New York.
Sankoff, D. (1972) Matching sequences under deletion/insertion constraints. Proc.
Natl. Acad. Sci. USA 69:4-6.
Sankoff, D., R. J. Cedergren, and G. Lapalme. (1976) Frequency of insertion-
deletion, transversion, and transition in the evolution of 5S ribosomal RNA.
J. Mol. Evol. 7:133-149.
Sankoff, D., and P. Rousseau. (1975) Locating the vertices of a Steiner tree in
arbitrary space. Math. Progr. 9:240-246.
Schopf, J. W., J. M. Hayes, and M. R. Walter. (1983) Evolution of earth's earliest
ecosystems: recent progress and unsolved problems. Pp. 361-384 in Earth's
Earliest Biosphere: Its Origin and Evolution (J. W. Schopf, ed.). Princeton
University Press, Princeton.
Sellers, P. (1974) An algorithm for the distance between two finite sequences. J.
Combination Theory 16:253-258.
Shoshani, J., M. Goodman, J. Czelusniak, and G. Braunitzer. (1985) A phylogeny
of Rodentia and other eutherian orders: parsimony analysis utilizing amino
acid sequences of alpha and beta hemoglobin chains. Pp. 191-210 in Ev-
olutionary Relationships Among Rodents: A Multidisciplinary Analysis. (W.
Patrick and J.-L. Hartenberger, eds.). Plenum Press, New York.
Swofford, D. L. (1989) PAUP: Phylogenetic Analysis Using Parsimony, Version
3.0. University of Illinois, Champaign.
Tanhauser, S. M. (1985) Evolution of Mitochondrial DNA: Patterns and Rate of
Change. Ph.D. Dissertation, University of Florida, Gainesville.
Thomas, W. K., and A. T. Beckenbach. (1989) Variation in salmonid mitochondrial
DNA: evolutionary constraints and mechanisms of substitution. /. Mol.
Evol. 29:233-245.
Van Etten, R. A., M. W. Walberg, and D. A. Clayton. (1980) Precise localization
and nucleotide sequence of the mouse mitochondrial rRNA genes and three
immediately adjacent novel tRNA genes. Cell 22:157-170.
ALIGNING DNA SEQUENCES 89

Van Valen, L. M. (1982) Homology and causes. /. Morphology. 173:305-312.

Waterman, M. S., and R. Jones. (1990) Consensus methods for DNA and protein
sequence alignment. Pp. 221-237 in Molecular Evolution: Computer Anal-
ysis of Protein and Nucleic Acid Sequences (R. F. Doolittle, ed.). Methods
in Enzymology, vol. 183. Academic Press, New York.
Wiley, E. O. (1975) Karl R. Popper, systematics, and classification: a reply to
Walter Bock and other evolutionary taxonomists. Syst. Zool. 24:233-243.
Wu, C.-I., and W.-H. Li. (1985) Evidence for higher rates of nucleotide substitution
in rodents than in man. Proc. Nad. Acad. Sci. USA 82:1741-1745.
6
Relative Efficiencies of Different Tree-
Making Methods for Molecular Data

MASATOSHI NEI

There are many different tree-making methods that can be used for mo-
lecular data (Nei, 1987a; Felsenstein, 1988). Each of these methods has
some advantages and disadvantages, and the overall relative efficiencies
of the methods in recovering the correct phylogenetic tree are still contro-
versial. The major problem in studying the relative efficiencies is that the
true tree is usually unknown for any set of real organisms or any set of
real DNA sequences, so that it is difficult to judge which tree is the correct
one. However, this problem can be avoided if we use computer simulation.
In the case of molecular data, particularly DNA sequences, the pattern of
evolutionary change is well-understood so that it is possible to simulate it
for any given phylogenetic tree. Sequences of DNA or other molecular
data thus generated can be used for reconstructing the tree, and the effi-
ciency of a tree-making method can be studied by examining how often it
recovers the correct tree.
This type of study was first conducted by Peacock and Boulter (1975)
who considered the evolution of amino acid sequences. However, they
compared only two different tree-making methods considering a constant
rate of evolution. Furthermore, they did not consider the property of the
genetic code in their simulation of amino acid changes. Therefore, their
conclusion has limited applicability to actual data. In the late 1970s we
initiated a comprehensive study of this problem, considering DNA se-
quences (Tateno et al., 1982). Initially, we studied relatively simple tree-
making methods but later extended our analysis to other methods. Similar
studies have also been done by a number of other authors. In this chapter,
a summary of the results of these studies is presented. Before the discussion
of these results, however, the theoretical basis of each tree-making method
that is used for molecular data will be presented.

90
RELATIVE EFFICIENCIES OF DIFFERENT TREE-MAKING METHODS 91

TREE-MAKING METHODS

As mentioned above, there are many methods for constructing trees from
molecular data. According to the type of data used, they can be divided
into two categories; that is, distance methods and discrete-character meth-
ods. In distance methods, evolutionary distance is computed for all pairs
of operational taxonomic units (OTUs; species or populations) or DNA
(or amino acid) sequences, and a phylogenetic tree is constructed by con-
sidering the relationship among these distances. Once distance values are
obtained, there are several ways of obtaining a tree. In discrete-character
methods, data with discrete character-states such as nucleotide states in
DNA sequences are used, and a tree is constructed by considering the
evolutionary relationships of OTUs or DNA sequences at each character
or nucleotide position.
It should be noted that some types of molecular data (e.g., DNA hy-
bridization data) exist only as distance data. Therefore, phylogenetic trees
for these data can be constructed only by distance methods. By contrast,
discrete-character data can usually be converted into distance data, so that
they can be analyzed either by distance methods or by discrete-character
methods. Some authors (e.g., Farris, 1981; Penny, 1982) have argued that
distance methods are inherently inferior to discrete-character methods (e.g.,
parsimony methods), but their arguments are apparently based on mis-
conceptions of distance methods (Felsenstein, 1986; Nei, 1987b). Actually,
some distance methods are often superior to parsimony methods in ob-
taining the correct tree, as will be mentioned below.
For both distance and discrete-character data, there are several different
tree-making methods that are based on different principles. The major
ones are as follows.

Distance Data
Average Distance (Linkage) Method [Unweighted Pair Group Method
with Arithmetic Means (UPCMA)]
The original idea of this method was presented by Sokal and Michener
(1958), but their procedure is different from the one currently in use. A
formal presentation of this method is given in Sneath and Sokal (1973).
To construct a phylogenetic tree by this method, it is necessary to assume
a constant rate of evolution.
Transformed Distance (TD) Method
Farris (1977) showed that if a distance matrix satisfies the condition of an
additive tree (all substitutions being counted) the UPGMA procedure pro-
duces the correct tree topology by using the following transformed distance
for species i and j.
d,; = (d,, - dir - dir)i2 + c, (i)
92 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

where r stands for a reference species (an outgroup or any given ingroup
species) and c is a constant to make d,t' positive. The TD method works
well whether the evolutionary rate varies from branch to branch. In prac-
tice, of course, the condition of an additive tree is not usually satisfied, so
this method may produce a wrong topology. Branch lengths cannot be
estimated by this method. However, once a topology is obtained, one can
easily estimate branch lengths by using Fitch and Margoliash's (1967) method
mentioned below, or some other method. There are several different al-
gorithms for this method (Klotz et al., 1979; Klotz and Blanken, 1981; Li,
1981).
Fitch and Margoliash's (FM) Method
Let dtl be the observed distance between species i and ; and et] be the
corresponding patristic distance, which is the sum of the lengths of all
branches connecting the two species in a reconstructed tree. Fitch and
Margoliash's (1967) method chooses a tree that minimizes the following
quantity.

where n is the number of species studied. A similar least-squares method

was also proposed by Cavalli-Sforza and Edwards (1967). In this method
the SFM value must be computed for all or a large number of topologies.
Minimum Evolution (ME) Method
In an unrooted bifurcating tree of n species, there are 2n — 3 possible
branches. Let €, be the branch length of the f-th branch for a given topology.
One can then compute the sum of all branches. That is,

In the ME method, this quantity is computed for all possible topologies,

and the topology that shows the minimum L value is chosen as the final
tree. This method was originally presented by Cavalli-Sforza and Edwards
(1967), but their method of computing €, was very complicated. Recently,
Saitou and Imanishi (1989) simplified the computation considerably by
using Fitch and Margoliash's (1967) approach.
One might think that this method is essentially the same as the maximum
parsimony method mentioned below. This is not true. Unlike the parsi-
mony method, this method is not affected by parallel or backward muta-
tions, as long as the evolutionary distance is properly measured. It recovers
the correct tree if the distance matrix satisfies the condition of an additive
tree.
Distance Wagner (DW) Method
This method was originally presented as a distance version of the Wagner
parsimony method (Farris, 1972). Farris, therefore, suggested that a meas-
RELATIVE EFFICIENCIES OF DIFFERENT TREE-MAKING METHODS 93

ure called "metric" (a distance measure that satisfies the principle of the
triangle inequality) be used for this method. However, this requirement
does not necessarily improve the performance of this method (Nei et al.,
1983; Sourdis and Krimbas, 1987). Tateno et al. (1982) modified this method
to make it more appropriate for molecular data. Swofford (1981) and Faith
(1985) also proposed modified versions of this method. Comparison of
different topologies is built in the method, and the final tree that is supposed
to be the best one is automatically produced. This method gives both
topology and branch lengths of the final tree.
Neighborliness (ST) Method
The principle of the neighborliness (ST) method (Sattath and Tversky,
1977; Fitch, 1981) is to use Buneman's (1971) four-point condition for an
additive tree. Consider a tree with n (> 4) OTUs and assume that OTUs
1 and 2 are a pair of neighbors, that is, two OTUs connected through a
single interior node in a bifurcating tree (see Saitou and Nei, 1987). Let
dl} be the distance between OTUs i and j. We then have the following
inequalities for an additive tree.
d12 + dv < d,, + d2r dl2 + dt) < d1} + d2l, (4)
where i and / are any OTUs (3 < i < j < n). For actual data, the above
condition may not hold because of disturbance of additivity of distances.
Sattath and Tversky (1977), and Fitch (1981) then proposed algorithms to
construct a tree by maximizing the number of cases in which condition (4)
holds. In practice, it is easier to use the ST algorithm of Sattath and Tversky
(1977).
Neighbor-Joining (NJ) Method
The principle of this method is the same as that of the ME method, but
the computational process is much simpler (Saitou and Nei, 1987). Com-
parison of different topologies is built in the algorithm, and the final tree
with both topology and branch lengths is automatically produced. Swofford
and Olson (1990) recently stated that this method is for estimating an
additive tree, but this statement is incorrect.

Discrete-Character Data
Maximum Parsimony (MP) Method
In this method, the DNA (or amino acid) sequences of ancestral species
are inferred from those of extant species, considering a particular tree
topology, and the minimum number of evolutionary changes that are re-
quired to explain all the observed differences among the sequences is
computed. This number is obtained for all possible topologies, and the
topology which shows the smallest number of evolutionary changes is cho-
sen as the final tree. This method is used mainly for finding the topology
of a tree, but branch lengths can be estimated under certain assumptions
(Fitch, 1971).
94 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

When the MP method is applied to morphological characters, it is cus-

tomary to assume that the primitive and derived character-states are known.
In the case of molecular data, this assumption generally does not hold,
and different character-states are often reversible. It is, therefore, impor-
tant to use the MP method, which permits reversible mutations (Eck and
Dayhoff, 1966; Fitch, 1971). In numerical taxonomy, this type of MP method
is sometimes called the Wagner parsimony method (Farris, 1970).
Evolutionary Parsimony (EP) Method
This method is primarily applied to four DNA (or RNA) sequences and
utilizes information on the transition/transversion bias in nucleotide sub-
stitution (Lake, 1987). The actual procedure is to compute three quantities,
X, Y, and Z, which are functions of the numbers of certain nucleotide
configurations among the four DNA sequences, and to determine which
of the three quantities is significantly different from 0. If only one of them
is significant, the tree topology corresponding to the quantity is regarded
as the correct one. If two or all of X, Y, and Z are significant, the splitting
pattern of DNA sequences is considered unresolvable. A procedure for
extending this method to the case of five or more sequences is presented
by Lake (1988).

Maximum Likelihood (ML) Method

In this method, the nucleotides of all DNA sequences at each nucleotide
site are considered separately, and the log-likelihood of having these nu-
cleotides are computed for a given topology by using a particular probability
model (Felsenstein, 1981a). This log-likelihood is added for all nucleotide
sites, and the sum of the log-likelihood is maximized to estimate the branch
length of the tree. This procedure is repeated for all possible topologies,
and the topology that shows the highest likelihood is chosen as the final
one.

METHODS OF COMPUTER SIMULATION

Generation of Simulated Sequences

In the case of DNA sequences, we know the basic rules of their evolutionary
change (Nei, 1987a: 79-88). There are two types of changes (i.e., nucleo-
tide substitution and deletion/insertion). The latter type of change occurs
haphazardly, and it is difficult to quantify the amount of change for a given
evolutionary time. Therefore, deletions and insertions are often neglected
in the reconstruction of phylogenetic trees. By contrast, the pattern of
nucleotide substitution has been studied extensively, and there are several
mathematical models that predict the evolutionary change of DNA se-
quences. These models are, of course, approximations to real changes of
DNA sequences, but their predictions are quite accurate unless the number
RELATIVE EFFICIENCIES OF DIFFERENT TREE-MAKING METHODS 95

of nucleotide substitutions per site is very high. Therefore, it is possible

to simulate the evolutionary change of DNA sequences.
The basic information for modeling the evolutionary change of DNA
sequences is the probability of change of a nucleotide (A, T, C, or G) to
another nucleotide during a short period of time (e.g., one year, one
generation, 1,000 years, etc.). Let Av be the probability that the z'-th nu-
cleotide changes to the j-ih nucleotide during the unit evolutionary time.
We then have the following transition matrix:

Wliere

and
Let g, be the probability that a given nucleotide site in a DNA sequence
is occupied by the i-th nucleotide at a given evolutionary time, and g be
the column vector of g1, g2, g3, and g4. We designate g at time f by g,. It
is then possible to compute gf by g, = Mlg0, where M is the matrix given
by (5) and g0 is the vector g at time 0. Therefore, if g0 is given for all
nucleotide sites of the DNA sequence under consideration, one can gen-
erate a DNA sequence at time t. Note that all nucleotide changes are
probabilistic. Therefore, the sequence generated for time t may vary from
replication to replication.
When A y 's are all the same and are equal to A, and the initial sequence
is a random sequence, the evolutionary change of DNA sequence becomes
much simpler. Consider two DNA sequences that diverged t time-units
ago. The probability that the two sequences have different nucleotides at
a given site is given by

(Jukes and Cantor, 1969). Similarly, the probability that a descendent

sequence is different from the ancestral sequence at a site becomes

Therefore, using pseudorandom numbers, one can easily generate a DNA

sequence at any time. We call this the one-parameter model.
Incidentally, equation (6) is useful for deriving an estimator of the total
number of nucleotide substitutions per site (d = 21t) between two se-
quences that diverged t time-units ago. It is given by
96 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

where p is the proportion of different nucleotides between the two se-

quences compared.
Actual data on nucleotide substitution indicate that transitional changes
(A ?^ G and T <=i C) are more frequent than transversional changes (all
other changes). Considering this difference, Kimura (1980) developed a
two-parameter model in which transitional and transversional changes can
be treated separately. If we use this model, the probability that two DNA
sequences that diverged t time-units ago show a transitional difference at
a given site becomes

whereas the probability that they show a transversional difference is

Here a and j8 denote the rates per unit of evolutionary time of transitional
and transversional changes, respectively. These equations can be used for
simulating the evolutionary change of DNA sequences. The two-parameter
model can also be used for estimating the number of nucleotide substitu-
tions between two sequences. It can be estimated by

(Kimura, 1980).
The transition probabilities A,/s in equation (5) have been estimated for
a number of groups of genes (e.g., Brown et al., 1982; Gojobori et al.,
1982; Aquadro and Greenberg, 1983; Graur, 1985; Saitou, 1987). Gen-
erally speaking, transitional changes are more frequent than transversional
changes, but actual substitution patterns are more complicated than the
two-parameter model. For this reason, several more complicated models
have been developed (see Nei, 1987a). However, both the one- and two-
parameter models are known to be quite robust. For estimating the number
of nucleotide subsitutions (d), equation (8) gives a good estimate, provided
that d is less than 0.5 and the transition/transversion bias is not extreme,
whereas the two-parameter model seems to work even for a higher value
of d if the number of nucleotides examined is large. For this reason, the
one- or two-parameter model is often used for computer simulation. If this
is not satisfactory, the matrix method presented earlier can be used (Jin
and Nei, 1990).
Recent data (Britten, 1986; Li et al., 1987b) suggest that the rate of
nucleotide substitution varies with evolutionary lineage. This can easily be
simulated by changing A,/s (or a and /3) with evolutionary lineage. Math-
ematically, however, the same result is obtained by changing t rather than
Ay's, and this is usually much simpler.
RELATIVE EFFICIENCIES OF DIFFERENT TREE-MAKING METHODS 97

Another important aspect of nucleotide substitution is that the substi-

tuion rate varies among nucleotide sites (Shoemaker and Fitch, 1989). The
first and second nucleotide positions of codons are usually less variable
than the third positions. Nucleotides in the active center regions of genes
are also usually less variable (Kimura, 1983). In some exceptional genes
such as major histocompatibility complex genes, the active center evolves
faster (Nei and Hughes, 1991). This problem can also be resolved by
considering spatially varying models of nucleotide substitution (Gojobori
et al., 1982; Tateno et al., 1982; Olson, 1987; Jin and Nei, 1990). A
commonly used method is to assume that Ay's (or a and /3) vary according
to a gamma-distribution or a lognormal distribution.

Model Trees and Estimation of the Trees

As mentioned earlier, if a model tree is given, we can simulate the evo-

lutionary changes of DNA sequences following the tree and generate the
present-day sequences. Once these sequences are obtained, a tree is re-
constructed by each tree-making method under consideration. In the case
of distance methods, the evolutionary distance between each pair of se-
quences must be computed. This distance can be computed by various
methods. When the number of nucleotide differences per site (p) is small
for all pairs of sequences, this number can be used as a distance measure.
We call this the p distance. When p is large, however, p is not a good
estimate of the number of nucleotide substitutions (d) because of backward
and parallel substitutions. In this case, d, is often estimated by equation
(8). We call this the d distance in this chapter, whereas the d value estimated
by equation (11) will be called the Kimura distance. However, there are
many other ways to estimate d, as mentioned earlier. Obviously, the most
desirable distance measure is the one that gives the best estimate of the
total number of nucleotide substitutions. In the case of discrete-character
models, the configuration of nucleotides among all sequences is considered
at each nucleotide site.
Since the evolutionary change of nucleotides is stochastic, the DNA
sequence generated for a given species varies from replication to replica-
tion. Therefore, computer simulation is repeated many times for a given
model tree, and the agreement between reconstructed trees and the model
tree is statistically determined.
There are two types of deviations of a reconstructed tree from the model
tree. One is topological error, and the other is the deviation of estimated
branch lengths from the true lengths. Topological error can be evaluated
by two measures. One is the proportion of replications in which the correct
topology is obtained. This is called the probability of obtaining the correct
topology (Pc). The other measure is the topological distance between a
reconstructed tree and the model tree. This distance, Tateno et al.'s (1982)
distortion index (dT), is measured by Robinson and Foulds' (1981) index.
Roughly speaking, this index is twice the number of branch interchanges
98 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

that are required to transform the topology of a tree to that of_ another.
When simulation is replicated, the mean topolpgical distance (dT) for all
replications is used as a criterion. In general, dT is highly negatively cor-
related to Pc. Therefore, many authors have used Pc only.
Branch-length errors can also be measured by several different methods
(Tateno et al., 1982). In this chapter, however, this problem will not be
considered.

RESULTS FROM COMPUTER SIMULATION

As mentioned earlier, relative efficiencies of tree-making methods have

been studied by many authors. Each author or group of authors has studied
a different set of tree-making methods. However, several tree-making
methods were examined repeatedly by different groups of authors; in these
cases the results obtained were usually consistent. In this review, the gen-
eral conclusions obtained from these studies (without going into details)
are presented.
Computer simulation of the reconstruction (estimation) of phylogenetic
trees is quite time consuming, particularly when the number of DNA se-
quences is large. Therefore, the number of DNA sequences used is usually
six to eight, though some authors (e.g., Tateno et al., 1982) examined up
to 32 sequences. The shape (topology and branch lengths) of the model
tree examined varies among authors, but they can be divided roughly into
the two types (A and B) given in Figure 6-1. Some methods show a better
performance in topology A, and others in topology B. In general, however,

Figure 6-1 Model trees (A) and (B) under the assumption of constant rate of nu-
cleotide substitution. U is the expected number of substitutions per site from the
common ancestral sequence to the extant sequences.
RELATIVE EFFICIENCIES OF DIFFERENT TREE-MAKING METHODS 99

the relative merits of different tree-making methods are similar for both
topologies (A and B).
In the following discussion, the cases of constant and varying rates of
nucleotide substitution will be considered separately.

Constant Rate of Substitution

Number of Nucleotide Substitutions—Effect on Recovery of Correct
Topology
When the expected number of nucleotide substitutions per site from the
common ancestral sequence to the extant sequence (U) is small (see Fig.
6-1), the FM, DW, and Tateno et al. 's modified Farris (MF) methods are
more efficient than UPGMA in recovering the correct topology. However,
when U is large and each branch length is large, UPGMA is better than the
former methods.
This was first shown by Tateno et al. (1982), but more general results
were obtained by Sourdis and Krimbas (1987). Table 6-1 shows the prob-
abilities of obtaining the correct tree (Pc) for various tree-making methods
when model tree A in Figure 6-1 is used. When the number of nucleotides
used is small and U is small, any method gives a very small Pc value, as
expected. This is true whether rooted or unrooted trees are constructed.
However, as long as U is equal to 0.1 or less, the FM, DW, and MF
methods are better than UPGMA in obtaining the correct rooted or un-
rooted tree. (The root for the trees obtained by the FM, DW, and MF
methods was given as the midpoint of the longest route connecting two
sequences.) The former three methods show more or less the same Pc
value. When U is large (e.g., U = 0.4), however, UPGMA often shows
a higher Pc value than the other methods. Somewhat similar results were
also obtained by Blanken et al. (1982). Table 6-1 shows that Pc's for rooted
trees are considerably smaller than those for unrooted trees, particularly
when the FM, DW, and MF methods are used. This indicates that a sub-
stantial amount of error in constructing a rooted tree occurs at the time
of rooting.

Comparison of Efficiency of ST and NJ Methods with Other Methods

The ST and NJ methods are almost always better than the UPGMA, DW,
MF, and TD methods whether the U value is small or large.
Table 6-2 shows the results of another computer simulation for a different
set of tree-making methods. This represents the case where U is relatively
small (U = 6a + b = 0.1 in model tree A, and U = 1.5a + c = 0.085
in model tree B of Figure 6-1). In this case, the MF and DW methods are
again better than UPGMA in obtaining the correct tree. When tree A is
used, they are as efficient as the TD (Li's algorithm), ST, and NJ methods.
However, when tree B is used, they are less efficient than the latter meth-
ods. Table 6-3 shows the Pc values for the case of a larger U value ([/ =
0.52 in tree A, and U = 0.465 in tree B) and a larger number of nucleotides
Table 6-1 Probabilities of Obtaining the Correct Topology (Pc x 100) for Rooted and Unrooted Trees for Four Different Tree-Making
Methods in the Case of a = fa in Model Tree A of Figure 6-1*
Rooted Trees Unrooted Trees
U R UPGMA FM DW MF UPGMA FM DW MF
m = 300
0.03 250 4 10 6 10 9 24 21 26
0.05 275 12 19 15 17 23 49 44 48
0.10 275 34 42 34 39 51 70 72 71
0.20 375 56 55 41 52 67 79 82 79
0.40 175 69 53 40 50 83 76 77 73
m = 900
0.03 275 38 43 42 44 48 82 80 81
0.05 275 60 64 61 64 74 94 95 94
0.10 275 88 84 75 84 92 99 99 99
0.20 250 96 91 85 91 98 98 100 99
0.40 100 99 95 82 95 100 99 99 100
m = 1,500
0.03 275 60 66 65 66 74 95 96 98
0.05 275 84 84 81 84 88 99 100 100
0.10 275 92 94 90 94 95 100 100 100
0.20 275 99 99 93 99 99 100 100 100
0.40 125 100 95 95 99 100 100 100 100
The abbreviations used are: U, expected number of nucleotide substitutions per site from the ancestral sequence to the extant sequence, m, number of nucleotidcs used,
R, number of replications; FM, Fitch and Margoliash's method, DW, distance Wagner method, MF, modified Farris method Trees were constructed by using the d distance
(Adapted from Sourdis and Knmbas, 1987.)
Table 6-2 Probabilities of Obtaining the Correct Unrooted Tree (Pc x 100) and Average Indices (dr) of Topological Errors (in
Parentheses) for Six Tree-Making Methods for the Case of a = 0.01, b = 0.04, and c = 0.07 in Model Trees in Figure 6-1*
Model Tree A Model Tree B
m 300 600 900 300 600 900
UP:/) 14 (3.18) 36 (1.72) 58 (0.98) 14 (4.54) 36 (2.74) 51 (1.68)
Vf:d 15 (3.18) 34 (1.74) 56(1.04) 13 (4.56) 35 (2.70) 52 (1.60)
UF:p 39 (1.76) 73 (0.58) 95 (0.10) 24 (2.86) 51 (1.30) 67 (0.76)
MF:d 38 (1.92) 72 (0.62) 95 (0.10) 19 (2.94) 48 (1.42) 64 (0.86)
DW:p 42 (1.70) 75 (0.54) 96 (0.08) 26 (2.36) 55 (1.12) 79 (0.48)
DW:d 37 (1.74) 74 (0.58) 95 (0.10) 28 (2.36) 58 (1.06) 79 (0.46)
TD:p 41 (1.58) 71 (0.70) 94 (0.12) 40 (2.04) 70 (0.78) 90 (0.22)
TD:d 36 (1.84) 66 (0.82) 89 (0.24) 39 (2.10) 70 (0.78) 90 (0 26)
ST.p 48 (1.26) 75 (0.54) 97 (0.06) 45 (1.66) 75 (0.62) 91 (0.22)
ST: d 44 (1.48) 70 (0.62) 96 (0.08) 43 (1.62) 74 (0.64) 91 (0.22)
NJ:p 48 (1.36) 76 (0.54) 97 (0.06) 46 (1.64) 76 (0.60) 91 (0.20)
NJ: d 41 (1.60) 70 (0.62) 96 (0.08) 45 (1.62) 75 (0.60) 91 (0.20)
'The abbreviations used are: UP, UPGMA; MF, modified Farris method; DW, distance Wagner method; TD, transformed distance method, Li's algorithm was used; ST,
Neighborhness, Sattath and Tversky's method; NJ, neighbor-joining method, p, trees reconstructed from p distances, d, trees reconstructed from d distances; m, number
of nucleotides used. Number of replications (R) is 100. (Adapted from Saitou and Nei, 1987.)
Table 6-3 Probabilities of Obtaining the Correct Unrooted Tree (Pc x 100) and Average Indices (dr) of Topological Errors (in
Parentheses) for Six Tree-Making Methods for the Case of a = 0.03, b = 0.34, and c = 0.42 in Model Trees in Figure 1 *
Model Tree A Model Tree B
m 500 1,000 2,000 500 1,000 2,000
UP:p 9 (3.78) 27 (2.10) 62 (0.86) 10 (5.20) 18 (3 76) 54 (1 32)
UP:d 9 (3.78) 27(2.10) 62 (0.88) 11 (5.30) 18 (3.74) 55 (1.26)
MF:/> 15 (4.02) 41 (1.82) 62 (0.92) 3 (5.68) 17 (3.64) 28 (2.40)
MF:d 13 (4.42) 34 (2.14) 55 (1.14) 3 (5.72) 13 (3.80) 26 (2.48)
DW:p 16 (3.78) 46 (1.54) 63 (0.82) 4 (5.42) 18 (3.28) 41 (1.72)
DW:rf 15 (4.22) 40 (1.96) 58 (0.98) 5 (5.50) 18 (3.48) 35 (1.82)
TD:p 3 (4.26) 37 (2.00) 53 (1.18) 15 (4.48) 28 (2.98) 70 (0.90)
TD:d 3 (4.84) 25 (2.60) 39(1.66) 12 (4.72) 27 (3.06) 66(1.02)
ST.p 10 (3.56) 44 (1.62) 68 (0.76) 13 (4.00) 36 (2.34) 74 (0.62)
ST:rf 6 (4.06) 40 (1.82) 56 (1.04) 10 (4.32) 34 (2.34) 71 (0.72)
NJ:p 11 (3.70) 44 (1.68) 67 (0.80) 13 (4.46) 34 (2.38) 75 (0.62)
NJ: d 5 (4.24) 38 (2.00) 57 (1.06) 14 (4 44) 32 (2.42) 73 (0.72)
'Abbreviations used are the same as those m Table 6-2 R = 100 (Adapted from Saitou and Nei, 1987.)
RELATIVE EFFICIENCIES OF DIFFERENT TREE-MAKING METHODS 103

examined. In this case, the MF and DW methods are as efficient as the

ST and NJ methods when tree A is used, but are less efficient than the
latter when tree B is used. The TD method is slightly less efficient than
the ST and NJ methods for both trees A and B.
It is known that the principle of triangle inequality applies to the p
distance but not to the d distance. Therefore it is interesting to compare
the Pc values for the trees obtained by using the p and d distances. Tables
6-2 and 6-3 show that there is no real difference in Pc between the p and
d distances when UPGMA is used. In other tree-making methods, the p
distance tends to give a slightly higher Pc than the d distance. However,
the difference between them is generally small, except when the number
of nucleotides examined (m) is small. As will be shown later, when the
rate of nucleotide substitution varies with evolutionary lineage, and p or
d is large, the d distance shows a better performance than the p distance.
Comparison of the MP Method with the ST, NJ, DW, and TD Methods
The MP method generally has a smaller Pc value than the ST and NJ meth-
ods. However, when U is small and the number of nucleotides examined is
very large, the MP method is as good as or slightly better than the latter.
As mentioned earlier, the MP method requires examination of all pos-
sible topologies, and since the number of possible topologies rapidly in-
creases as the number of OTUs increases, computer simulation has been
done for a relatively small number of OTUs. For example, the number of
possible topologies is 105 when the number of OTUs is six, whereas it is
10,395 when the number is eight. Table 6-4 shows the Pc values for the
model trees A and B in Figure 6-2. When model tree A is used and the
number of nucleotide substitutions is small (U = 0.05), the NJ and TD
methods are better than the MP and DW methods provided that the number
of nucleotides examined is <600. When the number of nucleotides is large

Table 6-4 Probabilities of Obtaining the Correct Tree (Pc x 100) for the
Maximum-Parsimony (MP) and Other Tree-Making Methods for Model Trees A
and B in Figure 6-2*
Constant Rate (Tree A) Varying Rate (Tree B)
m 300 600 1,200 300 600 1,200
U = 0.05
MP 46.3 79.0 97.0 16.7 39.3 58.0
NJ 58.7 82.0 95.3 30.0 46.7 63.3
DW 47.7 72.0 92.0 192 35.0 52.0
TD 54.7 82.0 95.0 25.3 42.3 62.3
U = 0.5
MP 44.7 69.3 84.0 16.7 22.3 31.0
NJ 62.7 84.3 97.0 25.0 34.7 49.3
DW 49.0 73.0 91.3 16.0 30.7 40.0
TD 64.3 84.3 96.0 25.0 33.3 48.0
'Abbreviations used are the same as those m Table 6-2. MP, maximum parsimony. K = 300. (Adapted
from Sourdis and Nei, 1988.)
104 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Figure 6-2 Model trees for six DNA sequences. (A) Case of constant rate of nu-
cleotide substitution. (B) Case where the substitution rate varies with evolutionary
lineage.

(m = 1,200), however, the MP method seems to be as good as or slightly

better than the other methods.
When the number of substitutions per site is large (i.e., U = 0.5), the
story is different. In this case, even for m = 1,200, the MP method shows
a smaller Pc value than the NJ and DW methods. This is apparently because
there are many backward and parallel substitutions in this case. Distance
methods are not affected by these substitutions as long as the distances are
correctly estimated.
The low performance of the MP method when m and U are both small
is due to the fact that in this method only so-called "informative sites" are
used, whereas in distance methods information for all variable sites is used.
(Note that noninformative variable sites in parsimony methods are actually
informative for constructing a tree in distance methods). When m and U
are small, the number of "informative sites" is also small. Therefore, the
MP method tends to produce many equally parsimonious trees, which often
(but not always) include the correct tree (Sourdis and Nei, 1988). In our
study, the case where the correct tree and other tied trees were obtained
was not included in the probability of obtaining the correct tree. This is
part of the reason why Pc became smaller in the MP method than in the
NJ and other methods. When m is large, however, the number of "in-
formative sites" becomes large, thus, Pc increases.
In this connection, one might argue that Pc is not an appropriate measure
for comparing the MP method with others because many tied trees, in-
cluding the correct one, are often produced in this method. When there
are many tied trees, it is customary to construct a consensus tree with
multifurcating branches in parsimony methods. A multifurcating consensus
tree can also be constructed with distance methods (e.g., Tajima, 1990).
However, a multifurcating tree is clearly wrong in the present case, and it
is not clear whether the comparison of consensus trees obtained by par-
simony and distance methods is actually meaningful.
When there are many tied trees obtained, a more reasonable way of
comparison is to compare ~dT for the MP and other methods. If many tied
RELATIVE EFFICIENCIES OF DIFFERENT TREE-MAKING METHODS 105

trees are close to the correct one, d r is expected to be close to 0; otherwise

it will be high. Since Sourdis and Nei did not compute dT, Li Jin and I
performed a small-scale computer simulation to evaluate this value using
model tree A in Figure 6-2. The results obtained for the MP and NJ methods
are presented in Table 6-5. The Pc values in this table are essentially the
same as those for the corresponding cases of Table 6-4. Table 6-5 also
includes the probability of obtaining the correct tree and other tied trees
(P2) and the probability of obtaining only incorrect tree(s) (P3). As in
Sourdis and Nei's study, P2 for the MP method is large when m is small
but decreases as m increases. For the NJ method, it is always 0. This result
is slightly different from that of Sourdis and Nei (1988), because in their
simulation some numbers with many digits were rounded to save computer
time. _
Table 6-5 shows that even if we use dT as the criterion of comparison,
the MP method is generally inferior to the NJ method. The only exceptions
are the cases of m = 600 and m = 1,200 with U = 0.05. Therefore, our
conclusion remains the same, whether we use Pc or dr as the criterion.
Comparison of ML Method with the MP, FM, ME, and N] Methods
The ML method is generally more efficient than the MP and FM methods
but is slightly less efficient than the ME and NJ methods.
The Pc values for the ML method for model trees A and B in Figure 6-
3 are presented in Table 6-6 in comparison with those for the MP, FM,
ME, and NJ methods. The trees for the distance methods were constructed
by using both p and d distances. It is clear that the FM method is poorest
in obtaining the correct tree among all the methods examined here. The
ML method is better than the MP method except for the case of U = 0.5
in tree A, but is slightly less efficient than the ME and NJ methods when
m = 300. When m = 600, it has a Pc value similar to that of the latter
two methods.

Table 6-5 Probabilities of Obtaining the Correct Unrooted Tree (Pc x 100), the
Correct Tree and Other Tied Trees (P2 x 100) and Incorrect Trees (P3 x 100),
and Average Indices of Topological Errors (c/r) for the MP and NJ Methods*
U = 0.05 U = 0.5
m 300 600 1,200 300 600 1,200
Maximum Parsimony (MP)
PC 44 79 96 40 63 88
Pi 41 14 3 13 8 2
P3 15 7 1 47 29 10
dT 0.96 0.31 0.05 1.30 0.72 0.22
Neighbor-Joining (NJ)
Pc 62 86 97 55 79 96
P2 0 0 0 0 0 0
P3 38 14 3 45 21 4
dT 0.87 0.31 0.07 1.05 0.41 0.07
*PC is the same as PI of Sourdis and Nei (1988). Model tree A in Figure 6-2 was used. R = 300.
106 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Figure 6-3 Model trees used by Saitou and Imanishi's (1989) computer simulation.

It should be noted that this computer simulation was conducted under

the conditions which are favorable for the ML method, but do not com-
pletely satisfy the assumptions of this method (Saitou and Imanishi, 1989;
N. Saitou, personal communication). Therefore, it seems that a slightly
lower performance of the ML method compared with the ME and NJ
methods is due to a small violation of the assumptions of this method. This
suggests that the ML method is sensitive to violations of its assumptions,
at least in this case.
Incidentally, Table 6-6 shows that the ME and NJ methods give essen-
tially the same Pc value, as mentioned earlier. Indeed, Saitou and Imanishi
(1989) showed that these two methods usually reconstruct the same tree
from a given set of data. This indicates that the NJ method usually finds
the minimum evolution tree, though its procedure is very simple.
Recently, Rohlf and Wooten (1988) reported results of their computer
simulations on the gene frequency maximum likelihood (GFML) method
(Felsenstein, 1981b). They considered situations where the assumptions of
the GFML method hold as closely as possible, yet their results showed
that UPGMA is generally better than the GFML method for gene-fre-
quency data. Curiously, however, these authors concluded otherwise, con-
sidering the possibility that the performance of the ML method would
increase as the number of genetic loci used increases. This conclusion is
not warranted because the performance of distance methods also increases
as the number of loci increases. A similar computer simulation was also
conducted by Kim and Burgman (1988). However, since this simulation
was conducted under the same assumptions (Gaussian process without
RELATIVE EFFICIENCIES OF DIFFERENT TREE-MAKING METHODS 107

Table 6-6 Probabilities of Obtaining the Correct Tree (Pc x TOO) for the
Maximum Likelihood (ML) and Other Tree-Making Methods for Model Trees A
and B of Figure 6-3*
p Distance d Distance
MP ML FM ME NJ FM ME NJ
Model Tree A
U = 0.05
m = 300 34 38 26 36 42 26 40 40
m = 600 76 80 58 80 78 58 80 82
U = 0.50
m = 300 60 48 30 56 58 22 42 46
m = 600 84 70 54 92 92 40 82 82
Model TreeB
U = 0.05
m = 300 54 62 42 68 68 40 68 70
m = 600 84 88 56 82 84 56 80 86
U = 0.50
m = 300 48 56 34 70 68 30 60 60
m = 600 58 76 36 76 76 36 70 70
•Abbreviations used are the same as those of previous tables; ML, maximum likelihood; ME; minimum
evolution method. R = 100 (Adapted from Saitou and Imanishi, 1989.)

mutation and selection) as those of the GFML model, the results of the
simulation are not very meaningful. Since actual gene-frequency changes
almost never follow the assumptions of the GFML model, the simulation
should have been conducted by using a more realistic model of gene fre-
quency change due to mutation, selection, and genetic drift (see Nei et
al., 1983).

Varying Rate of Substitution

Comparison of UPGMA with Other Tree-Making Methods
When the rate of nucleotide substitution varies with evolutionary lineage,
UPGMA is worse than most other tree-making methods.
Since UPGMA requires the assumption of a constant rate of evolution,
it is expected to have a poor performance in obtaining the correct tree
when the rate varies with evolutionary lineage. Indeed, computer simu-
lations have shown that it is very poor compared with other methods in
this case (Table 6-7). The Pc values in this table were obtained by using
model trees given in Figure 6-4.
Comparison Among Distance Methods
Among the distance methods available now, the ST, NJ, and ME methods
generally show a better performance than other methods. The ST, NJ, and
ME methods are nearly equally efficient.
This can be seen from Table 6-7 and Table 6-8. The Pc values in Table
6-8 were obtained by using model trees C and D in Figure 6-3. Note that
the total number of nucleotide substitutions between the two most distantly
108 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Table 6-7 Probabilities of Obtaining the Correct Unrooted Trees (Pc x 100) and
Average Indices (dT) of Topological Errors (in Parentheses) for the Case of Varying
Rate of Nucleotide Substitution*
Model Tree A Model Tree B
UP: 0 (8.06) 0 (9.74)
MF: 77 (0.50) 57 (1.46)
DW: 69 (0.72) 59 (1.26)
TD: 46 (1.30) 45 (1.68)
ST: 77 (0.50) 69 (0.82)
NJ: 75 (0.56) 72 (0.78)
•Abbreviations used are the same as those in Table 6-2 Model trees A and B are given in Figure 6-4 m
= 600; R = 100. p distances were used. (Adapted from Saitou and Nei, 1987.)

related sequences (dm) is relatively small in Table 6-7, whereas Table 6-8
includes both cases, of a small dm and a large dm. When dm is small, the
difference in Pc between the p and d distances is relatively small, though
in the case of model trees C and D of Figure 6-3, the d distance gives a
higher Pc value than the p distance (Table 6-8). At any rate, Tables 6-7
and 6-8 show that the Pc values for the ST, NJ, and ME methods are
generally higher than other distance methods.
Table 6-8 shows that when dm is large, the d distance gives a higher Pc
value than thep distance. This is because the p distance is seriously affected
by backward and parallel substitutions, whereas the d distance gives an
appropriate estimate of the total number of substitutions. As mentioned
earlier, many distance methods are capable of recovering the correct tree
as long as the total number of substitutions is correctly estimated.
Comparison of the MP Method with Other Methods
The MP method is worse than several distance methods such as the NJ, ST,
and ME methods and the ML method.

Figure 6-4 Model trees used for the simulations in Table 6-7. The value given to
each branch represents the expected number of nucleotide substitutions per site for
that branch.
RELATIVE EFFICIENCIES OF DIFFERENT TREE-MAKING METHODS 109

Table 6-8 Probabilities of Obtaining the Correct Unrooted Tree (Pc x 100) for
the MP, ML, and Other Methods in the Case of Varying Substitution Rate*
p Distance d Distance
MP ML FM ME NJ FM ME NJ
Model Tree C
a = 0.01
300 bp 64 78 26 56 56 34 72 72
600 bp 90 98 42 80 80 68 92 92
a = 0.05
300 bp 24 92 0 2 2 22 68 68
600 bp 20 100 0 0 0 60 96 96
Model TreeD
a = 0.01
300 bp 68 80 44 64 64 64 74 74
600 bp 94 96 56 90 90 88 92 92
a = 0.05
300 bp 26 96 0 6 4 30 78 78
600 bp 46 100 0 10 6 64 100 100
"Abbreviations used are the same as those in previous tables; bp, base pairs. R = 50. Model trees C and
D are those in Figure 6-3. (Adapted from Saitou and Imanishi, 1989.)

This statement is supported by the "varying rate" part of Table 6-4 and
Table 6-8. In all cases examined here, the MP method always gives a smaller
Pc value than the NJ, ME, and ML methods. However, it is better than
the FM method when dm is small (a = 0.01). The inferiority, of the MP
method is explained by Felsenstein's (1978) finding that when evolutionary
rate varies with lineage, two OTUs that have long branches tend to form
a cluster in the MP method even if this is not the correct branching pattern.
This does not happen in the NJ and ME methods as long as the distances
between sequences are correctly estimated. This also generally does not
occur with the ML method (Hasegawa and Yano, 1984; Saitou, 1988; Saitou
and Imanishi, 1989).
Some authors (e.g., Sober, 1985) claimed that the MP method is better
than other methods because no assumption is necessary about the process
of evolutionary change. This claim is not supported by simulation studies.
As mentioned earlier, for the MP method to work well, approximate con-
stancy of evolutionary rate and a small number of nucleotide substitutions
per site are necessary.
Comparison of the ML Method with the MP, ME, and NJ Methods
The ML method is generally better than the MP method and is nearly the
same as or slightly better than the ME and NJ methods in obtaining the
correct tree. However, the performance of this method should be investigated
more carefully by considering the cases where the underlying assumptions
are violated.
Table 6-8 clearly shows that the ML method is better than the MP method
when the evolutionary rate varies extensively. A similar result was also
obtained by Hasegawa and Yano (1984), and Saitou (1988) for the case of
four sequences. Table 6-8 also shows that the ML method is slightly better
110 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

than the ME and NJ methods. This is so despite the fact that the pattern
of nucleotide substitution simulated by Saitou and Imanishi (1989) was
slightly different from the assumptions of the ML method used.
Nevertheless, it should be noted that the actual pattern of nucleotide
substitution is much more complicated than the simple model used in the
ML method (Gojobori et al., 1982b; Jin and Nei, 1990), so the general
applicability of the above conclusion is questionable. Since the ML method
is expected to be more sensitive to violation of the assumptions than are
some distance methods, it is hoped that this method will be examined more
carefully, considering realistic patterns of nucleotide substitution, partic-
ularly variation in substitution rate among different nucleotide sites. Pre-
viously, Saitou (1988) showed that in the case of a = c = 1.0 and b = d
= 0.1 in model tree G of Figure 6-5, the NJ method (Pc = 0.74) is
significantly better than the ML method (Pc = 0.43). Recently, however,
Hasegawa and Saitou (personal communication) improved the perform-
ance of the ML method using a slightly different probability model. This
indicates that it is very important to use a proper probability model in the
ML method.
In the case of distance methods, the problem of substitution pattern
becomes important when one estimates the distances between different
sequences rather than when a tree is estimated from distance data. As long
as the total number of nucleotide substitutions is correctly estimated, many
distance methods are capable of producing correct trees. There are several
statistical methods of estimating distance that are quite robust unless the
distances are very large (see Nei, 1987a: 72). This problem will be discussed
later.
Comparison of the EP Method with the MP and NJ Methods
The EP method is inferior to the MP method when nucleotide substitution
occurs at random among the four nucleotides (one-parameter model) and

Figure 6-5 Unrooted model trees for four DMA sequences. G; general case. S,,
S2, S3; model trees used by Jin and Nei's (1990) computer simulation. The value
given to each branch represents the expected number of nucleotide substitutions
per site for that branch.
RELATIVE EFFICIENCIES OF DIFFERENT TREE-MAKING METHODS 111

the rate of nucleotide substitution is nearly constant. However, if there is a

transition/transversion bias and the rate of substitution varies extensively
with evolutionary lineage, the EP method is better than the MP method.
The NJ method is almost always superior to the EP method if a proper
distance measure is used.
As mentioned earlier, the EP method was developed primarily to be
applied to the case of four sequences. Therefore, all computer simulations
have been performed for this case. The model tree used is generally of the
form given in Figure 6-5. Table 6-9 shows the Pc values for the MP, NJ,
and EP methods for the case of Kimura's two-parameter model with model
trees Slt S2, and S3 in Figure 6-5. Kimura's model satisfies the assumption
of the EP method (Cavender, 1989; Jin and Nei, 1990). Therefore, the EP
method is supposed to show a good performance in this case. There are
three different sets of Fc's for the NJ method: NJP, NJD, and NJK rep-
resent the cases where the p distance, d distance, and Kimura distance
[equation (11)], respectively, are used. Since Kimura's two-parameter model
was used in generating DNA sequences, NJK is expected to show a better
performance than NJP and NJD except when the proportion of transitional
changes [B = a/(a + 2/?)] is Vs. This is indeed the case for almost all
parameter sets examined in Table 6-9.
Comparison of the MP and EP methods in Table 6-9 shows that when
all branch lengths of the model trees are relatively short (Sj), or when the
ratio of branch b to branch a is relatively small (S2), the former is better
than the latter even for the case of a high proportion of transitional changes
(B = 0.9). Otherwise (S3), the latter is better than the former. Essentially
the same results were obtained for various other sets of parameters (Li et
al., 1987a; Jin and Nei, 1990). The performance of the NJ method is
expected to depend on the distance measure used, as mentioned earlier.
When the appropriate distance (Kimura distance for this case) is used, the
NJ method is better than both the MP and EP methods. The same results
were obtained for several other sets of parameters as well as for the set of
substitution rates estimated from actual data (not the two-parameter model)
(Jin and Nei, 1990).
Recently, Sidow and Wilson (1990) modified the EP method, taking into
account the unequal frequencies of the four nucleotides (A, T, C, and G)
in the sequences. However, since the basic principle of the EP method
remains unchanged, our conclusion is expected to apply to this modified
version as well.

Effects of Variation in Substitution Rate Among Different Nucleotide

Sites
So far we have assumed that every nucleotide site evolves independently
and that the pattern of nucleotide substitution is the same for all sites. This
assumption is obviously not satisfied with actual data. In the case of protein-
coding genes, the third nucleotide positions of codons usually evolve faster
Table 6-9 Probabilities of Obtaining the Correct Unrooted Tree (Pc x 100) for the Two-Parameter Model of
Nucleotide Substitution in the Cases of Model Trees S,, S2, and S3 in Figure 6-5*
Tree-Making Method EP (x 2 )
Model Trees B MP NJP NJD NJK EP X y Z
S
a
lt 0.33 100 98 100 100 99 68 i 1
0.60 100 97 100 100 91 45 2 2
0.90 88 85 99 100 61 12 0 0
S t
•J2 0.33 91 88 100 100 63 12 6 6
0.60 86 84 98 97 51 6 4 4
0.90 67 69 84 86 40 5 1 1
c §
3 0.33 4 0 98 98 72 26 6 4
3
0.60 2 0 92 95 57 13 3 3
0.90 0 0 22 87 42 6 2 0
'Abbreviations used are the same as those of previous Tables. NJP and NJD, neighbor-joining method with p and d distances, respectively; NJK, neighbor-
joining with Kimura distance; EP, evolutionary parsimony, B, proportion of transitional changes [a/(a + 2/3)] EP(x 2 ) shows the proportion of replications
in which only the x 2 for X, Y, or Z was greater than 3 84 m = 1,000; R = 400.
'a = c = e = 0.05 and b = d = 0.25.
*a = c = 0.20, b = d = 0.40, and e = 0.05.
!a = c = e = 0.05 and b - d = 0.5.
(Adapted from Jin and Nei, 1990.)
RELATIVE EFFICIENCIES OF DIFFERENT TREE-MAKING METHODS 113

than the first and second positions, and codons encoding active centers of
proteins evolve at a slower rate than other codons except in certain special
genes (see Nei, 1987a; Hughes and Nei, 1988). Examining the pattern of
amino acid substitution in cytochrome c, Uzzell and Corbin (1971) showed
that the substitution rate varies from site to site, roughly following a gamma-
distribution. This distribution is close to the lognormal distribution for
certain parameter values (see Fig. 6-6).
Therefore it is important to examine the effects of variation in substi-
tution rate among different sites on the efficiencies of different tree-making
methods. Table 6-10 shows the results of one such study. In this case,
substitution rate was assumed to vary according to the lognormal distri-
bution with parameter a = 8 (Olsen, 1987). The expected branch lengths
used were the same as those for model trees S1( S2, and S3 of Figure 6-5.
Table 6-10 indicates that the MP method is better than the EP method for
trees Si and S2, but is worse than the latter for tree S3. Therefore, the
conclusion remains the same. The NJ method with the d distance and
Kimura distance also shows a better performance than either the MP or
EP method for Sx and S2. For S3, however, the EP method now shows a
better performance.
The poor performance of the NJ method for S3 is caused by the fact that
neither the d distance nor the Kimura distance is an additive measure of
the number of nucleotide substitutions in this case. However, an approx-

Figure 6-6 Gamma-distribution with a = 1 ( ) and a = 2 (—-) and the

lognormal distribution with a = 8 ( ). X = substitution rate per site.
Table 6-10 Probabilities of Obtaining the Correct Unrooted Tree (Pc x 100) for the Case of Varying Substitution Rate
Among Different Nucleotide Sites With the Two-Parameter Model*
Tree-Making Method EP (X 2 )
Model Trees B MP NJD NJK NJG1 NJG2 EP X Y Z
s, 0.33 97 98 98 100 99 89 44 3 2
0.60 94 99 99 100 100 77 22 3 3
0.90 74 90 92 98 96 53 6 0 0

S2 0.33 83 90 90 94 93 55 12 4 5
0.60 83 86 86 90 89 50 11 5 4
0.90 65 77 75 80 78 41 4 3 3

S3 0.33 7 30 31 96 75 64 21 5 5
0.60 5 25 29 94 72 55 9 3 3
0.90 1 10 27 85 64 41 5 2 0
'Abbreviations used are the same as those m previous tables. NJG1, NJ method with the gamma-distance of a = 1. NJG2, NJ method with the gamma-
distance of a = 2, m = 1,000; R = 400. Model trees are presented in Figure 6-5 (Adapted from Jin and Nei, 1990 )
RELATIVE EFFICIENCIES OF DIFFERENT TREE-MAKING METHODS 115

imately additive measure can be attained by considering the gamma-

distribution (Nei and Gojobori, 1986; Jin and Nei, 1990). Since the log-
normal distribution with a = 8 is rather close to the gamma-distribution
with parameter a = 1 or 2 (Fig. 6-6), we developed the following gamma-
distance for the two-parameter model (Jin and Nei, 1990).

where P and Q are the estimates of P and Q in equations (9) and (10),
respectively, and a is the square of the inverse of the coefficient of variation
of substitution rate (Nei, 1987a: 234-235). In the case of a = 1, d becomes

whereas for a = 2 it is

Table 6-10 shows that when the gamma-distance with a = 1 is used, the
NJ method outperforms the EP method, even for S3. This indicates that
as long as a distance measure that gives an approximate additivity of nu-
cleotide substitutions is used, the NJ method performs very well. The same
can be said for the ST and ME methods, since these methods are expected
to give the correct tree provided that additive distance measures are used.
No study has been done on the effect of intersite variation of substitution
rate on the performance of the ML method. However, since it is difficult
for this method to take this variation into account, the method will be
affected considerably. It is desirable to study this effect quantitatively in
the near future.

STATISTICAL TESTS OF THE TREES OBTAINED

Test of Topological Differences

When a tree (topology) is obtained by a tree-making method, one is nat-
urally interested in the accuracy of the tree (i.e., how good the tree is,
compared with other alternative trees). A number of authors (e.g., Cav-
ender, 1981; Templeton, 1983; Felsenstein, 1985a,b; Lake, 1987; Prager
and Wilson, 1988; Kishino and Hasegawa, 1989) have proposed statistical
tests for topological differences. However, all these methods depend on
assumptions which in reality do not necessarily hold. Therefore, one has
to be cautious in using these methods. The main problem in these tests is
that the evaluation of statistical significance of the difference between two
topologies depends on the true tree (both topology and branch lengths),
which is usually unknown, and the data sets used, which do not necessarily
satisfy the assumptions required for a given tree-making method.
116 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

For example, Prager and Wilson's (1988) test is for comparing two to-
pologies obtained by parsimony methods. In this test, the significance of
the difference between the number of sites in which one topology wins,
and the number of sites in which the other topology wins is examined by
a binomial test. When the number of substitutions per site is small for all
branches, and the trees to be compared are bifurcating, this test seems to
be acceptable. In other cases, however, it may identify a wrong tree as the
correct one and regard it as statistically established. This is obvious from
the computer simulation given for S3 in Tables 6-9 and 6-10, where a wrong
tree was chosen as the correct one in most replications. Templeton's (1983)
test for restriction-site data has the same problem as Wilson and Prager's.
In this case, even if the rate of nucleotide substitution is constant, it may
choose a wrong tree and regard it as statistically established (Nei and
Tajima, 1985, 1989; Pamilo, 1990).
Lake (1987) proposed a statistical test in association with his evolutionary
parsimony method. As mentioned earlier, this tests examines whether or
not quantity X, Y, or Z is significantly different from 0. In practice, all of
X, Y, and Z may become nonzero under certain patterns of nucleotide
substitution. Thus, this is not a test for comparing different topologies (Jin
and Nei, 1990). Indeed, this test may identify a wrong tree as the correct
one, as is clear from Tables 6-9 and 6-10. Therefore, this test should not
be used.
There is one test that is statistically sound if the rate of nucleotide
substitution is constant (Felsenstein, 1985a). It applies to a rooted tree for
three DNA sequences (Fig. 6-7). In this case there are three possible trees:
(A), (B), and (C) in Figure 6-7. If tree (A) is correct, one would expect
that the number of nucleotide differences (nl2) between sequences 1 and
2 to be smaller than that (nl3) between 1 and 3 or that («23) between 2
and 3. Therefore, if nn is significantly smaller than «13 and n23 by a binomial
test, one can conclude that it is statistically established. In this case, the
tree corresponding to the null hypothesis (n12 = nl3 = n23) is the trifur-
cation tree given in Figure 6-7D. Extension of this method to the case of
four or more sequences is complicated, and no studies have been done.
Note also that even with three DNA sequences, the above test breaks
down if the rate of nucleotide substitution is not constant.

Figure 6-7 Three different rooted trees (A, B, C) for three DMA sequences (1, 2,
and 3). (D); null hypothesis tree.
RELATIVE EFFICIENCIES OF DIFFERENT TREE-MAKING METHODS 117

Williams and Goodman (1989) recently extended this approach to the

case where the molecular clock does not apply. In this method, however,
one must infer the ancestral nucleotide at each "informative site" by par-
simony analysis. If this inference is wrong, their test is not reliable. Never-
theless, this test is a promising approach when there are useful outgroup
species. This test also applies only to the case of three species of which
the branching order is to be determined.
At any rate, it is a very difficult task to assess the significance level for
the differences between topologies. Since all statistical methods depend
on various assumptions, one should be prepared for the possibility that
even a tree statistically established at the 5% or 1% level by a certain
method may later turn out to be incorrect. One such example is the phy-
logeny for humans, chimpanzees, gorillas, orangutans, and gibbons. Tem-
pleton (1983) conducted a nonparametric test for restriction-site data of
mitochondrial DNAs (mtDNAs) from these organisms and concluded that
humans and gorillas are evolutionarily closer than humans and chimpan-
zees. However, later studies of DNA hybridization data for single-copy
genomic DNA have not supported this conclusion (Sibley and Ahlquist,
1984; Caccone and Powell, 1989).
In general, one should not be too confident about a tree obtained from
any tree-making method unless the number of nucleotides examined is very
large. If there is any doubt about the tree obtained, the first thing to do
is to increase the amount of data. Unless the amount of data is large, any
sophisticated statistical method may lead to an erroneous conclusion.

Accuracy of the Topology Estimated

Although it is very difficult to test topological differences under realistic
conditions, it is easier to test the accuracy of branch lengths estimated for
a given topology. This test is primarily for examining the statistical signif-
icance of a given branch length, but if a branch length is not significantly
different from 0, it casts doubt on the clustering pattern associated with
the branch. For example, Figure 6-8 shows a phylogenetic tree for humans,
chimpanzees, gorillas, orangutans, and gibbons. If branch length a in this
tree is not significantly different from 0, the branching order among hu-
mans, chimpanzees, and gorillas becomes questionable.
This type of test was initiated by Nei et al. (1985) for a UPGMA tree.
Application of this method to Brown et al.'s (1982) mtDNA data, from
which the phylogeny in Figure 6-8 was constructed, indicates that the branch
length a is not statistically significant. Therefore, Brown et al.'s data are
not sufficient for resolving the branching pattern of humans, chimpanzees,
and gorillas. Li (1989) extended this method to the case of other distance
methods, though his primary interest was to discriminate among the three
possible unrooted trees for four DNA sequences.
Theoretically, the standard error of any branch length can be computed
for a UPGMA or any other distance method tree. However, the compu-
tation becomes quite complicated when the number of sequences is large.
118 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Figure 6-8 UPCMA tree for humans, chimpanzees, gorillas, orangutans, and gib-
bons. This tree was constructed by using mtDNA sequence data from Brown et al.
(1982). Branch length estimates ± SE are given for all internal branches, whereas
only branch-length estimates are given for exterior branches. Branch length b is
significantly larger than 0, whereas branch lengths a and c are not significant.

In this case, a simpler method would be to use the jackknife or bootstrap

method (Efron, 1982). The jackknife method is particularly useful for
evaluating the standard error of a given branch length when the number
of nucleotides or restriction sites examined is relatively small. Mueller and
Ayala (1982) used this method for a tree obtained from gene-frequency
data (see also Pamilo, 1990). When this number is large, however, it would
be easier to use bootstrapping.
Felsenstein (1985b) used bootstrapping to evaluate the accuracy of a tree
obtained by a parsimony method. His method is not to evaluate the stand-
ard error of any particular branch length, but to examine how often a
particular cluster in a tree appears when nucleotide sites are resampled
with replacement many times. This method can be used for any other
method. This test can easily be applied, particularly for the NJ method,
because the computational time for reconstructing a tree is usually very
short.
Nevertheless, one should be cautious about the outcome of this test,
because if the data set used does not satisfy the assumption underlying a
tree-making method, one may identify a wrong cluster as a correct one.
This is particularly so for a maximum-parsimony tree when the rate of
substitution varies extensively with evolutionary lineage. It should also be
noted that if the original data set is biased for some reason, a cluster may
be regarded as statistically significant even if it is a wrong one. This is
because the original bias cannot be corrected by the resampling process.

NUMBER OF NUCLEOTIDES TO BE EXAMINED

From the computer simulations mentioned earlier, it is clear that a large

number of nucleotides must be examined to obtain the correct phylogenetic
tree. The number of nucleotides required to establish a tree statistically
RELATIVE EFFICIENCIES OF DIFFERENT TREE-MAKING METHODS 119

depends on the topology and branch lengths of the true tree, the pattern
of nucleotide substitution, the tree-making method, and other factors.
Saitou and Nei (1986) studied this problem considering mtDNAs from
humans, chimpanzees, gorillas, orangutans, and gibbons.
Mammalian mtDNA is known to evolve about ten times faster than
nuclear DNA (Brown et al., 1979), and the phylogenetic tree for humans,
chimpanzees, gorillas, orangutans, and gibbons (Fig. 6-8) is close to those
given in Fig. 6-9. Suppose that tree A or B in Fig. 6-9 is the correct tree.
How many nucleotides then must be examined to obtain the correct tree
with probability P? The mathematical technique for computing the mini-
mum number of nucleotides required (m*) for the UPGMA, TD, DW,
FM, and MP methods is provided by Saitou and Nei (1986), and it can be
shown that m* for the NJ and ME methods is identical with that of the
TD and DW methods for the case of n < 5 (Saitou and Nei, 1987). One
can therefore determine m* for all these methods.
Some of the results obtained are presented in Table 6-11 for the one-
and two-parameter models of nucleotide substitution under the assumption
that (/) only orangutans are used as an outgroup (four species) and (ii)
both orangutans and gibbons are used as outgroups. In the case of the
two-parameter model, a transition/transversion ratio (a/ft) appropriate to
mtDNA is assumed to be 20. Table 6-11 shows that m* for P = 0.95 is
smallest for the NJ, ME, TD, and DW methods, and largest for the FM
method (m* for the MP method being usually close to that of the former).
However, if tree A is correct and a/ft = 20, 2,600 ~ 3,100 nucleotides
must be examined even in the NJ method, depending on the number of
outgroup species used. This number is much larger than the number of
nucleotides examined (895) by Brown et al. (1982). Therefore, Brown et
al.'s data are unlikely to resolve the branching pattern of humans, chim-
panzees, and gorillas. This conclusion is the same as that obtained from
the statistical test of branch length a in Figure 6-8 mentioned earlier.
Saitou and Nei (1986) have also examined the number of nucleotides
required when orangutans and gibbons are not available as outgroups. In

Figure 6-9 Two model trees for the case of four or five species. A, B, C, any of
humans, chimpanzees, and gorillas; D, orangutans; E, gibbons. Numbers refer to
the expected number of nucleotide substitutions per site per branch.
120 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Table 6-11 Numbers (m*) of Nucleotides Required for Obtaining the Correct
Tree with a Probability of 0.95*

Tree-Making Four Species Five Species

Method 1-pt 2-p 1-p 2-p
Tree A of Figure 6-9
NJ, ME, TD, DW 2,100 3,100 1,700 2,600
MP, CP 2,100 3,300 1,700 2,700
UPGMA 4,200 4,700 4,200 4,700
FM 5,000 8,200 3,000 4,900
Tree B of Figure 6-9
TD, DW 760 1,200 640 890
MP, CP 790 1,300 680 980
UPGMA 1,400 1,500 1,400 1,500
FM 1,700 2,800 900 1,700
'Abbreviations used are the same as those used in previous tables
*l-p and 2-p indicate the one- and two-parameter models, respectively. CP, compatibility method (Adapted
from Saitou and Nei, 1986.)

this case, m* becomes about 4,700 under the assumption of the molecular
clock.
Recently, Miyamoto et al. (1987) and Maeda et al. (1988) used about
7,100 and 3,100 nucleotides, respectively, from the nuclear genome, to
study the branching pattern of humans, chimpanzees, and gorillas. The
total number of nucleotides available is now more than 10,000. However,
since the rate of nucleotide substitution in the nuclear genome is about
one-tenth that of mtDNA, this number still does not seem to be sufficient
for establishing the branching order unquestionably (Maeda et al., 1988).
In this connection, it is interesting that when all cladistically informative
sites from these data are used, Williams and Goodman's (1989) test leads
to the conclusion that humans and chimpanzees are closer to each other
than to gorillas at the 3% level. If we consider the possibility that the
ancestral nucleotide inferred is incorrect, however, this conclusion does
not seem to be sufficient for establishing the branching order of the three
species involved. Using a different statistical method, Li (1989) could not
reach the same conclusion from the same set of data. Of course, if we
consider all data from DNA hybridization, mtDNA, and nuclear DNA
sequences, humans and chimpanzees seem to be genetically closer to each
other than to gorillas.

DISCUSSION

As mentioned above, many studies have been conducted on the relative

efficiencies of tree-making methods. They indicate that the relative effi-
ciencies depend on various factors such as the shape of the true tree, the
numbers of nucleotide substitutions, transition/transversion ratios, and varying
rate of nucleotide substitution. However, we can make some general con-
clusions.
RELATIVE EFFICIENCIES OF DIFFERENT TREE-MAKING METHODS 121

One of the general conclusions is that the maximum parsimony method,

which is currently very popular among numerical taxonomists, is as efficient
as the ME, NJ, and ML methods only when the number of nucleotide
substitutions per site is very small and the number of nucleotides examined
is very large. Otherwise, the latter methods are better than the MP method.
This seems to be true whether or not the transition/transversion bias is
taken into account in the parsimony method (Jin and Nei, 1990). So far,
no studies have been conducted on the relative efficiency of such recent
parsimony methods as dynamic weighting parsimony (e.g., Williams and
Fitch, 1990). Therefore, more simulation work is necessary.
The NJ method is an approximate method of obtaining the minimum
evolution tree, but produces the same tree as that obtained by the ME
method in most cases; if it does not, the tree obtained is usually very close
to the ME tree. The advantage of the NJ method over the ME method is
the short computer time required. The ST method depends on a procedure
that is similar to that of the NJ method, and these two methods seem to
be equally efficient in obtaining the correct tree. The NJ method, however,
has one advantage over the ST method; it gives not only the topology but
also the branch-length estimates of the tree obtained.
The ML method is also very efficient in obtaining the correct tree when
the underlying assumptions are satisfied. When these assumptions are not
satisfied, it seems to be less efficient than the NJ or ME method because
the former depends on the details of the probability model, whereas the
latter are known to be quite insensitive to various aspects of nucleotide
substitution (see Table 6-10). Another disadvantage of the ML method is
that it requires enormous computer time, even for a small number of DNA
sequences examined.
Some investigators (e.g., Czelusniak et al., 1990) seem to be dissatisfied
with the fact that the NJ method gives only one topology. They are inter-
ested in knowing how good the NJ tree is compared with alternative to-
pologies. Since the NJ method is intended to obtain the minimum evolution
tree, comparison of alternative trees can be made by using the total sum
of branch lengths, i.e., L in equation (3). In the case of the ME method,
L is computed for all topologies, so that one can choose a topology that
gives the smallest L value, though it usually takes a large amount of com-
puter time. In practice, however, it is sufficient to examine the L values
for several alternative topologies that are close to that of the NJ tree. A
simple way to find alternative topologies is to use Robinson and Foulds'
(1981) measure (dT) of computing topological differences. This measure
takes a value of 2, 4, 6, 8, etc. I suggest that all topologies which are
different from the NJ tree by dT = 2 be examined for this purpose. Since
the number of such topologies is not large (Sourdis and Nei, 1988), it would
be much simpler to examine alternative topologies in this way. Once the
topologies are identified, one can easily compute L for each of the topol-
ogies by using Fitch and Margoliash's (1967) method of estimating branch
lengths. If this procedure finds a tree which has a smaller L value than
that of the NJ tree, then this tree should be used as the final tree. Otherwise,
122 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

the NJ tree will be used as the final tree. (Note that the ME tree is not
necessarily the correct tree).
Although the statistical test of topological differences is besieged with
many problems as mentioned earlier, a rough test of the difference in L
between two topologies can be made by computing the standard error of
the differences by using either the jackknife or bootstrap method. Con-
ceptually, this test is similar to Hasegawa et al.'s (1987) test of maximum-
likelihood values, but the computation is much simpler.
When the ME or NJ method is used, one must use an appropriate
distance measure. As mentioned earlier, provided that the distance meas-
ure gives the correct number of nucleotide substitutions, these methods
produce the correct tree. In practice, the pattern of nucleotide substitution
is quite complicated, and there is no universally accurate distance measure
for this purpose. However, I recommend the following guidelines for meas-
uring nucleotide substitutions (Jin and Nei, 1990).
1. When the Jukes-Cantor estimate of the number of nucleotide sub-
stitutions per site (d) between different sequences is about 0.1 or less,
use the Jukes-Cantor distance whether there is a transition/transver-
sion bias or the substitution rate (A) varies with nucleotide site. In
this case, the Kimura distance or the gamma-distance gives essentially
the same value as the Jukes-Cantor distance. One may also use the
p distance (proportion of different nucleotides) for constructing a
topology.
2. When d is greater than 0.1 but less than about 0.3, use the Jukes-
Cantor distance unless the transition bias is high (e.g., B > 0.5).
When this bias is high, use the Kimura distance.
3. When 1.0 > d > 0.3 and there is evidence that A varies extensively
with site, use the gamma-distance. In general, we suggest that the
gamma-distance with a = 1 be used. However, one may choose a
different gamma-distance, estimating a from data. Wilson et al. (1989)
recently used a distance with a — l/2 for restriction-site data of mtDNA
in hominoids.
4. When 1.0 > d > 0.3 and the frequencies of the four nucleotides (A,
T, C, and G) deviate substantially from equality, use Tajima and
Nei's (1984) distance.
5. When d > 1.0 for many pairs of sequences, the phylogenetic tree
estimated is not reliable for a number of reasons (e.g., large standard
errors of d's, and sequence alignment errors). We therefore suggest
that these sets of data should not be used. In this case, one may
eliminate the portion of the gene that evolves very fast and use only
the remaining region as is often done in studies of the evolution of
different kingdoms or phyla using ribosomal RNA genes (e.g., Gouy
and Li, 1989). If a coding region of DNA is examined, amino acid
sequences rather than DNA sequences should be used. One may also
use a different gene which evolves more slowly.
6. When a phylogenetic tree is constructed from the coding regions of
a gene, the distinction between synonymous (ds) and nonsynonymous
RELATIVE EFFICIENCIES OF DIFFERENT TREE-MAKING METHODS 123

(dN) substitutions (Miyata and Yasunaga, 1980; Li et al., 1985; Nei

and Gojobori, 1986) will be helpful, because the rate of synonymous
substitution is usually much higher than that of nonsynonymous sub-
stitution. When relatively closely-related species with ds < 1.0 are
studied for a large number of codons, one may use ds for constructing
a tree. This procedure is expected to reduce the effect of variation
in substitution rate among different sites, because synonymous sub-
stitutions are apparently largely neutral in higher organisms (Nei,
1987a: 79-86). However, when relatively distantly related species are
studied, the use of dN is recommended.
Finally, it should be mentioned that the simulation study of tree-making
methods is far from complete. In most studies, the number of DNA se-
quences considered is quite small (mostly four to eight) to save computer
time. In some tree-making methods (e.g., MP method), the increase in
the number of sequences considered may enhance the probability of ob-
taining the correct tree. Therefore, a more detailed study is necessary to
solve this problem. The effects of different patterns of nucleotide substi-
tution, particularly varying rate among different sites, should also be stud-
ied more carefully. Another problem that should be studied by computer
simulation is the statistical methods for testing topological differences or
the accuracy of a tree obtained. Since these methods depend on a number
of assumptions about complex evolutionary processes, it would be impor-
tant to establish the validity of the method by computer simulation.
In this chapter we considered only gene trees. A gene tree is not nec-
essarily the same as the phylogeny of the species or populations from which
the gene sequences are sampled (Nei, 1987a; Pamilo and Nei, 1988). The
chance that the gene and population trees are different is high when the
gene used is polymorphic or when the populations considered split rela-
tively recently. In these cases, one must examine many independently
evolving genes from the genome to estimate the correct population tree
(Saitou and Nei, 1986; Pamilo and Nei, 1988). In this connection, it is
important to note that a phylogenetic tree constructed from mitochondrial
or chloroplast DNAs is a gene tree though they include many genes. This
is because the genes in the mitochondrial or chloroplast DNA are inherited
as a single entity without recombination. Therefore, some caution should
be exercised in inferring population phylogenies from mitochondrial or
chloorplast DNA. Of course, gene trees are not always studied just for
inferring a population phylogeny. In such a case as the study of the evo-
lution of multigene families, a gene tree provides important information.

COMPUTER PROGRAMS

Computer programs for computing various distances for DNA sequences

(numbers of nucleotide substitutions per site, synonymous substitutions
per synonymous site, nonsynonymous substitutions per nonsynonymous
site, etc.) and for constructing (NJDRAW) and testing (NJBOOT) a neigh-
124 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

bor-joining tree are available upon request from the Institute of Molecular
Evolutionary Genetics, Penn State University, 328 Mueller Laboratory,
University Park, PA 16802-5303. Please send IBM compatible 3.5 or 5.25
inch floppy diskettes. Each program requires one 360 Kb diskette.

ACKNOWLEDGMENTS

I thank Li Jin for his comments. This study is supported by research grants
from the National Institutes of Health and the National Science Founda-
tion.

REFERENCES

Aquadro, C. F., and B. D. Greenberg. (1983) Human mitochondrial DNA vari-

ation and evolution: analysis of nucleotide sequences from seven individuals.
Genetics 103:287-312.
Blanken, R. L., L. C. Klotz, and A. G. Hinnebusch. (1982) Computer comparison
of new and existing criteria for constructing evolutionary trees from sequence
data. J. Mol. Evol. 19:9-19.
Britten, R. J. (1986) Rates of DNA sequence evolution differ between taxonomic
groups. Science 231:1393-1398.
Brown, W. M., M. George, Jr., and A. C. Wilson. (1979) Rapid evolution of
animal mitochondrial DNA. Proc. Nad. Acad. Sci. USA 76:1967-1971.
Brown, W. M., E. M. Prager, A. Wang, and A. C. Wilson. (1982) Mitochondrial
DNA sequences of primates: tempo and mode of evolution. /. Mol. Evol.
18:225-239.
Buneman, P. (1971) The recovery of trees from measurements of dissimilarity. Pp.
387-395 in Mathematics in the Archeological and Historical Sciences (F. R.
Hodson, D. G. Kendall, and P. Tautu, eds.). Edinburgh University Press,
Edinburgh.
Caccone, A., and J. R. Powell. (1989) DNA divergence among hominoids. Evo-
lution 43:925-942.
Cavalli-Sforza, L. L., and A. W. F. Edwards. (1967) Phylogenetic analysis: models
and estimation procedures. Am. J. Hum. Genet. 19:233-257.
Cavender, J. A. (1981) Tests of phylogenetic hypotheses under generalized models.
Math. Biosci. 54:217-229.
Cavender, J. A. (1989) Mechanized derivation of linear invariants. Mol. Biol. Evol.
6:301-316.
Czelusniak, J., M. Goodman, N. O. Moncrief, and S. M. Kehoe. (1990) Maximum
parsimony approach to construction of evolutionary trees from aligned ho-
mologous sequences. Pp. 601-615 in Molecular Evolution: Computer Anal-
ysis of Protein and Nucleic Acid Sequences (R. F. Doolittle, ed.). Methods
in Enzymology, vol. 183. Academic Press, New York.
Eck, R. V., and M. O. Dayhoff. (1966) Atlas of Protein Sequence and Structure.
National Biomedical Research Foundation, Silver Spring. Md.
Efron, B. (1982) The Jackknife, the Bootstrap and Other Resampling Plans. Society
for Industrial and Applied Mathematics, Philadelphia.
RELATIVE EFFICIENCIES OF DIFFERENT TREE-MAKING METHODS 125

Faith, D. P. (1985) Distance methods and the approximation of most-parsimonious

trees. Syst. Zool. 34:312-325.
Farris, J. S. (1970) Methods for computing Wagner trees. Syst. Zool. 19:83-92.
Farris, J. S. (1972) Estimating phylogenetic trees from distance matrices. Am.
Natur. 106:645-668.
Farris, J. S. (1977) On the phenetic approach to vertebrate classification. Pp. 823-
850 in Major Patterns in Vertebrate Evolution (M. D. Hecht, P. C. Goody,
and B. M. Hecht, eds.). Plenum Press, New York.
Farris, J. S. (1981) Distance data in phylogenetic analysis. Pp. 3-23 in Advances
in Cladistics. Proceedings of the First Meeting of the Willi Hennig Society
(V. A. Funk and D. R. Brooks, eds.). New York Botanical Garden, Bronx.
Felsenstein, J. (1978) Cases in which parsimony or compatibility methods will be
positively misleading. Syst. Zool. 27:401-410.
Felsenstein, J. (1981a) Evolutionary trees from DNA sequences: a maximum like-
lihood approach. /. Mol. Evol. 17:368-376.
Felsenstein, J. (1981b) Evolutionary trees from gene frequencies and quantitative
characters: finding maximum likelihood estimates. Evolution 35:1229-1242.
Felsenstein, J. (1985a) Confidence limits on phylogenies with a molecular clock.
Syst. Zool. 34:152-161.
Felsenstein, J. (1985b) Confidence limits on phylogenies: an approach using the
bootstrap. Evolution 39:783-791.
Felsenstein, J. (1986) Distance methods: a reply to Farris. Cladistics 2:130-143.
Felsenstein, J. (1988) Phylogenies from molecular sequences: inference and reli-
ability. Ann. Rev. Genet. 22:521-565.
Fitch, W. M. (1971) Toward defining the course of evolution: minimum change
for a specific tree topology. Syst. Zool. 20:406-416.
Fitch, W. M. (1981) A non-sequential method for constructing trees and hierar-
chical classifications. J. Mol. Evol. 18:30-37.
Fitch, W. M., and E. Margoliash. (1967) Construction of phylogenetic trees. Science
155:279-284.
Gojobori, T., K. Ishii, and M. Nei. (1982a) Estimation of average number of
nucleotide substitutions when the rate of substitution varies with nucleotide.
/. Mol. Evol. 18:414-423.
Gojobori, T., W.-H. Li, and D. Graur. (1982b) Patterns of nucleotide substitution
in pseudogenes and functional genes. J. Mol. Evol. 18:360-369.
Gouy, M., and W.-H. Li. (1989) Molecular phylogeny of the kingdoms Animalia,
Plantae, and Fungi. Mol. Biol. Evol. 6:109-122.
Graur, D. (1985) Pattern of nucleotide substitution and the extent of purifying
selection in retroviruses. /. Mol. Evol. 21:221-231.
Hasegawa, M., H. Kishino, and T. Yano. (1987) Man's place in Hominoidea as
inferred by molecular clocks of DNA. /. Mol. Evol. 26:132-147.
Hasegawa, M., and T. Yano. (1984) Maximum likelihood method of phylogenetic
inference from DNA sequence data. Bull. Biometric Soc. Japan 5:1-7.
Hughes, A., and M. Nei. (1988) Pattern of nucleotide substitution at major his-
tocompatibility complex class I loci reveals overdominant selection. Nature
335:167-170.
Jin, L., and M. Nei. (1990) Limitations of the evolutionary parsimony method of
phylogenetic analysis. Mol. Biol. Evol. 7:82-102.
Jukes, T. H., and C. R. Cantor. (1969) Evolution of protein molecules. Pp. 21-
132 in Mammalian Protein Metabolism (H. N. Munro, ed.). Academic Press,
New York.
126 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Kim, J., and M. A. Burgman. (1988) Accuracy of phylogenetic-estimation methods

under unequal evolutionary rates. Evolution 42:596-602.
Kimura, M. (1980) A simple method for estimating evolutionary rate of base
substitutions through comparative studies of nucleotide sequences. /. Mol.
Evol. 16:111-120.
Kimura, M. (1983) The Neutral Theory of Molecular Evolution. Cambridge Uni-
versity Press, Cambridge.
Kishino, H., and M. Hasegawa. (1989) Evaluation of the maximum likelihood
estimate of the evolutionary tree topologies from DNA sequence data, and
the branching order in Hominoidea. /. Mol. Evol. 29:170-179.
Klotz, L. C., and R. L. Blanken. (1981) A practical method for calculating evo-
lutionary trees from sequence data. /. Theor. Biol. 91:261-272.
Klotz, L. C., N. Komar, R. L. Blanken, and R. M. Mitchell. (1979) Calculation
of evolutionary trees from sequence data. Proc. Natl. Acad. Sci. USA 76:4516-
4520.
Lake, J. A. (1987) A rate-independent technique for analysis of nucleic acid se-
quences: evolutionary parsimony. Mol. Biol. Evol. 4:167-191.
Lake, J. A. (1988) Origin of the eukaryotic nucleus determined by rate-invariant
analysis of rRNA sequences. Nature 331:184-186.
Li, W.-H. (1981) A simple method for constructing phylogenetic trees from distance
matrices. Proc. Natl. Acad. Sci. USA 78:1085-1089.
Li, W.-H. (1989) A statistical test of phylogenies estimated from sequence data.
Mol. Biol. Evol. 6:424-435.
Li, W.-H., M. Tanimura, and P. M. Sharp. (1987b) An evaluation of the molecular
clock hypothesis using mammalian DNA sequences. J. Mol. Evol. 25:330-
342.
Li, W.-H., K. H. Wolfe, J. Sourdis, and P. M. Sharp. (1987a) Reconstruction of
phylogenetic trees and estimation of divergence times under nonconstant
rates of evolution. Cold Spring Harbor Symp. Quant. Biol. 52:847-856.
Li, W.-H., C.-I. Wu, and C.-C. Luo. (1985) A new method for estimating syn-
onymous and nonsynonymous rates of nucleotide substitution considering
the relative likelihood of nucleotide and codon changes. Mol. Biol. Evol.
2:150-174.
Maeda, N., C.-I. Wu, J. Bliska, and J. Reneke. (1988) Molecular evolution of
intergenic DNA in higher primates: pattern of DNA changes, molecular
clock and evolution of repetitive sequences. Mol. Biol. Evol. 5:1-20.
Miyamoto, M. M., J. L. Slightom, and M. Goodman. (1987) Phylogenetic rela-
tionships of human and African apes from DNA sequences of the vH-globin
region. Science 238:369-373.
Miyata, T., and T. Yasunaga. (1980) Molecular evolution of mRNA: A method
for estimating evolutionary rates of synonymous and amino acid substitutions
from homologous nucleotide sequences and its application. J. Mol. Evol.
16:23-36.
Mueller, L. D., and F. J. Ayala. (1982) Estimation and interpretation of genetic
distance on empirical studies. Genet. Res. 40:127-137.
Nei, M. (1987a) Molecular Evolutionary Genetics. Columbia University Press, New
York.
Nei, M. (1987b) Genetic distance and molecular phylogeny. Pp. 193-223 in Pop-
ulation Genetics and Fishery Management (N. Ryman and F. Utter, eds.).
University of Washington Press, Seattle.
Nei, M., and T. Gojobori. (1986) Simple methods for estimating the numbers of
RELATIVE EFFICIENCIES OF DIFFERENT TREE-MAKING METHODS 127

synonymous and nonsynonymoiis nucleotide substitutions. Mol. Biol. Evol.

3:418-426.
Nei, M., and A. L. Hughes. (1991) Polymorphism and evolution of the major
histocompatibility complex loci. Pp. 222-247 in Evolution at the Molecular
Level (R. K. Selander, A. G. Clark, and T. S. Whittam, eds.). Sinauer
Associates, Sunderland. I
Nei, M., J. C. Stephens, and N. Saitou. (1985) Methods for computing the standard
errors of branching points in an evolutionary tree and their application to
molecular data from humans and apes. Mol. Biol. Evol. 2:66-85.
Nei, M., and F. Tajima. (1985) Evolutionary change of restriction cleavage sites
and phylogenetic inference for man and apes. Mol. Biol. Evol. 2:189-205.
Nei, M., F. Tajima, and Y. Tateno. (1983) Accuracy of estimated phylogenetic
trees from molecular data. II. Gene frequency data. J. Mol. Evol. 19:153-
170.
Olsen, G. J. (1987) Earliest phylogenetic branchings: comparing rRNA-based ev-
olutionary trees inferred with various techniques. Cold Spring Harbor Symp.
Quant. Biol. 52:825-837.
Pamilo, P. (1990) Statistical tests of phenograms based on genetic distances. Ev-
olution 44:689-697.
Pamilo, P., and M. Nei. (1988) Relationships between gene trees and species trees.
Mol. Biol. Evol. 5:568-583.
Peacock, D., and D. Boulter. (1975) Use of amino acid sequence data in phylogeny
and evaluation of methods using computer simulation. J. Mol. Biol. 95:513-
527.
Penny, D. (1982) Towards a basis for classification: the incompleteness of distance
measures, incompatibility analysis and phenetic classification. J. Theor. Biol.
96:129-142.
Prager, E. M., and A. C. Wilson. (1988) Ancient origin of lactalbumin from'
lysozyme: analysis of DNA and amino acid sequences. /. Mol. Evol. 27:326-
335.
Robinson, D. F., and L. R. Foulds. (1981) Comparison of phylogenetic trees.
Math. Biosci. 53:131-147.
Rohlf, F. J., and M. C. Wooten. (1988) Evaluation of the restricted maximum-
likelihood method for estimating phylogenetic trees using simulated allele
frequency data. Evolution 42:581-595.
Saitou, N. (1987) Patterns of nucleotide substitutions in influenza A virus genes.
Jpn. J. Genet. 62:439-444.
Saitou, N. (1988) Property and efficiency of the maximum likelihood method for
molecular phylogeny. /. Mol. Evol. 27:261-273.
Saitou, N., and T. Imanishi. (1989) Relative efficiencies of the Fitch-Margoliash,
maximum-parismony, maximum-likelihood, minimum-evolution, and neigh-
bor-joining methods of phylogenetic tree construction in obtaining the cor-
rect tree. Mol. Biol. Evol. 6:514-525.
Saitou, N., and M. Nei. (1986) The number of nucleotides required to determine
the branching order of three species with special reference to the human-
chimpanzee-gorilla divergence. /. Mol. Evol. 24:189-204.
Saitou, N., and M. Nei. (1987) The neighbor-joining method: a new method for
reconstructing phylogenetic trees. Mol. Biol. Evol. 4:406-425.
Sattath, S., and A. Tversky. (1977) Additive similarity trees. Psychometrika 42:319-
345.
Shoemaker, J. S., and W. M. Fitch. (1989) Evidence from nuclear sequences that
128 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

invariable sites should be considered when sequence divergence is calculated.

Mol. Biol. Evol. 6:270-289.
Sibley, C. G., and J. E. Ahlquist. (1984) The phylogeny of the hominoid primates,
as indicated by DNA-DNA hybridation. /. Mol. Evol. 20:2-15.
Sidow, A., and A. C. Wilson. (1990) Compositional statistics: an improvement of
evolutionary parismony and its application to deep branches in the tree of
life. /. Mol. Evol. 31:51-68.
Sneath, P. H. A., and R. R. Sokal. (1973) Numerical Taxonomy. Freeman, San
Francisco.
Sober, E. (1985) A likelihood justification of parsimony. Cladistics 1:209-233.
Sokal, R. R., and C. D. Michener. (1958) A statistical method for evaluating
systematic relationships. Univ. Kansas Sci. Bull. 28:1409-1438.
Sourdis, J., and C. Krimbas. (1987) Accuracy of phylogenetic trees estimated from
DNA sequence data. Mol. Biol. Evol. 4:159-166.
Sourdis, J., and M. Nei. (1988) Relative efficiencies of the maximum parsimony
and distance-matrix methods in obtaining the correct phylogenetic tree. Mol.
Biol. Evol. 5:298-311.
Swofford, D. L. (1981) On the utility of the distance Wagner procedure. Pp. 25-
43 in Advances in Cladistics. Proceedings of the First Meeting of the Willi
Hennig Society (V. A. Funk and D. R. Brooks, eds.). New York Botanical
Gardens, Bronx.
Swofford, D. L., and G. J. Olsen. (1990) Phylogeny reconstruction. Pp. 411-501
in Molecular Systematics (D. M. Hillis and C. Moritz, eds.). Sinauer As-
sociates, Sunderland.
Tajima, F. (1990) A simple graphic method for reconstructing phylogenetic trees
from molecular data. Mol. Biol. Evol. 7:578-588.
Tajima, F., and M. Nei. (1984) Estimation of evolutionary distance between nu-
cleotide sequences. Mol. Biol. Evol. 1:269-285.
Tateno, Y., M. Nei, and F. Tajima. (1982) Accuracy of estimated phylogenetic
trees from molecular data. I. Distantly related species. J. Mol. Evol. 18:387-
404.
Templeton, A. R. (1983) Phylogenetic inference from restriction endonuclease
cleavage site maps with particular reference to the evolution of humans and
the apes. Evolution 37:221-244.
Uzzell, T., and K. W. Corbin. (1971) Fitting discrete probability distributions to
evolutionary events. Science 172:1089-1096.
Williams, P. L., and W. M. Fitch. (1990) Phylogeny determination using dynam-
ically weighted parsimony methods. Pp. 615-626 in Molecular Evolution:
Computer Analysis of Protein and Nucleic Acid Sequences (R. F. Doolittle,
ed.). Methods in Enzymology, vol. 183. Academic Press, New York.
Williams, S. A., and M. Goodman. (1989) A statistical test that supports a human/
chimpanzee clade based on noncoding DNA sequence data. Mol. Biol. Evol.
6:325-330.
Wilson, A. C., E. A. Zimmer, E. M. Prager, andT. D. Kocher. (1989) Restriction
mapping in the molecular systematics of mammals: a retrospective salute.
Pp. 407-419 in The Hierarchy of Life. Molecules and Morphology in Phy-
logenetic Analysis (B. Fernholm, K. Bremer, andH. Jornval, eds.). Elsevier
Science Publishers, B. V., Amsterdam.
7
Compositional Statistics Evaluated by
Computer Simulations

AREND SIDOW AND ALLAN C. WILSON

Phylogenetic hypotheses testable with DNA sequences span the entire

range of evolutionary relationships, from trees relating individuals within
a species (Vigilant et al., 1989), to the tree of life (Fox et al., 1980; Lake,
1988; Gouy and Li, 1989; Sidow and Wilson, 1990). A number of methods
are available for inferring evolutionary relationships among homologous
and aligned DNA sequences. This chapter is about a new method called
compositional statistics, which is most suitable for elucidating relationships
among highly diverged sequences. In contrast to most other methods, it
takes into account the sequences' base compositions.
Our discussion emphasizes the idea that biases in the base composition
of the compared sequences may affect a phylogenetic analysis and produce
systematic errors if not properly corrected for. Such biases are especially
strong in animal mitochondrial DNA (mtDNA) and in the heavy com-
partments of homoiotherm genomes (Perrin and Bernard!, 1987). Biases
may arise as a consequence of selection (Bernardi et al., 1985) or mutation
pressure (Sueoka, 1988), but there is some uncertainty as to which of these
two processes is actually responsible for producing them (Li and Graur,
1991; M. Bulmer, personal communication). Compositional statistics was
developed to take into account biases that are due to global selection on
the DNA's primary structure, but it may also function in cases where
mutation pressure produces the bias.
In order to point out the most useful applications of compositional sta-
tistics, we first discuss some important strengths and weaknesses of the
most commonly used methods of phylogenetic inference.

METHODS OF PHYLOGENETIC INFERENCE

Parsimony
Parsimony, the most widely used method of phylogenetic inference, lacks
a stochastic model of sequence evolution; it simply chooses an evolutionary

129
130 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

tree among all trees as the correct one if it requires the smallest number
of substitutions. Parsimony would be the natural method for inferring
phylogenetic relationships among organisms if parallelisms and reversals
were impossible in evolution. The correct phylogeny could immediately be
inferred from the existence of shared characters between related organisms.
It is now clear that DNA sequences generally do not meet this requirement
(Felsenstein, 1978; Shoemaker and Fitch, 1989; Saccone et al., 1989).
Parsimony is most likely to give positively misleading results when the
amount of evolution is very unequal in the branches of the phylogenetic
tree (Felsenstein, 1978; Lake, 1987). Nevertheless, it is useful and reliable
for closely related sequences in which parallelisms and reversals are rare.

Distance Methods
Distance methods are also widely used for inferring evolutionary relation-
ships. Sophisticated algorithms are available that, with a high probability,
find the best fit of evolutionary relationships to a distance matrix (Saitou
and Nei, 1987; Saitou and Imanishi, 1989; Jin and Nei, 1990).
For sequences that are not closely related, a major complication is that
observed pairwise distances are insufficient in an analysis. Multiple-hit
corrections are necessary to approximate actual pairwise distances between
the sequences (Gojobori et al., 1990). A simple two-parameter model,
distinguishing transitions from transversions (Kimura, 1980), can give suf-
ficient correction in many cases. For distantly related sequences, proper
corrections require extensive knowledge about the substitution process and
evolutionary constraints (Fitch, 1986; Olsen, 1987; Shoemaker and Fitch,
1989; Jin and Nei, 1990). In practice, it is usually impossible to take into
account differences in the rate of evolution among the sites, or estimate
the number of variable sites in a sequence—independent from the data
set that is used for the analysis—to make such corrections.

Maximum Likelihood Methods

These methods (Felsenstein, 1981) compute the likelihoods of all possible
trees, based on stochastic models of sequence evolution whose parameters
can either be specified by the user or estimated by the algorithm. It is then
tested whether the tree with the greatest likelihood is significantly better
than others. Potentially, maximum likelihood is the most powerful and
statistically most reliable method of phylogenetic inference. When likeli-
hood calculations are based on multiple-parameter models of sequence
evolution (e.g., Kishino and Hasegawa, 1990), the base composition of
the sequences is considered. Present limitations are partly computational
because of the complexity of likelihood calculations. Drawbacks of avail-
able maximum likelihood methods include the assumption that every site
in the sequence undergoes evolution at the same rate or that rate differences
among sites have to be specified by the user (Felsenstein, 1989).
COMPOSITIONAL STATISTICS EVALUATION 131

Evolutionary Parsimony
Since highly divergent sequences are less likely to obey simple models of
evolution, solving distant evolutionary relationships requires the most gen-
eral methods available. A major step in this direction was the invention
of evolutionary parsimony (EP) (Lake, 1987). Evolutionary parsimony's
property of rate invariance renders it particularly invulnerable to any var-
iation in the number of substitution events among branches of the evolu-
tionary tree, as well as variations in the evolutionary rate among different
sites in the sequences. For a thorough explanation of how EP achieves
rate invariance, see Sidow and Wilson (1990).
Evolutionary parsimony, like parsimony, relies on the concept of shared
states of homologous characters (Fig. 7-1). Briefly, the crucial differences
are the following: (1) In EP, informative nucleotide positions are all those
in which two sequences differ from the other two by a transversion. In
parsimony, transitional differences are informative as well. (2) In parsi-
mony, all informative patterns are at the same time supportive. In EP,
informative patterns need not be supportive. Evolutionary parsimony tests
the parsimony assignments relative to other informative patterns (referred
to as "background patterns") present in the aligned sequences. Parsimony
allows no test independent of the parsimony positions (Fig. 7-2).
Evolutionary parsimony statistically distinguishes misleading nucleotide
patterns that arose on peripheral branches, from those patterns that arose
on the central branch of the phylogenetic tree, as explained in Figure 7-2
(Holmquist et al., 1988; Sidow and Wilson, 1990). Parsimony (as well as
distance methods), because it lacks an internal comparison, may support
an incorrect phylogenetic tree when the substitution process in peripheral
branches creates parallel substitutions (Felsenstein, 1978).
Compositional Statistics
Compositional statistics (CS) is a new method of phylogenetic inference
that is related to EP but also incorporates base compositional components

Figure 7-1 Phylogenetically informative positions in four aligned, hypothetical se-

quences. Those nucleotide positions are shown whose patterns are informative for
the grouping of sequences 1 with 2, and of 3 with 4. On the left, are those patterns
which are informative in parsimony, on the right, are those informative for EP and
CS.
132 PHYLOCENETIC ANALYSIS OF DNA SEQUENCES

Figure 7-2 Comparison of parsimony and invariant methods of tree analysis. (Top)
The correct, rooted evolutionary tree for taxa 1, 2, 3, and 4. (Bottom) For both
parsimony and invariant methods (CS and EP), events on the central branch lead to
correct support of the network relating sequences 1 and 3 to the exclusion of 2 and
4 (as shown in the left half of each panel). Parallel events on the peripheral branches
support the incorrect topology when parsimony is used but not when the invariant
methods (EP and CS) are used (as shown in the right half of each panel).

of the stochastic models that underlie maximum likelihood calculations.

Since describing this method and applying it to two sets of DNA sequences
(Sidow and Wilson, 1990), we have conducted computer simulations aimed
at comparing its performance with that of EP. Besides briefly reviewing
CS and its relation to EP, this chapter gives the results of some of these
COMPOSITIONAL STATISTICS EVALUATION 133

simulations. Compositional statistics is shown to be less likely than EP to

make errors when base compositions are biased.
In comparison with other methods, a drawback of CS is that it can be
applied to only four sequences at a time. In practice, the solution to this
problem is to test all four-taxon trees that pertain to the phylogenetic
hypothesis to be solved. An example showing how CS was applied to ten
taxa appears in Sidow and Wilson (1990).

Limitations of Evolutionary Parsimony

Evolutionary parsimony assumes the two transversions from one base (e.g.,
G; -> T, and G, -» C,) to be roughly equally probable. That is,

where subscript i designates the position in the sequence. An exception to

this assumption is if T and C are maintained at different frequencies because
of selective constraints on the base composition. For example, if
F(T) > F(C), (2)
then

in order to maintain proper frequencies of T and C. Compositional con-

straints, like those in (2), are found to exist in a wide range of sequences.
Examples include animal mtDNA (where the coding strand displays a
significant bias against G) or the heavy compartments of homoiotherm
genomes, which show a significant bias toward a high G + C content
(Anderson et al., 1981; Bernardi et al., 1985; Saccone.et al., 1989). Thus,
certain kinds of sequences may render EP inapplicable to their analysis.
The same limitation holds true for distance and parsimony methods.

CS's Model of Sequence Evolution and its Phylogenetic Test

The central difference to EP is CS's model of sequence evolution. It is
identical to the stochastic model described in Hasegawa et al. (1985) and
Kishino and Hasegawa (1990). It distinguishes explicitly between mutation
and selection on the base composition of the evolving sequence. It also
distinguishes between transitions and transversions. In base compositional
equilibrium, transitions are assumed to obey the compositional balance
within the pool of purines, or pyrimidines:

and
134 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Transversions also obey compositional balance, and they are assumed

to proceed as follows:
1. Random choice of a position in the sequence to undergo a transver-
sion. The probability of a given base to be chosen for a transversion
mutation (tv) thus depends linearly on its frequency in the sequence:
P(tvN) = F(N). (6)
2. Choice of the particular substitution. The conditional probabilities of
the two competing transversions are directly proportional to the fre-
quencies of the two possible bases. If G, for example, undergoes a
transversion, selection will probabilistically balance the two possible
outcomes.

and

Thus,

and

where 6 is a proportionality coefficient for the overall probability of any

transversion. Remarkably, the model leads to "reversibility." For example,
a substitution from G to T becomes as likely as a substitution from T to
G because

The two steps of the substitution process are modeled after two processes
that exist in sequence evolution: mutation and selection.1
1. Mutation as the initial step in sequence evolution. The frequency of
a given mutation is dependent on the frequency of the original base
and on the error rate of the DNA replication and repair machinery.
Assuming that the latter distinguishes only transitions from transver-
sions, our model accurately describes the process of mutation.
'This model may not apply in cases where the compositional bias is due to mutation pressure
instead of selection. Under those circumstances, the condition necessary for CS to work is
that one type of base (e.g., G) does not prefer to change to one specific base (e.g., C) while
its counterpart (in this example, A) prefers to change to the other one (here, T). The condition
may be expressed as

As long as this condition is met, CS should also apply to sequences whose base compositional
biases are caused by directional mutation pressure (Sueoka, 1988; M. Bulmer, personal
communication), rather than selection.
COMPOSITIONAL STATISTICS EVALUATION 135

2. Global selection on the base composition. Given that the base com-
position has to be maintained in equilibrium, modeling selective pres-
sure according to the base frequencies is the simplest way of describing
this step of sequence evolution.
Based on this model, CS calculates the expected frequencies of phylo-
genetically informative patterns from the additional information present
in the sequences. First, a base composition matrix is constructed. The
frequencies of all bases at those positions in which one and only one
sequence differs by a transversion from the other three sequences are
entered in this matrix. The expected frequencies of occurrence of all in-
formative nucleotide patterns are then calculated. We recognize that in
this step, a multiple-hit correction should be incorporated. Such a correc-
tion is not available at present, however.
We stress that these frequencies are expected under evolution on the
peripheral branches only. Thus, the null hypothesis of the test can be
paraphrased as "no central branch." If there is no central branch, the
distribution of informative patterns of the topology which is being tested
will not exhibit significant deviations from the expected values. However,
if one of the topologies is correct, there will have been transversions on
its central branch. These contribute a number of parsimony patterns that
shift the observed distribution of informative patterns away from the ex-
pected one. Given a long enough central branch, the null hypothesis ("no
central branch") will be rejected, leading to significant support for that
topology.
The statistical test of CS is the following chi-square (x2) test:

where "pars" and "back" are observed numbers of parsimony and back-
ground patterns, respectively; ppaK and pback are the expected frequencies
of parsimony and background patterns (as calculated by CS), with/>pars +
Pback = 1; a°d n is the sample size with n = pars + back.
In CS and EP, the difference between observed and expected back-
ground patterns will follow a distribution whose mean should be 0 for an
incorrect topology. If t equivalent tests are conducted (for example, with
t different genes from the same four organisms), for one half of all differ-
ences dt,

where 0 < i ^ t, the inequality

should hold. We call this an "overrepresentation of background patterns."

For the other half, dt < 0. By contrast, none of the tests of the correct
topology should show an overrepresentation of background patterns be-
cause the substitutions on the central branch moved the distribution of d,
136 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

away from its mean at 0. Thus, for all i,

should hold, unless multiple transitions on the informative patterns ran-

domized the distribution of parsimony and background patterns. It is there-
fore possible to identify the correct topology by comparing the number of
times equivalent topologies display an overrepresentation of background
patterns. An example is given in Sidow and Wilson (1990) where two
incorrect topologies of the RNA polymerase tree were identified using this
approach. This approach thus adds power to the statistical tests of EP and
CS.

RESULTS

Comparison of CS and EP

In order to determine whether CS's modification of EP is valuable under

a range of conditions of sequence evolution, we conducted a series of
computer simulations. Table 7-1 and Figure 7-3 show the parameters used
in the simulations. The probability of a transition occurring was twice that
of a transversion.2 The branch lengths were chosen such that the total
number of topology-relevant patterns (i.e., the sample size in the statistical
test) was, on average, the same for both the correct tree and the (incorrect)
tree that was favored by parallel events on paraphyletic branches. After
each simulation, the generated sequences were analyzed with both EP and
CS.
The performance of both methods was evaluated by three criteria: (1)
rates of failure (i.e., the percentage of simulations in which the incorrect
topology received a x2 value of greater than 3.84, which is the boundary
for 95% confidence with one degree of freedom); (2) rates of success (i.e.,
the percentage of simulations in which the evolutionarily correct topology
received a x2 value of greater than 3.84); and (3) overrepresentation of
background patterns (i.e., the percentage of simulations in which the in-
correct topology shows an overrepresentation of background patterns).
We are showing only the results for the simulations in which all 10,000

2
We also conducted simulations using a fivefold transition bias. Under the conditions used,
the error rates of CS and EP are very similar to those that we found for the twofold transition
bias. However, there is a substantial loss of information for the correct topology because a
high transition bias randomizes the distribution of true parsimony patterns that were created
during evolution on the central branch. Real sequences with a high transition bias, such as
third positions of animal mtDNA codons, thus pose a serious dilemma to the phylogeneticist.
In such cases it is probably advisable to convert the sequences to two-character-state data,
distinguishing only between purines and pyrimidines. A distance analysis may then give
sufficient resolution, but it has to be borne in mind that "the proper distance measure" (sensu
Jin and Nei, 1990) may be impossible to find in real data. Compositional statistics and EP
should be immune to the problem if applied to nuclear genes because their mode of evolution
does not show a high transition bias.
COMPOSITIONAL STATISTICS EVALUATION 137

Table 7-1 Simulation Parameters in the Comparison

of Evolutionary Parsimony and Compositional Statistics
Parameter Value
Constant
Number of replications 200
Transition bias 2*
Number of bases 10,000
Fraction of purines 0.5f
Variable
Base composition (G or T) 25% to 45%*
Variability of positions 1 or 1/100 to 1s
Transition probabilities Calculated from"
base composition
'For every transversion, two transitions were simulated. On average,
therefore, each transition was four times as likely as each transversion.
'The sequences always contained 50% purines and 50% pyrimidines;
setting the frequency of one base within each class therefore constrains
the frequency of the other one.
'The proportions of G and T were changed in increments of 5%. There-
fore, there are 25 possible base compositions, from [25%G ; 25%T] to
[45%G ; 45%T]. Ten of them (e.g., [30%G ; 25%T] and [25%G ;
30%T]) are twofold redundant, and simulations were only performed
once for each pair.
!
In the case of equal variability, all 10,000 positions had a probability
of one, of accepting a substitution. When the variability varied, an equal
number of positions in the sequence accepted substitutions with a prob-
ability of 100/100, 99/100, 98/100, . . ., 1/100. If a substitution was
rejected, the search for a position was repeated until a position accepted
a substitution.
"The conditional probabilities of transitions were calculated by the for-
mula P(G -> A | ts) = P(A -> G | ts) = F(G) F(A) / (F(G) F(A) + F(T)
F(C)) and analogously for T and C. Therefore, they are also dependent
on the base composition. If, for example, G is much more abundant
than A, there will be significant mutation pressure toward equilibration.
In our model, this process will be counterbalanced by selection.

Figure 7-3 The branch lengths used in

all analyses. Shown are the number of
transversions on each branch. In the
simulations in which the probability of
nucleotide positions to change varied
100-fold, branches leading to 3 and 4
were slightly shorter than for the other
analyses to generate a roughly equal
number of informative patterns for the
correct and incorrect topologies.
138 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Figure 7-4 Three-dimensional plots

showing the low rate of failure of CS
compared to EP. The percentage of
times the incorrect topology was cho-
sen as shown on the vertical (Z) axis is
plotted against a range of different base
compositions (X and Y axes). (A) Ex-
pected distribution of failure when the
confidence limit is set to 95%. (B) Fail-
ure by CS. (C) Failure by EP.

nucleotide sites were equally variable. The results for those simulations in
which the variability differed 100-fold among positions were very similar
to those shown (see below).
Rates of Failure
In Figure 7-4, the number of times a method chose the incorrect topology
at statistical significance is plotted against the base compositional bias. Five
percent failure is expected because the confidence interval is set to 95%
(Fig. 7-4A). The error rate of CS exceeds the 5% ceiling only under extreme
compositional inequalities (Fig. 7-4B). This is probably due to the lack of
a multiple-hit correction during construction of the base composition ma-
trix. By contrast, EP fails at a higher frequency than 5% (Fig. 7-4C), even
under relatively mild compositional inequalities.
Rates of Success
The results (with respect to the correct topology) are shown with the same
kind of plot as for the incorrect topology (Fig. 7-5). The performance of
EP (Fig. 7-5A) has to be considered in context with the rates of failure in
COMPOSITIONAL STATISTICS EVALUATION 139

Figure 7-5 Success plots depicting the

percentage of times EP and CS calcu-
late a x2 value of greater than 3.84 for
the correct topology. (A) EP. (B) CS.
These plots are analogous to those in
Figure 7-4, but note the reversal of the
values on the X and Y axes (B).

Figure 7-4C. As the base compositional biases become more pronounced,

EP chooses both the correct and the incorrect topology at statistical sig-
nificance. This reflects the systematic error due to violation of HP's as-
sumptions. Generally, the higher x2 value still identifies the correct to-
pology, but with less consistency than does CS (not shown). Compositional
statistics chooses the correct topology at statistical significance, but the
performance decreases as base compositional inequalities become extreme
(Fig. 7-5B). The latter phenomenon can be attributed to the loss of power
of the method when base compositional inequalities, like the ones we
simulated, cause convergent evolution in conjunction with large rate dif-
ferences among lineages. Naturally, no method of phylogenetic inference
is robust against convergent evolution. Therefore, data sets with extreme
base compositional biases should be avoided.
Rejection of the Incorrect Topology by an Overrepresentation of
Background Patterns
The simulations are in agreement with the theoretical considerations above.
The expected 50% value (Fig. 7-6A) is not approximated by HP's plot
140 PHYLOCENETIC ANALYSIS OF DNA SEQUENCES

Figure 7-6 Overrepresentation of

background patterns for the incorrect
topology plotted against base composi-
tion. Note the reversal of the X and Y
axes. (A) Expected distribution of 50%.
(B) Approximation of the expected dis-
tribution by CS. (C) Deviation from the
expected distribution by EP.

(Fig. 7-6C). Compositional statistics, however, shows a distribution closer

to the expected one (Fig. 7-6B) and identifies an Overrepresentation of
background patterns in about 50% of all incorrect topologies. When the
correct topology was tested with CS, the values rose above 10% only under
the most extreme base compositional inequalities (not shown). The dis-
tribution remained generally low despite the significant probability of mul-
tiple hits under the branch lengths used. Thus, CS can still identify incorrect
topologies even when the correct one is not chosen at statistical significance
any more.

Comparison of CS With a Distance Method

Using the same simulation parameters as those for comparing EP and CS
(see Table 7-1; Fig. 7-3), we repeated the simulations for analysis by a
COMPOSITIONAL STATISTICS EVALUATION 141

Figure 7-7 Results from simulations addressing the performance of a distance method
(neighbor-joining) as a function of base compositional bias. X axis, base composi-
tional bias as calculated by formula (16). Y axis, length of the central branch as
calculated by the distance method. (Actual length = 500 substitutions.) Solid sym-
bols denote central branch lengths for the correct topology, open symbols denote
central branch lengths for the incorrect topology. The continuous line connects data
points of simulations in which all positions in the sequences were equally likely to
change. The dashed line connects data points of those simulations in which the
positions varied in their likelihood to undergo change (Table 7-1). Arrows show
critical values of the base compositional bias. For values to the right of the arrows,
the method begins to assign a central branch to the incorrect topology.

distance method. Because of the computational time required, we decided

to use a shortcut in assessing the performance of the method. From 50
simulations, the average of the observed distances was corrected for mul-
tiple hits, distinguishing transitions from transversions (Kimura, 1980). This
is a valid approach because the length of the sequences and the number
of substitutions are great enough for the variance in observed distances to
become small, when individual simulations are compared. The resulting
distance matrix was analyzed by the neighbor-joining program (Saitou and
Nei, 1987). The length of the central branch as calculated by the neighbor-
joining algorithm was plotted against the base compositional bias (Fig. 7-
142 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

7) as calculated by the formula

where subscript / denotes the bases (G, A, T, or C), and b, the frequency
of base / in the sequence. The analysis was then repeated for the same 15
base compositions as used for CS and EP and both types of positional
variabilities (see Table 7-1). The following observations are of importance.
1. When all positions are equally variable (Fig. 7-7, solid line), the
method underestimates the length of the central branch of the correct
topology as soon as the base composition begins to be skewed. When
the base compositional bias becomes greater than 0.1, it starts choos-
ing the incorrect tree (solid arrow).
2. When the positions in the sequence are assumed to vary with respect
to probability of change (Fig. 7-7, dashed line), the distance method
significantly underestimates the length of the central branch in all
cases, even when the base composition is not skewed at all. This is
in agreement with Lake's simulation study, which specifically ad-
dressed positional variability (J. Lake, unpublished). When base com-
positional biases become greater than 0.06, the distance method chooses
the incorrect tree3 (dashed arrow). Thus, under conditions of com-
positional bias, CS appears to be significantly more robust than a
distance method.
DISCUSSION

Comparing CS and EP

Our computer simulations show that CS is a robust and general method

of phylogenetic inference. Just like EP, it functions under a variety of
different conditions. Most importantly, it is unaffected by a wide range in
variability among different positions of the sequence. This is a major strength
of both CS and EP. Distance measures, by contrast, require corrections
based on significant assumptions about the substitution process to achieve
additivity (Jin and Nei, 1990).
Another characteristic of both CS and EP is that three separate tests are
required to identify the correct topology. This is due to the nature of the
informative nucleotide patterns. The sets of informative positions do not
overlap, and are independent of one another. This may seem to be a
drawback, but it has an advantage: systematic errors are usually identified
when there are significant results for more than one topology. Other meth-
ods choose a topology by a process that cannot determine whether other
hypotheses would be equally consistent with the data.
3
Any reduction in the length of the central branch of the simulated tree would increase the
likelihood of the distance method choosing the other (incorrect) tree. Evolutionary parsimony
or CS would not choose an incorrect tree but would simply indicate that no decision can be
made.
COMPOSITIONAL STATISTICS EVALUATION 143

The main difference from EP is that CS takes into account base com-
positional biases in its model of sequence evolution. We have shown that
CS performs significantly better than EP under biased conditions. Com-
positional statistics also appears to be robust in cases in which the substi-
tution process satisfies reversibility independent of the base composition.
That is, when P(A -* T) = P(T -» A), P(G -» T) = P(T -» G), etc., for
all transversions, CS's performance was very similar to that reported here
(not shown; see footnote 1).
When there are no base compositional biases, and/or transversions are
balanced, CS becomes identical to EP. In this case, CS gives ppars = pback
= 0.5, the value assumed by EP. Thus, CS is unlikely to do worse than
EP. In addition, since EP tends to fail more often than in 5% of all tests
even in cases when the base composition is not highly biased (see Fig. 7-
4), we recommend that CS be used, if applicable. In practice, however, it
may be easier to use EP, if a preliminary analysis using CS's algorithm
shows that there is no deviation from the equality ppars = pback.
Compositional statistics' performance also appeared to be more reliable
than that of a distance method, when base compositional bias was present.
It is especially noteworthy that the bias and different positional variability
affect the performance of the distance method synergistically, as can be
seen in Figure 7-7.
Under conditions of base compositional disequilibrium, CS's assump-
tions (as well as those of most other methods of phylogenetic inference)
are violated. Since disequilibrium is common between distantly related
sequences, this is a potential problem to be aware of. Compositional sta-
tistics may identify certain kinds of disequilibrium (Sidow and Wilson, 1990:
appendix II), but the specific circumstances under which disequilibrium
affects a phylogenetic analysis have not been studied.

Applications for CS

The robustness of CS renders it particularly useful for:

1. orthologous genes with a wide range of variability among nucleotide
positions within the sequences;
2. paralogous genes in which unequal rate effects, base compositional
inequalities, or other phenomena may render other methods inappl-
icable; and
3. analyses involving highly divergent sequences with strongly varying
branch lengths.
The recent advancement of phylogenetic inference into the realm of
statistical analysis, together with an explosion of nucleotide sequences, is
generating high standards for phylogenetic analysis. Statistically significant
results are now important to lend support to one's conclusions. A drawback
of both CS and EP is that they are conservative methods; they need a lot
of sequence data to attain sufficient sample sizes for their statistical tests
144 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

to be reliable. However, the conditions under which we tested them are

rather extreme and the length of the central branch was small compared
to the peripheral branches (see Fig. 7-3). The number of homologous
nucleotides needed to obtain a significant result with real sequences de-
pends on a number of factors; for example, the length of the central branch
compared to the peripheral ones, the base compositional biases, and the
number of available taxa.
As more sequences become available from a range of organisms, the
relatively low sensitivity of CS and EP will be less of a limitation. An
example for this is our phylogeny of RNA polymerase sequences repre-
senting the three deepest branches in the tree of life (Sidow and Wilson,
1990). By two criteria, CS was successful in that analysis: (1) only one
topology was chosen more often than the expected 5% and was therefore
inferred to be the correct one; and (2) for both other topologies, back-
ground patterns were overrepresented in approximately 50% of all tests,
in contrast to the selected one which did not show this pattern. Of course,
not nearly all problems in phylogenetic analysis will require application of
CS. Maximum likelihood methods are better if the sequences' mode of
evolution is closely approximated by the parameters used in the analysis—
a requirement that is usually hard to meet. For closely related sequences,
parsimony or distance methods are clearly superior (Saitou and Imanishi,
1989; Jin and Nei, 1990). Yet when distant evolutionary relationships are
addressed, several phenomena may cause such methods to produce serious
systematic errors. Such errors can be avoided by using a more general
method such as CS, the assumptions of which are less likely to be violated
by the substitution process.

SUMMARY

The performances of EP and CS were directly compared using computer-

generated DNA sequences. Both methods are robust under two adverse
(but frequently occurring) conditions: (1) a wide range in variability among
nucleotide positions; and (2) large differences in lengths of the terminal
branches. Therefore both methods are generally rate-invariant. A weak-
ness is their requirement for large data sets.
The major difference between CS and EP is their performance under
conditions that require maintenance of base compositional inequalities
(caused in our model by selective constraints on the DNA's primary struc-
ture). As the bias becomes more pronounced, two effects become apparent:
(1) EP begins to choose the incorrect tree at higher frequencies than al-
lowed by the confidence interval of its statistical test (CS, in contrast,
remains robust in such cases); and (2) identification of the incorrect to-
pology by an overrepresentation of background patterns becomes impos-
sible with EP, whereas CS retains this ability. A distance method also
showed decreased performance under the same simulation conditions.
COMPOSITIONAL STATISTICS EVALUATION 145

ACKNOWLEDGMENTS

We thank David Irwin, James Lake, Trang Nguyen, Cecilia Saccone, Na-
ruya Saitou, and Terry Speed for helpful discussion, and Thomas Kocher
for the formula for base compositional bias. Supported by grants from the
National Science Foundation and National Institutes of Health to ACW.

REFERENCES

Anderson, S., A.T. Bankier, E.G. Barrell, M.H.L. de Bruijn, A.R. Coulson, J.
Drouin, I.C. Eperon, D.P. Nierlich, B.A. Roe, F. Sanger, P.M. Schreier,
A.J.H. Smith, R. Staden, and I.G. Young. (1981) Sequence and organi-
zation of the human mitochondrial genome. Nature 290:457-465.
Bernardi, G., B. Olofsson, J. Filipski, M. Zerial, J. Salinas, G. Cuny, M. Meunier-
Rotival, and F. Rodier. (1985) The mosaic genome of warm-blooded ver-
tebrates. Science 228:953-958.
Felsenstein, J. (1978) Cases in which parsimony or compatibility methods will be
positively misleading. Syst. Zool. 27:401-410.
Felsenstein, J. (1981) Evolutionary trees from DNA sequences: a maximum like-
lihood approach. /. Mol. Evol. 17:368-376.
Felsenstein, J. (1989) Phylogenetic Inference Programs (PHYLIP), Manual 3.2.
University of Washington, Seattle, and University Herbarium of the Uni-
versity of California, Berkeley.
Fitch, W. M. (1986) The estimate of total nucleotide substitutions from pairwise
differences is biased. Philos. Trans. R. Soc. Land. [Biol. Sci.] 316:317-324.
Fox, G.E., E. Stackebrandt, R.B. Hespell, J. Gibson, J. Maniloff, T.A. Dyer,
R.S. Wolfe, W.E. Balch, R.S. Tanner, L.J. Magrum, L.B. Zablen, R.
Blakemore, R. Gupta, L. Bonen, B.J. Lewis, D.A. Stahl, K.R. Luehrsen,
K.N. Chen, and C.R. Woese. (1980) The phylogeny of prokaryotes. Science
209:457-463.
Gojobori, T., E.N. Moriyama, and M. Kimura. (1990) Statistical methods for
estimating sequence divergence. Pp. 531-550 in Molecular Evolution: Com-
puter Analysis of Protein and Nucleic Acid Sequences (R. F. Doolittle, ed.).
Methods in Enzymology, vol. 183. Academic Press, New York.
Gouy, M., and W.-H. Li. (1989) Phylogenetic analysis based on rRNA sequences
supports the archaebacterial rather than the eocyte tree. Nature 339:145-
147.
Hasegawa, M., H. Kishino, and T. Yano. (1985) Dating of the human-ape splitting
by a molecular clock of mitochondrial DNA. J. Mol. Evol. 22:160-174.
Holmquist, R., M.M. Miyamoto, and M. Goodman. (1988) Analysis of higher-
primate phylogeny from transversion differences in nuclear and mitochon-
drial DNA by Lake's methods of evolutionary parsimony and operator met-
rics. Mol. Biol. Evol. 5:217-236.
Jin, L., and M. Nei. (1990) Limitations of the evolutionary parsimony method of
phylogenetic analysis. Mol. Biol. Evol. 7:82-102.
Kimura, M. (1980) A simple method for estimating evolutionary rates of base
substitutions through comparative studies of nucleotide sequences. J. Mol.
Evol. 16:111-120.
146 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Kishino, H., and M. Hasegawa. (1990) Converting distance to time: application

to human evolution. Pp. 550-570 in Molecular Evolution: Computer Anal-
ysis of Protein and Nucleic Acid Sequences (R. F. Doolittle, ed.). Methods
in Enzymology, vol. 183. Academic Press, New York.
Lake, J.A. (1987) A rate-independent technique for analysis of nucleic acid se-
quences: evolutionary parsimony. Mol. Biol. Evol. 4:167-191.
Lake, J.A. (1988) Origin of the eukaryotic nucleus determined by rate-invariant
analysis of rRNA sequences. Nature 331:184-186.
Li, W.-H., and D. Graur. (1991) Fundamentals of Molecular Evolution. Sinauer
Associates, Sunderland.
Olsen, G. (1987) Earliest phylogenetic branchings: comparing rRNA-based evo-
lutionary trees inferred with various techniques. Cold Spring Harbor Symp.
Quant. Biol. 52:825-837.
Perrin, P., and G. Bernardi. (1987) Directional fixation of mutations in vertebrate
evolution. J. Mol. Evol. 26:301-310.
Saccone, C., G. Pesole, and G. Preparata. (1989) DNA microenvironments and
the molecular clock. /. Mol. Evol. 29:407-411.
Saitou, N., and T. Imanishi. (1989) Relative efficiencies of the Fitch-Margoliash,
maximum-parsimony, maximum-likelihood, minimum-evolution, and neigh-
bor-joining methods of phylogenetic tree construction in obtaining the cor-
rect tree. Mol. Biol. Evol. 6:514-525.
Saitou, N., and M. Nei. (1987) The neighbor-joining method: a new method for
reconstructing phylogenetic trees. Mol. Biol. Evol. 4:406-425.
Shoemaker, J.S., and W.M. Fitch. (1989) Evidence from nuclear sequences that
invariable sites should be considered when sequence divergence is calculated.
Mol. Biol. Evol. 6:270-289.
Sidow, A., and A.C. Wilson. (1990) Compositional statistics: an improvement of
evolutionary parsimony and its application to deep branches in the tree of
life. /. Mol. Evol. 31:51-68.
Sueoka, N. (1988) Directional mutation pressure and neutral molecular evolution.
Proc. Natl. Acad. Sci. USA 85:2653-2657.
Vigiliant, L., R. Pennington, H. Harpending, T.D. Kocher, and A.C. Wilson.
(1989) Mitochondrial DNA sequences in single hairs from a southern African
population. Proc. Natl. Acad. Sci. USA 86:9350-9354.
8
Weighted Parsimony: Does it Work?

WALTER M. FITCH AND JIA YE

In 1969, Farris suggested a relatively unbiased way of weighting the value

of a character for systematic purposes based upon the proposition that
characters that frequently change their state are unreliable guides to re-
lationships, whereas those that change only once are perfect guides to
relationships.
In nucleotide sequences, one can apply the same philosophy not only to
the various characters (homologous positions in the sequences), but to the
character changes as well. Thus, if transition changes (e.g., C «-» U) are
more common than transversion changes (e.g., C «-» A), then one might
give less weight to the former events than to the latter. In its extreme form
this leads to transversion parsimony where transition changes are not counted
at all.
Williams and Fitch (1989) devised a scheme to generalize such weighting
of character-state changes which proved to be identical to that of Sankoff
and Cedergen (1983).
A computer program has been developed that permits one to perform
either kind of weighting, or both simultaneously, plus some other options
which are explained in detail in Williams and Fitch (1990). (Some further
details of the program are given in the section, Weighted Tree Reconstruc-
tion.) The program is freely available by writing to the senior author.
In principle, weighted parsimony should give superior results to par-
simony with uniform weighting of characters and of state changes, because
weighting reduces the impact of noisy positions by giving emphasis to the
rarer events. In this work, we use simulation to test whether the principle
works in practice. More tests will be required before the value of weighting
can be assessed reliably, although the tests presented here suggest the
procedure is useful.

147
148 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

METHODS

The Simulation Program

Descent with change of an ancestral nucleotide sequence was accomplished
by computer. An ancestral sequence of 300 nucleotides is obtained by
random sampling of a pool of nucleotides of whatever composition the
experimenter chooses. In these tests, all nucleotides were equiprobable.
The descent was according to the tree shown in Figure 8-1 where the
numbers on the branches are the number of nucleotide substitutions that
are accepted between branchings. The experimenter may choose other
values.
The choice of the nucleotide substitution was random with the probability
of the transition being 0.8, and the probability of the two transversions
being equal to 0.1 each. Again the experimenter may choose other values.
The choice of the nucleotide to be substituted was random, but only
from among the variable positions. The experimenter may decide on any
fraction, v, of the nucleotides to be variable and thus 1 - v of them are
invariable. A set of nucleotide sites from the sequence is randomly assigned
to the variable set.
To model the property of the variable sites changing as nucleotide sub-
stitutions are fixed [concomitantly variable codons or "covarions," (Fitch
and Markowitz, 1970); concomitantly variable nucleotides or "covario-
tides," (Fitch, 1986; Shoemaker and Fitch, 1989)1, a persistence (of vari-
ability) parameter, p, is set by the experimenter. Thus, 1 - p is the
probability that, after each substitution of a variable nucleotide, one ran-
domly chosen member of the currently variable set will be exchanged with
a randomly chosen member of the currently invariable set of nucleotides.
There is no restriction on the number of times a position may change its
variability status. This program is also available from the senior author.

Figure 8-1 The simulated tree. An ancestral sequence of 300 nucleotides was
allowed to evolve as depicted, the number of nucleotide substitutions between nodes
being shown on the figure. The resulting sequences at the 12 tips were used to test
weighted and uniform parsimony procedures that attempt to recover the correct tree.
Branch lengths are proportional to the number of substitutions thereon so that for
this model the molecular clock is perfect. The length of the tree is 880 nucleotide
substitutions.
WEIGHTED PARSIMONY: DOES IT WORK? 149

Weighted Tree Reconstruction

As a result of the preceding process, one obtains a set of sequences (in

this case 12) corresponding to the tips of whatever tree one gives the
program (Fig. 8-1). It is necessary then to try to recover, from the sequences
alone, the phylogeny depicted in Figure 8-1. The process normally involves
providing a starting tree, either by the program or the investigator, and
determining that tree's length, that is, determine the minimum number of
nucleotide substitutions required to account for the descent of these se-
quences given the topology of the tree. This is a simple and well-known
procedure (Fitch, 1971). This tree topology is then modified by neighbor-
branch swapping, and the process repeated to see if a more-parsimonious
tree (= shorter length = fewer substitutions) can be found. The process
of topology modification continues until a tree is found for which none of
the allowable swaps produces a more-parsimonious tree. That minimal tree
is then considered to be the best estimate of the true topology.

Establishing Parameters

In the initial phases, we wished to determine the value of v (fraction of

sequence that are covariotides),p (persistence), and the lengths of branches
that would produce errors when uniform parsimony was applied, so that
we could see if weighted parismony did better. Thus, we always gave the
program the correct tree to start from so that we would be guaranteed that
its shortest tree was no longer than the correct tree. Once we had such
values, a similar procedure was used for the test of the method, but now
we examined all ten sets of sequences produced whether or not uniform
parsimony obtained the correct tree. This is because we need to know not
only how often weighted parismony improves upon the errors of uniform
parsimony, but also how often it does worse than uniform parsimony.

RESULTS

The effect of the number of covariotides and their persistence seemed to

have little effect in terms of causing uniform parsimony to get the wrong
tree compared to the effect of branch length. Analysis showed that until
the number of positions changing more than once in two contiguous branches
of the tree was sizeable, uniform parsimony kept getting the correct tree.
Hence, we ended up with a tree containing nearly three substitutions per
site in order to get incorrect trees with any frequency.

Nature of the Test

The test actually involves only ten simulations. These represent our first
efforts, and the small sample size means that all inferences are weak and
are only suggestive. To understand the results, a few details should be
150 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

understood about the weighting procedure. Subject to the initial conditions,

the procedure looks for the best tree it can find. On the basis of that tree,
it establishes new weights for the positions and for the nucleotide substi-
tutions. Using those new weights, it takes a starting tree (in this work the
starting tree of successive passes was always the best tree of the previous
pass) and again looks for a better tree. The succession of passes continues
until the same (best) tree topology is obtained on two passes in a row.
(There is provision for detecting the rare occurrence of cycling which did
not occur in this study.)
Our weighting procedure has 20 variant forms in addition to uniform
parsimony. In uniform parsimony, which is what most investigators use,
only one pass is performed and all positions have the same weight, as do
all nucleotide changes. The other 20 variants are a function of three prop-
erties: (1) initial conditions; (2) substitutional change symmetry; and (3)
type of position weighting.
Initial (first pass) conditions for nucleotide change weighting are of three
forms. The first is all changes equal (uniform). The second is based upon
the frequency of positions that have different pairs of nucleotides (exis-
tential). The third is like the second except the count is over the number
of ways of selecting a nucleotide pair in each position, summed over all
positions (combinational) (see Williams and Fitch, 1989, 1990, for further
explication).
Substitutional change symmetry is whether or not one requires that the
weight of a change in one direction must be equal to the weight in the
reverse direction, (e.g., A —» C = C —» A?).
Type of position weighting can be uniform, but after the first pass can
be related inversely to the number of changes in that position observed in
the best tree of the previous pass. For the latter, positions can be weighted
inverse linearly or inverse quadratically. In these tests, we performed the
procedure for all 21 parameter combinations and asked how many of the
20 non-uniformly-weighted ways gave the correct tree. Unlike uniform
weighting, two different topologies seem never to give the same weighted
score, although we could synthesize data that would.

Outcome of the Tests

The ten tests gave results summarized in Table 8-1. Although there are
four combinations of p and V, we held V(\ — p) = constant = 25. Tests
1, 2, 5, and 6 represent a standoff in that both uniform parsimony and all
20 versions of weighted parsimony were all right or all wrong. Test 3 might
seem to be a bad case for weighting in that three of the 20 weighted trees
were incorrect, but they differed only by the single swap of sequences 10
and 11.
Test 4 is in fact a worse case for weighting. Four of the five most-
parsimonious uniform trees are, of course, wrong, but none of the nine,
wrong, weighted, best trees are among the four most-parsimonious trees
WEIGHTED PARSIMONY: DOES IT WORK? 151

Table 8-1 Uniform versus Weighted Parsimony*

Uniform Weighted
Test V P + - + - E Length
1 100 0.75 1 20 220 340
2 100 0.75 1 20 220 370
3 50 0.5 1 17 3 440 315
4 25 0.0 1 4 11 9 880 353
5 100 0.75 1 20 220 352
6 50 0.5 3 20 440 313
7 100 0.75 1 14 6 220 331
8 50 0.5 1 14 6 440 313
9 250 0.9 1 14 6 88 502
10 100 0.75 1 18 2 220 341
"Test is a simple numbering of the ten cases; V is the number of covanotides among the 300 nucleotides;
p is the persistence of variability; under "uniform" and "weighted" are given the number of times the
correct ( + ) and incorrect ( — ) tree was obtained; E is the expected number of times a covanotide exchanged
places with an invariable nucleotide; length is the number of substitutions in the most-parsimonious tree,
using uniform weighting irrespective of whether it was the correct tree. Tests 4 and 6 had five and three
equally most-parsimonious trees under uniform weighting; one of the five was the correct tree. The correct
tree was never worse than the best tree by more than four substitutions, nor by more than three neighbor
branch swaps.

that are incorrect. In fact, they are one extra branch swap farther away
topologically from the correct tree than are the four incorrect most-
parsimonious trees of uniform parsimony. Tests 7 to 10 get increasingly
larger numbers of correct trees (where uniform parsimony gets the wrong
tree) and thus support weighting.

DISCUSSION

The first observation is that parsimony, even unweighted, is clearly robust

for the tree presented. We had to simulate a tree with many substitutions
before parsimony gave the wrong answer, and even then the tree was nearly
right, usually being only one branch swap from the correct tree. Of course
the trees would have gone wrong sooner if we had had fewer substitutions
on the internodal branches.
On the other hand, there was poor recovery of the actual substitutions,
generally only about 40% of the total, and not all of those were the correct
changes in the correct positions. Of course, recovery would have been
better had there been fewer substitutions along the branches.
We had thought that reducing the number of covariotides would, by
increasing the frequency of second substitutions in the same site in neigh-
boring branches, cause the parsimony procedure to go wrong with fewer
substitutions. This appears to be true, but with less strength than antici-
pated. We suspect that this arises, at least partly, because the occasional
removal of a changed covariotide from the variable group makes it im-
possible to destroy the evidence it bears by additional homoplasious changes.
The second observation is that weighting indeed improves the perform-
152 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

ance. Excluding the standoff cases (tests 1, 2, 5, and 6), weighting more
commonly got a better tree. Even in test 6, where none of the weighted
trees was correct, in all 20 cases the tree was the same and was only one
branch swap removed from the correct tree, whereas two of the three
incorrect best uniformly weighted trees were two branch swaps away from
the correct tree. Only in test 4 were the weighted trees two branch steps
removed from the correct tree, while the incorrect uniform trees were but
one swap away.
What would be of greater interest is to discover if the benefits of weight-
ing are more pronounced for some of the 20 options than others. This
turns out to be the case, although, not surprisingly, no single option is
invariably superior to all other options.
Table 8-2 shows how well each of the 21 choices fared for the six non-
trivial cases. The pair of values in the upper left corner (two right, four
wrong) are the results of uniform weighting (usually misleadingly called
unweighted) of both positions and transformations. The column, as it con-
tinues below that cell, shows the effects of nonuniform-transformation
weighting while maintaining uniform positional (character) weighting. The
top row to the right of that cell shows the effect of nonuniform-character
weighting while maintaining uniform-transformation weighting. The re-
mainder of the table shows the effect of combinations of nonuniform po-
sitional and transformation weighting.
From this, it appears that: (1) if there is uniform transformational weight-
ing, nonuniform positional weighting does not help. We believe this may
be because the simulation did not have a sufficiently great enough variance

Table 8-2 Frequencies of Getting the Correct Tree

Positions*
U L Q
Matrix* + _ + _ + _ Initial*
V 2 4 2 4 2 4 u

4 2 3 3 3 3 u
Ls 4 2 6 0 5 1 e
4 2 5 1 5 1 c

5 1 3 3 3 3 u
La 5 1 6 0 6 0 e
5 1 5 1 5 1 c
'Positions may be weighted uniformly (If), inversely proportional to the number of substitutions they
have (linear, L), or inversely proportional to the square of the number of substitutions they have (quadratic,
0.
The transformation matrix may be weighted uniformly (U) or inversely proportional to the number of
times a particular kind of nucleotide changed to any particular other nucleotide (linear, L) except that it
can be either symmetric, Ls, or asymmetric, La, depending upon whether or not one counts, e.g., A —»
C with C —* A. For Ls, one adds them together and divides by two.
•For the initial pass, when there is no count of changes from a best tree, the starting transformation matrix
may be uniform (u), based on the number of positions containing a pair (existential, e), or based on the
number of ways a particular pair of nucleotides might be drawn at random (combinatorial, e)
WEIGHTED PARSIMONY: DOES IT WORK? 153

in the number of substitutions per nucleotide. Nevertheless, such weighting

often improves the result when there is also transformational weighting;
(2) transformational weighting was effective even in the absence of posi-
tional weighting; (3) having the initial transformation matrix set to some-
thing other than all one's (rows with e or c on the right) was beneficial;
(4) both positional and substitutional weighting should be used; (5) quad-
ratic weighting does not seem more likely to lead to the correct tree than
linear weighting; and (6) an asymmetric substitutional weighting does slightly
better than symmetric weighting. The last choice may, however, be a con-
sequence of using the correct root. In cases where there is not a clear
outgroup, asymmetric weighting might be counterproductive as the direc-
tionality of each change will remain questionable.
The third observation, to be detailed elsewhere, is that application of
the Fitch and Markowitz (1970) procedure to estimate the number of cov-
ariotides in these trees almost always obtained the correct value within
10%. Thus, the parameter that complicates the model, the number of
covariotides, is in fact estimable from the sequence data.
Our conclusion from these tests, although subject to the performing of
many more tests, is that weighted parsimony is more likely to get the correct
or a better tree. There is added force to this conclusion to the extent that
the model is more realistic because of its inclusion of covariotides, a feature
we regard as an important and necessary reflection of biological reality
and a feature not heretofore an important aspect of simulations. The model
is unrealistic in other ways, particularly in its clocklike regularity, a con-
dition imposed to cause all trees in the test to have the same underlying
structure.

ACKNOWLEDGMENTS

This work was supported by National Science Foundation grant BSR-

9096152.

REFERENCES

Farris, J.S. (1969) A successive approximations approach to character weighting.

Syst. Zool. 18:374-385.
Fitch, W.M. (1971) Toward defining the course of evolution: minimum change for
a specific tree topology. Syst. Zool. 20:406-416.
Fitch, W.M. (1986) The estimate of total nucleotide substitutions from pairwise
differences is biased. Philos. Trans. R. Soc. Lond. [Biol. Sci.] 316:317-324.
Fitch, W.M., and E. Markowitz. (1970) An improved method for determining
codon variability in a gene and its application to the rate of fixations of
mutations in evolution. Biochem. Genet. 4:579-593.
Sankoff, D. and R.J. Cedergren. (1983) Simultaneous comparison of three or more
sequences related by a tree. Pp. 253-263 in Time Warps, String Edits, and
154 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Macromolecules: The Theory and Practice of Sequence Comparison (D.

Sankoff and J.B. Kruskal, eds.). Addison-Wesley, Reading.
Shoemaker, J.S., and W.M. Fitch. (1989) Evidence from nuclear sequences that
invariable sites should be considered when sequence divergence is calculated.
Mol. Biol. Evol. 6:270-289.
Williams, P.L., and W.M. Fitch. (1989) Finding the minimal change in a given
tree. Pp. 453-470 in The Hierarchy of Life. Molecules and Morphology in
Phylogenetic Analysis (B. Fernholm, K. Bremer, and H. Jornvall, eds.).
Elsevier Science Publishers B.V., Amsterdam.
Williams, P.L., and W.M. Fitch. (1990) Phylogeny determination using a dynam-
ically-weighted parsimony method. Pp. 615-625 in Molecular Evolution:
Computer Analysis of Protein and Nucleic Acid Sequences (R. F. Doohttle,
ed.). Methods in Enzymology, vol. 183. Academic Press, New York.
9
Testing the Theory of Descent

DAVID PENNY, MICHAEL D. HENDY, AND MICHAEL A. STEEL

If it could be demonstrated that any complex organ existed which could not
possibly have been formed by numerous, successive, slight modifications, my
theory would absolutely break down (Darwin, 1859: 189).
The comment of Popper (1976:168) that "Darwinism is not a testable
scientific theory, but a metaphysical research program—a possible frame-
work for testable scientific theories" is sometimes used to question the
scientific status of evolutionary theory. The original comment was not in
any way "antievolution," and indeed, Popper noted "the strange similarity
between my theory of the growth of knowledge and Darwinism" (Popper,
1976:169). Later, Popper (1978, 1984) limited these criticisms, but this has
not always satisfied critics. One of our interests has been to determine,
without ambiguity, if evolutionary theory could meet Popper's criteria for
the demarcation of science.
The introductory quote given above is one of many examples of how
Charles Darwin considered his theory of evolution to be falsifiable. Al-
though Darwin is remembered today for the general aspects of theory of
evolution, in his day-to-day research, he made many predictions from his
theory—then sought (or made observations) to test the predictions. Some
of the best known involve his experiments on plants, including work on
orchid flowers (both in structure and function), pin and thrum flowers of
primrose (and other dimorphic and trimorphic forms of flowers), and on
the power of movement of plants (Ghiselin, 1969; Allan, 1977; Penny,
1985). This last example of the power of movement in plants is particularly
interesting in that he had left a record of his reasoning. The first extract
is from his autobiography.
For in accordance with the principles of evolution, it was impossible to account
for climbing plants having been developed in so many widely different groups,
unless all kinds of plants possess some slight power of movement of an anal-
ogous kind.

155
156 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

And again, the delightful comment in a notebook from 1839,

Is there any very sleepy mimosa, nearly allied to the Sensitive Plant?
(Details of sources are in Penny, 1985). The prediction that all plants should
have some "slight power of movement" led to two books on his research,
The Movements and Habits of Climbing Plants and The Power of Movement
in Plants. Many other examples of how Darwin used predictions to direct
research could be given, including features of human evolution.
It was not just a coincidence that Charles Darwin was looking for pre-
dictions from theories. During the early stages of the development of his
theory (after the voyage of the Beagle), Darwin read widely in many areas
of science. This was partly in order to understand what was required of a
good scientific theory. His reading included works on the philosophy of
science by Herschel, Whewell and Comte (Ruse, 1975; Schweber, 1977).
Each of these authors emphasized the importance of prediction in science
and, judging from letters to Charles Lyell (Schweber, 1977), Darwin was well
aware of the importance of a good scientific theory leading to predictions.
In this century, Karl Popper (1963, 1972) has developed this theme
further and emphasized the potential falsifiability of scientific theories. A
theory is more useful (and therefore better) if it prohibits (excludes) a
larger proportion of possible observations. The approach claims to be both
descriptive and prescriptive. It is descriptive in that it claims to describe
how the most effective scientists have worked,'and it is prescriptive in that
it advocates how scientists should aim to work. Hypotheses (or conjectures)
are considered tools that are judged on their effectiveness in helping sci-
entists devise new and more powerful tests whether the tests be observa-
tional, experimental, or analytical. None of this denies that sociological
and cultural factors play a role in science—as long as such factors are
considered descriptive of scientific procedures and not prescriptive of how
scientists should select theories. During the past two decades there has
been a strong anti-intellectual movement suggesting that hypotheses are
largely determined by cultural factors, and consequently, to this extent are
arbitrary. Such a movement has failed to develop falsifiable predictions
and is not scientific by Popper's criterion for the delimitation of science.
In this chapter we review our approach to the study of evolutionary
trees. This has been developed within a strong Popperian framework (Rid-
diford and Penny, 1984) of aiming to develop falsifiable hypotheses. After
discussing some of the general issues involved, we then discuss the question
of how good methods are for inferring trees, particularly from molecular data.

IS EVOLUTION A SCIENTIFIC THEORY?

Similar Trees From Different Sequences

The simple prediction from the "theory of descent" is that, because the
sequences share the same tree pattern of ancestry, the optimal trees from
TESTING THE THEORY OF DESCENT 157

different sets of data should be similar. (From the proposed stochastic

nature of the mechanism of mutation and selection it would be surprising
if the trees were identical. Indeed, it would be more devastating to Dar-
winism if different sets of short sequences always gave identical trees).
Comparing results from different data sets had been used previously (e.g.,
Mickevich, 1978) and our interest was in getting quantitative results.
Our first major project was to compare trees derived for the same 11
mammalian taxa but from different sets of sequence data. We saw three
requirements for being able to test the prediction of similar trees from
different sequences. These were the ability to:
1. find the optimal tree(s) for 10 or more taxa;
2. find a tree comparison measure to compare trees objectively; and
3. derive the distribution of this tree comparison measure.
Finding the Optimal Tree
Sufficient taxa were required so that it would be most unlikely to get the
same tree by chance. Fortunately, sequence data for 11 taxa were available
for five proteins or peptides.
For 11 taxa there are 34,459,425 (17!!) binary trees. The number of trees
increases exponentially with the number of taxa; therefore, any method
which considers all trees cannot be efficient. It was shown quite early in
the project that searching trees for the optimal tree(s) is an example of a
set of problems known to be NP-complete (Graham and Foulds, 1982).
This result implies that it is most unlikely that an efficient method will ever
be found for a complete search on a large number of taxa. William Day
has extended this analysis to other related problems with trees (see Day
etal., 1986).
Finding optimal trees requires a program which is unbiased toward any
subclass of trees. For example, it could happen that a program tended to
group together adjacent taxa in the data matrix. An error of this nature
could lead to the trees from different sequences being more similar then
expected, not through common descent, but through program limitations.
A quite different problem is that an optimality criterion, for example par-
simony, could tend to favor some trees by bringing together isolated taxa.
This will be discussed later.
In 1980, a search through all trees for 11 taxa was estimated to require
55 days with the computers we had available. The solution to this problem
was the development of a branch and bound algorithm (Hendy and Penny,
1982), which is now widely used. It reduced the computing time for these
data to just under five minutes, while still guaranteeing to have found all
optimal trees. This met our first objective of finding optimal trees for the
five sets of data.
Finding a Tree Comparison Measure
When the parismony branch and bound program was applied to each of
the five different proteins, the minimal trees were not identical, but did
158 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Figure 9-1 Application of symmetric tree metric. Two trees selected from the min-
imal trees from /3-hemoglobin (top) and a-hemoglobin (bottom) sequences. On the
symmetric difference metric there are six differences. These are indicated by a small
cross on the three nonequivalent edges of each tree.

indeed look "similar" (Penny et al., 1982). The second objective was to
find a useful tree comparison metric that would allow an objective com-
parison of these trees. The symmetric difference metric had been developed
(Robinson and Foulds, 1981) from a parallel interest in trees and had an
efficient method for its calculation (Penny and Hendy, 1985b). A more
efficient method was developed by Day (1985).
The symmetric difference metric on two trees counts the number of edges
that occur in one, but not both, trees. Edges are equivalent if they partition
the taxa into the same two subsets (see Edge Bipartitions). This is repeated
for each edge of the tree in turn. Figure 9-1 shows two minimal trees from
different sequences which are distance six differences apart. The differences
between the two trees are indicated by a small cross on edges which have
no equivalent edge on the other tree.
TESTING THE THEORY OF DESCENT 159

Deriving the Distribution of the Tree Comparison Measure

At this point, it is unclear whether finding six differences is significant.
What is the probability that two randomly selected trees would have six
differences? In order to answer that question, the problem of deriving the
distribution of the tree comparison metric must be solved. The expected
distribution has been calculated for up to 16 taxa (Hendy et al., 1984) and
is shown graphically in Penny and Hendy (1986) and Hendy et al., (1988).
The distribution is highly asymmetric, which is useful for many biological
applications because it is particularly sensitive for closely-related trees. We
find the probability of randomly selecting trees on 11 taxa with six or fewer
differences (as in Fig. 9-1) is 4 x 10 "5.
Similar results were found from comparisons of minimal trees from other
sequences (Penny et al., 1982), where each pair of minimal trees from
different sets of sequences was more similar than expected by chance. Thus,
it is fair to claim that the original prediction that minimal-length trees from
different data sets would be similar, is supported. The method of analysis
allowed the possibility for the theory of descent to fail. We have more
confidence in the theory if it passes quantitative tests.
Our conclusion is that the theory of descent can meet the same quan-
titative standards as expected in other areas of science. There is no need
for "special pleading" that evolution is hard to quantify. The project led
to improved techniques in several areas for studying trees. The problem
of finding optimal trees was shown to be NP-complete (Graham and Foulds,
1982). Branch and bound methods were developed (Hendy and Penny,
1982) so that optimal trees could be found in reasonable time for up to at
least 16 taxa. A tree comparison metric (Robinson and Foulds, 1981) was
implemented by showing it could be calculated efficiently (Penny and Hendy,
1985b). The expected distribution of this metric was derived (Hendy et
al., 1984) so that quantitative tests can be made. The analysis of a complex
scientific problem in order to get falsifiable predictions can be a productive
approach to science.

Additional Predictions
Are Shorter Trees Better?
The parsimony criterion for an optimal tree assumes that shorter trees
(those requiring fewer changes) are better estimates of evolution than
longer trees. As a corollary of this we would expect that trees shorter for
one data set should also be shorter on other data sets. This is different
from the previous section which compared minimal trees from independent
data sets.
A test of this prediction is shown in Figure 9-2 where trees requiring 124
to 133 changes on the /3-hemoglobin data were selected. The lengths of
these trees were then determined on the combined sequences from cyto-
chrome c, fibrinopeptides A and B, and a-hemoglobin.
160 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Figure 9-2 Lengths of trees with ^-hemoglobin and other sequences. The shortest
trees for /f-hemoglobin sequences were selected. The lengths of these trees were
then recalculated on the combined sequences from cytochrome c, fibrmopeptides
A and B, and a-hemoglobin. On average, trees that require fewer mutations with
/3-hemoglobin, require fewer mutations with the new sequences. The bars are twice
the standard error of the mean. (Adapted from Penny and Hendy, 1985a.)

The results in Figure 9-2 show that, on average, trees requiring fewer
mutations with /3-hemoglobin require fewer mutations with other sequence
data. The test was repeated for each of the other four sequences (Penny
and Hendy, 1985a), giving even better results. Evolution is a stochastic
process, and the shortest tree on an individual sequence cannot be guar-
anteed correct.
Sampling Error and Convergence
It is generally recognized that the sequences used in a particular study may
be too short to give an accurate prediction. If the only problem is that the
sequences are too short, then it is expected that the optimal tree should
become a better estimate as the sequences become longer. We cannot, of
course, measure this difference directly. However, what can be measured
is whether optimal trees from different data sets become more similar as
sequences become longer.
Perhaps the best method for selecting subsets of data is by the random
resampling of columns. This can be done by either bootstrapping or jack-
knifing. In bootstrapping (Felsenstein, 1985; Penny and Hendy, 1985a),
subsets of columns from the data matrix are randomly selected with re-
placement. These subsets can have the same number of columns as the
original data. A column may be omitted from a particular subset, or se-
TESTING THE THEORY OF DESCENT 161

lected more than once. With jackknifing (Penny and Hendy, 1985a, 1986),
the subsets are randomly selected without replacement. Consequently, the
subsets must be shorter than the original sequence. Jackknifing methods
can also be divided into those where subsets may overlap and those with
disjoint subsets (where no column occurs in both subsets). In this latter
group, which we call hobbits or halflings, each subset contains no more
than half the columns.
Trees can be formed from these subsets by standard methods and the
results analyzed using a tree comparison metric. There have been two ways
of comparing results. Felsenstein (1985) determined the internal edges that
occur in at least 95% of the trees. Penny and Hendy (1985a, 1986) studied
the rate of convergence as longer sequences (subsets) were used.
With either form of jackknifing we have measured the average distance
between optimal trees (using the symmetric difference tree comparison
metric) from each subset of columns. What we would expect is that trees
from different subsets would become more similar as the subsets contained
more columns. This indeed is the case as is shown in Figure 9-3. Instead
of showing the value from the symmetric difference metric directly, we
have converted it, using the calculated distribution of the metric (Hendy
et al., 1984, 1988), to the equivalent number of trees. These results show
that even the five sequences for each mammal is insufficient to allow

Figure 9-3 Convergence with larger subsets of columns. Minimal length trees were
found for jackknife samples from the combined sequences for six proteins. The
samples contained 12.5%, 25%, 50% or 92% of the columns in the data matrix.
The symmetric difference metric was used to find the average distance between
optimal trees from each group of subsets. Optimal trees became more similar as the
samples became longer. The results allow four different methods to be compared
(see Penny and Hendy, 1986 for an explanation of the symbols and further dis-
cussion). (Adapted from Penny and Hendy, 1986.)
162 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

convergence to a single tree. In the cases we have studied (Penny and

Hendy, 1985a, 1986), the trees, as expected, become more similar as the
subsets become larger. This is independent evidence for evolutionary in-
formation in the sequences.
These resampling approaches allow tentative answers to several inter-
esting questions. Is it likely that a different optimal tree will be found if
more information is gathered for each taxon? If so, which trees? How large
a subset of trees is necessary to be confident of including the correct tree?
Do some methods for inferring trees converge faster than others as longer
sequences are used?
Is bootstrapping better than jackknifing with larger-and-larger samples?
The answer may depend on the application. In a taxonomic study, an author
may not wish to propose a new taxonomic category unless confident that
future work will support it. This is using stability of a single edge of the
tree as a criterion. In such a case, bootstrapping may be the preferred test
since it does not accept an edge if there is reasonable doubt. If, however,
the intent is to find the best estimate of the phylogeny (the tree which
gives the best prediction as more data become available), then convergence
may be suitable. Under these circumstances the aim may be to find how
large a subset of trees is required to be confident that the subset includes
the correct tree. It must be noted that although the test allows a decision
as to whether convergence has occurred, it is still possible for methods to
converge to an incorrect tree. This is discussed later (see Hadamard Trans-
formations) .

Testing Other Models

From a Popperian viewpoint, an important feature of the scientific ap-
proach is that scientists should be able to give rational explanations of why
they use a particular theory. There is an asymmetry here in that it is not
claimed that arriving at the original theory was necessarily rational—the
creative process is much more complex than that. The aim is for decisions
between competing ideas to be rational, and that we should be able to give
conditions under which we would reject a currently favored hypothesis.
In the first section we discussed testing the prediction that trees from
different sequences led to similar trees. Could we test nonevolutionary
models? One well-known astronomer, Sir Fred Hoyle, had ideas of the
earth being bombarded with influenza viruses from passing comets (Hoyle,
1984). We called this the unhealthy falling object (UFO) model. From the
UFO theory it is not expected that viral sequences would arrive in a par-
ticular order consistent with an evolutionary tree. The details of the theories
examined are described in Henderson et al. (1989).
One step of the analysis required the comparison of how well a column
of data fitted a binary tree, compared to the null model of the star (or big
bang) tree (Fig. 9-4). On this null method we could imagine a single ances-
tor from which all existing species had been independently derived. That
is, no pair of species is more closely related than any other pair. Even with
TESTING THE THEORY OF DESCENT 163

Figure 9-4 (A) The star (or big bang) tree which is a common null hypothesis
(Thompson, 1975). (B) is used to illustrate that even if sequences were generated
by a star tree process, it is still possible to fit the data with fewer changes to a binary
tree.

data generated independently from a common origin (the star tree model)
it would still fit more closely to a binary tree. For example, Figure 9-4 has
a 3,5 character [three of one color (code or character state), five of the
second] and this would require at least three changes on a star tree (Fig.
9-4A). But even if the star tree was correct, many binary trees could be
drawn that required only one change on the binary tree. The number can
be calculated from Figure 9-4B by forming a tree with eight taxa from the
two subtrees with a new edge. It thus became important to know the
expected number of changes required to fit a column of data to a random
binary tree. If a binary tree model is a good representation of the data,
then the number of changes required to best fit the data to a tree should
be significantly less than the number required for data generated from the
star tree model.
For 2-state colors (codes or character states) with frequencies a and b,
on a single column of n taxa (n = a + b), we found fm(a,b) binary trees
whose minimal coloring required m changes (Carter et al., 1990).
fm(a,b) = (m-l)!(2«-3m)(2«-5)!W(a,m)N(&,m)/(2n-2m-l)!! (1)
where, a is the number of pendant vertices with the first color (y in Fig.
9-4),
b is the number of pendant vertices with the second color (x in Fig.
9-4),
a > b, a + b = n, the number of taxa,
m (si) is the minimal number of changes on the tree,
!! is the double factorial [(2n-5)!! = 1 x 3 x 5 . . . x(2«-5);
0!! = -1!! = 1], and
\rf . _ \(2n-m-l}\l(n-m)\(m-l)\2n-m if n > m, and
yviAz./wi
^ ' — \ r\0 if• e n <^ m.
Table 9-1 Values of Fm(a,b) for Four to Eight Taxa*

Single Column Distribution Weighted Average Distribution

Number of Taxa # Trees for m = 1,2,3,4 t #Trees for m = 1,2,3,4
a b 1 2 3 4 1 2 3 4
4 2 2 1 2 z 1 2
5 3 2 3 12 z 3 12
6 4 2 15 90 z 12 78
y 3 12
3 3 9 54 42 z 9 45 36
y 0 9 6
7 5 2 105 840 z 60 570
y 45 270
4 3 45 360 540 z 36 234 360
y 9 126 180
8 6 2 945 9,450 Z 360 4,680
y 270 2,250
X 270 2,250
w 45 270
5 3 315 3,150 6,930 z 180 1,440 3,420
y 45 855 1,620
X 90 720 1,710
w 0 135 180
4 4 225 2,250 5,544 2,376 2 144 1,152 2,592 1,152
y 72 432 1,440 576
X 0 648 1,296 576
w 9 18 216 72
*Fm(a,b) is the number of trees with m changes with a and b occurrences of two character-states. The values for the single column distribution are
calculated from equation (1). For the weighted average distribution, t indicates the topology (unrooted binary tree) using the convention in Henderson
et al. (1989); z is the caterpillar (unbranched) tree (Fig. 9-6); y the topology with a single branch at the end (Fig 9-6), x has a single branch m the
middle, and topology w has two branches. The numbers for the weighted average distribution are normalized to the number of times each topology
occurs (see Fig 9-6). For any pair of values a and b (e.g., 5 and 3), the columns of the weighted average method will sum to the values for the single
column distribution
TESTING THE THEORY OF DESCENT 165

In the example in Figure 9-4 [n = 8, a = 5,b = 3 (eight taxa, one color

occurring five times and the other three times)] the probability of observing
m = 1,2, and 3 changes on the tree is 0.030, 0.303, and 0.667, respectively
(Table 9-1). Le Quesne (1989) has derived similar results by direct count-
ing.
Little progress has been made for r > 2 colors (codes). If the third and
subsequent codes occur only once, then the problem is trivial, for example,
fm(a,b,\) = (2n-3)fm_l(a,b). The single third color can be added to any
of the 2n - 3 edges on the trees containing the first two codes. With the
third or subsequent colors occurring more than once, the only cases derived
are when m = r- 1 (i.e., with only the least possible number of changes
on the tree) and where m = r.
/ r _ 1 rfli,«2 . . - « , ) = (2n-5)!!JV(«1,l) . . . . N(ar,l)l(2n-2r-l)\\ (2)
(Carter et al., 1990), and
fr(ai,a2 . . . ar) = (2n-5)!W(fll,l) . . . . N(a,,l)
(r-l)(4(rt-r) 2 -2rt + r)/(2rc-2r+l)!! (3)
(Steel, 1992).
Distributions for all cases up to 16 taxa and four colors have been es-
timated by simulation.
Equations (1), (2), and (3) are restricted to a single column (character)
which limits their usefulness. The reason for this limitation is that the
probabilities for each column are not independent. For example, consider
a simple two-color (code) case with six taxa where each color occurs three
times (aaabbb). Figure 9-5 shows the two unrooted topologies for six taxa.
(We define topologies as unlabeled trees, rooted or unrooted.) It is not
possible to fit such a column to the second topology (Fig. 9-5B) with only
a single a <-» b change on the tree. If such a column (three a's and three
£>'s) does fit a tree with only one a «-» b change, then the tree must be the
topology shown in Figure 9-5 A. This knowledge will affect the probabilities
for subsequent columns of data. Consequently, the above formulae only

Figure 9-5 Nonindependence of characters. The figure gives the two unrooted
topologies (unlabeled trees) for six taxa. A 2-state character with three a's and three
b's cannot be fitted to the second topology (B) with only one a**b change on the
tree. If such a column (three a's and three b's) does fit a tree with only one a<-»fa
change, then the tree must be topology (A) whereas four a's and two fa's can be
fitted to both topologies with a single a<-»b change. This indicates how the distribution
of changes on different columns of data cannot be independent. The formulae
described in the text can only be applied to single columns of data.
166 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

apply to single columns of data, although they give good approximations

when the data set is large (Steel et al., 1991). We refer to this as the single
column distribution.
An alternative approach has been developed by Steel et al. (1991), which
takes into account the lack of independence between columns. This eval-
uates the probabilities for each topology separately. A weighted average
is made using the number of trees that can be derived from each topology.
The formula for the number of trees from a topology is,
n\l2'ff (3)
(Hendyetal., 1984),
where x is the number of twofold centers of symmetry in the topology and
y (^1) is the number of threefold centers of symmetry (which occurs only
with n = 3,6,9, . . .). Figure 9-6 gives the six topologies for n = 9 taxa,
identifies the centers of twofold and threefold symmetry, and applies the
above formula to each topology. As a check, it is shown that the total
number of trees over the six topologies sums to (2«-5)H or 135,135. This
"weighted average" approach is more difficult to calculate as it requires a
separate calculation for each topology, but the results can be combined
for many columns of data. It was the method referred to earlier (Henderson
et al., 1989), and used for testing nonevolutionary models.
Results of both approaches for up to eight taxa are given in Table 9-1.
Weighted aveiage values for n = 9 are given in Henderson et al. (1989).
Archie (1989) has recently used simulation to estimate these values.
These methods that calculate the probabilities of finding columns of given
lengths on the tree allow some models of evolution that do not assume an
evolutionary tree to be tested. In the influenza virus case referred to above,
several nonevolutionary models could be eliminated. In the present con-
text, the important point is that a tree is a falsifiable model.
One additional point needs to be considered when calculating the ex-
pected number of changes on a tree. Formula (3) assumes that all trees
have the same chance of occurring, the "all trees equiprobable" model.
(In this context, trees are binary unrooted trees with end-points labeled.)
An alternative "Markov model" assumes the trees are derived from a
process that includes random speciation and extinction (Simberloff, 1987).
Under these circumstances, it is not valid to use formula (3). This second
model has been used in biogeographic studies where there are a small
number of areas being analyzed. However, the assumptions of the Markov
model are not met in many (most?) phylogenetic studies where only a small
proportion of possible taxa are used. The subset of taxa in such studies
are not chosen at random.
In the set of 11 mammals referred to earlier, the taxa used were not
randomly selected from all mammalian species. For example, there is only
one rodent and no bat sequences. In a random sample of mammalian
species, most would be drawn from these two orders. Rather, the taxa
have been "selected" to cover a wider range of taxa and so the assumptions
TESTING THE THEORY OF DESCENT 167

Figure 9-6 Application of the formula for counting trees from topologies. The six
unrooted topologies for n = 9 taxa are given. The x centers of twofold symmetry
are marked with a small arrow, and the one center of threefold symmetry (y = 1)
is marked by a large arrow. The numbers of phylogenetic trees that can be derived
from each topology (n\l2x&) are shown, together with the observation that these
sum to the expected number of trees, (2n — 5)!!.

of the Markov model are not met. Under these conditions the neutral "all
trees equiprobable" assumption is more appropriate.
As an additional precaution, it should be noted that the probabilities
found for the symmetric difference metric vary with the topology. The
published values (Hendy et al., 1984, 1988) are a weighted average over
all topologies. This is the reason the distribution was initially calculated
only for up to 16 taxa where there are about 500 topologies, although more
recently some properties have been determined for larger numbers of taxa
(Steel, 1988). It is not valid to find one tree and then use the distribution
168 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

to estimate the probability that a second tree has x differences. The dis-
tribution refers to a pair of trees, selected at random.

VALIDITY OF METHODS FOR INFERRING TREES

Hadamard Transformations
One major conclusion from Popper's approach to science is that more
progress is made by trying to find the limits of a hypothesis or conjecture,
rather than "testing" the hypothesis in areas where it is expected to apply.
In the present context, the hypothesis or conjecture is that a particular
tree-building method is expected, given sufficient data, to reconstruct the
correct tree. In retrospect, the previous sections are examples of the lim-
itations of making and testing simple predictions. In each case we would
have been very surprised if, for example, the trees had not become more
similar as longer sequences were used. In our own defense we would say
that it is useful to be able to estimate the number of trees that should still
be considered possible. Nevertheless, a far more powerful test would have
been to try to find conditions under which a method of tree reconstruction
would fail.
The work of Felsenstein (1978) and Cavender (1978) has introduced an
improved approach which does this. It allows a search for models of ev-
olution where tree building methods would not be expected to find the
correct tree. That is, it allows a search for conditions where a tree recon-
struction method will fail. A method is said to be inconsistent if, under
some conditions of the model, it can converge to an incorrect solution as
longer sequences are taken. The paradox is that a method may by chance
find the correct tree with short sequences, but as longer sequences are
used, the probability that the method finds the correct tree goes to 0.
A model of evolution has three parts,
1. a tree (or more generally a graph),
2. an assumed "mechanism" of change to the sequences, and
3. "edge lengths" (probabilities of change along the edges of the tree).
A frequently assumed mechanism (Farris, 1973; Cavender, 1978) is that
changes occurring in the sequence are "independent and identically dis-
tributed" (i.i.d.). Changes at any position along the sequence, and any-
where on the tree, are independent. All positions (nucleotide or amino
acid) have the same chance of changing state (identically distributed). In
addition, some mechanisms assume the same rate of change along each
edge of the tree—the molecular clock. We will use the term "standard
model" for a mechanism of independent and identically distributed changes
on a tree but which does not assume the molecular clock. Felsenstein (1978)
showed that under some conditions, a model with four taxa and uneven
TESTING THE THEORY OF DESCENT 169

rates of evolution could give data for which parsimony would be expected
to reconstruct the wrong tree. Thus, parsimony is, in general, an incon-
sistent method even though there are many specific models where it will
work correctly. An important question is whether the conditions that lead
to inconsistent performance are common (and therefore, parsimony should
seldom be used) or unusual (in which case the inconsistency may not be
a problem in practice).
To extend Felsenstein's analysis to n > 4 taxa we developed the Had-
amard transformation (Hendy and Penny, 1989; Hendy, 1991) for 2-state
characters, a and b (Table 9-2). We will describe this in three parts, each
of which is now straightforward (this was probably not true of the original

Table 9-2 Bipartitions and Vectors for the Hadamard Transformations, Illustrated
for the Five Taxa in the Model in Figure 9-8B*
Probabilities
Index Subsets P q S r s
1 {1} 0.100 0.1116 0.1116 0.0000* 0.0583 +
2 {1,2} 0.005 0.0050 0.0050 0.2232 0.0113
3 {1,3} 0.0000 0.2332 0.0081
4 {1,2,3} 0.0000 0.2332 0.0155
5 {1,4} 0.0000 0.2332 0.0081
6 {1,2,4} 0.0000 0.2332 0.0155
7 {1,3,4} 0.0000 0.2332 0.0155
8 {1,2,3,4} 0.200 0.2554 0.2554 0.4463* 0.1301 +
9 {1,5} 0.0000 0.3720 0.0155
10 {1,2,5} 0.005 0.0050 0.0050 0.3720 0.0113
11 {1,3,5} 0.0000 0.3720 0.0081
12 {1,2,3,5} 0.100 0.1116 0.1116 0.5952* 0.0583 +
13 {1,4,5} 0.0000 0.3720 0.0081
14 {1,2,4,5} 0.100 0.1116 0.1116 0.5952* 0.0583 +
15 {1,3,4,5} 0.100 0.1116 0.1116 0.5952* 0.0583 +
16 {1,2,3,4,5} -0.7118 0.5952* 0.5197 +
0.0000 1.0000
*The 16 (2" ') subsets containing 1, of {1,2,3,4,5} for n = 5 taxa. In each case the subset with its complement
(the remaining taxa) forms a bipartition. Each edge of a tree splits the set of taxa into complementary
subsets. Similarly, any 2-state character also splits the set of taxa. We refer to these as "edge bipartitions"
and "character bipartitions." The edge bipartitions of the tree of Figure 9-8B have indices 1, 2, 8, 10, 12,
14, and 15. For each edge e, we list the probability p, that a character will have different states at its
endpoints. The <j, value is the number of changes expected under a Poisson model on that edge. 8 is
obtained from q by inserting 0 where there is no corresponding q, value and one negative value for
<52n-i so that Z<5, = 0. r is an intermediate vector which contains the actual distance, per nucleotide, of
the minimal path between even-sized subsets of taxa. The Hadamard transformation (Hendy and Penny,
1989) allows us to compute s from 8 where s, is the probability of obtaining the character bipartition with
index j. Thus, for c characters, cs, will be the expected frequency of this character bipartition. Note in
this case SK = 0 5197, so we would expect approximately 52% of characters to be constant, while bipartition
3 should occur for less than 1% (0.81%) of characters. The inverse Hadamard transformation produces
a vector S, with q, = 8, > 0 identifying the edges e, of the tree
Parsimony uses only "informative" characters, those which group taxa together (that is, ignoring sin-
gleton and constant characters). Their corresponding s, values are marked " +." In choosing the parsimony
tree(s) for these data we find four minimal length trees (expected length 0.0902c), none of which is the
tree used to generate the data. The expected length of the correct tree is 0.0934c, the third longest of the
15 possible binary trees. This is an example of inconsistency in parsimony, even with equal rates of
evolution.
170 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

description). The three parts are the model, the expected form of the data,
and a method for interconverting between these two (that is, calculating
the data from parameters of the model, and vice versa.
We note that each edge e, of a tree partitions the taxa into two subsets,
the set of taxa to the left of the edge and the set of taxa to the right. This
pair of sets is called an edge bipartition (split). There are 2""1 possible
bipartitions. Also, each column of the sequences induces a bipartition of
the taxa, the two subsets being those taxa with state a and those with b.
These we call sequence bipartitions.
We express the bipartitions by the taxa (represented by numbers) they
contain. For example, if the character-states for four taxa are a, b, b and
b (summarized as abbb), this is the bipartition {1} and {2, 3, 4}. (The
character-states baaa also give this bipartition.) There is no need to list
both subsets of the bipartition. We normally list only the subset containing
taxon 1.
There are four (2"~2) ways in which the sequence bipartition {1} {2, 3,
4} can be generated on a tree and these are shown in Figure 9-7. They
correspond to the four ways of labeling two (n - 2) internal points with the
two codes. The probabilities for each of the eight (2""1) bipartitions can
be calculated in this way. In all there are 22"~3 sets of calculations (2"~2
ways for 2""1 partitions). However, the whole procedure can be carried
out using Hadamard transformations. Assume a tree T where for each edge
e, there is a probability p, that the character-states at each end of the edge
are different. With these values given we use the Hadamard transformation
to calculate the probability of obtaining any particular sequence bipartition.

Figure 9-7 Ways of forming a single bipartition for four taxa. The figure shows four
combinations of changes along edges of a tree that all lead to the bipartition {1}
{2,3,4}.
TESTING THE THEORY OF DESCENT 171

The transformation is

where H is a Hadamard matrix of 2"~l rows and columns, and s and 8 are
defined in Table 9-2. Equation (4) is easily inverted as H"1 = HV2""1
giving

The calculations are illustrated for a case with n = 5 in Table 9-2 with
an example derived from Figure 9-8B. In practice it is not necessary to
construct the Hadamard matrix, as there is now a simple algorithm to carry
it out. From equation (4) we can find values of edge lengths on a tree
where parsimony will be inconsistent; that is, it will be guaranteed to select
the wrong tree as sequences become longer.
8 and s are equivalent for a specified mechanism of change in that each
can be obtained from the other, and are not dependent on a particular
tree. Calculating the expected data in this way quickly allows the tree to
be identified to which a tree-building method will converge. Then the rate

Figure 9-8 Trees used in calculations. (A)

A tree of four taxa (t,_ 4 ) with p, being the
probability, for any column, of observing
a change along an edge, p, can take val-
ues between 0 (no change) and 0.5 (com-
plete randomization), co, = (1 — 2p,).
Conditions under which several tree-
building methods will converge to the
correct tree are given in the text. (B) The
unrooted tree used to illustrate the Hada-
mard transformations (see Table 9-2). The
edge length of the central edges are 0.01.
The tree is consistent with the molecular
clock if rooted on the central pendant
edge.
172 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

of convergence to this tree can be studied separately by randomly selecting

larger data sets from s.
The computational advantage of the Hadamard transformations over
existing maximum likelihood methods is that they allow the data (sequences
or bipartitions) to be used directly for estimation of parameters of the
model. This is indicated in Figure 9-9a to show that it is possible to go
from the model to give predicted sequence bipartitions, or from sequence
bipartitions to estimate lengths of edges on trees.
By contrast, existing maximum likelihood methods start with a single
tree and some initial guesses for edge lengths. It then repeatedly (maybe
thousands of times) makes slight adjustments to edge lengths and repeats
the calculation to see if the observed data are now more likely. The process
is indicated in a general way in Figure 9-9b to indicate the repeated cal-
culations. Figure 9-9b has been modified from the usual calculation for
maximum likelihood to allow an easier comparison to the Hadamard trans-
formations. One interesting development would be to combine the Had-
amard transformations with maximum likelihood as the optimality crite-
rion. Such a process would set a good estimate of the optimal branch lengths
with a single computation.
The optimality criterion we have favored is finding the closest tree (Hendy,
1991). Its advantages come from its being characterized mathematically.

tree
mechanism data

rates
A Hadamard transformations

tree
mechanism data

rates

B maximum likelihood
Figure 9-9 The advantage in being able to invert the calculation. With the Had-
amard transformations each calculation need only be calculated once, not repeatedly
as for maximum likelihood.
TESTING THE THEORY OF DESCENT 173

This allows the global optimal tree to be found exactly (by branch and
bound methods). Thus far, the closest tree method is the only one based
on the idea of general invariants for n > 4. We have used the closest tree
criterion to find the optimal tree for 20 taxa (Penny et al., 1990). For 20
taxa there are > 1020 trees! By contrast, current maximum likelihood
methods would not search more than 100 (102) trees—an advantage of
1018 times for methods based on the Hadamard transformations.
Conditions for Consistency—Four Taxa
Some progress has been made with four-taxa, 2-state codes (colors) on the
general conditions under which a method will, or will not, converge to an
incorrect tree on the standard model. Using the terminology of Figure 9-
8A and defining &>, = ( ! - 2p,) we find some common tree-building
methods will converge to the correct tree (be consistent) if and only if, the
following conditions are met (Steel, 1989).

Cluster Analysis. For a simple clustering procedure which joins the pair
of taxa having the smallest observed distance, the condition for consistency
is:

It can be shown that this is a special case of the next condition shown
below. Thus, for example, if parsimony fails to converge to the correct
tree with four taxa, so too will a simple clustering procedure. The converse
does not hold. The special case, with/?! = p3 andp2 = p\= PS was derived
by Felsenstein (1978).

Parsimony or Compatibility Methods and Methods Using the Four-Point

Distance Criterion. This last method selects min {dl_2 + d3>4, dl3 +
^2,4> ^1,4 + ^2,3}- These are consistent, if and only if,

With equal rates of evolution (the molecular clock) this condition is always
met and parsimony will be consistent with four taxa.

Maximum Likelihood, Closest Tree (from Hadamard Transformations),

and the Corrected Four-Point Distances Metric. These methods will always
be consistent with four taxa on the standard model. The corrected four-
point condition is to select,
min (dlf + dkl - 2-dtj dkllc) : {i, /, k, /} = {!, 2, 3, 4}
where c is the number of columns.
174 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

It should be noted that, at least with parsimony, the conditions for

consistency are more restrictive as the number of taxa increases. Parsimony
with:
four taxa, can only be inconsistent with unequal rates of evolution;
five taxa, with equal rates of evolution (the molecular clock) can be
inconsistent if the root is on the central pendant edge which is adjacent
to a short and long edges;
six taxa, with arbitrarily low rates of change can be inconsistent;
n large, with all branch lengths (internal and pendant) are both small
and equal can be inconsistent.
The first three cases (four to six taxa) are from Hendy and Penny (1989)
and the last is from Steel (1989). The conclusion is that the range of cases
where parismony is inconsistent increases as more taxa are added.
It is desirable that this type of study be extended to other optimality
criteria. With four taxa, methods using the four-point distance criterion
have identical performance (with respect to consistency) as parismony (see
above). Again with four taxa, simple clustering methods are worse than
parsimony in that they will fail under a wider range of conditions. We
know less about the performance of cluster analysis with additional taxa
except that it is not identical to parsimony. With five or more taxa we can
find examples where parsimony fails and simple clustering is consistent,
and vice versa.
Loss of Information
We have commented elsewhere (Penny et al., 1990) that all methods ignore
some of the information in sequences. With four taxa, distances use nearly
all the information, but the proportion of information in distances declines
very rapidly (Steel et al., 1988) as the number of taxa increases.
Table 9-2 also indicates information that is omitted by parsimony and
standard distance methods. Singletons and constant partitions (marked
" + " in Table 9-2) are omitted by parsimony methods. Entries in r which
do not correspond to a path between a pair of taxa (marked "*") are
omitted by distance methods. There is no general correspondence between
these omissions in r and s. The losses of information in parsimony and
distances are not equivalent. Parsimony methods ignore any singletons and
constant columns, corresponding to n + 1 elements in s. Distance methods
use all the s values to construct a distance matrix which corresponds to
only n(n — l)/2 of the r values. The Hadamard transformations use all of
these 2""1 values. Table 9-3 contrasts the increase in the numbers of bi-
partitions used by the Hadamard transformation with those used by par-
simony and distance methods for small values of n. The example in Table
9-2 and Figure 9-8B is a case where the loss of information is sufficient for
parsimony to converge to the incorrect tree (Hendy and Penny, 1989). The
correct tree found by selecting partitions {1, 2} and {1, 2, 5} is not the
shortest, but the third longest tree (out of 15).
TESTING THE THEORY OF DESCENT 175

Table 9-3 Comparative Use of Information by Parsimony and

Distance Methods*
Bipartitions
Used in Entries in
Taxa Total Parsimony Distance Matrix
4 8 3 6
5 16 10 10
6 32 25 15
7 64 56 21
8 128 119 28
9 256 246 36
10 512 501 45
11 1024 1012 55
12 2048 2035 66
13 4096 4082 78
14 8192 8177 91
15 16384 16368 105
16 32768 32751 120
*For n = 4-16 we list the number of bipartitions, together with the numbers
considered by parsimony and distance matrix methods. Parsimony methods omit
n + 1 bipartitions from s which rapidly become a minute proportion of the total.
The omitted bipartitions include the n singletons and the number of constant
columns. These omitted bipartitions are important under any model that includes
estimates of rates of change. Because they occur so often, the bipartitions omitted
by parsimony are those with the most accurate estimates of their frequency. Dis-
tance methods include some information from all bipartitions, but the proportion
of entries of r that are used, rapidly becomes very small (Steel et al., 1988).

Twelve Mammals, Seven Sequences

In earlier papers, we used a data set from 11 mammals and five (later
extended to six) sequences. The data are amino acid sequences converted
to "best-guess" nucleotide sequences. These data have been particularly
useful for developing and testing new methods of analysis. However, even
six sets of sequences were insufficient for the optimal tree to be stable (that
it would not change as longer sequences became available). When the
minimal and near-minimal trees were analyzed (Penny and Hendy, 1985b)
it was found that the position of the carnivore (dog) was the least stable.
The carnivore was attached by a long "unbranched edge" which can be
(Hendy and Penny, 1989) comparatively unstable on the tree. Would add-
ing a second carnivore improve the stability? Would we then get a tree
that would be expected to remain constant as longer sequences became
available? In other words, would we have convergence to a single tree?
We have now added a second carnivore to the tree and an additional
sequence, a-crystallin, for all 12 taxa. Thus the data set now has seven
sets of sequences from 12 mammals (Table 9-4). The sequences from the
earlier studies were a- and /3-hemoglobins, fibrinopeptides A and B, cy-
tochrome c, and myoglobin. In some cases a "taxon" was made up of
sequences from two species. One example is using either mouse or rat to
represent the rodent taxon. The additional carnivore taxon is made up
Table 9-4 Combined Data from Seven Sequences for 12 Mammals*
10 20 30 40 50 60 70 80 90 100

1. Monkey CU G CU GGAAA G CAA. .A. . . .C. . .U.C G C...C.A...G C

2. Sheep . .G C GGC G. . . .A. A. .G.AACAGA. .AA U G A. . .U. .A G.G.A. .G.G.C. . .UCAU.C.
3. Horse G CU.G A.GU.AAC. . . .U. .GU. . .UU. . . .A.C. . . AC C.UGAGC.C. . . G C . . . G . G . . C . . . C . . C . .
4. Kangaroo .GG.A CA. .A. . .GC. .AA. . .AAA.G. . .A. .GC AA. .AG.AAACCA.A.A. .C. .U.G.A. . .A AAAA A.GC. .
5. Rodent .U.A. .A C. . .G. .G. .C A. .C AU. . .U. .G U. . .A. .GAC C. . .OC.AAA. .C.C. .C. .U. . .
6. Rabbit C. .A.AA GAA G A..AA.U....A....G.AU A....U..A.AA..C..A..GC A C C.
7. Dog C. .AAAA. .CCA C U.GAA. . .GOT U.U. .G. . .AC C. .A.GA.U A A. . .G C AU.
8. Pig .6 C GA. .G G.A AACA.A U. . .UU.G A U.A AG.G.CCA.UC. . .GO
9. Cat .U. .U.A. . .CG. . . .C. .C. . .A.AC. . .U.GA. . . .GCUU.O G.A.G U. . .AA.GA.U A.G G. .CCAA. . .U.GU
10. Human C C G C.GAAA G CU C G GCC C G CC U.
11. Cow .GG C GGC A A AGC. .A. .AA GUA. . .CA AA. . .U. .AC. . .GGGC. . .6. GCC. . .UCAU.C.
12. Ape C C.C G C.GAAA G CU C G GCC CA. . .G CC U.
110 120 130 140 150 160 170 180 190
1234567890123456789012345678901234567890123456789012345678901234567890123456789012345678901
GGGAGGAACCGAGCCGGAGAGGAUUGAGCAGOnjAOlCGAGCGUGAGCAGGUACaU^CC^
1. Monkey . C . AA . .A CC.AGAAUU.A.C.CUA C GU U U. . . .G.A.C. .A.UA.
2. Sheep .U.G CA.CUCC A.C.G. .C. .AC GAGA.C. .CCGA.G C.U. 555555555555
3. Horse .A A ACA. .U.C A C. .A C. .AC. . .CC. . .AC. .CC. . .C. .A G C.A. .G
4. Kangaroo AA. . .AC. A UA.A. .GG A.UC A.AG.G UC C C.A UC. . .AA.A
5. Rodent . .ACA. . .AAC. . .GCA. .U.A. . .AG G C C.G.AA. . .6 A G. -A G. .U.G
6. Rabbit AAC. AAC. . .GC UU.C G C.A C . . .AC CC. .CG. .A.C. . .G.A. .6
7. Dog AA A A. .AU.UA. .CG CA.A.AC. . .AAA.GG. .C C C.A. A G.U.GC.G. . .
8. Pig CAA CCA. .UCAG A. . .G C G G. . .GC.G. . .
9. Cat GCA A U. .G.G CA. . .AC. . .A. . .GG. .C. . .A. . .CCG.C.A.A A. G.U.GC.G. ..
10. Human C. . .UA.AA U. .AGAATJU.A.CUCUA C GC UG A. U.C. .G.A.C. .A.UAA
11. Cow CCAG. . .C CCA. .UGG. .G A.C.G. .C. .AC GAGAAC. .CCGA.G. .U. . .UA.GA. . .C.G. . .
12. Ape C. . .UA.AA U. .AGAAUU.A.CUCUA C GC UG A. U.C. .666666666666
'Columns 1-42 and 43-97 are derived from a- and/3-hemoglobms, 98-110 and 111-127 from fibnnopeptides A and B; 128-142 from cytochrome c; 143-179 from myoglobins;
and 180-191 from a-crystalhn Columns 1-142 are (with the addition of cat) from Penny et al. (1982), columns 143-179 from Penny and Hendy (1986), and columns 180-191
are derived from deJong et al (1977). a-crystalhn sequences were not available for sheep or apes and so additional singleton character-states ('5 and 6') were assigned to these
taxa. This does not distinguish between any tree, and so does not affect the results obtained. The ammo acid sequence for lion cytochrome c was supplied by Professor R.P.
Ambler, Edinburgh. The data, with some additional information, are available by e-mail from D.Penny @ massey.ac.nz.
TESTING THE THEORY OF DESCENT 177

Figure 9-10 Trees for 12 mammals, based on seven sequences, (a) is the minimum
length tree (472 mutations) based on the 191 columns in Table 9-4. It is also the
consensus or median tree (Penny et al., 1982). (b) is three changes longer (475
mutations) but may be correct if the "long edges attract" problem of parsimony is
leading to incorrect convergence. One additional change results from transferring
carnivores and two changes from separating the rabbit-rodent neighboring pair.

from six feline sequences (five from cat and lion cytochrome c) plus one
from a seal. The procedure of using composite laxa is legitimate in this
study as it cannot result in trees from different sequences being more similar
than expected. It can only introduce more dissimilarity between trees if
the composite taxon was phylogenetically incorrect. For the reasons dis-
cussed below we suggest the best tree is that shown in Figure 9-1 Ob.
Another approach is possible by analyzing the frequency of the edges
in minimal trees from bootstrap samples. This is illustrated in Figure 9-11
with results from 132 minimal trees (in this case using parsimony) from
bootstrap samples. Only the relationships between three groups; primates,
lagomorphs (rabbit), and rodents are shown. The most common subtree
is (rabbit, rodent), primate (Fig. 9-llb) which occurs 74 times. The next
most frequent is (primate, rabbit), rodent (Fig. 9-lla), 44 times; and (pri-
mate, rodent), rabbit subtree (Fig. 9-llc), which occurs seven times. Figure
9-llb conforms to the Glires model (Novacek, 1985) which links rodents
and lagomorphs (rabbits).
178 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Figure 9-11 Frequency of partitions in bootstrap and jackknife samples, (a), (b),
and (c) are the three subtrees for primates, rabbit (rab), and rodents (rat), and the
frequency of occurrence of these subtrees in the minimal length trees from 132
bootstrap samples, (d) is the unresolved trichotomy. An anomaly is the low frequency
of just one of the three subtrees, (c) compared to (a). This can be explained as an
example of the error with parsimony involving a path of long-short-long edges, as
in the heavy lines in (a). With parsimony these long edges would tend to "attract"
(Hendy and Penny, 1989) leading to tree (b).

The question is to interpret these frequencies (74, 44, 7). If there was
a short edge linking these three groups (that is, close to a trichotomy, Fig.
9-lld) then all three subtrees (Figs. 9-lla-c) would be expected to occur
with about equal frequency. If Figure 9-1 Ib was correct, then we would
expect Figures 9-lla and 9-llc with equal frequency. The observed fre-
quencies are significantly different from those expected for either the tri-
chotomy (Fig. 9-lld) or Glires (Fig. 9-llb) models.
Qualitatively we can account for the observed distribution of Figure
9-lla-c if Figure 9-lla ((primate, rabbit), rodent) is the correct relation-
ship. This tree has a long-short-long series of edges (rabbit, internal edge,
rodent) which is shown as dark lines in Figure 9-lla. Parsimony would
lead to the long edges being drawn together to give Figure 9-llb. This
would account for two of the three subtrees being found more frequently
than the third. A prediction from this hypothesis is that bootstrap samples
from data sets which include sequences from a quite different lagomorph
(or rodent) should resolve the issue. Note that convergence from the other
two trees (Fig. 9-llb and 9-llc) could not lead to the observed frequencies
in that neither should lead to the observed frequencies. If Figure 9-llb
TESTING THE THEORY OF DESCENT 179

were correct, then any tendency for incorrect convergence should have
equal affects on both rodents and lagomorphs, leading equally to Figures
9-lla and 9-llc. We suggest the tree in Figure 9-10b shows the correct
relationship of the three groups (rodents, lagomorphs, and primates).
A similar problem occurs with the position of the two carnivores, relative
to the ungulates and the metatherians (kangaroo). The bipartition with
the two carnivores is almost always found, but their position on the tree
varies. It is on the first division of the eutherians (83 times), after the
separation of ungulates (40 times), or on the first division of the ungulate
lineage (eight times). Again we do not have the pattern expected from a
near-trichotomy (all three equal, or one more common and the next two
equal). The difference is statistically significant in a x2 test.
Is this another case of incorrect convergence? Consider what would
happen if the correct tree had the carnivores separating after the ungulate
line has diverged (that is, in the position shown in Fig. 9-10b). We again
have a case of a long lineage (kangaroo), a short internal edge, and a
medium lineage leading to the two carnivores. This may be sufficient to
indicate incorrect convergence. Again, if either of the other two possibil-
ities for kangaroo, ungulates, and carnivores was the correct tree, they
would not have the long-short-long sequence of edges that can give incor-
rect convergence.
These conclusions need to be made quantitative. A possible correction
for parsimony is available (Penny and Hendy, in preparation). However
it requires a reasonable estimate of rates of change along the edges of a
tree. This depends on the numbers of singletons and constant columns.
We note that the "long edges attract" problem that leads parsimony to
converge to a wrong tree may be more frequent than anticipated. The
initial work by Felsenstein (1978) with four taxa suggested the problem
may occur with very unequal rates. With five taxa it can occur with equal
rates (Hendy and Penny, 1989) but required a short edge(s) between long
edges. Even this knowledge is inadequate to indicate whether or not in-
correct convergence occurs in a specific case.
Even with these data we were unable to reach a firm conclusion. The
time of separation of these 12 mammals is probably 60 to 80 million years
ago. If seven sequences are insufficient to resolve such a recent divergence,
would we expect a single sequence (even quite a long sequence) to be
sufficient for a divergence that occurred much earlier—say 1 billion, or
even 3 billion years ago? Such a rhetorical question is expected to be
answered in the negative. It is of course possible that we are very fortunate
and a single sequence [e.g., small-subunit ribosemal RNA (rRNA)] is
sufficient. But we have no rational reason to assume it without testing.
Work on the small-subunit rRNA has vastly improved our knowledge of
the phylogeny of organisms but there is no reason to assume it is sufficient.
What concerns us is the scientific myth (and at present, it can only be
considered a myth) that a single sequence is sufficient to reconstruct the
whole history of life. We call this the "Myth of a Universal Tree from One
180 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Gene" (MUTOG). The myth is strong enough to get evolutionary trees

into textbooks which lack qualifying statements. The main problem is that
the myth inhibits testing of ideas by not taking additional sequences and
testing for convergence. If the tree is "believed" to be correct, there is
little motivation to undertake new work to test the tree. Such a state of
affairs is disturbing to anyone using a Popperian framework for science.
An idea (trees in this case) should be a stimulus to new measurements and
better, more rigorous tests.

DISCUSSION

By discussing our work from a Popperian viewpoint, we do not wish to

imply a concept of a single monolithic framework in which scientists work.
Evolutionists in particular should be well aware of the problem of essen-
tialism. They are aware of the importance of diversity in evolution (Mayr,
1982), and consequently should be skeptical of any attempt to define "one
true scientific method." Medawar, who supported a Popperian view of
science (Medawar, 1974), has commented that
Among scientists are collectors, classifiers and compulsive tidiers-up. Many
are detectives by temperament and many are explorers; some are artists and
others artisans. There are poet-scientists and philosopher-scientists and even
a few mystics (Medawar, 1967:132).
Attempts by philosophers to describe a single mechanism of research that
all successful scientists follow, are doomed to failure. A diversity of ap-
proaches to science is still accommodated on the Popperian model in which
no scientific hypothesis is ever absolutely proven, where no hypothesis
should be "believed." A similar comment on the diversity of approaches,
using the analogy of diversity within a species appears in Hull (1988). The
important point is the attitude toward hypotheses: they should never be
accepted as beyond questioning. They are tools to aid in the design of
harder and more rigorous tests.
We have recently discussed tree-building methods as requiring at least
five criteria. Methods should be consistent, robust, efficient, fast, and
falsifiable (CREFF) (Penny et al., 1990). These criteria appear incompat-
ible in that efficient methods (which increase in a polynomial manner with
the number of taxa) for real data are not known. Little is known on the
robustness of methods relative to deviations from the assumptions about
the mechanism of change. Improvements are still needed to increase the
power of tree-building methods. Perhaps the most urgent is an analog of
the Hadamard transformations for four-state characters. So far, an initial
approach has been developed (Penny et al., 1990) which still needs further
development. Another area that is important to test is the effect of devia-
tions from the "standard model." This would allow a test of the robustness
of tree building.
Whatever the approach that is taken, it appears to us that it is important
to find the limits of any tree-building method and to find under what
TESTING THE THEORY OF DESCENT 181

conditions it will break down. Such a Popperian approach should help

identify problem areas and assist the search for better methods. Improved
methods will almost certainly depend on a better understanding of their
mathematical basis. This makes the study of evolutionary trees an exciting
part of modern biology.

REFERENCES

Allan, M. (1977) Darwin and His Flowers: The Key to Natural Selection. Faber
and Faber, London.
Archie, J.W. (1989) A randomization test for phylogenetic information in system-
atic data. Syst. Zool. 38:239-252.
Carter, M., M.D. Hendy, D. Penny, L.A. Szekely, and N.C. Wormald. (1990)
On the distribution of lengths of evolutionary trees. SIAM J. Disc. Math.
3:38-47.
Cavender, J. (1978) Taxonomy with confidence. Math. Biosci. 40:271-280.
Darwin, C. (1859) On the Origin of Species by Means of Natural Selection, or the
Preservation of Favoured Races in the Struggle for Life. John Murray,
London.
Day, W.H.E. (1985) Optimal algorithms for comparing trees with labelled leaves.
J. Classif. 2:7-28.
Day, W.H.E., D.S. Johnson, and D. Sankoff. (1986) The computational complexity
of inferring rooted phylogenies by parsimony. Math. Biosci. 81:33-42.
de Jong, W.W., J.T. Gleaves, and D. Boulter. (1977) Evolutionary changes of a-
crystallin and the phylogeny of mammalian orders. J. Mol. Evol. 10:123-
135.
Farris, J.S. (1973). A probability model for inferring evolutionary trees. Syst. Zool.
22:250-256.
Felsenstein, J. (1978) Cases in which parsimony or compatibility methods will be
positively misleading. Syst. Zool. 27:401-410.
Felsenstein, J. (1985) Confidence limits on phylogenies: an approach using the
bootstrap. Evolution 39:783-791.
Ghiselin, M.T. (1969) The Triumph of the Darwinian Method. University of
Chicago Press, Chicago.
Graham, R.L., and L.R. Foulds. (1982) Unlikelihood that minimal phylogenies
for a realistic biological study can be constructed in reasonable computa-
tional time. Math. Biosci. 60:133-142.
Henderson, I.M., M.D. Hendy, and D. Penny. (1989) Influenza viruses, comets,
and the science of evolutionary trees. /. Theor. Biol. 140:289-303.
Hendy, M.D. (1991). A combinatorial description of the closest tree algorithm for
finding evolutionary trees. Disc. Math, (in press).
Hendy, M.D., C.H.C. Little, and D. Penny. (1984) Comparing trees with pendant
vertices labelled. SIAM J. Appl. Math. 44:1054-1067.
Hendy, M.D., and D. Penny. (1982) Branch and bound algorithms to determine
minimal evolutionary trees. Math. Biosci. 59:277-290.
Hendy, M.D., and D. Penny. (1989) A framework for the quantitative study of
evolutionary trees. Syst. Zool. 38:297-309.
Hendy, M.D., M.A. SteeJ, D. Penny, and I.M. Henderson. (1988) Families of
trees and consensus. Pp. 355-362 in Classification and Related Methods of
182 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Data Analysis (H.H. Bock, ed.). Elsevier Science Publishers. B.V., Am-
sterdam.
Hoyle, F. (1984) Living Comets. Cardiff University Press, Cardiff.
Hull, D. (1988) Science as a Process: An Evolutionary Account of the Social and
Conceptual Development of Science. University of Chicago Press, Chicago.
Le Quesne, W.J. (1989) Frequency distribution of lengths of possible networks
from a data matrix. Cladistics 5:395-407.
Mayr, E. (1982) The Growth of Biological Thought: Diversity, Evolution and
Inheritance. Harvard University Press, Cambridge.
Medawar, P. A. (1967). The Art of the Soluble. Methuen, London.
Medawar, P. A. (1974) Hypothesis and imagination. Pp. 241-273 in The Philosophy
of Karl Popper, Vol. 1 (P. A. Schlipp, ed.). Open Court, LaSalle.
Mickevich, M.F. (1978) Taxonomic congruence. Syst. Zool. 27:143-158.
Novacek, M. (1985) Cranial evidence for rodent affinities. Pp. 59-81 in Evolu-
tionary Relationships Among Rodents: A Multidisciplinary Analysis (W.P.
Luckett and J.-L. Hartenberger, eds.). Plenum Press, New York.
Penny, D. (1985) The evolution of meiosis and sexual reproduction. Biol. J.
Linn. Soc. 25:209-220.
Penny, D., L.R. Foulds, and M.D. Hendy. (1982) Testing the theory of evolution
by comparing phylogenetic trees constructed from five different protein
sequences. Nature 297:197-200.
Penny, D., and M.D. Hendy. (1985a) Testing methods of evolutionary tree con-
struction. Cladistics 1:266-278.
Penny, D., and M.D. Hendy. (1985b) The use of tree comparison metrics. Syst.
Zool. 34:75-82.
Penny, D., and M.D. Hendy. (1986) Estimating the reliability of evolutionary
trees. Mol. Biol. Evol. 3:403-417.
Penny, D., M.D. Hendy, E.A. Zimmer, and R.K. Hamby. (1990) Trees from
sequences: panacea or Pandora's box? Aust. Syst. Bot. 3:21-38.
Popper, K.R. (1963) Conjectures and Refutations: The Growth of Scientific Knowl-
edge. Routledge and Kegan Paul, London.
Popper, K.R. (1972) Objective Knowledge: An Evolutionary Approach. Oxford
University Press, Oxford.
Popper, K.R. (1976) Unended Quest: An Intellectual Autobiography. Fontana,
London.
Popper, K.R. (1978) Natural selection and the emergence of mind. Dialectica
32:339-355.
Popper, K.R. (1984) Erkenntnis und gestaltung der wirklichkeit: die suche nach
einerbesserenwelt. Pp. 11-40 in VortrageundAufsatzeAusDreissigJahren.
R. Piper, Munich. English translation: (1991) Knowledge and the shaping
of reality: the search for a better world. New Zealand Sci. Rev. 48 (in press).
Riddiford, A., and D. Penny. (1984) The scientific status of evolutionary theory.
Pp. 1-38 in Evolutionary Theory: Paths Into the Future (J.W. Pollard, ed.).
John Wiley and Sons, London.
Robinson, D.F., and L.R. Foulds. (1981) Comparison of phylogenetic trees. Math.
Biosci. 53:131-147.
Ruse, M. (1975) Darwin's debt to philosophy: an examination of the influence of
the philosophical ideas of John F. W. Herschel and William Whewell on the
development of Charles Darwin's theory of evolution. Stud. Hist. Philos.
Sci. 6:159-181.
TESTING THE THEORY OF DESCENT 183

Schweber, S.S. (1977) The origin of the Origin revisited. J. Hist. Biol. 10:229-316.
Simberloff, D.S. (1987) Calculating probabilities that cladograms match: a method
of biogeographical inference. Syst. Zool. 36:175-195.
Steel, M. A. (1988) Distribution of the symmetric difference metric on phylogenetic
trees. SIAM J. Disc. Math. 1:541-551.
Steel, M.A. (1989) Distributions on blcoloured evolutionary trees. Ph.D. Disser-
tation, Massey University, Palmerston North.
Steel, M.A. (1992) Distributions on bicoloured binary trees arising from the prin-
cipal of parsimony. Discr. Appl. Math, (in press).
Steel, M.A., M.D. Hendy, and D. Penny. (1988) Loss of information in genetic
distances. Nature 336:118.
Steel, M.A., M.D. Hendy, and D. Penny. (1991) Significance of the length of the
shortest tree. J. Classif. (in press)
Thompson, E.A. (1975) Human Evolutionary Trees. Cambridge University Press,
Cambridge.
10
Parsimony and Phylogenetic Inference
Using DNA Sequences: Some
Methodological Strategies

JOEL CRACRAFT AND KATHLEEN HELM-BYCHOWSKI

Sequence data of DNA are becoming increasingly important within sys-

tematic and evolutionary biology. With this expanding emphasis, many
investigators have rising expectations that sequence variation offers an
almost unlimited potential to resolve phylogenetic relationships and thereby
provide us with a highly robust, if not definitive, description of the pattern
of life's history (e.g., Goodman et al., 1987; Sibley and Ahlquist, 1987;
Goodman, 1989). Various reasons have been given in support of this ex-
pectation. One of the most common is that because DNA is the physical
information system recording all inherited attributes of organismal history,
direct comparison of its structure provides the most basic of all data for
phylogenetic reconstruction. Inasmuch as all historical change is encoded
in DNA, it is reasoned, it makes sense to examine that change directly.
A second argument is one of large numbers: we can obtain far more
character-state data from DNA sequences than we can, say, from tradi-
tional morphological comparisons, and therefore, statistically, as additional
data are examined, true patterns of phylogenetic relationships will even-
tually emerge. Furthermore, it is often claimed that because morphology
is so readily subject to convergence when compared to sequence data,
obtaining more and more of the latter represents our best hope for resolving
the phylogenetic pattern.
Technological innovations are also fueling the rush to obtain sequence
data. Comparative sequence data are much easier and less expensive to
collect than they were even a year ago. Just as importantly, there has been
a rapid growth in the amount of computer hardware and software that
makes the analysis of comparative sequence data readily accessible to work-
ers at all levels of expertise (Platnick, 1989; see also many of the papers

184
PARSIMONY AND PHYLOGENETIC INFERENCE 185

in Doolittle, 1990). In particular, the availability of tree-building algorithms

for microcomputers has been a contributing factor to the marked increase
in the numbers of papers that use DNA sequences to answer phylogenetic
questions (see summary in Swofford and Olsen, 1990). At its simplest,
therefore, it is relatively straightforward to collect comparative sequence
data, analyze them with a given tree-building method, and publish the
results.
Yet, it is clear from the literature that phylogenetic inference using
sequence data is anything but simple. There is, for example, substantial
disagreement over the best method to resolve relationships from sequence
data. And even if agreement could be reached on a general method, it still
is not clear how that method might be applied in order to obtain a phy-
logenetic hypothesis that is judged reliable, or best, by some criterion.
These and other problems make phylogenetic analysis of DNA sequences
extremely complex.
There are three main classes of methods of phylogenetic inference ap-
plied to sequence data. First, with distance analysis, a distance (similarity/
dissimilarity) coefficient is computed for all pairwise comparisons of raw
sequence data, or for data corrected for multiple substitutions, and the
taxa are then clustered based on the values in the distance matrix (Fitch
and Margoliash, 1967; Farris, 1972; Felsenstein, 1984, 1988; Saitou and
Nei, 1987). In general, the tree having the shortest overall branch length
relative to that implied by the original data is chosen as the preferred tree.
See Swofford and Olsen (1990) for details about optimization procedures,
which differ among distance methods. Second, the method of maximum
likelihood uses the original sequence data and, working from a prior ev-
olutionary model of nucleotide substitution, computes a likelihood score
for each tree given the original data (Felsenstein, 1981; Bishop and Friday,
1985; Saitou, 1988, 1990). That tree with the highest likelihood is the most
preferred. The third method, cladistics or parsimony analysis, also uses
the original sequence data, and from calculations of character-state changes
(nucleotide substitutions) on alternative trees, the investigator chooses the
one having the fewest character-state transformations, that is, the one for
which homoplasies—parallelisms and reversals—are minimized (Fitch, 1971,
1977; Hendy and Penny, 1982; Cedergren et al., 1988).
There are several important reasons for preferring a parsimony approach
to phylogenetic inference when using DNA sequences. First, parsimony
methods are applied to the original sequence data rather than to some
transformed distance metric, thus avoiding an inevitable loss of information
(Farris, 1981, 1985; Penny, 1982). Second, while not entirely assumption
free, parsimony methods are either less dependent upon assumptions about
sequence evolution than are other methods; or they at least permit a more
informed investigation of any assumptions that might be made (Cedergren
et al., 1988). Third, algorithms that implement parsimony procedures are
far more sophisticated than distance or maximum-likelihood alternatives
and consequently allow a deeper analysis of the phylogenetic structure of
186 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

one's data and of the dynamics of sequence evolution. All three of these
reasons mean that parsimony provides the investigator with enhanced in-
terpretability of the results as compared to other methods.
Nevertheless, as noted above, the application of parsimony procedures
to sequence data is complex and has not yet received sufficient attention.
This chapter addresses two of the more general problems involving the use
of parsimony procedures in phylogenetic inference: (1) given that there
are physico-chemical/functional constraints on sequence evolution, espe-
cially in sequences coding for proteins or structural RNAs, how might
parsimony be applied in order to infer phylogenetic relationships, and (2)
how might we judge the phylogenetic informativeness of sequence data?
Our inquiry into these problems will be based on an analysis of mitochon-
drial DNA (mtDNA), but the findings should be applicable to other parts
of the genome as well.
Although it is a simple enough procedure to obtain a global parsimony
estimate of the phylogenetic structure inherent in any given data set, rec-
ognition that rates of transitions are, at least in vertebrate mtDNA, much
higher than those for transversions (Brown et al., 1982; Brown, 1983,1985;
Moritz et al., 1987), that there may be biases in base composition, and
that silent substitutions in the third codon position are more common than
those at first and second positions, have suggested to some investigators
that parsimony procedures will have difficulties when used to resolve phy-
logenetic relationships, especially among more distantly related taxa (Mor-
itz et al., 1987; Hayasaka et al., 1988a; Irwin et al., 1991). If homoplasy
is too extensive, it is argued, obtaining a strongly corroborated phylogenetic
tree might be difficult because even a minimum-length tree will have a
high number of homoplastic changes (see Archie, 1989; Smith, 1989). To
the extent that this is true, it raises several important questions: How might
we undertake parsimony analysis so as to reduce, or at least identify, the
influence of homoplasy? Second, how do we identify those parts of the
tree that seem to be less strongly supported and at the same time specify
which components of the data are less informative phylogenetically? Fi-
nally, in any given study, how do we determine how much sequence must
be obtained in order to reveal a reliable phylogenetic signal?
These questions speak to the notion of reliability and informativeness,
terms that have been used in different ways in the literature. Some inves-
tigators speak of "reliability of methods," referring to that method which
finds the best-fit tree for a given set of data. Others view reliability in
terms of an "ideal" result: the ability of a method, or a set of data, to
enable us to infer a tree that reflects the one true phylogeny. It is well to
keep the distinction of these two views of reliability in mind, for often it
is implied that once a best-fit tree is obtained, so too has an accurate
estimate of the one true phylogeny. It is sometimes argued, moreover, that
under a particular model of nucleotide sequence change, one or more
methods will fail to be reliable; that is, it would be expected that the best-
fit tree generated by those methods will not be an accurate estimate of the
PARSIMONY AND PHYLOGENETIC INFERENCE 187

true phylogeny (e.g., Felsenstein, 1978). All of these views on reliability,

however, are more arguments over methods, or perhaps of models of
evolutionary change, inasmuch as the one true phylogeny can never be
specified with certainty. Nor is it clear that a "realistic" evolutionary model
can be developed without some prior hyothesis of relationships with which
to investigate character-state change. Therefore, given only the data set
of interest, it is difficult to say whether the best-fit tree produced by any
method is or is not an accurate representation of the true phylogeny; it is
simply the most economical explanation for the available data.
Ideally, "reliability" is best thought of as a parsimony problem and
consequently should be judged in terms of congruence of data. "Inform-
ativeness," likewise, is also a parsimony problem. The two concepts are
related. A given tree, calculated to be the most-parsimonious, or best-fit,
for some set of data, can be taken to be a reliable estimate of the true
phylogenetic history of the taxa if most-parsimonious trees for other, in-
dependent data are congruent with it. A given sample of sequence can be
judged phylogenetically informative if it corroborates the most-parsimon-
ious tree derived from other data. This reciprocal duality between the
structure of data and what is inferred from them manifests the application
of parsimony (and congruence analysis) to hypotheses at two different
levels of inference: that of the data (character homology or hypotheses of
synapomorphy), and that of the phylogenetic hypotheses themselves (Cra-
craft and Mindell, 1989).

THE DATA

This study uses a single data set, and therefore, informativeness and re-
liability will be assessed in relation to the "stability" seen in the cladistic
signal across partitioned subsets of the data. That is, to the extent to which
clades are corroborated by these different subsets, the data will be judged
informative and the phylogenetic hypotheses themselves reliable. "Noise,"
on the other hand, will be judged to exist when the data do not corroborate
a consistent (or stable) cladistic signal.
The sequence data analyzed in this study include an 898 base pair (bp)
fragment of mtDNA that encompasses 458 bp of the 3' end of the NADH-
dehydrogenase subunit 4 (ND4 gene), 198 bp of three transfer RNA genes
(tRNAHis, tRNASer, and tRNA Leu ), and 242 bp of the 5' end of the ND5
gene (Brown et al., 1982; Hayasaka et al., 1988a: fig. 1:629). Comparative
sequences were available for 12 primate taxa: human (Homo sapiens),
common chimpanzee (Pan troglodytes), gorilla (Gorilla gorilla), orangutan
(Pongo pygmaeus), gibbon (Hylobates lar), Japanese macaque (Macaca
fuscata), rhesus macaque (M. mulatto), crab-eating macaque (M. fascicu-
laris), Barbary macaque (M. sylvanus), squirrel monkey (Saimiri sclureus),
Philippine tarsier (Tarsius syrlchta), and ring-tailed lemur (Lemur catta).
188 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

METHODS OF ANALYSIS

All analyses were undertaken using the Phylogenetic Analysis Using Par-
simony (PAUP) computer program of D. L. Swofford (1990). Version 3.0
of this program for the Macintosh system is especially well suited for anal-
ysis of sequence data because it readily permits investigation of the effects
of transitions versus transversions, the inclusion/exclusion of first, second,
or third codon positions in coding sequences, and of nucleotide substitution
parsimony of amino acid replacements. In addition, the program can com-
pute consensus trees by different methods, perform bootstrap analyses,
and calculate numerous measures of the information content for a given
tree structure.
In order to investigate the effects of sample size of characters on the
cladistic structure of the data, the 898 bp fragment was partitioned into
four sequential subsets: (1) the first 225 bp of the ND4 gene; (2) a 458 bp
segment including all of the ND4 gene that was sequenced; (3) a 656 bp
fragment including the 458 bp ND4 fragment plus the 198 bp of the three
tRNAs; and (4) the entire 898 bp fragment representing the ND4 fragment,
tRNAs, and the 242 bp fragment of the ND5 gene. The effects of data set
size were studied using this approach (rather than by random subsamples
of different sizes) because systematists add sequence data sequentially and
because it permitted an analysis of regions of the data set having different
functional domains. In all analyses, the trees were rooted by designating
the lemur as the outgroup. The results would have been the same if the
tarsier had been designated the outgroup; therefore, this study is only
investigating patterns of relationships within the clade consisting of the
other primate species (the anthropoids; terminology of clades follows Hay-
asaka et al., 1988a). A nohprimate outgroup was not used in order to make
our results more comparable to those of Hayasaka et al. (1988a).
Hayasaka et al. (1988a) proposed a phylogenetic hypothesis for the 12
primate taxa based upon substitution differences among those taxa and
using the neighbor-joining method of Saitou and Nei (1987). A single tree
was obtained (Fig. 10-1), which according to Hayasaka et al. (1988a:633),
was also obtained using the unweighted pair-group method of distance
analysis and by the distance-Wagner method. Hayasaka et al. concluded
that "Because these three different methods give phylogenetic trees with
the same topology, the phylogenetic relationships derived from these mtDNA
sequence comparisons [see our Fig. 10-1] appear reliable." It is not the
purpose of this chapter to speak to the issue of primate relationships per
se; rather the goal is to examine influences on the phylogenetic structure
of these mtDNA sequences, to explore the implications of that structure
for assessing the "reliability" of sequence data in general, and to investigate
how parsimony analysis might be effectively applied to sequence data.
Two series of analyses are used to examine and describe the extent to
which the primate mtDNA data set is phylogenetically informative. The
first series employs parsimony analysis on different subsets of the data with
the purpose of revealing the cladistic structure of most-parsimonious and
PARSIMONY AND PHYLOGENETIC INFERENCE 189

Figure 10-1 Phylogenetic hypothesis for 12 species of primates using a distance

analysis (neighbor-joining method; Saitou and Nei, 1987) of an 898 bp fragment of
mtDNA (from Hayasaka et al., 1988a). Names applied to clades discussed in this
chapter follow the terminology of Hayasaka et al. (1988a): (1) hominines, (2) great
apes, (3) hominoids, (4) catarrhines, (5) anthropoids, (6) macaques, and (7) prosi-
mians.

near-most-parsimonious trees. The most-parsimonious tree, or collection

of equally parsimonious trees, was found. Once its length was known, all
trees within five steps of that length were retained, and a majority-rule
consensus tree (Margush and McMorris, 1981) computed (majority-rule
consensus trees were adopted because they permit an analysis of the re-
lationship between the amount of phylogenetic signal and data set size and
because they are comparable to the results of bootstrap analyses as dis-
cussed below). The collection of trees having lengths within 1% of the
minimum-length tree was also examined and will be mentioned briefly.
A second series of analyses used the same subsets of data and constructed
a majority-rule bootstrap tree (Felsenstein, 1985) from 100 randomly sam-
pled (with replacement) replications of each of the subsets. The bootstrap,
in this case, is not intended as a test of statistical significance but as an
additional means of describing and evaluating the structure of the phylo-
genetic signal inherent in the data.

RESULTS

Global Parsimony Analysis

Most-Parsimonious and Near-Most-Parsimonious (Five-Step) Trees
A global parsimony analysis was undertaken using all nucleotide substi-
tutions in each of the four subsets of the 898 bp fragment. This analysis
was designed to answer two principle questions: Can a global parsimony
analysis, in which transitions and transversions are equally (uniformly)
190 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

weighted, reveal a stable phylogenetic signal? And, second, does the sta-
bility of that signal change as sample size is increased? Most-parsimonious
trees are described but not figured. Primary attention is given to the set
of trees within five steps of the most-parsimonious tree (or trees), but the
collection of trees within 1% of the minimum-length tree is also examined.

1. ND4:225 bp Fragment. A parsimony analysis of the smallest data set

of 225 characters yielded six equally parsimonious trees of 264 steps. These
trees have only two topological ambiguities: (1) the hominine trichotomy;
and (2) a trichotomy among the hominoids, macaques, and the squirrel
monkey. All other relationships were as postulated by Hayasaka et al.
(1988a: fig. 3:636) (see Fig. 10-1).
There are 95 trees within five steps of the shortest trees. The consensus
tree exhibits relatively poor structure except within the macaques (Fig.
10-2A). There is no resolution within the great apes clade or among the
basal lineages of the anthropoids.

2. ND4:458 bp Fragment. When the analysis is expanded to include all of

the ND4 fragment, a single most parsimonious tree of 608 steps is found
that is similar to the distance tree except for a sister-group relationship
between the chimp and gorilla.
In the 458 bp data set for ND4 there are 17 trees within five steps of
the single most-parsimonious tree (Fig. 10-2B). Resolution within the hom-
inoids has improved substantially, although the human-chimpanzee-gorilla
trichotomy, while recognized in 88% of the trees, is not resolved internally.
The hominoid relationships of the gibbon are still somewhat ambiguous,
occurring in only 53% of the trees. The additional data do not change the
resolution within the macaques, but do now place them closer to the hom-
inoids than to the squirrel monkey in 65% of the trees.

3. ND4 + tRNAs: 656 bp Fragment. A parsimony analysis of the ND4

fragment and the three tRNA genes produced a single most-parsimonious
tree of 766 steps identical to that of the preceding analysis in which the
gorilla is joined to the chimpanzee.
There were 9 trees within five steps of the most-parsimonious tree (Fig.
10-2C). With the addition of the 198 bp tRNA fragment to the ND4 se-
quence, resolution within the hominoid clade is improved, although the
human-chimpanzee-gorilla trichotomy remains. Resolution in other parts
of the tree remains relatively the same as for the 458 bp fragment.

4. ND4 + tRNAs + ND5: 898 bp Fragment. A parsimony analysis of the

entire data set yielded two equally most-parsimonious trees (1,157 steps),
one uniting chimpanzee and human, the other uniting chimpanzee and
gorilla; all other relationships were as in the distance tree (Fig. 10-1).
Within five steps of the most-parsimonious trees, five trees were found.
PARSIMONY AND PHYLOGENETIC INFERENCE 191

The consensus tree (Fig. 10-2D) is fully resolved but a human-chimpanzee

relationship was supported in only three of those trees (60%). Interestingly,
only the branch uniting the orangutan to the hominines received less sup-
port as compared to the 656 bp fragment five-step consensus tree. Overall
there was a marked improvement in resolution with the larger data set.

5. ND4 + ND5 Fragments Compared With the tRNA Fragment. In order

to assess the relative influence on phylogenetic structure of mtDNA se-
quences that code for proteins versus structural RNAs, the two regions
were analyzed separately. A parsimony analysis of the 700 bp encompassing
the two coding regions of ND4 and ND5 taken together produced two
equally parsimonious trees (999 steps) that differ only in the relationships
of the chimpanzee to humans or to the gorilla; in all other respects the
two trees are identical to the distance tree (Fig. 10-1).
There were five trees within five steps of the two shortest trees, and
their consensus is shown in Figure 10-2E. Resolution is improved over the
656 bp fragment except in the placement of the orangutan with the hom-
inines. Compared to the entire 898 bp data set (Fig. 10-2D), the five-step
tree of ND4 + ND5 has less resolution within the hominines but the
relationship between the hominoids and the macaques is more strongly
supported by the ND4 + ND5 data set.
An analysis of the 198 bp fragment that includes the three tRNAs sug-
gests that their contribution to instability is minimal. A single most-par-
simonious tree was found (154 steps) that resolved relationships within the
hominoids and macaques identical to those of the distance tree. This tRNA
tree, however, placed the macaques outside the other taxa to yield: (ma-
caques (tarsier (squirrel monkey + hominoids))).
There are 254 trees five steps away from the shortest tRNA tree and,
not unexpectedly, the consensus tree (Fig. 10-2F) exhibits relatively little
resolution.

Near-Most (1 %)-Parsimonious Trees

The collection of trees within 1% of the length of the most-parsimonious
tree has been used as a method to investigate the cladistic structure of
sequence data (Smith, 1989) and was also examined in this study. Table
10-1 summarizes the results of this analysis.
The number of trees recovered within 1% of the shortest tree(s) did not
vary noticeably among data sets despite the fact that larger data sets had
trees of longer branch lengths. Only two clades, hominines and hominoids,
consistently gained cladistic support as the size of the data set increased.
All others showed decreasing support or no marked change. Some clades
(anthropoids and macaques) were well defined no matter what size the
data set. Interestingly, a global parsimony analysis shows strong resolution
within the macaques across the four primary data sets, but no resolution
is seen within the hominines.
E. ND4+ND5: 700 bp F. tRNAs:198bp
5 trees (1004 steps or less) 254 trees (159 steps or less)
shortest tree = 999 steps (2 trees) shortest tree = 154 steps (1 tree)

Figure 10-2 Majority-rule consensus trees of a global-parsimony analysis of all trees within five steps of the
most-parsimonious solution for subsets of the ND4 + tRNAs + ND5 fragment of mtDNA, as discussed in the
text. Numbers next to internal branches represent frequencies of occurrence for those clades among the most-
and near-most-parsimonious solutions.
Table 10-1 Global Parsimony Analysis of Near-Most (1%)-Parsimonious Trees: Summary of Majority-Rule Consensus Trees
Clade
Japanese +
Fragment Human + Great Japanese + Rhesus +
(no. of trees) Chimpanzee Hominines Apes Hominoids Catarrhines Anthropoids Rhesus Crab-eating Macaques
ND4: 225 bp n.r.* 52' 100 55 62 100 100 100 100
(29)
ND4:458 bp n.r. 79 96 58 71 100 92 96 100
(24)
ND4 + tRNAs: 656 bp n.r. 86 95 73 64 100 100 82 100
(22)
898 bp fragment n.r. 100 59 100 59 100 100 88 100
(17)
ND4 + ND5: 700 bp n.r. 100 69 100 69 100 85 100 100
(13)
tRNAs: 198 bp 56 100 n.r. 100 n.r.* 60 100 52 100
(25)
*n r., clade not resolved at 50% level
'Percentage of trees in which clade is resolved in majority-rule consensus trees
'Squirrel monkey clustered with the hommoids.
PARSIMONY AND PHYLOGENETIC INFERENCE 195

Summary of Global Parsimony Analysis

Numerous cladistic groupings are consistently revealed in a global parsi-
mony analysis. With the exception of the smallest data set, the first 225
bp fragment of ND4, global parsimony analysis of each of the other data
sets was able to resolve all relationships consistently except for the tricho-
tomy within the hominines. This is also true of the combined ND4 + ND5
fragment of 700 bp, and, as noted above, even the most-parsimonious tree
of the tRNAs was able to resolve most of the relationships.
These results indicate that this relatively small 898 bp fragment of mtDNA
contains significant phylogenetic signal for taxa whose divergences extend
as far back as 50 to 60 million years, that even "noise" produced by the
high transition rate of mtDNA cannot obscure the signal, and that a global
parsimony analysis is sufficiently powerful to resolve a consistent cladistic
pattern even in data that are generally thought to be "noisy."
These conclusions are reinforced when the collections of near-most-
parsimonious trees are examined (Fig. 10-2). In general, there is improved
resolution as more data are added (Fig. 10-2A through 10-2E), but the
trend is not always strong. Most importantly, a consistently strong cladistic
signal is found in all but the very smallest data sets.
When the structure of trees within 1% of the most-parsimonious solution
is examined, however, far fewer of the relationships are identified by a
strong cladistic signal (e.g., clusters supported in 90% or more of the trees;
see Table 10-1). Only the monophyly of the macaques and of the anthro-
poids is consistently revealed across all four primary data sets. As noted
earlier, furthermore, saving 1% of the most-parsimonious trees does not
produce a clear convergence in the phylogenetic signal as the data set is
enlarged. This results from the fact that in each subset of the data there
are many trees within 1% of the minimum-length tree, including some of
extremely variant tree topology.

Transversion Parsimony Analysis

Most-Parsimonious and Near—Most-Parsimonious (Five-Step) Trees

Does the large bias for transition substitutions within mtDNA increase the
"noise" and lead to instability of the phylogenetic signal within the primate
data set? Can the signal be improved by examining transversions alone?
And, does the stability of a phylogenetic signal inferred from a parsimony
analysis of transversions improve with increasing sample size? These ques-
tions were investigated by repeating the above analysis using transversion
differences.

1. ND4: 225 bp Fragment. An analysis of the first 225 bp of the ND4

fragment produced two equally parsimonious trees (79 steps). In both trees,
the macaques were not fully resolved, with only the Japanese-rhesus ma-
caque clade being held in common. The major difference between the two
196 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

trees was found in the placement of Homo: in one tree it was with gorilla,
in the other with chimpanzee.
There were 681 trees within five steps of the most-parsimonious solu-
tions. Their majority-rule consensus (Fig. 10-3A) shows very strong support
for three clades, the anthropoids, catarrhines, and macaques, but fails to
resolve the relationships of any other group.

2. ND4:458 bp Fragment. A transversion parsimony analysis of the entire

ND4 fragment yielded two equally parsimonious trees (196 steps). One is
identical to the distance tree (Fig. 10-1), and the other failed to resolve a
trichotomous relationship at the base of the macaques: (Barbary, crab-
eating, Japanese + rhesus).
Although 114 trees were found within five steps of the most-parsimonious
solution, the majority-rule consensus (Fig. 10-3B) shows a marked im-
provement in resolution compared to the 225 bp fragment. Clades are
resolved within the hominoids, but not within macaques.

3. ND4 + tRNAs: 656 bp Fragment. A transversion analysis of the ND4

fragment along with the tRNAs produced two equally parsimonious trees
(239 steps) that have identical structure to the two most-parsimonious trees
of the preceding analysis.
There were 59 trees within five steps of the two most-parsimonious trees.
Compared to the 458 bp fragment, some additional, slight resolution is
apparent in the hominines and the macaques (Fig. 10-3C).

4. ND4 + tRNAs + ND5: 898 bp Fragment. A transversion parsimony

analysis of the entire data set yielded a single most-parsimonious tree (381
steps). It has the same structure as the distance tree (Fig. 10-1).
For the transversions in the total data set, 46 trees were found within
five steps of the most-parsimonious tree. Although there is no resolution
within the hominines or macaques, compared to the previous consensus
trees, there is a marked increase in support for the hominines and catar-
rhines (Fig. 10-3D).

5. ND4 + ND5 Fragments Compared With the tRNA Fragment. In a trans-

version parsimony analysis of the ND4 and ND5 fragments combined (700
bp), a single most-parsimonious tree was obtained (338 steps) that is iden-
tical to the distance tree. A transversion parsimony analysis of the three
tRNAs (198 bp) found seven equally parsimonious trees (42 steps). The
consensus tree for these showed the chimpanzee united with the human
in 77% of the cases, but with no resolution among the other hominoids,
including the gorilla. The Japanese and rhesus macaques were clustered
as sister-species, but no other nodes were resolved within macaques, and
the squirrel monkey was united with the macaques.
Transversion analysis of the 700 bp coding region finds 66 trees within
five steps of the most-parsimonious tree. Their majority-rule consensus
PARSIMONY AND PHYLOGENETIC INFERENCE 197

tree (Fig. 10-3E) is very similar to that of the entire 898 bp data set (Fig.
10-3D) except that the great apes are much less supported (76% versus
91%). Not unexpectedly, there are a large number of trees within five
steps of the most-parsimonious solutions for the tRNA fragment. The
majority-rule consensus for 3,000 trees (computer memory was exhausted;
many more could have been found) is shown in Figure 10-3F. Only the
macaques as a group are well defined.

Near-Most (1 %)-Parsimonious Trees

Table 10-2 summarizes majority-rule consensus tree support for all trees
with lengths within 1% of the most-parsimonious transversion solutions.
A consistent pattern is observed: except for the tRNA fragment, virtually
all of the "deep" cladistic events are strongly supported in each of the five
primary analyses, and this also includes the shortest 225 bp fragment. In
contrast, strong resolution of the "more recent" cladistic events, including
those within the macaques and hominines, is lacking.

Summary of Transversion Parsimony Analysis

The most general conclusion to be drawn from a comparison of the global
parsimony analysis (Fig. 10-2; Table 10-1) with the transversion parsimony
analysis (Fig. 10-3; Table 10-2) is that the use of transversions, by them-
selves, reveals a strong cladistic signal for the relatively older clades. This
is especially evident when examining those trees having a length within
1% of the most-parsimonious solutions (Table 10-2), which in most of the
data sets included all those trees within two or three steps of the minimum-
length solution. Since many more trees were examined in the five-step
analysis (Fig. 10-3), support for some of these deeper branches declined,
but in general remained fairly strong. None of these methods of analysis,
when applied to transversion differences alone, were able to reveal a strong
cladistic signal within the hominines or within the macaques.
The five step analysis revealed a slight increase in the strength of the
cladistic signal as the data set increased in size. The 656 bp fragment (Fig.
10-3C) is not much more informative than the 458 bp fragment (Fig.
10-3B), but the total data set (Fig. 10-3D) and the ND4 + ND5 fragment
(700 bp; Fig. 10-3E) exhibit stronger cladistic support within the hominoids.

Global Parsimony of First and Second Codon Positions

Does the high rate of transitions in third positions contribute noise and
therefore result in decreased stability of the phylogenetic signal? This
question was investigated by a global parsimony analysis of the ND4 and
ND5 fragments in which the third positions were excluded. The analysis
yielded a single most-parsimonious tree of 519 steps. There were only two
differences from the distance tree. The first had the chimpanzee united
with the gorilla, but the most unexpected result (given the preceding anal-
Figure 10-3 Majority-rule consensus trees of a transversion parsimony analysis of all trees within five steps of
the most-parsimonious solution for subsets of the ND4 + tRNAs + ND5 fragment of mtDNA, as discussed in
the text. Frequencies as in Figure 10-2.
Table 10-2 Transversion Parsimony Analysis of Near-Most (1%)-Parsimonious Trees: Summary of Majority-Rule Consensus Trees
Clade
Japanese +
Fragment Human + Great Japanese + Rhesus +
(no. of trees) Chimpanzee Hominines Apes Hominoids Catarrhmes Anthropoids Rhesus Crab-eating Macaques

ND4: 225 bp n.r.* 65* 100 100 100 100 53 n.r. 100
(17)
ND4:458 bp n.r. 100 100 100 100 100 n.r. n.r. 100
(14)
ND4 + tRNAs: 656 bp 75 100 100 100 100 100 n.r. n.r. 100
(8)
898 bp fragment 59 100 96 100 100 100 n.r. n.r. 100
(27)
ND4 + ND5: 700 bp n.r. 100 100 100 100 100 n.r. n.r. 100
(38)
tRNAs: 198 bp 77 n.r. n.r. 74 n.r.* 100 64 n.r. 100
(77)
*n.r., clade not resolved at 50% level.
'Percentage of trees in which clade is resolved in majority-rule consensus trees
^Squirrel monkey clustered with macaques.
PARSIMONY AND PHYLOGENETIC INFERENCE 201

yses) was the closer relationship of the gibbon, rather than the orangutan,
to the great apes.
There were eight trees within five steps of the shortest tree (Fig.
10-4A). The consensus tree showed little resolution within either the hom-
inoids or the macaques, but it did resolve the monophyly of both of these
clades and clustered them relative to the squirrel monkey. This analysis
suggests that attempting to enhance the phylogenetic signal by eliminating
transitions in the third position was unsuccessful. It had no effect on im-
proving the resolution of the older cladistic events, which were strongly
resolved by global parsimony, and not unexpectedly, it led to a decrease
in the signal within the macaques (compare Fig. 10-4A with Fig. 10-2E).

Amino Acid Replacements: Substitution Parsimony

Amino acid replacements are often taken as being more informative of
relationships than nucleotide substitutions, especially for more distant
branching events. The structure of the phylogenetic signal of replacements
was examined by parsimony analysis of the numbers of substitutions that
account for amino acid differences among the taxa. A transformation ma-
trix for vertebrate mtDNA provided by D. L. Swofford was used to convert
amino acid differences to substitution differences and then each tree was
evaluated.
A single most-parsimonious tree of 387 steps was found for the 231
codons of the ND4 and ND5 fragments arranged tandemly. Relationships
were similar to those of the distance tree except within the hominoids
where gorilla and human were united and the gibbon (not orangutan) was
the sister-group of the hominines.
There were 49 trees within five steps of the most-parsimonious tree. The
majority-rule consensus tree (Fig. 10-4B) shows relatively little resolution,
either within the macaques or the hominoids, thus suggesting that this
method is less effective in revealing phylogenetic signal than is a global
parsimony analysis.

Bootstrap Analysis
Global Parsimony Bootstrap

1. ND4:225 bp Fragment. A bootstrap analysis of the first 225 bp of ND4

produced a majority-rule consensus tree identical to the distance tree ex-
cept for chimpanzee being united with gorilla (Fig. 10-5A). The latter
relationship means little, however, given that all three alternatives within
the hominines are about equally supported. Clades having substantial sup-
port in this data set include: (1) the macaques and their inclusive groupings;
(2) the great apes; and (3) the anthropoids.

2. ND4: 458 bp Fragment. A bootstrap analysis of the entire ND4 frag-

ment produced a tree very similar to the preceding, except that human
Figure 10-4 (A) A majority-rule consensus tree of a global parsimony analysis of all trees within five steps of
the most-parsimonious solution for the tandemly arranged fragments of ND4 and ND5 genes of mtDNA with
third codon positions excluded. (B) A majority-rule consensus tree of a global parsimony analysis of nucleotide
substitutions resulting in amino acid replacements in 231 codons of the ND4 and ND5 genes of mtDNA.
Frequencies in both (A) and (B) as described in Figure 10-2.
PARSIMONY AND PHYLOGENETIC INFERENCE 203

and chimpanzee are narrowly united (Fig. 10-5B). Support for each of the
clades has generally improved, especially for the catarrhines, the homi-
nines, and the macaques.

3. ND4 + tRNAs: 656 bp Fragment. The majority-rule consensus tree for

this data set (Fig. 10-5C) is identical to the preceding. In general, support
for each clade increases, although the catarrhine clade is found in only
50% of the replicates.
4. ND4 + tRNAs + ND5: 898 bp Fragment. The majority-rule consensus
tree for all the data keeps the same pattern of relationships as the preceding,
except for uniting chimpanzee and gorilla (50% of the replicates; Fig.
10-5D). Bootstrap support for the other clades has increased, sometimes
substantially (e.g., hominoids, catarrhines).
5. ND4 + ND5 Fragments Compared With the tRNA Fragment. A boot-
strap majority-rule consensus tree of the protein-coding region (Fig.
10-5E) shows a similar pattern to that of the distance tree. Levels of support
are close to those of the 898 bp tree (Fig. 10-5D). Bootstrapping the 198
bp of the rRNAs (Fig. 10-5F) suggests that they contribute less noise to
the data than might be expected, except at the base of the anthropoids
where the squirrel monkey is united with the hominoids. The tRNAs, how-
ever, strongly support the monophyly of the macaques and a rhesus +
Japanese macaque relationship.

Summary of the Global Parsimony Bootstrap

Not unexpectedly, the phylogenetic signal, as measured by the percentage
of congruent bootstrap replicates, improves as the sample size increases.
Thus, for the 898 bp fragment (Fig. 10-5D), six clades were supported
more than 97% of the time. Nevertheless, if one preferred to interpret the
"significance" of these results stringently by collapsing all branches that
failed to meet a 90 or 95% criterion, the bootstraps of characters for each
subset of the data would reveal a cladistic structure not very much different
from that seen in previous analyses (Fig. 10-2).
Transversion Parsimony Bootstrap
1. ND4: 225 bp Fragment. A bootstrap of transversions contained in the
first 225 bp of the ND4 fragment produces a majority-rule consensus tree
much like the distance tree except for the lack of resolution at the base of
the macaques (Fig. 10-6A). The monophyly of the macaques, the catar-
rhines, and the anthropoids is strongly supported by a bootstrap analysis
of these data.

2. ND4: 458 bp Fragment. A bootstrap of the entire ND4 fragment pro-

duced a tree identical to the distance tree (Fig. 10-6B). Relative to the 225
bp fragment, the pattern of support varies. Clades within the hominoids
are better supported with this expanded data set, although the monophyly
Figure 10-5 Majority-rule bootstrap trees of a global parsimony analysis for subsets of the ND4 + tRNAs +
ND5 fragment of mtDNA, as discussed in the text. Frequencies of occurrence among the bootstrap trees are
presented for the internal branches.
Figure 10-6 Majority-rule bootstrap trees of a transversion parsimony analysis for subsets of the ND4 + tRNAs
+ ND5 fragment of mtDNA, as discussed in the text. Frequencies as in Figure 10-5.
208 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

of the catarrhines is much less so. Support for a human + chimpanzee

relationship has increased to 66% of the replicates.

3. ND4 + tRNAs: 656 bp Fragment. The relationships implied by this

data set (Fig. 10-6C) are the same as in the preceding analysis. There is
an increase in the support of all clades except that of the catarrhines.

4. ND4 + tRNAs + ND5: 898 bp Fragment. A bootstrap of the trans-

versions in the entire data set yielded a tree showing no changes in cladistic
structure from the preceding analysis, but with a general increase in the
cladistic support at most nodes (Fig. 10-6D).

5. ND4 + ND5 Fragments Compared to the tRNA Fragment. A trans-

version bootstrap analysis of the combined ND4 and ND5 sequences once
again produced the same tree as did using all the data and, in general,
with approximately the same level of support (Fig. 10-6E). An analysis of
transversions in the tRNAs shows that this data set retains much of the
phylogenetic structure present in the protein-coding regions, but virtually
all of the nodes are poorly supported (Fig. 10-6F). In the tRNA data set,
the squirrel monkey clusters with the macaques, showing ambiguity at the
base of the anthropoids.
Summary of the Transversion Bootstrap Analysis
As with the global parsimony bootstrap analysis (Fig. 10-5), there was a
general increase in support for the nodes as sample size was increased.
Neither bootstrap analysis provided significantly better support than the
other. Although the transversion bootstrap produced much better reso-
lution within the hominines, the global bootstrap analysis resolved the
relationships within the macaques much more convincingly. Both did about
equally well with the other parts of the tree.
The trees of Figure 10-6 are all majority-rule consensus trees, and if the
level of stringency of cladistic support was taken to be 90 or 95% of the
bootstrap replicates, the resulting topologies would not differ very much
from those of the transversion parsimony analysis (Fig. 10-3).

DISCUSSION

Informativeness, Noise, and the Primate Data Set

A major concern when analyzing any systematic data is to assess the extent
to which they are phylogenetically informative. Systematists are also in-
terested in knowing whether a particular analytical method is effective in
maximizing informativeness and thus results in phylogenetic hypotheses
that are more strongly supported than those produced by other methods.
These problems are especially acute with sequence data, not only because
PARSIMONY AND PHYLOGENETIC INFERENCE 209

of the large numbers of characters that are normally generated, but also
because the evolutionary dynamics of DNA are such that a large amount
of noise (i.e., homoplasy) would not be unexpected.
How might the informativeness and noisiness of the primate data set be
evaluated? One approach might examine sets of near-most-parsimonious
trees and judge informativeness by the degree to which the cladistic struc-
ture is corroborated across those trees. A second method might examine
consensus trees produced by resampling of the data themselves and ask to
what extent each node is supported; that is, in what percentage of the trees
forming the consensus does a given node occur? Both methods examine
the structure of data, but in different ways.
A consideration of the amount of support for each node as seen in the
majority-rule consensus trees (Figs. 10-2 and 10-3; Tables 10-1 and 10-2)
permits a detailed description of the phylogenetic signal contained in the
data. In the global majority-rule consensus of trees within five steps of the
most-parsimonious solutions (Fig. 10-2), the relationships within the ma-
caques are fully resolved using the largest data sets. The hominines always
cluster relative to the orangutan, and these two groups cluster relative to
the gibbon. The major lineages of the tree—hominoids, catarrhines, an-
thropoids, and macaques—are also well supported by the largest data sets.
In the transversion majority-rule trees (Fig. 10-3), there is little or no
phylogenetic signal within the macaques or hominines, but the data un-
ambiguously resolve all other clades, especially using the largest data sets.
Taken together, these two analyses reveal a consistent phylogenetic signal
for all clades except those within the hominines.
Bootstrap analysis constitutes a second approach to characterizing the
amount of phylogenetic signal in the data. Bootstrap sampling of the total
data set (Fig. 10-5) reveals a strong phylogenetic signal within the ma-
caques, and comparison with the bootstrap analysis of transversions alone
(Fig. 10-6) leads to the conclusion that a substantial component of that
signal is in fact coming from transition differences. Within the hominines,
however, it is transversions, not transitions, that provide the most con-
sistent signal. In both bootstrap analyses, the major lineages are all well
supported using the largest data sets, although the phylogenetic position
of the squirrel monkey is the least well-corroborated.
How do the two approaches compare? Not unexpectedly, the results
using these two different methods parallel one another to a very great
extent (e.g., compare Figs. 10-2 and 10-5; and 10-3 and 10-6). Whenever
a clade is strongly supported in a near-most-parsimonious consensus tree,
it is also strongly supported in the comparable bootstrap tree. Differences
arise in those clades that lack strong support. Because bootstrap trees tend
to be fully resolved, they have the appearance of being more informative,
yet there is generally little character support underlying that increased
resolution. Still, when examination of the structure of near-most-
parsimonious trees often yields an unresolved polytomy, bootstrap resam-
pling of the data seems more sensitive to finding the "correct" signal, even
210 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

if that signal is not especially strong. Again, one need not interpret boot-
strap results in a strict statistical sense, but simply as another method of
evaluating the presence and strength of signal in the data. Thus, boots-
trapping of transversions across the entire data set provides resolution of
relationships within the macaques (Fig. 10-6D), whereas the majority-rule
consensus analysis of near-most-parsimonious trees (Fig. 10-3D) does not.
This result generally holds for most subsets of the data, not only for re-
lationships within the macaques, but also within the hominines.

Does Phylogenetic Signal Stabilize with More Data?

A critical practical problem of phylogenetic inference using DNA se-
quences is deciding how much sequence must be obtained before relation-
ships can be resolved. In attempting to answer this question, factors other
than sequence length per se cannot be ignored. It is generally assumed,
for example, that certain segments of DNA will be more informative than
others and therefore less sequence will be required to reveal a phylogenetic
signal. This supposition usually entails considerations of evolutionary rate,
which is discussed in the next section. Here we ask whether phylogenetic
signal within the primate data set is related to data set size and which
method of analysis best reveals that signal.
The global parsimony analysis of near-most-parsimonious trees shows
an overall increase in the strength of signal as sample size is increased (Fig.
10-2A through 10-2E). There is a marked improvement in the 458 bp tree
as compared to the 225 bp tree. With respect to the 656 bp consensus tree
(Fig. 10-2C), some nodes are slightly better supported compared to the
458 bp tree, others are not. Generally, the 898 bp and 700 bp trees have
stronger corroboration than the 656 bp tree.
The transversion parsimony analysis of near-most-parsimonious trees
also exhibits increasing strength of signal with increasing sample size (Fig.
10-3). That increase in signal, however, is confined to the deeper branches
and to the hominoids. Resolution within the macaques is not affected by
increases in sample size using transversions alone.
As noted earlier there is a general increase in phylogenetic signal, as
determined by bootstrap support, when sample size increases. This is true
for both the global parsimony and transversion parsimony analyses (Figs.
10-5, 10-6). Not all nodes show increased support between successive sub-
sets of data, and not all nodes are well supported, but there is a definable
convergence toward a single tree. These results suggest that bootstrap
analysis may be somewhat more informative as a descriptor of the effects
of sample size on phylogenetic signal than is examination of near-most-
parsimonious trees.

Evolutionary Rate and Phylogenetic Signal

The degree to which a segment of DNA is considered likely to be inform-
ative is generally thought to be related to its rate of base substitution
PARSIMONY AND PHYLOGENETIC INFERENCE 211

relative to the divergence times that the investigator is trying to resolve.

Segments with high rates of change are assumed to be less informative for
deeper branches because the accumulation of homoplasy (parallel and back
mutations) erases any signal (Hayasaka et al., 1988a; Smith, 1989). As
substitutions "saturate," genetic distances among the taxa approach an
asymptote relative to divergence time (Brown, 1983; Moritz et al., 1987),
and this observation has led to the conclusion that mtDNA sequence var-
iation will be less useful in resolving relationships among relatively more
distantly related taxa (Moritz et al., 1987).
To the extent that data are available (Brown, 1985), the genes examined
in this study (ND4 and ND5) might be categorized as "relatively fast"
compared to many other protein-coding genes, tRNAs, or ribosomal RNA
genes. Hayasaka et al. (1988a:637-638) estimated that the relationship
between divergence time and the number of substitutions among taxa [cor-
rected for multiple mutations (Gojobori et al., 1982)] is linear only up to
about 30 million years and that older divergence times (e.g., those deeper
than the base of the hominoids) cannot be resolved "conclusively." Yet,
such a conclusion is method-dependent. Even though ND4 and ND5 are
"fast," the present analysis suggests their rate may have only marginal
influence on the relationship between phylogenetic signal and divergence
time (at least within primates). Depending upon how parsimony is applied
to the data, these genes are just as capable of resolving the more ancient
divergences within the primates as they are the most recent. Thus, by
examining the relative influence of transitions and transversions, a rea-
sonably consistent phylogenetic signal is revealed across the tree (rela-
tionships within the hominines constitute a slight exception; see below).
These conclusions deserve to be tempered, however, by noting that tran-
sition/transversion differences do not become equal until the very most
distant relationships between the prosimians and the other taxa are con-
sidered (Hayasaka et al., 1988a; see our Table 10-3). It could be argued,
therefore, that relationships within the anthropoids are not affected by a
saturation effect (although saturation is evident at the base of the anthro-
poids).

The Resolving Power of Most-Parsimonious Trees

It is often thought within molecular systematics that parsimony methods
will often fail to be informative because of the high degree of homoplasy
in sequence data (see next section). Therefore, it is useful to examine that
supposition by asking whether the most-parsimonious trees, based on a
consideration of all of the data, were in fact accurate estimates of the
phylogenetic relationships of the primates, or whether the "noise" in these
"fast" evolving genes made that resolution difficult or impossible.
The answer to this question may be surprising to some: the relationships
of the primates were consistently revealed by the most-parsimonious trees
of each of the data sets. In general, the only ambiguity involved resolution
Table 10-3 Transition/Transversion Differences Among 12 Primate Taxa (Sequence Data Derived From Hayasaka et al., 1988a)
Clade 1 2 3 4 5 6 7 8 9 10 11 12
1. Human —
2. Chimpanzee 74/5 —
3. Gorilla 84/8 86/9 —
4. Orangutan 108/35 119/34 116/33 —
5. Gibbon 116/45 124/44 123/45 115/52 —
6. Japanese macaque 134/74 144/73 139/72 143/75 145/77 —
7. Rhesus macaque 131/77 145/76 135/75 141/78 133/80 32/3 —
8. Crab-eating macaque 149/74 166/73 162/72 158/75 154/77 71/4 81/5 —
9. Barbary macaque 155/74 149/73 146/72 137/77 139/77 103/8 100/9 104/6 —
10. Squirrel monkey 146/98 154/99 145/96 139/111 135/106 156/102 158/104 154/102 154/102 —
11. Tarsier 150/136 147/137 145/132 136/131 143/133 145/134 144/134 144/132 148/136 137/151 —
12. Lemur 133/142 134/141 123/138 129/129 128/137 126/126 128/128 140/126 128/128 119/135 101/126 —
PARSIMONY AND PHYLOGENETIC INFERENCE 213

of relationships within the hominines, which has been a long-standing prob-

lem that even very large amounts of data have had difficulty in resolving.
The most parsimonious solution of the smallest data set, the 225 bp frag-
ment, also showed a trichotomy at the base of the anthropoids, yet was
still remarkably effective in resolving relationships among the other taxa.
It therefore seems that a straightforward parsimony analysis of un-
weighted data, including both transitions and transversions, is an effective
method for estimating what appears to be the "correct" tree, even in subsets
of the data. It may be that because most of the data are not saturated with
transition substitutions, there remains relatively little noise in the data and,
therefore, parsimony is effective in revealing the signal. On the other hand,
it may be that a simple parsimony approach is in fact a very powerful
method and can discern phylogenetic signal, even in data having a strong
component of noise [consistency indexes (Kluge and Farris, 1969) calcu-
lated after excluding uninformative characters typically had values slightly
less than 0.600].

Parsimony and Phylogenetic Inference Using DNA Sequences

Many workers have long advocated parsimony analysis as a method of
phylogenetic inference for DNA sequence data (Fitch, 1971, 1977; Hendy
and Penny, 1982). At the same time, there has been concern within the
molecular systematic community that once the algorithmic problems have
been solved—and large advances have been made in recent years—there
will still remain important questions as to exactly how parsimony proce-
dures are to be applied to sequence data. The focus of most applications
of parsimony has been to find the most-parsimonious tree or trees. Yet,
systematic (comparative) and functional analyses of DNA and its evolution
have established a body of knowledge that creates a broad spectrum of
choices and challenges for the systematist. Included in this body of knowl-
edge are the observations that the frequency of transitions and transver-
sions are not always equal; third position changes are more frequent than
first or second; and secondary structure configurations can bias the fre-
quency and direction of nucleotide change.
The investigator is thus faced with procedural difficulties from the outset,
difficulties that raise important philosophical and empirical questions. At
their most basic level, these questions involve the problem of what con-
stitutes evidence in phylogenetic inference, a subject that has received too
little attention in the literature (a recent notable exception is Kluge, 1989).
Do we base our best estimate of phylogenetic relationships on all the
sequence evidence that is available, or, if on the basis of other evidence
we have reason to think that some of the data are likely to be misleading,
should those be eliminated from the analysis? If the latter, what are the
criteria for choosing that portion to be eliminated? Moreover, what jus-
tifications do we advance to establish the validity of those criteria? Do we
treat all characters (in this case, nucleotide positions) equally ("weighted
214 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

uniformly") or should some be weighted more than others (eliminating

characters might also be viewed as a form of weighting)? Do we treat all
character-state transformations at any given nucleotide position equally,
or are they to be weighted differently? In both cases, how is weighting
itself to be justified and how might weights be determined objectively?
These questions can only be a prelude to a full discussion of how parsimony
is to be applied inasmuch as many other issues (e.g., identifying signal,
assessing topological reliability, the notion of congruence) also have a
bearing on our choices regarding systematic procedure (Cedergren et al.,
1988; Kluge, 1989).
One theme of this chapter has been to ask what might be the best
approach for applying parsimony procedures to sequence data. Because
the problem involves so many complex issues, not the least of which are
the philosophical tendencies of the investigator, it is unlikely that the sys-
tematic community will agree on a unified answer. If one sees the goal of
systematic research to be the discovery of the one true history of life, and
accepts the precept that we will never have certain knowledge of that
history, then our only recourse would seem to be to accept the hypothesis
of relationships which is most consistently revealed by the data regardless
of which analytical techniques might have been employed (this does not
imply, of course, that all methods of analysis are necessarily equally useful
or appropriate). This is essentially an appeal for acceptance of stability of
cladistic signal and to identify as noisy those cladistic patterns that are
ambiguous or less consistently revealed.
There has been substantial support in the literature for weighting among
nucleotide positions (e.g., within codons) or among alternative character-
state transformations within positions, and many weighting procedures
have been suggested (e.g., Farris, 1969; Altschul et al., 1989; Williams and
Fitch, 1989,1990). Although there is a rationale for thinking that weighting
will improve the resolution of cladistic signal, what is less clear is which
method of weighting is "best" in any case. In principle, at least, a tree
derived from weighted data should be a better or more reliable estimate
of the true phylogeny than a tree derived from unweighted data, simply
because the effect of weighting should be to emphasize the best available
evidence (Carpenter, 1988). Yet, a final conclusion regarding the reliability
of the resulting tree can only be reached by comparison to results based
on more data and an assessment of congruence of their cladistic signal. It
is possible, after all, that various assumptions about the evolutionary pro-
cess that may underlie a particular method of weighting are themselves
incorrect and thus may have led to spurious results. An unfortunate tend-
ency within molecular systematics has been the application of a weighting
scheme and the uncritical acceptance of the resulting tree. Currently avail-
able parsimony algorithms are designed to produce a single point solution;
that is, a single most-parsimonious tree or a set of equally most parsimon-
ious trees. Yet, it is difficult to judge from a single most-parsimonious tree
whether a cladistic grouping represents a strong (stable) signal found in
PARSIMONY AND PHYLOGENETIC INFERENCE 215

the data or whether the data are noisy with respect to that group's resolution
(e.g., within the hominines). Thus, even if the data are weighted, questions
about the reliability of the result still exist in addition to those raised by
any justification of the weighting methods themselves.
The analyses undertaken in this study explore ways of revealing phy-
logenetic signal in sequence data through parsimony procedures. Taken
together, the results of this analysis suggest that if there is any phylogenetic
signal inherent in the data, parsimony methods will find it. The ultimate
arbiter of this conclusion must be congruence with phylogenetic relation-
ships inferred from other data sets, which in the case of the primates
examined here support the conclusions of this chapter (e.g., Hayasaka et
al., 1988b; Miyamoto and Goodman, 1990). Taken separately, each ap-
proach to parsimony reveals a significant portion of that signal, but reso-
lution in some parts of the tree may remain ambiguous. Most-parsimonious
trees based on the entire 898 bp are essentially congruent with the distance
tree of Figure 10-1, there being two equally most-parsimonious trees dif-
fering only in the placement of chimpanzee with human or with gorilla;
the single most-parsimonious tree using transversions alone resolved that
ambiguity in favor of chimpanzee + human. An assessment of the strength
of the phylogenetic signal is best described, however, by comparison of
the majority-rule consensus trees derived from near-most-parsimonious
solutions using global (Fig. 10-2) and transversion (Fig. 10-3) parsimony,
along with those derived from bootstrap analysis of global (Fig. 10-5) and
transversion (Fig. 10-6) parsimony solutions. Indeed, a comparison using
these four methods of analysis demonstrates a strong phylogenetic signal in
even the smallest subsets of the data, including the first 225 bp fragment and
the 198 bp tRNA fragment. Caution must be exercised in generalizing these
results inasmuch as this is only one data set, but parsimony analyses such
as these appear to constitute a powerful methodological tool for resolving
phylogenetic signal even in a small amount of data. Hayasaka et al.
(1988a:637) concluded that the high rate of homoplasy in mtDNA makes
the more distant relationships of the tree less reliable than those that are
more recent, yet parsimony analysis revealed a strong phylogenetic signal
across the entire tree.

Phylogenetic Inference Using Sequence Data: Some Comments on

Alternative Methods
There has been considerable discussion in recent years about which method
of phylogenetic inference is the most appropriate for DNA sequence data.
Efforts have been made to adjudicate this issue by simulation experiments
in which a data set is generated based on a "known true" phylogeny and
given assumptions as to how sequence evolution proceeds (Tateno et al.,
1982; Sourdis and Krimbas, 1987; Saitou and Imanishi, 1989). Different
methods are then applied to the data set to see if they recover the "true
216 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

phylogeny." These simulations are perhaps useful for identifying conditions

under which a given method may have limitations, but they are encumbered
with philosophical and empirical problems. Because certain knowledge of
phylogenetic relationships is beyond us, it is questionable whether some
artificially generated "truth" can serve as the arbiter of the best method
for recovering that which we cannot hope to know with certainty. The
evolutionary models underlying these simulations also suffer from a critical
methodological circularity: it is reasonable to assume that a detailed and
empirical understanding of the mechanisms of molecular change will not
be obtained unless it is based on prior hypotheses about the phylogenetic
relationships of organisms. Consequently, to incorporate assumptions about
the evolutionary process into a simulation procedure that will be used to
choose a method of phylogenetic inference may bias our recovery of those
relationships, which themselves will bias and confound our understanding
of processes of molecular evolution. Perhaps as important, these simulation
experiments generally take as their model of evolutionary change one that
most closely applies to noncoding, randomly evolving DNA. In practice,
however, the majority of molecular systematists are attempting to infer
relationships from sequences that will seldom, if ever, evolve under the
constraints specified by those models.
In real-world cases, we might expect that when there is a strong phy-
logenetic signal inherent in the data, virtually any method will recover it.
When data are extremely noisy, say as a result of too little data being used
to resolve very close internodal distances or of too much homoplasy, then
perhaps no method would be expected to resolve a reliable signal. Some-
where in between these extremes we might expect that different methods
will sometimes yield different phylogenetic hypotheses for a given set of
data. Choice among those hypotheses should not, however, be based on
an appeal to method, but on comparison and congruence with phylogenetic
inferences derived from other data.
At this time, three primary methods of phylogenetic inference are ap-
plied to sequence data; namely, parsimony analysis, distance analysis, and
maximum likelihood. Distance analysis has had both its defenders (Fel-
senstein, 1984; Nei, 1987) and critics (Farris, 1981, 1985). Tree-building
algorithms that work on distance matrices are necessary for certain kinds
of data, such as DNA-DNA hybridization or immunological distances, but
it is questionable whether they should be applied to discrete character data
such as sequences. At issue is not whether distance analysis of sequences
is able to reveal phylogenetic structure, for cleary it can. More to the point
is the inevitable loss of information in converting sequence data to distances
(Farris, 1981, 1985; Penny, 1982), their sensitivity to saturation effects so
that they must be corrected, the loss of flexibility to examine the deep
structure of the data underlying the results, and the inability of most dis-
tance algorithms to examine the structure of the large suite of trees having
near minimum-length fit. Taking the distance analysis of the primate data
PARSIMONY AND PHYLOGENETIC INFERENCE 217

set as an example, Hayasaka et al. (1988a) did not cluster using original
distances but corrected them for multiple substitutions (Gojobori et al.,
1982) and then clustered on those new distances. A single tree was produced
by the neighbor-joining method (Saitou and Nei, 1987) (Fig. 10-1) but
without an assessment of whether it was the best-fit tree. That tree, more-
over, was not compared quantitatively or qualitatively to other near-
minimum-length trees. Many different models have been proposed to "cor-
rect" for multiple substitutions. Yet, even though substitution probabilities
can be estimated from observed nucleotide frequencies, there is no guar-
antee that those probabilities actually reflect the dynamics of processes
that produced the observed nucleotide variation. That being the case, it
remains to be seen to what extent these different models that are being
applied to sequence data will influence inferred phylogenetic structure, not
only of the minimum-length tree, but of those trees close to that solution.
Maximum-likelihood methods are more developed in theory than in
practice (Felsenstein, 1981; Bishop and Friday, 1985; Saitou, 1988, 1990;
Fukami and Tateno, 1989), and we will not discuss them extensively here.
A choice among trees using maximum-likelihood methods depends upon
accepting an underlying model of evolutionary change. Often, that model
incorporates equal probabilities of substitutional change from one nucleo-
tide to another, but some have been developed to account for differences
in rate between transitions and trans versions (Saitou, 1990). Calculation
of likelihood functions for trees is computationally difficult, and the method
cannot be extended easily beyond cases for five or six taxa. It seems that
if more realistic, and thus more complicated, models of evolutionary change
are to be incorporated, computational difficulty is sure to increase.
Cedergren et al. (1988:102) note that parsimony methods have virtue in
phylogenetic inference using sequence data because they result "in the
most economical reconstruction of mutational history, with no assumptions
and with the minimum of coincidence and unobserved changes. . . . " Al-
though it is arguable whether parsimony is assumption free (all scientific
methods designed to discover pattern involve some assumptions), the
strengths of parsimony analysis over other methods of analyzing sequence
data are easy to enumerate. With parsimony, it is possible to use a minimum
of assumptions about the processes governing molecular evolution and still
obtain interpretable results. Weighting of characters or character-states
involves assumptions, but that too can be accomplished conservatively such
that narrowly constrained assumptions about molecular processes are avoided.
Today's parsimony algorithms are fast and efficient at searching among all
possible trees, thus guaranteeing that not only will the most-parsimonious
tree be found, but that the cladistic structure of near-most-parsimonious
trees can be examined. Moreover, with parsimony it is possible to examine
and evaluate the data supporting each cladistic component on a tree. For
these reasons alone, parsimony is the preferred method for analyzing DNA
sequences.
218 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

ACKNOWLEDGMENTS

We would like to thank T. E. Dowling, D. M. Hillis, T. Jones, D. Mindell,

and especially M. M. Miyamoto for their helpful comments on the man-
uscript. We thank E. Chipouras and M. M. Miyamoto for providing the
calculations for Table 10-3. This research was supported by the National
Science Foundation through grants BSR-8805957 and BSR-9007652.

REFERENCES

Altschul, S. F., R. J. Carroll, and D. J. Lipman. (1989) Weights for data related
by a tree. J. Mol. Biol. 207:647-653.
Archie, J. W. (1989) A randomization test for phylogenetic information in system-
atic data. Syst. Zool. 38:239-252.
Bishop, M. J., and A. E. Friday. (1985) Evolutionary trees from nucleic acid and
protein sequences. Proc. R. Soc. London 2268:271-302.
Brown, W. M. (1983) Evolution of mitochondrial DNA. Pp. 62-88 in Evolution
of Genes and Proteins (M. Nei and R. K. Koehn, eds.). Sinauer Associates,
Sunderland.
Brown, W. M. (1985) The mitochondrial genome of animals. Pp. 95-130 in Mo-
lecular Evolutionary Genetics (R. J. Maclntyre, ed.). Plenum Press, New
York.
Brown, W. M., E. M. Prager, A. Wang, and A. C. Wilson. (1982) Mitochondrial
DNA sequences of primates: tempo and mode of evolution. J. Mol. Evol.
18:225-239.
Carpenter, J. M. (1988) Choosing among multiple equally parsimonious clado-
grams. Cladistics 4:291-296.
Cedergren, R., M. W. Gray, Y. Abel, and D. Sankoff. (1988) The evolutionary
relationships among known life forms. /. Mol. Evol. 28:98-112.
Cracraft, J., and D. P. Mindell. (1989) The early history of modern birds: a com-
parison of molecular and morphological evidence. Pp. 389-403 in The Hi-
erarchy of Life. Molecules and Morphology in Phylogenetic Analysis (B.
Fernholm, K. Bremer, and H. Jornvall, eds.). Elsevier Science Publishers
B. V., Amsterdam.
Doolittle, R. F. (1990) (ed.). Molecular Evolution: Computer Analysis of Protein
and Nucleic Acid Sequences. Methods in Enzymology, vol. 183. Academic
Press, New York.
Farris, J. S. (1969) A successive approximations approach to character weighting.
Syst. Zool. 18:374-385.
Farris, J. S. (1972) Estimating phylogenetic trees from distance matrices. Am. Nat.
106:645-668.
Farris, J. S. (1981) Distance data in phylogenetic analysis. Pp. 2-23 in Advances
in Cladistics. Proceedings of the First Meeting of the Willi Hennig Society.
(V. A. Funk and D. R. Brooks, eds.). New York Botanical Garden, Bronx.
Farris, J. S. (1985) Distance data revisited. Cladistics 1:67-85.
Felsenstein, J. (1978) Cases in which parsimony or compatibility methods will be
positively misleading. Syst. Zool. 27:401-410.
Felsenstein, J. (1981) Evolutionary trees from DNA sequences: a maximum like-
lihood approach. J. Mol. Evol. 17:368-376.
PARSIMONY AND PHYLOGENETIC INFERENCE 219

Felsenstein, J. (1984) Distance methods for inferring phylogenies: a justification.

Evolution 38:16-24.
Felsenstein, J. (1985) Confidence limits on phylogenies: an approach using the
bootstrap. Evolution 39:783-791.
Felsenstein, J. (1988) Phylogenies from molecular sequences: inference and reli-
ability. Ann. Rev. Genet. 22:521-565.
Fitch, W. M. (1971) Toward defining the course of evolution: minimum change
for a specific tree topology. Syst. Zool. 20:406-416.
Fitch, W. M. (1977) On the problem of discovering the most parsimonious tree.
Am. Nat. 111:223-257.
Fitch, W. M., and E. Margoliash. (1967) Construction of phylogenetic trees. Science
155:279-284.
Fukami, K., and Y. Tateno. (1989) On the maximum likelihood method for es-
timating molecular trees: uniqueness of the likelihood point. /. Mol. Evol.
28:460-464.
Gojobori, T., K. Ishii, and M. Nei. (1982) Estimation of average number of nu-
cleotide substitutions when the rate of substitution varies with nucleotide.
/. Mol. Evol. 18:414-423.
Goodman, M. (1989) Emerging alliance of phylogenetic systematics and molecular
biology: a new age of exploration. Pp. 43-61 in The Hierarchy of Life.
Molecules and Morphology in Phylogenetic Analysis (B. Fernholm, K. Bre-
mer, and H. Jornvall, eds.). Elsevier Science Publishers B. V., Amsterdam.
Goodman, M., M. M. Miyamoto, and J. Czelusniak. (1987) Pattern and process
in vertebrate phylogeny revealed by coevolution of molecules and mor-
phology. Pp. 141-176 in Molecules and Morphology in Evolution: Conflict
or Compromise? (C. Patterson, ed.). Cambridge University Press, Cam-
bridge.
Hayasaka, K., T. Gojobori, and S. Horai. (1988a). Molecular phylogeny and
evolution of primate mitochondrial DNA. Mol. Biol. Evol. 5:626-644.
Hayasaka, K., S. Horai, T. Gojobori, T. Shotake, K. Nozawa, and E. Matsunaga.
(1988b) Phylogenetic relationships among Japanese, rhesus, Formosan, and
crab-eating monkeys, inferred from restriction-enzyme analysis of mito-
chondrial DNAs. Mol. Biol. Evol. 5:270-281.
Hendy, M. D., and D. Penny. (1982) Branch and bound algorithms to determine
minimal evolutionary trees. Math. Biosci. 59:277-290.
Irwin, D. M., T. D. Kocher, and A. C. Wilson. (1991) Evolution of the cytochrome
b gene of mammals. J. Mol. Evol. 32:128-144.
Kluge, A. G. (1989) A concern for evidence and a phylogenetic hypothesis of
relationships among Epicrates (Boidae, Serpentes). Syst. Zool. 38:7-25.
Kluge, A. G., and J. S. Farris. (1969) Quantitative phyletics and the evolution of
anurans. Syst. Zool. 18:1-32.
Margush, T., and F. R. McMorris. (1981) Consensus w-trees. Bull. Math. Biol.
43:239-244.
Miyamoto, M. M., and M. Goodman. (1990) DNA systematics and evolution of
primates. Ann. Rev. Ecol. Syst. 21:197-220.
Moritz, C., T. E. Dowling, and W. M. Brown. (1987) Evolution of animal mito-
chondrial DNA: relevance for population biology and systematics. Ann.
Rev. Ecol. Syst. 18:269-292.
Nei, M. (1987) Molecular Evolutionary Genetics. Columbia University Press, New
York.
220 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Penny, D. (1982) Towards a basis for classification: the incompleteness of distance

measures, incompatibility analysis and phenetic classification. J. Theor. Biol.
96:129-142.
Platnick, N . I . (1989) An empirical comparison of microcomputer parsimony pro-
grams. II. Cladistics 5:145-161.
Saitou, N. (1988) Property and efficiency of the maximum likelihood method for
molecular phylogeny. /. Mol. Evol. 27:261-273.
Saitou, N. (1990) Maximum likelihood methods. Pp. 584-598 in Molecular Evo-
lution: Computer Analysis of Protein and Nucleic Acid Sequences (R. F.
Doolittle, ed.). Methods in Enzymology, vol. 183. Academic Press, New
York.
Saitou, N., and T. Imanishi. (1989) Relative efficiencies of the Fitch-Margoliash,
maximum-parsimony, maximum-likelihood, minimum-evolution, and
neighbor-joining methods of phylogenetic tree construction in obtaining the
correct tree. Mol. Biol. Evol. 6:514-525.
Saitou, N., and M. Nei. (1987) The neighbor-joining method: a new method for
reconstructing phylogenetic trees. Mol. Biol. Evol. 4:406-425.
Sibley, C. G., and J. E. Ahlquist. (1987) Avian phylogeny reconstructed from
comparisons of the genetic material, DNA. Pp. 95-121 in Molecules and
Morphology in Evolution: Conflict or Compromise? (C. Patterson, ed.).
Cambridge University Press, Cambridge.
Smith, A. B. (1989) RNA sequence data in phylogenetic reconstruction: testing
the limits of its resolution. Cladistics 5:321-344.
Sourdis, J., and C. Krimbas. (1987) Accuracy of phylogenetic trees estimated from
DNA sequence data. Mol. Biol Evol. 4:159-166.
Swofford, D. L. (1990) PAUP: Phylogenetic Analysis Using Parsimony, Version
3.0. Illinois Natural History Survey, Champaign.
Swofford, D. L., and G. J. Olsen. (1990) Phylogeny reconstruction. Pp 411-501
in Molecular Systematics (D. M. Hillis and C. Moritz, eds.). Sinaurer As-
sociates, Sunderland.
Tateno, Y., M. Nei, and F. Tajima. (1982) Accuracy of estimated phylogenetic
trees from molecular data. I. Distantly related species. /. Mol. Evol. 18:387-
404.
Williams, P. L., and W. M. Fitch. (1989) Finding the minimal change in a given
tree. Pp. 453-470 in The Hierarchy of Life. Molecules and Morphology in
Phylogenetic Analysis (B. Fernholm, K. Bremer, and H. Jornvall, eds.).
Elsevier Science Publishers B.V., Amsterdam.
Williams, P. L., and W. M. Fitch. (1990) Phylogeny determination using dynam-
ically weighted parismony method. Pp. 615-626 in Molecular Evolution:
Computer Analysis of Protein and Nucleic Acid Sequences (R. F. Doolittle,
ed.). Methods in Enzymology, vol. 183. Academic Press, New York.
11
Evolutionary Analysis of Length-Variable
Sequences: Divergent Domains of
Ribosomal RNA

ALLAN LARSON

It is important in molecular phylogenetic studies to match the timescale of

the evolutionary divergence events being studied to the evolutionary rate
and pattern of the molecules used to resolve them. In many taxa, and
especially in vertebrates, divergence events occurring approximately 70 to
300 million years (Myr) before present have been particularly difficult to
resolve. Comparatively rapidly evolving molecules whose study has clari-
fied phylogenetic relationships on a more recent timescale (mitochondrial
DNA, serum albumin, various proteins compared electrophoretically) are
often too highly saturated with change to be useful for resolving divergence
events in this time range. The highly conserved sequences that have been
used to investigate much older divergences [cytochrome c, small-subunit
ribosomal RNA (rRNA)] usually do not offer sufficient variation to resolve
divergence events within the 70 to 300 Myr range. The "divergent domains"
of the large subunit of nuclear-encoded rRNA (Hassouna et al., 1984)
constitute a group of sequences whose evolutionary rate and pattern fa-
cilitate phylogenetic resolution in this time range (Qu et al., 1988; Larson
and Wilson, 1989; Larson, 1991), and also occasionally for more recent
phylogenetic divergences (Gonzalez et al., 1990). These are the most highly
variable segments of rRNA. Divergent domains may differ extensively both
in sequence and length among even closely related genera, but they occupy
homologous positions relative to the more conserved parts of the large
ribosomal subunit (Hassouna et al., 1984; Gerbi, 1985; Clark, 1987).
Neither the higher-order structures nor the cellular functions of the di-
vergent domains are well documented. Larson and Wilson (1989) found
indirect estimates of higher-older structures of divergent domains to be
highly conflicting, and the secondary structures of these sequences are

221
222 PHYLOCENETIC ANALYSIS OF DNA SEQUENCES

likely to be quite variable on an evolutionary timescale (Gonzalez et al.,

1985; Gorski et al., 1987). Divergent domains have been hypothesized to
be remnants of mobile elements that inserted into the ribosomal genes
(Clark et al., 1984), or alternatively, the remnants of linkers that connected
different "functional segments" during the evolutionary assembly of the
ribosome, and which were subsequently eliminated from all but the nuclear-
encoded eukaryotic ribosomes (Clark, 1987; see also Gonzalez et al., 1985;
Spencer et al., 1987). Larson and Wilson (1989) concluded, however, that
the evolutionary patterns of at least four of these domains are inconsistent
with the hypothesis that they are evolving free of functional constraint.
Because the ribosomal divergent domains are potentially of great im-
portance for molecular systematic studies, it is necessary to investigate in
detail the evolutionary pattern demonstrated by these sequences and to
evaluate whether this pattern is prone to support or to violate the require-
ments of commonly used molecular phylogenetic methods. The superpo-
sition of information contained in base substitutions with information from
length mutations is particularly important for analyzing these sequences.
In this chapter, I present an investigation of the evolutionary properties
of 11 of the 12 divergent domains of the large ribosomal subunit using
comparisons of salamanders and several closely-related outgroup taxa. A
schematic diagram of the positions of these divergent domains in the large
ribosomal subunit of the frog Xenopus laevis is shown in Figure 11-1. The
data used in this analysis (Fig. 11-2) represent a subset of the information
presented in my recent phylogenetic studies of salamanders (Larson and
Wilson, 1989; Larson, 1991), aligned with sequences published elsewhere
for Mus domesticus (Hassouna et al., 1984) and Xenopus laevis (Ware et
al., 1983).
This evaluation of evolutionary pattern requires partitioning the changes

Figure 11-1 Schematic diagram of the large ribosomal subunit of Xenopus laevis
showing positions of divergent domains 1-12 (solid lines). For domains D2 and D8,
only the segments marked with diagonal bars are included in this study.
EVOLUTIONARY ANALYSIS OF LENGTH-VARIABLE SEQUENCES 223

Figure 11-2 Aligned sequences for 11 divergent domains of the large subunit of
rRNA (after Larson, 1991). Samples include 11 salamanders (1-11), acaecilian (12),
a frog (13), a mammal (14), and a fish (15). Numbers at the tops of the columns are
the standard position numbers for the published Xenopus laevis (13) sequence (Ware
et al., 1983). Positions not present in the Xenopus sequence are denoted alphabet-
ically. Dots denote identity to the sequence in row 1; dashes denote gaps in the
aligned sequence. Divergent domains are marked above the aligned sequences
(D1—»|, D2—»|, etc., to D12). Vertical bars mark breaks in the reported sequence.
Base designations follow IUB recommendations: A, adenine; C, cytosine; C, guan-
ine; U, uracil; S, A or C; W, G or U; n, any or unknown. Lower case base letters
(a, g, u) denote uncertain positions in the published Xenopus sequence (Ware et
al., 1983). Large inserts in the aligned sequences are denoted by letters J, K, L, M,
224 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Figure 11-2 (continued).

N, and O and are identified below the appropriate segment. Alignments are tentative
for sites 2664-2688. Sequences reported for domains D2 and D8 are partial se-
quences (see Fig. 11-1) and the sequence reported for domain D5 includes a short
segment outside the 3' end of that domain (as identified by Hassouna et al., 1984).
(Figure continues.)
EVOLUTIONARY ANALYSIS OF LENGTH-VARIABLE SEQUENCES 225

Figure 11-2 (continued).

226 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Figure 11-2 (continued).

on a phylogenetic tree topology using parsimony (Fitch, 1971). The effects

of varying tree topology are evaluated to assess the inaccuracies that may
result from basing the study on an incorrect tree topology. The three
topologies investigated correspond to: (1) the topology favored by the
parsimony analysis of Larson (1991); (2) a tree representing an alternative
hypothesis of salamander relationships (Duellman and Trueb, 1986) but
retaining the same topology as tree 1 for outgroup taxa (the fish Scaphi-
rhynchus, the mammal Mus, the frog Xenopus, and the caecilian Typhlo-
EVOLUTIONARY ANALYSIS OF LENGTH-VARIABLE SEQUENCES 227

Figure 11-2 (continued).

228 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Figure 11-2 (continued).

EVOLUTIONARY ANALYSIS OF LENGTH-VARIABLE SEQUENCES 229

Figure 11-2 (continued).

nectes); and (3) a tree retaining the salamander relationships of tree 1 but
shuffling the relationships of outgroups (see Fig. 11-3).

MOLECULAR EVOLUTIONARY PATTERN AND PHYLOGENETIC

RECONSTRUCTION

The methods used to infer phylogenies from molecular sequence data are
reviewed by Felsenstein (1988). Despite repeated criticism, the most prev-
alent criterion used for constructing molecular phylogenetic trees is par-
simony, which favors the topology that minimizes the total amount of
homoplastic change in the characters being studied. A related criterion,
character compatibility, favors the topology specified by the largest group
of nonconflicting characters. Felsenstein (1988) notes that both of these
methods are subject to statistical inconsistency under certain conditions;
when these conditions occur, the parsimony and compatibility criteria will
favor an incorrect topology more strongly as the size of the data set in-
creases. Inequality of rates of molecular evolution among different lineages
coupled with high rates of molecular evolution will promote inconsistency.
What constitutes a high rate of change is relative to the divergence events
being investigated. A molecule that evolves too rapidly for investigating
one set of evolutionary branching events may be found to evolve conserv-
atively when evaluating more recent ones.
In the context of parsimony, compatibility, and alternative criteria, it is
important to investigate the following parameters of molecular evolution-
ary pattern: (1) variance of evolutionary rate among lineages; (2) distri-
bution of change among sites in the molecular sequence; (3) evidence for
saturation of substitution (multiple substitutions occurring at the same site
between lineages being compared); (4) ratio of transitions to transversions;
(5) evolutionary rate for base substitutions versus length mutations; (6)
overall evolutionary rate; and (7) base composition.
Variance of evolutionary rate among lineages is assessed using the index
230 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Figure 11-3 Three tree topologies used to investigate molecular evolutionary pat-
terns. Tree 1 is the topology favored by a parsimony analysis of rRNA sequence
variation by Larson (1991). Tree 2 is the topology presented by Duellman and Trueb
(1986) for salamander samples with the same configuration of outgroups used in
tree 1. Tree 3 is the same configuration of salamander samples used in tree 1 with
an alternative topology for outgroup taxa. Nodes are numbered (1 -13) for reference
to Table 11-2. For each branch, the inferred number of substitutions is given followed
by the inferred number of length mutations. (Figure continues.)

of dispersion (R = variance/mean rate; see Gillespie, 1986a) for all paired

sister lineages in the tree. If the molecular clock is a Poisson process, then
R = 1, although the values measured for the proteins used in molecular
clock studies are usually higher (approximately 2 to 3; Kimura, 1983), and
values exceeding 30 have been reported in more extensive surveys of pro-
tein and nucleic acid evolution (Gillespie, 1986b, 1989). Interpretations of
this outcome regarding the clocklike nature of molecular evolution are
highly controversial. Compare, for example, Kimura (1983) and Gillespie
(1986a, 1986b, 1989).
The density distribution of changes among sites in the molecular se-
quence is investigated using the negative binomial (Bliss and Fisher, 1953;
EVOLUTIONARY ANALYSIS OF LENGTH-VARIABLE SEQUENCES 231

Figure 11-3 (continued).

Uzzell and Corbin, 1971). The negative binomial distribution is defined by

the mean (m) number of changes per site (as measured on the phylogenetic
tree) and k = m2/(s2 — m) where s2 is the variance. Values of k are very
large if the distribution of substitutions is Poisson; when k is small, some
molecular sites are much more likely than others to undergo evolutionary
change. A similar parameter, alpha (= m2/s2) has been used for estimating
molecular similarity of homologous sequences from restriction map com-
parisons (Nei and Li, 1979).
Saturation of sites in molecular sequence comparisons can be investigated
using patterns in the ratio of transitions (substitutions replacing purines
with purines and pyrimidines with pyrimidines) to transversions (substi-
tutions replacing pyrimidines with purines and purines with pyrimidines).
Comparison of recently diverged molecular sequences demonstrates an
approximately twofold excess of transitions over transversions for cellular
genomes (Fitch, 1980; Jukes, 1980; Nichols et al., 1980), and the transition
bias is even larger for some organellar genomes (see review by Moritz et
al., 1987). As more distantly related (i.e., more extensively saturated)
comparisons are made, the measured transition/transversion ratio drops
232 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Figure 11-3 (continued).

(DeSalle et al., 1987). The expected equilibrium ratio depends upon the
base composition of the molecular sequence and can be evaluated using
the equations of Holmquist (1983). Comparisons of observed and expected
transition/transversion ratios are used to evaluate the degree of saturation
of ribosomal divergent domain sequences among species separated by ap-
proximately 70 to 200 Myr.
Evolutionary rates of the divergent domains are estimated using com-
parisons for which approximate evolutionary divergence dates are available
from paleontological data and previous molecular studies. The paleontol-
ogy of salamanders is reviewed by Estes (1981), and its relevance to the
molecular phylogeny of salamanders is discussed by Larson (1991). Fossil
evidence suggests that the separation ofAmbystoma and Dicamptodon was
complete by 65 Myr before present. Maxson and Wilson (1979) placed the
split between Ambystoma and the amphiumid and plethodontid lineages
(represented here by Amphiuma, Aneides, and Desmognathus samples) at
200 Myr ago based upon comparative immunological results. According
to the molecular phylogeny of Larson (1991), this comparison spans the
oldest divergence event in the history of the extant salamander families.
EVOLUTIONARY ANALYSIS OF LENGTH-VARIABLE SEQUENCES 233

Relative rates of substitutions and length mutations are evaluated without

exact dates by comparing the numbers of changes inferred for each category.
Where the rate assumptions of parsimony methods are violated, alter-
native approaches that are rate-invariant can be used. The invariant method
of Lake (1987) is designed for use in sequence comparisons where rates
are likely to be unequal among lineages and where saturation is a problem.
Its application requires larger amounts of evolutionary change than are
available in the data set on which this study is based, and it is less likely
than parsimony to be informative under conditions where parsimony anal-
ysis is appropriate (Prager and Wilson, 1988). Felsenstein (1988) notes that
Lake's invariant method contains the assumption that the alternative out-
comes of transversion events occur with equal probability (that G to C
changes and G to U changes are equally likely, etc.) and that the method
is subject to error when this assumption is violated. This assumption can
be tested using the data for ribosomal divergent domains evaluated here,
and it is relevant for the possible use of the divergent domains for evaluating
evolutionary divergences older than the ones considered here.
Compositional statistics is a rate-invariant method presented by Sidow
and Wilson (1990) as an alternative to Lake's evolutionary parsimony. It
shares with evolutionary parsimony the potential to resolve deep branches
in the tree of life using sequence comparisons that demonstrate substitu-
tional saturation and variable rates of evolution. Unlike evolutionary par-
simony, however, it utilizes information on base composition and does not
assume balanced transversions. The evolution of base composition within
the divergent domains is analyzed here with respect to the application of
compositional statistics.

RESULTS

The boundaries of the divergent domains used in this analysis correspond

to those identified by Hassouna et al. (1984) (see Fig. 11-1). A short region
located immediately 3' to domain 5 shows extensive length variation in
salamanders and is included with domain 5 for this analysis. Molecular
changes are partitioned on the alternative tree topologies using the com-
puter program, Phylogenetic Analysis Using Parsimony (PAUP) (Swof-
ford, 1984). The types of molecular evolutionary change inferred for each
topology in Figure 11-3 are summarized in Table 11-1.

Indices of Dispersion

Indices of dispersion are presented for each of the 13 nodes of each tree
depicted in Figure 11-3 (Table 11-2). Because of the relatively small num-
bers of changes recorded within the individual domain segments, R values
are calculated for all divergent domains together, including base substi-
234 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Table 11-1 Summary of Evolutionary Changes for Each Divergent Domain*

Domaint T V I Del L Total N Total/N
1 (122a-274) 39 17 15 11 1 83 160 0.5
2 (821-950) 44 23 6 14 0 87 121 0.7
3 (992-1170) 24 23 4 4 1 56 132 0.4
4 (1358a-1370) 2 7 6 5 1 21 20 1.1
5 (1449-1500) 5 8 13 13 3 42 53 0.8
6 (1727-1775) 13 5 3 3 3 27 50 0.5
7a (1947-2007) 12 5 0 2 0 19 61 0.3
7b (2043-2122) 9 5 0 0 0 14 36 0.4
8 (2596-2777) 32 32 17 19 1 101 152 0.7
9 (3137-3164) 2 8 0 2 2 14 28 0.5
10 (3204-3291) 10 8 0 0 0 18 73 02
12 (3857-4023) 55 52 8 16 2 133 155 09
Total Tree 1 247 193 72 89 14 615 1,041 0.6
Total Tree 2 283 212 81 83 14 673 1,041 0.6
Total Tree 3 234 192 89 79 16 610 1,041 0.6
*Changes are inferred from the lineages on tree topology 1 (Fig 11-3) for each of the divergent domains
Total changes (all domains summed) are given for tree topologies 1-3 (Fig. 11-3) at the bottom of the
table.
Abbreviations used are: T, transitions, V, transversions, I, insertions, Del, deletions, L, unpolarized length
changes; Total, sum of mutational events; N, number of bases per segment.
'Domains are numbered as recognized by Hassouna et al. (1984). Numbers in parentheses denote the sites
scored (as numbered in Fig. 11-2). Only partial sequences are given for domains 2 and 8 (see Fig. 11-1)
and sequences reported for domain 5 include a short segment located 3' to this domain as recognized by
Hassouna et al. (1984).

tutions and length mutations. For each tree, R ranges from values much
lower to values much larger than the value of 1 expected if molecular
change through time is Poisson distributed. On tree 1 (the topology favored
by Larson, 1991), the median value (1.5) is similar to observations for
protein sequences on which molecular clock studies are based (Wilson et
al., 1977; Kimura, 1983). The mean value (3.65 ± its standard error of
1.16) exceeds the median because a few nodes have values that are much
higher than the others. Use of the alternative tree topologies increases the
median value of R (2.6 for tree 2, and 2.3 for tree 3).
Indices of dispersion greater than 1 indicate some degree of clumping
of changes on the phylogenetic tree. There are only a few examples where
particular taxa are associated consistently with rate inequalities on all three
topologies tested. The terminal lineages leading to Amphiuma, Necturus,

Table 11-2 Indices of Dispersion (R) Representing the Variance/Mean Ratios of

the Inferred Numbers of Changes Occurring on Sister Lineages of Each of the Tree
Topologies Shown in Figure 11-3 and the Median (M)*
Node
1 2 3 4 5 6 7 8 9 10 11 12 13 M
Tree 1 1.5 0.3 0.0 0.1 3.9 6.9 10.1 1.6 10.2 1.4 9.5 1.5 0.5 1.5
Tree 2 1.0 0.0 3.4 2.8 2.6 10.7 2.2 3.7 1.6 0.1 13.6 0.0 3.2 2.6
Tree 3 1.5 0.0 1.1 6.3 21.6 6.0 0.1 7.6 9.5 2.3 9.8 0.2 1.7 2.3
'Nodes are as numbered on Figure 11-3.
EVOLUTIONARY ANALYSIS OF LENGTH-VARIABLE SEQUENCES 235

Rhyacotriton, and Typhlonectes are short relative to their sister lineages

on all three trees. The lineage leading to Siren is anomalously long on trees
1 and 3, but nearly identical in length to its sister lineage on tree 2. A
partitioning of changes for each divergent domain individually on tree 1
(not shown) reveals that most domains have changes distributed throughout
the tree rather than being anomalously clumped on one or a few lineages.
There are, however, a few apparent exceptions to this generalization. Five
of the 11 domains examined (1, 5, 6, 8, and 12) have a somewhat higher
than expected total amount of change on the lineage leading to Siren.
Domains 1 and 4 show a somewhat high concentration of change on the
lineage immediately ancestral to all salamanders. Domain 12 also shows a
higher than expected amount of change on the Notophthalmus lineage,
and for domain 9 the majority of the observed changes are on the Xenopus
lineage.

Density Distribution of Substitutions and Length Mutations

For each divergent domain, the negative binomial parameters (m, k, and
alpha) are evaluated separately for base substitutions (Table 11-3), length
mutations (Table 11-4), and for all changes together (Table 11-5). For
making these calculations, a "site" is defined as a column of single-base
width in the block of aligned sequences (although the individual sequences
in the alignment will not always have a base present at each site; see Fig.
11-2). Multiple-base deletions are counted separately for each site covered
in the sequence alignment. Multiple-base insertions are counted as single
events and the position of the insertion is treated as a single site for this
analysis. The total number of sites counted for evaluating base substitutions
alone (Table 11-3) does not include "sites" produced by insertions unique
to single sequences and therefore differs from the totals used to evaluate
length mutations (Table 11-4) and total changes (Table 11-5).
For tree 1, the overall mean number of base substitutions per molecular
site (m) is 0.44. Estimates for tree 2 are generally higher (mean = 0.49)
reflecting the fact that it is less parsimonious overall. Estimates derived
from tree 3, which rearranges only the outgroup taxa, are closer overall
to values obtained with tree 1 (mean = 0.42). The mean number of length
changes per molecular site (Table 11-4) is less than the mean number of
substitutions by a factor of approximately 1.5. Overall, the mean number
of changes per site (substitutions and length mutations) is 0.68 for tree 1,
with the other trees differing only slightly. Mean values estimated for m,
k, and alpha (Tables 11-3-11-5) differ by no more than a factor of 1.5
when different tree topologies are used. For length mutations, large dif-
ferences in the estimates of these parameters are seen for some individual
domains when tree topology 3 is compared to topologies 1 and 2 (see, for
example, domains 6 and 9 in Table 11-4). This results mainly from differ-
ences in partitioning a few large length mutations occurring in some of the
smaller domains among the caecilian, frog, salamander, and mammalian
Table 11-3 Density Distribution of Base Substitutions Per Site Within Divergent Domains of the Large Ribosomal Subunit Estimated
Separately Using Tree Topologies 1, 2, and 3 (Fig. 11-3)
Domain*
1 2 3 4 5 6 7a 7b 8 9 10 12 Meanf
Total sites 152 119 128 12 52 47 61 36 147 28 73 151 83.83
Mean (1) 0.37 0.56 0.37 0.75 0.25 0.38 0.28 0.39 0.44 0.36 0.25 0.71 0.44
(2) 0.39 0.61 0.43 0.67 0.31 0.38 0.34 0.44 0.56 0.36 0.23 0.78 0.49
(3) 0.34 0.55 0.38 0.75 0.21 0.36 0.26 0.36 0.43 0.36 0.26 0.66 0.42
k (1) 0.40 0.94 0.76 0.93 0.48 0.76 0.36 0.52 0.39 — 0.25 1.18 0.63
(2) 0.31 0.73 0.32 1.14 0.33 0.76 0.21 0.38 0.29 — 0.28 0.96 0.47
(3) 0.42 0.93 0.53 0.93 0.41 0.62 0.54 0.64 0.34 — 0.23 1.39 0.63
Alpha (1) 0.19 0.35 0.25 0.42 0.16 0.25 0.16 0.22 0.20 0.42 0.12 0.44 0.26
(2) 0.17 0.33 0.18 0.42 0.16 0.25 0.13 0.20 0.19 0.42 0.13 0.43 0.24
(3) 0.19 0.35 0.22 0.42 0.14 0.21 0.18 0.23 0.19 0.42 0.12 0.45 0.26
'Calculations for domain 2 omit one region of the Mas sequence in Figure 11-2 that is not alignable with the others. Values of k are omitted for domain 9 because very
low values of the variance to the mean produce negative values of k.
•Values of parameters for the individual domains are weighted by the length of the domain to obtain mean values.
Table 11-4 Density Distribution of Sites Covered by Length Mutations Within Divergent Domains of the Large Ribosomal Subunit
Estimated Separately Using Tree Topologies 1, 2, and 3 (Fig. 11-3)*

Domaint
1 2 3 4 5 6 7a 7b 8 9 10 12 Mean*

Total sites, 160 121 132 20 53 50 61 — 152 28 — 155 93.20

Mean (1) 0.18 0.24 0.12 0.95 0.64 0.22 0.10 — 0.39 1.29 — 0.19 0.29
(2) 0.21 0.22 0.12 0.95 0.72 0.22 0.10 — 0.39 1.29 — 0.19 0.29
(3) 0.20 0.24 0.14 0.65 0.79 0.14 0.10 — 0.41 0.07 — 0.21 0.26
k (1) 0.10 0.22 0.48 — 0.57 1.53 0.42 — — 0.26 0.38
—
— —
— —
(2) 0.09 0.43 0.48 — 0.45 1.53 0.45 — 0.26 0.41
(3) 0.10 0.22 0.26 — 0.61 0.20 — — 0.47 — — 0.24 0.28
Alpha (1) 0.06 0.12 0.10 1.39 0.30 0.19 0.11 — 0.20 1.80 — 0.11 0.21
(2) 0.06 0.15 0.10 1.39 0.28 0.19 0.11 — 0.21 1.80 — 0.11 0.21
(3) 0.07 0.12 0.09 0.80 0.34 0.08 0.11 — 0.22 0.08 — 0.11 0.14
*No Length Mutations Were Recorded in Domains 7b and 10.
tCalculations for domain 2 omit one region of the Mas sequence in Figure 11-2 that is not au'gnable with the others. Values of fc are omitted for domains 4, 7a, and 9
because very low values of the variance to the mean produce negative values of k.
'Values of parameters for the individual domains are weighted by the length of the domain to obtain mean values.
Table 11-5 Density Distribution of Changes (Length Mutations and Base Substitutions) Within the Divergent Domains of the Large
Ribosomal Subunit Estimated Separately Using Tree Topologies 1, 2, and 3 (Fig. 11-3)
Domain*
1 2 3 4 5 6 7a 7b 8 9 10 12 Meant
Total sites 160 121 132 20 53 50 61 36 152 28 73 155 86.75
Mean (1) 0.53 0.78 0.48 1.40 0.89 0.54 0.38 0.39 0.82 1.68 0.25 0.88 0.68
(2) 0.58 0.82 0.54 1.35 1.02 0.54 0.44 0.44 0.92 1.68 0.23 0.95 0.73
(3) 0.52 0.79 0.51 1.10 0.98 0.48 0.36 0.36 0.83 0.43 0.26 0.85 0.64
k (1) 0.48 1.38 0.94 8.17 0.98 1.55 0.57 0.52 0.66 — 0.25 1.79 1.10
(2) 0.36 1.33 0.47 10.27 0.75 1.55 0.33 0.38 0.49 — 0.28 1.47 0.95
(3) 0.19 1.45 0.71 1.75 1.00 0.92 0.79 0.64 0.65 0.47 0.23 2.34 0.96
Alpha (1) 0.25 0.50 0.32 1.20 0.47 0.40 0.23 0.22 0.37 3.02 0.12 0.59 0.46
(2) 0.22 0.51 0.25 1.19 0.43 0.40 0.19 0.20 0.32 3.02 0.13 0.58 0.43
(3) 0.14 0.51 0.30 0.68 0.50 0.32 0.25 0.23 0.37 0.23 0.12 0.62 0.36
'Calculations for domain 2 omit one region of the Mas sequence in Figure 11-2 that is not alignable with the others. Some values of k are omitted for domain 9 because
very low values of the variance to the mean produce negative values of k
'Values of parameters for the individual domains are weighted by the length of the domain to obtain mean values.
EVOLUTIONARY ANALYSIS OF LENGTH-VARIABLE SEQUENCES 239

lineages. This has a relatively small impact, however, on the estimated neg-
ative binomial parameters for all domains combined (Tables 11-4 and 11-5).
The values of k and alpha estimated from all three trees are too small
for the density distribution of changes (for both site substitutions and length
mutations) to be Poisson. Some sites are more likely than others to evolve.
This replicates the observation of Larson and Wilson (1989) that some
segments of sequence within the divergent domains are relatively highly
conserved, whereas others undergo extensive change.

Transition/Transversion Ratios

Estimates of the ratio of transitions to transversions are given for the 11

divergent domains in Table 11-6. The results obtained using different tree
topologies are within the same order of magnitude, generally differing by
not more than a factor or two. Four domains (1, 2, 6, and 7a/7b) have
values of transitions/transversions very close to 2, consistent with prior
empirical comparisons of nuclear-encoded sequences where site saturation
has not occurred. Four additional domains (3, 8, 10, and 12) have values
closer to 1. All of these are well above the expected equilibrium value of
approximately 0.4 calculated using the method of Holmquist (1983). Three
small domains (4, 5, and 9) have values that approach the expected ratios
for saturated sequences. The transversions observed in these three small
domains occur mainly in comparisons to the Mus sequence, however, which
is relatively distantly related to the amphibians and therefore more likely
to demonstrate substitutional saturation. Additional sampling of the mam-
malian lineage may clarify the substitutional changes that separate the
amphibian and mouse sequences. With this exception, it appears that mo-
lecular sites in the divergent domains do not demonstrate substitutional
saturation within the scope of this study.

Bias in Transversion Substitutions

Polarized transversion substitutions were tallied for the divergent domains

and adjacent regions reported by Larson (1991) for tree topology 1 (Fig.
11-3). For 16 observed transversions replacing an A base, 12 replaced it
with C and only four replaced it with U. For 76 transversions replacing a
G base, 51 replaced it with C and 25 replaced it with U. For 29 transversions
replacing a U base, 23 replaced it with G and only six replaced it with A.
Using a x2 goodness-of-fit test, all three of these cases are significantly
different from the prediction of balanced transversions, at least at the 0.05
level. Only the transversions replacing the C base occur with approximately
equal frequencies; of 61 changes, 32 replace it with G and 29 replace it
with A. Overall, there is a more than twofold preference for replacing a
purine with C rather than U, and an almost twofold preference for replacing
a pyrimidine with G rather than with A.
Table 11-6 Ratios of Transition/Transversion Substitutions Measured According to Three Alternative Tree Topologies (Fig. 11-3),
Compared With the Expected Equilibrium Ratio (Calculated According to Holmquist, 1983)
Domain
1 2 3 4 5 6 7a 7b 8 9 10 12 Mean'
Observed (1) 2.3 1.9 1.0 0.3 0.6 2.6 2.4 1.8 1.0 0.3 1.3 1.1 1.4
(2) 2.2 2.1 1.2 0.1 0.9 3.3 3.2 1.3 1.0 0.3 1.4 1.1 1.5
(3) 2.0 1.8 1.1 0.3 0.6 2.8 2.2 1.6 0.9 0.3 1.1 1.0 1.4
Expected 0.4 0.3 0.4 0.5 0.4 0.4 0.4 0.5 0.4 0.2 0.3 0.4 0.4
Substitutions*
(1) 56 67 47 9 13 18 17 14 64 10 18 107 36.67
(2) 60 74 55 8 17 17 21 16 82 10 17 118 41.25
(3) 51 66 49 9 11 19 16 13 63 10 19 100 35.50
'Mean values of the transition/transversion ratio are calculated by weighting the values for individual domains by the inferred number of base substitutions.
'The number of base substitutions inferred for each divergent domain for each of the tree topologies (Fig. 11-3).
EVOLUTIONARY ANALYSIS OF LENGTH-VARIABLE SEQUENCES 241

Rates of Change

Approximately 1,040 bases of sequence are tallied for the divergent do-
mains, although this is not a complete assessment of the divergent domain
sequences in rRNA. Calculation of the rate of divergence between lineages
(counting changes accumulated on both lineages descended from a node)
per molecular site per million years ago (Mya) is done using 65 Mya as
the estimated separation time of Ambystoma and Dicamptodon based on
the fossil record (see review by Estes, 1981). Fifteen total changes are
observed on these lineages, giving an estimated divergence rate of 0.02%
change/site/Mya for the pair of lineages. The 65 Mya divergence is a min-
imum estimate of divergence time, making the rate estimate a maximum
one. Within the resolution of the phylogenetic study of Larson (1991), the
divergence of the amphiumids, plethondontids, and remaining extant sal-
amanders constitutes an unresolvable three-way split. The divergence of
the amphiumids and plethodontids from Ambystoma is estimated by Max-
son and Wilson (1979) to be 200 Mya. Using this as a date for the hy-
po thethical three-way split, the following rates of change are estimated:
(1) amphiumids (Amphiuma) to plethodontids (Aneides and Desmogna-
thus) (0.0002 changes/site/Mya); (2) amphiumids to ambystomatids (Am-
bystoma and Dicamptodon) (0.0002 changes/site/Mya); (3) plethodontids
to ambystomatids (0.0003 changes/site/Mya). In nearly all cases, the changes
measured here are single base-length mutations or substitutions, thereby
representing approximately 0.02% divergence/site/Mya. This is comparable
to the value estimated by Larson and Wilson (1989) for evolutionary change
in the divergent domains for salamanders. Larson and Wilson (1989) in-
terpreted these values, which were based on comparisons within pletho-
dontids, to be underestimates for salamanders as a whole, but the results
given here indicate that they are probably accurate. Values calculated by
Larson and Wilson (1989) and Schmickel et al. (1990) for mammals are
approximately eight to ten times larger. A possible acceleration of evo-
lutionary rate in mammals may help to explain why the saturation inferred
from three divergent domains discussed above (based on transition/trans-
version ratios) occurs mainly in comparisons to the Mus sequence.
The ratio of base substitutions to length mutations for the divergent
domains (based on tree topology 1) is approximately 2.5:1 (440/175 mu-
tations). This difference is judged not to be great enough to warrant sep-
arate weighting of the two kinds of events for phylogenetic reconstruction.

Evolution of Base Composition

For applying the rate invariant phylogenetic method of Sidow and Wilson
(1990), and for estimating the expected equilibrium ratio of transitions/
transversions for a molecular sequence comparison (Table 11-6), it is nec-
essary to measure base compositions. Base compositions are determined
for the 11 divergent domains for samples from the caecilian, fish, frog,
salamander, and mammal lineages (Table 11-7). The mean G + C content
Table 11-7 Nucleotide Base Compositions of Divergent Domains*
Domain
Base 1 2 3 4 5 6 7a 7b 8 9 10 12
Salamander A 13 10 14 33 14 13 17 19 8 0 7 14
C 37 42 32 33 32 33 30 28 38 50 33 32
G 36 37 44 25 43 46 40 42 38 40 46 36
U 14 10 10 8 11 9 13 11 16 10 14 18
Caecilian A 9 8 15 25 12 11 17 19 8 0 8 16
C 37 42 32 25 26 30 31 28 43 50 35 33
G 40 40 43 25 45 48 38 42 35 40 46 37
U 14 9 10 25 17 11 15 11 15 10 11 14
Frog A 9 11 14 8 16 8 15 17 6 11 11 14
C 40 45 31 33 29 31 31 31 44 46 35 35
G 41 36 44 33 41 47 41 44 38 36 44 38
U 11 8 11 25 14 14 13 8 12 7 9 13
Mammal A 10 8 15 15 13 9 16 22 8 4 8 16
C 37 41 32 40 28 34 29 25 43 56 35 31
G 39 40 45 25 45 48 41 42 37 37 47 37
U 13 11 8 20 13 9 14 11 12 4 9 16
Fish A 14 8 16 10 13 18 20 22 11 0 13 16
C 32 38 30 40 30 36 30 28 35 50 32 28
G 36 40 42 20 43 40 34 39 37 40 43 31
U 18 14 13 30 15 7 16 11 18 10 13 25
Mean A 11.0 9.0 14.8 18.2 13.6 11.8 17.0 19.8 8.2 3.0 9.4 15.2
C 36.6 41.6 31.4 34.2 29.0 32.8 30.2 28.0 40.6 50.4 34.0 31.8
G 38.4 38.6 43.6 25.6 43.4 45.8 38.8 41.8 37.0 38.6 45.2 35.8
U 14.0 10.4 10.5 21.6 14.0 10.0 14.2 10.4 14.6 8.2 11.2 17.2
Standard deviation A 2.3 1.4 0.8 4.2 1.5 4.0 1.9 2.2 1.8 4.8 2.5 1.1
C 2.9 2.5 0.9 6.2 2.2 2.4 0.8 2.1 3.9 3.6 1.4 2.6
G 2.3 1.9 1.1 4.7 1.7 3.3 2.9 1.8 1.2 1.9 1.6 2.8
U 2.5 2.3 1.8 8,4 2.2 2.6 1.3 1.3 2.6 2.7 2.2 4.8
*A11 values are given as percentages. Values are based on the sequences presented for Aneides flavipunctatus (salamander), Typhionectes compressicauda (caecilian),
Xenopus laevis (frog), Mus domesticus (mammal), and Scaphirhynchus platorynchus (fish).
EVOLUTIONARY ANALYSIS OF LENGTH-VARIABLE SEQUENCES 243

of the divergent domains ranges from approximately 60 to 90%. Three

small domains (4, 6, and 9) have an atypically large variance of base
composition.

DISCUSSION

Assessment of molecular evolutionary pattern is important for deciding

which analytical method will extract the maximum amount of phylogenetic
information from comparative molecular sequence data without system-
atically incurring errors. The divergent domains of the large ribosomal
subunit are perhaps unusual in the degree to which base-substitutional and
length-mutational changes are superimposed. These sequences are very
important for phylogenetic studies, however, because of their potential for
resolving evolutionary branching events on an intermediate timescale and
because they can be studied in all eukaryotic taxa. The results of this study
are important not only for interpreting direct sequence comparisons in-
volving the divergent domains, but also for evaluating phylogenetic studies
based upon restriction mapping of nuclear ribosomal genes. Conclusions
regarding the utility of different methods of phylogenetic analysis are ex-
pected to be of general significance for studies of ribosomal evolution.
The evolutionary pattern revealed by the divergent domains does not
meet the assumptions of parsimony analysis exactly, but the approximation
is judged to be close enough that statistical tests based upon parsimony
should be robust for the evolutionary timescale being investigated. The
molecular clock provides a good expectation for the distribution of mo-
lecular evolutionary changes on the phylogenetic tree. It is anticipated,
however, that a few lineages will show substantially larger or smaller amounts
of change than predicted by the molecular clock in studies of rRNA. For
a strictly stochastic clock, the index of dispersion (R) should equal 1,
whereas the median value measured for the divergent domains in this study
is 1.5. Some parts of the tree have a much smaller variance/mean ratio,
but others are anomalously high. The use of incorrect tree topologies will
inflate estimates of/?, thereby overestimating the deviations from clocklike
evolution.
When rate variation produces inconsistency of the parsimony criterion,
lineages demonstrating unusually large amounts of change will tend to
cluster together, thereby obscuring the true relationships of the species
being studied. Larson (1991) notes that the lineages demonstrating rela-
tively large amounts of change in this data set are not grouped by parsi-
mony. Although the observed rate asymmetry appears not to produce
inconsistency of the parsimony criterion on the timescale investigated, it
cautions against the use of parsimony analysis of divergent domains to
investigate divergences much older than these. The observed rate asym-
metry also cautions against making phylogenetic inferences from un-
weighted pair group clustering of taxa (UPGMA; see Felsenstein, 1988),
244 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

which is more sensitive than parsimony to rate asymmetry. The UPGMA

analysis is discouraged, furthermore, by the difficulty of converting to
distance values comparisons of paired sequences that differ by both length
mutations and site substitutions. The presence of length mutations that
affect more than a single site has the consequence that not all evolutionary
events produce equivalent amounts of divergence among paired sequences.
The observation that the overall mean number of changes inferred per
site is substantially less than one (Tables 11-3-11-5) supports the conclusion
that the evolutionary saturation of sites, which also can produce inconsis-
tency of parsimony analysis, is not a problem here. This conclusion is
compromised somewhat by the observation that the density distribution of
changes along the molecular sequence is highly nonrandom. For example,
a few small domains average more than one change per site (Tables
11-4-11-5). Furthermore, values of the negative binomial parameter, k,
and the related parameter, alpha (Tables 11-3-11-5) show that evolution-
ary change in the divergent domains is not Poisson distributed, but that
some sites are much more likely than others to undergo evolutionary change.
If changes along the molecular sequence were Poisson distributed, evo-
lutionary saturation of sites would occur less rapidly than it does when
substitutions are concentrated in a subset of the sites. This problem can
be avoided somewhat by using as taxonomic characters only the more
conservatively evolving sites in the alignment. Alternatively, rate-invariant
approaches (Lake, 1987; Sidow and Wilson, 1990) could be used for making
phylogenetic inferences from heavily substituted sequences, although these
methods are applicable only to sites that have not undergone length
mutation.
The magnitude of alpha is important for estimating sequence divergence
or distance from restriction map data (Nei and Li, 1979). Values obtained
for protein sequences (1-2) are too high for use with divergent domains.
This would overestimate the total divergence among paired sequences by
inferring that evolutionary changes are more evenly distributed among sites
than they are. The values reported here (0.46 for all mutations, 0.26 for
base substitutions alone) are more compatible with the values of 0.45-0.50
used for DNA sequences by Carr et al. (1987) and Wilson et al. (1989).
Even these are too high, however, for some divergent domains.
The observed ratio of transitions to transversions in sequence compar-
isons is also useful for assessing substitutional saturation of molecular sites.
Except for three very small domains (4, 5, and 9), transition/transversion
ratios exceed those expected for saturated sequences and approximate the
value of 2, characteristic of recently diverged molecular sequences com-
pared in earlier empirical studies (Table 11-6). These studies of molecular
evolutionary pattern suggest that the divergent domains meet the criterion
of parsimony that sequences are not heavily saturated. This result suggests
also that the rate-invariant methods developed for analysis of highly sat-
urated sequences (Lake, 1987; Sidow and Wilson, 1990) are less likely than
EVOLUTIONARY ANALYSIS OF LENGTH-VARIABLE SEQUENCES 245

parsimony to be informative for phylogenetic analysis of ribosomal diver-

gent domains on the timescale addressed here.
To resolve divergence events much older than the ones addressed in this
study using the ribosomal divergent domains requires a method that ac-
commodates rate variation, substitutional saturation of sites, and length
mutations. Lake's invariant method (Lake, 1987) meets the first two re-
quirements but not the third, and its assumption of balanced transversions
is violated by the pattern observed here for the divergent domains. The
latter problem is remedied by the compositional statistics of Sidow and
Wilson (1990). Neither of these methods, however, is able to utilize in-
formation from the length-variable sites for phylogenetic inference.
Further theoretical work is needed to make more precise quantitative
statements regarding the degree to which violations of the assumptions of
parsimony will promote errors in phylogenetic inference. For evaluating
the usefulness of the parsimony criterion, it is particularly important to
measure the effects on phylogenetic reconstruction of deviations from the
following optimal conditions: (1) the median index of dispersion (R) should
equal 1; (2), the mean density of substitutions per molecular site should
not exceed 1; (3) the negative binomial parameter k should be large; and
(4) the ratio of transition to transversion substitutions should be approx-
imately 2:1. It is suggested that future studies address the interactions of
these factors with respect to consistency of the parsimony criterion.

SUMMARY

The molecular evolutionary patterns of 11 divergent domains of the large

ribosomal subunit are assessed to evaluate the appropriateness of these
sequences for phylogenetic analyses using parsimony and other criteria.
The taxa studied represent nine families of salamanders covering approx-
imately 70 to 200 million years of divergence, and four vertebrate out-
groups. The rate of evolution of the divergent domain sequences is esti-
mated to produce approximately 0.02% sequence divergence between
lineages per site per Mya. Consistency of the parsimony criterion for phy-
logenetic inference requires that rates of molecular evolution be stochast-
ically constant. Variance of evolutionary rates is assessed using the index
of dispersion (R = variance/mean rate of change along branches of the
phylogenetic tree). The median value of R is 1.5, which indicates a higher
than stochastic variance (for which R = 1.0), but which is well within the
range used in earlier molecular clock studies, in which R may be as large
as 2 or 3.
Parsimony analysis requires also that the sequences being compared are
not heavily saturated with multiple substitutions at the same sites. The
average total number of changes per molecular site is 0.68 for the ribosomal
divergent domains. Substitutional saturation therefore appears not to be
246 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

a problem at the evolutionary timescale studied, although the distribution

of changes along the molecular sequence is rionrandom, thereby increasing
the likelihood of substitutional saturation at the most rapidly evolving sites.
The average transition/transversion ratio for the divergent domains is 1.9:1,
approximating the value observed empirically for comparisons of relatively
recently diverged molecular sequences (2.0) and substantially exceeding
the value predicted for highly saturated sequences (0.4).
Parsimony analysis of the ribosomal divergent domains appears to be
justified for investigating phylogenetic divergence events occurring ap-
proximately 70 to 200 Mya. Rate-invariant methods are preferable for
investigating much older divergences. The method of compositional sta-
tistics is recommended because it is able to accommodate an observed bias
in transversion substitutions. To extract the maximum amount of phylo-
genetic information from these sequence comparisons, however, the su-
perposition of base substitutions and length mutations also must be
addressed.

ACKNOWLEDGMENTS

I thank Norma Gieg-Sinclair for collecting data and Chris DeHaan for
making calculations and figures. Joel Cracraft, Michael M. Miyamoto,
Ellen Prager, Elizabeth Zimmer, and an anonymous reviewer provided
helpful suggestions and discussion. Financial support was provided by NSF
grant BSR-8708393 and BRSG grants from Washington University.

REFERENCES

Bliss, C. I., and R. A. Fisher (1953) Fitting the negative binomial distribution to
biological data. Biometrics 9:176-200.
Carr, S. M., J. A. Brothers, and A. C. Wilson. (1987) Evolutionary inferences
from restriction maps of mitochondrial DNA from nine taxa of Xenopus
frogs. Evolution 41:176-188.
Clark, C. G. (1987) On the evolution of ribosomal RNA. /. Mol. Evol. 25:
343-350.
Clark, C. G., B. W. Tague, V. C. Ware, and S. A. Gerbi. (1984) Xenopus laevis
28S ribosomal RNA: a secondary structure model and its evolutionary and
functional implications. Nucleic Acids Res. 12:6197-6220.
DeSalle, R., T. Freedman, E. M. Prager, and A. C. Wilson. (1987) Tempo and
mode of sequence evolution in mitochondrial DNA of Hawaiian Drosophila.
J. Mol. Evol. 26:157-164.
Duellman, W. E., and L. Trueb. (1986) Biology of Amphibians. McGraw-Hill,
New York.
Estes, R. (1981) Gymnophiona, Caudata. Handb. Palaherpetol. 2:1-115.
Felsenstein, J. (1988) Phylogenies from molecular sequences: inference and reli-
ability. Ann. Rev. Genet. 22:521-565.
EVOLUTIONARY ANALYSIS OF LENGTH-VARIABLE SEQUENCES 247

Fitch, W. M. (1971) Toward defining the course of evolution: minimum change

for a specific tree topology. Syst. Zoo/. 20:406-416.
Fitch, W. M. (1980) Estimating the total number of nucleotide substitutions since
the common ancestor of a pair of homologous genes: comparison of several
methods and three beta hemoglobin messenger RNAs. J. Mol. Evol. 16:
153-209.
Gerbi, S. A. (1985) Evolution of ribosomal DNA. Pp. 419-517 in Molecular
Evolutionary Genetics (R. J. Maclntyre, ed.). Plenum Press, New York.
Gillespie, J. H. (1986a) Statistical aspects of the molecular clock. Pp. 255-272 in
Evolutionary Process and Theory (S. Karlin and E. Nevo, eds.). Academic
Press, London.
Gillespie, J. H. (1986b) Variability of evolutionary rates of DNA. Genetics 113:
1077-1091.
Gillespie, J. H. (1989) Lineage effects and the index of dispersion of molecular
evolution. Mol. Biol. Evol. 6:636-647.
Gonzalez, I. L., J. L. Gorski, T. J. Campen, D. J. Dorney, J. M. Erickson, J. E.
Sylvester, and R. D. Schmickel. (1985) Variation among human 28S ribo-
somal RNA genes. Proc. Natl. Acad. Sci. USA 82:7666-7670.
Gonzalez, I. L., J. E. Sylvester, T. F. Smith, D. Stambolian, and R. D. Schmickel.
(1990) Ribosomal RNA gene sequences and hominoid phylogeny. Mol. Biol.
Evol. 7:203-219.
Gorski, J. L., I. L. Gonzalez, and R. D. Schmickel. (1987) The secondary structure
of human 28S rRNA: the structure and evolution of a mosaic rRNA gene.
/. Mol. Evol. 24:236-251.
Hassouna, N., B. Michot, and J.-P. Bachellerie. (1984) The complete nucleotide
sequence of mouse 28S rRNA gene: implications for the process of size
increase of the large subunit rRNA in higher eukaryotes. Nucleic Acids Res.
12:3563-3583.
Holmquist, R. (1983) Transitions and transversions in evolutionary descent: an
approach to understanding. /. Mol. Evol. 19:134-144.
Jukes, T. H. (1980) Silent nucleotide substitutions and the molecular evolutionary
clock. Science 210:973-978.
Kimura, M. (1983) The Neutral Theory of Molecular Evolution. Cambridge Uni-
versity Press, Cambridge.
Lake, J. A. (1987) A rate-independent technique for analysis of nucleic acid
sequences: evolutionary parsimony. Mol. Biol. Evol. 4:167-191.
Larson, A. (1991). A molecular perspective on the evolutionary relationships of
the salamander families. Evol. Biol. 25:211-277.
Larson, A., and A. C. Wilson. (1989) Patterns of ribosomal RNA evolution in
salamanders. Mol. Biol. Evol. 6:131-154.
Maxson, L. R., and A. C. Wilson. (1979) Rates of molecular and chromosomal
evolution in salamanders. Evolution 33:734-740.
Moritz, C., T. E. Dowling, and W. M. Brown. (1987) Evolution of animal mito-
chondrial DNA: relevance for population biology and systematics. Ann.
Rev. Ecol. Syst. 18:269-292.
Nei, M., and W.-H. Li. (1979) Mathematical model for studying genetic variation
in terms of restriction endonucleases. Proc. Natl. Acad. Sci. USA 76:
5269-5273.
Nichols, B. P., G. F. Miozzari, M. Van Cleemput, G. N. Bennett, andC. Yanofsky.
(1980) Nucleotide sequences of the trp G regions of Escherichia coli, Shigella
248 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

dysenteriae, Salmonella typhimurium, and Serratia marcescens. J. Mol. Biol.

142:503-517.
Prager, E. M., and A. C. Wilson. (1988) Ancient origin of lactalbumin from
lysozyme: analysis of DNA and amino acid sequences. /. Mol. Evol. 27:
326-335.
Qu, L.-H., M. Nicoloso, and J.-P. Bachellerie. (1988) Phylogenetic calibration of
the 5' terminal domain of large rRNA achieved by determining twenty
eucaryotic sequences. /. Mol. Evol. 28:113-124.
Schmickel, R. D., J. Sylvester, D. Stambolian, and I. L. Gonzalez. (1990) The
ribosomal DNA sequence for the study of the evolution of human, chim-
panzee, and gorilla. Pp. 11-32 in DNA Systematics Volume HI: Human and
Higher Primates (S. K. Dutta and W. P. Winter, eds.). CRC Press, Boca
Raton.
Sidow, A., and A. C. Wilson. (1990) Compositional statistics: an improvement of
evolutionary parsimony and its application to deep branches in the tree of
life. J. Mol. Evol. 31:51-68.
Spencer, D. F., J. C. Collings, M. N. Schnare, and M. W. Gray. (1987) Multiple
spacer sequences in the nuclear large subunit ribosomal RNA gene of Crith-
idia fasciculata. EMBOJ. 6:1063-1071.
Swofford, D. L. (1984) Phylogenetic Analysis Using Parsimony. Illinois Natural
History Survey, Champaign.
Uzzell, T., and K. W. Corbin. (1971) Fitting discrete probability distributions to
evolutionary events. Science 172:1089-1096.
Ware, V. C., B. W. Tague, C. G. Clark, R. L. Course, R. C. Brand, and S. A.
Gerbi. (1983) Sequence analysis of 28S ribosomal DNA from the amphibian
Xenopus laevis. Nucleic Acids Res. 11:7795-7817.
Wilson, A. C., S. S. Carlson, and T. J. White. (1977) Biochemical evolution. Ann.
Rev. Biochem. 46:573-639.
Wilson, A. C., E. A. Zimmer, E. M. Prager, andT. D. Kocher. (1989) Restriction
mapping in the molecular systematics of mammals: a restrospective salute.
Pp. 407-419 in The Hierarchy of Life. Molecules and Morphology in Phy-
logenetic Analysis. (B. Fernholm, K. Bremer, and H. Jornvall, eds.) Elsevier
Science Publishers B.V., Amsterdam.
12
Statistical Methods for Testing
Molecular Phytogenies

WEN-HSIUNG LI AND MANOLO GOUY

Although numerous methods have been developed for reconstructing phy-

logenetic trees, there exist few methods for evaluating the statistical con-
fidence of an inferred phylogeny or for testing whether one phylogeny is
significantly better than another. In fact, the statistical methodology for
testing phylogenies is in a rather primitive state. This has occurred for two
reasons. First, although phylogenetic reconstruction has long been rec-
ognized as a problem in statistical inference (Edwards and Cavalli-Sforza,
1964), few authors have formulated the problem in a statistical framework.
Indeed, most current methods yield one or a few trees and do not provide
information concerning the confidence level of estimated phylogenies. Sec-
ond, the problem is extremely complex, largely because the number of
possible alternative trees is large even when only a moderate number of
taxa are involved. For this reason, most current statistical tests are heuristic
when the number of taxa involved is five or larger. Fortunately, the rapid
accumulation of DNA sequence data has stimulated a strong trend to make
phylogenetic reconstruction more statistical.
Statistical tests can be classified as analytical or resampling. Resampling
methods (e.g., bootstrapping, jacknifing) resample the data to infer em-
pirically the variability of the estimate obtained by a tree-making method.
For a recent review of resampling methods, see Felsenstein (1988). In this
chapter, we discuss only analytical methods. Analytical tests can be based
on parsimony methods, distance methods, likelihood methods, or invariant
methods. The last approach, which includes the evolutionary parsimony
method, has been reviewed in Felsenstein (1988). We discuss only the
other approaches.

PARSIMONY TESTS

This analytic approach was initiated by Cavender (1978, 1981). He studied

the confidence limits that can be placed on a phylogeny inferred from a

249
250 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

four-species data set. The results were extended by Felsenstein (1985), who
obtained narrower confidence limits by adding the assumption of a constant
rate of evolution (i.e., an evolutionary clock). For ease of representation,
we shall discuss Felsenstein's results before discussing Cavender's. We shall
also discuss the case of more than four species.

Four Species With a Molecular Clock

We assume that species 4 is a known outgroup, so that it can be used as
a reference to infer the branching order of the other three species. The
three possible rooted trees are shown in Figure 12-la-12-lc. In terms of
parsimony, a nucleotide site is informative (useful) for choosing among
the three trees only if it is in the same state in two of the four species, and
in another state in the other two species. For example, if the site has the
configuration AAGG among species 1-4, respectively, then it supports
tree la because this tree requires (at least) only one nucleotide substitution
to explain the configuration, whereas trees Ib and Ic each require two
substitutions. As another example, if the configuration is ATGG, then it
is not informative because each of the three trees requires two substitutions.
We assume that all informative sites are equivalent and that each in-
formative site supports tree la with probability p, tree Ib with probability
q, and tree Ic with probability r. Then the probability that among n random
informative sites there are / sites supporting tree la, / sites supporting tree
Ib, and k sites supporting tree Ic is trinomial, that is,

For the moment, let us assume that tree la is the true tree. Under the
assumption of rate-constancy, an informative site has a higher probability
of supporting the true tree than supporting an incorrect tree. So, p > q
= r. The probability that among n random informative sites, the number
(C) of sites supporting a particular incorrect tree is m or larger is given by

which is obtained by expanding [q + (p + q)]". Since p + 2q = 1 and

since q < p, formula (2) assumes the maximum value when p = q = Vs,
that is, if the three species represent a true trichotomy (Fig. 12-Id).
Now suppose that we do not know which tree is the true tree. Suppose
further that in a sequence data set with n informative sites, the best sup-
ported tree is favored by m sites. Is this tree the true tree? Let us assume
that this tree is incorrect and has by chance been supported by so many
sites. The probability for a particular incorrect tree to be supported by m
or more sites is given by equation (2). Therefore, the probability for one
of the two incorrect trees to be supported by m or more sites is
STATISTICAL METHODS FOR TESTING MOLECULAR PHYLOGENIES 251

Figure 12-1 a, b, and c represent the three possible rooted, bifurcating trees for
three taxa with one outgroup and d represents the case of a trichotomy. (From Li
and Gouy, 1990.)

If P is smaller than a, then the support for the best tree is significant at
the level of a.
In practice, p is unknown, and it is necessary to use the least favorable
value, Vs. Table 12-1 shows the minimum m value required for equation
(2) to be smaller than or equal to a given probability (d), assuming that n
informative sites are observed and that q = p = Vs. It can be used to
determine the significance level. For example, if n = 10 and m = 8, then
the significance level is approximately 0.5% from Table 12-1 and 2 x 0.5%
= 1% from formula (3), and so the inferred tree is significant at the 1%
level. If n is greater than 100, equation (2) can be computed by assuming
that mln follows a normal distribution with mean equal to l/s and variance
equal to (l/3 x %)/«; the approximation is very rough, however.
In addition to the above test, Felsenstein (1985) has also developed a
statistic for testing whether the best supported tree is significantly better
than the next best tree. Suppose that these two trees are supported by n2
and n2 sites, respectively. Let
S = n1 - n2.
Then the probability that the best supported tree is at least s steps better
is given by

(ij,k)
j - max(/,&) > s
or k - max(/,y) ^ s
If Prob (S > s) is smaller than a, then the best supported tree is significantly
better than the next best tree at the a level. Again we do not know the p
value and so we accept the worst-case analysis assuming p = q = r = 1A.
Table 12-2 shows some results computed from this analysis. For example,
if there are n = 10 informative sites, then the best supported tree must
be at least five steps better than the next best tree in order to be significantly
better at the 5% level.
The test given by formula (4) is computationally somewhat more com-
plicated than that given by formula (3), but the powers of the two tests
Table 12-1 Minimum m Value Required for Prob(C a m) in Equation (2) to be Smaller Than or Equal to <5*
5(%) 5(%)
f
n 0.05 0.30 0.50 1.00 2.50 n 0.05 0.30 0.50 1.00 2.50
10 9 9 8 8 7 34 22 20 20 19 18
11 10 9 9 8 8 36 23 21 20 20 19
12 11 10 9 9 8 38 24 22 21 21 19
13 11 10 10 9 9 40 24 23 22 21 20
14 12 11 10 10 9 43 26 24 24 23 22
15 12 11 11 10 10 47 28 26 25 24 23
16 13 12 11 11 10 50 29 27 26 26 24
17 13 12 12 11 11 53 30 28 28 27 26
18 14 13 12 12 11 57 32 30 29 28 27
19 14 13 13 12 11 60 33 31 31 30 28
20 15 14 13 13 12 63 35 33 32 31 29
21 15 14 14 13 12 67 36 34 34 33 31
22 16 15 14 14 J3 70 38 35 35 34 32
23 16 15 15 14 13 73 39 37 36 35 33
24 17 16 15 15 14 77 41 38 38 36 35
25 17 16 16 15 14 80 42 40 39 38 36
26 18 16 16 15 14 83 43 41 40 39 37
27 18 17 17 16 15 87 45 42 42 40 39
28 19 17 17 16 15 90 46 44 43 42 40
29 19 18 17 17 16 93 47 45 44 43 41
30 20 18 18 17 16 97 49 46 46 44 43
32 21 19 19 18 17 100 50 48 47 46 44
"It is assumed that q = p = 'A
r
n = number of informative sites observed
(Taken from Li and Gouy, 1990 )
STATISTICAL METHODS FOR TESTING MOLECULAR PHYLOGENIES 253

Table 12-2 Difference (S) in the Number of Supporting Sites Required for
Significance at the 5% Level
Sites Informative (Clock) All Sites
2 — 2
3 — 3
4 4 3
5 5 3
10 5 5
13 5 6
15 6 6
20 6 8
25 7 9
30 8 10
40 9 13
50 9 15
100 13 26
200 17 48
500 27 109
1,000 — 209
2,000 — 405
5,000 — 984
10,000 — 1,940
(From Felsenstein, 1988.)

are about the same (Felsenstein, 1985). This is not surprising, because if
the best supported tree is significant, it is likely to be significantly better
than the next best tree (and also the worst tree), and vice 'versa.
Although the above results were derived under the assumption that one
of the four species is known to be an outgroup, they apply equally well in
the absence of this knowledge if we consider unrooted trees and if the rate-
constancy assumption holds.

Four Species Without a Molecular Clock

Although in the above example we have assumed a molecular clock, the

analysis was made under the worst situation, namely that the probability
for an informative site to support an incorrect tree is Va rather than < Vs.
Thus, the results are applicable even under nonconstant rates as long as
the rates among the three lineages are not too different or the sequences
have not diverged greatly so that the probability for an informative site to
support an incorrect tree is ^ 1A. However, if this condition does not hold,
then a different analysis is necessary. Cavender (1978) has carried out such
an analysis, considering morphological characters with two states. Felsen-
stein (1983) extended the calculation to the case of four states, such as
nucleotide sequences.
Cavender's formulation is as follows. Suppose that the tree in Figure 12-
2 is the true tree. Assume that all characters are equivalent and can be
characterized by two states, 0 and 1. Assume further that a change of a 0
254 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Figure 12-2 A hypothetical tree for four

sequences.

to a 1 over the course of time may be followed by a change back to 0 and

that a change one way is just as probable as a change the other way. In
Figure 12-2, the branch length a denotes the probability that a character
at vertex 1 differs in state from that at vertex 5; the other branch lengths
are defined in the same manner. In the worst-case analysis, Cavender
assumes that the rate of evolution is rapid in both lineages 1 and 4 so that
their characters have been randomized, whereas the rates in the other
lineages are extremely slow so that no change has occurred. That is, b =
c = e = 0 and a = d = l/2. If we assume that the state is 0 at vertices 2
and 3, then for the site to be informative, the state should be 1 at both
vertices 1 and 4. This occurs with probability Vz x l/i = 1A, and in this
case the site supports a wrong tree. Thus, in the worst case, one-fourth of
all characters can support an incorrect tree with none supporting the correct
tree. For this reason one can conclude in favor of the most-parsimonious
tree only if the evidence is stronger than that.
In the case of nucleotide sequences, with a simple symmetric model of
base change, Felsenstein (1983) showed that 3/ie of all characters can be
"phylogenetically informative" and favor a wrong tree. Therefore, in this
case one asks whether the proportion of sites supporting a tree is signifi-
cantly greater than 3/ie of all sites. This probability can be obtained by a
binomial expansion of (13/i6 + 3/ie)". The third column of Table 12-2 gives
the difference in steps required for significance at the 5% level.

More Than Four Species

In theory, the above analysis can be extended to the case of more than
four species. However, the situation now becomes very complicated be-
cause the number of possible alternative trees increases rapidly with the
number of species; for example, for five species there are 15 possible
unrooted trees. Moreover, the above formulation is based on the worst-
case analysis, so when the number of possible alternative trees is large,
the test developed is likely to have very low power.
In view of the complexity of the problem, Kishino and Hasegawa (1989)
STATISTICAL METHODS FOR TESTING MOLECULAR PHYLOGENIES 255

developed a heuristic test in a manner similar to that of Templeton (1983).

They considered whether two trees X and Y differ significantly in terms
of the number of substitutions. Assume that all nucleotide sites are inde-
pendent and identical. Let the minimum number of base substitutions at
the i-th site for tree X be X, and let
X = Xl + X2 + + Xn. (5)
We define Y in the same manner. Let Z = X — Y. The mean of Z can
be estimated by the inferred values of X, and Yf. The variance of Z can
be estimated by the sample variance over sites, that is,

where Z, = X, — Y, and N is the total number of sites on the sequence

When N is large, [Z - £(Z)]/V(Z) approaches the standard normal dis-
tribution and the standard normal test can be used to test the significance
of Z. In practice it is difficult to know how large N should be for Z to be
close to a normal distribution. We note that actually the number of in-
formative sites is more relevant than the total number of sites.
It should also be noted that when one tree is significantly more parsi-
monious than another tree, it does not mean that it is the most parsimonious
tree. To draw this conclusion, one must show that the tree is more par-
simonious than all of the other possible trees. In practice, however, when
the number of possible trees is large, we cannot compare all of them and
so instead we usually compare only those that seem to be most plausible.
In this respect, one should note that two trees usually differ only in a small
part of the whole topology, and in comparing two trees, we implicitly
assume that the other part of the tree topology is correct. Snce this as-
sumption may not be true, it can complicate the test.

TESTS OF INTERNAL BRANCH LENGTHS

Instead of testing the significance of an inferred phylogeny, it is simpler

to test the significance of estimated internodal distances. As examined
below, in the case of four taxa significance of the internodal distance can
be taken as significance of the inferred phylogeny. When the number of
taxa under study is more than four, the two problems are no longer equiv-
alent and the requirement of all internodal distances being significantly
greater than 0 seems to be a too stringent test for the significance of the
inferred branching order. A simple way to test the significance of internodal
distances is to study their variances.

Variances of Internodal Distances

Mueller and Ayala (1982) proposed to compute these variances by the
jackknife method, while Nei et al. (1985) derived analytic formulas for the
256 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

case of a UPGMA tree. That is, a tree estimated by the unweighted pair-
group method (UPGMA) (Sneath and Sokal, 1973). The UPGMA method
assumes a constant rate of evolution, but there is now strong evidence that
this assumption is often violated (e.g., Li et al., 1987a). It is therefore
desirable to consider an approach that does not make this assumption. Li
(1989) proposed a two-step approach. The first step is to infer the branching
order. One can use the transformed distance method (Farris, 1977; Klotz
et al., 1979; Li, 1981), the neighbor-joining method (Saitou and Nei, 1987),
or any other method that does not assume rate-constancy and that has
been shown to be effective for obtaining the correct tree. The second step
is to estimate the branch lengths by the least-squares method (Cavalli-
Sforza and Edwards, 1967; Chakraborty, 1977). The variances of internodal
distances are then obtained from the equations derived from the least-
squares method. The variances for the cases of four and five taxa are given
below.

Four Taxa
Suppose that the inferred tree topology is as shown in Figure 12-3a. The
branch lengths estimated by the least-squares method are

and the variance of c is

where V(d,j) denotes the variance of the estimate of the distance between
sequences i and /'.
First, consider protein sequence data. The mean and variance of the
number (dv) of amino acid replacements per site between sequences / and
/ can be estimated by

where b = 19/20, p is the proportion of different amino acids between the

two sequences, and N is the number of residue sites compared.
STATISTICAL METHODS FOR TESTING MOLECULAR PHYLOCENIES 257

Figure 12-3 Model trees used in the derivation of the mean and variance of the
branch lengths. (From Li, 1989.)

Next, consider nucleotide sequence data. Under the assumption of ran-

dom substitution among the four types of nucleotides (i.e., the one-
parameter model), the mean and variance of the number of substitutions
per nucleotide site between sequences i and j are also given by (13) and
(14), except that now b = 3/i, p is the proportion of different nucleotides
between the two sequences, and N is the number of nucleotide sites com-
pared (Jukes and Cantor, 1969; Kimura and Ohta, 1972). Under the two-
parameter model (Kimura, 1980), the formulas corresponding to (13) and
(14) are

where P and Q are, respectively, the proportions of transitional and trans-

versional differences between sequences i and ;', A = l/2tn(x) - lA£n(y)
is the number of transitional substitutions per site, B = l/zfn(y) is the
number of transversional substitutions per site, x = 1 / (1 - 2P - Q),
y = 1 / (1 - 20, and z = (x + y)/2.

Five Taxa
Suppose that the inferred branching order is as shown in Figure 12-3b.
The branch lengths are then given by
258 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

If two of the five taxa, say taxa 4 and 5, are known to be outgroups, then
one needs to obtain only the variance of c.

If only one or no outgroup is available, then one needs also to obtain the
variance of e.

Computer programs for the above formulas are available upon request
by sending an IBM PC-compatible floppy disk to the authors.

Test of Significance of an Inferred Phylogeny

In the case of three taxa with one or two outgroups, the above results can
be used to test the significance of an inferred phylogeny. Since in this case
there is only one internal branch (i.e., branch c), testing the significance
of the internal branch is equivalent to testing the significance of the inferred
phylogeny. More explicitly, the null hypothesis is that the true phylogeny
STATISTICAL METHODS FOR TESTING MOLECULAR PHYLOCENIES 259

is a trichotomy (i.e., the three taxa diverged at the same time). This hy-
pothesis is the same as the hypothesis of c = 0. Therefore, if the estimated
c is significantly greater than 0, the null hypothesis of a trichotomy is
rejected and the inferred branching order can be taken as statistically
significant. If one considers unrooted trees, the same argument applies to
the case of four taxa with no outgroup, because in this case there is also
only one internal branch (Fig. 12-3a).
The above formulas for the mean and variance of c were derived under
the assumption that the inferred branching order of the three taxa was ((1,
2), 3), which means that lineage 3 branched off earlier than did lineages
1 and 2. If, instead, the inferred branching pattern is ((1, 3), 2), then the
subscripts 2 and 3 in the above formulas should be exchanged, and if the
inferred branching pattern is ((2, 3), 1), subscripts 1 and 3 should be
exchanged. Under the null hypothesis of a trichotomy, the three branching
patterns ((1, 2), 3), ((1, 3), 2), and ((2, 3), 1) occur with equal probability.
However, in each data set only one pattern can occur and only one c can
be positive and is tested for significant deviation from 0. Regardless of
which pattern occurs, the probability for c to assume a particular (non-
negative) value is the same. If the distribution of c is the same as the
distribution of \x\ where x is a standard normal random variate, then the
standard statistical test based on the standard normal distribution can be
applied. In particular, the estimated c is significantly greater than 0 at
the 5% level if the ratio of mean to standard error is > 2, and is signifi-
cant at the 1% level if the ratio is > 2.6. Obviously, the case of four taxa
with no outgroup can be treated in the same manner, if one considers
unrooted trees. A computer simulation (Li, 1989) for the case of three
taxa with one outgroup shows that the significance level defined by the
above criteria is generally applicable, though the distribution of c is not
strictly normal.
When the number of taxa under study is more than four, the situation
becomes complicated. For example, in the case of five taxa there are two
internal branches (Fig. 12-3b), and the probability for (only) one of them
to become by chance significantly greater than 0 at the level of a = 5%
is 2a = 10%. Thus, in this case one cannot reject the null hypothesis that
all the internal branches have 0 length (i.e., all the taxa diverged at the
same time point and form a "star" phylogeny); of course, this null hy-
pothesis can be rejected if a s 2.5%. On the other hand, the probability
for both internal branches to be by chance significantly greater than 0 at
the level of a = 5% is only approximately a2 = 0.0025 (it is approximate
because the two internodal distances are not estimated independently).
Hence the requirement of all internal branches being significant seems to
be a too stringent test for the significance of the inferred topology. How-
ever, one cannot draw a conclusion about the significance of an inferred
tree topology as long as one or more of the internodal distances are
nonsignificant; of course, the uncertainty can be restricted to a subset of
taxa.
260 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Numerical Examples

The following examples assume that the rate of nucleotide substitution is

constant over time and that the observed number of substitutions between
each pair of sequences is equal to the expected value.
Table 12-3 shows the case of three species with an outgroup (taxon 4)
(Fig. 12-3a). In the table, a denotes the proportion of transitional changes;
a = 1A if substitutions occur randomly. The standard error (SE), which is
the square root of V(c), is larger for a = 2/s than for a = Vs. Since
transitional changes generally occur more often than transversional changes
(Brown et al., 1982; Li et al., 1984) the two-parameter model is more
realistic than the one-parameter model; the former is applicable to all a
values whereas the latter only to a = Vs.
The ratio c/SE can be used to test whether c is significantly greater than
0. A ratio of 2 can be taken ,as significant at the 5% level. All the values
in Table 12-3 were obtained for N = 1,000. When c = 0.01, the ratio is
2 or larger if a < 2/s. Thus, this case requires only a small amount of
sequence data to resolve the branching order of the three species. When
c is 0.005, then the ratio is considerably smaller than 2 (e.g., the ratio is
1.28 for a = Vs). Formulas (14) and (16) imply that V(c) is inversely
proportional to N. Therefore, for the ratio to increase from 1.28 to 2, the
TV value should increase from 1,000 to N' = 1,000 x (2/1.28)2 = 2,500
approximately. The other N' values in Table 12-3 were obtained in the
same manner. If c is 0.001, then the number of nucleotide sites needed to
be studied is rather large, greater than 50,000. Saitou and Nei (1986) have
considered this problem from a different angle. They studied the proba-
bility of obtaining the correct topology as a function of the number of
nucleotides studied under various tree-making methods.
Table 12-4 shows the amount of reduction in V(c) when a second out-
group (taxon 5) is added (Fig. 12-3b); the V(c) value for the case of one
outgroup is denoted by V^c) and that for the case of two outgroups by

Table 12-3 Standard Error (SE) of the Estimate of the Length of Branch c in
Figure 12-3a*
c J SE* c/SE L'«
0.010 '/3 0.0046 2.17 850
% 0.0050 2.00 1,000
0.005 !/3 0.0039 1.28 2,500
% 0.0043 1 16 3,000
0.001 V3 0.0032 0.31 41,000
% 0.0037 0.27 54,000
'The branch lengths in Figure 12-3a are a = b = 0.05, d = 0.05 + c, e = 0.15 - c.
*a = proportion of transitional substitutions.
*SE was computed under the assumption that the number of nucleotide sites (L) studied is 1,000
5
Z/ is the number of nucleotide sites required for the ratio c/SE to be 2 or larger (i.e., to be approximately
5% significant)
(Taken from Li, 1989 )
STATISTICAL METHODS FOR TESTING MOLECULAR PHYLOGENIES 261

V2(c). The reduction increases as V^c) becomes larger. Since V(c) is in-
versely proportional to N, a reduction in V(c) can also be achieved by
increasing N. Is it more advantageous to increase TV or to add a second
outgroup? The total number of nucleotides sequenced is 4/V for the case
of one outgroup and 57V for the case of two outgroups, the latter being
1.25 times the former. Therefore, if the same total number of nucleotides
are to be sequenced, it is less advantageous to add a second outgroup than
to increase TV if V^(c)IV2(c) is smaller than 1.25, whereas the reverse is
true if the ratio is larger than 1.25. In Table 12-4 the ratio is smaller or
equal to 1.25 for the first six cases and is larger than 1.25 for the last six
cases. Since the ratio tends to increase with Fj(c), in general it is more
advantageous to increase N if V\(c) is relatively small, but more advan-
tageous to add a second outgroup if V^c) is relatively large. In all the
cases in Table 12-4, the distances from sequences 4 and 5 to the other three
are the same (i.e., g = fin Fig. 12-3b), so that the fifth sequence is as
good a reference as the fourth one. If the fifth is more distantly related to
the other three than the fourth sequence is, then the reduction in V(c) is
expected to be smaller than those shown in Table 12-4. Further, the effect
will also be reduced if sequences 4 and 5 are closely related to each other.

NEIGHBORS RELATION TESTS

Let us first define the concept of "neighbors." In an unrooted tree, two

operational taxonomic units (OTUs) are said to be neighbors if they are
connected through a single internal node. For example, in Figure 12-4,
OTUs 1 and 2 are a pair of neighbors, and OTUs 5 and 6 are another pair.
In comparison, neither OTUs 1 and 3 nor OTUs 2 and 3 are neighbors.
However, if we combine OTUs 1 and 2 into a composite OTU, then the
combined OTU (12) and OTU 3 become a new pair of neighbors.
We explain below how the neighbors relation concept can be used to
reconstruct phylogenetic trees. We then propose methods for testing the
significance of neighbor pairs.
Four Taxa

Assume that the tree shown in Figure 12-3a is the true tree. The branch
lengths are said to be additive, if the following equalities hold.
dl3 + d24 = d14 + d23 = a + b + 2c + d + e
= dl2 + d34 + 2c, (26)
where dtj is the distance (number of substitutions) between OTUs I and j
estimated from sequence data while a, b, c, d, and e are the inferred branch
lengths. Therefore, under additivity the following two conditions hold:
dl2 + d^ < dl3 + d24 (27a)
dl2 + d34 < dl4 + d23. (27b)
Table 12-4 Variance (V(c)) of the Estimate of the Length of Branch c in Figure 12-3b
V,(c)' VJ.c)
a=b d /=* e c a* (xlO- 4 ) (xlO-«) V,(c)/V2(C)
0.020 0.025 0.050 0.025 0.005 '/3 0.072 0.066 1.09
% 0.079 0.071 1.11
0.029 0.001 >/3 0.028 0.023 1.22
2
/3 0.033 0.027 1.22
0.050 0.055 0.100 0.045 0.005 V, 0.153 0.126 1.21
% 0.188 0.150 1.25
0.049 0.001 '/3 0.104 0.079 1.32
% 0.135 0.100 1.35
0.100 0.110 0.150 0.040 0.010 '/3 0.445 0.351 1.27
% 0.577 0.444 1.30
0.045 0.005 '/3 0.376 0.287 1.31
% 0.501 0.374 1.34
*a = proportion of transitional substitutions
*V,(c) was obtained under the assumption that the second outgroup (taxon 5 in Fig. 12-3b) is not available.
(Taken from Li, 1989.)
STATISTICAL METHODS FOR TESTING MOLECULAR PHYLOGENIES 263

Figure 12-4 A model tree for illustrating

the neighbors relation tests.

These two conditions are known as "the four-point condition" (Buneman,

1971). They may hold even if the additivity holds only approximately.
Conversely, for four OTUs with unknown relationships, if the above
two conditions hold, then we may assume that OTUs 1 and 2 are neighbors,
and so are OTUs 3 and 4. The question is then whether the topology
inferred is statistically significant.
A simple way to test the significance of the topology is to test whether
both x and y are significantly greater than 0, where

x = (dl3 + d24) - (d12 + d34) (28a)

y = (d14 + d23) - (dl2 + d34). (28b)

However, our computer simulation shows that this test is too stringent,
and a more reasonable test is to consider whether z = x + y is significantly
greater than 0.
If the degree of sequence divergence is large, one should use the cor-
rected number of nucleotide substitutions per site for dt] (Saitou and Im-
anishi, 1989; Jin and Nei, 1990). In this case, one can compute the variance
of z and use it to test the significance of z in a manner similar to that in
the preceding section.
However, if the degree of sequence divergence is not large, say 50% or
less, then it is slightly better to use the observed proportion of differences
for dg (Saitou and Nei, 1987). Assume that all nucleotide sites are identical
and independent. Let zk be the z value for the k-th site and Z = 2zfc; the
dij values used are the observed numbers of differences between sequences
i and / at site k. Then the variance of z = Z/N can be estimated by

where z = ^z^N, and N is the number of nucleotide sites on the sequence.

This formulation is computationally simple and can be easily extended to
the case of many OTUs (see below).
264 PHYLOCENETIC ANALYSIS OF DNA SEQUENCES

Instead of formula (29), the variance of Z can be computed as follows.

Since Z = dl3 + d24 + dl4 + d23 - 2(d12 + d34),

Let ptj be the proportion of different nucleotides between sequences i and

j. Then V(dij) = ptj(l - ptl)/N, which is the binomial variance. For the
covariances we note that (with m and n referring to the other two se-
quences)
Cov(^,d_) = E(d,,dmn) - E(dlJ)E(dmn)
= E(dljdmn) - ptpmn.

The term E(dtidmn) can be estimated by computing the dijdmn value for
each nucleotide site and then taking an average over the sites in the se-
quence, as in the computation of equation (29).
In the case of five taxa with taxa 4 and 5 as known outgroups, we have

where dl(]K} = (dt] + dlk)/2. The variance of z = x + y can be obtained

as above, though more tediously.

More Than Four Taxa

Sattath and Tversky (1977) proposed the following method of phylogenetic

reconstruction for the case of more than four OTUs (see also Fitch, 1981).
First, compute a distance matrix. For each OTU pair (if), examine all other
pairs (mn) and count the number of quadruples in which the four-point
condition holds (i.e., dt] + dmn is smaller than both dim + djn and d,n +
d/m). The pair with the highest score is selected and considered as a single
(composite) OTU. Second, compute a new distance matrix and search for
the next neighbor pair following the above procedure. The process is re-
peated until all OTUs are clustered. In this method, the distance between
two OTU clusters is computed by the arithmetic mean of all the pairwise
STATISTICAL METHODS FOR TESTING MOLECULAR PHYLOGENIES 265

distances involved. For example, the distance between clusters (ijk) and
(mri) is defined as
d(f,k)(mn) = (dim + 4m + djm + d,n + dkm + dkn)!6.
After the topology is inferred by the above procedure, we propose to
test its significance by testing whether all the subsets in the tree are sig-
nificantly supported. Suppose that the inferred tree is that shown in Figure
12-4. We may first test the significance of the subset (12) by testing whether
+ 2
Z = dl3 + 42(456) 4>3 + ^1(456) ~ [^12 + 4s(456)] (32)

is significantly greater than 0. We can test the significance of the subset

(56) in the same manner. Then there are only four OTUs left: (12), 3, 4,
and (56). Since the hypothesis is that (12) and 3 form one pair of neighbors,
and 4 and (56) form another pair, we test whether
2
= [4(12)(56) + 4(4] + [4(12)4 + 43(56)] ~ 2[4(12)3 + ^4(56)] ( 33 )

is significantly greater than 0.

Another way of testing the significance of a subset is as follows. For
every possible pair of OTUs i and / in the subset, and for every possible
pair of OTUs m and n not in the subset, we test whether
z = dm + d,n + dm + djm - 2K, + dmn] (34)
is significantly greater than 0. If every test is significant, then the subset
is taken to be significant. This test is more stringent than the preceding
test.
Some of the above formulas look complicated, but computer compu-
tation is straightforward. Computer programs are available from the au-
thors.

LIKELIHOOD TESTS

The first application of the maximum likelihood (ML) method to tree

reconstruction was made by Cavalli-Sforza and Edwards (1967), who used
gene-frequency data. Later, Felsenstein (1973, 1981) developed ML al-
gorithms for amino acid or nucleotide sequence data (see also Neyman,
1971; Kashyap and Subas, 1974). In the past, however, the ML method
has not been used frequently, largely because of computational difficulties.
Fortunately, computer computation is now much faster, and in recent years
a number of authors have become interested in this method (Hasegawa et
al., 1985, 1987; Felsenstein, 1987, 1988; Saitou, 1988, 1990; Smouse and
Li, 1987; Kishino and Hasegawa, 1989, and references therein).

Likelihood Function

To use the ML method, one must first have a probabilistic model for the
process of nucleotide substitution. That is, one must specify the transition
266 PHYLOCENETIC ANALYSIS OF DNA SEQUENCES

Figure 12-5 Model trees for the derivation of the likelihood function.

probability from one nucleotide state to another in a time interval. For

example, in the one-parameter (or Jukes-Cantor) model, if the rate of
substitution is A per site per unit time, then the probability of changing
from nucleotide / to nucleotide ; per unit time is A/3. Furthermore, given
nucleotide i at time 0, the probability that the nucleotide at time t is also
i is given by

and the probability that the nucleotide at time t is j , j + i, is given by

In general, we can write

where P(t) is the matrix of transition probabilities P,,(i) and the matrix R
consists of the rate of change rit from nucleotide i to nucleotide / per unit
time; the diagonal element ru is defined so that the sum of the elements
in each row is equal to 0.
The next step is to set up the likelihood function. Let us use the case
of four sequences (taxa) as an example. For the hypothetical tree given in
Figure 12-5a, the likelihood function for a nucleotide site with nucleotides
i,;', k, and / in sequences 1, 2, 3, and 4, respectively, can be computed as
follows. Assume a constant rate of substitution. If the nucleotide at the
ancestral node (the root) was x, the probability of having nucleotide / in
sequence 4 is Pxt(ti + t2 + t3), the probability of having nucleotide y at
the common ancestral node of sequences 1, 2, and 3 is Pxy(t^), and so on.
STATISTICAL METHODS FOR TESTING MOLECULAR PHYLOGENIES 267

Therefore, given x, y, and z at the ancestral node and the two other interioi
nodes, the probability of observing i, j, k, and / at the tips is equal to
P*t Cl + '2 + 'a) P,y Cl) Pyk (h + h} Pyz (h) Pzi fe) PZJ fe).
All the above transition probabilities can be computed by using equations
(35) and (36). Since in practice we do not know the ancestral nucleotide
we can only assign a probability gx, which is usually assumed to be the
frequency of nucleotide x in the sequence. Noting that x, y, and z can be
any of the four nucleotides, we sum over all possibilities and obtain the
following likelihood function.

This equation is the same as equation (14) of Saitou (1988), except that
the term Pyz (t2) is now placed inside rather than outside the summation
over z (i.e., the last summation).
A formulation without the assumption of rate-constancy, though more
tedious, can be done in the same manner. In this case, it is usually more
convenient to consider the transition probability in terms of the branch
length; for example, we consider Ptl (va) instead of P,; (ta), where va =
hata and Aa and ta are the values of A and t for the a-th branch.
For an unrooted tree, we usually do not assume rate-constancy. For the
unrooted tree given in Figure 12-5b, we obtain the following likelihood
function:

if we assume that the interior node connecting taxa 3 and 4 is the ancestral
node (Felsenstein, 1981; Saitou, 1988). If the process is time-reversible so
that the transition probability from i to / in a time interval is the same as
that from; to i (i.e., Pi}(i) = P,,(f), which is the case for the one-parameter
model and many other current models of nucleotide substitution), then
any node or point in the tree can be taken as the "ancestral" node. This
is the so-called pulley-principle (Felsenstein, 1981). However, if the process
is not time-reversible, then we must assume or infer the root of the tree.
In the above we considered a single site. The likelihood for all sites is
the product of the likelihoods for individual sites if we assume that all the
nucleotide sites involved are independent. Suppose that there are s ho-
mologous sequences each with N nucleotides. We can represent the data
by a matrix X, in which the element Xtl is the nucleotide at they'-th site in
the z'-th sequence. Let Xk be the vector (Xlk, , Xsk) (i.e., the
transpose of the k-th column of the matrix X). Then for a given phylogenetic
tree, the likelihood for the entire sequence is
268 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

where /(x|0) = f(xlt x2, , xs\d) is the probability that at the nu-
cleotide site under study, sequence i has the nucleotide x,, and the vector
0 denotes the unknown parameters such as the branching dates and the
rates of nucleotide substitution along the branches of a tree (Kishino and
Hasegawa, 1989).
Under the assumption that all sites are independent and equivalent, the
likelihood function can also be derived by considering the nucleotide con-
figurations among the sequences (Saitou, 1988). A nucleotide configuration
at a site is defined as the pattern of nucleotide difference at that site among
the sequences involved. Let us consider the case of four sequences. As
shown in Table 12-5 there are 15 configuration patterns. The first pattern
(i, /', /, z) means that the nucleotide is the same in all four sequences, the
second pattern (/, /, j,;') means that the nucleotide is the same for the first
three sequences, but is different for the fourth sequence, and so on. For
the unrooted tree given in Figure 12-5b, and under the one-parameter
substitution model, we can show that the probability of obtaining the first
pattern is equal to
I/! = 4 h(i, i, i, /),
where h(i, i, i, i) is given by equation (39), and the factor 4 arises because
there are four possible nucleotides. The probability of obtaining the second
pattern is equal to
U2 = 12 h(i, i, i, j),
in which the factor 12 arises because there are four possibilities for i, and
for each i there are three possibilities for /. In the same manner, we can
obtain the probabilities for the other patterns in Table 12-5 (Saitou, 1988).

Table 12-5 Nucleotide Configurations for Four Sequences

Configuration*
Number A B c D
1 j i I j
2 i i i 1
3 i i ] i
4 i i i i
5 j i i i
6 i i 1 1
7 I j i I
8 i J i i
9 i i 1 k
10 i j I k
11 1 i I k
12 I k I i
13 i i k i
14 i i k i
15 i i k I
*A, B, C, and D are sequences, and j, /, k, and / are different nucleotides.
(From Saitou, 1988.)
STATISTICAL METHODS FOR TESTING MOLECULAR PHYLOGENIES 269

The likelihood function, Lp for the ;'-th topology is

where Uf] is the probability of obtaining the /-th pattern for the ;'-th to-
pology, Nf is the observed number of the j'-th pattern, C = N\/(N!\ N2\
A^!) and N is the total number of nucleotide sites examined and
is equal to the sum of N,'s.

Statistical Tests
After the likelihood function L is derived under a topology, one can search
for the ta or va values that maximize L. In the tree reconstruction procedure
proposed by Felsenstein (1981), one computes the ML value for all or
many topologies and chooses the one with the highest ML value as the
most likely candidate for the true tree. It has been noted, however, that
the likelihood function varies from topology to topology so that the ML
values for different topologies are conditional and cannot be compared in
the traditional statistical sense (Nei, 1987; Smouse and Li, 1987; Saitou,
1988). That is, the ML values for different topologies would be equivalent
to the ML values computed under different hypotheses and thus cannot
be compared in the traditional sense. However, Kishino and Hasegawa
(1989) suggest that the comparison is meaningful in view of Akaike's (1974,
1985) theory. Akaike has shown that an information criterion (AIC) de-
fined as
AIC = - 2 €«(maximum likelihood) + 2(number of parameters) (42)
can be used as a measure of badness of a model. This theory is based on
the Kullback-Leibler information, and two models can be compared even
if they are non-nested; the one with a lower AIC is preferred (Akaike,
1974, 1985). Unfortunately, although Akaike's theory provides a logical
basis for using the ML value as a criterion for choosing the best tree
topology, in terms of statistical testing the theory does not remove all
logical problems involved in non-nested models (see below).
In the case of four taxa, the situation is relatively simple. There are only
three possible unrooted trees (Fig. 12-6A). Since each tree has only one
internal branch, we may test its significance by testing the significance of
the internal branch length. We use tree 1 as an example. We note that if
the central branch is assumed to be 0, then tree 1 reduces to the starlike
tree in Figure 12-6B. Thus, testing the significance of the central branch
in tree 1 is equivalent to testing whether tree 1 is significantly better than
the starlike tree. Since the latter is a submodel of the former, the test can
be done in the traditional sense. The former model has one more parameter
than does the latter, and so twice the logarithm of the ratio of the ML
values for the two models is asymptotically distributed as a x2 with one
degree of freedom (see Smouse and Li, 1987; Felsenstein, 1988). Strictly
270 PHYLOCENETIC ANALYSIS OF DNA SEQUENCES

Figure 12-6 (A) Three possible topologies for an unrooted tree of four taxa. (B) The
starlike tree for four taxa. (From Saitou, 1988.)

speaking, the starlike model is not completely inside tree 1, but is only on
the boundary of the probability space. However, since the likelihood func-
tion is continuous at the boundary, this probably does not create a problem
(E. Thompson, cited in Felsenstein, 1988). Usually, only one of the three
trees in Figure 12-6A will have an ML value greater than the ML value
for the starlike tree, and the other two trees assume the ML value when
their internal branch is 0 (Saitou, 1988). However, in some cases more
than one tree may have an ML value larger than that for the starlike tree.
When such a situation occurs, we may follow Felsenstein's (1988) sugges-
tion to test whether the difference in the ML values between the two trees
is significant. Here a logical problem arises because the two ML values are
estimated with the same number of parameters (and under different hy-
potheses), so there is no degree of freedom left for statistical testing.
Felsenstein suggests to assume one degree of freedom and that twice the
logarithm of the ratio of their ML values follows the x2 distribution. This
test tends to be more stringent than the above test.
When more than four taxa are involved, the situation becomes compli-
cated. In this case, testing of the significance of all internal branches tends
to be more stringent than testing the significance of an inferred phylogeny,
but is simpler and may be the only feasible way. Because of the large
amount of computation involved, we consider only the ML tree (i.e., the
tree with the highest ML value among all possible trees). Actually, even
searching for this tree is extremely tedious. Therefore, Felsenstein (1981)
and Saitou (1988) have proposed step-by-step search methods. Saitou's
method for the case of five taxa is illustrated in Figure 12-7. At level I,
one considers a starlike tree and computes the ML value. At level II, one
STATISTICAL METHODS FOR TESTING MOLECULAR PHYLOGENIES 271

Figure 12-7 Three levels of unrooted trees for five taxa. (From Saitou, 1988.)

considers trees with a bifurcating node and a trifurcating node. There are
ten such trees, three of which are shown in Figure 12-7. One computes
the ML values for all ten trees and chooses the one with the highest ML
value. Suppose it is tree 2a. This tree can be resolved into three bifurcating
trees, trees 3a, 3b, and 3c (Fig. 12-7). One then computes the ML values
for all three trees and chooses the tree with the highest ML value as the
final tree. This search is not exhaustive, but is likely to lead to the ML
tree. After the ML tree is inferred, we can test the significance of each
internal branch. For example, to test the significance of the second internal
branch of tree 3a, we note that when this branch becomes 0, tree 3a reduces
to tree 2a. Since tree 2a is a submodel of tree 3a, we can test whether tree
3a is significantly better than tree 2a by the •£• test described in the preceding
paragraph.
Instead of testing whether a particular tree is significant, we may test
some plausible alternative trees against each other. Two approaches have
272 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

been proposed. First, Felsenstein (1987, 1988) suggests in the case where
the two topologies are adjacent (i.e., they each have a branch which, if its
length is shrunk to 0, results in the same trifurcation) that we could test
one topology against the other by pretending that there is one degree of
freedom. For example, trees 3a and 3b in Figure 12-7 are two adjacent
trees because they both reduce to tree 2a when the second internal branch
is shrunk to 0. Three situations can occur in testing tree 3a against tree
3b. The first situation is that both trees assume an ML value when the
length of the second internal branch is 0. In this case, both trees are not
plausible. The second situation is that one tree, say tree 3a, assumes an
ML value when the second internal branch is greater than 0, while tree 3b
assumes an ML value when the second internal branch is 0. In this case,
testing tree 3a against tree 3b is equivalent to testing the significance of
the second internal branch of tree 3a or testing tree 3a against tree 2b.
The third situation is that both trees 3a and 3b assume their ML value
when the second internal branch is greater than 0. In this case, the testing
is more stringent than testing the significance of the second internal branch
in either tree 3a or 3b. The third situation should arise rarely, because at
least one of the two trees is erroneous and would have a low probability
of having a positive second internal branch. As noted above, Felsenstein's
approach has a logical problem because the ML values for both trees are
estimated with the same number of parameters and under different hy-
potheses. However, the test tends to be conservative.
Second, Kishino and Hasegawa (1989) propose to test one tree against
another by computing the mean and variance of the difference in the ML
value between two trees. Assume that all nucleotide sites in the sequence
are independent and equivalent. Consider trees 1 and 2. Let L,(0,|X) be
the likelihood function of tree i, and ^(0,|X) be the logarithm of the ML
value for tree i, where 0 is the maximum likelihood estimator of 0,. Then
the difference in mean is given by f^O^X) — €2(02|X). Kishino and Has-
egawa (1989) propose to estimate the variance by the sample variance,
that is,

vfhere fi(Xj\8,) is given by equation (40). If the sample size A' is large, then
the standard normal test may be used. Note that like Felsenstein's ap-
proach, this one is also heuristic because the two likelihood functions are
not constructed from the same topology (i.e., are not from the same prob-
ability space). However, an application of this method to the DNA se-
quence data from apes and man seems to give reasonable results (Kishino
and Hasegawa, 1989).
Finally, it should be noted that all the above tests are based on the large-
sample theory and so a large amount of data is required.
STATISTICAL METHODS FOR TESTING MOLECULAR PHYLOGENIES 273

DISCUSSION

Problem of Statistical Inconsistency

In phylogenetic reconstruction, as well as in any other statistical inference
problems, it is important for a method to be statistically consistent in the
sense that the probability for the method to give the correct tree increases
with the amount of data. When the degree of sequence divergence is small
or moderate (e.g., 50% or lower), the maximum parsimony method is
likely to be consistent. However, when the degree of divergence is high,
and when the assumption of rate-constancy is violated, it may not be
consistent (Felsenstein, 1978; Li et al., 1987b). It should be noted that a
statistical test does not correct the inconsistency problem inherent in the
tree-reconstruction method used, and so a strongly supported tree may in
fact be an erroneous one. The evolutionary parsimony (EP) method was
intended to overcome the effects of unequal rates of evolution. However,
the EP method can also become statistically inconsistent if the transver-
sional rates are unequal (Jin and Nei, 1990). Indeed, in an application of
this method to the small-subunit rRNA sequences from human, Droso-
phila, rice, and Physarum, Gouy and Li (1989) obtained a x2 of 4.8 for
the tree with human and rice in a clade, the ^'s for the other two trees
being 0.6 and 0.3. The favored tree is obviously erroneous as humans are
much more closely related to Drosophila than to rice.

Heterogeneous Data
Phylogenetic studies often use sequence data from different DNA regions.
If all the regions studied have similar rates of nucleotide substitution, then
all the data can be combined together into a single set. However, if con-
siderable variation in rates exist, regions with different rates should be
treated separately, particularly when the distance matrix approach is used,
because the internodal distances are dependent on the rate of evolution.
The question then arises as to how to test the significance when the results
from different data sets are combined. A simple test procedure is the
inverse x2 method (Fisher, 1932). Suppose that there are k different data
sets. Let p, be the significance level (probability) estimated from the /-th
data set. If the null hypothesis is true, then -2€«(/?,) has a x1 distribution
with two degrees of freedom and

has a \2 distribution with 2k degrees of freedom.

In the case of likelihood tests, it is easy to combine results from different
regions because the log ML values from different regions can be added.
A more difficult problem arises when the substitution rate is uneven
over a sequence, but it is unclear how to divide the sequence into separate
274 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

regions. Such a situation would occur often in coding sequences. The neigh-
bors relation tests proposed above would be robust against rate hetero-
geneity because the variances are estimated from sample variances, which
should include variation due to heterogeneous rates. The tests based on
the variance of internal branch lengths would be less robust because the
variances are computed under the assumption of a homogeneous sequence
and so would tend to be underestimates. In the likelihood test proposed
by Kishino and Hasegawa (1989), the variances are estimated from sample
variances and so the test may be robust, but it is not clear whether the
other likelihood tests discussed above are robust. It has often been argued
that parsimony methods are robust against rate heterogeneity. This may
be true when the degree of sequence divergence is small or moderate, but
may not be true for divergent sequences. Since highly divergent sites can
lead to inconsistent results, they should be given less weight in both phy-
logenetic reconstruction and statistical tests.

Strengths and Weaknesses of Current Tests

To date, no study seems to have been made that compares the relative
strengths and weaknesses of different tests. Therefore, our discussion below
involves intuitive arguments and speculations. A rigorous comparison should
be made in the future by using both computer simulation and empirical
data.
The parsimony tests developed by Cavender (1978) and Felsenstein (1985)
are not difficult to compute. However, since Cavender's test assumes ex-
tremely unfavorable conditions, it has low power. Felsenstein's tests assume
an equal probability (i.e., Vs) for an informative site to support each of
the three possible unrooted trees. This assumption is close to reality if the
branching order for the three taxa under study is close to a trichotomy,
and if the rate of evolution is approximately the same among the three
lineages. In this situation, the tests would perhaps perform better than the
tests based on distance methods or likelihood methods because the con-
fidence level can be calculated more precisely than in the other tests.
However, if the central branch (c in Figure 12-3a) is substantially long,
Felsenstein's assumption would not hold well, and his tests may not perform
as well as the tests based on distance methods or likelihood methods.
Parsimony tests are difficult to develop when the number of taxa involved
is larger than four. The performance of the variance test by Kishino and
Hasegawa (1989) has not been studied. This test is computationally difficult
when the number of taxa is large.
The test based on the variance of internal branch lengths is rather easy
to compute when the number of taxa under study is five or fewer. This
test would tend to be more powerful than the neighbors relation test be-
cause it is formulated under more specific assumptions on the model of
nucleotide substitution and because it takes better account of the corre-
lation between sequences. However, since it is dependent on more as-
STATISTICAL METHODS FOR TESTING MOLECULAR PHYLOGENIES 275

sumptions, it is less robust than the neighbors relation test. When the
number of taxa becomes more than five, the former test becomes com-
putationally tedious. Moreover, the estimates of mean and variance of a
branch length depends on the assumption of the tree topology. For ex-
ample, the mean and variance of the first internal branch in Figure 12-4
would not be accurate if the true branching order of OTUs 4, 5, and 6 is
different from that shown in the figure. These problems are less serious
(or do not arise) if we use the neighbors relation test.
The maximum likelihood is a widely used statistical method. Its use in
phylogenetic reconstruction can be justified, although there may be some
logical problems as discussed above. The major problem with this approach
is computational difficulties. When the number of taxa under study is five
or fewer, the computation is not too tedious and one can even use a
sophisticated model of nucleotide substitution. Unfortunately, as the num-
ber of taxa increases, the computation becomes prohibitively tedious. If
the computational difficulties can be overcome, then the likelihood ap-
proach should enjoy more attention than its current status.

ACKNOWLEDGMENTS

We thank N. Saitou for sending us his figures. This study was supported
by NIH grant GM30998.

REFERENCES

Akaike, H. (1974) A new look at the statistical model identification. IEEE Trans.
Autom. Contr. AC-19-.716-723.
Akaike, H. (1985) Prediction and entropy. Pp. 1-24 in A Celebration of Statistics
(A.C. Atkinson and S.E. Fienberg, eds.). Springer-Verlag, New York.
Brown, W.M., E.M. Prager, A. Wang, and A.C. Wilson. (1982) Mitochondrial
DNA sequences of primates: tempo and mode of evolution. J. Mol. Evol.
18:225-239.
Buneman, P. (1971) The recovery of trees from measurements of dissimilarity. Pp.
387-395 in Mathematics in the Archeological and Historical Sciences (F.R.
Hodson, D.G. Kendall, and P. Tautu, eds.). Edinburgh University Press,
Edinburgh.
Cavalli-Sforza, L.L., and A.W.F. Edwards. (1967) Phylogenetic analysis models
and estimation procedures. Am. J. Hum. Genet. 19:233-257.
Cavender, J.A. (1978) Taxonomy with confidence. Math. Biosci. 40:271-280.
Cavender, J.A. (1981) Tests of phylogenetic alternatives under generalized models.
Math. Biosci. 54:217-229.
Chakraborty, R. (1977) Estimation of time of divergence from phylogenetic studies.
Can. J. Genet. Cytol. 19:217-223.
Edwards, A.W.F., andL.L. Cavalli-Sforza. (1964) Reconstruction of evolutionary
trees. Pp. 67-76 in Phenetic and Phylogenetic Classification (V.H. Heywood
and J. McNeill, eds.). Syst. Assn. Publ. No. 6., Systematics Association,
London.
276 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Farris, J.S. (1977) On the phenetic approach to vertebrate classification. Pp. 823-
850 in Major Patterns in Vertebrate Evolution (M.K. Hecht, P.C. Goody,
and B.M. Hecht, eds.). Plenum Press, New York.
Felsenstein, J. (1973) Maximum-likelihood and minimum-steps methods for esti-
mating evolutionary trees from data on discrete characters. Syst. Zoo/. 22:240-
249.
Felsenstein, J. (1978) Cases in which parsimony or compatibility methods will be
positively misleading. Syst. Zool. 27:401-410.
Felsenstein, J. (1981) Evolutionary trees from DNA sequences: a maximum like-
lihood approach. /. Mol. Evol. 17:368-376.
Felsenstein, J. (1983) Inferring evolutionary trees from DNA sequences. Pp. 133-
150 in Statistical Analysis of DNA Sequence Data (B.S. Weir, ed.). Dekker,
New York.
Felsenstein, J. (1985) Confidence limits on phylogenies with a molecular clock.
Syst. Zool. 34:152-161.
Felsenstein, J. (1987) Estimation of hominoid phylogeny from a DNA hybridization
data set. J. Mol. Evol. 26:123-131.
Felsenstein, J. (1988) Phylogenies from molecular sequences: inference and reli-
ability. Ann. Rev. Genet. 22:521-565.
Fisher, R.A. (1932) Statistical Methods for Research Workers. Oliver and Boyd,
London.
Fitch, W.M. (1981) A non-sequential method for constructing trees and hierarchical
classifications. J. Mol. Evol. 18:30-37.
Gouy, M., and W.-H. Li (1989) Phylogenetic analysis based on rRNA sequences
supports the archaebacterial rather than the eocyte tree. Nature 339:145-
147.
Hasegawa, M., H. Kishino, and T. Yano. (1985) Dating of the human-ape splitting
by a molecular clock of mitochondrial DNA. /. Mol. Evol. 22:160-174.
Hasegawa, M., H. Kishino, and T. Yano. (1987) Man's place in Hominoidea as
inferred from molecular clocks of DNA. J. Mol. Evol. 26:132-147.
Jin, L., and M. Nei. (1990) Limitations of the evolutionary parsimony method of
phylogenetic analysis. Mol. Biol. Evol. 7:82-102.
Jukes, T.H., and C.R. Cantor. (1969) Evolution of protein molecules. Pp. 21-132
in Mammalian Protein Metabolism (H. N. Munro, ed.). Academic Press,
New York.
Kashyap, R.L., and S. Subas. (1974) Statistical estimation of parameters in a
phylogenetic tree using a dynamic model of the substitutional process. J.
Theor. Biol. 47:75-101.
Kimura, M. (1980) A simple method for estimating evolutionary rates of base
substitutions through comparative studies of nucleotide sequences. J. Mol.
Evol. 16:111-120.
Kimura, M., and T. Ohta. (1972) On the stochastic model for estimation of mu-
tational distance between homologous proteins. J. Mol. Evol. 2:87-90.
Kishino, H., and M. Hasegawa. (1989) Evaluation of the maximum likelihood
estimate of the evolutionary tree topologies from DNA sequence data, and
the branching order in Hominoidea. /. Mol. Evol. 29:170-179.
Klotz, L.C., N. Komar, R.L. Blanken, and R.M. Mitchell. (1979) Calculation of
evolutionary trees from sequence data. Proc. Natl. Acad. Sci. USA 76:4516-
4520.
Li, W.-H. (1981) Simple method for constructing phylogenetic trees from distance
matrices. Proc. Natl. Acad. Sci. USA 78:1085-1089.
STATISTICAL METHODS FOR TESTING MOLECULAR PHYLOGENIES 277

Li, W.-H. (1989) A statistical test of phylogenies estimated from sequence data.
Mol. Biol. Evol. 6:424-435.
Li, W.-H., and M. Gouy. (1990) Statistical tests of molecular phylogenies. Pp.
645-659 in Molecular Evolution: Computer Analysis of Protein and Nucleic
Acid Sequences (R.F. Doolittle, ed.). Methods in Enzymology, vol. 183.
Academic Press, New York.
Li, W.-H., M. Tanimura, and P.M. Sharp. (1987a) An evaluation of the molecular
clock hypothesis using mammalian DNA sequences. J. Mol. Evol. 25:330-
342.
Li, W.-H., K.H. Wolfe, J. Sourdis, and P.M. Sharp. (1987b) Reconstruction of
phylogenetic trees and estimation of divergence times under nonconstant
rates of evolution. Cold Spring Harbor Symp. Quant. Biol. 52:847-856.
Li, W.-H., C.-I. Wu, and C.-C. Luo. (1984) Nonrandomness of point mutation as
reflected in nucleotide substitutions in pseudogenes and its evolutionary
implications. /. Mol. Evol. 21:58-71.
Mueller, L.D., and F.J. Ayala. (1982) Estimation and interpretation of genetic
distance in empirical studies. Genet. Res. 40:127-137.
Nei, M. (1987) Molecular Evolutionary Genetics. Columbia University Press, New
York.
Nei, M., J. C. Stephens, and N. Saitou. (1985) Methods for computing the standard
errors of branching points in an evolutionary tree and their application to
molecular data from humans and apes. Mol. Biol. Evol. 2:66-85.
Neyman, J. (1971) Molecular studies of evolution: a source of novel statistical
problems. Pp. 1-27 in Statistical Decision Theory and Related Topics (S.S.
Gupta and J. Yackel, eds.). Academic Press, New York.
Saitou, N. (1988) Property and efficiency of the maximum likelihood method for
molecular phylogeny. /. Mol. Evol. 27:261-273.
Saitou, N. (1990) Maximum likelihood methods. Pp. 584-598 in Molecular Evo-
lution: Computer Analysis of Protein and Nucleic Acid Sequences (R.F.
Doolittle, ed.). Methods in Enzymology, vol. 183. Academic Press, New
York.
Saitou, N., and T. Imanishi. (1989) Relative efficiencies of the Fitch-Margoliash,
maximum-parsimony, maximum-likelihood, minimum-evolution, and neigh-
bor-joining methods of phylogenetic tree construction in obtaining the cor-
rect tree. Mol. Biol. Evol. 6:514-525.
Saitou, N., and M. Nei. (1986) The number of nucleotides required to determine
the branching order of three species with special reference to the human-
chimpanzee-gorilla divergence. J. Mol. Evol. 24:189-204.
Saitou, N. and M. Nei. (1987) The neighbor-joining method: a new method for
reconstructing phylogenetic trees. Mol. Biol. Evol. 4:406-425.
Sattath, S., and A. Tversky. (1977) Additive similarity trees. Psychometrika 42:319-
345.
Smouse, P.E., and W.-H. Li. (1987) Likelihood analysis of mitochondrial restric-
tion-cleavage patterns for the human-chimpanzee-gorilla trichotomy. Evo-
lution 41:1162-1176.
Sneath, P.H.A., and R.R. Sokal. (1973) Numerical Taxonomy. W.H. Freeman,
San Francisco.
Templeton, A.R. (1983) Phylogenetic inference from restriction endonuclease
cleavage site maps with particular reference to the evolution of humans and
the apes. Evolution 37:221-244.
 13
Discriminating Between Phylogenetic
Signal and Random Noise in DNA
Sequences

DAVID M. HILLIS

Most parsimonious trees are really neat. I am all for them, but have you any
idea about the distribution of all the trees in the universe that you might have
sampled? Fitch (1984)
Nucleic acid sequences have become a major source of phylogenetic in-
formation. In order to use these data appropriately, it is critical to distin-
guish sequences that are saturated by change from those that are phylo-
genetically informative. In this chapter, I will call variation that is potentially
informative about phylogenetic history "signal," and use "noise" to de-
scribe variation that is not informative. Distinguishing between signal and
noise is not something that can be accomplished by use of alignment cri-
teria, because two sequences may be identical at a large number of func-
tionally constrained sites, and yet essentially randomized at sites where
variation is tolerated (e.g., the third positions of codons). This chapter
explores the use of the distributions of tree lengths as a guide to the
detection of phylogenetic signal in comparative data sets.
In parsimony analysis, changes at nucleotide positions among aligned
sequences are mapped onto a tree, and the number of evolutionary changes
required to accommodate that tree with the data is calculated as the tree
length. For any given data set, this procedure may be repeated for many
thousands of trees; the tree that yields the shortest length is designated
the optimal tree under the parsimony criterion. The optimal tree is thus
the one that requires the fewest number of evolutionary changes. In this
chapter, I argue that the shape of the distribution of tree lengths contains
information useful in deciding whether or not the data set contains phy-
logenetic signal.

278
DISCRIMINATION BETWEEN PHYLOGENETIC SIGNAL AND NOISE 279

BACKGROUND

Fitch (1979) and Goodman et al. (1979a) were among the first authors to
explore tree-length distributions in a phylogenetic context. These authors
compared the relative phylogenetic information in two protein sequence
data sets that had been used to infer relationships among orders of mam-
mals. As can be seen in Figure 13-1, the distribution of tree lengths from
the ot-hemoglobin sequences is nearly perfectly symmetrical, whereas that
from the a-crystallin sequences is strongly skewed with a long left tail.
Therefore, there are many fewer solutions near the optimal (most-parsi-
monious) solution for the a-crystallin data set than for the a-hemoglobin
data set. The shape of these tree-length distributions suggests that the a-

Figure 13-1 Tree-length distributions compared by Fitch (1979, 1984) and Good-
man et al. (1979a).
280 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

crystallin data set is more informative (because it allows better discrimi-

nation among near-optimal solutions) than the a-hemoglobin data set.
Fitch (1984) suggested that symmetric distributions such as that for a-
hemoglobin are "most likely for bushy trees where all lines emanate from
a single point," whereas asymmetric distributions like that for a-crystallin
are "most likely for the 'stringy' trees, those which are maximally asym-
metric in their splitting and have long intervals between splits." Several
authors have presented strongly skewed tree-length distributions and have
suggested that such distributions are evidence for strong phylogenetic signal
(e.g., Fitch, 1984; Goodman et al., 1979a; Hillis, 1985; Werman, 1986;
Hillis and de Sa, 1988; Hillis and Dixon, 1989). I will argue that skewness
of tree-length distributions can be a useful indicator of phylogenetic signal
and is largely (but not completely) independent of tree topology.
Having looked at tree-length distributions for many data sets, I devel-
oped the view illustrated in Figure 13-2. If all tree topologies for a given
set of sequences are equally optimal (i.e., of equal length), then there is
neither signal nor noise in the data set. Such a data set would result in the
completely unresolved bush shown in Figure 13-2a. If the real phylogeny
is in the form of a bush, however, then an equal length for all possible
trees is an unlikely outcome; random variation would result in some trees
being shorter than others by chance alone. The distribution of trees would
include some topologies that were shorter than average, and a similar
number that were longer than average, so the distribution of these lengths
would produce a nearly symmetrical distribution, such as those shown in
Figure 13-2b and 13-2b'. The optimal topology (or topologies) in this case
might be a completely symmetrical tree (Fig. 13-2b), a completely asym-
metrical tree (Fig. 13-2b'), or something in between. However, if the true
phylogeny was not a single polytomy and the sequences were constrained
by history (i.e., contain phylogenetic information), then the tree-length
distributions shown in Figures 13-2b and 13-2b' are highly unlikely. Se-
quences constrained by history would produce distributions of tree lengths
similar to those in Figure 13-2c and 13-2c': highly asymmetrical distribu-
tions with few trees near the optimal solution. These asymmetric distri-
butions are a consequence of congruence among characters as a result of
a common phylogenetic history. Thus, many characters support the to-
pology that reflects this history, and conflict with a large number of alter-
native trees.
If the view described above is correct, then it suggests a means of escaping
a common problem in phylogenetic analysis. Any comparative data set can
be subjected to phylogenetic analysis, even if the data contain no historical
information (i.e., they are too noisy for meaningful phylogenetic analysis).
Nonetheless, random variation will usually result in one or a relatively few
optimal topologies (at least compared to the much larger number of pos-
sibilities) . What is needed is a means of distinguishing such data sets from
those that are informative about phylogenetic history. This chapter explores
the causes of skewed tree-length distributions, examines the proposition
DISCRIMINATION BETWEEN PHYLOGENETIC SIGNAL AND NOISE 281

Figure 13-2 Hypothesis of the behavior of tree-length distributions, under several

model phylogenies with differing amounts of signal and noise. See text for expla-
nation.

that random sequence data produce symmetrical tree-length distributions,

explores possible tests for skewness, evaluates the amount of signal nec-
essary to produce significant skewness, compares real data sets to random
simulations, suggests a possible "stopping algorithm" to prevent the
over-resolution of phylogenies, and points to areas in need of additional
research.

THE CAUSES OF SKEWED TREE-LENGTH DISTRIBUTIONS

To understand why data sets with strong phylogenetic signal produce skewed
tree-length distributions, consider a data set on eight taxa (Figs. 13-3-13-
282 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Figure 13-3 Classes of binary characters for eight taxa and their effects on tree-
length distributions.

5).1 For eight taxa, there are only three types of binary characters that
affect tree length (Fig. 13-3): those for which two taxa have one state and
the others have the alternative state ("two/six characters"); those for which
the alternative states are distributed between three and five taxa, respec-
tively ("three/five characters"); and those for which the alternative states
are evenly divided between two subsets of four taxa each ("four/four char-
]
Because tree lengths are unaffected by the rooting point of the tree, all examples in this
discussion will concern unrooted trees. Also, for simplicity, examples in this section will
involve ditypic (binary) characters, although the concepts can be extended easily to multistate
characters.
DISCRIMINATION BETWEEN PHYLOGENETIC SIGNAL AND NOISE 283

acters"). Characters with states limited to a single taxon and characters

that are uniform among all species are compatible with all possible tree
topologies and thus do not affect the shape of the tree-length distribution.
Figure 13-3 shows the distributions of tree lengths for data sets limited
to each of these types of characters. All three types of characters produce
a tree-length distribution with a left-hand skew. Two/six characters produce
distributions with the strongest skew but the smallest range; this is because
only a very few trees connect the pair of species with the unique state,
whereas all other trees require the character-state to evolve twice. Three/
five characters also produce a skewed distribution, but more of the possible
trees are compatible or partially compatible with this type of character.
Although few trees unite the three species with the same state, many more
unite at least two of the three species on the tree, and most trees split all
three species. This produces a skewed distribution with a range of two
steps (compared to a range of one step for a two/six character; see Fig.
13-3). The four/four characters produce the distribution with the greatest
span but the least skewness, because there are fewer trees that unite none
of the similar taxa than unite only two of the four similar taxa (Fig. 13-3).
Therefore, any informative character can add to the skewness of a tree-
length distribution, although different types of characters do not contribute
to skewness equally.
The effects of tree topology on tree-length distributions can be assessed
by considering the kinds of characters that would be produced by the
different topologies. For eight taxa, there are only four unrooted, unlabeled
topologies (Fig. 13-4). Each internal branch in the topology is supported
by either a two/six, a three/five, or a four/four character; the numbers of
each of these types of internal branches are shown in Figure 13-4. If one
assigns two characters to each internal branch, the distributions for each
topology would be as shown in Figure 13-5. Note that all of these distri-
butions are strongly skewed, although not all to the same degree. A com-
mon measure of skewness is the gl statistic—the third central moment
divided by the cube of the standard deviation (Sokal and Rohlf, 1981).
For n trees of length T, gl is calculated as

where s is the standard deviation of the tree lengths. This statistic is negative
for distributions with left-skew, 0 for symmetric distributions, and positive
for distributions with right-skew. The skewness of the distributions in Fig-
ure 13-5 varies from gt = -1.096 to g1 = -0.6637. Therefore, although
tree topology has a quantitative influence on skewness, the qualitative
result (strong left-skew) is the same for all topologies. It is interesting to
note that symmetry of the tree topology appears to have little influence
on skewness.
284 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Figure 13-4 Distinct unrooted tree topologies for eight taxa, numbers of internal
branches supported by each class of binary character, and the relative spans of the
tree-length distributions.

EXPERIMENTS WITH RANDOM SEQUENCES

To address the behavior of tree-length distributions in the absence of phy-

logenetic signal, random data matrices were constructed in the following
manner (Fig. 13-6). One hundred matrices each with six, seven, or eight
sequences, each 100 nucleotides long, were created using the random data
generator in the computer program MacClade (version 2.97; Maddison and
Maddison, 1991). This version of the program uses the random number
generator in the Macintosh tool box to generate random data (D. Mad-
dison, personal communication). The four nucleotides were given equal
probabilities of occurring, so the frequency of each nucleotide was ap-
proximately 0.25 in each sequence. All possible tree topologies were an-
alyzed for all 300 random matrices using Swofford's (1990) Phylogenetic
Analysis Using Parsimony computer software package (PAUP version 3.0).
To address the effects of the length of sequences on skewness, two
additional sets of 100 matrices each of random data were generated for
eight taxa. These data sets consisted of sequences 30 positions long and
200 positions long, respectively.
Two examples of tree-length distributions from the eight-taxa data sets
are shown in Figure 13-7; figured are the distributions with the strongest
left-hand and right-hand skews, respectively. The distribution of gt scores
for the tree-length distributions from all the random data sets are shown
DISCRIMINATION BETWEEN PHYLOGENETIC SIGNAL AND NOISE 285

Figure 1 3-5 Effect of tree topology on tree-length distributions. Each internal branch
is assigned a length of two character changes.
286 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

generate random
sequence data
(6 X 100, 7 X 100, 8 X 100,
8 X 30, 8 X 200)

I
examine all possible trees
6 taxa: 105 topologies repeat 100
7 taxa 945 topologies times
8 taxa: 10,395 topologies

I
calculate g for each
distribution Figure 13-6 Protocol for the gener-
ation of random sequence-data mat-
rices and analysis of skewness.

in Figure 13-8. Two trends are clear from these scores. First, the mean
and modal values of g1 across data sets are slightly negative, as predicted
for binary data by Le Quesne (1989). Even with random data, approxi-
mately twice as many tree-length distributions are skewed slightly to the
left as to the right. Second, the average value of the gl scores approaches
0 with increasing number of taxa: the tree-length distributions become
closer to symmetrical.
Although the only published tests for skewness of tree-length distribu-
tions have used the normal distribution as a null model (e.g., Werman,
1986), the random data simulations indicate that this is an inappropriate
test (Fig. 13-9). Only a small percentage of tree-length distributions based
on random sequences fall into the 95% confidence limits of a normal
distribution. The percentage of distributions that fall within the 95% con-
fidence limits for skewness of a normal distribution drops with increasing
number of taxa, despite the fact that the distributions become more sym-
metrical (Fig. 13-9). This is because the degree of expected departures
from symmetry for a normal distribution decreases with increasing sample
size, and the sample size (number of tree topologies) increases very rapidly
with an increase in the number of taxa (Felsenstein, 1978). Thus, less than
20% of the tree-length distributions produced from random sequences fall
within the expectations of a normal distribution with just eight taxa (Fig.
13-9).
The random data simulations provide a simple means for testing tree-
length distributions for significantly greater left-handed skewness than would
be expected from random data, without making any a priori assumptions
about the shape of the distribution of expected g1 values. The lower 5%
and 1% of the g1 distributions, shown in Figure 13-8, can be used to obtain
DISCRIMINATION BETWEEN PHYLOGENETIC SIGNAL AND NOISE 287

Figure 13-7 Examples of tree-length distributions from analysis of random se-

quence-data matrices for eight taxa. Shown are the most right-skewed (a; g, =
0.3055) and the most left-skewed (b; g, = —0.5278) distributions produced from
100 random matrices.

critical values of g1 for a given number of taxa. A test employing these

values (Table 13-1) requires a minimum of 100 variable sequence positions.
For fewer than 100 positions, the critical values are slightly lower (more
negative); there is very little change in the critical value for data sets with
more than 100 sequence positions (Fig. 13-10).
How much phylogenetic signal is needed to produce significant skew-
ness? To examine this question, I added characters consistent with the
branches of a single tree (signal) to the eight-taxa random matrices (noise).
For most data sets, 10% signal (e.g., 11 characters out of 110) was sufficient
to produce tree-length distributions that were significantly skewed, as in-
dicated by the test discussed above (see Fig. 13-11). The most right-skewed
distribution examined (gl = 0.3055) required 20% signal to become sig-
nificantly left-skewed (gl = -0.4037). Therefore, the test appears to be
fairly sensitive; even small percentages of phylogenetic signal can be de-
tected in otherwise random data sets.
288 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Figure 13-8 The distributions of skewness statistics (g,) for the 300 random se-
quence matrices (six, seven, and eight taxa). The dashed vertical lines indicate the
g, score for symmetrical distributions. The lower 5% (shaded) and 1% (black) of
each distribution is indicated.

Figure 13-9 The proportions (shaded area) of skewness scores for the random
sequence-data matrices that are not significantly different from scores for normal
distributions (six, seven, and eight taxa).

Table 13-1 Critical Values for g, Statistics of Tree-Length Distributions for Six,
Seven, and Eight Taxa*
Number of Taxa
P 6 7 8
0.05 -0.51 -0.45 -0.34
0.01 -0.67 -0.60 -0.47
'Scores lower than those shown are outside of the 95% or 99% confidence limits for distributions derived
from random data.
DISCRIMINATION BETWEEN PHYLOGENETIC SIGNAL AND NOISE 289

Figure 13-10 The effect of length of

sequence on g, statistics. The three
samples each contain 100 matrices of
eight taxa; the lengths of the random
sequences are 30, 100, and 200 nu-
cleotides. The lowest 1 % of the distri-
bution is indicated by a single line, the
lowest 5% of the distribution is indi-
cated by an open bar plus the line, and
the remaining portion of the distribu-
tion is indicated by a solid bar.

ANALYSIS OF REAL SEQUENCES

I have examined skewness of tree-length distributions from several real

sequence data sets, all with at least eight taxa. Except for the a-hemoglobin
data discussed by Fitch (1984) and above, all of these were significantly
more skewed than the distributions from random data matrices (p < 0.05;
Table 13-1 and Fig. 13-12). Only one of these distributions (that for mi-
tochondrial cytochrome b genes) was even within the observed bounds of
the gl values for the random matrices. The mitochondrial cytochrome b
data set consists of sequences on eight species of vertebrates, including an
actinopterygian fish, two amphibians, a bird, and four mammals. If one
of the two most closely related taxa (sheep and cow) are removed from

Figure 13-11 The effect on skew-

ness of adding signal to otherwise
random sequence matrices (see
text). The g, scores represented by
open boxes in the lower portion of
the figure correspond to distribu-
tions produced from random data
(the most left-skewed, symmetrical,
and right-skewed distributions are
used as examples). The shaded
boxes indicate the g, scores for the
distributions after the addition of
10% or 20% signal, as indicated.
290 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Figure 13-12 Skewness (g,) scores

for tree-length distributions pro-
duced from real data matrices
(lower) compared to the distribu-
tions of g, scores produced from
analysis of random data matrices
(upper) for eight taxa. The rRNA
data on deuterostomes (taxa shown
in Fig. 13-13) and metazoans (rep-
resentatives of eight phyla) are from
Field et al. (1988); the lens a-crys-
tallin data on eight orders of mam-
mals are from De Jong et al. (1977);
the a-hemoglobin data on eight or-
ders of mammals are from Good-
man et al. (1979b); the anuran and
salamander data are from Hillis
(1991); the ribosomal DNA data on
tetrapods are from Hillis et al.
(1991); and the mitochondrial cyto-
chrome b sequences are eight spe-
cies of vertebrates from Kocher and
White (1989) and GenBank
(IntelliGenetics, Mountain View,
CA).

this data matrix, the tree-length distribution becomes much less skewed,
indicating that it is the inclusion of these relatively closely related species
that accounts for most of the phylogenetic signal in these sequences.
The observations from the mitochondrial cytochrome b data set suggest
an algorithm that could be used to prevent over-resolution of phytogenies
(reading beyond the signal) in sequence data sets. The execution of this
algorithm is shown in Figure 13-13 for the ribosomal RNA (rRNA) se-
quences presented by Field et al. (1988) for eight species of deuterostomes.
1. Determine the skewness of the tree-length distribution for all possible
trees (or a random sample thereof—MacClade and PAUP each have
options for this purpose).
2. If the distribution is significantly skewed, then find the best-supported
branch that unites two or more of the taxa (e.g., by bootstrapping
or counting synapomorphies) and continue to step 3. If the distri-
bution is no more skewed than would be expected from random data,
then stop.
3. Repeat step 1 for all remaining tree topologies, given the branch
determined in step 2.
4. Repeat step 2, saving any branches previously determined (continue
repeating steps 1 and 2 until the distribution of all remaining trees is
no more skewed than expected from random data).
DISCRIMINATION BETWEEN PHYLOGENETIC SIGNAL AND NOISE 291

Figure 13-13 Resolution of relationships among eight species of deuterostomes

(Field et al., 1988) using the "stopping algorithm" described in the text.

One objection to this algorithm is that tree resolution cannot proceed

beyond a tree of five unresolved lineages (Fig. 13-13), because the number
of possible tree topologies (15) is too small to examine skewness (see Table
13-2). However, I have found that for most sequence data sets, the tree-
length distribution is no longer skewed well before this point; therefore,
this limitation may not be too severe. Another objection is that the overall
probability of making a type-I error increases as the number of resolved
branches increases, so the initial level of a must be set quite low in order
to keep the overall a below 0.05. This reduces the power of the test. There
also may be objections that a single branch is determined at each step,
whereas the phylogenetic information may be distributed across more than
one branch. Nonetheless, I present this algorithm as an example of the
possible uses of tree-length distributions, although I do not advocate its
use until tree-length distributions have been studied in greater detail.
292 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Table 13-2 Number of Distinct Unrooted Trees for Three or More Taxa, and the
Recommended Methods of Analyzing Tree-Length Distributions
Number of Taxa Number of Unrooted Trees Method of Analysis
3 1 Too few trees
4 3 Too few trees
5 15 Too few trees
6 105 Exact enumeration
7 945 Exact enumeration
8 10,395 Exact enumeration
9 135,135 Exact enumeration or
random sampling
10 2,027,025 Exact enumeration or
random sampling
11 34,459,425 Exact enumeration or
random sampling
>12 £654,729,075 Random sampling

IMPLEMENTATION AND DIRECTIONS FOR FUTURE RESEARCH

This study suggests that tree-length distributions can provide an accurate

and sensitive indication of the presence of phylogenetic signal in compar-
ative sequence data sets. Skewness statistics provide a simple means of
evaluating these distributions, which are often determined as a by-product
of phylogenetic analysis. In response to these findings, David L. Swofford
and David R. Madison have recently modified PAUP and MacClade, re-
spectively, so that these programs now provide gl statistics for tree-length
distributions.
Although I have used the conventional measure of skewness (g:) as a
guide to the detection of phylogenetic signal, other measures (e.g., the
other odd central moments) may prove to be even better indicators of
signal. However, considerable additional research needs to be conducted
before any measures of skewness can be used meaningfully with regularity.
Simulations with random data (or exact solutions) need to be conducted
for greater numbers of taxa (Table 13-2), the effects of base composition
and length of sequence need to be studied in greater detail, and additional
studies need to be conducted on the effects of different levels of signal and
noise (or the effects of two sets of conflicting signal). Despite the need for
these additional studies, this study indicates the general usefulness of tree-
length distributions in phylogenetic analysis.

ACKNOWLEDGMENTS

I thank J. Coddington, B. I. Crother, M. T. Dixon, J. S. Farris, J. Fel-

senstein, K. Halanych, A. I. Hillis, W. M. Fitch, and M. White for helpful
discussion, comments on the manuscript, and other assistance. D. R. Mad-
dison and D. L. Swofford provided test versions of their respective pro-
DISCRIMINATION BETWEEN PHYLOGENETIC SIGNAL AND NOISE 293

grams, without which this study would not have been possible. J. J. Bull
has provided constant input and advice on this work, and I am especially
grateful for his efforts. This study was supported by National Science Foun-
dation grants BSR-8657640 and BSR-8796293.

REFERENCES

De Jong, W. W., J. T. Gleaves, and D. Boulter. (1977) Evolutionary changes of

a-crystallin and the phylogeny of mammalian orders. J. Mol. Evol. 10:123-
135.
Felsenstein, J. (1978) The number of evolutionary trees. Syst. Zool. 27:27-33.
Field, K. G., G. J. Olsen, D. J. Lane, S. J. Giovannoni, M. T. Ghiselin, E. C.
Raff, N. R. Pace, and R. A. Raff. (1988) Molecular phylogeny of the animal
kingdom. Science 239:748-753.
Fitch, W. M. (1979) Cautionary remarks on using gene expression events in par-
simony procedures. Syst. Zool. 28:375-379.
Fitch W. M. (1984) Cladistic and other methods: problems, pitfalls, and potentials.
Pp. 221-252 in Cladistics: Perspectives on the Reconstruction of Evolutionary
History (T. Duncan and T. F. Stuessy, eds.). Columbia University Press,
New York.
Goodman, M., J. Czelusniak, and G. W. Moore. (1979a) Further remarks on the
parameter of gene duplication and expression events in parsimony recon-
structions. Syst. Zool. 28:379-385.
Goodman, M., J. Czelusniak, G. W. Moore, A. Romero-Herrera, and G. Matsuda.
(1979b) Fitting the gene lineage into its species lineage, a parsimony strategy
illustrated by cladograms constructed from globin sequences. Syst. Zool.
28:132-163.
Hillis, D. M. (1985) Evolutionary genetics of the Andean lizard genus Pholidobolus
(Sauria: Gymnophthalmidae): phylogeny, biogeography, and a comparison
of tree construction techniques. Syst. Zool. 34:109-126.
Hillis, D. M. (1991) The phylogeny of amphibians: current knowledge and the role
of cytogenetics. Pp. 7-31 in Amphibian Cytogenetics and Evolution (D. M.
Green and S. K. Sessions, eds.). Academic Press, New York.
Hillis, D. M., and R. de Sa. (1988) Phylogeny and taxonomy of the Ranapalmipes
group (Salientia: Ranidae). Herpetol. Monographs 2:1-26.
Hillis, D. M., and M. T. Dixon. (1989) Vertebrate phylogeny: evidence from 28S
ribosomal DNA sequences. Pp. 355-367 in The Hierarchy of Life. Molecules
and Morphology in Phylogenetic Analysis (B. Fernholm, K. Bremer, and
H. Jornvall, eds.). Elsevier Science Publishers B. V., Amsterdam.
Hillis, D. M., M. T. Dixon, and L. K. Ammerman. (1991) The relationships of
coelacanths: evidence from sequences of vertebrate 28S ribosomal RNA
genes. In press in Coelacanth Biology and Evolution (J. A. Musick, ed.).
Special volume of Environ. Biol. Fishes 32.
Kocher, T. D., and T. J. White. (1989) Evolutionary analysis via PCR. Pp. 137-
147 in PCR Technology. Principles and Applications for DNA Amplification
(H. A. Erlich, ed.). Stockton Press, New York.
Le Quesne, W. J. (1989) Frequency distributions of lengths of possible networks
from a data matrix. Cladistics 5:395-407.
Maddison, W. P., and D. R. Maddison. (1991) MacClade: Interactive Analysis of
Phylogeny and Character Evolution. Sinauer Associates, Sunderland.
294 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Sokal, R. R., and F. J. Rohlf. (1981) Biometry. 2nd ed. W. H. Freeman, San
Francisco.
Swofford, D. L. (1990) PAUP: Phylogenetic Analysis Using Parsimony, Version
3.0. Illinois Natural History Survey, Champaign.
Werman, S. D. (1986) Phylogenetic relationships of the true viper, Eristocophis
mcmahoni, based on parsimony analysis of allozyme characters. Copeia
1986:1014-1020.
14
When are Phylogeny Estimates From
Molecular and Morphological Data
Incongruent?

DAVID L. SWOFFORD

Many sources of information may be available for use in inferring phylo-

genetic relationships among a group of taxa. Because different character
sets share a common evolutionary history, we expect reliable methods of
phylogenetic analysis to recover the correct evolutionary tree regardless
of the kind of data employed. To the extent that character sets "tell the
truth" about their past, phylogenies inferred from different character sets
should be congruent with the true tree and therefore with each other. This
line of reasoning suggests that congruence among data sets might provide
the strongest achievable evidence that a proposed phylogeny is accurate
(e.g., Penny and Hendy, 1986). In practice, however, the ideal of perfect
congruence is frequently not achieved. Phylogenies estimated from addi-
tional character sets typically disagree in minor details and sometimes
contain major discrepancies.
When phylogeny estimates from two character sets disagree in nontrivial
ways, we are confronted with a dilemma: both of the estimates cannot be
correct, so how do we reconcile the differences? One explanation is that
one of the data sets is simply unreliable and that no method of phylogenetic
reconstruction could be expected to recover the correct tree given the poor
quality of the data. At one time or another, most of us have witnessed or
participated in vigorous debates about the relative strengths of alternative
sources of information or about the power versus futility of particular
techniques. The increase in prominence of molecular approaches to sys-
tematics has increased the frequency of such exchanges, especially when
molecular evidence controverts traditional notions of phylogenetic rela-

295
296 PHYLOCENETIC ANALYSIS OF DNA SEQUENCES

tionship within a set of taxa. A fundamental question that must be ad-

dressed when apparent conflicts arise is whether the incongruence is "real"
(i.e., unavoidable) or merely "spurious" (explicable due to sampling error
or to inappropriate assumptions or analytical methods) (Hillis, 1987).
This chapter has two purposes. First, in keeping with the general theme
of this volume, I will review several methods currently being used to assess
levels of congruence, presenting my own views on the strengths and lim-
itations of each. Second, I will suggest some additional procedures, facil-
itated by recent improvements in computer software, that allow a more
comprehensive examination of the question posed in the title.

Terminology
I will deliberately avoid excessive formalism. In the remainder of the chap-
ter, collections of terminal taxa are referred to simply as groups. Nontrivial
or Informative groups contain at least two terminal taxa, but not all of the
terminal taxa under study (the universal group). (In general, the term
"group" as used herein can be considered synonymous with the terms subset
and component used by some authors.) Attention is restricted to rooted
trees with labeled terminal nodes, representing actual terminal taxa, and
unlabeled internal nodes, representing ancestral taxa that are purely hy-
pothetical. Under these restrictions, the terms "tree" and "cladogram"
can safely be used interchangeably. ("Terminal taxa" are equivalent to the
tips, leaves, terms, or OTUs of some authors; hypothetical ancestral taxa
are sometimes referred to as HTUs.) If the word "taxon" is used without
a modifier, it is assumed to mean "terminal taxon."
The term monophyletic is used in the sense of Hennig (1966): a group
is monophyletic on a tree if, and only if, it contains all of the descendants
of its most recent common ancestor. A statement that a group "appears"
or "occurs" on a tree, or that a tree "contains" a group, implies that the
group is monophyletic on the tree. A clade is a monophyletic section
(subtree) of a tree, and is identified by the names of its terminal taxa;
these terminal taxa constitute a cluster.

"TAXONOMIC" CONGRUENCE: CONSENSUS TREES, CONSENSUS

INDICES, AND TREE COMPARISON METRICS

Taxonomic congruence refers to the extent to which independent classifi-

cations for the same set of taxa support the same groupings. The most
commonly used approach for assessing taxonomic congruence, pioneered
by Mickevich (1978), involves the calculation of consensus trees that sum-
marize areas of agreement among classifications. The amount of structure
retained by a consensus tree provides a measure of the congruence among
the original classifications. The general approach is illustrated in Figure
14-1. First, two (or more) rival trees are constructed using different char-
WHEN ARE PHYLOGENY ESTIMATES INCONGRUENT? 297

Figure 14-1 Illustration of

basic strategy used by Mick-
evich to compare methods
with respect to taxonomic
congruence. Two trees (a,b)
calculated using method one
and a consensus (c). Two
trees (d,e) calculated using
method two and a consensus
(f). Because the consensus
tree from method two is less
resolved, one can conclude
that for this data set, method
one generates more congruent
classifications.

acters sets (Fig. 14-la,b). Second, a consensus method summarizes the

groupings shared by the rival trees, producing a consensus tree (Fig.
14-lc). A highly resolved consensus tree reflects substantial congruence
among the rival trees, whereas a poorly resolved consensus tree may in-
dicate areas of significant disagreement. If another method of tree esti-
mation is applied to the same pair of data sets, different rival trees may
result (e.g., Fig. 14-ld,e), perhaps yielding a more poorly resolved con-
sensus (Fig. 14-lf). In this example, the first method generated more con-
gruent classifications than the second.
Mickevich's (1978) study compared various phenetic and numerical clad-
istic methods with respect to their ability to construct congruent classifi-
cations from different data sets. She used the only consensus method avail-
able at the time, that of Adams (1972). Since then, a number of additional
consensus methods have been proposed (Nelson, 1979; Margush and
McMorris, 1981; Sokal and Rohlf, 1981; McMorris and Neumann, 1983;
Neumann, 1983; Shao, 1983; Stinebrickner, 1984; Barthelemy and McMorris,
1986; Bremer, 1990). Some of the more commonly used methods are de-
scribed below.

Strict and Majority-Rule Consensus

Of the existing consensus tree methods, the strict consensus is conceptually
the simplest. The strict consensus tree, apparently used first by Schuh and
Polhemus (1981), was defined by Sokal and Rohlf (1981) as the unique
tree that contains only those groups that appear on all of the rival trees.
Unfortunately, strict consensus trees were called "Nelson trees" (after
Nelson, 1979) by Schuh and Farris (1981). Page (1989a) clearly demon-
strates that "strict" and "Nelson" trees are not equivalent, but the improper
synonymy has become entrenched in some circles.
Examples of strict consensus trees are shown in Figure 14-2. The greatest
advantage of the strict consensus is simplicity of interpretation. If a group
298 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Figure 14-2 Examples of strict consensus. The top row shows two trees (a,b) and
their strict consensus tree (c). The bottom row shows two trees (d,e) that are identical
except for the placement of taxon H. Their consensus tree is completely unresolved
(f), indicating that no groups are shared by the two trees.

appears on the strict consensus tree, we can immediately be certain that

it appears on all of the rival trees. However, strict consensus also has a
serious liability. Two trees that support identical relationships for all but
one of the taxa (e.g., taxon H in Fig. 14-2d,e) may share no groupings
when this last taxon is included, so that the consensus tree is a completely
unresolved bush (Fig. 14-2f). Adams (1986) and Funk and Brooks (1990)
show similar examples.1 This property has caused some workers to conclude
that strict consensus trees are "too strict."
When several trees are being compared, a majority-rule consensus (Mar-
gush and McMorris, 1981) may be preferable to a strict consensus. Instead
of including in the consensus tree only those groups that occur on all of
the rival trees, we retain all groups that appear on more than some pre-
specified proportion of the rival trees. Typically, a proportion of 0.5 is
used, so that all groups occurring on more than half of the rival trees are
retained in the "50% majority-rule" consensus. Note that a group must
be present on more than half of the trees to be retained in the consensus,
because two groups occurring in exactly half of the trees might not be able
to coexist on the same tree. Of course, if only two trees are being compared,
the strict and majority-rule methods are equivalent.

Adams Consensus
To my knowledge, the consensus method of Adams (1972) predates all
others. Adams actually described two variants on his consensus method,
1
I am not sure who first called attention to this property. I first heard it described by Joseph
Felsenstein at the October 1981 meeting of the Numerical Taxonomy Conference in Ann
Arbor, Michigan.
WHEN ARE PHYLOGENY ESTIMATES INCONGRUENT? 299

the first applying when the rival trees are "fully labeled" (i.e., internal
nodes represent actual ancestral taxa rather than purely hypothetical ones),
and the second when all internal nodes are "unlabeled" as is usually the
case in phylogenetic analysis. Consequently, Adams consensus trees are
sometimes more precisely called Adams-2 trees, eliminating any uncertainty
about which of Adams' (1972) methods is being used.
Adams consensus trees often preserve more of the structure found in
the rival trees than do strict consensuses. For example, Figure 14-3c shows
the Adams consensus of the trees in Figure 14-3a and b. It reflects the
following statements about relationship made by both rival trees: (1) taxa
A and C are more closely related to each other than either is to taxa D,
E, or F; (2) taxa E and F are more closely related to each other than either
is to taxa A, C, or D; and (3) taxon D is more closely related both to taxa
E and F than it is to either taxa A or C. However, in no case does the
same group of taxa appear on both of the rival trees; consequently, the
strict consensus is a completely unresolved bush. Funk and Brooks (1990)
present a similar example (their figs. 14-16). Regrettably, the same authors
also present another tree (their fig. 9) that is incorrectly called an "Adams
consensus" (of the trees in their figs. 7 and 8), stating that "the Adams
tree is better resolved than either of the [rival] cladograms because neither
. . . conflicts with the other." Actually, the preservation of all uncontra-
dicted groups by a consensus technique is a property of the combinable
component (Bremer, 1990) and Nelson (1979) consensus methods (see
below). Although this example was intended to illustrate the differences
between Adams and strict consensus methods, when performed correctly
the two methods yield the same consensus trees in this case. Furthermore,
because they consider strict and Nelson trees to be equivalent, the tree
referred to as a Nelson consensus tree (NCT:6) is actually an Adams
consensus tree, and vice versa. Thus, Funk and Brooks' (1990) conclusion
that "an Adams tree answers those questions about which the clades are
not in conflict" is less appropriate than the characterization of Mickevich
and Platnick (1989): "In an Adams consensus . . . any taxonomic state-
ments shared by the classifications being compared are included in the
consensus, regardless of whether they constitute completely uncontradicted
components." That is, the statement that taxa E and F are more closely
related to each other than to any of the remaining taxa, made by Funk

Figure 14-3 An example of Adams (1972) consensus. Tree (c) is the Adams con-
sensus of trees (a) and (b).
300 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

and Brooks' cladogram 8, is not shared by their cladogram 7, with the

result that an EF cluster is not represented in the consensus.
McMorris et al. (1983) suggested that Adams' method lacks a compelling
justification and has achieved popularity primarily owing to historical pre-
cedence. However, Adams (1986) addressed that criticism by character-
izing the Adams consensus as the "best tree that represents all the infor-
mation shared among a set of trees." In this second paper, Adams suggests
that a tree be thought of as a set of "leaf subset nestings" rather than as
a "set of clusters." In biological terms, one group of taxa nests within a
larger one if the more recent common ancestor of the smaller group is a
descendant of the most recent common ancestor of the larger group. (Note
that this definition does not require monophyly of either group.) He then
proves that the Adams consensus tree, as defined in his earlier (1972)
paper, is the unique tree that satisfies the following two conditions: (1)
any nesting found in all of the rival trees must also occur in the consensus
tree; and (2) any nesting that reflects clusters of the consensus tree (i.e.,
a nesting involving the inclusion of one monophyletic group within a larger
monophyletic group) must be found in all of the rival trees.
Adams' (1972) method has also been criticized (Sokal and Rohlf, 1981;
Rohlf, 1982; Rohlf et al., 1983) because it may produce consensus trees
containing clusters not found on any of the rival trees (see Fig. 14-4). This
property mildly complicates the interpretation of Adams consensus trees,
but it must be accepted if one agrees that the structure of a tree encompasses
more information than a simple listing of its clusters (Adams, 1986). Under
Adams' interpretation, this additional information comes in the form of
nestings, which are loosely analogous to the "taxonomic statements" of
Mickevich and Platnick (1989).

Combinable-Component and Nelson Consensus

When one or more of the rival trees is not fully dichotomous, groups that
are never contradicted may occur on some, but not all, of the trees. For
example, suppose that three trees contain a clade (A,B) and a fourth tree
exhibits an unresolved (A,B,C) trichotomy. In this case, the AB group
would not be retained in a strict consensus of the four trees, even though
it occurred in three-fourths of the rival trees and was not contradicted by
any of them. In this case, we might choose to include the group AB in the

Figure 14-4 An example illustrating that an Adams consensus tree (c) may contain
groups not found on any of the rival trees (a, b).
WHEN ARE PHYLOGENY ESTIMATES INCONGRUENT? 301

consensus. Bremer (1990) formally described this approach as "combinable-

component consensus;" essentially the same idea was put forth by Hillis
(1987). Two groups are "combinable" (Nelson, 1979) if either (1) they
have no taxa in common ("exclusion"); (2) they are identical ("replica-
tion"); or (3) one group is a proper subset (further resolution) of the other
("inclusion"). The combinable-component consensus tree is defined by the
set of all combinable groups (i.e., each group retained in the consensus is
equal to or combinable with all groups of every rival tree). When all rival
trees are fully dichotomous, the strict and combinable-component consen-
sus methods are equivalent.
Considerable confusion exists as to exactly what is meant by "Nelson
consensus." Bremer (1990) argues (correctly, in my view) that when exactly
two trees are compared, the "Nelson tree" becomes equivalent to the
combinable-component consensus tree. Likewise, Mickevich and Platnick
(1989:40) state: "In a Nelson consensus, only those components [groups]
that are uncontradicted in the classifications being compared are included
in the consensus tree." This definition would include combinable, but not
necessarily universally occurring, groups in the consensus. But two pages
later (Mickevich and Platnick, 1989:42) they assert that "a Nelson con-
sensus includes only those subtaxa exactly replicated in the cladograms
being compared" (emphasis mine), which is the definition of a strict con-
sensus tree.
Nelson (1979) described his method for obtaining a "general cladogram"
as a two-stage procedure. First, a tree containing only those groups found
on at least two trees ("replicated components") is constructed. Then, if
there are any other unreplicated groups that are combinable with all of
the replicated groups, they are included in the consensus as well. If all
unreplicated groups are combinable with all replicated groups, as will al-
ways be the case when exactly two trees are compared, Nelson's approach
becomes equivalent to Bremer's (1990) combinable-component consensus.
Thus, Page's (1989a) contention that Nelson's method is equivalent to strict
consensus for the two-tree case is true only if the two trees are both fully
dichotomous, in which case any unreplicated group will be noncombinable
with at least one group appearing on the other tree. For more than two
trees, however, Nelson's method can produce results that differ strikingly
from both strict and combinable-component consensus trees. Figure 14-5
shows three rival trees in which the replicated groups are AB (trees 5a
and 5c), ABC (trees 5a and 5b), and ABCD (all three trees). The two
unreplicated groups (AC and ABD) are both noncombinable with at least
one of the replicated groups, therefore the Nelson consensus tree includes
only the three replicated groups (Fig. 14-5d). On the other hand, the only
group retained by the strict (and combinable-component) consensus is
ABCD (Fig. 14-5e). In this example, the Nelson and majority-rule con-
sensus trees are identical, but this equivalence does not hold in general.
Unfortunately, Nelson (1979) did not consider the possibility of repli-
cated, but noncombinable, groups, leading McMorris et al. (1983) to con-
302 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

elude that his method need not even specify a tree. Page (1989a) addressed
this oversight, providing a formal, unambiguous definition of "Nelson con-
sensus" that is equivalent to Nelson's original procedure when all replicated
groups are combinable, but also treats the case of noncombinable replicated
groups. Page's method uses standard graph theoretical techniques (e.g.,
Bron and Kerbosch, 1973) to find all cliques of combinable groups: sets
of groups such that every group in the set is combinable with every other
group in the set. Each clique is assigned a score equal to the sum, over
groups included in the clique, of the number of times each group is rep-
licated in the rival trees. (The replication count for a group is simply one
less than the number of times the group occurs on the rival trees, so that
unreplicated groups do not contribute to the score.) If there is a single
clique with the highest replication score, the groups included in this clique
define the Nelson consensus tree. If more than one such clique exists, the
Nelson consensus contains only those groups found in all of the maximal-
replication cliques. In this case, groups found in some, but not all, of the
maximal-replication cliques are classified as "ambiguous."
Although no one seems to have suggested it, an alternative procedure
would be to include unreplicated as well as replicated groups in the score
for each clique, directly analogous to character compatibility analysis (Le
Quesne, 1969; Estabrook et al., 1976). Nelson's (1979) emphasis on rep-
licated groups was motivated by his calculations of the small probability
of replicating particular groups by chance alone. However, it is not obvious
that this preoccupation with replication is justified. Suppose that in a study
of 20 taxa, two completely asymmetrical (pectinate) trees contain the group
AB. According to Nelson, the probability that the second tree would rep-
licate the AB group found on the first tree equals one divided by the
number of distinct ways of choosing a pair of objects (without replacement)
from a pool of 20:2

But if taxa A and B were known to be part of a larger "true" group ABC,
then the probability that two random resolutions of a trichotomy involving
these three taxa would generate an AB group in both cases is 1/9, greater
by a factor of 10. Nelson uses similar logic to establish that combinability
of unreplicated groups with replicated groups is not particularly surprising;
it is unclear why this line of reasoning should not be extended to some
replicated groups as well.

2
1 do not agree with his procedure for calculating these probabilities, but that is a separate
issue. A more appropriate procedure (Simberloff, 1987) is to generate a large number of
pairs of random trees under a Markovian model and count the number of times a given
component is replicated due to chance alone.
WHEN ARE PHYLOGENY ESTIMATES INCONGRUENT? 303

Figure 14-5 An example of

Nelson (1979) consensus.
Tree (d) is the Nelson consen-
sus of trees (a), (b), and (c).
The strict tree (e) is much less
resolved than the Nelson con-
sensus, demonstrating that the
two methods are not equiva-
lent.

Median Consensus
The median consensus procedure (Barthelemy and Monjardet, 1981; Bar-
thelemy and McMorris, 1986) is closely related to majority-rule consensus.
In this method, a tree comparison metric that reflects the topological dis-
tance (i.e. , disagreement) between any pair of trees (see below) is defined.
If we represent by d(T,,Tj) the distance between a pair of trees T, and T},
then the total distance of any tree T to a set of k rival trees is given by

A tree Tm is a median if its total distance dm to the rival trees is less than
that for any other tree. If the tree comparison metric is the symmetric-
difference distance (Robinson and Foulds, 1981; Hendy et al., 1984; see
below), the 50% majority-rule consensus tree is a median tree (Barthelemy
and McMorris, 1986). If the number of rival trees (K) is odd or if there
are no groups that appear in exactly half of the rivals, the majority-rule
tree is the only median tree. If k is even, any tree representing a combi-
nation of the majority-rule consensus tree with one or more (combinable)
groups appearing in exactly half of the rival trees is also a median tree.
Barthelemy and McMorris (1986) stated that the consensus tree pub-
lished by Penny et al. (1982) was an example of a median consensus tree.
Although Penny et al. (1982) did not use the term "median," D. Penny
(personal communication) has also recently suggested that it be used for
the consensus method described in their paper. However, the procedures
of Penny et al. (1982) and of Barthelemy and McMorris (1986) are different.
Penny et al.'s (1982) method chooses the binary (fully dichotomous) tree
that minimizes the total distance to the rival (binary) trees. The "consensus
tree" of Penny et al. (1982) and the majority-rule consensus tree have total
symmetric-difference distances of 216 and 192, respectively, to the 39 rival
trees, demonstrating that the former is not a median tree under the criterion
304 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

of Barthelemy and McMorris (1986). To avoid further confusion, I suggest

that "median tree" be restricted to its original usage and that "median
binary tree" be used for the Penny et al. (1982) procedure. The latter
seems particularly useful as a means for selecting a "representative" tree
from a set of rival binary trees.

Largest Common Pruned Trees

In all of the consensus methods discussed above, the consensus tree con-
tains all of the taxa included in the rival trees. A rather different approach
(Gordon, 1980; Finden and Gordon, 1985) is to find the largest subset (or
subsets) of terminal taxa for which the rival trees specify identical rela-
tionships. In this method, taxa and their associated branches are removed
("pruned") from the rival trees until the reduced trees become topologi-
cally equivalent, yielding a common pruned tree. By removing the smallest
possible number of terminal taxa, one obtains a largest common pruned
tree (there may be more than one such tree).
This approach is particularly useful when only a few taxa are responsible
for the incongruence among trees, thereby providing a means of identifying
"unstable" taxa. For example, the strict consensus of the two trees shown
in Figure 14-6a and b is poorly resolved (Fig. 14-6c). However, these two
trees are equivalent in topology if taxon B is pruned from both trees (Fig.
14-6d). Terminal taxa not included in a common pruned tree can subse-
quently be "regrafted" to the tree. An obvious way to perform the re-
grafting (Gordon, 1980) is to reconnect removed taxa such that any group
appearing on all rival trees is preserved in the regrafted tree. The resulting
"common pruned-and-regrafted tree" is identical to or a resolution of the
strict consensus tree (e.g., Fig. 14-6e). In this example, the pruned-and-
regrafted tree is also equivalent to an Adams consensus tree, but the two
methods do not always give the same results.
Like other consensus methods, the largest-common-pruned-tree ap-
proach has some undesirable properties that limit its utility. One drawback

Figure 14-6 An example of

largest common pruned trees.
Two rival trees (a, b). Strict
consensus (c) of these two
trees. Largest common pruned
tree (d). Common pruned-
and-regrafted tree (e) obtained
by reattaching taxon B.
WHEN ARE PHYLOGENY ESTIMATES INCONGRUENT? 305

Figure 14-7 An example demonstrating that largest common pruned trees are not
always preferable. Tree (c), the largest common pruned tree for rival trees (a) and
(b), conveys much less information than does a smaller common pruned tree (d).

of the common-pruned-tree method is the difficulty of identifying largest

common pruned trees. An obvious "brute-force" approach is to examine
all ways of pruning k taxa and checking the resulting pruned trees for
equivalence, increasing k until a common pruned tree is found. As noted
by Finden and Gordon (1985), this method is computationally infeasible
for all but small data sets. These authors describe heuristic algorithms that
appear to be reasonably effective, but an exact algorithm that avoids the
computational disadvantages of the brute-force approach would be more
satisfying.
Even if finding them were easy, largest common pruned trees are not
always preferable, as demonstrated by the example of Figure 14-7. The
largest common pruned tree for trees 7a and 7b (Fig. 14-7c), which retains
six taxa, is clearly less useful than is the smaller common pruned tree of
Figure 14-7d, which retains only five taxa. This example suggests that
consideration of all common pruned trees of nontrivial size may provide
useful information that is lost if attention is restricted to the largest common
pruned tree(s).
Common pruned trees have received little attention from systematists,
probably due to their unavailability in widely distributed computer pack-
ages such as COMPONENT (Page, 1989b), Hennig86 (Farris, 1988), PAUP
(Swofford, 1991), and PHYLIP (Felsenstein, 1990). I hope to include a
facility for computing them in a future version of PAUP.

Consensus Indices
Examination of a consensus tree can provide a rough idea of the extent to
which a collection of rival trees agrees upon relationships; a consensus
index provides a quantitative measure of this agreement. Consensus indices
typically vary between 0 (implying no agreement among the rival classi-
fications) and 1 (implying the maximum agreement possible). A variety of
consensus indices have been proposed, reflecting different perceptions among
authors as to what aspects of "consensus" should be measured and how
306 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

these aspects should be quantified. In this regard, Day and McMorris (1985)
argue that the term "consensus index" should be reserved for measures
that strictly measure agreement among rival trees. Other measures that
consider additional attributes such as information content (including all of
the "consensus indices" commonly used by systematists) are called con-
sensus object invariants. This distinction highlights an interesting contrast
in the perspectives of mathematicians versus biologists. To Day and McMorris
(1985), a group of identical trees exhibits "perfect" agreement even if the
trees are totally unresolved. These authors consider it axiomatic that a
consensus index should take the value of 1 in the presence of such agree-
ment, regardless of the nature of the "vote." On the other hand, a biologist
views unanimous agreement on a "bush" as agreement on nothing at all
(e.g., Mickevich and Platnick, 1989), and therefore wants the index to take
the value 0. In the following discussion, I will use the term "consensus
index" in the less restrictive sense. Rohlf (1982) and Mickevich and Platnick
(1989) have presented thorough and readable discussions of consensus
indices, including computational details. Consequently, I will confine my
remarks to brief characterizations of some of the more commonly used
measures.
The simplest consensus index simply quantifies the amount of resolution
of a consensus tree and is equal to the number of nontrivial clusters that
the tree contains (= component information of Nelson, 1979) divided by
the maximum possible such number (t - 2, where t is the number of
terminal taxa). Colless (1980, 1981) has referred to this quantity as the
"consensus fork" index. However, the degree of resolution of a classifi-
cation does not adequately reflect the amount of information it contains.
For example, both trees in Figure 14-8 contain two nontrivial groups, but
the tree of Figure 14-8a clearly allows more statements about relationship
to be made, and therefore conveys more information (see Mickevich and
Platnick, 1989 for an eloquent discussion of this issue).
Several measures that combine information and agreement into a single
index have been proposed. These indices all contain a numerator (score)
representing information of some sort divided by a limiting value (nor-

Figure 14-8 Demonstration that

consensus trees with the same
degree of resolution (number of
nontrivial groups) convey differ-
ent amounts of information. Tree
(a) allows more statements of re-
lationship to be made than does
tree (b).
WHEN ARE PHYLOGENY ESTIMATES INCONGRUENT? 307

malizing factor) representing the highest achievable score for any tree. The
earliest measure, Mickevich's (1978) CI, was defined in terms of the num-
ber of "extra steps" required for a set of artificial cluster variables (de-
scribing the nontrivial clusters on the consensus tree) to evolve on a bush
under the Wagner parsimony criterion (Farris, 1970). Rohlf (1982) showed
that the numerator of Mickevich's index could be expressed equivalently
as a sum, over all nontrivial clusters contained in the consensus, of cluster
"weights:" the weight for the z'-th cluster equals min(n, — 1, t - «,), where
n, is the number of terminal taxa in cluster i and t is the total number of
terminal taxa. Colless (1980) suggested simply using n, as the weight for
each cluster, which yields a weighted consensus fork index proportional to
Nelson's (1979) total information? Mickevich and Platnick (1989) appar-
ently rediscovered the same index, calling it Pm, the "proportion of the
maximal total information." They noted that total information could be
defined using artificial variables, analogously to Mickevich's original for-
mulation: total information is equivalent to "WISS total steps," the total
number of steps required for the artificial variables to evolve on a bush
under the Camin-Sokal parsimony criterion (Camin and Sokal, 1965; Farris
et al., 1970). Yet another related measure is the "levels sum" (Schuh and
Farris, 1981), which counts the number of times distinct pairs of terminal
taxa occur together in the same informative cluster. It is computed by
summing, over all pairs of taxa, the number of informative clusters that
contain both members of each pair. Rohlf (1982) demonstrated that the
procedure was equivalent to using a cluster weight of nt(n, - l)/2 and
determined the limiting value to be t(t - l)(f - 2)/6 (note correction of
typographical error in his formula).
A property of all of the indices described in the above paragraph is that
they are sensitive to the symmetry or balance of the tree (Colless, 1980;
Rohlf, 1982; Shao and Rohlf, 1983; Shao and Sokal, 1990). The normalized
total-information measure and the levels sum can achieve their maximum
value only when the consensus tree is fully resolved and maximally asym-
metrical (Colless, 1980; Rohlf, 1982). Rohlf (1982) makes the same claim
for Mickevich's (1978) CI. However, at least as originally defined by Mick-
evich, CI can in fact achieve its maximum value on a tree that is maximally
symmetrical (Fig. 14-9).
An important implication of asymmetry bias in a consensus index is that
two different asymmetrical trees can achieve a higher score on the index
than two identical symmetrical trees. Whether this property is viewed as
an asset or a liability depends upon one's point of view. If we want a
consensus index to reflect only agreement among trees, we clearly do not
want to use a measure that is also sensitive to asymmetry. If we instead
want an index to reflect, in some sense, the amount of information shared
by a set of trees, then we must accept the possibility that different, asym-
3
The formula given by Colless for calculating the limiting value of his index contained a
typographical error and should instead be (n + l)(n - 2)/2. It is clear from his calculations
that he used the correct formula.
308 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Figure 14-9 Demonstration that Mickev-

ich's consensus index can achieve its
maximum value on a tree that is not max-
imally asymmetrical. C/ = 1.0 for both
trees.

metrical trees may allow more shared statements about relationship than
trees that are identical but symmetrical, and we should therefore design
our index accordingly. If we then use this index to evaluate the extent to
which alternative methodological approaches generate congruent classifi-
cations (e.g., Mickevich, 1978), we must be careful to avoid overstating
our conclusions. Specifically, the finding that two different data sets pro-
duce the same estimate of phylogenetic relationships using method "X"
says nothing about the truth of those relationships (i.e., the "false confi-
dence" problem cited by Faith, 1988). For example, two data sets will
always produce identical estimates of a phylogeny if the "estimation" pro-
cedure consists of placing the terminal taxa in alphabetical order on a
completely pectinate tree. Conversely, if we are willing to assume that a
method is reliable in general, and we want to use taxonomic congruence
as a means of assessing confidence in a particular result, we would not
want to use a consensus index that penalized the method for getting the
correct result if that result happened to correspond to a symmetrical tree.
It is possible to design consensus indices that preserve the desirable
group-size properties without suffering from excessive tree asymmetry bias.
For example, Rohlf s (1982) C/7 always takes the value 1 for a fully resolved
consensus tree (i.e., identical, fully resolved rival trees), regardless of the
degree of asymmetry. Mickevich and Platnick's (1989) statement that Rohlf s
measure amounted to "mutating the consensus information index (Mick-
evich, 1978) to exclude considerations beyond resolution alone" represents
a misunderstanding of Rohlf s index. It is clear that Rohlf wants the index
to reflect more than resolution alone, but he does not want it to take a
value less than 1 if two or more fully resolved rival trees agree completely.
As such, CIj does not deserve to be tossed onto the scrap heap along with
Colless' consensus fork index. Indeed, for the trees shown in Figure 14-
lOa and b, which are equally well resolved and therefore have equal con-
sensus fork indices, both Rohlf's C/; and Mickevich's CI are exactly three
times greater for the "more informative" tree lOa than for tree lOb. How-
ever, CIj is greater for tree lOc, which implies complete agreement of the
rival trees, than for tree lOd, which might represent the consensus of two
trees that specified different relationships for taxa A, B, and C. In contrast,
both Mickevich's CI and the weighted consensus fork (Pm) are higher for
tree lOd than for tree lOc.
Rohlf (1982) also suggested another intriguing index: the proportion of
all possible binary trees that contain the clusters found in the consensus
WHEN ARE PHYLOGENY ESTIMATES INCONGRUENT? 309

Figure 14-10 Effect of tree symmetry and resolution on consensus indices. Con-
sensus trees (a) and (b) both contain a single nontrivial group, but measures such
as Rohlf's (1982) C/,, Mickevich's (1978) C/, and Mickevich and Platnick's (1989)
Pm (= Colless' weighted consensus fork) take higher values for tree (a) than for tree
(b), reflecting the higher information content of (a). Although consensus tree (c) is
fully resolved, it takes lower values for CI and Pm than the partially unresolved, but
more asymmetrical consensus tree (d). (CF = Colless' unweighted consensus fork.)

tree. The number of binary trees that contain the clusters of the consensus
tree is equivalent to the "number of fully resolved cladograms allowed by
a [strict] consensus" (Mickevich and Platnick, 1989; see also Mickevich
and Farris, 1981). For example, suppose that a strict consensus tree with
eight terminal taxa contains two dichotomies, one trichotomy, and one
tetrachotomy. We can fully resolve the trichotomy in three different ways
and the tetrachotomy in 15 different ways, so there are 3 x 15 = 45 binary
trees allowed by the consensus. (The number of ways to "fully resolve" a
fc-tomy is equivalent to the number of rooted binary trees for k taxa; e.g.,
Felsenstein, 1978.) Since there are 135,135 possible rooted binary trees for
eight taxa, Rohlf's (1982) C/2 is then 45/135,135 = 3.33 x 10~4. Thus,
the number of trees allowed by the consensus is a tiny fraction of the
number of possible trees. Mickevich and Platnick (1989) use the comple-
ment of C/2, corresponding to the "number of prohibited resolutions,"
which follows the usual convention of letting the value 1 indicate maximum
resolution. Mickevich and Platnick also tackle the more difficult problem
of computing the number of resolutions prohibited by an Adams (1972)
consensus tree. They recommend using the product of the proportion of
prohibited Adams resolutions and the normalized total-information index
as a "general information index."
310 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

A legitimate criticism of all of the consensus indices described above is

that they depend only on the consensus tree; once this tree has been
calculated, all contact with the original rival trees is abandoned (Day and
McMorris, 1985). Thus, for example, it is not possible to determine whether
a consensus index is low because there was substantial disagreement among
a set of well-resolved trees or because there was general agreement among
a set of poorly resolved trees. Day and McMorris (1985) categorize all of
these indices as "type 1 paradigms." In paradigms of "type 2" the consensus
index is computed after calculation of a consensus tree, but the index
depends on both the form of the consensus tree and on the original trees.
An example is the a index of Day (1983), which does not appear to have
an exact algorithm for its calculation. It has been ignored by systematists,
perhaps because it satisfies the Day-McMorris axiom specifying that it
always take the value 1 for identical rival trees, even if they are unresolved.
In "type 3 paradigms," the consensus tree and index are specified simul-
taneously by an incremental process in which addition of groups to the
consensus tree determines corresponding changes to the consensus index
(e.g., 5-consensus index of Stinebrickner, 1984, 1986; consensus index cin
of Adams, 1986). Unfortunately, none of these indices provides what we
really need: an index that is sensitive to both agreement among and in-
formation content of the original trees, but that also allows us to quantify
the relative contributions of each of these aspects to lack of resolution in
the consensus. We cannot hope to achieve such precision and versatility
in a one-dimensional index, but further work in this area may prove fruitful.

What Does the Magnitude of a Consensus Index Really Mean?

What does it mean to say that a consensus index equals, say, 0.7? Tech-
nically, for many indices, this value implies that the rival trees share 70%
as much information on relationships as does a set of identical, maximally
asymmetrical trees. But does a value of 70% indicate substantial agreement
and/or shared information, or should we be dejected over the failure to
find the congruence we anticipated? This problem of interpretation pre-
sents a serious obstacle to the use of consensus methods in quantifying
levels of congruence. It may be useful to test, using Monte Carlo methods,
whether a given consensus index is greater than that expected for a set of
randomly chosen trees (Shao and Rohlf, 1983; Shao and Sokal, 1986). For
the case of two rival trees, Shao and Sokal (1986) provide critical-value
tables for eight different consensus indices over a range of taxon numbers.
A similar approach has been used by Simberloff (1987). One difficulty with
these methods is that the null hypothesis is relatively uninteresting; it would
be extremely unlikely for the amount of congruence to be so low that we
would be unable to reject the hypothesis that our trees were no more
congruent than a random sample of trees. Unfortunately, it is not obvious
how a "more interesting" null hypothesis would be formulated.
WHEN ARE PHYLOGENY ESTIMATES INCONGRUENT? 311

General Criticisms of Consensus Methods

Miyamoto (1985) and Carpenter (1988) have criticized the use of consensus
trees as a basis for classification because, in general, they provide a less
parsimonious explanation of the original character data—and therefore
less "explanatory power"—than do the original trees from which the con-
sensus is derived. Technically, this criticism does not pertain to the use of
consensus methods for the study of congruence, because there is no nec-
essary connection between the calculation of a consensus tree as the basis
of a consensus index and the establishment of a classification based on that
consensus tree. However, the fitting of character data to consensus trees
in this way is inappropriate in any case. Consensus trees are simply state-
ments about areas of agreement among trees; they should not be inter-
preted as phylogenies. Polytomous nodes on a consensus tree do not in-
dicate simultaneous cladogenetic events ("hard" polytomies; Maddison,
1989), as assumed by the character-state reconstruction methods used by
Miyamoto (1985) and Carpenter (1988). Rather, polytomies on the con-
sensus tree indicate areas of uncertain resolution ("soft" polytomies). Mad-
dison (1989) has described algorithms for reconstructing character changes
under the uncertain-resolution (soft polytomy) interpretation that effec-
tively allow a polytomy to be resolved in the way that is most favorable
for each character considered independently. These algorithms thus pro-
vide a more appropriate mechanism for fitting character data to a consensus
tree, and the criticisms of Miyamoto (1985) and Carpenter (1988) no longer
apply if they are used.
More to the point is Miyamoto's (1985) argument that because the orig-
inal character data are no longer relevant once the "fundamental" trees
have been constructed, all information about the relative amounts of char-
acter support favoring alternative resolutions is discarded (see also Cra-
craft, 1983). For example, suppose that the groupings AB, AC, and BC
are supported by three, two, and one nucleotide position characters, re-
spectively. Further suppose that the AC group is supported by 20 characters
in a morphological data set, with no characters supporting either the AB
or BC groupings. A consensus method would then see only the conflicting
AB and AC groupings from the molecular and morphological trees, re-
spectively, and would be forced to accept an (A,B,C) trichotomy, despite
the weak, ambiguous support for AB in the molecular data, and the strong
unambiguous support for AC in the morphological data. Because of these
considerations, Miyamoto (1985) and others (e.g., Kluge, 1989) have rec-
ommended combining all of the available data sets into a single pooled
data set rather than computing trees separately for each. Again, while this
strategy may be desirable for classificatory purposes, it is less useful if we
wish to examine the extent to which different data sets support congruent
taxonomic relationships.
Yet another difficulty with the use of consensus approaches to evaluate
congruence concerns the selection of trees for comparison. If we compare
312 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

only the optimal (i.e., most-parsimonious, or best-fitting according to some

other criterion) trees for each data set, we effectively ignore the uncertainty
associated with the estimate of the tree. For example, the best tree for a
molecular data set may be highly incongruent with the best tree for a
morphological data set. Yet, there may be another tree that explains the
molecular data nearly as well as the "best" one that is virtually identical
to the tree derived from the morphological data. Evidence is accumulating
that this possibility is a very real one. Maddison (1991) shows several
examples of data sets that generate two or more "islands" of equally par-
simonious trees: trees within an island are (by definition) relatively similar
in topology to other trees in the same island, but trees in different islands
can be strikingly different, as shown in Figure 14-11. Hendy et al. (1988)
have also discussed this phenomenon. Clearly, before reaching a conclusion
that phylogeny estimates from two data sets are incongruent, we need to
examine the possibility that consideration of near-optimal trees for one or
both data sets may reduce the apparent amount of incongruence. The
hypothetical example of Figure 14-12 illustrates this point. The most-
parsimonious trees for a morphological (Fig. 14-12b) and a molecular (Fig.
14-12c) data set are relatively incongruent. However, a third tree (Fig.
14-12d), while not the best tree for either data set, is the second best tree
for both, and is only one step longer than the most-parsimonious tree for
the combined data set.

Tree Comparison Metrics

Another approach to assessing taxonomic congruence is through the use
of tree comparison metrics that quantify differences in the shapes of pairs
of trees. These methods have the advantage of not requiring preliminary
calculation of a consensus tree. On the other hand, the distance between
two trees—just a number—is more difficult to interpret than a consensus
tree that immediately indicates areas of agreement and a consensus index
that varies between predictable bounds. Penny and Hendy (1985) outline
a number of interesting applications of tree comparison metrics. For in-
stance, if the probability distribution of a particular metric is known, we
can ask whether any pair of trees is more similar than would be expected
by chance alone (e.g., Penny et al., 1982). We can also test whether the
positioning of one taxon (or a small set of taxa) is particularly unstable by
watching the decrease in tree-to-tree distances as we prune one (or a few)
taxa from the trees.
The most widely used tree comparison metric has been the symmetric-
difference distance, or partition metric (Robinson and Foulds, 1979;
Robinson and Foulds, 1981), defined as the number of groups that appear
on one tree or the other but not on both. (It can be defined equivalently
as the minimum number of branch contractions and decontractions, in any
order, required to transform one tree into the other, see Robinson and
Foulds, 1981.) Its chief advantages are that it is easy to calculate and it
Figure 14-11 Two trees (a,b) from different "islands" for James Carpenter's "ROPA10" data set (see Maddison,
1991). Their strict consensus is shown in (c); see text.
314 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Figure 14-12 Hypothetical example illustrating that although a tree may not be
optimal for either of two data sets, it may explain both data sets reasonably well. A
morphological data set (MORPH) containing three characters and a molecular data
set (MOLEC) containing six characters (a). Most-parsimonious tree (b) for the mor-
phological data set (length for MORPH = 3; length for MOLEC = 10; total length
= 13). Most-parsimonious tree (c) for the molecular data (length for MORPH = 5;
length for MOLEC = 6; total length = 11). Second most-parsimonious tree (d) for
both data sets (length for MORPH = 4; length for MOLEC = 8; total length = 12).

works equally well for binary and nonbinary trees. Because of its com-
putational tractability, Hendy et al. (1984) have been able to calculate its
probability distribution for up to 16 taxa on binary trees and 12 taxa al-
lowing nonbinary trees. It does possess a drawback similar to that of strict
consensus trees, however, in that two trees that are identical except for
the placement of one terminal taxon can take the maximum possible dis-
tance value. Penny and Hendy (1985) suggest strategies for coping with
this problem.
A well-studied, but relatively little used tree comparison metric is the
"crossover" or "nearest-neighbor interchange" metric (e.g., Robinson,
1971; Waterman and Smith, 1978). This distance measure lacks an efficient
exact algorithm, but useful approximations have been developed (Brown
and Day, 1984). Its probability distribution is known for up to eight taxa
on binary trees (Jarvis et al., 1983). Other measures of tree dissimilarity
have been suggested by Farris (1973), Estabrook et al. (1985), and Crozier
et al. (1986).

'CHARACTER" CONGRUENCE

Character congruence (Kluge, 1989) refers to the extent to which all avail-
able characters make a unified, internally consistent statement about the
WHEN ARE PHYLOGENY ESTIMATES INCONGRUENT? 315

relationships of the taxa under study. For the purpose of studying congru-
ence between data sets, placing the emphasis squarely on the primary data
(characters) rather than on an abstraction of the primary data (trees derived
from those characters) has tremendous appeal. In particular, such an ap-
proach permits a more direct assessment of the relative strength of support
for incongruent character distributions.

Measures of Character Fit

The consistency index (Kluge and Farris, 1969) has traditionally been used
to evaluate the fit of character data on phylogenetic hypotheses. Following
the notation and terminology of Farris (1989), the consistency index for a
single character, c, equals mis, where m is the minimum amount of change
that the character may show on any tree and s is the amount of change
required by the character on the tree being evaluated. The quantity (1 —
c) then represents the proportion of change which can only be explained
as homoplasy. An "ensemble" consistency index C for a suite of characters
(e.g., those included in a data matrix) can be calculated as MIS, where M
and 5 are the sums over all characters in the suite of the individual-character
m and s values, respectively. Ideally, one could use the consistency index
as a direct measure of character congruence, but several complications
arise. For instance, as more taxa are added to a study, our ability to detect
instances of homoplasy improves. The natural prediction that the consist-
ency index would be negatively correlated with the number of taxa in a
data set has been confirmed empirically (Archie, 1989; Sanderson and
Donoghue, 1989), demonstrating that C cannot stand alone as a measure
of character incongruence.
Another problem with the consistency index is that it is strongly affected
by the distribution of character-state frequencies. For example, C is always
equal to 1 for a character for which a single taxon possesses state 1 and
the rest state 0. Consequently, it is often suggested that autapomorphies
be excluded from the data set before calculating the ensemble consistency
index. But that strategy just shifts the problem slightly: binary characters
with a "(2, t — 2)" distribution of states cannot possibly have a consistency
index less than 1/2, and so on. Farris (1989) proposed two new indices,
the retention index and the rescaled consistency index. The retention index,
r, is defined as (g - s)/(g - m), where s and m are as defined above, and
g is the maximum possible amount of change that a character could possibly
require on any tree (equal to the length of the character on a completely
unresolved bush). Thus when a character fits the tree as poorly as possible,
its retention index will be 0, so that it can be more meaningfully interpreted
as a congruence statistic. Note that for uninformative (e.g., autapomorphic)
characters, m = g so that r is undefined. The ensemble retention index
R is defined analogously to the ensemble consistency index; that is,
(G - S)/(G - M) = (Sg - 2s)/(2g - 2m). Farris (1989) recommends
using r as a factor for scaling c between 0 and 1, hence re and RC are
316 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

called the "rescaled consistency index" and the "ensemble rescaled con-
sistency index," respectively.
Archie (1989) has adopted an alternative approach to measuring char-
acter incongruence based on random permutations of rows (taxa) within
columns (characters) of a data matrix. He estimates a statistic called the
homoplasy excess ratio (HER) which can be thought of as a modified
ensemble retention index in which G, the length of the longest possible
tree, is replaced by expected length of a tree from the randomized data
(his "observed homoplasy excess"). He suggests that the HER-Maximum,
equivalent to Farris's retention index, can be used to approximate HER.

Character Incongruence Indices

Mickevich and Farris (1981) outlined a method for measuring character

incongruence that partitions the total incongruence 1T (extra steps or hom-
oplasies) into a within-data-set component iw and a between-data-set
component iB. The partitioning is accomplished by computing most-
parsimonious (minimal) trees for each data set separately and after pooling
the data into a single combined data set. Let E(X) and E(Y) represent
the number of extra steps required by minimal trees for two data sets X
and Y, respectively, and let iT represent the total incongruence (number
of extra steps required by a minimal tree for the combined data set). The
within-data-set incongruence, iw, is then equal to E(X) + E(Y) and the
extra length iB required due to incongruence between data sets is given by
iT - iw. In other words, any excess homoplasy required by the combined
data beyond that needed to explain each data set separately is attributed
to incongruence between the two data sets. The Mickevich-Farris incon-
gruence index 1MF can be expressed as the proportion of the total incon-
gruence iT that is due to between-data-set incongruence:
IMF — IB/IT-

While this approach has intuitive appeal, it is not without problems. One
difficulty is the need to combine the data sets into a single matrix. If, for
example, one data set were cranial morphology and the other ribosomal
RNA (rRNA) sequences, the morphological data set could easily be over-
whelmed by the large number of molecular characters. Miyamoto (1985)
suggested that when the numbers of characters in each data set become
too unbalanced, character weights could be assigned so that each data set
would exert the same total influence. A potentially more serious limitation
is illustrated in the following example. Suppose that a given data set for
taxa A, B, C, and an outgroup may contain only two types of characters:
"type I" characters that support the tree ((AB)C) and "type II" characters
that support the tree (A(BC)). Further, suppose that the distribution of
character types is 10:0 in data set X (i.e., 10 characters of type I, and 0
characters of type II) and 5:5 in data set Y. The most-parsimonious tree
for the combined data set, ((AB)C), requires five extra steps, and IMF
WHEN ARE PHYLOGENY ESTIMATES INCONCRUENT? 317

equals [5 - (0 + 5)]/5 = 0. We would therefore conclude that there is

no between-data-set incongruence, even though half of the characters in
data set Y contradict the tree unanimously supported by the characters in
data set X. This conclusion seems unreasonable.
M. Miyamoto (in personal communication to Kluge, 1989) suggested an
alternative procedure for partitioning between- from within-data-set in-
congruence. As for Farris and Mickevich's method, minimal trees are first
constructed for each data set treated separately. However, rather than
combining the two data sets, each data set is mapped onto the tree com-
puted for the other data set (the possibility of equally parsimonious trees
is ignored for the moment). Let F(A*b) denote the number of extra steps
required to explain the evolution of data set A on tree b, a minimal tree
for data set B. The within-data-set incongruence is calculated as before:
iw = F(X*x) + F(Y*y),
but the total incongruence is defined as
# = F(X*y) + F(Y*x).
The extra length due to incongruence between two data sets X and Y,
expressed as a proportion of the total incongruence, is
IM = (IT - iw)/ir
(Kluge, 1989). Thus, if tree y explains data set X nearly as well as does
tree x, and if tree x explains data set Y nearly as well as does tree y, then
the incongruence value will be small, regardless of how dissimilar trees x
and y are. Miyamoto's technique therefore survives one of the objections
to consensus tree approaches raised above. A slight complication arises if
there is more than one minimal tree for one or both data sets. In this case,
a data set may require a different length depending on which tree from
the opposite data set is chosen. One could use either the best-fitting tree,
or average the lengths over all of the trees; the first alternative seems more
in keeping with the spirit of parsimony.
Unfortunately, IM shares some of the undesirable properties of IMF.
Modifying slightly the example given above (M. Miyamoto, personal com-
munication), suppose that the distribution of character types I:II is 9:0 in
data set X and 5:4 in data set Y. Tree ((AB)C) is then the minimal tree
for both data sets, with 7M equal to [(0 + 4) - (0 + 4)]/4 = 0. As for
IMF, it does not seem reasonable to conclude that there is no between-
data-set incongruence when nearly half of the characters in one data set
contradict a tree supported by all of the characters in the other data set.
Furthermore, if the distribution of character types were instead 4:5 in data
set Y, the most-parsimonious tree for Y would be (A(BC)), with which
all of the characters in X would be incongruent. In this case, IM = [(9 +
5) - (0 + 4)]/(9 + 5) = 0.71. The relatively minor change from a 4:5 to
a 5:4 distribution of character types I:II changes the between-data-set in-
congruence measure from none to over 70%! Because of these consider-
318 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

ations, it is probably unreasonable to expect the complex notion of be-

tween-data-set incongruence to be adequately captured by a simple index.

A MULTIFACETED APPROACH TO THE STUDY OF CONGRUENCE: AN

EXAMPLE USING KLUGE'S EPICRATES DATA

Kluge (1989) examined congruence in two data sets for snakes of the genus
Epicrates: a biochemical data set composed of 24 skin and scent gland lipid
characters (Tolson, 1987), and a morphological data set composed of 53
skeletal and external phenotypic characters. Although the biochemical data
are not, strictly speaking, "molecular," this example is particularly useful
for illustrating many of the concepts discussed above. Kluge calculated
minimal trees separately for the morphological and biochemical data (first
two trees in Fig. 14-13 and first 10 trees in Fig. 14-14, respectively; see
also consensus trees in Figs. 14-15a-c) and for the combined data set (see
Kluge, 1989). The morphological, biochemical, and combined data sets
required 35, 4, and 44 extra steps, respectively. The Mickevich-Farris in-
congruence index (IMF) is thus equal to [44 - (35 + 4)]/44 = 0.114. On
this basis, Kluge (1989:16-17) concluded:
Only 11.4% (5/44) of the total character incongruence is due to the disparity
between data sets . . . Thus, disagreement between the biochemical and mor-
phological characters is small relative to incongruence among the characters
within each set. The two qualitatively independent sources of information are
not strikingly contradictory in this example.
Although this conclusion seems reasonable on the surface, it deserves
further scrutiny. For example, the group (E. chrysogaster, E. exsul) is
supported by six unique and unreversed synapomorphies on all of the
shortest trees for the biochemical data. However, this group is well sep-
arated on both of the shortest trees for the morphological data, with E.
exsul more closely related to five other taxa than it is to E. chrysogaster.
Also, the group (E. inornatus, E. subflavus) is supported by three unique
and unreversed synapomorphies on both minimal trees for the morphol-
ogical data but does not appear in half of the minimal trees for the bio-
chemical data. In order to examine the question of congruence in Epicrates
in more detail, I have conducted a number of additional analyses that
extend those of Kluge (1989). Many of these analyses were facilitated by
features new to version 3.0 of PAUP (Swofford, 1991).

The Incongruence Index IM

The incongruence index 1M can be calculated using PAUP by first "ex-
cluding" the biochemical character set from consideration and finding min-
imal trees for the remaining data in order to obtain the within-data-set
incongruence for the morphological characters. The incongruence of the
biochemical characters with the trees obtained from the morphological
WHEN ARE PHYLOGENY ESTIMATES INCONGRUENT? 319

data can be calculated by reincluding the biochemical character set, then

excluding the morphological character set and requesting the length re-
quired by the biochemical characters on the same set of trees. An analogous
procedure is used to quantify incongruence within the biochemical data
set and incongruence of morphological characters on the trees computed
for the biochemical data. For the Epicrates data, the biochemical characters
require 34 and 36 steps, respectively, on the two minimal trees for the
morphological data (Fig. 14-13) and the morphological data require from
105 to 116 steps on the 10 minimal trees for the biochemical data (Fig. 14-
14). Subtracting the minimum possible number of steps for each data set,
we find that at least 34 - 24 = 10 extra steps are required by the bio-
chemical data on the "morphological" trees and at least 105 - 65 = 40
extra steps are required by the morphological characters on the "biochem-
ical" trees. Thus, IM equals [(10 + 40) - (35 + 4)]/(10 + 40) = 0.220,
which suggests about twice as much between-data-set incongruence as does
IMF-

Consensus Trees and Indices

A strict consensus tree for the two shortest morphological and 10 shortest
biochemical trees is shown in Figure 14-15d. Relatively little structure is
preserved, as reflected by the low consensus index values in Table 14-1.
The Adams consensus tree (Fig. 14-15e) is somewhat more resolved, how-
ever. The presence of a (E. chrysogaster, E. exsul) group in the Adams
consensus may seem surprising given the marked separation of these two
taxa on both morphological trees. However, this pair of taxa consistently
nests within a larger group containing all taxa exclusive of E. cenchria and
E. angulifer, and is therefore retained.
One of several largest common pruned trees for the same 12 trees is
shown in Figure 14-15f. (Others can be obtained by substituting E. inor-
natus for E. subflavus and/or E. fordii for either E. gracilis or E. monensis.)
While it is useful to know that all 12 of the trees agree on the relationships
for these subsets of taxa, it is also disheartening that obtaining a common
pruned tree requires the removal of nearly half of the taxa.

Suboptimal Trees

As discussed above, incongruence between two data sets may be more

apparent than real if a tree can be found that is at most slightly suboptimal
for both data sets. In the case of parsimony analysis, it is relatively simple
to determine how many near-minimal trees exist by examining tree-length
frequency distributions. Because the number of rooted binary trees for 10
taxa is so large (= 34,459,525), I approximated the shape of the frequency
distribution of tree lengths by computing the lengths of 10,000 randomly
chosen trees for each data set (Fig. 14-16). The distributions are strongly
left-skewed for both, indicating that both contain strong phylogenetic signal
Figure 14-13 Twenty-five trees of length <104 steps for the Ep/crates morphological data. Length required by
the biochemical characters on each tree is shown in parentheses. Abbreviations: A, E. angu//fer; CE, E. cenchria;
CH, £. chrysogaster; E, E. exsul; F, E. fore//';; G, E. gracilis; M, E. monensis; I, E. inornatus; ST, E. striatus; SU,
E. subflavus.
Figure 14-14 Forty-one trees of length <29 steps for the Epicrates biochemical data. Length required by the
morphological characters on each tree is shown in parentheses. Abbreviations as in Figure 14-13.
324 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Figure 14-15 Consensus trees for Epicrates example. Strict and Adams consensus
(a) of two minimal trees for morphological data. Strict (b) and Adams (c) consensus
of 10 minimal trees for biochemical data. Strict (d) and Adams (e) consensus of two
minimal trees for morphological-data plus 10 minimal trees for biochemical data.
A largest common pruned subtree (f) for the set of trees used for (d) and (e). Abbre-
viations as in Figure 14-13.
Table 14-1 Consensus Indices for Consensus Trees in Figure 14-15
Tree
Consensus Index a b c d e
Colless CF 0.875 0.625 0.750 0.375 0.625
Mickevich-Platnick Pm 0.886 0.545 0.659 0.455 0.614
(= Colless' weighted CF)
Mickevich CI 0.800 0.350 0.550 0.250 0.500
Rohlf Cf, 0.970 0.559 0.719 0.472 0.688
Rohlf C/2 8.71 x 10~8 3.05 x 10-" 2.61 x 10-7 8.23 x 10-= 7.84 x 10-7
326 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Figure 14-16 Approximate frequency distributions of tree lengths for Epicrates mor-
phological (a) and biochemical (b) data. Each distribution was obtained by random
sampling of 10,000 trees under a model in which every possible tree is equally
likely.
(see Chapter 13). However, it is apparent from Figure 14-17, which shows
the exact frequencies of tree lengths in the left tails of the distributions,
that signal strength is a relative concept. For example, the vast majority
of trees are longer than 120 steps for the morphological data (Fig. 14-16a),
but there are 4,961 trees of length 120 or less (Fig. 14-17a). Obviously,
there is no set of trees that is clearly better than the rest for either
data set.
In order to determine whether trees exist that explain both data sets
reasonably well, I initially decided to consider all trees that were less than
5% longer than the shortest trees for each data set. There are 25 trees of
length less than or equal to 104 steps for the morphological data (Fig. 14-
13) and 41 trees of length less than or equal to 29 steps for the biochemical
data (Fig. 14-14). Surprisingly, these two sets of trees contain no tree in
common! Additional analysis reveals that intersecting trees are not found
until 105-step trees are included for the morphological data (two trees of
length 28 steps for the biochemical data), or 30-step trees are included for
the biochemical data (three trees of length 104 steps for the morphological
data); these trees can be found in Figures 14-13 and 14-14.
Are phylogenies estimated for Epicrates from the two data sets incon-
gruent? I will leave the final answer to this question to the reader, but I
would merely suggest that the issue is much more complex than Kluge's
(1989) assessment. I find it mildly distressing that the best 25 trees for one
data set and the best 41 trees for another data set do not share even one
tree in common. On the other hand, the number of trees in the union of
all trees less than or equal to 105 steps for the morphological data, and all
trees less than or equal to 30 steps for the biochemical data, is still a tiny
fraction (188/34,459,525 = 0.00000546) of the total number of trees for 11
taxa. This observation suggests that there may indeed be substantial con-
gruence between the two data sets, but that "congruence" is not quite
what we had hoped it would be.
WHEN ARE PHYLOGENY ESTIMATES INCONCRUENT? 327

Figure 14-17 Exact frequency distributions of tree lengths for Epicrates morphol-
ogical (a) and biochemical (b) data for lengths in the left tail of each distribution.

DISCUSSION AND CONCLUSIONS

That I have failed to provide definitive answers to the question posed in

the title of this paper should not be surprising given the complexity of the
problem. When phylogenies estimated from multiple data sets disagree, I
believe that we should use a multiplicity of approaches to evaluate the
incongruence, marshaling evidence for groups that are well supported and
identifying those which should be subjected to further scrutiny. In addition
to the techniques discussed in this chapter, additional methods for assessing
the reliability of phylogenetic hypotheses (Penny and Hendy, 1986; Fel-
senstein, 1988; West and Faith, 1990) must be incorporated. A large num-
ber of researchers is just beginning to acquire molecular data for the purpose
of "evaluating congruence of molecular phylogenies with morphologically
based classifications;" it is imperative that these studies entail more than
a simple comparison of the optimal trees for each data set.

To Combine or Not to Combine

One of the most controversial issues involving the handling of multiple

data sets is whether or not they should be combined into a single data set
for analysis. Kluge (1989) takes a strong position that classifications should
be based on "total evidence." In his view, the preferred classification
corresponds to the most-parsimonious tree for a data set containing every
independent character from every available source of information. Miya-
moto (1985) has also argued that most-parsimonious trees for combined
data sets have maximum explanatory power and are therefore to be
preferred.
Although there are situations in which combining data sets can be useful,
I find compelling reasons to analyze logically independent data sets sep-
arately as well. Virtually all methods of phylogenetic inference assume
328 PHYLOCENETIC ANALYSIS OF DNA SEQUENCES

independence of characters. This assumption is difficult if not impossible

to verify, but it is usually much more plausible for characters in different
data sets than for characters within the same data set. For example, the
discovery that trees constructed separately from sequences for four differ-
ent genes agree in most respects would inspire a reasonably high level of
confidence in the conclusions. The corroboration provided by the inde-
pendent estimates would not be available if the data sets were combined.
On the other hand, the observation that three of the data sets agree com-
pletely with each other but disagree strikingly with the fourth would suggest
an interesting problem for subsequent molecular evolutionary studies that
could easily be missed if analyses were restricted to a pooled data set.
The problem of potential nonindependence can clearly be seen in the
Epicrates example. Six lipid-compound characters are present in E. chry-
sogaster and E. exsul that are absent in the remaining taxa. Only one
biochemical character (7) is incompatible with these six characters. A (E.
chrysogaster, E. exsul) clade is therefore strongly supported by the bio-
chemical data set: the shortest tree on which this group does not appear
is 33 steps long, five steps longer than a minimal tree for the biochemical
data. In contrast, this group is rather strongly rejected by the morphological
data set: the shortest tree containing a (E. chrysogaster, E. exsul) clade
requires 104 steps, four steps longer than a minimal tree. When the two
data sets are combined, the cost of rejecting a (E. chrysogaster, E. exsul)
group (six parallelisms involving independent acquisition of the same lipid
compound in different lineages) outweighs the cost of including it (the four
extra morphological steps cited above). Thus, the (E. chrysogaster, E.
exsul) clade is included in the minimal trees for the combined data set.
Furthermore, as Kluge (1989) emphasizes, it is "confirmed" by six "unique
and unreversed diagnostic synapomorphies." However, as just demon-
strated, this diagnosis is costly; four additional homoplasies must be pos-
tulated for the morphological data in order to preserve the (E. chrysogaster,
E. exsul) clade in the trees for the combined data.
Although it is impossible to prove without a complete characterization
of the genetic basis of the lipid variation, the incongruence between the
biochemical characters supporting a (E. chrysogaster, E. exsul) grouping,
and the morphological characters at least suggests the possibility that the
biochemical characters are not truly independent. This prospect might have
been overlooked entirely had the data sets not been analyzed separately.

Do "Classes" of Data Even Exist?

Another reason given for restricting analyses to combined data sets is that
there are no truly recognizable "classes" of characters. Proponents of this
argument suggest that "data are data," that nothing is gained by classifying
data into categories such as molecular versus morphological, mitochondrial
versus nuclear, or even small-subunit versus large-subunit rRNA. This is
WHEN ARE PHYLOGENY ESTIMATES INCONGRUENT? 329

a philosophical position that I find unreasonable but cannot entirely refute.

Presumably, even the most extreme advocate of this position would agree
that the genes responsible for differentiation of skeletal elements are in-
dependent of the gene coding for lactate dehydrogenase. For the reasons
discussed above, I believe that the advantages of comparing independent
estimates of phylogeny outweigh the disadvantages.

Should We Ever Consider Suboptimal Trees?

Ardent defenders of the principle of maximum parsimony sometimes object

to any consideration of nonminimal trees, suggesting that in doing so, one
begins to "slide down a slippery slope" that does not end until every
possible tree has been accepted as a candidate. Although I support the use
of parsimony as a criterion for evaluating trees, I believe that much can
be learned from approaches that examine nonminimal trees as in the Ep-
icrates example above. Obviously, if one of the minimal trees for the
morphological data set were the "true" tree and only the biochemical data
were available, we would reject the true tree if we failed to consider
nonminimal trees for the biochemical data.
I take the admittedly non-Popperian position that an ambiguous solution
that contains the truth is, in many situations, preferable to an unambiguous
solution that is wrong. If the ultimate goal is only to erect a classification
for a group of taxa, then I accept the argument that a fully resolved, most-
parsimonious tree is useful as a bold, testable hypothesis. It is becoming
more and more common, however, for a systematist's hypothesis to become
a comparative biologist's assumption (see, e.g., Donoghue, 1989 and ref-
erences cited therein). In this context, systematists do a disservice to their
fellow biologists if they do not provide any clues as to how well alternative,
slightly suboptimal trees might explain a given set of data. For example,
knowledge that a conclusion regarding adaptation receives some support
from all trees within three steps of the shortest may be much more satisfying
than the discovery that the same conclusion is strongly supported by the
most-parsimonious tree, but decisively rejected by a tree one step longer.

ACKNOWLEDGMENTS

I am especially grateful to Mike Miyamoto and David Maddison for dis-

cussion and constructive criticism of the ideas presented herein. I also thank
Joel Cracraft, Wayne Maddison, and participants in the University of Il-
linois Molecular Evolution Journal Club for their comments, and Jim Car-
penter for permission to publish results based on his "RopalO" data set.
Finally, I thank Arnold Kluge and Peter Tolson for their efforts in obtaining
the data on Epicrates used in the example.
330 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

REFERENCES

Adams, E. N. III. (1972) Consensus techniques and the comparison of taxonomic

trees. Syst. Zool. 21:390-397.
Adams, E. N. III. (1986) N-trees as nestings: complexity, similarity, and consensus.
/. Classif. 3:299-317.
Archie, J. W. (1989) Homoplasy excess ratios: new indices for measuring levels
of homoplasy in phylogenetic systematics and a critique of the consistency
index. Syst. Zool. 38:253-269.
Barthelemy, J.-P., and F. R. McMorris. (1986) The median procedure for n-trees.
J. Classif. 3:329-334.
Barthelemy, J.-P., and B. Monjardet. (1981) The median procedure in cluster
analysis and social choice theory. Math. Soc. Sci. 1:235-267.
Bremer, K. (1990) Combinable component consensus. Cladistics 6:369-372.
Bron, C., and J. Kerbosch. (1973) Algorithm 457: finding all cliques of an undi-
rected graph. Comm. ACM 16:575-577.
Brown, E. K.,and W. H. E. Day. (1984) A computationally efficient approximation
to the nearest neighbor interchange metric. J. Classif. 1:93-124.
Camin, J. H., and R. R. Sokal. (1965) A method for deducing branching sequences
in phylogeny. Evolution 19:311-326.
Carpenter, J. M. (1988) Choosing among multiple equally parsimonious clado-
grams. Cladistics 4:291-296.
Colless, D. H. (1980) Congruence between morphometric and allozyme data for
Menidia species: a reappraisal. Syst. Zool. 29:288-299.
Colless, D. H. (1981) Predictivity and stability in classifications: some comments
on recent studies. Syst. Zool. 30:325-331.
Cracraft, J. (1983) Commentary. Pp. 456-467 in Perspectives in Ornithology (A. H.
Brush and G. A. Clark, Jr., eds.). Cambridge University Press, Cambridge.
Crozier, R. H., P. Pamilo, R. W. Taylor, and Y. C. Crozier. (1986) Evolutionary
patterns in some putative Australian species in the ant genus Rhytidoponera.
Aust. J. Zool. 34:535-560.
Day, W. H. E. (1983) The role of complexity in comparing classifications. Math.
Biosci. 66:97-114.
Day, W. H. E., and F. R. McMorris. (1985) A formalization of consensus index
methods. Bull. Math. Biol. 47:215-229.
Donoghue, M. J. (1989) Phylogenies and the analysis of evolutionary sequences,
with examples from seed plants. Evolution 43:1137-1156.
Estabrook, G. F., C. S. J. Johnson, and F. R. McMorris. (1976) A mathematical
foundation for the analysis of cladistic character compatibility. Math. Biosci.
29:181-187.
Estabrook, G. F., F. R. McMorris, and C. A. Meacham. (1985) Comparison of
undirected phylogenetic trees based on subtrees of four evolutionary units.
Syst. Zool. 34:193-200.
Faith, D. P. (1988) Consensus applications in the biological sciences. Pp. 325-332
in Classification and Related Methods of Data Analysis (H. H. Bock, ed.).
Elsevier Science Publishers B. V., Amsterdam.
Farris, J. S. (1970) Methods for computing Wagner trees. Syst. Zool. 19:83-92.
Farris, J. S. (1973) On comparing the shapes of taxonomic trees. Syst. Zool. 22:50-
54.
Farris, J. S. (1988) Hennig86, Version 1.5. Distributed by the author, Port Jefferson
Station, N. Y.
WHEN ARE PHYLOGENY ESTIMATES INCONGRUENT? 331

Farris, J. S. (1989) The retention index and the rescaled consistency index. Cladistics
5:417-419.
Farris, J. S., A. G. Kluge, and M. J. Eckardt. (1970) A numerical approach to
phylogenetic systematics. Syst. Zool. 19:172-191.
Felsenstein, J. (1978) The number of evolutionary trees. Syst. Zool. 27:27-33.
Felsenstein, J. (1988) Phylogenies from molecular sequences: inference and reli-
ability. Ann. Rev. Genet. 22:521-565.
Felsenstein, J. (1990) PHYLIP Manual, Version 3.3. University Herbarium, Uni-
versity of California, Berkeley.
Finden, C. R., and A. D. Gordon. (1985) Obtaining common pruned trees. J.
Classif. 2:255-276.
Funk, V. A., and D. R. Brooks. (1990) Phylogenetic Systematics as the Basis of
Comparative Biology. Smithsonian Institution Press, Washington, D.C.
Gordon, A. D. (1980) On the assessment and comparison of classifications. Pp.
149-160 in Analyse de Donnees et Informatique (R. Tomassone, ed.). IN-
RIA, Le Chesnay.
Hendy, M. D., C. H. C. Little, andD. Penny. (1984) Comparing trees with pendant
vertices labelled. SIAMJ. Appl. Math. 44:1054-1065.
Hendy, M. D., M. A. Steel, D. Penny, and I. M. Henderson. (1988) Families of
trees and consensus. Pp. 355-362 in Classification and Related Methods of
Data Analysis (H. H. Bock, ed.). Elsevier Science Publishers B. V., Am-
sterdam.
Hennig, W. (1966) Phylogenetic Systematics. University of Illinois Press, Urbana.
Hillis, D. M. (1987) Molecular versus morphological approaches to systematics.
Ann. Rev. Ecol. Syst. 18:23-42.
Jarvis, J. P., J. K. Luedeman, and D. R. Shier. (1983) Comments on computing
the similarity of binary trees. /. Theor. Biol. 100:427-433.
Kluge, A. G. (1989) A concern for evidence and a phylogenetic hypothesis of
relationships among Epicrates (Boidae, Serpentes). Syst. Zool. 38:7-25.
Kluge, A. G., and J. S. Farris. (1969) Quantitative phyletics and the evolution of
anurans. Syst. Zool. 18:1-32.
Le Quesne, W. J. (1969) A method of selection of characters in numerical tax-
onomy. Syst. Zool. 18:201-205.
Maddison, D. R. (1991) The discovery and importance of multiple islands of most-
parsimonious trees. Syst. Zool. 40:in press.
Maddison, W. P. (1989) Reconstructing character evolution on polytomous cla-
dograms. Cladistics 5:365-377.
Margush, T., and F. R. McMorris. (1981) Consensus n-trees. Bull. Math. Biol.
43:239-244.
McMorris, F. R., D. B. Meronk, and D. A. Neumann. (1983) A view of some
consensus methods for trees. Pp. 122-126 in Numerical Taxonomy (J. Fel-
senstein, ed.). Springer-Verlag, Berlin.
McMorris, F. R., and D. Neumann. (1983) Consensus functions defined on trees.
Math. Soc. Sci. 4:133-136.
Mickevich, M. F. (1978) Taxonomic congruence. Syst. Zool. 27:143-158.
Mickevich, M. F., and J. S. Farris. (1981) The implications of congruence in
Menidia. Syst. Zool. 30:351-370.
Mickevich, M. F., and N. I. Platnick. (1989) On the information content of clas-
sifications. Cladistics 5:33-47.
Miyamoto, M. M. (1985) Consensus cladograms and general classifications. Clad-
istics 1:186-189.
332 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Nelson, G. J. (1979) Cladistic analysis and synthesis: principles and definitions,

with a historical note on Adanson's Families des Plantes (1763-1764). Syst.
Zool 28:1-21.
Neumann, D. A. (1983) Faithful consensus methods for n-trees. Math. Biosci.
63:271-287.
Page, R. D. M. (1989a) Comments on component-compatibility in historical bio-
geography. Cladistics 5:167-182.
Page, R. D. M. (1989b) COMPONENT User's Manual (Release 1.5). University
of Auckland, Auckland.
Penny, D., L. R. Foulds, and M. D. Hendy. (1982) Testing the theory of evolution
by comparing phylogenetic trees constructed from five different protein
sequences. Nature 297:197-200.
Penny, D. and M. D. Hendy. (1985) The use of tree comparison metrics. Syst.
Zool. 34:75-82.
Penny, D. and M. D. Hendy. (1986) Estimating the reliability of evolutionary
trees. Mol. Biol. Evol. 3:403-417.
Robinson, D. F. (1971) Comparison of labeled trees with valency three. /. Comb.
Theory 11:105-119.
Robinson, D. F., and L. R. Foulds. (1979) Comparison of weighted labelled trees.
Pp. 119-126 in Lecture Notes in Mathematics, Vol. 748. Springer-Verlag,
Berlin.
Robinson, D. F., and L. R. Foulds. (1981) Comparison of phylogenetic trees.
Math. Biosci. 53:131-147.
Rohlf, F. J. (1982) Consensus indices for comparing classifications. Math. Biosci.
59:131-144.
Rohlf, F. J., D. H. Colless, and G. Hart. (1983) Taxonomic congruence re-ex-
amined. Syst. Zool. 32:144-158.
Sanderson, M. J., and M. J. Donoghue. (1989) Patterns of variation and levels of
homoplasy. Evolution 43:1781-1795.
Schuh, R. T., and J. S. Farris. (1981) Methods for investigating taxonomic con-
gruence and their application to the Leptopodomorpha. Syst. Zool. 30:331-
351.
Schuh, R. T., and J. T. Polhemus. (1981) Analysis of taxonomic congruence among
morphological, ecological, and biogeographic data sets for the Leptopo-
domorpha (Hemiptera). Syst. Zool. 29:1-26.
Shao, K. (1983) Consensus Methods in Numerical Taxonomy. Ph.D. Dissertation,
State University of New York, Stony Brook.
Shao, K., and F. J. Rohlf. (1983) Sampling distribution of consensus indices when
all bifurcating trees are equally likely. Pp. 132-136 in Numerical Taxonomy
(J. Felsenstein, ed.). Springer-Verlag, Berlin.
Shao, K.-T., and R. R. Sokal. (1986) Significance tests of consensus indices. Syst.
Zool. 35:582-590.
Shao, K.-T., and R. R. Sokal. (1990) Tree balance. Syst. Zool. 39:266-276.
Simberloff, D. (1987) Calculating probabilities that cladograms match: a method
of biogeographical inference. Syst. Zool. 36:175-195.
Sokal, R. R., and F. J. Rohlf. (1981) Taxonomic congruence in the Leptopodo-
morpha reexamined. Syst. Zool. 30:309-325.
Stinebrickner, R. (1984) s-Consensus <rees and indices. Bull. Math. Biol. 46:923-
935.
Stinebrickner, R. (1986) s-Consensus index method: an additional axiom. /. Classif.
3:319-327.
WHEN ARE PHYLOGENY ESTIMATES INCONGRUENT? 333

Swofford, D. L. (1991) PAUP: Phylogenetic Analysis Using Parsimony, Version

3.0. Illinois Natural History Survey, Champaign.
Tolson, P. J. (1987) Phylogenetics of the boid snake genus Epicrates and Caribbean
vicariance theory. Occ. Papers Mus. Zoo/., Univ. Michigan 715:1-68.
Waterman, M. S., and T. F. Smith. (1978) On the similarity of dendrograms. J.
Theor. Biol. 73:789-800.
West, J. G., and D. P. Faith. (1990) Data, methods and assumptions in phylogenetic
inference. Aust. Syst. Bot. 3:9-20.
 15
Congruence Among Data Sets: A
Bayesian Approach

WARD C. WHEELER

When several independently derived data sets agree, producing the same
cladogram, we have high confidence in the result. Such complete agree-
ment, however, is rare and becoming rarer with the multiplication of mo-
lecular data sets.
Extreme disagreement has arisen between morphological and molecular
studies. Zimmer et al. (1989) have produced results on monocot origins
that contradict the morphological results of Donoghue and Doyle (1989).
Field et al. (1988) have suggested relationships within and among inver-
tebrate groups that counter the great majority of work on invertebrate
systematics. Mammal studies have produced further disagreement (Good-
man et al., 1985; Wyss et al., 1987), and studies of bird phylogeny (Cracraft,
1974; Sibley and Ahlquist, 1981) have had their differences. While these
conflicts are not limited to comparisons between molecules and morphol-
ogy, herein lie some of the most prominent cases. I do not mean to overstate
the extent of the disagreement or of the acrimony accompanying it. In fact,
congruence among data sets may well be more common than incongruence.
Nonetheless, in these works, there are important differences between hy-
potheses of relationships which require explanation.
When molecular data disagree with hypotheses based on morphology,
there are two common responses. Either one of the studies is dismissed
or some consensus is attempted. When consensus prevails, a consensus
"tree" is most often produced. There are, however, shortcomings to this
approach.
The most frequently used consensus procedures are Adams (1972) and
strict (Rohlf, 1982) consensus. Neither of these methods is entirely satis-
fying. The Adams procedure may generate groupings of taxa that are found
nowhere in the initial cladograms, while the strict method is so ruthlessly
conservative that all resolution may be lost by a shift in the position of a
single taxon. In the strict consensus, only groups present in all input trees
334
CONGRUENCE AMONG DATA SETS: A BAYESIAN APPROACH 335

are included. Lately, a modified procedure has become popular (Margush

and McMorris, 1981). The modified method includes groups that appear
in some fraction (greater than one half) of the cladograms. In this way,
"majority rule" consensus or percentage consensus trees are produced.
There is, however, a problem intrinsic to consensus analysis. These pro-
cedures tend to combine the weaknesses of the data sets; disagreements
based on few data are accorded equal weight with agreements based on a
great deal of information, since only topologies are compared (Miyamoto,
1985). A desirable method would combine the strengths of the data instead,
making allowances for the degree to which separate data sets support
different nodes. Here, I propose a method that attempts to do this using
Bayesian decision theory. The Bayesian approach draws on a philosophy
and an analytical procedure that directly apply to the problem of congru-
ence in phylogenetic analysis.

THE MONTY HALL PROBLEM

Before discussing the particulars of Bayesian decision theory, I would like

to start with an everyday example of the purpose and power of this type
of analysis, known as the "Monty Hall Problem." Imagine yourself on the
television game show "Let's Make a Deal." Monty Hall, the host, presents
you with three doors. Concealed behind one of these is a fabulous prize,
while the others hide objects considerably less attractive. You are invited
to choose a door, guessing which one hides the best prize. Suppose you
select door number 1. At this point, Monty Hall opens door number 3,
revealing an undesirable bauble (by the dictates of the game, Hall never
reveals the best prize). You are now faced with a dilemma: do you stay
with your original pick or do you go for what's behind door number 2?
The proper Bayesian would switch, and double her chances of winning.
This may, at first, seem counterintuitive. After all, the prize is equally
likely to be behind any of the doors. How are your odds improved by
switching? The trick comes from the use of additional information. Initially,
your pick ha$ a one in three chance of being correct. Hence, the probability
that the prize is behind one of the other doors is the complement of this—
two-thirds. When Monty Hall reveals that there is nothing worthwhile
behind one of the two doors you did not pick, the probability of that door
being the one with the desirable prize collapses to 0. The total probability,
however, that the two unchosen doors contain the best prize is unchanged;
and it remains two-thirds. At this point, there is a one-third chance you
have picked the correct door and a two-thirds chance that the prize is
behind the other door. Of course you should switch.
Put another way, if your first choice, door number 1, is correct and you
stand pat, you win. But if the desirable prize lies behind door number 2
or number 3 you must switch to win. Hence, by switching you win two out
336 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

of three times. It is true that your chances of finding the prize behind door
number 1 do not change when Monty Hall discloses the prize behind door
number 3. Eut your chances of finding the one prize behind door number
2 increase dramatically.
The Bayesian decision to switch is mandated by two conditions: (1) the
prior information that no door is more likely than any other to contain the
desirable prize; and (2) the information gathered by observing the contents
of one of the doors not chosen. These are the basic components of a
Bayesian analysis, the information prior to the observation and the ob-
servations themselves. This type of procedure is the only one to include
both kinds of information in the decision process.
The point I will make here is that in the use of morphological and
molecular data to determine the credibility of a cladogram, morphology
can provide the prior evidence to interpret molecular data.

DEFINITIONS OF PROBABILITY

Since this type of analysis deals with the probabilities of phylogenetic

schemes, it is necessary to first define what is meant by the probability of
a cladogram. There are three ways of defining probability: logical, empir-
ical, and subjective (Hartigan, 1983). A logical probability is a rational
degree of belief in a hypothesis in light of certain evidence, whereas an
empirical probability is a statement of fact about the world. The third type
of definition, subjective, is similar to the logical in that it represents a
degree of belief, but it is an individual degree of belief. Two people may
assign different probabilities based on the same evidence.
An example of these different concepts of probability can be found in
betting ratios. Hartigan (1983) makes the point that a logical bettor (Baye-
sian) wagers what they ought to, while the empiricist will bet what is
profitable in the long run, and the subjectivist will bet what they are willing.
In the procedure described here, I will use the logical definition of prob-
ability. The probability of a particular cladogram, then, is our degree of
belief in that cladogram, in light of evidence from molecular and mor-
phological characters.

BAYESIAN INFERENCE AND DECISION THEORY

Just as there are three types of probability definitions, there are, broadly
speaking, three types of inference: "classical," likelihood, and Bayesian.
I am using classical inference to refer to most of the standard methods of
statistical inference, those involving the calculation of p values and con-
fidence intervals to approximate final probabilities. Classical methods have
at least two shortcomings: first, they assume that the uncollected data can
affect the distribution of the estimates; and second, they base their con-
CONGRUENCE AMONG DATA SETS: A BAYESIAN APPROACH 337

elusions in part on the manner in which the data were collected. Thus, if
the collection procedure is not known, or is modified during the gathering
of observations, p values have no meaning, and classical methods are
invalidated. Likelihood methods, in contrast to classical methods, rely
solely on the observed data; their conclusions are not influenced by the
way in which the data are gathered. These methods, however, do not make
use of all relevant information. Specifically, they ignore information about
prior distributions. This omission is quite deliberate. Due to the philo-
sophical difficulties in determining priors—they may be entirely subjec-
tive—likelihood approaches pointedly avoid their use, or at least assume
that all priors are equal and therefore irrelevant. Bayesian procedures, not
bothered by these restrictions, extend likelihood through the inclusion of
prior information.
All Bayesian procedures rely on an elementary relation of conditional
probability (Bayes, 1763):

The "final" or posterior probability that the parameter @ has the value 0,
given the data D is constructed by multiplying the probability of the data
given 0,-, pr(D|0,), with its prior probability, pr(0,), and normalized by the
sum of these terms for all values of 6 (dk).
If tne priors are known or can be reasonably calculated, all statisticians
would agree that Bayesian methods should be used to determine the best
estimate of ®. Some priors seem inherently reasonable. An example would
be the proportion of defective items on an assembly line. In this case, the
performance of the line in the past gives a prior probability of the future
production of defective items. Yet, when priors are less clearly linked to
observation, there is disagreement about their use. In fact, those who hold
to likelihood argue that in all but the most trivial cases, prior probabilities
cannot be calculated; thus only likelihoods (or their ratios) can be used.

THE SENSITIVITY OF THE RESULT TO THE PRIOR

One way to measure the effect of prior probability is through the use of
simplex space. Take, for example, a parameter 0, which may have three
values, 0j, 92, or 63. The sensitivity of the final probability to the prior
probabilities can be graphed (Fig. 15-1). The space is divided into three
regions. These contain the prior probabilities that determine the choice of
61, 92, or 03 as the estimate of @. The values of the priors are plotted
(pr(0j) + pr(02) + Pr(0s) = 1) and their propinquity to the decision lines
noted. If this point is very close to one or several of these lines, small
variations in the priors can affect the estimate of ®, undermining our faith
in the result. If, however, the priors map to areas well within a decision space,
338 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Figure 15-1 Simplex space of a

simple inference problem showing
the areas in which estimates 1, 2,
and 3 are chosen.
we have greater confidence in the conclusions. A more precise measure
of the robustness of results is offered by decision theory.

DECISION THEORY

A Bayesian decision is one which minimizes risk. In order to make this

type of decision, an appropriate loss function must be specified for all
possible values of 6. This loss function, coupled with prior probability and
data probability values, allows us to determine the cost or risk of any
decision. By definition, a decision is to accept some value 6, as the estimate
of a parameter @, or to embark upon a course of action 0, out of all possible
actions.
Unfortunately, the losses must be specified for all possible correct and
incorrect decisions. This task may be overwhelming or impossible, given
the complexity of the options and their costs. If, however, we are willing
to accept that all incorrect hypotheses are equally undesirable and all
correct hypotheses are equally desirable in that they are correct, a simple
loss function can be used.
In this type of function, the cost of a correct decision is 0 while the cost
of an incorrect choice is one (Fig. 15-2). The risk of a decision, then, is
the sum of the cost of all the decisions:

The risk of the decision i (R,) is equal to the sum over all / decisions of
the product of the cost of that decision (Cv—if the parameter has the value
i) and its final probability. With the simple cost matrix suggested above
(Fig. 15-2), the decision risk is merely the complement of its final proba-
bility. Since
Cy = 1 if i * /
C!t = 0 if i = ;,
the risk reduces to:
CONGRUENCE AMONG DATA SETS: A BAYESIAN APPROACH 339

Figure 15-2 A simple cost matrix

used in phylogenetic decisions with
"n" decisions possible.

for a single cladogram or:

for several cladograms {k} in i. By these manipulations, the risk of a decision

with a final probability of 0.9 would be 0.1. The minimum risk occurs when
the probability is maximized.
I propose that this notion of risk be used to evaluate the support for
cladograms. Possibly no single cladogram will stand out as having accept-
ably low risk; in these cases, several cladograms should be considered until
their collective risk is sufficiently low. By analogy with p values, we might
choose the "risk" of type-I error to be no greater than 5%. Here, we would
accept the most probable cladograms until 95% of the risk had been re-
moved.

THE PROCEDURE

Determination of Prior Probabilities (pr(0j)}

The first step in this type of analysis is to determine the prior probabilities
for all of the possible values of 6. Usually, prior probabilities are calculated
using known distributions of continuous or discrete variables. In this case,
however, the prior probability of a cladogram in a molecular analysis is its
ability to explain morphological data, which is an entirely discrete value.
There are many measures of the relative abilities of cladograms to explain
data, such as the length (number of evolutionary steps) of the scheme or
its consistency index (Kluge and Farris, 1969). Here, a length difference
of one step (or postulated event) will be considered equal to a difference
in probability of a factor of e, the base of natural logarithms. (Since these
probabilities will be normalized, they need not sum to unity.) This factor
of probability reflects the relative degree of belief in the different hy-
potheses before the molecular observations are made.
This assessment of probability may seem arbitrary at first. However, it
can be justified by the link between natural logarithms and phylogenetic
character weights (Farris, 1977). Of course, an individual investigator may
use different methods to determine prior probabilities based on their con-
340 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

fidence in that topology. In an unweighted parsimony analysis, all changes

increment the cladogram cost by one step; similarly here, all changes are
considered to be equally probable until there is evidence to the contrary.
Thus the prior probability of cladogram i with a length L, from some
previous analysis is:

or to sum to unity:

Determination of Data Probability {pr(O = 6,\D)}

The second phase of the procedure involves determining the probability
of the data given a certain 9,, or topology. Any method that yields this
value can be used. There are three methods, however, which seem espe-
cially appropriate: unweighted or maximum parsimony, weighted parsi-
mony, and maximum likelihood.
The first of these may seem, initially, to be unsuited to statistical inter-
pretation. As with the establishment of priors, a factor of e change in
probability can be assigned to each event. In this way, all changes are
treated equally, and a quantitative degree of belief is determined for each
topology. As with Keynes' (1921) argument about the frequency of colored
balls in an urn, in the absence of knowledge to the contrary, we must
assume all events are equally likely. Total symmetry is a very simple model
of character change, but it offers a starting point for examining the data.
One effect of using the same probability factor in determining priors and
data probabilities is that the result is the same as if all the data had been
collected at one time and pooled. The absolute value of the natural log-
arithm of the final probabilities will be the lengths of cladograms con-
structed from these pooled data. The most probable (least risk) cladogram
will be the most-parsimonious.
From this entirely symmetrical model of character transformation, the
next step is to vary the weights assigned to different types of transformation.
Just as these models of character transformation are used in weighted
parsimony analyses, they can be used to assign data probabilities. Along
these lines, Farris (1977) and DeBry and Slade (1985) have offered a Dollo-
type method based on probabilistic thinking. In the analysis of nucleic acid
sequences, many have suggested that transitions should be weighted dif-
ferently from transversions (Brown et al., 1982). Specific cost ratios have
been proposed, with obvious probabilistic interpretations. Elsewhere, I
have put forward a combinatorial weight scheme for nucleic acids that
yields weights for different types of character transformation (Wheeler,
1990). In each of these methods, only the different types of change receive
varying probabilities. No assessments of the probability of change itself
occurring are attempted (in a character transformation matrix all au = 0).
Instead, these procedures rely on explicit minimization of "cost" values.
 ONGRUENCE AMONG DATA SETS: A BAYESIAN APPROACH 341

Figure 15-3 Markov type character trans-

formation model showing the probability
of change from one character state to an-
other.

Likelihood methods differ, in that they seek to incorporate the rate of

evolution. Not only the possible avenues of change, but the probability of
change is included in the analysis. Smouse and Li (1987) use this approach
in their examination of restriction fragment character change (and even
discuss prior probabilities from other studies). Markov character transfor-
mation models (Fig. 15-3) chart rates of evolution as diagonal values. Since
parsimony models lack these values, they offer no information on the
probability of change.
With the prior probabilities and the data probabilities in hand, the final
probability of each cladogram is calculated by equation (1).

Determination of Utility and Risk

The loss function suggested above and the final probabilities are used to
determine the risks of various single and compound decisions [equation
(3)]. A single cladogram, or several, may be accepted until risk is suffi-
ciently minimized.

An Example

In order to demonstrate this congruence procedure, the case of the mon-

ophyly of the Eumetabola will be examined in the light of morphological
(Hennig, 1969, 1981; Kristensen, 1975, 1981; Boudreaux, 1979) and mo-
lecular (Wheeler, 1989) evidence. There are 15 dichotomous unrooted
arrangements of the five insect taxa (Paleoptera, Orthoptera, Hemiptera,
Diptera, and Coleoptera, as shown in Fig. 15-4). In the morphological data
set, there were five presence/absence characters: wing flexion (neoptery),
the jugal bar, larval stemmata, holometaboly, and larval ocelli. The number
of evolutionary events required by each of the possible arrangements is
shown in Table 15-1. The sequence data of Wheeler (1989) were applied
to these same cladograms; their lengths are also shown in Table 15-1, along
with the weighted parsimony values. The procedure used to determine the
weighted cladogram lengths is described in detail in Wheeler (1990). The
likelihoods, or more accurately, their absolute values (Felsenstein, 1978,
1979), are also shown in this table.
342 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Figure 15-4 The 15 possible topologies for the Coleoptera, Diptera, Hemiptera,
Orthoptera, and Paleoptea, with Paleoptera as the outgroup.

The final probabilities determined from the three types of cladogram

costs and the morphologically based priors are shown in Table 15-2. Table
15-3 gives the topology decisions and associated risks.
The three types of data probabilities, maximum parsimony, weighted
parsimony, and likelihood, agree for the most part. Topology III is the
most favored (or at least one of the most favored hypotheses) in each of
the analyses. However, the likelihood method includes topology XV and
CONGRUENCE AMONG DATA SETS: A BAYESIAN APPROACH 343

Table 15-1 Cladogram Length

Sequence Data
Unweighted Weighted
Cladogram Morphology Parsimony Parsimony Likelihood
I 5 133 317 494
II 6 137 327 494
III 6 132 314 489
IV 7 135 321 494
V 8 140 332 494
VI 8 136 322 494
VII 7 137 327 494
VIII 8 139 327 494
IX 8 140 334 494
X 8 137 328 493
XI 8 134 321 493
XII 8 132 316 489
XIII 8 136 322 493
XIV 8 138 329 493
XV 8 133 318 490

Table 15-2 Final Probabilities

Unweighted Weighted
Cladogram Parsimony Parsimony Likelihood
I 4.5 x 10-' 1.2 x 10-' 1.5 x 10-2
II 3.0 x 10-3 2.0 x 10-6 5.5 x 10-3
III 4.5 x ID'1 8.7 x 10-' 8.1 x 10->
IV 8.2 x 10-3 2.9 x 10-4 2.0 x 10-3
V 2.0 x 10-5 1.8 x 10-" 7.4 x 10~4
VI 1.1 x 10-3 3.9 x 10-5 7.4 x 10-"
VII 1.1 x 10-3 7.2 x 10-7 2.0 x ID' 3
VIII 5.5 x 10~5 2.6 x 10~7 7.4 x 10-4
IX 2.0 x 10~5 2.4 x 10-10 7.4 x 10-"
X 4.1 x 10~4 9.7 x 10-8 2.0 x 10-3
XI 8.2 x 10-3 1.1 x 10-" 2.0 x 10-3
XII 6.1 x 10~2 1.6 x 10-2 1.1 x 10-'
XIII 1.1 x 10-3 3.9 x 10-5 2.0 x 1Q-3
XIV 1.5 x 10~4 3.6 x 10-8 2.0 x 10-3
XV 2.2 x 10-2 2.1 x 10-3 4.1 x 10~2

Table 15-3 Topology Decisions

Cladograms
Unweighted Weighted
Risk Level Parsimony Parsimony Likelihood
20% I, III III III
10% I, III I, III III, XII
5% I, HI, XII I, HI III, XII, XV
344 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

excludes topology I in the 5% risk decision, while the two parsimony-based

methods include I at the expense of XV.

CONCLUSIONS

At present there are two commonly used congruence procedures: global

parsimony and consensus (Miyamoto, 1985; Kluge, 1989). In a global par-
simony analysis, all of the characters are lumped together, and the most-
parsimonious solution for the combined data is chosen as the most-favored
hypothesis. Consensus procedures, on the other hand, treat the data sets
distinctly; individual characters and their combinations do not play a direct
role in the topology decision made from comparing the data sets. In this
respect, consensus procedures treat characters as if they were members of
noncomparable categories; hence, only topologies are compared. The method
presented here attempts to extract the best qualities of both approaches
through Bayesian analysis.
If the simple e factor transformation is used for both the prior distribution
and the determination of the data probability, the result will be the most-
parsimonious solution for the combined data. Since all characters are weighted
equally, it does not matter from which analysis (morphological or molec-
ular) they were derived. If, however, the investigator believes that char-
acter classes exist, or feels that some distinction can be made between the
types of data used, this method allows him to separately determine the
probabilities that influence the final decision.
The procedure draws its strength from this flexibility. These probabilities
offer the investigator a quantitative assessment of the degree of belief. If
several analyses favor a certain hypothesis, but each only marginally, the
final probability of that hypothesis should reflect the increased confidence
which comes from this corroboration.
The risk inherent in the decision to accept or reject topologies gives a
measure of the level of agreement between data sets. If there is little
discrimination, or the various observations differ greatly, no hypothesis
will have sufficiently low risk to be accepted by itself. In these cases, several
or many topologies will have to be considered to reduce the risk to a
satisfactory level.
Congruence is the final test of any hypothesis. The type of procedure
presented here gives a logical means of assessing the agreement and dis-
agreement between different sets of data that bear on the same phyloge-
netic question.

ACKNOWLEDGMENTS

I would like to acknowledge Kirk Fitzhugh, Paul Vrana, and especially

Elise Broach for reading drafts of this manuscript. I would also like to
CONGRUENCE AMONG DATA SETS: A BAYESIAN APPROACH 345

thank two anonymous reviewers. I thank Christopher Urheim for making

clear to me the subtleties of the "Monty Hall Problem." This research was
supported by the Alfred P. Sloan Foundation.

REFERENCES

Adams, E. N. III. (1972) Consensus techniques and the comparison of phylogenetic

trees. Syst. Zool. 21: 390-397.
Bayes, T. (1763) An essay towards solving a problem in the doctrine of chances.
Philos. Trans. R. Soc. 53: 370-418.
Boudreaux, H. B. (1979) Arthropod Phytogeny with Special Reference to Insects.
John Wiley & Sons, New York.
Brown, W. M., E. M. Praeger, A. Wang, and A. C. Wilson . (1982) Mitochondrial
DNA sequences of primates: tempo and mode of evolution. /. Mol. Evol.
18: 225-233.
Cracraft, J. (1974) Phylogeny and evolution of the ratite birds. Ibis 116: 494-521.
DeBry, R. W., and N. A. Slade. (1985) Cladistic analysis of restriction endonu-
clease cleavage maps with a maximum likelihood framework. Syst. Zool.
34: 21-34.
Donoghue, M. J., and J. A. Doyle. (1989) Phylogenetic studies of seed plants and
angiosperms based on morphological characters. Pp. 181-194 in The Hi-
erarchy of Life. Molecules and Morphology in Phylogenetic Analysis (B.
Fernholm, K. Bremer, and H. Jornvall, eds.). Elsevier Press Science Pub-
lishers B.V., Amsterdam.
Farris, J. S. (1977) Phylogenetic analysis under Dollo's law. Syst. Zool. 26: 77-
88.
Felsenstein, J. (1978) Cases in which parsimony or compatibility methods will be
positively misleading. Syst. Zool. 27: 401-410.
Felsenstein, J. (1979) Alternate methods of phylogenetic inference and their in-
terrelationship. Syst. Zool. 28: 49-62.
Field, K. G., G. J. Olsen, D. J. Lane, S. J. Giovannoni, M. T. Ghiselin, E. C.
Raff, N. R. Pace, and R. A. Raff. (1988) Molecular phylogeny of the animal
kingdom. Science 239: 748-753.
Goodman, M., J. Czelusniak, and J. E. Beeber. (1985) Phylogeny of primates and
other eutherian orders: a cladistic analysis using amino acid and nucleotide
sequence data. Cladistics 1: 171-185.
Hartigan, J. A. (1983) Bayes Theory. Springer-Verlag, New York.
Hennig, W. (1969) Die stammesgeschichte der Insekten. Seckenberg-Buchern. 49:
1-436.
Hennig, W. (1981) Insect Phylogeny. John Wiley & Sons, New York.
Kluge, A. G. (1989) A concern for evidence and a phylogenetic hypothesis of
relationships among Epicrates (Boidae, Serpentes). Syst. Zool. 38: 7-25.
Kluge, A. G., and J. S. Farris. (1969) Quantitative phyletics and the evolution of
anurans. Syst. Zool. 18: 1-32.
Kristensen, N. P. (1975) The phylogeny of hexapod "orders": a critical review of
recent accounts. Z. Zool. Syst. Evol. Forsch. 13: 1-44.
Kristensen, N. P. (1981) Phylogeny of insect orders. Ann. Rev. Entomol. 26: 135-
157.
346 PHYLOGENETIC ANALYSIS OF DNA SEQUENCES

Margush, T. and F. R. McMorris. (1981) Consensus n-trees. Bull. Math. Biol. 43:
239-244.
Miyamoto, M. M. (1985) Consensus cladograms and general classifications. Clad-
istics 1: 186-189.
Rohlf, F. J. (1982) Consensus indices for comparing classifications. Math. Biosci.
59: 131-144.
Sibley, C., and J. E. Ahlquist. (1981) The phylogeny and relationships of the ratite
birds as indicated by DNA-DNA hybridization. Pp. 301-335 in Evolution
Today (G. G. E. Scudder and J. L. Reveal, eds.) Carnegie-Mellon Uni-
versity, Pittsburgh.
Smouse, P. E., and W.-H. Li. (1987) Likelihood analysis of mitochondrial restric-
tion-cleavage patterns for the human-chimpanzee-gorilla trichotomy. Evo-
lution 41: 1162-1176.
Wheeler, W. C. (1989) Evolution and systematics of insect rDNA. Pp. 307-321
in The Hierarchy of Life. Molecules and Morphology in Phylogenetic Anal-
ysis (B. Fernholm, K. Bremer, and H. Jornvall, eds.). Elsevier Science
Publishers B. V., Amsterdam.
Wheeler, W. C. (1990) Combinatorial weights in phylogenetic analysis: a statistical
parsimony procedure. Cladistics 6: 269-275.
Wyss, A. R., M. J. Novacek, and M. C. McKenna. (1987) Amino acid sequence
versus morphological data and the interordinal relationships of mammals.
Mol. Biol. Evol. 4: 99-116.
Zimmer, E. A., R. K. Hamby, M. L. Arnold, D. A. Leblanc, and E. L. Theriot.
(1989) Ribosomal phylogeny and flowering plant evolution. Pp. 205-214 in
The Hierarchy of Life. Molecules and Morphology in Phylogenetic Analysis
(B. Fernholm, K. Bremer, and H. Jornvall, eds.) Elsevier Science Publishers
B. V., Amsterdam.
Index
Acrylamide sequencing stock solutions, 37, of homiotherm genomes, 133
40 of mitochondrial DNA, 133
Actinopterygian fish, 289 relative to evolutionary parsimony, 233
Adams consensus. See Consensus methods for salamanders and their outgroups,
and consensus trees 241-43
Adams-2 trees, 299 Bayesian
Additive tree, 93 decision theory, 335-44
Additivity, 261 inference, 337-44
Akaike's information criterion (AIC) Beckman, 26
definition, 269 Beneckia harveyi, 55 ribosomal RNA
relative to Kullback-Leibler information, alignment, 63
269 Bethesda Research Laboratories (BRL),
relative to maximum likelihood method, 23, 28-29, 32, 51
269 Bird, 289, 334
Aldridch, 32 Boehringer-Mannheim, 28-30
Alignment scores, 75, 83 Bootstrapping
of random sequences, 67-68 definition, 160
statistical distribution, 67-69 of distances, 118, 122
two-sequence problem, 60 of global parsimony trees, 201, 203-5
a Index, 310 problem of long edges attracting, 178-79
Ambystoma, 232 relative to trichotomy, 177-79
californiense, 223-29 as test of reliability, 118, 209, 290
Amino acid sequences, to test sampling error and convergence,
alignment, 60 160-61
evolutionary distance, 256 of transversion parsimony trees, 203,
mammal phylogeny, 175-76 206-8
tree length distribution, 279-80 versus jackknifing, 162
Amphibians, 289 Bos taunts, 81-84
Amphwxus, 291 Branch
Amphmma, 232 and bound algorithm, 157
means, 223-29 length variances
Andrias davidianus, 223-29 for five taxa, 257-58
Aneides, 232 for four taxa, 256-57
flavipunctatus, 223-29 by jackknifing, 255
Anthropoids, 189 by the least-squares method, 256, 258
Antlered deer, 12 relative to amount of sequence data,
Anurans, 290 260-62
Applied Biosystems (ABI), 48, 57 relative to second outgroup, 261-62
Artiodactyla, 84 for testing phylogenies, 258-59
Atomergic Chemical Corporation, 32 Brittle star, 291
Autoradiogram, 47
Average distance (linkage) method
[unweighted pair group method with
arithmetic means (UPGMA)],
algorithm, 91 Camin-Sokal parsimony criterion, 307
and inconsistency, 173 Carnivores, 175-77, 179
simulation analysis, 99-103, 106-8, 117 Catarrhines, 189
statistical test of, 117-18 C.B.S. Scientific Inc., 37
Character
methods, 9, 91, 93-94
BDH Chemicals, 32 weighting. See Weighting
Base composition biases, 129, 133-35, 141 Characters, sequence versus morphological,
calculation, 142 13,73

347
348 INDEX

Chemical sequencing for sequence alignments, 70, 75,

base-specific reactions, 24, 32-33 78-79, 83
kit, 29 "stopping" algorithm, 290-91
to obtain initiation points, 20-21, 25 for testing internal branch lengths, 258
protocol, 32-34 simulation
of shorter fragments, 33 character weighting, 147-53
solutions, 30-31 and compositional statistics, 132-33
stock chemicals, 32 compositional statistics versus distance
strand breakage, 24-25, 33 methods, 140-42
trouble-shooting, 25 compositional statistics versus
C index, 315 evolutionary parsimony, 136-40
c Index, 315 congruence and tree-length
CI index, 307, 325 distributions, 280
cin Index, 310 random data and tree length
C/j index, 308-9, 325 distributions, 284-86
C/2 index, 308-9, 325 of tree-making methods, 97-99
Clade, 296 Congruence
Cladistics, 4 Bayesian approach, 335-44
Cladogram, probability of, 336 Eumetabola data (example), 341-44
Classes of data, 344 character, 314, 318
Classical inference, 336-37 consistency indices, 315-16
Cliques, 296, 302 definition, 314-15
Cluster, 296 extension of methodological
Coleoptera, 341-42 parsimony, 11-12, 187
Cole-Parmer, 38 incongruence indices, 316-18
Combinable component consensus. See informativeness, 187-88
Consensus methods measures, 315-16
Common molecular versus morphological, 334
chimpanzee, 117, 119, 187 test of reliability, 11-14, 187-88,
pruned trees. See Consensus methods, 214-16
Consensus trees multifaceted approach (Epicrates
Comparative biology, 3 example), 318-26, 328
Compatibility method, 229 taxonomic, 296-314
and inconsistency, 173 consensus indices, 305-12
COMPONENT computer package, 305 consensus trees; 296-305
Compositional statistics definition, 296
algorithm, 131-36 tree comparison metrics, 312, 314
applications, 143 versus simulation analysis, 12
effects of mutation, 134 Consensus
effects of selection, 134-35 fork index, 306, 308, 325
evolutionary model, 133-35 indices
mutation versus selection pressure, 129 a index, 310
relative to base composition biases, 233 cin (Adams), 310
statistical test, 135-36 CI (Mickevich), 307, 308, 325
strengths/weaknesses, 131-33 CA (Rohlf), 308-9, 325
versus distance methods, 140-42 C/2 (Rohlf), 308-9, 325
versus evolutionary parsimony combining information and agreement,
computer simulation, 136-40 306-7
differences, 143 criticisms, 307, 310-12
model tree, 137 "false confidence" problem, 308
rates of failure/success, 138-39 general description, 305-6
similarities, 142 levels sum, 307
versus parsimony method, 244-45 meaning of magnitude, 310
Computer Pm (proportion of maximum total
programs information), 307, 325
for calculating distances, 123 s-consensus index, 310
for distance methods, 123 sensitivity to tree asymmetry, 307-9
MacClade, 292 total information, 307
for the neighborliness method, 265 type 1, 2, and 3 paradigms, 310
Phylogenetic Analysis Using Parsimony weighted consensus fork index, 307-8,
(PAUP), 188, 233, 284, 292 325
INDEX 349

Consensus (continued) Diptera, 341-42

methods, 78, 344 Distance methods, 79-80, 91-93, 185
Adams, 297-300, 304, 309, 334-35 disadvantages, 9-10, 130, 185-86,
combinable component, 300-1 196-97
criticisms, 311-12, 334-35, 344 and heterogeneous data, 273-74
and "islands" of parsimonious trees, loss of information, 174-75
312-14 statistical justification, 10
largest common pruned tree, 304-5 statistical tests of reliability, 258-59,
majority rule, 189, 297-98, 301, 303, 263-65
335 Distance-Wagner (DW) method
median consensus, 303-4 algorithm, 92-93
Nelson, 297-303 relative to number of nucleotides, 119-
and pooling data, 311, 327-28 20
strict, 297-99, 334-35 simulation analysis, 99-104, 108
and suboptimal trees, 312-14, 319, Distortion index (dT) (See symmetric-
326, 329 difference distance)
object invariants, 306 Distribution of tree lengths,
sequence, 70, 82-84 test of signal, 278-92
trees DNA
Adams-2 trees, 299 cloned, 20, 22
discussion of types, 296-305 core facilities
examples benefits, 47, 56, 57
Adams, 299-300, 324 services, 46
Nelson, 303 cosmid, 20, 22, 26-27
pruned, 304, 305, 315, 324 hybridization, 91, 117, 120, 216
strict, 298, 303-4, 324 lambda, 20
suboptimal trees, 314 plasmid, 26-27
general approach, 296-97 polymerase I (Klenow fragment), 18
median versus median binary, 300—4 purification
relationship to phylogeny, 311 alkaline extraction procedure, 25
Consistency index, 315-16 CsCl gradients, 25, 27
C, 315 protocols, 26-27
c, 315 solutions, 26-27
HER-maximum, 316 sequences
homoplasy excess ratio (HER), 316 increasing availability, 5, 59, 184-85
for primate phylogenies, 213 of large-subunit nuclear ribosomal
rescaled consistency index, 315-16 RNA, 223-29
retention index, 315-16 limitations for phylogenetic study,
Covarions, 148 179-80, 185
Covariotides, 148-49, 151, 153 phylogenetic potential, 59, 184
Cow, 289 signal versus noise, 278-92
Crinoid, 291 sequencing
Crossover metric, 314 automated, 45-58
a-Crystallin, 279-80, 290 band detection, 48-49
Cytochrome b, 289-90 benefits, 53, 57
Cytochrome c, 59, 113, 159, 176, 221 costs, 46
definition, 47
Darwinism description, 48
and climbing plants, 155-56 drawbacks, 53, 57
falsifiability of theory, 155-56 history, 47
as scientific theory, 155 labeling techniques, 48
Data probability present weaknesses, 19
determination, 340-41 sample output, 49, 54
for Eumetabola data, 341-43 on screen data analysis, 53, 57
DDBJ, 59 as a service, 46
Deaza-7-GTP, 51 strategies, 48-49, 52
Desmognathus, 232 technological advances, 18-19, 42
ochrophaeus, 223-29 versus manual, 50-55, 57
Deuterostomes, 290-91 of both complements, 6
Dicamptodon, 232 chemical method, 24-25, 29, 31-34
species, 223-29 collapse regions, 51, 53-54
350 INDEX

DNA (continued) for protein sequences, 256

of complete genes, 6-7 and saturation, 80, 82, 84
compression, 51 two-parameter model, 257
effective reading lengths, 37, 41, 51 variances, 256-57
enzymatic method, 35-36, 47 parsimony (EP) method
error rates, 52-53, 57 algorithm, 94, 131
film readers, 24, 42 simulation analysis, 110-15
manual statistical inconsistency, 273
description, 47 strengths/weaknesses, 131, 133
strategies, 51 tests of assumptions, 233, 239
in molecular systematics, 5 versus parsimony method, 131, 233,
overlap failures, 52-53 244-45
premature chain termination, 53 rate
random versus directional sequencing, for divergent domains, 232
19, 41-42 for salamanders, 241
sequencing gels, 37-41 versus time scale of divergence,
strategy for large inserts, 19-24 210-11, 221
verification, 52-53
Drosophila, 273
yakuba, 81 Falcon, 27
Drummond Scientific Co., 39 "False confidence" problem, 308
DuPont, 48-52, 57 Fibrinopeptides A and B, 159, 176
Dynamic Finding optimal trees
programming branch and bound algorithm, 157
of 5S ribosomal RNA, 63-65 number of trees, 157
of full sequences, 62-65 Fisher Scientific, 32
of maximum segments, 65 Fitch and Margoliash's (FM) method
of multiple sequences, 64-70 algorithm, 92
optimization technique, 62 relative to number of nucleotides,
weighting parsimony, 121 119-20
simulation analysis, 99-100, 105, 107,
109
Eastman Kodak, 32, 41 5'-end-labeling
185 ribosomal RNA gene, 50-52 problem of small RNAs, 27
Electrophoretogram, 48 protocol, 28-29
EMBL, 59 relative to restriction sites, 27
Enzymatic sequencing solutions, 28-30
current popularity, 18, 35 with [32P-"/]-ATP, 27-29
DNA polymerases, 18, 35 Fotodyne, Inc., 37, 41
kit, 35 Four-point condition, 93, 261, 263
obtaining initiation points, 19-20 Frog, 291
protocol, 36 Full sequence alignment
solutions, 31, 36 computation, 62-63
Epicrates, 318, 320-24 computer time, 63
Eppendorf, 28-29, 32 of 55 ribosomal RNA, 63-64
Erdos-Renyi law, 68-69
Escherichia coli g, Statistic, 283-92
DNA polymerase I (Klenow fragment), distribution, 287-88
18 effect of added signal, 289
55 ribosomal RNA alignment, 63-64 effect of sequence length, 289
Euler-Mascheroni constant, 68 real data versus random data, 190
Eumetabola, 341-42 Callus gallus, 81-84
Evolutionary Gamma distribution, 113, 115, 122
distance. See also Nucleotide substitution GenBank, 59, 79, 290
d, 97, 122 Gene
and divergence time, 80-82 frequency maximum likelihood (GFML)
gamma distribution, 122 method, 106-7
Jukes/Cantor, 95, 122 trees, 7, 123
Kimura, 97, 122 Gibbon, 117, 119, 187
one-parameter model, 257 Gingko biloba, 50-51, 57
p, 97, 122 Glires, 177, 179
INDEX 351

Global parsimony analysis Homoplasy excess ratio (HER), 316

of amino acid replacements, 201-2 Homo sapiens, 81-84, 117, 119, 187
with bootstrapping, 201, 203-5 Human, 117, 119, 187, 273, 291
of first and second codon positions, 197, Human Genome Initiative, 18, 59
201-2 Hylobates lar, 117, 119, 187
of mitochondrial DNA subsets, Hynobius sp., 223-29
190-94, 201, 203-5 Hypotheses
with near-most-parsimonious (1%) trees, and Charles Darwin, 156
191, 194 diversity of approaches in science, 180
with near-most-parsimonious (five-step) falsifiability of, 156
trees, 190-93 finding the limits of, 168,
of protein-coding genes, 191, 193, 203, nonevolutionary models, 162, 166
205 Popperian framework, 156
summary of results, 195, 203, 209-10 as rational explanations, 162
of transfer RNA genes, 191, 193, 203, and sociological and cultural factors, 156
205
Gorilla gorilla, 117, 119, 187
Great apes, 189 1M index, 317-19
Group, nontrivial, informative, but not IMF index, 316-18, 325
universal, 296 Incongruence of multiple data sets,
295-329
HER-maximum index, 316 incongruence indices, 316-18
Hadamard transformation recommendations, 327-29
advantages over maximum likelihood, Index
172-73 a, 310
bipartitions, 169-70 of dispersion (K)
loss of information, 174—75 definition, 230
probability of partition, 169-71 for salamander phylogeny, 234-35
and statistical inconsistency, 123 Inference. See also Phylogenetic inference
Hairpins, 53 definition of types, 336-37
"Hard" polytomies, 311 Informativeness. See also Reliability
Hemiptera, 341-42 between most- and near-most-
Hemoglobin, 59, 158-59, 176, 279-80, 289 parsimonious trees and bootstrap
a-hemoglobin, 158-59, 176, 279-80, 289 trees, 209-10
(3-hemoglobin, 158-59, 176 definition, 186-87
Hennig86 computer package, 305 of most-parsimonious trees, 211, 213
Heterogeneous data relative to congruence, 214-15
among data sets, 273 relative to data set size, 210
effects on individual methods, 274 testing of, 209
inverse x2 method, 273 versus evolutionary rate, 210-11
within a single sequence, 273-74 Initiation points (IP), 20
Hominines, 189 Insecta, 341-42
Hominoids, 189 Insertion and deletions (indels), 60, 75, 94
Homiotherm genomes, heavy Intelligenetics, 290
compartments, 129 Interdisciplinary Center for Biotechnology
Homology Research, University of Florida,
by congruence, 8, 74 45-49
definition, 7-8, 73-74 International Union of Pure and Applied
in morphological context, 8, 74 Chemistry (IUPAC), 83
nonserial, 74 Intraspecific polymorphism
nucleotide base position, 7-8, 75-76 and on-screen analysis, 55-56
percent homologous, 75 problem of, 7, 123
phylogenetic, 74 Invertebrates, 334
recognition criteria, 8, 74-75 "Islands" of near-most-parsimonious trees,
repetitive, 74 312-14
serial, 74 ITP, 51
in sequence context, 7-8, 61, 74-76
by similarity, 8, 74
by synapomorphy, 8, 73 Jackknifing
Homoplasy, 8 definition, 161
effect on trees, 80 hobbits (halfings), 161
352 INDEX

Jackknifing (continued) Maximum

to test sampling error and convergence, likelihood (ML) method, 9, 185
160-61 algorithm, 94
versus bootstrapping, 162 computational difficulties, 217, 275
Journal of Molecular Evolution, 59 first uses, 265
likelihood function, 266-69
search for maximum likelihood tree,
Kangaroo, 179 270-71
kb DNA size standard, 23 simulation analysis, 105-10, 115, 121
KiloBase Sequencing System, 51 and statistical inconsistency, 173
statistical tests of, 122, 265-72
strengths/weaknesses, 130, 217
Lagomorphs, 177-79 transition probabilities, 265-67
Lemur, ring-tailed, 187 parsimony (MP) method. See Parsimony
Lemur catta, 187 method
Length mutations, segment alignment
of divergent domains, 221 computation, 65
in parsimony analysis, 235, 237 computer program, 70
relative to evolutionary parsimony and of E colt and M. capricolum, 64-65
compositional statistics, 245 next-best alignments, 65
"Let's Make a Deal", and Bayesian statistical distribution, 67-69
decision theory, 335-36 Measure of skewness (g, statistic), 283
Levels sum, 307 Median consensus. See Consensus methods;
Likelihood Consensus trees
functions, Metazoans, 290
with constant rate, 266-67 Methodological parsimony
by nucleotide configurations, 268-69 and congruence, 11
without rate-constancy, 268 definition, 10
Likelihood (continued) and hypothesis testing, 10
inference, 336-37 versus statistical inference, 10
methods, 341 MHC class II A(J gene, 50, 53, 55-56
Long branches attracting, problem of, Minimum evolution (ME) method
177-79 algorithm, 92
relative to maximum parsimony method,
92
relative to number of nucleotides,
Macaco 119-20
fascicularis, 187 simulation analysis, 105-9, 115, 121-22
fuscata, 187 Mitochondrial DNA (mtDNA)
mulatto, 187 alignment, 75
sylvanus, 187 base compositional biases, 129, 133
Macaque, 187, 189 for cervids, 12-13
Barbary, 187 cytochrome b, 289-90
crab-eating, 187 higher primate phylogeny, 117, 120
Japanese, 187 for mammals, 82-85
rhesus, 187 molecular evolution, 186
MacClade computer program, 284, 290, 292 NADH-dehydrogenase genes, 187
Majority-rule consensus trees. See also and phylogenetic reliability, 12-13,
Consensus methods 84-85, 215
of bootstrap trees, 201, 203-8 for phylogenetic weighting of alignments,
description, 298 82-85
of global parsimony trees, 190-95, 197, for primate phylogeny, 195
201-2 for primates, 117-120, 187
of transversion parsimony trees, 195, 200 for ribosomal RNA, 80-85
Mallinckrodt, 30 for studying tree distributions, 289-90
Mammal phylogeny transfer RNA genes, 187
incongruence, 334 M9 media, 26
test of convergence, 175, 177 Molecular Biology and Evolution, 59
Mammals, 289-90 Molecular
Markov evolution
character transformations, 341 and molecular systematics, 4-7,
model of generating random trees, 302 229-33
INDEX 353

Molecular (continued) simulation analysis, 99, 101-3,

relative to alignments, 60-61, 75 107-8, 115, 121
systematics statistical test of, 263-65
emergence of, 4 Neighbors, 261
future prospects, 13 Nelson consensus. See Consensus methods;
relative to molecular evolution, 4, Consensus trees
229-33 Notophthalmus viridescens, 223-29
Monocot, 334 Nucleic acid sequences,
Monophyletic groups alignments, 60
definition, 4, 296 estimating nucleotide substitutions, 257
identification, 73 increasing importance, 13
Monty Hall problem signal versus noise, 278-92
and Bayesian decision theory, 335-36 Nucleotide
Most-parsimonious trees configuration,
and homology, 76 definition, 268
relation to tree distribution, 278 probabilities of, 268
Mouse DNA, 50, 53, 55-56 substitution. See also Evolutionary
Movements and Habits of Climbing Plants, distance
The, 156 nonsynonymous, 122-23
Multiple one-parameter model, 95
data sets synonymous, 122-23
and classes of data, 328-29 transition matrix, 95, 341
and congruence, 295-96 transition/transversion bias, 96, 122
and independence of characters, two-parameter model, 96
327-29 Nucleotides, number of
and pooling, 311, 327-28 based on simulation analysis, 119-20
sequence alignment relative to higher primates, 119-20
computer relative to tree-making methods, 119-20
program, 70
time, 69
consensus alignment, 70 Oligomer primers, 5-6, 21, 48, 50
related by tree, 69-70 Orangutan, 117, 119, 187
Mus Orthology, 7, 74
domesticus, 223-29, 239, 241 Orthoptera, 341-42
musculus, 82-84 Overrepresentation of background patterns,
Mycoplasma capricolum, 55 ribosomal 135-36, 139-40
RNA alignment, 64
Myoglobin, 175
Myth of a Universal Tree from One Gene Paleoptera, 341-42
(MUTOG), 179-80 Pan troglodytes, 81, 117, 119, 187
Paralogy, 7, 74
partial, 74
Nearest neighbor interchange metric, 314 Parsimony
Necturus beyeri, 223-29 and congruence, 11
Negative binomial and hypothesis testing, 4
for all mutations, 235, 238 methodological parsimony, 10-11
for base substitutions, 235-36 strengths/weaknesses, 129-31
definition, 230-31 Parsimony method, 9, 79, 83-85, 93,
for length mutations, 235, 237 129-31. See also Phylogenetic
relative to restriction map data, 244 Analysis Using Parsimony
Neighbor-joining (NJ) method, advantages, 185-86, 217
algorithm, 93, 121 application to
number of nucleotides, 119-20 Epicrates, 318-26
and primate phylogeny, 188-89 Eumetabola phylogeny, 341-44
relative to minimum evolution, 93 mammal phylogeny, 175-79
simulation analysis, 99, 101-15, 121-22, mitochondrial DNA sequences, 82-85,
141 188, 289-90
statistical test of, 118 primate phylogeny, 188-89
Neighborliness (ST) method, 261, 263-65 ribosomal RNA sequences,
algorithm, 93 82-85, 226, 290-91
for more than four taxa, 264-65 salamander phylogeny, 229-33
354 INDEX

Parsimony method (continued) by parsimony justification, 9

distinguished from methodological for parsimony methods, 249-55
parsimony, 10 by simulation analysis, 9, 90, 215-16
as estimate of evolution, 159 by statistical justification, 9
and inconsistency, 169, 171, 173-74, 273 systematics, 4
informative positions, 250 weighting of alignments, 73, 77-79
limitations, 229, 243 Phylogeny
loss of information, 169, 174-75 of deuterostomes, 291
number of nucleotides, 119-20 of Epicrates, 320-24
optimal conditions, 245 of Eumetabola, 342
and phylogenetic weighting, 82-85 of higher primates, 118
possible correction for, 179 of mammals, 84, 177-78
power of, 211, 213, 215 of primates, 189, 192-93, 198-99, 202
problem of long edges attracting, 178-79 probability of, 336
relative to molecular evolution, 186, 213, of resampled mitochondrial DNA data,
229, 243-45 192-93, 198-99, 202, 204-7
simulation analysis, 103-5, 107-14, 121, of salamanders, 230-32
148-49, 151-53, 284, 287 Physarum, 273
statistical tests, 116-18, 249-55 />„, 307, 325
testing reliability, 186-87 Polymerase chain reaction (PCR), 45, 53,
and tree distributions, 284-87 57
versus invariant methods, 132 asymmetrical PCR, 5-6
weighted parsimony, 147-53 PCR mapping, 21, 23
weighted versus uniform parsimony, potential applications, 5-6
147-53 problems with, 6
Partition metric. See Symmetric-difference universal primer-pairs, 6
metric uses in molecular systematics, 5
PAUP. See Phylogenetic Analysis Using Pongo pygmaeus, 81, 117, 119, 187
Parsimony Power of Movement in Plants, The, 156
Pc (probability of obtaining the correct Prediction
topology), 97 of more similar trees with longer
Persistence, 148-49 sequences
Photomultipliers, 49-50 outcome, 161-62
PHYLIP computer package, 305 relevance to other questions, 162
Phylogenetic Analysis Using Parsimony the test, 160-61
(PAUP) computer package, 83, 188, of similar trees
233, 284, 290, 292, 305, 318 outcome, 159
Phylogenetic as test of theory of descent, 156-57
hypotheses, of similarity among shorter trees
and parsimony, 4 outcome, 160
prerequisites for historical analysis, rationale, 159
3-4 Primates, 84
inference mitochondrial DNA, 84, 118, 187
informativeness, 186-87 number of nucleotides, 119-20
nature of evidence, 213-14 phylogeny of, 117-18, 177-79, 189,
relative to evolutionary models, 8-9, 192-93, 198-99, 202, 204-7
129-33 Prior
relative to molecular evolution, 229, evidence, use in phylogeny, 335-36
243 probability, determination, 339-40
relative to quality of data, 13 Probability,
reliability, 186-87 definitions of types, 336
role of congruence, 215 for particular cladogram, 336
role of stability, 214-15 Problem of short internodes, 80, 82
and weighting, 147, 213-14 Progressive alignment program, 78
reliability. See also Informativeness Prosimians, 189
by branch length variances, 258-59 Pulley-principle, 267
and heterogeneous data, 274
for maximum likelihood method,
265-72 Rabbit, 177-79
for neighborliness method, 261, Random binary trees. See also Trees from
263-65 topologies
INDEX 355

Random binary trees (continued) Rice, 273

fit with r > 2-state characters, 165 Risk of decision, 338-39
fit with two-state characters, 163 Rival trees, 296-97
nonindependence of characters, 165-66 Rodent, 177-79
versus star phylogeny, 163 Rodentia, 84
Rate
heterogeneity,
and alignments, 76-77 Saimiri sciureus, 187
among nucleotide sites, 97, Salamanders, 230-32, 290
230-31, 274 Sanger dideoxy sequencing, 35-36, 47
among positions. Savant, 28-29, 33, 35
gamma distribution, 113, 115, 122 Scaphirhynchus platorynchus, 223-29
heterogeneous data, 273-74 Sea urchin, 291
negative binomial, 230-31, 235-39 Sequenase, 36, 51
for salamanders and their outgroups, Sequence alignment
235-39 alignment scores, 61, 75
simulation analysis, 111, 113-15 alternative alignments, 78
statistical testing, 274 computer-assisted, 7, 61, 63, 75
simulation analysis, 103, 107-15 computer programs, 60-61, 70, 75,
among taxa, 80, 96, 130, 229-30, 241 78-79, 83
index of dispersion, 230, 234-35 database searches, 60, 78
for salamanders and their outgroups, dynamic programming, 62-70
234-35, 243-44 of 55 ribosomal RNA, 63-65
simulation analysis, 107-11 of full sequences, 62-65
invariance, 131 hashing, 60
invariant methods and homology, 7-8, 74-76
compositional statistics, 133 as initial step, 7
evolutionary parsimony, 94, 131 insertions and deletions (indels), 60, 75
versus parsimony method, 132, 244-45 maximum likelihood heuristic, 61
simulation analysis, 110-12 of maximum segments, 65
Rattus norvegicus, 81 of multiple sequences, 69-70, 78-79, 82
Reliability, 186-87 number of alignments, 61
Repelcote, 32, 37, 40 phylogenetic weighting, 73, 77-79
Rescaled consistency index, 315-16 with phylogeny, 7-8, 69-70, 79
Retention index, 315-16 probabilistic versus statistical, 75-76
Reverse transcriptase, 18 problems, 79-82
Rhyacotriton olympicus, 223-29 rate variation, 76-77
Ribonuclease (RNase) A, 27 relative to molecular evolution,
Ribosomal RNA (rRNA) sequences, 60-61, 76-77
for aligning, 63-65, 82-83 sensitivity to input order, 82-84
for Beneckia harveyi, 63 statistical significance, 67-69
for comparing manual and automated two-sequence problem, 60-61
sequencing, 50-53 variation in taxa, 76-77
divergent domains, 221-22 weighting, 83
185, 50-53 Sequencing
for Eschenchia coli, 63-64 contig, 21
55, 63-65 gel
for Ginkgo biloba, 50-53 acrylamide solutions, 40
for illustrating phylogenetic weighting autoradiography, 41
82-85 bellbottom gel system, 37
large-subunit of nuclear ribosomal RNA, cleaning plates, 40
221 electrophoresis conditions, 39-40
for mammals, 82—85 loading dye, 31, 35-36
mitochondrial DNA, 12, 80-85 pouring gels, 37-39
for Mycoplasma capricolum, 64 resuspending samples, 33
for recognizing short internodes, 80-82 sequencing artifacts, 36
for salamanders, 223-29 thermostated plates, 38-40
and stopping algorithm, 290 treating plates, 40
for studying molecular evolution, 222 wedge gel system, 37
in studying tree distributions, 290 large DNA inserts
for vertebrates, 80-82 advantages, 23-24
356 INDEX

Sequencing (continued) "Soft" polytomies, 311

costs, 24 Species trees, 7, 123
disadvantages, 24 Splinkers, 20
protocol, 20, 22 Squirrel monkey, 187
strategy, 19-24 Standard model of evolution, 168
Sheep, 289 Starfish, 291
Short internodes, 79-80 Statistical
Sigma, 29-30, 32 inconsistency
Signal versus noise, 278-92. See also additional work needed, 174
Informativeness of cluster analysis, 173
Simplex space, 337-38 of compatibility methods, 173
Simulation analysis of evolutionary parsimony, 273
criticisms of, 11, 215-16 with four taxa, 173
future studies, 123 heterogeneous data, 274
model trees, 97, 137, 148 of parsimony methods, 118,
rationale, 10, 90 173-74, 273
relative to molecular evolution, 11, 94, and problem of long edges attracting,
137, 148 179
results of relative to four-point criterion, 173
aligning sequences, 67-69 relative to Hadamard transformation,
for compositional statistics versus other 173
methods, 111, 136-42 inference
for constant rate of substitution, criticisms of, 10
99-107, 137, 148 justification for distance methods,
for distance methods, 107-9, 140-42 9-10
for the evolutionary parsimony method versus methodological parsimony, 10
versus other methods, 110-11, test of neighborliness method,
136-40 for four taxa, 263-64
for the gene-frequency maximum and heterogeneous data, 274
likelihood (GFML) method, 106-7 for more than four taxa, 264-65
for the maximum likelihood method using variance, 263-65
versus other methods, 105-6, versus test of branch length variance,
109-10, 121 274-75
for the maximum parsimony method tests of internal branch lengths
versus other methods, 103-5, estimating variance, 255-58
108-9, 121 for five taxa, 257-59
for the neighborliness and neighbor- for four taxa, 256-57
joining methods versus other and heterogeneous data, 274
methods, 99, 101, 103, 121 by their variance, 117-18, 258-59
for number of nucleotide substitutions, versus neighborliness test, 274-75
99-100 tests of maximum likelihood method
for one-parameter, two-parameter, and against the star phylogeny, 269-70
Kimura models, 111-12 Akaike's information criterion, 269
forp and d distances, 101-3, 108-9 between alternative trees, 270-72
for studying tree distributions, 284-88 by the difference in maximum
to test phylogenetic reliability, likelihood scores, 272
9-11, 90, 136-38, 149, 289-90 with heterogeneous data, 273
underlying evolutionary assumptions, by their internal branches, 269-71
11, 94, 137, 148 by maximum likelihood ratio, 269-70
for unweighted pair group method for more than four taxa, 270-72
versus other methods, 107-8 problems with, 269-70, 272
for varying rate of substitution among tests for parsimony
lineages, 107-11, 112 by bootstrapping, 118
for varying rate of substitution among Cavender's worst-case situation,
nucleotide sites, 111, 113-15 253-54
for weighted versus uniform confidence limits, 249-50
parsimony, 147-53 for four species, 249-54
simulated sequences, 67, 94-97, 137, for more than four species, 254-55
148, 284 relative merits of the different tests,
time constraints, 98, 141 274
Siren intermedia, 223-29 relative to next best tree, 241-53
INDEX 357

Statistical (continued) relative to number of nucleotides,

relative to trichotomy, 250-53 119-20
standard normal test, 255 simulation analysis, 99, 101-3, 108
winning sites method, 116 Transition/transversion bias
with a clock, 116, 250-53 and character weighting, 76, 83-84, 195,
without a clock, 253-54 340
tests of phytogeny and compositional statistics, 133-35
analytical methods, 249 for divergent domains, 239-40
assumptions, 115-18, 255 and evolutionary parsimony, 94
by bootstrapping, 118, 122, 160, 249 and Kimura distance, 96, 122, 257
by branch length variances, 117, for primates, 212
255-59 relative to sequence saturation,
current state, 249, 274 231-32, 244-45
for distance tree, 117-18, 255-65 for salamanders, 239-40
for evolutionary parsimony, 116 and two-parameter model, 96, 122, 130,
given a clock, 116, 250-53 257
and heterogeneous data, 273-74 Transversion parsimony analysis
for invariant methods, 136 with bootstrapping, 203, 206-8
by jackknifing, 118-22 of mitochondrial DNA subsets,
for maximum likelihood method, 195-200, 203
269-72 with near-most-parsimonious (1%) trees,
for the neighborliness method, 261-65 197, 200
for parsimony methods, 116, 249-55 with near-most-parsimonious (five-step)
problems, 115-18, 273-74 trees, 195-99
relative merits of different tests, and phylogenetic weighting, 84-85
274-75 of protein-coding genes, 196-97, 199,
resampling methods, 249 207-8
statistical inconsistency, 273 of ribosomal RNA genes, 84-85
for unweighted pair group method, 117 summary of results, 197, 208-10
Williams/Goodman test, 117, 120 of transfer RNA genes, 196-97, 199,
"Stopping algorithm", 290-91 207-8
objections, 291 Tree
Strict consensus. See Consensus methods; -building methods
Consensus trees average distance (linkage) method, 91
Subjective probability, 336 compatibility method, 229
Substitutional change symmetry, 150 compositional statistics, 131-33
Symmetric-difference distance (partition distance-Wagner method, 92-93
metric) evolutionary parsimony, 94, 131
definition, 158, 312-14 Fitch and Margoliash's method, 92
and distortion index, 97-98 maximum likelihood method, 94
distribution of metric, 159 minimum evolution method, 92
and median consensus, 303 neighbor-joining method, 93, 121
relationship to topology, 167-68 neighborliness method, 93
Symplesiomorphies, 4 parsimony method, 93, 129-30
Synapomorphies, 4 requirements for, 180
definition, 4 testing their limits, 168
and homology, 73, 76 transformed distance method, 91
Systematics, 3 comparison metrics, 312-14
"crossover" metric, 314
"nearest neighbor" interchange metric,
314
Taq polymerase, 5, 18 symmetric-difference distance
Tarsier, Philippine, 187 (partition metric), 303, 312, 314
Tarsius syrichta, 187 length distribution
Terminal taxa, 296 approximating the distribution, 292,
Total 319, 326
evidence, 327 based on random sequences, 285-89
information, 307 based on real sequences, 289-92
T7 polymerase, 18 for Epicrates, 319, 326-27
Transformed distance (TD) method number of trees, 166-67
algorithm, 91 phylogenetic context, 278-92
358 INDEX

Tree (continued) relative to amount of sequence data,

recommended analysis, 292 260-62
skewness, 279-92 of sequence distance, 256-57
causes of, 281-83 for testing phylogenies, 117-18, 255,
symmetric versus asymmetric, 279-92 258-59, 263-65
topology Vertebrates, 223, 230-32, 289-90
effects on consensus index, 307-9
effects on tree length distribution,
280-81, 283, 285
Trees Wagner parsimony method, 92, 94
number of, 166-67 Weighted consensus fork index, 307-8, 325
from topologies. See also Finding optimal Weighting
trees according to evolutionary assumptions, 9
"all trees equiprobable" model, 166 of alignments, 73-85
to calculate fit of data, 165-66 character transformation, 340
from Markov model, 166-67 combinatorial, 150, 152
and tree lengths, 283-84 computer simulation, 147-53
Triangle inequality, 93 existential, 150, 152
Tunicate, 291 by higher-level structure, 6-7
Typhlonectes compressicauda, 223-29 inverse linear, 150, 152-53
inverse quadratical, 150, 152-53
multiple data sets, 316
Ungulates, 179 nucleotide-data, 147-53
"Unified approach" program, 79 rationale, 147, 149
Unweighted pair group method. See reliability, 214-15
Average distance (linkage) method successive approximations, 85
US Biochemicals, 35-36 transformational, 152-53
uniform versus weighted parsimony,
Variances 149-53
of branch lengths, 117-18, 255-56, 258
in maximum likelihood method, 272
in neighborliness test, 263-65
number of nucleotide sites versus second Xenology, 74
outgroup, 261-62 Xenopus laevis, 81, 223-29