Article

Microenvironment drives cell state, plasticity, and

drug response in pancreatic cancer
Graphical abstract Authors
Srivatsan Raghavan, Peter S. Winter,
Andrew W. Navia, ..., William C. Hahn,
Andrew J. Aguirre, Alex K. Shalek

Correspondence
[email protected] (P.S.W.),
[email protected]
(A.J.A.),
[email protected] (A.K.S.)

In brief
Systematic profiling of metastatic
pancreatic cancer biopsies and matched
organoid models provides a view of
cellular states, their regulation by the
tumor microenvironment, and the ability
to modulate these states to impact drug
responses.

Highlights
d scRNA-seq of metastatic pancreatic cancer and matched
organoid models

d Ex vivo to in vivo comparisons reveal loss of cell state

heterogeneity in models

d Cell state is shaped by the microenvironment in vivo and can

be controlled ex vivo

d Cell state drives drug response

Raghavan et al., 2021, Cell 184, 6119–6137

December 9, 2021 ª 2021 The Author(s). Published by Elsevier Inc.
https://doi.org/10.1016/j.cell.2021.11.017 ll
 ll
OPEN ACCESS

Article
Microenvironment drives cell state,
plasticity, and drug response in pancreatic cancer
Srivatsan Raghavan,1,2,3,4,20 Peter S. Winter,1,2,5,6,7,20,* Andrew W. Navia,1,2,5,6,7,20 Hannah L. Williams,1,3,20
Alan DenAdel,8,9 Kristen E. Lowder,1,2 Jennyfer Galvez-Reyes,2,5,6 Radha L. Kalekar,1,2 Nolawit Mulugeta,2,5,6
Kevin S. Kapner,1,2 Manisha S. Raghavan,1,2 Ashir A. Borah,2 Nuo Liu,2,5,6 Sara A. Väyrynen,1,3 Andressa Dias Costa,1,3
Raymond W.S. Ng,1,2 Junning Wang,1 Emma K. Hill,1 Dorisanne Y. Ragon,1 Lauren K. Brais,1 Alex M. Jaeger,6
Liam F. Spurr,1,2 Yvonne Y. Li,1,2 Andrew D. Cherniack,1,2,3 Matthew A. Booker,10 Elizabeth F. Cohen,10
Michael Y. Tolstorukov,10 Isaac Wakiro,1 Asaf Rotem,1,2,11 Bruce E. Johnson,1,3,4,11 James M. McFarland,2

(Author list continued on next page)

1Department of Medical Oncology, Dana-Farber Cancer Institute, Boston, MA 02215, USA
2Broad Institute of MIT and Harvard, Cambridge, MA 02142, USA
3Harvard Medical School, Boston, MA 02115, USA
4Department of Medicine, Brigham and Women’s Hospital, Boston, MA 02115, USA
5Institute for Medical Engineering & Science, Massachusetts Institute of Technology, Cambridge, MA 02139, USA
6Koch Institute for Integrative Cancer Research, Massachusetts Institute of Technology, Cambridge, MA 02139, USA
7Department of Chemistry, Massachusetts Institute of Technology, Cambridge, MA 02139, USA
8Center for Computational Molecular Biology, Brown University, Providence, RI 02912, USA
9Division of Applied Mathematics, Brown University, Providence, RI 02912, USA
10Department of Informatics and Analytics, Dana-Farber Cancer Institute, Boston, MA 02215, USA

(Affiliations continued on next page)

SUMMARY

Prognostically relevant RNA expression states exist in pancreatic ductal adenocarcinoma (PDAC), but our
understanding of their drivers, stability, and relationship to therapeutic response is limited. To examine these
attributes systematically, we profiled metastatic biopsies and matched organoid models at single-cell reso-
lution. In vivo, we identify a new intermediate PDAC transcriptional cell state and uncover distinct site- and
state-specific tumor microenvironments (TMEs). Benchmarking models against this reference map, we
reveal strong culture-specific biases in cancer cell transcriptional state representation driven by altered
TME signals. We restore expression state heterogeneity by adding back in vivo-relevant factors and show
plasticity in culture models. Further, we prove that non-genetic modulation of cell state can strongly influence
drug responses, uncovering state-specific vulnerabilities. This work provides a broadly applicable frame-
work for aligning cell states across in vivo and ex vivo settings, identifying drivers of transcriptional plasticity
and manipulating cell state to target associated vulnerabilities.

INTRODUCTION Byrne et al., 2017; Drost and Clevers, 2018; Gillet et al., 2013).
Assessment of mutational fidelity is feasible and useful for two
Recent advances in high-throughput genomic sequencing have reasons: first, bulk measurements of DNA alterations have favor-
led to a detailed understanding of the genetic alterations that able signal-to-noise profiles (i.e., somatic mutations are present
underlie human tumors (Garraway and Lander, 2013). These or absent); and, second, some cancers respond to therapies in a
reference maps have driven a ‘‘mutation-centric’’ view of can- genotype-directed manner (Garraway and Lander, 2013; Hyman
cer that informs our current approach to precision medicine. et al., 2017). However, a growing body of evidence indicates that
In this framework, DNA alterations are used as biomarkers to using mutations alone to assign therapies has limitations (Nam
guide therapy selection (Hyman et al., 2017), and ex vivo et al., 2021; van de Haar et al., 2021). The advent of single-cell
models are used to validate mutational associations and power genomic technologies has confirmed extensive mutational het-
therapeutic discovery efforts. To maintain translational rele- erogeneity in human tumors but also revealed that the
vance, the fidelity of ex vivo models to in vivo attributes is complexity of cancer extends to variation in cell transcriptional
paramount. state. The relationship between cell state and therapeutic sensi-
Driven by the mutation-centric view of cancer, model fidelity is tivity represents a new but poorly understood opportunity for
typically assessed via genomic similarity (Ben-David et al., 2019; cancer therapeutic development (Hahn et al., 2021).

Cell 184, 6119–6137, December 9, 2021 ª 2021 The Author(s). Published by Elsevier Inc. 6119
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
 ll
OPEN ACCESS Article

Ewa T. Sicinska,3,12 Tyler E. Jacks,6 Ryan J. Sullivan,3,13 Geoffrey I. Shapiro,1,3,4 Thomas E. Clancy,1,3,14
Kimberly Perez,1,3,4 Douglas A. Rubinson,1,3,4 Kimmie Ng,1,3,4 James M. Cleary,1,3,4 Lorin Crawford,8,15,16
Scott R. Manalis,6,17 Jonathan A. Nowak,3,11,18 Brian M. Wolpin,1,3,4,21,22 William C. Hahn,1,2,3,4,21,22
Andrew J. Aguirre,1,2,3,4,21,22,* and Alex K. Shalek2,3,5,6,7,19,21,22,23,*
11Center for Cancer Genomics, Dana-Farber Cancer Institute, Boston, MA 02215, USA
12Department of Oncologic Pathology, Dana-Farber Cancer Institute, Boston, MA 02215, USA
13Massachusetts General Hospital Cancer Center, Boston, MA 02114, USA
14Department of Surgery, Brigham and Women’s Hospital, Boston, MA 02115, USA
15Department of Biostatistics, Brown University, Providence, RI 02912, USA
16Microsoft Research New England, Cambridge, MA 02142, USA
17Department of Biological Engineering, Massachusetts Institute of Technology, Cambridge, MA 02142, USA
18Department of Pathology, Brigham and Women’s Hospital, Boston, MA 02115, USA
19Ragon Institute of MGH, MIT, and Harvard, Cambridge, MA 02139, USA
20These authors contributed equally
21These authors contributed equally
22Senior author
23Lead contact

*Correspondence: [email protected] (P.S.W.), [email protected] (A.J.A.), [email protected] (A.K.S.)

https://doi.org/10.1016/j.cell.2021.11.017

Cell state, as measured by RNA expression, is a complex rep- can be used as a tractable biomarker for therapy selection re-
resentation of tumor phenotype because it integrates inputs from mains of critical importance.
cell-intrinsic (e.g., mutational background, epigenetic state) and To parse the instructive roles of cell-intrinsic and cell-extrinsic
cell-extrinsic (e.g., cell-to-cell interactions, tissue architecture) inputs to cancer cell state, we developed and employed an opti-
sources. Although the field has generated high-resolution sin- mized translational workflow to perform both high-resolution
gle-cell RNA sequencing (scRNA-seq) maps of cancer cell tran- profiling of metastatic PDAC patient tissue using scRNA-seq
scriptional states across diverse contexts (Filbin et al., 2018; (Gierahn et al., 2017; Hughes et al., 2020) and derivation of
Hovestadt et al., 2019; Kim et al., 2018; Neftel et al., 2019; Patel matched organoid models (Boj et al., 2015; Tiriac et al., 2018)
et al., 2014; Puram et al., 2017; Sade-Feldman et al., 2019; Suvà from the same core needle biopsy. Using this approach, we
and Tirosh, 2019; Tirosh et al., 2016a; Tirosh et al., 2016b; van generated a single-cell map of metastatic PDAC and used it as
Galen et al., 2019; Venteicher et al., 2017), we have not mapped a reference to benchmark cell states in matched organoid
their stability or the relative influences of cell-intrinsic and cell- models. We identify a new intermediate co-expressor (IC)
extrinsic factors in specifying them. Moreover, we have a limited PDAC cell state, uncover distinct site- and subtype-specific
understanding of the degree to which models accurately recapit- TMEs, and demonstrate that TME signals are critical regulators
ulate the distribution of cancer cell states seen in patients. of cancer cell state, plasticity, and response to therapy.
Despite extensive genomic characterization of pancreatic
ductal adenocarcinoma (PDAC), most cancers do not harbor RESULTS
therapeutically tractable alterations (Aguirre et al., 2018; Bailey
et al., 2016; Cancer Genome Atlas Research Network, 2017). Model systems retain genetic fidelity but lose
However, RNA subtypes (states) derived from bulk measure- expression state heterogeneity
ments have emerged as an important clinical biomarker (Aguirre Current precision medicine pipelines focus on preserving muta-
et al., 2018; Aung et al., 2018; Bailey et al., 2016; Cancer tional fidelity in cancer models; however, it is unclear how well
Genome Atlas Research Network, 2017; Chan-Seng-Yue et al., these same models represent prognostic RNA states (Figure 1A).
2020; Collisson et al., 2019; Collisson et al., 2011; Connor We compared bulk DNA and RNA-sequencing profiles of primary
et al., 2019; Hayashi et al., 2020; Moffitt et al., 2015; O’Kane and metastatic patient tumors with established cell lines and a
et al., 2020; Porter et al., 2019; Tiriac et al., 2018). While PDAC cohort of newly generated patient-derived organoids to under-
subtyping studies have largely recovered common expression stand how each model system represents the distribution of
features, it remains unclear whether these bulk RNA-seq state mutational and RNA phenotypes seen in patients. We observed
measurements mask heterogeneity at the single-cell level, which no significant difference in driver oncogene alteration fre-
features derive from malignant versus non-malignant cells, and quencies among the groups, suggesting that model systems
how well they are recapitulated in laboratory models. Moreover, are relatively faithful genetic representations of patient cohorts
the PDAC tumor microenvironment (TME) includes numerous (Fisher’s exact test; Figure 1B, far right). Next, we assessed
non-malignant immune and stromal cell types (Balachandran PDAC subtypes derived from several recent bulk RNA-
et al., 2019; Bernard et al., 2019; Elyada et al., 2019; Grünwald sequencing studies (Figure 1C) (Bailey et al., 2016; Chan-Seng-
et al., 2021; Ligorio et al., 2019), but their variation across Yue et al., 2020; Collisson et al., 2011; Moffitt et al., 2015).
different sites of disease and their effects on malignant cell state Subsets of signatures from each study were highly concordant
and therapeutic response is not well characterized. Given the and separated primarily into classical-like (clade 1) and basal-
lack of mutational biomarkers for PDAC, understanding how like (clade 2) groups. Clade 3 signatures were generally lower in
cell state is shaped by the local TME and whether cell state expression but tended to associate with clade 1, while clade 4

6120 Cell 184, 6119–6137, December 9, 2021

 ll
Article OPEN ACCESS

A B

F G

Figure 1. Assessing transcriptional states in patient tumors and cancer models

(A) Precision medicine pipelines assess model fidelity for genetics but typically do not evaluate RNA states.
(B) Alterations in PDAC driver genes across primary resections (TCGA), metastatic biopsies (Panc-Seq), and organoid and cell line (CCLE) models. Grey indicates
where genomic data were not available. P-values by Fisher’s exact test.
(C) Comparison of PDAC expression signatures from bulk RNA-sequencing in primary and metastatic tumors, cell lines, and organoid models in (B). Rows are
clustered, columns are sorted by average basal-classical score difference. P-values indicate differences between patient tumors, cell lines, and organoids
by ANOVA.
(D) Schematic of contributors to RNA state that may lead to differences between in vivo and ex vivo expression patterns.
(E) Metastatic patient samples were collected via core needle biopsies and dissociated. Biopsy cells were allocated for scRNA-seq, and patient-matched or-
ganoids were developed with downstream serial scRNA-seq sampling.
(F and G) t-distributed stochastic neighbor embedding (t-SNE) for biopsy (F) and matched patient-derived organoid cells (G).
See also Figure S1; Tables S1 and S2.

signatures represented exocrine pancreas expression patterns In contrast to the mutational data, we observed significant dif-
and tended to be expressed only in primary disease, suggesting ferences for all RNA signatures when comparing tumors to model
that they may be due to contributions from the primary PDAC types (Figure 1C). Whereas primary and metastatic samples
TME. Relatedly, the documented low malignant cellularity in include tumors with both classical-like (clade 1) and basal-like
PDAC may obstruct malignant cell-specific signature identifica- (clade 2) signatures, cell lines exhibited predominantly basal-
tion in bulk RNA-sequencing datasets (Collisson et al., 2019). like (clade 2) subtypes while organoids were nearly entirely

Cell 184, 6119–6137, December 9, 2021 6121

 ll
OPEN ACCESS Article

classical-like (clade 1) with partially overlapping expression of expression programs to in vivo PDAC biology (Figure 2A). The re-
some clade 3 signatures (Figures 1C and S1A). These observa- maining literature-derived signatures had low expression in
tions demonstrate that despite preservation of genomic alter- these metastatic cells, suggesting either specificity for primary
ations, neither PDAC cell lines nor organoids represent the full disease or influence from non-malignant expression in bulk
repertoire of expression subtypes seen in patient cohorts. RNA profiles. Interestingly, a large subset of cells showed low
expression for all previously proposed signatures, highlighting
Single-cell profiling of metastatic PDAC and matched a gap in our understanding of PDAC expression states (Figure 2A;
organoid models ‘‘Low for bulk subtypes’’).
These findings highlight a critical need for new approaches to Given this large subset of cells that score weakly for previously
identify the determinants of cancer cell state both in vivo and in identified signatures, we sought to understand the spectrum of
model systems. Cell state is a complex attribute since it integrates in vivo expression states through an unbiased analysis of our sin-
cell-intrinsic as well as TME-dependent features; therefore, mul- gle-cell dataset. Principal component analysis (PCA) on all ma-
tiple mechanisms (e.g., clonal selection and/or plasticity) may lignant cells revealed that genes enriched for signatures of
contribute to the divergence between in vivo and ex vivo expres- epithelial/mesenchymal transition (EMT) (PC1) (Gröger et al.,
sion patterns (Figure 1D). We hypothesized that a dataset at sin- 2012), basal-like and classical PDAC states (PC1 and PC2),
gle-cell resolution allowing for matched comparisons of in vivo and cell cycle (PC3 and PC8) (Tirosh et al., 2016a) drove the ma-
and ex vivo attributes would enable a better understanding of jor axes of variation in the dataset (Figures S2B and S2C). The
cell state drivers, stability, and functional relevance. To this end, PC1-PC2 score difference was correlated with the basal-clas-
we established a pipeline to generate matched scRNA-seq pro- sical signature axis previously identified in bulk studies (r =
files and organoid models using core needle biopsies from pa- 0.8). We identified 1,909 genes significantly correlated with
tients with metastatic PDAC (n = 23) (Figures 1E and S1B; Table either end of the basal or classical-enriched continuum within
S1). Most samples were obtained from metastatic lesions residing our single-cell cohort (Figure S2D; Table S3; STAR Methods). In-
in the liver (19/23), and the majority (21/23) were analyzed by tar- spection of these genes revealed that more basal-like cells are
geted DNA-sequencing, yielding the expected mutational pat- associated with transforming growth factor beta (TGF-b)
terns (Figure S1B). signaling, WNT signaling, EMT, and cell cycle progression.
Our pipeline generated approximately 1,000 high-quality sin- Epithelial and pancreatic lineage programs are enriched in the
gle cells per biopsy and successful early-passage organoid cul- cells with more classical-like attributes (Figures S2E and S2F).
tures from 70% of patient samples (16/23 samples reaching at We term the signatures derived from our single-cell cohort as
least passage 2) (Figures 1E, 1F, and S1B–S1D). Consistent single-cell basal (scBasal) and single-cell classical (scClassical)
with other studies, we observed patient-specific and admixed (Figure 2B).
clusters of single cells suggesting the presence of both malig- Single-cell analysis enabled a dissection of expression states
nant and non-malignant cells in each biopsy (Figures S1E and that are confounded by contaminating non-malignant cells in
S1F; STAR Methods) (Kim et al., 2018; Puram et al., 2017; bulk measurements. For example, in bulk RNA-seq studies
Sade-Feldman et al., 2019; Tirosh et al., 2016a). Inferred copy (Aguirre et al., 2018; Cancer Genome Atlas Research Network,
number variation (CNV) alteration scores separated putative 2017), EMT programs are strongly correlated with markers of fi-
cancerous and non-cancerous cells in each biopsy and demon- broblasts (r = 0.9), inversely related to purity metrics (r =
0.5),
strated high concordance with reference targeted DNA-seq (Fig- and poorly correlated with basal-like expression (r =
0.08). At
ures S1G and S1H) (Patel et al., 2014; Tirosh et al., 2016b). CNV single-cell resolution, we observed a subset of cells that express
analysis and manual inspection of expression patterns for known both scBasal and EMT programs, while other scBasal cells had
markers across single cells supported the identification of malig- low expression of EMT programs (Figure 2C, left). Similarly, the
nant cells as well as 11 unique non-malignant cell types (Figures origin of cytokine response signatures can be difficult to interpret
S1I–S1K; Table S2). Thus, despite the documented low malig- from bulk studies, as interferon (IFN) response gene signatures
nant cellularity in PDAC (Aguirre et al., 2018; Cancer Genome are positively correlated with markers of several cell types
Atlas Research Network, 2017; Chan-Seng-Yue et al., 2020), including macrophages (r = 0.6) and T cells (r = 0.4), negatively
we established a robust pipeline that retrieved high-quality ma- associated with purity metrics (r =
0.4), and not associated
lignant (n = 7,740) and non-malignant (n = 15,302) single-cell with either basal-like or classical scores (r = 0.01 and
0.04,
transcriptomes from metastatic PDAC needle biopsies, as well respectively). In single cells, we observe clear patterns of asso-
as those of matched organoids (n = 24,995) (Figures 1E–1G). ciation between IFN response and a subset of cells with interme-
diate scBasal expression (Figure 2C, right). These findings sug-
Single-cell resolution identifies an IC cancer cell state in gest not all basal-like cells have the same underlying attributes
metastatic PDAC and highlight the importance of utilizing scRNA-seq to under-
After separating the CNV-confirmed PDAC cells from non-malig- stand how malignant cells sense and respond to their local TME.
nant cells (excluding one neuroendocrine sample; Figure S2A; Although some studies suggest that basal and classical pro-
see STAR Methods), we first interrogated whether previously grams exist only in discrete cell populations (Chan-Seng-Yue
described RNA subtypes (Figure 1C) are represented in single et al., 2020), we observed that the scBasal and scClassical pro-
metastatic PDAC cells. Subsets of metastatic cells scored highly grams were not mutually exclusive. Rather, we identified cells
for either classical-like or basal-like signatures derived from in- that are intermediate for scBasal and scClassical gene expression
dependent bulk studies indicating the general relevance of these signatures, co-expressed features of both programs to varying

6122 Cell 184, 6119–6137, December 9, 2021

 ll
Article OPEN ACCESS

A C

F G H

Figure 2. An intermediate co-expressor state bridges basal and classical phenotypes

(A) Signature scores (rows) for bulk derived expression subtypes in malignant cells (columns).
(B) Heatmaps depict the expression of scBasal and scClassical expression programs and highlight a co-expressing cell population.
(C) Variation in EMT and IFN response signature expression within malignant cells that have scBasal expression.
(D) The intermediate co-expressor (‘‘IC’’) expression program is enriched in co-expressing cells. Enrichment adjusted P-values (hypergeometric test) for EMT,
KRAS, and AKT gene sets are indicated at right in (B) and (D).
(E) Gene set enrichment analysis for the 115 genes specific to the intermediate co-expressor program.
(F) Malignant cell state diagram for PDAC. ScBasal-scClassical commitment score (x axis) and IC score (y axis).
(G) Frequency of co-expressing cells is related to increased mixing of scBasal and scClassical cell populations. Log ratio of % scBasal and scClassical cells in
each sample (x axis; dotted line at 0 indicates equal percentages of scBasal and scClassical cells) versus their % IC cells (y axis).
(H) Multiplex immunofluorescence analysis identifies co-expressing cells in matched metastatic samples. White box indicates region for co-expression insets at
bottom. Scale bar represents 10 mm.
See also Figures S2, S3, and S4; Table S3.

degrees, and were poorly described by previously identified bulk gene sets (Figures 2B, 2D, 2E, and S2F). We also assessed
RNA subtypes (Figures 2B [‘‘Co-expressing cells’’], S3A, and whether the IC state overlapped with any phenotypes recently
S3B). We identified 115 genes whose expression was correlated reported in a study of normal pancreas progenitors (Qadir et al.,
with scBasal-scClassical co-expression and termed this gene 2020). We found that both scBasal and scClassical gene
set the IC state (Figures 2D, S3C, and S3D; Table S3; STAR expression signatures are expressed by pro-ductal progenitor
Methods). cells, while the IC gene expression program is enriched in an
The IC state is enriched for developmental, RAS signaling, undifferentiated, stress-responsive progenitor population (Fig-
and inflammation/stress response gene sets (Figure 2E). Signa- ure S3E) (Qadir et al., 2020). Although the IC state may not
tures of RAS signaling are enriched in the IC state even represent a distinct step along a developmental trajectory, it
compared with scBasal and scClassical programs, and, by may represent a similar stress-induced transition state in a
contrast, scClassical states are enriched for AKT-associated cancer context.

Cell 184, 6119–6137, December 9, 2021 6123

 ll
OPEN ACCESS Article

We next assessed how this new IC state relates to previously lower rates of long-term propagation than models derived from
proposed bulk signatures to clarify potential inter-relationships. scClassical tumors (Figure 3A). When comparing early passage
Pairwise correlation of our PDAC single-cell and established CNV-confirmed organoid cancer cells to their cognates from pa-
bulk RNA-seq signatures in malignant cells revealed that the IC tient tissues, culture in an ex vivo microenvironment caused
state is unique and not well described by prior signatures (Fig- greater deviation in cell state than CNV-defined genotype (Fig-
ure S3F). Our findings suggest that malignant PDAC cells orga- ure 3A; STAR Methods).
nize in a tripartite cell state framework that spans committed We next examined how ex vivo transcriptional states in our or-
basal and classical phenotypes, with considerable signature ganoid cohort differed from their matched patient samples (Fig-
co-expression in single cells (Figure 2F). Similar to the variation ure 3B). We observed a striking loss of scBasal gene expression
in EMT scores observed in scBasal-expressing cancer cells (Fig- and to a lesser extent, the IC program. By contrast, aggregate
ure 2C), we noted heterogeneity among co-expressing cells for scClassical gene expression remained largely unchanged in or-
the IC program (Figure 2F). ganoid conditions. This comparative analysis also nominated a
Classification of tumors by their malignant pseudo-bulk signa- set of upregulated organoid-specific genes that were not pre-
ture expression stratified the cohort into those that expressed sent in vivo, including markers of epithelial identity, oxidative
predominantly scBasal, scClassical, or IC signatures (Fig- stress response pathways (e.g., NRF2 target genes), and amino
ure S3G–S3I). Individual patient specimens still exhibited signif- acid metabolism (hereafter collectively referred to as ‘‘organoid-
icant heterogeneity at the cellular level, containing at least two specific’’ gene expression) (Figure 3C, bottom; Table S4). These
and sometimes all three malignant cell states (Figure S3J). Sam- findings suggest that changes to microenvironmental growth
ples with greater malignant cell state diversity (i.e., higher pro- conditions significantly alter cellular transcriptional states and
portions of both scBasal and scClassical cells) also exhibited a induce culture-specific expression programs, highlighting the
higher proportion of cells expressing the IC state, suggesting importance of benchmarking ex vivo models using matched
the IC state may serve as a transition between scBasal and in vivo states as a reference.
scClassical poles (Figure 2G).
Transcriptional state heterogeneity is shaped by the
Multiplex immunofluorescence confirms a tripartite cell ex vivo microenvironment
state framework in metastatic and primary PDAC Regardless of the cell state distribution in the original biopsy, indi-
To validate this extensive heterogeneity and the presence of co- vidual models assumed more scClassical or organoid-specific
expressing cells in our metastatic cohort, we used a subtype- cell states over time in culture (Figure 3D), mirroring the results
specific multiplex immunofluorescence (mIF) panel to categorize in our larger bulk RNA-seq cohort (Figures 1B and 3C [top]). To
single malignant cells by their patterns of marker detection in 10 better understand the drivers of this cell state bias, we first inves-
matched cases from our single cell study (Figure S4A; Table S3; tigated genetic alterations associated with either the basal or
STAR Methods). We observed overlap of basal and classical classical subtypes. Prior work has suggested that KRAS amplifi-
markers within single cells at the protein level, corroborating cations associate with basal features (Chan-Seng-Yue et al.,
the existence of co-expressing cells (Figures 2H and S4B). More- 2020; Miyabayashi et al., 2020), while amplifications of lineage
over, we observed a significant correlation between malignant transcription factors like GATA6 associate with classical sub-
phenotypes assessed by mIF protein detection for samples types (Chan-Seng-Yue et al., 2020). In agreement with prior
from the same RNA subtype (average r = 0.52) compared to bulk studies, we observed a significant association between sin-
those of different subtypes (average r = 0.06, p < 10
7, Student’s gle-cell inferred KRAS copy number gain and the scBasal state in
t test) (Figure S4C, white dots). Deep sampling of each matched metastatic cells (p < 0.03 Fisher’s exact test) (Figure S5E). Four
biopsy using mIF (mean = 13,078 cells per sample) also KRAS-amplified samples proliferated ex vivo and maintained
confirmed the increased fraction of co-expressing cells in tu- this alteration, but their malignant cell state shifted from scBasal
mors with equal basal and classical cell state mixing (Fig- in vivo to scClassical in organoid culture (Figures 3D [black
ure S4D). Finally, we used mIF to identify co-expressing cells dots] and S5F), including one sample where the same KRAS-
in primary tumor samples, suggesting that the IC state may be amplified clone exhibited a similar scBasal (biopsy) to scClassical
a general feature of PDAC (Figure S4E). (organoid) state shift (Figure 3E). These findings demonstrate that
KRAS amplification alone is not sufficient to lock the scBasal
Benchmarking model expression state using matched expression state, PDAC cells are plastic, and the microenviron-
in vivo reference maps ment can influence cell state independent of genotype in this
With this high-resolution reference map of in vivo malignant cell context.
states, we next asked whether matched ex vivo models retained We next compared genetic heterogeneity (inferred CNV sub-
the cell state distribution of the tissues from which they were clones) and transcriptional states from matched biopsy tissue
derived. An unbiased comparison of all malignant cells (biopsy and organoid samples from iterative passages. ScClassical tu-
and organoid; 32,073 cells) revealed separation of biopsy and or- mors tended to maintain their genotype and transcriptional state
ganoid profiles, while organoid samples from iterative passages both early in culture and at later passages (e.g., PANFR0631)
ex vivo clustered together (Figures S5A and S5B). After removing (Figures 3D and S5G [clone A]). In contrast, most models derived
low-frequency non-malignant cells (Figures S5B–S5D; STAR from scBasal or IC tumors exhibited early cell-state deviation
Methods), we observed that models attempted from biopsies and cessation of growth within 100 days of initiation (e.g.,
with high malignant-cell-averaged scBasal or IC states exhibited PANFR0552) (Figures 3D and S5H). In some cases, we observed

6124 Cell 184, 6119–6137, December 9, 2021

 ll
Article OPEN ACCESS

A B

Figure 3. Organoid culture selects against the scBasal state with transcriptional evolution over time
(A) Outgrowth and similarity between organoids and matched biopsy samples. Red fill indicates successful early (Early) and long-term (Estab.) culture. Right gray
scale indicates similarity between each biopsy-early organoid pair for inferred CNVs (genotype, Geno.) or cell state (State). P-value for Geno. versus State
differences determined by Student’s t test.
(B) Schematic for matched in vivo malignant cell and organoid comparison.
(C) Single-cell and average expression of malignant programs (top) and organoid-specific genes (bottom) in biopsy cells and matched, early passage organoid
cells (n = 13 models). P-values determined by Student’s t test. Parenthetical P-values (left) indicate hypergeometric test for pathway enrichments.
(D) Swimmer’s plot depicting organoid state evolution in culture. Pie charts indicate the fraction of cell states at each time point.
(E) Clonal fractions from KRAS-amplified PANFR0575 biopsy (gray) and organoid (red) cells. Clone A (green) is present in both. Heatmap shows expression of
scBasal and scClassical states in clone A from both contexts.
See also Figure S5.

outgrowth of an scClassical sub-clone in culture. For example, states emerged at later passages (Figures S5I and S5J). By
the in vivo scBasal clones from PANFR0489R rapidly decreased contrast, in vivo scBasal clones from PANFR0575 were main-
in abundance while clones with scClassical or organoid-specific tained initially in culture, but their cell state was highly plastic

Cell 184, 6119–6137, December 9, 2021 6125

 ll
OPEN ACCESS Article

and changed to scClassical when measured as organoids at Although maintaining organoids in minimal or cell line media
passages 2 and 3 (Figures 3E and S5K). After > 100 additional conditions resulted in partial recovery of scBasal and IC expres-
days in ex vivo culture, PANFR0575 regained scBasal expres- sion, we failed to observe fully polarized models (Figures S5P
sion and clones with inferred TP63 amplifications, a squa- and S5Q), suggesting that these conditions lack critical TME
mous-specifying transcription factor (Somerville et al., 2018), components to fully specify cell state. We used differential
dominated the culture (Figure S5K [clones D and E]). Our detailed expression across each biopsy-organoid pair to nominate genes
analysis reveals that both plasticity and selection occur in the that were expressed in vivo but missing ex vivo (Figure 3B; STAR
same model, where certain clones demonstrate plasticity in Methods). Genes differentially expressed by malignant cells
response to microenvironmental signals, and certain genotypes, in vivo were related to soluble cytokine signaling, cell-cell
though rare, may still exert a strong effect despite opposing cues communication, and microenvironmental interactions (Fig-
from the microenvironment. In sum, these findings underscore ure 4G). Hierarchical clustering revealed state-dependent
the need to consider both mutational and transcriptional state expression patterns for in vivo-specific genes (Figure 4H; Table
to ensure faithful representation of in vivo cancer cell S4). For example, IFN response and EMT genes were signifi-
phenotypes. cantly upregulated in scBasal and IC malignant cells in vivo (clus-
ters 1 and 2), while genes associated with cell-cell interactions
Media formulation influences PDAC transcriptional and surface glycoproteins were more strongly expressed in IC
states ex vivo and scClassical cells (cluster 3). Genes related to biological
Given the evidence for selection and plasticity in standard orga- adhesion were more uniform in their expression across the sub-
noid conditions, we tested whether specific aspects of the types (cluster 4) (Figure 4H). The relative absence of these genes
ex vivo culture environment governed cell state determination. in organoid culture and their differences in expression across
We first evaluated the effects of extracellular matrix dimension- transcriptional subtypes in vivo suggests that TME signals might
ality by culturing established 3-dimensional (3D) organoid play a role in fully specifying cancer cell transcriptional state and
models as 2-dimensional (2D) cell lines in the same organoid me- potentially therapeutic responses.
dium. This did not affect transcriptional subtype across the 4
models tested (Figure S5L). Next, we asked whether culturing Composition of the metastatic TME and site-specific
established organoid models in altered media conditions could differences in mesenchymal populations
rescue expression heterogeneity (Figure 4A). We cultured 4 or- The presence of TME-associated expression patterns in cancer
ganoid models in media without any additives for 6 days (‘‘Min- cells in vivo suggested there may be subtype-dependent structure
imal’’ medium) (Table S5; STAR Methods) and observed a robust to, and instructive signaling from, the metastatic TME; however,
increase in scBasal expression, a decrease in organoid-specific relatively little is known about the composition of the metastatic
gene expression, and stable CNV profiles (Figures 4B, S5M, and TME in PDAC. We analyzed the non-malignant cells (n = 14,811)
S5N). With greater time in minimal medium, the distribution of in the metastatic niche to further subclassify cell types and provide
cell states shifted even further toward IC and scBasal (Figure 4C). a more complete picture of the immune/stromal composition of
Since minimal medium lacks serum and mitogens to maintain metastatic disease (Figure 5A). Sub-clustering of T cells and nat-
prolonged cell growth, we tested a reduced organoid media ural killer (NK) cells revealed 4 cell types (CD4+ T, CD8+ T, NK, and
formulation (‘‘OWRNA’’) (Table S5; STAR Methods) and found CD16+ [FCGR3A+] NK cells) (Figures S6A and S6B; STAR
that OWRNA supported proliferation, strengthened scClassical, Methods) Similarly, an unbiased analysis within the monocyte/
and allowed for scBasal expression (Figure S5O). macrophage compartment revealed 3 subsets of tumor associ-
Given the divergence in cell state for cell line versus organoid ated macrophages (TAMs) (FCN1+ ‘‘monocyte-like’’ TAMs,
models of PDAC (Figures 1A and 1B), we took an established cell C1QC+ TAMs, and SPP1+ TAMs) (Figures S6C and S6D; Table
line and an organoid model and cultured them in the reciprocal S2) (Zhang et al., 2020; Zilionis et al., 2019). Marker expression
media condition. Organoid cells grown in standard cancer cell across all previously described non-malignant cells is summa-
line medium (‘‘Cell line media,’’ RPMI-1640 with 10% fetal bovine rized in Figure S6E.
serum) (STAR Methods) gained scBasal expression, while Whether the TME differs between primary PDAC and different
CFPAC1, an established pancreatic cancer cell line, lost scBasal metastatic sites is not well understood. Although we found equal
features when grown in complete organoid media (Figures 4D, distribution of immune cells across different metastatic sites,
4E, S5P, and S5Q). To understand whether these state changes mesenchymal cells clustered by the site of disease (Figures
are functionally significant, we exposed CFPAC1 cells grown in S6F and S6G). Despite uniform expression of a previously
basal media conditions (‘‘cell line media’’) or classical conditions described myofibroblast signature (Elyada et al., 2019; Ӧhlund
(complete organoid media) to SN-38 and paclitaxel. Strikingly, in et al., 2017), an unbiased analysis revealed divergent mesen-
both instances, we observed greater sensitivity among cells chymal states favoring expression of either fibroblast-like or
grown in organoid media, suggesting that environmental factors pericyte-like genes (Figures S6H–S6K; Table S2) (Bartoschek
can shape therapeutic response through state changes (Fig- et al., 2018; Di Carlo and Peduto, 2018; Hosaka et al., 2016; Pe-
ure 4F). These findings demonstrate that even established lon et al., 2020). While biopsies from each location contained
models are sensitive to changes in culture conditions and that both subsets, we detected a strong association between liver bi-
a combination of up-front selective pressures and culture me- opsies and the pericyte-like mesenchymal state (Figures 5B,
dium-imprinting shapes the cell-state biases in extant model S6L, and S6M). We validated this association in larger cohorts
systems, influencing their functional attributes. and observed a similar relationship between pericyte-like

6126 Cell 184, 6119–6137, December 9, 2021

 ll
Article OPEN ACCESS

A C

B
D E

H
F

Figure 4. Modulation of the media microenvironment allows recovery of scBasal states

(A) Strategies to recover scBasal expression in different media conditions.
(B) Tied dot plot for sample average scBasal score (left) and organoid-specific score (right) in the indicated conditions. Color outline indicates sample identity. P-
value compares single cell distributions within models and was calculated by Student’s t test.
(C) Cell state diagrams for organoid cells cultured in minimal media. P-values are for that time point versus the complete media condition and compare B/C
commitment (top) and IC scores (right) by ANOVA followed by Tukey’s HSD.
(D and E) Violin plots depict scBasal and organoid-specific expression scores in PANFR0562 organoid cells (D) or CFPAC1 cell line (E) after 6 days in Complete
organoid medium or in ‘‘Cell line’’ medium. P-values for differences were calculated by Student’s t test.
(F) CFPAC1 cell line growth rate-adjusted dose response curves to SN-38 and paclitaxel after culture in standard ‘‘Cell line’’ medium or in Complete organoid
medium. Points are the mean ± SD of 3 technical replicates. Curves are representative of 2 independent experiments.
(G) Significant pathway enrichments (P-value < 10
12) for top in vivo differentially expressed genes (143 genes).
(H) Average expression in biopsy (left) and organoid cells (right) for the 143 top in vivo differentially expressed genes (rows), organized by originating tumor’s
overall transcriptional subtype (colored dots). Parenthetical P-values (left) for enrichment of indicated pathways are by hypergeometric test. Overall biopsy versus
organoid expression difference is determined by Student’s t test (bottom). P-values computed by one-way ANOVA followed by Tukey’s HSD (center) are for
differences in average expression between biopsy transcriptional subtypes (*P-value < 10
8; **P-value < 10
16).
See also Figure S5; Tables S4 and S5.

Cell 184, 6119–6137, December 9, 2021 6127

 ll
OPEN ACCESS Article

A E F

G
B

C I

D J

Figure 5. Distinct mesenchymal phenotypes and transcriptional state-specific immune heterogeneity exist in the liver metastatic
microenvironment
(A) t-SNE visualization of non-malignant cells identified in the metastatic microenvironment, abbreviations as in Figure S1J (TAM, tumor associated macrophage).
(B) Expression of Fibroblast-like (PC2 low) and Pericyte-like (PC2 high) mesenchymal (Mes.) cell programs across different metastatic sites (top).
(C) Density plots for mesenchymal cell phenotype score in single cells from our metastatic cohort (top) or previously published PDAC bulk RNA-seq profiles
(bottom), fill indicates anatomic site. P-value determined by Student’s t test (top) or by ANOVA followed by Tukey’s HSD (bottom).
(D) Summary of mesenchymal phenotypes in primary versus liver metastatic PDAC.

(legend continued on next page)

6128 Cell 184, 6119–6137, December 9, 2021

 ll
Article OPEN ACCESS

expression and liver metastases, while in contrast, tumors in the cells (‘‘paracrine’’ factors) (Figure 6A). We first identified secreted
pancreas favored fibroblast-like mesenchymal expression (Fig- factors differentially expressed by cancer cells in each transcrip-
ure 5C). Thus, we observed diverse immune and stromal cell tional state (‘‘autocrine’’ signals) and applied these to rescue IC
types in the metastatic TME and identified mesenchymal fea- and scBasal expression in ex vivo models (Figure 6B; Table S6;
tures unique to the liver niche compared with primary disease STAR Methods). TGFB2 was the top differentially expressed fac-
(Figure 5D). tor shared by malignant cells in both scBasal and IC TMEs (Fig-
ure 6B), and organoids cultured with TGF-b ligands exhibited a
Transcriptional subtypes associate with distinct pronounced shift toward IC and scBasal states (Figure 6C).
immune microenvironments The reemergence of scBasal transcriptional heterogeneity in
We next searched for associations between malignant cell state both minimal media (Figure 5C) and TGF-b conditions (Figure 6C)
and the composition of the TME. Five samples were excluded suggested that different types of microenvironmental pressure
from this analysis on the basis of low cell counts (< 200 cells) can lead to the basal-like cell state. These experiments also indi-
or indeterminant transcriptional subtype (Figure S6N). We cate that culture conditions can be tuned to achieve composi-
applied Simpson’s diversity index to define each tumor’s overall tional differences spanning scClassical, heterogenous, and
microenvironmental composition (STAR Methods). Tumors with scBasal expression, akin to those seen in vivo.
higher average malignant scClassical or IC expression harbored We next assessed whether TME signals, like media formula-
greater TME diversity, while strongly scBasal tumors exhibited tion (Figure 4F), could influence drug sensitivity through altering
more homogeneous TMEs (Figure 5E). We observed a similar transcriptional state. In an isogenic organoid system (STAR
pattern in bulk samples where we inferred diversity by using Methods), models induced to adopt scBasal expression through
the ‘‘immunogenic’’ signature (Bailey et al., 2016) to indicate exposure to TGF-b for 3 weeks were less sensitive to several
greater immune cell infiltration (Figure 5F). Clustering over the standard-of-care chemotherapeutic agents including gemcita-
cell type fractions in each biopsy revealed the non-malignant bine, paclitaxel, and SN-38, the active analog of irinotecan (Fig-
cell types driving overall diversity differences (Figures 5G and ures 6D, S7A, and S7B; Table S7). The duration of exposure to
5H). C1QC+ TAMs dominated the TME of strongly scBasal tu- TGF-b corresponded with the degree of state shift, and these
mors, and both CD8+ and CD4+ T cells were significantly states were highly plastic, as withdrawal of TGF-b resulted in a
depleted in scBasal contexts compared to the rest of the sam- return to scClassical expression (Figure 6E). Drug sensitivity
ples in the cohort (Figures 5H and 5I). By contrast, T cells were tracked with cancer cell state, as models were re-sensitized to
most often isolated from, and their abundance positively corre- chemotherapeutic agents upon shifting back to the scClassical
lated with, higher IC malignant fractions in our cohort (Figures state (Figures 6E and 6F). These observations agree with recent
5I and S6O). We observed similar patterns in bulk RNA-seq clinical trial data showing that patients with basal phenotypes
data from The Cancer Genome Atlas (TCGA) (Weinstein et al., tend to have poorer outcomes with combination chemotherapy
2013; Vivian et al., 2017), noting reduced levels of immune- (Aung et al., 2018; Bailey et al., 2016; Collisson et al., 2011; Con-
related gene expression in other epithelial tumors with high nor et al., 2019; Moffitt et al., 2015). Together, these findings
basal/squamous gene expression (Figure S6P [cluster 4]). highlight the remarkable phenotypic and functional plasticity
Together, these findings suggest coordination between cancer inherent in tumor models.
cell states and the local TME, with decreased immune cell diver- To assess for state-specific vulnerabilities more broadly, we
sity in basal contexts (Figure 5J). tested a 24-drug panel in isogenic organoid systems with the
scBasal state induced via two distinct routes, either by culturing
State-specific TME signals drive transcriptional in minimal or TGF-b-containing medium (Figure 6G; Table S7).
heterogeneity and drug response Within the same model, scClassical states were on average
Based on these observations, we hypothesized that soluble fac- more sensitive to both chemotherapy and agents targeting
tors specific to the TME of each transcriptional subtype may DNA-damage repair pathways while scBasal states were more
drive cancer cell states and potentially influence their therapeutic sensitive to mitogen-activated protein kinase (MAPK) pathway
responses (Figure 6A). In vivo, the secreted factor milieu sur- inhibitors (MEK and ERK inhibitors) (Figures 6G, S7A, and
rounding cancer cells originates from at least two sources: ma- S7B). These findings provide direct evidence that transcriptional
lignant cells themselves (‘‘autocrine’’ factors) and non-malignant state can be a major determinant of drug response and that

(E) Correlation between microenvironment diversity (Simpson’s index, x axis) and the average malignant scBasal-scClassical commitment score (y axis) for each
scRNA-seq sample.
(F) Correlation between TME diversity as inferred by immunogenic signature score (x axis) versus average tumor scBasal-scClassical commitment score (y axis)
in primary and metastatic bulk RNA-seq samples.
(G and H) Sample-level (columns) variation in Simpson’s index (G, dot plot), average malignant scBasal-scClassical (G, top heat bar) and IC (G, bottom heat bar)
expression, and fraction of non-malignant cell types (H). Samples were clustered and ordered within metastatic site (liver versus other) by their fractional rep-
resentation of cell types. Dots indicate top statistically significant cell type frequency differences calculated using a Kruskal-Wallis test with multiple hypothesis
correction.
(I) Boxplots compare cell type fractions between the scBasal predominant tumors with low diversity (PANFR0593, 575, 545) and all others. P-value determined by
Student’s t test.
(J) Summary of associations between microenvironmental diversity, non-malignant cell types, and malignant cell state.
See also Figure S6; Table S2.

Cell 184, 6119–6137, December 9, 2021 6129

 ll
OPEN ACCESS Article

A B

C
E

D F

Figure 6. Tumor state-specific factors rescue malignant transcriptional heterogeneity and reveal state-specific drug sensitivities
(A) Schematic describing microenvironmental inputs to tumor cell state in vivo (left, ‘‘Metastatic environment’’) versus ex vivo (center, ‘‘Organoid environment’’)
and a strategy to recover malignant transcriptional heterogeneity ex vivo (right, ‘‘State-supportive environment’’).
(B) Differential expression (Wilcoxon rank sum test) of secreted factors between in vivo tumor cells scored as scBasal versus scClassical (x axis) and IC malignant
cells versus the rest (y axis). State-specific genes that pass significance after multiple hypothesis correction (p < 0.05) are colored by their group association.
(C and D) Cell state diagrams with marginal density plots (C) and growth rate-adjusted dose response curves to gemcitabine and SN-38 (D) for organoid model
PANFR0562 cultured for 3 weeks in control medium (OWRNA) or in control medium with TGF-b. P-values in (C) for group differences between B/C commitment
(top) and IC scores (right) were calculated by ANOVA followed by Tukey’s HSD. In (D), points are the mean ± SD of 3 technical replicates. Curves are repre-
sentative of 2 independent experiments.
(legend continued on next page)

6130 Cell 184, 6119–6137, December 9, 2021

 ll
Article OPEN ACCESS

differential targeting of cell states represents an actionable ther- in vivo tissue as a reference for culture model state fidelity, an
apeutic paradigm. approach that has been applied in normal organoid systems (Fu-
jii et al., 2018) but not in cancer, enabled a structured dissection
Paracrine signals from the local TME direct cancer cell of cell-intrinsic versus TME-induced contributions to malignant
transcriptional phenotypes cell state. As the community continues to catalog these tran-
We next searched for paracrine factors differentially expressed scriptional states in clinical samples at single-cell resolution,
by the non-malignant cells in each subtype. We noted an we anticipate this framework will be broadly applicable for un-
increasing number of differentially expressed factors in the IC derstanding their functional significance across a variety of
and scBasal contexts and mapped each paracrine factor to its cancers.
cognate cell type to summarize the overall secreted factor com- Single-cell resolution enabled us to appreciate the layering of
binations that shape subtype-specific TMEs in metastatic PDAC malignant states whereby single cells can be unified in their sim-
(Figures 7A and 7B; Table S6; STAR Methods). Interestingly, ilarity for an expression subtype (e.g., scBasal or IC) but differ in
IFNG from CD8+ T cells was most highly expressed in the IC their expression of TME-influenced programs (e.g., IFN
TME, consistent with a higher T cell fraction in IC tumors (Figures response). Future studies across more patients and different
5G and S6O) and the relative increase in IFN-responsive gene anatomic disease sites at single-cell resolution will be needed
expression in IC and scBasal malignant cells (Figures 4G and to fully parse which invariant ‘‘core’’ genes mark archetypal
4H). Correspondingly, cells from organoid models exposed to cell states in PDAC and which expression features are superim-
interferon g (IFNg) showed increased IFN response gene expres- posed by the TME. Clinically annotated datasets will also aid in
sion (IFN response score) and a concomitant shift toward the IC the assessment of these expression states for prognostic value
state (Figures 7C, S7C, and S7D). and their utility in nominating cell state-specific therapeutic
We next examined two in vivo scenarios for evidence that the liabilities.
TME can drive cell state variation by analyzing samples from (1) We also uncovered formerly unappreciated relationships be-
distinct metastatic sites within the same patient and (2) the same tween cancer cell transcriptional states and the local TME.
metastatic site before and after immunotherapy. In the first case Similar to the relationship between inflammation and tumorigen-
(PANRF0473), a larger fraction of T cells within the liver metasta- esis (Alonso-Curbelo et al., 2021; Li et al., 2021), our data
tic niche expressed high levels of IFNG (IFNG expression score) support a model wherein as tumors become inflamed and im-
relative to the lung metastatic niche. Correspondingly, malignant mune-activated, malignant cells display enhanced state plas-
cells in the liver showed evidence of IFN response and higher IC ticity. These relationships may have implications for PDAC
state expression (Figure 7D, S7E,F). Similarly, in serial samples immunotherapy strategies given that a productive CD8 T cell
from the same liver lesion in patient PANFR0489, the post-pro- response may promote more aggressive basal-like states, as
gression biopsy harbored a higher fraction of T cells with high suggested by our paired pre- and post-immunotherapy sam-
IFNG expression scores, consistent with changes to the TME ples. Coordination between basal/mesenchymal malignant cell
stemming from immunotherapy. Post-progression malignant states and immune responses may be a broadly relevant phe-
cells again expressed higher IFN response scores, and their nomenon given our observations in other basal-like tumors and
transcriptional phenotypes were shifted toward the IC and recent work in glioblastoma (Hara et al., 2021). A recent publica-
scBasal states relative to the pre-treatment malignant cells (Fig- tion further supports the idea that local TME variation can influ-
ures 7E, 7F, S7G, and S7H). These paired biopsies allow a win- ence clinical outcomes in PDAC (Grünwald et al., 2021), while
dow into the complex interplay between the TME and cancer our work provides direct evidence that TME signals are critical
cells within the same patient and, consistent with our organoid drivers of malignant cell state and that cell state dictates drug
studies, illustrate the critical role that TME-supplied cytokines sensitivity. Larger cohorts of longitudinal single-cell measure-
and immune activation may play in directing cancer cell state. ments, both spatially resolved and transcriptome-wide, from in-
dividual patients undergoing therapy are needed to fully assess
DISCUSSION variation in the TME, the kinetics of cell state plasticity, and their
consequences for therapeutic response.
Single-cell atlases of cancer have revolutionized our under- While the genetic evolution of ex vivo models is an established
standing of human malignancies and revealed that mutational phenomenon (Ben-David et al., 2019; Ben-David et al., 2018), our
and transcriptional heterogeneity are common. Critical next study shows similar ex vivo evolution along a transcriptional axis
steps include understanding what drives cancer cell state and that causes discordance between a patient sample and its
whether it can be targeted therapeutically. Here, we provide a matched avatar. For example, biopsies taken from PANFR0489R
systematic framework for assessing cancer cell states, identi- and PANFR0576 establish long-term cultures ex vivo, but they do
fying drivers of transcriptional plasticity, and evaluating their not represent the dominant cell state of the corresponding patient
functional significance in model systems. The use of matched tissue in vivo (Figure 3D). Given our demonstrated link between

(E) Cell state diagram time series for PANFR0562 organoids cultured with TGF-b or after TGF-b removal.
(F) Growth rate-adjusted dose response curves to gemcitabine and paclitaxel for models in (E). Points are the mean ± SD of 3 technical replicates.
(G) State-specific drug sensitivities in isogenic organoid model pairs skewed toward scBasal or scClassical states by altering media formulation. Points are the
mean ± SEM of 2–6 biologic replicates for the difference in growth rate-corrected Area Over the Curve (AOC) between each scBasal-scClassical model pair.
See also Figure S7; Tables S5, S6, and S7.

Cell 184, 6119–6137, December 9, 2021 6131

 ll
OPEN ACCESS Article

Figure 7. Malignant transcriptional states respond to TME alterations in organoid models and in vivo
(A) Differential expression (Wilcoxon rank sum test) of secreted factors by non-malignant cells (paracrine) grouped by their sample-averaged malignant cell
expression state in scBasal and scClassical (x axis) tumors and IC biopsies and the rest (y axis). State-specific genes that pass significance after multiple hy-
pothesis correction (p < 0.05) are colored by their group association.
(legend continued on next page)

6132 Cell 184, 6119–6137, December 9, 2021

 ll
Article OPEN ACCESS

cell state and therapeutic response, these findings establish the factors and state. In addition, we tracked CNV alterations to
necessity for preserving transcriptional fidelity in personalized define genetic clones, but future studies using barcode-based
medicine pipelines (Hahn et al., 2021). Since models from lineage tracing approaches are needed to confirm whether in-
different patients may also harbor differential plasticity and adapt- dividual cells transition between scClassical and scBasal states
ability, future efforts will need to evaluate models in multiple con- via the IC state as our data suggest (Wagner and Klein, 2020).
ditions to account for this latent property, define its drivers (e.g., While we focused predominantly on metastatic biopsies from
genetics, epigenetics), and accurately map vulnerabilities. Impor- the liver, it will be important in future studies to more broadly
tantly, manipulating the soluble microenvironment may offer a analyze how site-specific cues drive state plasticity. Even
more tractable approach for state-specific high-throughput with our limited number of samples from outside the liver, we
screening compared to more complex heterotypic co-cultures uncovered site-specific differences in mesenchymal popula-
or patient-derived xenograft systems. tions which revealed important differences between metastatic
Within the context of PDAC, clinical studies are ongoing to eval- and primary disease. Given the pivotal role that has been sug-
uate the efficacy of gemcitabine/nab-paclitaxel and FOLFIRINOX gested for the fibrotic TME in primary disease (Ho et al., 2020;
in patients with basal- versus classical-predominant metastatic Sahai et al., 2020), these findings carry important implications
disease (e.g., PASS01 – NCT04469556). Our observations sug- for targeting the stromal compartment in primary versus meta-
gest that basal tumors may exhibit broadly decreased sensitivity static PDAC. Future studies with larger sample and cell
to chemotherapy and highlight the need for new strategies to numbers will be needed to make comparisons across cell types
target this transcriptional subtype of PDAC. Importantly, we in primary and metastatic disease to fully understand how the
show that the transition to an scBasal state may render cells sen- TME regulates transcriptional phenotypes in these distinct
sitive to other classes of inhibitors. Since most PDAC tumors are niches. Here, advances in spatial transcriptomics will likely
heterogenous, combination strategies that suppress distinct cell empower a deeper understanding of these relationships across
states may be necessary for maximal synergistic effect (Palmer compartments (Longo et al., 2021). Finally, future studies that
et al., 2019; Palmer and Sorger, 2017). Furthermore, the ability broadly map the landscape of RNA state-dependent drug
of the TME to drive malignant cell state transitions suggests that response across various cancer models will be needed to fully
next generation therapeutic strategies may also need to target define the links between malignant cell transcriptional plasticity
site-specific supportive cells in the TME to control cell-state evo- and therapeutic response.
lution during therapy.
In sum, we provide a widely applicable framework to bench-
STAR+METHODS
mark cell states in patient-derived model systems, identify the
drivers of malignant transcriptional heterogeneity, and examine
Detailed methods are provided in the online version of this paper
the functional significance of cell state. As efforts to characterize
and include the following:
cell states across malignancies provide new and increasingly
higher resolution maps of patient tissues, an important next d KEY RESOURCES TABLE
step will be to understand better how to control and target cell d RESOURCE AVAILABILITY
state. We anticipate that our approach will provide a path toward B Lead contact
the systematic evaluation of cell state as a targetable feature in B Materials availability
cancer. B Data and code availability
d EXPERIMENTAL MODEL AND SUBJECT DETAILS
Limitations of the study B Human specimens
We focused on understanding how the TME supports cell state, B Human PDAC patient-derived organoid models
and thus we cannot comment directly on the epigenetic mech- d METHOD DETAILS
anisms through which these cell state transitions occur. Addi- B Sample preparation for single-cell RNA-sequencing of
tional studies into the regulatory mechanisms underlying clinical samples, organoids, and cell lines
PDAC state transitions will be a critical next step in further B Assessing organoid and cell line transcriptional states
delineating the relationships between cell-intrinsic and extrinsic under different matrix and media conditions

(B) Dot plot for state-specific significant differentially expressed paracrine factors (rows) by subtype-specific non-malignant cell types (columns). Dot size
represents that cell type’s fraction within tumors of each subtype, and fill color indicates average expression. Only cell types with a fractional representation > 5%
from each subtype are visualized.
(C) Density plots of IFN response score (top) and IC score (bottom) in control organoid cells and after addition of IFNg for 6 days. P-values were calculated by
Student’s t test.
(D) Biopsy samples from distinct metastatic sites (liver, dark gray; lung, light gray) in the same patient (PANFR0473) demonstrate co-variation in T cell IFNG
expression (top), malignant cell IFN response score (middle), and malignant IC score (bottom). P-values for density plot differences were calculated by Student’s
t test.
(E) Biopsy samples from the same lesion pre- and post-immunotherapy (checkpoint inhibitor plus a macrophage-targeting agent; pre-, PANFR0489, pink; post-,
PANFR0489R, blue) demonstrate coordinated changes with treatment in T cell IFNG expression (top), malignant cell IFN response score (middle), and malignant
IC score (bottom). P-values for density plot differences were calculated by Student’s t test.
(F) Heatmap for malignant cell state shifts from samples in (E).
See also Figure S7; Tables S5, S6, and S7.

Cell 184, 6119–6137, December 9, 2021 6133

 ll
OPEN ACCESS Article
B Compound sensitivity testing in cell lines and organoid AUTHOR CONTRIBUTIONS
models
B Single-cell RNA-seq (scRNA-seq) data library genera- Conceptualization, S.R., P.S.W., B.M.W., W.C.H., A.J.A., A.K.S.; Methodol-
ogy, S.R., P.S.W., A.W.N., H.L.W., L.C., J.A.N., B.M.W., W.C.H., A.J.A.,
tion, sequencing, and alignment
A.K.S.; Validation, S.R., P.S.W., A.W.N., H.L.W., K.E.L., J.A.N., B.M.W.,
B Bulk RNA- and DNA-sequencing of organoids W.C.H., A.J.A., A.K.S.; Formal analysis, S.R., P.S.W., A.W.N., H.L.W., A.D.,
B Multiplex immunofluorescence imaging K.S.K., A.A.B., N.L., L.F.S., Y.Y.L., A.D. Cherniack, M.A.B., E.F.C., M.Y.T.,
d QUANTIFICATION AND STATISTICAL ANALYSIS J.M.M., L.C.; Investigation, S.R., P.S.W., A.W.N., H.L.W., K.E.L., J.G.-R.,
B Mutation and CNV identification from bulk DNA- R.L.K., N.M., M.S.R., R.W.S.N., J.W., I.W.; Resources, S.A.V., A.D. Costa,
sequencing E.K.H., D.Y.R., L.K.B., A.R., B.E.J., E.T.S., R.J.S., G.I.S, T.E.C., K.P., D.A.R.,
K.N., J.M.C., J.A.N., B.M.W.; Data Curation, P.S.W., S.R., A.W.N., H.L.W.,
B Bulk RNA-sequencing analysis
K.S.K.; Writing – original draft, S.R., P.S.W., A.W.N., H.L.W.; Writing – review
B Single-cell data quality pre-processing and initial cell
& editing, S.R., P.S.W., A.W.N., H.L.W., A.M.J., L.C., J.A.N., S.R.M., B.M.W.,
type discovery W.C.H., A.J.A., A.K.S.; Visualization, S.R., P.S.W., A.W.N., H.L.W.; Supervi-
B Single-cell CNV identification sion, B.E.J., T.E.J., L.C., J.A.N., S.R.M., B.M.W., W.C.H., A.J.A., A.K.S.; Proj-
B Subclonal analysis with single-cell inferred CNVs ect administration, S.R., P.S.W., B.M.W., W.C.H., A.J.A., A.K.S.; Funding
B Subclustering of malignant and non-malignant cells acquisition, S.R., P.S.W., S.R.M., B.M.W., W.C.H., A.J.A., A.K.S.
B Generation of expression signatures/scores
DECLARATION OF INTERESTS
B Analysis of normal pancreas progenitor data
B Matched organoid clustering and cell-typing
B.M.W. reports research support from Celgene and Eli Lilly and consulting
B Correlation distances for genotype and transcriptional for BioLineRx, Celgene, G1 Therapeutics, and GRAIL. W.C.H. is a consultant
cell state for Thermo Fisher, Solasta Ventures, MPM Capital, Tyra Biosciences, iTeos,
B Matched biopsy versus organoid malignant cell Frontier Medicines, Function Oncology, KSQ Therapeutics, Jubilant Thera-
comparison peutics, RAPPTA Therapeutics, and Paraxel. A.K.S. reports compensation
for consulting and/or SAB membership from Merck, Honeycomb Biotechnol-
B TME associations
ogies, Cellarity, Repertoire Immune Medicines, Ochre Bio, Third Rock Ven-
B Biopsy paracrine and autocrine subtype-specific fac-
tures, Hovione, Relation Therapeutics, FL82, and Dahlia Biosciences.
tor analysis A.J.A. has consulted for Oncorus, Arrakis Therapeutics, and Merck & Co.
B Tumor phenotyping from mIF data and has research funding from Mirati Therapeutics, Syros, Deerfield, and
Novo Ventures that are unrelated to this project. S.R.M. is a founder of Tra-
vera. S.R. has equity in Amgen. J.M.C. has received research funding from
SUPPLEMENTAL INFORMATION
Merck, Tesaro, AstraZeneca, Bayer, and Esperas Pharma; has served as a
consultant to Bristol Myers Squibb; and has received travel funding from
Supplemental information can be found online at https://doi.org/10.1016/j.cell.
Bristol Myers Squibb. A.R. is an employee of AstraZeneca and an equity
2021.11.017.
holder in Celsius Therapeutics and NucleAI. Y.Y.L. reports equity from
g.Root Biomedical Services. A.D. Cherniack reports research support from
ACKNOWLEDGMENTS Bayer. R.J.S reports research support from Merck and consulting and/or
SAB membership for Bristol Myers Squibb, Merck, Novartis, and Pfizer.
This work was funded by the Lustgarten Foundation Dedicated Laboratory G.I.S. reports sponsored research support from Eli Lilly, Merck KGaA/
(B.M.W. and A.J.A.), Dana-Farber Cancer Institute Hale Center for Pancreatic EMD-Serono, Merck, and Sierra Oncology and has served on advisory
Cancer Research (A.J.A., B.M.W., and W.C.H.), the Doris Duke Charitable boards for Pfizer, Eli Lilly, G1 Therapeutics, Roche, Merck KGaA/EMD-Se-
Foundation (A.J.A.), Pancreatic Cancer Action Network (A.J.A. and B.M.W.), rono, Sierra Oncology, Bicycle Therapeutics, Fusion Pharmaceuticals, Cy-
NIH-NCI K08 CA218420-02 (A.J.A.), P50 CA127003 (A.J.A., B.M.W., and brexa Therapeutics, Astex, Almac, Ipsen, Bayer, Angiex, Daiichi Sankyo, Se-
G.I.S.), U01 CA176058 (W.C.H.), U01 CA224146 (W.C.H. and A.J.A.), U01 attle Genetics, Boehringer Ingelheim, ImmunoMet, Asana, Artios, Atrin,
CA210171 (B.M.W.), U01 28020510 (W.C.H. and A.K.S.), 1U2CCA23319501 Concarlo Holdings, Syros, Zentalis, CytomX Therapeutics, Blueprint Medi-
(A.K.S.), U54 CA217377 (W.C.H., S.R.M., and A.K.S.), U01 CA250549 cines, and Kymera Therapeutics. T.J. is a member of the Board of Directors
(W.C.H., S.R., and A.J.A.), Stand Up To Cancer (B.M.W.), Noble Effort Fund of Amgen and Thermo Fisher Scientific; a co-Founder of Dragonfly Thera-
(B.M.W.), Wexler Family Fund (B.M.W.), Promises for Purple (B.M.W.), Hope peutics and T2 Biosystems; serves on the Scientific Advisory Board of Drag-
Funds for Cancer Research Postdoctoral Fellowship (S.R.), Harvard Cata- onfly Therapeutics, SQZ Biotech, and Skyhawk Therapeutics; and is the
lyst/The Harvard Clinical and Translational Science Center UL1 TR002541 President of Break Through Cancer. None of these affiliations represent a
(S.R.), Koch Institute for Integrative Cancer Research at MIT and the Dana- conflict of interest with respect to the design or execution of this study or
Farber/Harvard Cancer Center Bridge Project (W.C.H., S.R.M., and S.R.), interpretation of data presented in this manuscript. T.J.’s laboratory currently
Dana-Farber Cancer Institute Gloria Spivak Faculty Advancement Fund also receives funding from the Johnson & Johnson Lung Cancer Initiative and
(S.R.), Hopper-Belmont Foundation Inspiration Award (S.R.), Marcotte Center The Lustgarten Foundation for Pancreatic Cancer Research, but this funding
for Cancer Research (B.E.J.), Finnish Cultural Foundation (S.A.V.), Orion did not support the research described in this manuscript. K.N. reports
Research Foundation sr (S.A.V.), Damon Runyon Cancer Research Founda- research support from Revolution Medicines, Evergrande Group, Pharma-
tion (A.M.J.), NIH K99 Award CA241072-01 (A.M.J.), Koch Institute Cancer vite, and Janssen; is a member of SABs for SeaGen, BiomX, and Bicara
Core Grant NIH P30 CA14051 (T.E.J. and A.M.J.), Broman Family Fund for Therapeutics; and has consulted for X-Biotix Therapeutics and Redesign
Pancreatic Cancer Research (K.N.), Ludwig Center for Molecular Oncology Health. A.K.S., P.S.W., A.W.N., S.R., W.C.H., A.J.A., B.W.M., and J.G.R.
at MIT (A.W.N.), Searle Scholars Program (A.K.S.), Beckman Young Investi- have filed a patent (Publication number WO/2021/081491; International
gator Program (A.K.S.), Sloan Fellowship in Chemistry (A.K.S.), and the Pew- application number PCT/US2020/057342) related to this work.
Stewart Scholars Program for Cancer Research (A.K.S.). This work was sup-
ported by the National Cancer Institute’s Office of Cancer Genomics Cancer INCLUSION AND DIVERSITY
Target Discovery and Development (CTD^2) initiative. We also thank SciSto-
ries for assistance with schematics and Bryan Bryson for helpful discussions We worked to ensure gender balance in the recruitment of human subjects. We
of macrophage phenotypes. worked to ensure ethnic or other types of diversity in the recruitment of human

6134 Cell 184, 6119–6137, December 9, 2021

 ll
Article OPEN ACCESS

subjects. We worked to ensure diversity in experimental samples through the Initiative (2020). HNF4A and GATA6 Loss Reveals Therapeutically Actionable
selection of the cell lines. We worked to ensure diversity in experimental sam- Subtypes in Pancreatic Cancer. Cell Rep. 31, 107625.
ples through the selection of the genomic datasets. One or more of the authors Byrne, A.T., Alférez, D.G., Amant, F., Annibali, D., Arribas, J., Biankin, A.V.,
of this paper self-identifies as an underrepresented ethnic minority in science. Bruna, A., Budinská, E., Caldas, C., Chang, D.K., et al. (2017). Interrogating
One or more of the authors of this paper self-identifies as a member of the open issues in cancer medicine with patient-derived xenografts. Nat. Rev.
LGBTQ+ community. The author list of this paper includes contributors from Cancer 17, 632.
the location where the research was conducted who participated in the data
Cancer Genome Atlas Research Network (2017). Integrated Genomic Charac-
collection, design, analysis, and/or interpretation of the work.
terization of Pancreatic Ductal Adenocarcinoma. Cancer Cell 32, 185–
203.e13.
Received: March 12, 2021
Revised: September 24, 2021 Chan-Seng-Yue, M., Kim, J.C., Wilson, G.W., Ng, K., Figueroa, E.F., O’Kane,
Accepted: November 11, 2021 G.M., Connor, A.A., Denroche, R.E., Grant, R.C., McLeod, J., et al. (2020).
Published: December 9, 2021 Transcription phenotypes of pancreatic cancer are driven by genomic events
during tumor evolution. Nat. Genet. 52, 231–240.
REFERENCES Collisson, E.A., Sadanandam, A., Olson, P., Gibb, W.J., Truitt, M., Gu, S.,
Cooc, J., Weinkle, J., Kim, G.E., Jakkula, L., et al. (2011). Subtypes of pancre-
Abo, R.P., Ducar, M., Garcia, E.P., Thorner, A.R., Rojas-Rudilla, V., Lin, L., atic ductal adenocarcinoma and their differing responses to therapy. Nat.
Sholl, L.M., Hahn, W.C., Meyerson, M., Lindeman, N.I., et al. (2015). BreaKmer: Med. 17, 500–503.
detection of structural variation in targeted massively parallel sequencing data Collisson, E.A., Bailey, P., Chang, D.K., and Biankin, A.V. (2019). Molecular
using kmers. Nucleic Acids Res. 43, e19. subtypes of pancreatic cancer. Nat. Rev. Gastroenterol. Hepatol. 16, 207–220.
Aguirre, A.J., Nowak, J.A., Camarda, N.D., Moffitt, R.A., Ghazani, A.A., Hazar- Connor, A.A., Denroche, R.E., Jang, G.H., Lemire, M., Zhang, A., Chan-Seng-
Rethinam, M., Raghavan, S., Kim, J., Brais, L.K., Ragon, D., et al. (2018). Real- Yue, M., Wilson, G., Grant, R.C., Merico, D., Lungu, I., et al. (2019). Integration
time Genomic Characterization of Advanced Pancreatic Cancer to Enable of Genomic and Transcriptional Features in Pancreatic Cancer Reveals
Precision Medicine. Cancer Discov. 8, 1096–1111. Increased Cell Cycle Progression in Metastases. Cancer Cell 35, 267–
Alonso-Curbelo, D., Ho, Y.J., Burdziak, C., Maag, J.L.V., Morris, J.P., 4th, 282 e267.
Chandwani, R., Chen, H.A., Tsanov, K.M., Barriga, F.M., Luan, W., et al. Di Carlo, S.E., and Peduto, L. (2018). The perivascular origin of pathological fi-
(2021). A gene-environment-induced epigenetic program initiates tumorigen- broblasts. J. Clin. Invest. 128, 54–63.
esis. Nature 590, 642–648. Drost, J., and Clevers, H. (2018). Organoids in cancer research. Nat. Rev. Can-
Aung, K.L., Fischer, S.E., Denroche, R.E., Jang, G.H., Dodd, A., Creighton, S., cer 18, 407–418.
Southwood, B., Liang, S.B., Chadwick, D., Zhang, A., et al. (2018). Genomics- Elyada, E., Bolisetty, M., Laise, P., Flynn, W.F., Courtois, E.T., Burkhart, R.A.,
Driven Precision Medicine for Advanced Pancreatic Cancer: Early Results Teinor, J.A., Belleau, P., Biffi, G., Lucito, M.S., et al. (2019). Cross-Species Sin-
from the COMPASS Trial. Clin. Cancer Res. 24, 1344–1354. gle-Cell Analysis of Pancreatic Ductal Adenocarcinoma Reveals Antigen-Pre-
Bailey, P., Chang, D.K., Nones, K., Johns, A.L., Patch, A.M., Gingras, M.C., senting Cancer-Associated Fibroblasts. Cancer Discov. 9, 1102–1123.
Miller, D.K., Christ, A.N., Bruxner, T.J., Quinn, M.C., et al.; Australian Pancre- Fan, J., Lee, H.O., Lee, S., Ryu, D.E., Lee, S., Xue, C., Kim, S.J., Kim, K., Bar-
atic Cancer Genome Initiative (2016). Genomic analyses identify molecular kas, N., Park, P.J., et al. (2018). Linking transcriptional and genetic tumor het-
subtypes of pancreatic cancer. Nature 531, 47–52. erogeneity through allele analysis of single-cell RNA-seq data. Genome Res.
Balachandran, V.P., Beatty, G.L., and Dougan, S.K. (2019). Broadening the 28, 1217–1227.
Impact of Immunotherapy to Pancreatic Cancer: Challenges and Opportu- Filbin, M.G., Tirosh, I., Hovestadt, V., Shaw, M.L., Escalante, L.E., Mathewson,
nities. Gastroenterology 156, 2056–2072. N.D., Neftel, C., Frank, N., Pelton, K., Hebert, C.M., et al. (2018). Develop-
Bartoschek, M., Oskolkov, N., Bocci, M., Lövrot, J., Larsson, C., Sommarin, mental and oncogenic programs in H3K27M gliomas dissected by single-
M., Madsen, C.D., Lindgren, D., Pekar, G., Karlsson, G., et al. (2018). Spatially cell RNA-seq. Science 360, 331–335.
and functionally distinct subclasses of breast cancer-associated fibroblasts Fujii, M., Matano, M., Toshimitsu, K., Takano, A., Mikami, Y., Nishikori, S., Su-
revealed by single cell RNA sequencing. Nat. Commun. 9, 5150. gimoto, S., and Sato, T. (2018). Human Intestinal Organoids Maintain Self-
Ben-David, U., Siranosian, B., Ha, G., Tang, H., Oren, Y., Hinohara, K., Strath- Renewal Capacity and Cellular Diversity in Niche-Inspired Culture Condition.
dee, C.A., Dempster, J., Lyons, N.J., Burns, R., et al. (2018). Genetic and tran- Cell Stem Cell 23, 787–793 e786.
scriptional evolution alters cancer cell line drug response. Nature 560, Garcia, E.P., Minkovsky, A., Jia, Y., Ducar, M.D., Shivdasani, P., Gong, X., Li-
325–330. gon, A.H., Sholl, L.M., Kuo, F.C., MacConaill, L.E., et al. (2017). Validation of
Ben-David, U., Beroukhim, R., and Golub, T.R. (2019). Genomic evolution of OncoPanel: A Targeted Next-Generation Sequencing Assay for the Detection
cancer models: perils and opportunities. Nat. Rev. Cancer 19, 97–109. of Somatic Variants in Cancer. Arch. Pathol. Lab. Med. 141, 751–758.
Bernard, V., Semaan, A., Huang, J., San Lucas, F.A., Mulu, F.C., Stephens, Garraway, L.A., and Lander, E.S. (2013). Lessons from the cancer genome.
B.M., Guerrero, P.A., Huang, Y., Zhao, J., Kamyabi, N., et al. (2019). Single- Cell 153, 17–37.
Cell Transcriptomics of Pancreatic Cancer Precursors Demonstrates Epithelial Ghandi, M., Huang, F.W., Jané-Valbuena, J., Kryukov, G.V., Lo, C.C., McDo-
and Microenvironmental Heterogeneity as an Early Event in Neoplastic Pro- nald, E.R., 3rd, Barretina, J., Gelfand, E.T., Bielski, C.M., Li, H., et al. (2019).
gression. Clin. Cancer Res. 25, 2194–2205. Next-generation characterization of the Cancer Cell Line Encyclopedia. Nature
Bi, W.L., Greenwald, N.F., Ramkissoon, S.H., Abedalthagafi, M., Coy, S.M., 569, 503–508.
Ligon, K.L., Mei, Y., MacConaill, L., Ducar, M., Min, L., et al. (2017). Clinical Gierahn, T.M., Wadsworth, M.H., 2nd, Hughes, T.K., Bryson, B.D., Butler, A.,
Identification of Oncogenic Drivers and Copy-Number Alterations in Pituitary Satija, R., Fortune, S., Love, J.C., and Shalek, A.K. (2017). Seq-Well: portable,
Tumors. Endocrinology 158, 2284–2291. low-cost RNA sequencing of single cells at high throughput. Nat. Methods 14,
Boj, S.F., Hwang, C.I., Baker, L.A., Chio, I.I., Engle, D.D., Corbo, V., Jager, M., 395–398.
Ponz-Sarvise, M., Tiriac, H., Spector, M.S., et al. (2015). Organoid models of Gillet, J.P., Varma, S., and Gottesman, M.M. (2013). The clinical relevance of
human and mouse ductal pancreatic cancer. Cell 160, 324–338. cancer cell lines. J. Natl. Cancer Inst. 105, 452–458.
Brunton, H., Caligiuri, G., Cunningham, R., Upstill-Goddard, R., Bailey, U.M., Gröger, C.J., Grubinger, M., Waldhör, T., Vierlinger, K., and Mikulits, W. (2012).
Garner, I.M., Nourse, C., Dreyer, S., Jones, M., Moran-Jones, K., et al.; Glas- Meta-analysis of gene expression signatures defining the epithelial to mesen-
gow Precision Oncology Laboratory; Australian Pancreatic Cancer Genome chymal transition during cancer progression. PLoS ONE 7, e51136.

Cell 184, 6119–6137, December 9, 2021 6135

 ll
OPEN ACCESS Article
Grünwald, B.T., Devisme, A., Andrieux, G., Vyas, F., Aliar, K., McCloskey, Miyabayashi, K., Baker, L.A., Deschênes, A., Traub, B., Caligiuri, G., Plenker,
C.W., Macklin, A., Jang, G.H., Denroche, R., Romero, J.M., et al. (2021). D., Alagesan, B., Belleau, P., Li, S., Kendall, J., et al. (2020). Intraductal Trans-
Spatially confined sub-tumor microenvironments in pancreatic cancer. Cell plantation Models of Human Pancreatic Ductal Adenocarcinoma Reveal Pro-
184, 5577–5592.e18. gressive Transition of Molecular Subtypes. Cancer Discov. 10, 1566–1589.
Hafner, M., Niepel, M., Chung, M., and Sorger, P.K. (2016). Growth rate inhibi- Moffitt, R.A., Marayati, R., Flate, E.L., Volmar, K.E., Loeza, S.G., Hoadley, K.A.,
tion metrics correct for confounders in measuring sensitivity to cancer drugs. Rashid, N.U., Williams, L.A., Eaton, S.C., Chung, A.H., et al. (2015). Virtual
Nat. Methods 13, 521–527. microdissection identifies distinct tumor- and stroma-specific subtypes of
Hahn, W.C., Bader, J.S., Braun, T.P., Califano, A., Clemons, P.A., Druker, B.J., pancreatic ductal adenocarcinoma. Nat. Genet. 47, 1168–1178.
Ewald, A.J., Fu, H., Jagu, S., Kemp, C.J., et al.; Cancer Target Discovery and Nam, A.S., Chaligne, R., and Landau, D.A. (2021). Integrating genetic and non-
Development Network (2021). An expanded universe of cancer targets. Cell genetic determinants of cancer evolution by single-cell multi-omics. Nat. Rev.
184, 1142–1155. Genet. 22, 3–18.
Hara, T., Chanoch-Myers, R., Mathewson, N.D., Myskiw, C., Atta, L., Bus- Neftel, C., Laffy, J., Filbin, M.G., Hara, T., Shore, M.E., Rahme, G.J., Richman,
sema, L., Eichhorn, S.W., Greenwald, A.C., Kinker, G.S., Rodman, C., et al. A.R., Silverbush, D., Shaw, M.L., Hebert, C.M., et al. (2019). An Integrative
(2021). Interactions between cancer cells and immune cells drive transitions Model of Cellular States, Plasticity, and Genetics for Glioblastoma. Cell 178,
to mesenchymal-like states in glioblastoma. Cancer Cell 39, 779–792 e711. 835–849 e821.

Hayashi, A., Fan, J., Chen, R., Ho, Y.-j., Makohon-Moore, A.P., Lecomte, N., O’Kane, G.M., Grunwald, B.T., Jang, G.H., Masoomian, M., Picardo, S., Grant,
Zhong, Y., Hong, J., Huang, J., Sakamoto, H., et al. (2020). A unifying paradigm R.C., Denroche, R.E., Zhang, A., Wang, Y., Lam, B., et al. (2020). GATA6
for transcriptional heterogeneity and squamous features in pancreatic ductal Expression Distinguishes Classical and Basal-like Subtypes in Advanced
adenocarcinoma. Nature Cancer 1, 59–74. Pancreatic Cancer. Clin Cancer Res. 26, 4901–4910.

Ho, W.J., Jaffee, E.M., and Zheng, L. (2020). The tumour microenvironment in Öhlund, D., Handly-Santana, A., Biffi, G., Elyada, E., Almeida, A.S., Ponz-Sar-
pancreatic cancer - clinical challenges and opportunities. Nat. Rev. Clin. On- vise, M., Corbo, V., Oni, T.E., Hearn, S.A., Lee, E.J., et al. (2017). Distinct pop-
col. 17, 527–540. ulations of inflammatory fibroblasts and myofibroblasts in pancreatic cancer.
J. Exp. Med. 214, 579–596.
Hosaka, K., Yang, Y., Seki, T., Fischer, C., Dubey, O., Fredlund, E., Hartman,
Palmer, A.C., and Sorger, P.K. (2017). Combination Cancer Therapy Can
J., Religa, P., Morikawa, H., Ishii, Y., et al. (2016). Pericyte-fibroblast transition
Confer Benefit via Patient-to-Patient Variability without Drug Additivity or Syn-
promotes tumor growth and metastasis. Proc. Natl. Acad. Sci. USA 113,
ergy. Cell 171, 1678–1691 e1613.
E5618–E5627.
Palmer, A.C., Chidley, C., and Sorger, P.K. (2019). A curative combination can-
Hovestadt, V., Smith, K.S., Bihannic, L., Filbin, M.G., Shaw, M.L., Baumgart-
cer therapy achieves high fractional cell killing through low cross-resistance
ner, A., DeWitt, J.C., Groves, A., Mayr, L., Weisman, H.R., et al. (2019).
and drug additivity. eLife 8, e50036.
Resolving medulloblastoma cellular architecture by single-cell genomics. Na-
ture 572, 74–79. Patel, A.P., Tirosh, I., Trombetta, J.J., Shalek, A.K., Gillespie, S.M., Wakimoto,
H., Cahill, D.P., Nahed, B.V., Curry, W.T., Martuza, R.L., et al. (2014). Single-
Hughes, T.K., Wadsworth, M.H., 2nd, Gierahn, T.M., Do, T., Weiss, D., An-
cell RNA-seq highlights intratumoral heterogeneity in primary glioblastoma.
drade, P.R., Ma, F., de Andrade Silva, B.J., Shao, S., Tsoi, L.C., et al. (2020).
Science 344, 1396–1401.
Second-Strand Synthesis-Based Massively Parallel scRNA-Seq Reveals
Cellular States and Molecular Features of Human Inflammatory Skin Pathol- Pelon, F., Bourachot, B., Kieffer, Y., Magagna, I., Mermet-Meillon, F., Bonnet,
ogies. Immunity 53, 878–894 e877. I., Costa, A., Givel, A.M., Attieh, Y., Barbazan, J., et al. (2020). Cancer-associ-
ated fibroblast heterogeneity in axillary lymph nodes drives metastases in
Hyman, D.M., Taylor, B.S., and Baselga, J. (2017). Implementing Genome-
breast cancer through complementary mechanisms. Nat. Commun. 11, 404.
Driven Oncology. Cell 168, 584–599.
Porter, R.L., Magnus, N.K.C., Thapar, V., Morris, R., Szabolcs, A., Neyaz, A.,
Kim, J.H., Park, S.Y., Jun, Y., Kim, J.Y., and Nam, J.S. (2017). Roles of Wnt
Kulkarni, A.S., Tai, E., Chougule, A., Hillis, A., et al. (2019). Epithelial to mesen-
Target Genes in the Journey of Cancer Stem Cells. Int. J. Mol. Sci. 18, 1604.
chymal plasticity and differential response to therapies in pancreatic ductal
Kim, C., Gao, R., Sei, E., Brandt, R., Hartman, J., Hatschek, T., Crosetto, N., adenocarcinoma. Proc. Natl. Acad. Sci. USA 116, 26835–26845.
Foukakis, T., and Navin, N.E. (2018). Chemoresistance Evolution in Triple- Puram, S.V., Tirosh, I., Parikh, A.S., Patel, A.P., Yizhak, K., Gillespie, S.,
Negative Breast Cancer Delineated by Single-Cell Sequencing. Cell 173, Rodman, C., Luo, C.L., Mroz, E.A., Emerick, K.S., et al. (2017). Single-Cell
879–893 e813. Transcriptomic Analysis of Primary and Metastatic Tumor Ecosystems in
Li, B., Gould, J., Yang, Y., Sarkizova, S., Tabaka, M., Ashenberg, O., Rosen, Y., Head and Neck Cancer. Cell 171, 1611–1624 e1624.
Slyper, M., Kowalczyk, M.S., Villani, A.C., et al. (2020). Cumulus provides Qadir, M.M.F., Álvarez-Cubela, S., Klein, D., van Dijk, J., Muñiz-Anquela, R.,
cloud-based data analysis for large-scale single-cell and single-nucleus Moreno-Hernández, Y.B., Lanzoni, G., Sadiq, S., Navarro-Rubio, B., Garcı́a,
RNA-seq. Nat. Methods 17, 793–798. M.T., et al. (2020). Single-cell resolution analysis of the human pancreatic
Li, Y., He, Y., Peng, J., Su, Z., Li, Z., Zhang, B., Ma, J., Zhuo, M., Zou, D., Liu, X., ductal progenitor cell niche. Proc. Natl. Acad. Sci. USA 117, 10876–10887.
et al. (2021). Mutant Kras co-opts a proto-oncogenic enhancer network in Sade-Feldman, M., Yizhak, K., Bjorgaard, S.L., Ray, J.P., de Boer, C.G., Jen-
inflammation-induced metaplastic progenitor cells to initiate pancreatic can- kins, R.W., Lieb, D.J., Chen, J.H., Frederick, D.T., Barzily-Rokni, M., et al.
cer. Nature Cancer 2, 49–65. (2019). Defining T Cell States Associated with Response to Checkpoint Immu-
Ligorio, M., Sil, S., Malagon-Lopez, J., Nieman, L.T., Misale, S., Di Pilato, M., notherapy in Melanoma. Cell 176, 404.
Ebright, R.Y., Karabacak, M.N., Kulkarni, A.S., Liu, A., et al. (2019). Stromal Sahai, E., Astsaturov, I., Cukierman, E., DeNardo, D.G., Egeblad, M., Evans,
Microenvironment Shapes the Intratumoral Architecture of Pancreatic Cancer. R.M., Fearon, D., Greten, F.R., Hingorani, S.R., Hunter, T., et al. (2020). A
Cell 178, 160–175 e127. framework for advancing our understanding of cancer-associated fibroblasts.
Longo, S.K., Guo, M.G., Ji, A.L., and Khavari, P.A. (2021). Integrating single- Nat. Rev. Cancer 20, 174–186.
cell and spatial transcriptomics to elucidate intercellular tissue dynamics. Sholl, L.M., Do, K., Shivdasani, P., Cerami, E., Dubuc, A.M., Kuo, F.C., Garcia,
Nat. Rev. Genet. 22, 627–644. E.P., Jia, Y., Davineni, P., Abo, R.P., et al. (2016). Institutional implementation
Macosko, E.Z., Basu, A., Satija, R., Nemesh, J., Shekhar, K., Goldman, M., Tir- of clinical tumor profiling on an unselected cancer population. JCI Insight 1,
osh, I., Bialas, A.R., Kamitaki, N., Martersteck, E.M., et al. (2015). Highly Par- e87062.
allel Genome-wide Expression Profiling of Individual Cells Using Nanoliter Somerville, T.D.D., Xu, Y., Miyabayashi, K., Tiriac, H., Cleary, C.R., Maia-Silva,
Droplets. Cell 161, 1202–1214. D., Milazzo, J.P., Tuveson, D.A., and Vakoc, C.R. (2018). TP63-Mediated

6136 Cell 184, 6119–6137, December 9, 2021

 ll
Article OPEN ACCESS

Enhancer Reprogramming Drives the Squamous Subtype of Pancreatic Ductal Venteicher, A.S., Tirosh, I., Hebert, C., Yizhak, K., Neftel, C., Filbin, M.G., Hov-
Adenocarcinoma. Cell Rep. 25, 1741–1755 e1747. estadt, V., Escalante, L.E., Shaw, M.L., Rodman, C., et al. (2017). Decoupling
Spurr, L.F., Touat, M., Taylor, A.M., Dubuc, A.M., Shih, J., Meredith, D.M., genetics, lineages, and microenvironment in IDH-mutant gliomas by single-
Pisano, W.V., Meyerson, M.L., Ligon, K.L., Cherniack, A.D., et al. (2021). Quan- cell RNA-seq. Science 355, eaai8478.
tification of aneuploidy in targeted sequencing data using ASCETS. Bioinfor-
Vivian, J., Rao, A.A., Nothaft, F.A., Ketchum, C., Armstrong, J., Novak, A., Pfeil,
matics 37, 2461–2463.
J., Narkizian, J., Deran, A.D., Musselman-Brown, A., et al. (2017). Toil enables
Suvà, M.L., and Tirosh, I. (2019). Single-Cell RNA Sequencing in Cancer: Les- reproducible, open source, big biomedical data analyses. Nat. Biotechnol. 35,
sons Learned and Emerging Challenges. Mol. Cell 75, 7–12. 314–316.
Tiriac, H., Belleau, P., Engle, D.D., Plenker, D., Deschênes, A., Somerville,
Wagner, D.E., and Klein, A.M. (2020). Lineage tracing meets single-cell omics:
T.D.D., Froeling, F.E.M., Burkhart, R.A., Denroche, R.E., Jang, G.H., et al.
opportunities and challenges. Nat. Rev. Genet. 21, 410–427.
(2018). Organoid Profiling Identifies Common Responders to Chemotherapy
in Pancreatic Cancer. Cancer Discov. 8, 1112–1129. Weinstein, J.N., Collisson, E.A., Mills, G.B., Shaw, K.R., Ozenberger, B.A., Ell-
Tirosh, I., Izar, B., Prakadan, S.M., Wadsworth, M.H., 2nd, Treacy, D., rott, K., Shmulevich, I., Sander, C., and Stuart, J.M.; Cancer Genome Atlas
Trombetta, J.J., Rotem, A., Rodman, C., Lian, C., Murphy, G., et al. (2016a). Research Network (2013). The Cancer Genome Atlas Pan-Cancer analysis
Dissecting the multicellular ecosystem of metastatic melanoma by single- project. Nat. Genet. 45, 1113–1120.
cell RNA-seq. Science 352, 189–196.
Wöll, S., Schlitter, A.M., Dhaene, K., Roller, M., Esposito, I., Sahin, U., and Tür-
Tirosh, I., Venteicher, A.S., Hebert, C., Escalante, L.E., Patel, A.P., Yizhak, K., eci, Ö. (2014). Claudin 18.2 is a target for IMAB362 antibody in pancreatic neo-
Fisher, J.M., Rodman, C., Mount, C., Filbin, M.G., et al. (2016b). Single-cell plasms. Int. J. Cancer 134, 731–739.
RNA-seq supports a developmental hierarchy in human oligodendroglioma.
Nature 539, 309–313. Zhang, L., Li, Z., Skrzypczynska, K.M., Fang, Q., Zhang, W., O’Brien, S.A., He,
van de Haar, J., Hoes, L.R., Roepman, P., Lolkema, M.P., Verheul, H.M.W., Y., Wang, L., Zhang, Q., Kim, A., et al. (2020). Single-Cell Analyses Inform
Gelderblom, H., de Langen, A.J., Smit, E.F., Cuppen, E., Wessels, L.F.A., Mechanisms of Myeloid-Targeted Therapies in Colon Cancer. Cell 181, 442–
and Voest, E.E. (2021). Limited evolution of the actionable metastatic cancer 459 e429.
genome under therapeutic pressure. Nat. Med. 27, 1553–1563. Zilionis, R., Engblom, C., Pfirschke, C., Savova, V., Zemmour, D., Saatcioglu,
van Galen, P., Hovestadt, V., Wadsworth Ii, M.H., Hughes, T.K., Griffin, G.K., H.D., Krishnan, I., Maroni, G., Meyerovitz, C.V., Kerwin, C.M., et al. (2019). Sin-
Battaglia, S., Verga, J.A., Stephansky, J., Pastika, T.J., Lombardi Story, J., gle-Cell Transcriptomics of Human and Mouse Lung Cancers Reveals
et al. (2019). Single-Cell RNA-Seq Reveals AML Hierarchies Relevant to Dis- Conserved Myeloid Populations across Individuals and Species. Immunity
ease Progression and Immunity. Cell 176, 1265–1281 e1224. 50, 1317–1334 e1310.

Cell 184, 6119–6137, December 9, 2021 6137

 ll
OPEN ACCESS Article

STAR+METHODS

KEY RESOURCES TABLE

REAGENT or RESOURCE SOURCE IDENTIFIER

Antibodies
Mouse monoclonal anti-cytokeratin 17, Thermo Fisher Cat# MA5-13539; RRID:AB_10980102
clone E3
Rabbit monoclonal anti-S100A2, clone Abcam Cat# ab109494; RRID:AB_10859000
EPR5392
Rabbit monoclonal anti-Claudin18.2, clone Abcam Cat# ab241330
EPR19202-244
Rabbit monoclonal anti-GATA-6 XP, Cell Signaling Technology Cat# 5851; RRID:AB_10705521
clone D61E4
Rabbit monoclonal anti-TFF1 (estrogen Abcam Cat# ab92377; RRID:AB_10562112
inducible protein pS2), clone EPR3972
Mouse monoclonal anti-cytokeratin, clone Agilent/Dako Cat# M3515; RRID:AB_2132885
AE1/AE3
Mouse monoclonal anti-pan-keratin, Cell Signaling Technology Cat# 4545; RRID:AB_490860
clone C11
Opal polymer HRP anti-mouse and Akoya Biosciences Cat# ARH1001EA; RRID:AB_2890927
anti-rabbit secondary antibody
Biological samples
Human PDAC samples This study N/A
Chemicals, peptides, and recombinant proteins
Advanced DMEM/F12 Thermo Fisher Cat# 12634028
RPMI 1640 Corning Cat# 10-040-CV
Fetal bovine serum Sigma Cat# F4135
Penicillin/streptomycin Thermo Fisher Cat# 15140122
Primocin Invivogen Cat# ant-pm-1
HEPES Thermo Fisher Cat# 15630080
GlutaMAX Thermo Fisher Cat# 35050061
A83-01 Tocris Cat# 2939
Recombinant mouse Noggin Peprotech Cat# 250-38E
Recombinant mouse EGF Peprotech Cat# 315-09
Recombinant human FGF10 Peprotech Cat# 100-26
Human [Leu15]-Gastrin I Sigma Cat# G9145
N-acetylcysteine Sigma Cat# A9165
Nicotinamide Sigma Cat# N0636
B-27 supplement Thermo Fisher Cat# 17504044
Recombinant human TGF-b1 Peprotech Cat# 100-21
Recombinant human IFNg Peprotech Cat# 300-02
Growth factor reduced Matrigel Corning Cat# 356231
TrypLE Express Thermo Fisher Cat# 12604054
Trypsin-EDTA (0.25%) Thermo Fisher Cat# 25200056
CellTiterGlo 3D Promega Cat# G9683
Collagenase XI Sigma Cat# C7657
Dnase StemCell Technologies Cat# 07900
Y-27632 Selleck Chemicals Cat# S1049
ACK lysing biffer Thermo Fisher Cat# A1049201
Trypan blue solution, 0.4% Thermo Fisher Cat# 15250061
(Continued on next page)

e1 Cell 184, 6119–6137.e1–e12, December 9, 2021

 ll
Article OPEN ACCESS

Continued
REAGENT or RESOURCE SOURCE IDENTIFIER
Spectral DAPI Akoya Biosciences Cat# FP1490
Opal 520 reagent pack Akoya Biosciences Cat# FP1487001KT
Opal 540 reagent pack Akoya Biosciences Cat# FP1494001KT
Opal 570 reagent pack Akoya Biosciences Cat# FP1488001KT
Opal 620 reagent pack Akoya Biosciences Cat# FP1495001KT
Opal 650 reagent pack Akoya Biosciences Cat# FP1496001KT
Opal 690 reagent pack Akoya Biosciences Cat# FP1497001KT
Xylenes (histological) Fisher Scientific, X3P1GAL Cat# X3P-1GAL
BOND Epitope Retrieval Solution 1 Leica Biosystems Cat# AR9961
BOND Epitope Retrieval Solution 2 Leica Biosystems Cat# AR9640
Antibody diluent/block Akoya Biosciences Cat# ARD1001EA
1x Plus Automation Amplification Diluent Akoya Biosciences Cat# FP1609
ProLong Gold Antifade Mountant Fisher Scientific/Molecular Probes Cat# P36930
Gemcitabine Selleck Chemicals Cat# S1149
Paclitaxel Selleck Chemicals Cat# S1150
SN-38 Selleck Chemicals Cat# S4908
Trametinib Selleck Chemicals Cat# S2673
5-FU Selleck Chemicals Cat# S1209
Afatinib Selleck Chemicals Cat# S1011
AZD6738 (Ceralasertib) Selleck Chemicals Cat# S7693
Binimetinib Selleck Chemicals Cat# S7007
BVD-523 (Ulixertinib) Selleck Chemicals Cat# S7854
Dinaciclib Selleck Chemicals Cat# S2768
Everolimus Selleck Chemicals Cat# S1120
Gedatolisib Selleck Chemicals Cat# S2628
GSK126 Selleck Chemicals Cat# S7061
(+)-JQ1 Selleck Chemicals Cat# S7110
LY3023414 (Samotolisib) Selleck Chemicals Cat# S8322
MK-1775 (Adavosertib) Selleck Chemicals Cat# S1525
Navitoclax Selleck Chemicals Cat# S1001
Olaparib Selleck Chemicals Cat# S1060
Oxaliplatin Selleck Chemicals Cat# S1224
Palbociclib Selleck Chemicals Cat# S4482
Prexasertib Selleck Chemicals Cat# S6385
SHP099 Selleck Chemicals Cat# S8278
Tazemetostat Selleck Chemicals Cat# S7128
YKL-5-124 Selleck Chemicals Cat# S8863
2-mercaptoethanol Sigma Cat# M3148
Buffer RLT QIAGEN Cat# 79216
Buffer RLT Plus QIAGEN Cat# 1053393
Deoxynucleotide (dNTP) solution mix NewEngland BioLabs Cat# N0447L
Superase.In RNase Inhibitor Thermo Fisher Cat# AM2696
Maxima H minus reverse transcriptase Fisher Scientific Cat# EP0753
AMPure XP beads Beckman Coulter Cat# A63881
Guanidinium thiocyanate Thermo Fisher Cat# AM9422
N-Lauroylsarcosine sodium salt solution Sigma Cat# L7414
(Sarkosyl NL)
Exonuclease l New England BioLabs Cat# M0293S
Klenow Fragment New England BioLabs Cat# M0212L
Polycarbonate membrane filters 62x22 Fisher Scientific/Sterlitech Corporation Cat# NC1421644
(Continued on next page)

Cell 184, 6119–6137.e1–e12, December 9, 2021 e2

 ll
OPEN ACCESS Article

Continued
REAGENT or RESOURCE SOURCE IDENTIFIER
MACOSKO-2011-10 mRNA Capture Beads Fisher Scientific/ChemGenes Cat# NC0927472
ERCC RNA spike-in mix Thermo Fisher Cat# 4456740
Critical commercial assays
AllPrep DNA/RNA/miRNA Universal kit QIAGEN Cat# 80224
Nextera XT DNA Library Preparation Kit Illumina Cat# FC-131-1096
Nextseq 500/550 High output v2.5 kit (75 Illumina Cat# 20024906
cycles)
NovaSeq 6000 S2 kit (100 cycles) Illumina Cat# 20012862
TruSeq Stranded mRNA Library Prep kit Illumina Cat# 20020595
Kapa HiFi HotStart ReadyMix Kapa Biosystems Cat# KK2602
KAPA HyperPrep kit (PCR-free) Kapa Biosystems Cat# KK8505
High Sensitivity D5000 ScreenTape Agilent Cat# 5067-5592
Qubit dsDNA High-Sensitivity kit Thermo Fisher Cat# Q32854
Quant-iT Ribogreen RNA Assay kit Thermo Fisher Cat# R11490
Quant-iT PicoGreen dsDNA Assay kit Thermo Fisher Cat# P11496
Deposited data
Bulk and single-cell transcriptomic data This study https://singlecell.broadinstitute.org/
from PDAC patient samples and organoid single_cell/study/SCP1644 dbGaP:
models phs002712.v1.p1
Primary PDAC genomic and transcriptomic Cancer Genome Atlas Research https://portal.gdc.cancer.gov/projects/
data (TCGA, Pancreatic Ductal Network, 2017 TCGA-PAAD
Adenocarcinoma)
Metastatic PDAC genomic and Aguirre et al., 2018 dbGaP: phs001652.v1.p1
transcriptomic data (Panc-Seq)
CCLE transcriptomic data Ghandi et al., 2019 https://portals.broadinstitute.org/ccle
TCGA transcriptomic data (other Vivian et al., 2017 https://toil.xenahubs.net
malignancies)
Experimental models: Cell lines
Human PDAC organoids This study N/A
CFPAC-1 ATCC Cat# CRL-1918; RRID:CVCL_1119
L Wnt-3A cells for Wnt-3A conditioned ATCC Cat# CRL-2647; RRID:CVCL_0635
medium
Cultrex 293T cells for R-spondin1 Trevigen Cat# 3710-001-K; RRID:CVCL_RU08
conditioned medium
Oligonucleotides
Seq-Well ISPCR: AAG CAG TGG TAT CAA Integrated DNA Technologies N/A
CGC AGA GT
Custom Read 1 Primer: GCC TGT CCG Integrated DNA Technologies N/A
CGG AAG CAG TGG TAT CAA CGC AGA
GTA C
Seq-Well 5¢ TSO: AAG CAG TGG TAT CAA Integrated DNA Technologies N/A
CGC AGA GTG AAT rGrGrG
Seq-Well Custom P5-SMART PCR hybrid Integrated DNA Technologies N/A
oligo: AAT GAT ACG GCG ACC ACC GAG
ATC TAC ACG CCT GTC CGC GGA AGC
AGT GGT ATC AAC GCA GAG TAC
Seq-Well dN-SMRT oligo: AAG CAG TGG Integrated DNA Technologies N/A
TAT CAA CGC AGA GTG ANN NGG NNN B
Software and algorithms
R project for statistical computing v3.5.1 R Core Team https://www.r-project.org
R package – Seurat v2.3.4 GitHub https://github.com/satijalab/seurat
(Continued on next page)

e3 Cell 184, 6119–6137.e1–e12, December 9, 2021

 ll
Article OPEN ACCESS

Continued
REAGENT or RESOURCE SOURCE IDENTIFIER
R package – Circlize v0.4.8 CRAN https://CRAN.R-project.org/ package =
circlize
R package – infercnv v0.99.4 GitHub https://github.com/broadinstitute/inferCNV
R package – data.table v1.12.0 GitHub https://github.com/Rdatatable/data.table
R package – ggplot2 v3.2.1 CRAN https://CRAN.R-project.org/ package =
ggplot2
R package – ComplexHeatmap v2.7.3 Bioconductor https://bioconductor.org/packages/
ComplexHeatmap/
R package – dplyr v1.0.7 CRAN https://cran.r-project.org/web/
packages/dplyr/
STAR GitHub https://github.com/alexdobin/STAR
Cumulus Li et al., 2020 https://cumulus.readthedocs.io/
Broad Picard pipeline v1.90 GitHub https://broadinstitute.github.io/picard/
Genome Analysis Toolkit (GATK) Broad Institute https://gatk.broadinstitute.org/hc/en-us
v1.6-5-g557da77 and v.4.1.6.0
Python Programming Language v3.7.4 Python https://www.python.org
Other
Leica BOND RX Research Stainer Leica Biosystems https://www.leicabiosystems.com/
ihc-ish-fish/fully-automated-ihc-ish-
instruments/bond-rx/
Vectra 3.0 Automated Quantitative Imaging PerkinElmer/Akoya Biosciences https://www.akoyabio.com/phenoptics/
System mantra-vectra-instruments/vectra-3-0/

RESOURCE AVAILABILITY

Lead contact
Further information and requests for resources and reagents should be sent to and will be fulfilled by Dr. Alex K. Shalek (shalek@
mit.edu).

Materials availability
Organoid models generated in this study are available upon request with a materials transfer agreement.

Data and code availability

De-identified single-cell RNA-seq data are publicly available for download and visualization via the Single Cell Portal: https://
singlecell.broadinstitute.org/single_cell/study/SCP1644. Genomic and transcriptomic data will be available at the NCBI Database
of Genotypes and Phenotypes (dbGaP). This paper analyzes existing, publicly available data. The links and accession numbers
for these datasets are listed in the key resources table.
Code is available from the lead contact upon request.
Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

EXPERIMENTAL MODEL AND SUBJECT DETAILS

Human specimens
Eligible participants were recruited from outpatient clinics and inpatient units at the Dana-Farber Cancer Institute and the Brigham
and Women’s Hospital. Investigators obtained written, informed consent from patients at least 18 years old with pancreatic cancer
for Dana-Farber/Harvard Cancer Center Institutional Review Board (IRB)-approved protocols 11-104, 17-000, 03-189, and/or 14-408
for tissue collection, molecular analysis, and organoid generation. ScRNA-seq samples were collected from 23 patients between
October 2018 and December 2020, both male (n = 13) and female (n = 10). Organoid samples for bulk genomic and transcriptomic
analyses were collected between May 2015 and January 2018. Core needle biopsy specimens were collected and the first core was
sent for pathologic analysis. One or more additional cores were then allocated for scRNA-seq and/or organoid generation. Clinical
features of our patient cohort are included in Table S1.

Cell 184, 6119–6137.e1–e12, December 9, 2021 e4

 ll
OPEN ACCESS Article

Human PDAC patient-derived organoid models

Tissue samples were minced into small portions using a scalpel and then digested at 37 C for 15 min using digest medium that con-
sisted of human complete organoid medium (see below), 1 mg/mL collagenase XI (Sigma Aldrich), 10 mg/mL DNase (Stem Cell Tech-
nologies), and 10 mM Y27632 (Selleck) (Tiriac et al., 2018). After dissociation, a portion of the cells from the fresh tumor specimen were
allocated for scRNA-seq, and the remainder were initiated and maintained as patient-derived organoid cultures as previously
described (Boj et al., 2015; Tiriac et al., 2018). In brief, digested cells were seeded in 3-dimensional (3D) Growth-factor Reduced Ma-
trigel (Corning), fed with human complete organoid medium containing Advanced DMEM/F12 (GIBCO), 10 mM HEPES (GIBCO), 1x
GlutaMAX (GIBCO), 500 nM A83-01 (Tocris), 50 ng/mL mEGF (Peprotech), 100 ng/mL mNoggin (Peprotech), 100 ng/mL hFGF10 (Pe-
protech), 10 nM hGastrin I (Sigma), 1.25 mM N-acetylcysteine (Sigma), 10 mM Nicotinamide (Sigma), 1x B27 supplement (GIBCO),
RSPONDIN-1 conditioned media 10% final, WNT3A conditioned media 50% final, 100 U/mL penicillin/streptomycin (GIBCO), and 1x
Primocin (Invivogen) (Table S5), and maintained at 37 C in 5% CO2. 10 mM Y27632 (Selleck) was included in the culture medium of
newly initiated samples until the first media exchange. For propagation, organoids were dissociated with TrypLE Express (GIBCO)
before re-seeding into fresh Matrigel and culture medium.
After initial processing of fresh tissue specimens, we monitored samples closely for organoid growth. We did not passage organoids
at set time intervals, as there was significant variability in the time needed to establish relatively robust growth of organoids (Figure 3D).
Instead, we maintained early passage organoids until they reached relative confluence, and then passaged them at low split ratios
(1:1, 1:1.5, or 1:2 dilutions) in complete organoid medium to promote continued growth. In one case, PANFR0489R, cells persisted
as individuals and small organoids after initiation in complete organoid medium, but did not grow and expand cell numbers signifi-
cantly. Approximately 15 weeks after initiation, we switched a portion of the surviving cells to organoid medium without A83-01 or
mNoggin, and observed renewed growth of organoids under these media conditions but not of those that remained in complete or-
ganoid medium. Consequently, we expanded this sample in media without A83-01 or mNoggin, including performing early passage
scRNA-seq. After several additional passages, once the organoids were robustly growing, we were able to transition this model back
to complete organoid medium with no apparent change in growth rate, morphology, or transcriptional state. All other serially sampled
organoids were maintained and assessed in complete medium except as indicated when specific media alterations or experimental
perturbations were performed. The identify of organoid models was authenticated by comparison of their inferred CNV profiles with
targeted genomic sequencing and CNV profiles of matched patient tissue and with inferred CNV profiles from patient tissue and earlier
passage models in the case of samples serially assessed with scRNA-seq. The identify of cell line models was authenticated by short
tandem repeat (STR) analysis. Cell line and organoid cultures were routinely tested for mycoplasma contamination.

METHOD DETAILS

Sample preparation for single-cell RNA-sequencing of clinical samples, organoids, and cell lines
Patient specimens were dissociated for paired scRNA-seq and organoid generation as described above. In our initial process opti-
mization for fresh patient specimens, we found that dissociation times below 30 min, while not always completely digesting all biopsy
material and potentially affecting the representation of difficult to dissociate cell types (e.g., fibroblasts), resulted in greater cell
viability and improved RNA quality downstream. After tissue dissociation, cells were washed, treated with ACK lysing buffer (Thermo
Fisher) to lyse red blood cells, washed again, and counted using a hemocytometer with 0.4% Trypan blue (Thermo Fisher) added at
1:1 dilution for viability assessment. We allowed residual tissue chunks to settle before selecting a predominance of single cells for
counting and Seq-Well processing. We allocated between 10,000 and 15,000 viable cells per Seq-Well array, and where possible we
prepared two arrays per sample. Most samples were processed and loaded onto Seq-Well arrays within 2-3 h of biopsy acquisition.
Remaining cells and tissue chunks were allocated for patient-matched organoid generation.
For scRNA-seq of organoid samples and cell lines (CFPAC1), we passaged models and allowed them to grow for 6 days before
dissociating to single cells (organoids – TrypLE Express, Thermo Fisher; cell lines – 0.25% Trypsin-EDTA, Thermo Fisher), counting,
and allocating 15,000 viable cells for Seq-Well. By standardizing the collection of organoid scRNA-seq samples at 6 days after
passaging, we tried to minimize bias arising from cell cycle differences in samples at different degrees of confluence.

Assessing organoid and cell line transcriptional states under different matrix and media conditions
For adaptation of patient-derived organoids onto 2-dimensional (2D) culture surfaces as patient-derived cell lines, tissue culture
plates were pre-coated with 100 mg/mL Matrigel dissolved in basal media for 2 h at 37 C before washing with PBS. Established
organoid models were dissociated and seeded onto these Matrigel-coated culture wells in complete organoid media. In parallel,
a portion of these passage-matched organoid cells were re-seeded into Matrigel droplets as above. Cells were cultured in both ma-
trix conditions in complete organoid media until they were confluent, approximately 2-3 weeks. Cells were collected and lysed using
Trizol before snap freezing. RNA was isolated and purified as described below (‘‘Bulk RNA-sequencing of organoids’’ section) using
chloroform extraction, aqueous phase isolation, and processing using the QIAGEN AllPrep DNA/RNA/miRNA Universal kit before
being submitted for sequencing.
For scRNA-seq assessment of organoid cell states when cultured under different media conditions, established organoid models
were passaged as above by dissociating and reseeding into Matrigel droplets. A portion of the cells were cultured with complete
organoid media (‘‘Complete media’’), while a distinct portion of passage-matched cells were cultured in ‘‘Minimal’’ media, which

e5 Cell 184, 6119–6137.e1–e12, December 9, 2021

 ll
Article OPEN ACCESS

consisted of Advanced DMEM/F12 (Thermo Fisher), 10 mM HEPES (Thermo Fisher), 1x GlutaMAX (Thermo Fisher), 100 U/mL peni-
cillin/streptomycin (Thermo Fisher), and 1x Primocin (Invivogen) (Table S5). Cells were cultured for 6 days before being collected,
dissociated, and aliquoted for scRNA-seq. Images were taken with an Olympus XM10 camera mounted to an Olympus CKX41 mi-
croscope 1 day after seeding and again after 11 days in culture to assess organoid growth in both conditions. The portion of cells
cultured in minimal media were maintained in the same conditions for a longer duration and harvested again for scRNA-seq at
27 days and 59 days after the initial introduction of minimal media. To mirror the standard scRNA-seq workflow, the cells harvested
at the 27- and 59-day time points were collected 6 days after passaging.
In addition to the minimal media experiment, organoid cells were also cultured in standard cell line media (‘‘RP10’’), which contains
RPMI-1640 (Thermo Fisher) and 100 U/mL penicillin/streptomycin (Thermo Fisher) with 10% fetal bovine serum (Sigma), or in
reduced organoid media ‘‘OWRNA,’’ which consists of Advanced DMEM/F12 (Thermo Fisher), 10 mM HEPES (Thermo Fisher), 1x
GlutaMAX (Thermo Fisher), 50 ng/mL mEGF (Peprotech), 100 ng/mL hFGF10 (Peprotech), 10 nM hGastrin I (Sigma), 1.25 mM N-ace-
tylcysteine (Sigma), 10 mM Nicotinamide (Sigma), 1x B27 supplement (Thermo Fisher), 100 U/mL penicillin/streptomycin (Thermo
Fisher), and 1x Primocin (Invivogen) (i.e., complete organoid medium with removal of WNT3A, RSPONDIN-1, NOGGIN, and
A-8301; Table S5). Furthermore, OWRNA reduced organoid medium served as the baseline control medium when assessing the ef-
fect of specific factors (IFNGg and TGF-b1) from the TME on malignant cell states. Cells were cultured for 6 days before being
collected, dissociated, and aliquoted for scRNA-seq in each of the following conditions: RP10, OWRNA, OWRNA with 50 ng/mL
IFNGg (Peprotech), and OWRNA with 5 ng/mL TGF-b1 (Peprotech) (Table S5). The cells cultured under the IFNGg and TGF-b1 con-
ditions were maintained in culture and harvested again for scRNA-seq 38 days after being introduced to these new media conditions.
For these longer duration time points, cells were again passaged 6 days before collecting for scRNA-seq.
For scRNA-seq assessment of transcriptional states of the established pancreatic cancer cell line CFPAC1 under different media
conditions, CFPAC1 cells were cultured in parallel in either standard cell line medium RP10 or complete organoid medium for 6 days
before being collected, dissociated, and aliquoted for scRNA-seq.

Compound sensitivity testing in cell lines and organoid models

Organoids were propagated in their respective media formulations for the indicated times (e.g., 3 weeks, 6 weeks) before performing
drug sensitivity testing. After dissociation, cells were seeded into ultra-low attachment 384-well plates (Corning) at 1000 viable cells
per well with 20 mL of their respective medium containing 10% Matrigel by volume. After 24 h, compounds were added to each well
over 12-point dose curves along with DMSO controls using a Tecan D300e digital dispenser. A parallel untreated plate was harvested
at this 24 h time point for growth rate correction. Drug- or DMSO-treated cells were cultured for 5 days before assessing for viability.
Cell viability was measured by adding 20 uL of CellTiter-Glo 3D (Promega) to each well, incubating for 1 h at room temperature on a
shaker, and measuring luminescence using an EnVision plate reader (PerkinElmer). Given plate well randomization, raw luminescence
data were deconvoluted with an in-house Python script (Python v3.7.4). Each condition was tested in triplicate, and each dose point
was normalized to DMSO controls to estimate relative viability. Growth rate correction was performed as previously described, with
growth-rate adjusted dose response curves fit to a 3-parameter sigmoidal curve (Hafner et al., 2016). Each experiment was performed
using cells cultured in paired basal (OWRNA+TGF-b1 or Minimal) and classical (OWRNA) media conditions, and independent exper-
iments were performed as summarized in Figure S7A and Table S7. To quantify the relative sensitivity of basal versus classical models
to a given compound, the difference in growth-rate adjusted areas over the curve (AOC) for organoids cultured in paired media con-
ditions was calculated for each compound within each experiment, as illustrated in Figure 6G and Figure S7A.
CFPAC1 cells were cultured in standard cell line medium RP10 or in complete organoid medium for 2 or 5 weeks before seeding for
drug testing. Cells were dissociated and seeded into 384-well tissue culture plates (Greiner Bio-One) at 400 viable cells per well with
20 mL of their respective medium. Addition of compounds and sample harvesting were as above, except cell viability was measured
by adding 20 uL of CellTiter-Glo (Promega) to each well and incubating for 15 min at room temperature on a shaker before measuring
luminescence. Data analysis and dose response curve fitting were performed as described above.

Single-cell RNA-seq (scRNA-seq) data library generation, sequencing, and alignment

ScRNA-seq processing followed the Seq-Well protocol, uniquely compatible with low-input samples (Gierahn et al., 2017; Hughes
et al., 2020). Briefly, arrays were preloaded with RNA capture beads (Fisher Scientific/ChemGenes) and stored in quenching buffer
until used. Prior to cell loading, arrays were resuspended in 5 mL RPMI-1640 medium (Thermo Fisher) with 10% fetal bovine serum
(Sigma), hereafter referred to as RP10. After dissociation, single-cell suspensions were manually counted and diluted to 15,000 cells
per 200 mL of RP10 when cell numbers allowed. Excess RP10 was aspirated from the array and cells were loaded onto arrays. Excess
cells were washed off with PBS (4x5 mL), briefly left in RPMI (5 mL) and cell+bead pairs were sealed for 40 min at 37 C using a poly-
carbonate membrane (Fisher Scientific). Arrays were rocked in lysis buffer for 20 min and RNA was hybridized onto the beads for
40 min. Beads were removed and reverse transcription was performed overnight using Maxima H Minus Reverse Transcriptase
(Fisher Scientific). Prior to sequencing, the beads underwent an exonuclease treatment (New England BioLabs) and second strand
synthesis en masse followed by whole transcriptome amplification (WTA, Kapa Biosystems) in 1,500 bead reactions (50 mL). cDNA
was isolated using Agencourt AMPure XP beads (Beckman Coulter) at 0.6X SPRI (solid-phase reversible immobilization) followed by
a 1X SPRI and quantified using a Qubit dsDNA High Sensitivity assay kit (Thermo Fisher). Library preparation was performed using
Nextera XT DNA tagmentation (Illumina FC-131-1096) and N700 and N500 indices specific to a given sample. Tagmented and

Cell 184, 6119–6137.e1–e12, December 9, 2021 e6

 ll
OPEN ACCESS Article

amplified sequences were purified with a 0.6X SPRI, and cDNA was quantified (Qubit dsDNA High sensitivity assay kit, Thermo
Fisher) and the base pair distribution measured (High sensitivity D5000 screen tape, Agilent). cDNA was loaded onto either an Illumina
Nextseq (75 Cycle NextSeq 500/550 v2.5 kit) or Novaseq (100 Cycle NovaSeq 6000 S2 kit) at 2.4 pM. Regardless of platform, the
paired end read structure was 21 bases (cell barcode and UMI) by 50 bases (transcriptomic information) with an 8 base pair (bp)
custom read one primer. The demultiplex and alignment protocol was followed as previously described (Macosko et al., 2015). While
Novaseq data were directly output as FASTQs, Nextseq BCL files were converted to FASTQs using bcl2fastq2. The resultant Next-
seq and Novaseq FASTQs were demultiplexed by sample based on Nextera N700 and N500 indices. Reads were then aligned to the
hg19 transcriptome using the cumulus/dropseq_tools pipeline on Terra maintained by the Broad Institute using standard settings (Li
et al., 2020).

Bulk RNA- and DNA-sequencing of organoids

RNA was obtained for bulk RNA-sequencing from established organoids using one of two approaches. Dissociated organoids were
resuspended into cold Matrigel, added as droplets to tissue culture plates (Greiner BioOne), and allowed to polymerize for 30 min
before addition of media. Organoids were grown for 14-21 days (until confluent) under these conditions with regular media changes.
At the time of harvest, cells were washed with cold phosphate buffered saline (PBS) at 4 C, then lysed with Trizol (Invitrogen) before
snap-freezing. To isolate RNA, we performed chloroform extraction with isolation of the aqueous phase before processing RNA as
per protocols outlined in the QIAGEN AllPrep DNA/RNA/miRNA Universal kit.
In the second approach, used to obtain both RNA and DNA, dissociated organoids were resuspended in a solution of 10% Matrigel
in complete organoid media (volume/volume) and cultured in ultra-low-attachment culture flasks (Corning). Organoids were grown
for 14-21 days (until confluent) before pelleting, washing with cold PBS at 4 C until most Matrigel was dissipated, and then snap
frozen. Cell pellets were homogenized using buffer RLT Plus (QIAGEN) and a Precellys homogenizer. Samples were then processed
for both DNA extraction and RNA isolation as per the QIAGEN AllPrep DNA/RNA/miRNA Universal kit. Purified RNA and DNA were
then submitted for sequencing by the Broad Institute Genomics Platform.
For bulk RNA-sequencing, total RNA was quantified using the Quant-iT RiboGreen RNA Assay Kit (Thermo Fisher) and normalized to
5 ng/mL. Following plating, 2 mL of a 1:1000 dilution of ERCC RNA controls (Thermo Fisher) were spiked into each sample. An aliquot of
200 ng for each sample was transferred into library preparation which uses an automated variant of the Illumina TruSeq Stranded mRNA
Sample Preparation Kit. This method preserves strand orientation of the RNA transcript, and uses oligo dT beads to select mRNA from
the total RNA sample followed by heat fragmentation and cDNA synthesis from the RNA template. The resultant 400 bp cDNA then goes
through dual-indexed library preparation: ‘A’ base addition, adaptor ligation using P7 adapters, and PCR enrichment using P5
adapters. After enrichment, the libraries were quantified using Quant-iT PicoGreen (1:200 dilution; Thermo Fisher). After normalizing
samples to 5 ng/mL, the set was pooled and quantified using the KAPA Library Quantification Kit for Illumina Sequencing Platforms.
The entire process was performed in 96-well format and all pipetting was done by either Agilent Bravo or Hamilton Starlet.
Pooled libraries were normalized to 2 nM and denatured using 0.1 N NaOH prior to sequencing. Flowcell cluster amplification and
sequencing were performed according to the manufacturer’s protocols using the NovaSeq 6000. Each run was a 101 bp paired-end
with an eight-base index barcode read. Data were analyzed using the Broad Picard pipeline which includes de-multiplexing and data
aggregation (https://broadinstitute.github.io/picard/). FASTQ files were then processed as described below (see Bulk RNA-
sequencing analysis).
For whole genome sequencing, 350 ng of genomic DNA was fragmented using a Covaris focused-ultrasonicator targeting 385bp
fragments followed by size selection using SPRI cleanup. Library preparation was performed using a KAPA HyperPrep without ampli-
fication kit (KAPA Biosystems) with palindromic forked adapters with unique 8-base index sequences embedded within the adaptor
(Roche). Libraries were then quantified using quantitative PCR (kit purchased from KAPA Biosystems) with probes specific to the
adaptor ends on an Agilent Bravo liquid handling platform. Libraries were normalized to 2.2 nM, pooled into 24-plexes, combined
with NovaSeq Cluster Amp Reagents DPX1, DPX2, and DPX3, and loaded into single lanes of a NovaSeq 6000 S4 flowcell using
a Hamilton Starlet Liquid Handling system. Cluster amplification and sequencing occurred utilizing sequencing-by-synthesis kits
to produce 151bp paired-end reads. Output from Illumina software was processed by the Broad Picard pipeline (https://
broadinstitute.github.io/picard/) to yield BAM files containing demultiplexed, aggregated aligned reads. BAM files were then pro-
cessed as described below (see Mutation and CNV identification from bulk DNA-sequencing).

Multiplex immunofluorescence imaging

A multi-marker panel was developed to characterize malignant cell subtype in formalin-fixed paraffin-embedded (FFPE) 4 mm tissue
sections using multiplex immunofluorescence (mIF). The panel comprises markers associated with either a basal (Keratin-17: Thermo
Fisher MA513539 and S100A2: Abcam 109494) or classical (CLDN18.2: Abcam 241330, GATA6: Cell Signaling Technology 5851 and
TFF1: Abcam 92377) subtype. Additionally, DAPI (Akoya Biosciences FP1490) was included for identification of nuclei and pan-cy-
tokeratin (AE1/AE3: Dako M3515; C11: Cell Signaling Technology 4545) for identification of epithelial cells. Secondary Opal Polymer
HRP anti-mouse and anti-rabbit antibody (Dako ARH1001EA), Tyramide signal amplification, and Opal fluorophores (Akoya Biosci-
ences) were used to detect primary antibodies (Keratin-17, Opal 520; S100A2, Opal 650; GATA6, Opal 540; CLDN18.2, Opal 570;
TFF1, Opal 690; panCK, Opal 620).

e7 Cell 184, 6119–6137.e1–e12, December 9, 2021

 ll
Article OPEN ACCESS

These specific mIF markers were chosen for several reasons. KRT17, S100A2, and TFF1 are included in the original basal and
classical RNA gene signatures (Moffitt et al., 2015). GATA6 is also a classical marker in an extended RNA gene panel and has
been reported to be a potential driver of the classical phenotype (Brunton et al., 2020; Moffitt et al., 2015; O’Kane et al., 2020).
Single markers of S100A2 and GATA6 have also been used extensively in imaging experiments in past literature to mark cells in
the basal or classical state, respectively (Aung et al., 2018; Chan-Seng-Yue et al., 2020; Miyabayashi et al., 2020; O’Kane et al.,
2020; Somerville et al., 2018). CLDN18.2 has also been associated with classical phenotypes, and antibody-based therapies
targeting CLDN18.2 have been developed and tested in PDAC (Wöll et al., 2014). Furthermore, the markers chosen for the
mIF subtyping panel agreed with those selected as optimal markers to differentiate basal-like versus classical by an international
panel of experts at a workshop on pancreatic cancer subtyping held at Memorial Sloan Kettering in 2019. Here, we used multiple
markers for each state to provide greater confidence in cell state identification and to assess marker heterogeneity across our
mIF cohort.
Prior to use in multiplex staining, primary antibodies were first optimized via immunohistochemistry on control tissue to confirm
contextual specificity. Monoplex immunofluorescence and iterative multiplex fluorescent staining were then used to optimize stain-
ing order, antibody-fluorophore assignments and fluorophore concentrations. Multiplex staining was performed using a Leica BOND
RX Research Stainer (Leica Biosystems, Buffalo, IL) with sequential cycles of antigen retrieval, protein blocking, primary antibody
incubation, secondary antibody incubation, and fluorescent labeling. Overview images of stained slides were acquired at 10X magni-
fication using a Vectra 3.0 Automated Quantitative Imaging System (Perkin Elmer, Waltham, MA) and regions of interest (ROIs) were
selected for multispectral image acquisition at 20X. After unmixing using a spectral library of single-color references, each image was
inspected to ensure uniform staining quality and adequate tumor representation.

QUANTIFICATION AND STATISTICAL ANALYSIS

Mutation and CNV identification from bulk DNA-sequencing

For targeted DNA-sequencing of clinical samples, next-generation sequencing using a custom-designed hybrid capture library prep-
aration was performed on an Illumina HiSeq 2500 with 2x100 paired-end reads, as previously described (Garcia et al., 2017; Sholl
et al., 2016). Sequence reads were aligned to reference sequence b37 edition from the Human Genome Reference Consortium using
bwa, and further processed using Picard (version 1.90, http://broadinstitute.github.io/picard/) to remove duplicates and Genome
Analysis Toolkit (GATK, version 1.6-5-g557da77) to perform localized realignment around indel sites. Single nucleotide variants
were called using MuTect v1.1.45, insertions and deletions were called using GATK Indelocator. Copy number variants (CNV) and
structural variants were called using the internally-developed algorithms RobustCNV and BreaKmer followed by manual review
(Abo et al., 2015). RobustCNV calculates copy ratios by performing a robust linear regression against a panel of normal samples.
The data were segmented using circular binary segmentation, and event identification was performed based on the observed vari-
ance of the data points (Bi et al., 2017).
We computed the cytoband-level copy number calls and weighted (by length) average segment means across the covered regions
of each cytoband using ASCETS (Spurr et al., 2021). Briefly, cytobands were considered amplified/deleted if more than 70% of the
covered regions had a log2 copy ratio of greater than 0.2/less than
0.2, and were considered neutral if more than 70% of the
covered regions had a log2 copy ratio between
0.2 and 0.2.
For comparisons of driver mutation frequencies across patient tumors, cell lines, and organoid models, mutation and CNV calls for
KRAS, TP53, CDKN2A, and SMAD4 were either compiled from prior publications for patient samples and CCLE cell lines (Aguirre
et al., 2018; Cancer Genome Atlas Research Network, 2017; Ghandi et al., 2019) or generated from whole genome sequencing of
organoid models. Organoid cohort variants were called from tumor/germline pairs using GATK (v.4.1.6.0, Paired tumor-control
mode). Potential germline variants were additionally filtered using gnomAD (v2.1). Significance of short-nucleotide variants (SNVs)
was determined using MuSiC2 (v0.2, q-value < 0.1). CNVs were initially called using GATK (v.4.1.6.0) followed by analysis with GISTIC
(v2.0.23). Genomic alterations in KRAS, TP53, CDKN2A, and SMAD4 were binarized by counting a gene in the given sample as
altered if it contained at least one of the following alterations: missense mutation, nonsense mutation, splice site alteration, frameshift
insertion or deletion, in-frame insertion or deletion, a high amplification, or a homozygous deletion. Statistical significance of alter-
ation occurrence per gene across sample cohorts was determined using a Fisher’s exact test with multiple test correction using
the Benjamini-Hochberg procedure (R v4.0.3).

Bulk RNA-sequencing analysis

FASTQs for bulk RNA expression profiles were downloaded from the relevant repository (TCGA, https://toil.xenahubs.net; PDAC Cell
lines, https://portals.broadinstitute.org/ccle), available in-house (Panc-Seq, metastatic PDAC), or generated for this study (organoid
cohort) (Aguirre et al., 2018; (Weinstein et al., 2013) Cancer Genome Atlas Research Network, 2017; Ghandi et al., 2019; Vivian et al.,
2017). All were processed using the same pipeline. Briefly, each sample’s sequences were marked for duplicates and then mapped
to hg38 using STAR. After running QC checks using RNaseqQC, gene-level count matrices were generated using RSEM. Instructions
to run the pipeline are given in the Broad CCLE github repository https://github.com/broadinstitute/ccle_processing. Length-normal-
ized values (TPM) were then transformed according to log2(TPM+1) for downstream analysis. The entire dataset was scaled and

Cell 184, 6119–6137.e1–e12, December 9, 2021 e8

 ll
OPEN ACCESS Article

centered to allow relative comparisons across sample types (e.g., tumors, organoids, and cell lines). Signature scores were
computed as below (e.g., basal and classical; see Generation of expression signatures/scores below) (Puram et al., 2017).

Single-cell data quality pre-processing and initial cell type discovery

All single-cell data analysis was performed using the R language for Statistical Computing (v3.5.1). Each biopsy sample’s digital gene
expression (DGE) matrix (cells x genes) was trimmed to exclude low quality cells (< 400 genes detected; < 1,000 UMIs; > 50% mito-
chondrial reads) before being merged together (preserving all unique genes) to create the larger biopsy dataset. The merged dataset
was further trimmed to remove cells with > 8,000 genes which represent outliers and likely doublet cells. We also removed genes that
were not detected in at least 50 cells. The same metrics were applied to the organoid single-cell cohort (see below). On a per cell
basis, UMI count data was divided by total transcripts captured and multiplied by a scaling factor of 10,000. These normalized values
were then natural log transformed for downstream analysis (i.e., log-normalized cell x gene matrix). Initial exploration of the data was
performed using the R package Seurat (v2.3.4) and followed two steps: 1) SNN-guided quality assessment and 2) cell type compo-
sition determination. In step 1, we intentionally left cells in the DGE matrix of dubious quality (e.g., % mitochondrial reads > 25% but <
50%), performed PCA over the variable genes (n = 1,070 genes), and input the first 50 PCs (determined by Jackstraw analysis im-
plemented through Seurat) to build an SNN graph and cluster the cells (res = 1; k.param = 40). The inclusion of poor-quality cells
essentially acts as a variance ‘‘sink’’ for other poor-quality cells and they cluster together based on their shared patterns in qual-
ity-associated gene expression. This method helped to nominate additional low quality (e.g., defined exclusively by mitochondrial
genes) or likely doublet cells (e.g., clusters defined by co-expression of divergent lineage markers) which were removed from the
dataset (n = 1,678 cells). This led to an overall high-quality dataset of single-cells with a low overall fraction of mitochondrial reads
(median = 0.09) for downstream analysis (Figure S1D).
Using the trimmed dataset, we proceeded to step 2 using a very similar workflow as above but with slightly altered input conditions
for defining clusters. Here we used PCs 1-45 and their associated statistically significant genes for building the SNN graph and deter-
mining cluster membership (resolution = 1.2; k.param = 40). This identified the 36 clusters shown (visualized using t-SNE; perplexity,
40; iterations, 2,500) in Figure S1E. The expression of known markers was used to collapse clusters containing shared lineage infor-
mation. For example, clusters 1, 2, and 4 all express high levels of macrophage markers—CD14, FCGR3A (CD16), CD68—and were
accordingly collapsed for this first pass analysis (Figure S1E,I). To aid our cell type identification, we performed an ROC test imple-
mented in Seurat to confirm the specificity (power > 0.6) of the top marker genes used to discern the cell types. Combined with in-
ferred CNV information (see below), this analysis confirmed the presence of 11 broad non-malignant cell types in our biopsy dataset
(Table S2). Variation in the SNN graph parameters above did not strongly affect cell type identification.

Single-cell CNV identification

To confirm the identity of the putative malignant clusters identified in Figure S1F, we estimated single-cell CNVs as previously
described by computing the average expression in a sliding window of 100 genes within each chromosome after sorting the detected
genes by their chromosomal coordinates (Patel et al., 2014; Tirosh et al., 2016b). We used all T/NK, Fib, Hep, and Endo cells identified
above as reference normal populations for this analysis. Complete information on the inferCNV workflow used for this analysis can be
found here https://github.com/broadinstitute/inferCNV/wiki. To compare with bulk targeted DNA-sequencing, we collapsed individ-
ual probes to cytoband-level information (weighted average of log2 ratios across each cytoband, see above) within each sample.
ScRNA-seq-inferred CNVs showed high concordance across samples with the bulk measurements and suggests that, at least by
this metric, we are likely sampling the same dominant clones within sequential but distinct cores from each needle biopsy procedure
(Figure S1G). For plotting CNV profiles in putative malignant versus normal cells (Figure S1H), we computed the average CNV signal
for the top 5% of altered cells in each biopsy and correlated all cells in that biopsy to the averaged profile as has been previously
described (Tirosh et al., 2016a). Relation of this correlation coefficient to the CNV score (mean square deviation from diploidy) in
the single cells from each biopsy shows consistent separation of malignant from non-malignant cells, and, combined with member-
ship in patient-specific SNN clusters, substantiates the identification of malignant cells in our dataset. One patient sample,
PANFR0604, did not contain any malignant cells within the core biopsy used for scRNA-seq analysis.

Subclonal analysis with single-cell inferred CNVs

The inferCNV workflow can be used to call subclonal genomic variation with high sensitivity and is described at https://github.
com/broadinstitute/inferCNV/wiki (Fan et al., 2018; Patel et al., 2014; Tirosh et al., 2016b). Briefly, we used a six-state Hidden
Markov Model (i6-HMM) to predict relative copy number status (complete loss to > 3x gain) across putative altered regions in
each cell. A Bayesian latent mixture model then evaluated the posterior probability that a given copy number alteration is a
true positive. We set a relatively stringent cutoff for this step (BayesMaxPNormal = 0.2) to only include high probability alterations
for downstream clustering. The results of this filtered i6-HMM output were then used to cluster the single cells using Ward’s
method. We used inferCNV’s ‘‘random trees’’ method to test for statistical significance (p < 0.05, 100 random permutations
for each split) at each tree bifurcation and only retained subclusters that had statistical evidence underlying the presumed het-
erogeneity. To track subclonal heterogeneity between biopsy and matched organoid cells in Figure 3E and Figure S5E-K, the
above workflow was implemented within each biopsy and the relevant matched organoid samples, essentially treating all cells
as the same ‘‘tumor’’ and allowing the CNVs to determine cell sorting agnostic to sample-of-origin. The results of the HMM output

e9 Cell 184, 6119–6137.e1–e12, December 9, 2021

 ll
Article OPEN ACCESS

can be used to infer gene-level information based on which genes are in the affected window. This (like the rest of the HMM work-
flow) is computed over groups of cells (e.g., samples or sub-clones) and used to map KRAS and other alterations to samples or
sub-clones (Figure 3E, Figure S5E-K).

Subclustering of malignant and non-malignant cells

Detailed phenotyping required splitting the dataset into malignant and non-malignant fractions. After subsetting to only the malignant
cells, we re-scaled the data and ran PCA including the first 35 PCs for SNN clustering and t-SNE visualization. This PCA was used to
determine the PanNET identity for PANFR0580 (Figure S2A). After removing PANFR0580, we repeated the steps above and used this
new PCA for the remainder of PDAC malignant cell analysis. Subsequent phenotyping for malignant cells is discussed below (Gen-
eration of expression signatures/scores). A similar approach was used for calling the non-malignant subsets in Figure 5A. To deter-
mine the specific phenotypes within T/NK, macrophage, and mesenchymal populations, we separately subclustered these groups
using PCs 1-20 and a resolution of 0.6 in each case. Of note, subclustering within the macrophages revealed a distinct cluster of cells
co-expressing markers of both T/NK cells and macrophages (n = 491 cells). We discarded these cells as likely doublets, as have
others, and re-ran the macrophage PCA and clustering (Zhang et al., 2020; Zilionis et al., 2019). These cells are included in the
full dataset in case they are of interest to others. Each unbiased analysis helped to define the non-malignant phenotypes summarized
in Figure 5 and Figure S6.

Generation of expression signatures/scores

All expression scores were computed as previously described by taking a given input set of genes and comparing their average rela-
tive expression to that of a control set (n = 100 genes) randomly sampled to mirror the expression distribution of the genes used for the
input (Tirosh et al., 2016b). While all scores were computed in the same way, choosing the genes for input varied. We have outlined
the relevant approaches below. Where correlations (Pearson’s r) are performed over genes, we used the log-transformed UMI count
data for each case. Unless otherwise noted, we selected the top 30 statistically significant genes for each signature (> 3 SD above the
mean for shuffled data) for visualization and scoring.
Cell cycle
We utilized previously established signatures for G1/S (n = 43 genes) and G2/M (n = 55 genes) to place each cell along this dynamic
process (Tirosh et al., 2016a). After inspecting the distribution of scores in the complete dataset, we considered any cell > 1.5 SD
above the mean for either the G1/S or the G2/M scores to be cycling (van Galen et al., 2019).
PDAC bulk subtype signatures
We scored malignant cells within our single-cell cohort for expression of previously published signatures derived from bulk RNA-
sequencing of primary and metastatic tumors.
scBasal and scClassical programs
We first scored each malignant single cell for the basal-like and classical genes identified by Moffitt et al., 2015 as these were well
described by unbiased analysis in our data (PCA, Figure S2B,C, S3B). To derive refined single-cell basal (scBasal) and single-cell
classical (scClassical) signatures using our malignant cohort and determine programs associated with these cell states, we corre-
lated the aforementioned basal and classical scores to the entire gene expression matrix containing malignant cells and identified
the 1,909 genes significantly associated with either subtype (r > 0.1; > 3 SD above the mean for shuffled data, full data in Table
S3). Biological pathway correlates for scBasal and scClassical are summarized in Figure S2E,F [WNT signaling (Kim et al., 2017);
EMT (Gröger et al., 2012); cell cycle progression (Tirosh et al., 2016a)]. For visualization, we use the scBasal and scClassical genes
(top 30 correlated genes for each). In Figure 2C we score single cells for EMT (Gröger et al., 2012) and the union of Hallmark and
Reactome interferon response gene sets to show their divergence within cells expressing the scBasal state.
Intermediate co-expressor (IC) program
Ordering the cells by their polarization or ‘‘score difference,’’ simply the difference of the two scores, using these basal and classical
scores related to PC1 and PC2 revealed a significant fraction of cells co-expressing intermediate levels of both cell states (Figure 2B,
Figure S3A,B). Co-expressing cells showed associations with features across several additional PCs, but lacked a single dominant
axis. To define a consensus set of genes that are preferentially expressed by coexpressing cells, we computed the Euclidean dis-
tance to the line representing equal basal and classical co-expression for each cell. To limit the influence of cell quality on this analysis
and to specifically identify genes related to co-expression, we used cells from each group (basal, intermediate, and classical) with
fractionally low mitochondrial genes (< 0.2) and non-zero basal or classical expression (basal or classical score > 0) and correlated
their Euclidean distance (Figure S3C) to the entire gene expression matrix of malignant cells. Next, for each gene positively associ-
ated with this co-expressor state (Pearson’s r > 0), we subtracted the second highest correlation coefficient for each subtype-asso-
ciated gene (basal and classical), and then re-ranked the matrix by this corrected value. This enriched for genes more specific to the
co-expressor state by excluding those that were also associated with basal or classical programs. We then selected the 115 genes
with a corrected correlation value > 0.1 (p < 0.00001, shuffled data) as our intermediate co-expressor (IC) signature (Figure S3D, Table
S3). Single cells were classified based on Euclidian distance to co-expression, where cells with Euclidian distance < 0.2 are defined
as intermediate co-expressor and the remainder (Euclidian distance > 0.2) by their maximal of either scBasal or scClassical scores.
We binned each organoid cell (Figure 4C,D) by its maximal expression for one of the 3 in vivo scores (scBasal, scClassical, or IC).
Here a cell must be within 1 SD of the mean expression for a given subtype in vivo, else it was considered ‘‘organoid-specific’’ as

Cell 184, 6119–6137.e1–e12, December 9, 2021 e10

 ll
OPEN ACCESS Article

this program was superimposed on all organoid cells, regardless of their subtype identity (Figure 4C). We used these classifications to
summarize overall sample malignant cell composition and visualize the groups. Tumor heterogeneity measures were not significantly
affected by changing these cutoffs.
Non-Malignant programs
TAM signatures were determined similar to above and previous work (van Galen et al., 2019; Zhang et al., 2020; Zilionis et al., 2019).
Using PCA as an anchor (Figure S6C), we correlated expression within the TAM compartment to either FCN1, SPP1, or C1QC (top
loaded genes on each relevant PC) and merged the resultant correlation coefficients for every detected gene to the 3 subtypes into
one matrix (i.e., a 16,920 3 3 matrix). For each TAM type (i.e., vector of correlation coeffects to each marker), we first ranked the
matrix by decreasing correlation coefficient, selected only the most significantly associated genes to that type (r > 0.1; > 3 SD above
the mean for shuffled data), subtracted the second highest correlation coefficient for each subtype-associated gene, and then re-
ranked the matrix by this corrected value. We repeated this procedure for each TAM subtype independently. This ensures that
the genes selected are specific to a given TAM subset and do not describe general TAM features. The top 30 genes for each
were used for scoring and visualization (Table S2; Figure S6D).
Mesenchymal phenotypes were determined using a similar workflow. To examine mesenchymal heterogeneity, we removed a
subset of adrenal endocrine cells (cluster 4, 40 cells; Figure 5C) and then performed PCA on the remaining mesenchymal cells.
PC1 was driven by spillover genes (likely contributed from ambient RNA) and lacked any coherent biological program and was
not considered further. PCs 2 and 3 by contrast were consistent with variable mesenchymal (PC2) and inflammatory CAF (PC3) phe-
notypes. All these cells scored highly for previous myCAF gene expression programs so this phenotype did not fully explain the het-
erogeneity in mesenchymal cells. Again, using correlation, we determined the genes driving low PC2 scores (Fibroblast-like), and
high PC2 scores (Pericyte-like), as well as those associated with the high PC3 scores (Inflammatory). As before, we used the top
30 genes within each subset for scoring and visualization. These same genes (Fibroblast-like and Pericyte-like) were used to examine
bulk RNA-seq profiles and their difference in each sample quantifies which phenotype is favored in the bulk averages (Figure 5C).

Analysis of normal pancreas progenitor data

We obtained the genes by cells matrix of normal pancreas progenitors (Qadir et al., 2020) at https://www.ncbi.nlm.nih.gov/geo/
query/acc.cgi?acc=GSE131886. We clustered the original data and excluded a small subset of immune cells (CD45+). We then
collapsed cell types from the original paper into broad categories (Pro-Acinar, Pro-ductal, Undifferentiated, and Mesenchymal)
based on lineage marker expression. For analysis in Figure S3E we averaged the expression for scBasal, scClassical and IC genes
in each group. For use in Figure S3F we generated signatures for each population using differential expression (FindAllMarkers func-
tion in Seurat using the ‘‘wilcox’’ option) and scored our single cells for these normal-derived signatures as above.

Matched organoid clustering and cell-typing

After applying similar quality metrics as above, we performed PCA, SNN clustering, and t-SNE embedding for 32,073 cells including
organoid cells and all malignant cells from primary PDAC biopsies (PCs 1-50; resolution = 1.2; k.param = 45; perplexity = 45; max_
iter = 2,500), and identified 39 total clusters. Organoids clustered separately from their matched biopsies, suggesting expression
and/or CNV related drift in culture. Only two SNN clusters—clusters 4 and 32—were admixed by sample. We determined the specific
gene expression programs in these two clusters via differential expression testing by Wilcoxon rank sum test (p < 0.05, Bonferroni
correction; log(fold change) > 0.5). These comparisons were done in a ‘‘1 versus rest’’ fashion, testing for genes defining each cluster
(4 or 32) compared to the entire dataset. Their expression profiles were consistent with non-malignant cell types, likely fibroblasts
(cluster 32) and epithelial cells (cluster 4; Figure S5B,C).

Correlation distances for genotype and transcriptional cell state

To generate correlation distances for genotype and transcriptional cell state, each single cell in a biopsy-organoid pair was repre-
sented by two vectors of information: (i) a cell state vector containing expression values for scBasal and scClassical genes (n =
60 genes) and (ii) a genotype vector containing the average CNV score for each cytoband. The cell state and genotype distances
between every single cell within a biopsy/early organoid pair was computed from these vectors using a correlation-based (Pearson’s
r) distance metric of the form d = (1-r)/2. This resulted in two distance matrices of n x n dimension where n is the total number of cells
from each biopsy/early organoid sample pair. Values in Figure 3A are computed by averaging the values for d between only early
organoid and matched biopsy cells.

Matched biopsy versus organoid malignant cell comparison

For CNV-confirmed malignant cells from each biopsy and its matched organoid (earliest passage), we used differential expression
(Wilcoxon rank sum test; p < 0.05, Bonferroni correction; log(fold change) > 0.3) to understand the features lost from malignant cells in
the in vivo setting and gained when transitioning into growth in organoid culture. We required any gene to be significantly differentially
expressed in at least 3 model-biopsy comparisons to summarize the consistent changes. We repeated this same workflow for both
organoid- and biopsy-specific genes (Table S4) outlined in Figure 3C and Figure 4G,H, respectively.

e11 Cell 184, 6119–6137.e1–e12, December 9, 2021

 ll
Article OPEN ACCESS

TME associations
We determined the transcriptional-subtype-dependent composition of the TME (Figure 5E-I) following two steps. First, we computed
the Simpson’s Index (measure of ecological diversity) using the count of each non-malignant cell type captured from each sample as
input (Figure 5E,G) and correlated each biopsy’s diversity score to its scBasal versus scClassical commitment score. Importantly, the
number of non-malignant cells captured from each biopsy was not associated with basal versus classical commitment score (r =
0.09). Next, to understand which cell types drive these differences, we computed the fractional representation for every non-malig-
nant cell type in each core needle biopsy and determined their pairwise correlation distance (Pearson’s r) followed by hierarchical
clustering using Ward’s method (dendrogram in Figure 5G). For both analyses we only used samples with > 200 non-malignant cells
captured (Figure S6N).

Biopsy paracrine and autocrine subtype-specific factor analysis

Factors present in the TME but absent from organoid culture could originate from at least two sources, the malignant cells themselves
(autocrine) or non-malignant cells in the local microenvironment (paracrine). We examined any gene with gene ontology annotations
related to ‘‘cytokines,’’ ‘‘chemokines,’’ or ‘‘growth factors’’ and took the union of these lists, yielding 321 genes, 218 of which were
detected in our dataset. For ‘‘autocrine’’ factors we performed differential expression between malignant cells binned as scBasal and
scClassical, and then IC versus rest. A gene was considered differentially expressed if it passed a p < 0.05 with Bonferroni correction
and a log(fold change) > 0.2 in one of these comparisons. Genes were then assigned to subtypes based on the log fold change
direction (Figure 6B, Table S6). ‘‘Paracrine’’ factors were determined in a similar manner with slight modifications. We grouped
non-malignant cells into basal, classical or IC based on the average expression and clustering for malignant programs from their
respective tumor samples (Figure S3G). We then assessed for differential expression between all cells from a given group and the
rest using the same cutoffs as above and sorted factors into subtypes based on their log fold change directionality (Figure 7A, Table
S6). We visualized which cell type contributed the highest average expression for each factor among the cell types from each of the
respective cell state-specific TMEs (Figure 7B). We note that our use of ‘‘paracrine’’ and ‘‘autocrine’’ here is somewhat inexact as
these secreted factors could act in either manner depending on the context. We merely use this nomenclature to reflect a ‘‘cancer
cell centric’’ view, i.e., factors secreted by malignant cells are autocrine and those deriving from the TME are paracrine.

Tumor phenotyping from mIF data

Supervised machine learning algorithms were applied for tissue and cell segmentation (inForm 2.4.1, Akoya Biosciences). Single-
cell-level imaging data were exported and further processed and analyzed using R (v3.6.2). To assign phenotypes to individual ma-
lignant epithelial cells, mean expression intensity in the relevant subcellular compartment was first used to classify cells as positive or
negative for each of the 5 markers. Combinatorial expression patterns for the five markers were then used to phenotypically classify
cells as basal, classical, co-expressing / IC or marker negative (3 combinations of 2 basal markers, 7 combinations of 3 classical
markers, 1 pan-marker negative, 21 combinations of co-expression of basal and classical markers, Figure S4A, Table S3). Tumor
subtype composition was assessed by calculating the fraction of total malignant cells positive for each cell phenotype (Figure S4B,
excluding pan-marker negative cells).

Cell 184, 6119–6137.e1–e12, December 9, 2021 e12

 ll
Article OPEN ACCESS

Supplemental figures

(legend on next page)

 ll
OPEN ACCESS Article

Figure S1. Clinical and genomic features, quality metrics, unsupervised cell type identification, and malignant cell confirmation across the
biopsy cohort, related to Figure 1
(A) Distribution of PDAC patient tumor and cell model phenotypes along the basal-classical spectrum for bulk RNA-seq samples in Figure 1C. P-value for group
differences was calculated by ANOVA.
(B) Clinical and molecular features for all patients included in the dataset (Rx = Therapy; Other = Adrenal (PANFR0637), Omentum (PANFR0635, PANFR0598),
Peritoneum (PANFR0588); Org. at P2 = Organoid measured at passage 2). Mutations were determined by bulk targeted DNA-seq (Red, Altered; White, wild type;
Grey, Data not available). Number of single cells captured per biopsy and their malignant and non-malignant fraction is visualized at the right. The scRNA-seq
biopsy for sample PANFR0604 did not contain any malignant cells.
(C) Distribution of unique molecules and genes captured in quality cells per biopsy, median values are indicated for each metric (dotted line) and violin plots are
colored by patient (top, Log10(UMIs); bottom, number of genes).
(D) Distribution of fraction mitochondrial reads across the entire trimmed biopsy dataset (n = 23,042 cells). Red dotted line denotes the median.
(E) t-SNE visualization of the entire single-cell biopsy dataset colored by the identified SNN clusters (inset numbers).
(F) Distribution of single cells captured per biopsy across the identified and putative malignant and non-malignant SNN clusters.
(G) Heatmaps represent select scRNA-seq-derived copy number profiles where expression across the transcriptome is organized by chromosome (columns) for
each single putative malignant cell (rows) from a given biopsy. Top bar indicates reference bulk targeted DNA-seq for the same patient.
(H) CNV correlation (averaged top 5% of altered cells per biopsy) versus CNV score (mean square of modified expression) for each single putative malignant
(colored points) and reference normal (empty black circles) cell within a given biopsy. One sample, PANFR0604, did not contain any malignant cells.
(I) Overview of cell-typing for all cells in the biopsy dataset. Cells are ordered by SNN cluster and separated by cell types. Top heatmap represents expression
levels for a subset of select markers (n = 73 genes) used to identify cell types. Color bar indicates cell types and binarized cell cycle phenotypes are labeled (black,
cycling; white, not). CNV scores (mean square of alterations per cell) used to parse malignant from non-malignant are shown using T/NK, endothelial, fibroblasts,
and hepatocytes as reference; gray boxes denote normal cell types where we did not compute reference CNV scores. Bottom panel shows biopsy of origin for
each cell. The data are split by non-malignant (n = 15,302) and malignant (7,740) identity.
(J) t-SNE visualization as in S1E but colored by cell types identified. Endo, Endothelial; Mes, Mesenchymal; B, B cell; Hep, Hepatocyte; DC, Dendritic cell; pDC,
Plasmacytoid dendritic cell; Mac, Macrophage; T, T cell; NK, Natural killer cell.
(K) Fraction of each cell type contributed by each biopsy sample (color fill, patient ID as in S1B), cell type totals are noted at the top of each bar.
 ll
Article OPEN ACCESS

Figure S2. Identifying and contextualizing single-cell scBasal and scClassical associated biology, related to Figure 2
(A) Principal component analysis (PCA) and scatterplot for PC1 and PC2 across all malignant cells (n = 7,740, colored by patient ID as in Figure S1B) separates
PANFR0580’s malignant cells (n = 662) from the rest. Heatmap for genes with the strongest negative loading on PC1 (n = 30) denotes a neuroendocrine identity
(TTR, CHGB). This tumor was later classified by histology as a pancreatic neuroendocrine tumor (PanNET).
(B) Principal component analysis (PCA) and scatterplot for PC1 and PC2 across all PDAC malignant cells (n = 7,078, colored by patient ID as in Figure S1B),
excluding PANFR0580’s PanNET malignant cells. Heatmap for genes with the strongest positive and negative loadings on PC2 (n = 30 each) shows overlap with
genes from the Moffitt basal and classical signatures.
(C) Principal component (PC) elbow plot (top) showing the standard deviation for the first 20 components calculated over the verified PDAC malignant cell variable
genes. Line is drawn at the putative ‘‘elbow’’ (black versus gray points) as inclusion of additional PCs described overlapping information or quality metrics. Cross-
correlational analysis (bottom) for each single-cell’s embeddings across first 9 PCs (black points) and scores for literature curated gene sets describing EMT,
classical and basal, and cell cycle phenotypes.

(legend continued on next page)

 ll
OPEN ACCESS Article

(D) Pairwise correlation of genes significantly enriched in scBasal or scClassical expression states within single-cell cohort. Left bar indicates the subtype as-
sociation of each gene (orange, scBasal-enriched; blue, scClassical-enriched).
(E) Tied dot plots depict the correlation coefficient for each gene (points) from select literature-derived gene sets, indicated at the top of each plot, to either
scBasal or scClassical cell states. Dotted lines represent significance threshold (3 SD above the mean of shuffled data), points and lines are colored if that gene
passes the threshold.
(F) GSEA pathway enrichments for the top 100 genes correlated to either scBasal or scClassical expression scores.
 ll
Article OPEN ACCESS

Figure S3. Cells with co-expression of scBasal and scClassical states express a distinct intermediate co-expressor gene program, related to
Figure 2
(A) Expression of scBasal and scClassical gene programs, with cells ordered by their scBasal-scClassical score difference. Quality metrics, EMT scores and the
binarized cell cycle program are shown for each single cell below the heatmap. P-value for binarized cycling group differences was calculated using Fisher’s
Exact test. P-value for EMT score was calculated by Kruskal-Wallis test with multiple hypothesis correction.
(B) Difference between PC1 and PC2 (top) and scClassical–scBasal score difference (bottom) are shown. Cells with equal scBasal and scClassical expression are
associated with intermediate PC scores and cells are ordered as in S3A. P-values for group differences were calculated by Kruskal-Wallis test with multiple
hypothesis correction.

(legend continued on next page)

 ll
OPEN ACCESS Article

(C) Euclidean distance for each cell to co-expression (y = x) of scBasal (x) and scClassical (y) expression scores. Bottom track indicates the score derived from the
genes specific to the intermediate co-expressor state shown in S3D. P-values for group differences were calculated by Kruskal-Wallis test with multiple hy-
pothesis correction.
(D) Gene correlation to either scBasal or scClassical score (x axis) or the corrected intermediate co-expressor correlation (Euclidean distance in S3C). Green
highlighted genes have corrected intermediate co-expressor correlation > 0.1 (p < 0.00001 above shuffled).
(E) Averaged expression of all three malignant programs in normal pancreatic progenitor niche subsets (Acinar (Pro Ac.), Ductal (Pro Duct.), or Undifferentiated
(Undiff.) and mesenchymal (Mes.) cells defined in Qadir et al., 2020. P-values for each set of genes are computed by Kruskal-Wallis test with multiple hypothesis
correction.
(F) Cross-correlation between new and previously proposed expression signatures in our PDAC single cells. Average expression for each signature (rows) is
shown at the right for cells in the malignant subtypes from our cohort and the normal pancreatic progenitor cells from Qadir et al., 2020. White dot indicates the
subset with the highest average significant expression for each signature (Kruskal-Wallis test); no white dot indicates no significant expression.
(G) Pairwise correlation for biopsies with malignant cells (n = 22). Data are correlation coefficients for the average expression of all signature genes in the malignant
cells from a given biopsy. Clade identities are at left with the one PanNET tumor (PANFR0580) included for comparison.
(H) Average expression for the 184 genes used for clustering in S3G. Clade identity colors match text color in S3G, and individual samples (columns) are ordered
as in S3G with sample ID numbers provided below.
(I) Scores for the expression of genes in S3H (gray scale heat) across the 4 main cell types found in the pancreatic progenitor niche (Qadir et al., 2020). White dot
indicates the normal subset with the highest average expression for each malignant program (Kruskal-Wallis test), none of the normal subsets significantly
express the neuroendocrine gene signature.
(J) Frequencies of scBasal, scClassical, and IC states in individual tumors demonstrates intratumoral transcriptional heterogeneity.
 ll
Article OPEN ACCESS

Figure S4. Multiplex immunofluorescence is concordant with scRNA-seq and demonstrates intratumoral heterogeneity with the presence of
intermediate co-expressor cells, related to Figure 2
(A) Schematic for comparison of the matched datasets by combinatorial marker phenotypes.
(B) Marker detection in each single cell from the 10 samples in the mIF (top, 130,784 cells) and matched scRNA-seq datasets (bottom, 3,062 cells). Cells are
sorted by their combinatorial phenotype outlined in S4A.
(C) Comparison within and between modalities on matched samples. Samples are sorted by their pseudo-bulk RNA subtype identity (color bar, top; dendrogram
from Figure S3G). Pearson correlations (red to blue heat) were performed over the fractional representation of each marker pattern in S4A for each biopsy using
either modality (scRNA-seq or mIF). The upper left and lower right quadrants cross-compare cell state composition from individual biopsies assessed by mIF and
scRNA-seq, respectively, with both showing agreement with averaged RNA subtype (top color bar). The upper right and lower left quadrants cross-compare

(legend continued on next page)

 ll
OPEN ACCESS Article

modalities (scRNA-seq versus mIF) for each sample (white dots), demonstrating that individual marker patterns for each sample are similar despite being
measured by different methods in different cores from the same patient.
(D) Frequency of co-expressing cells is related to increased mixing of basal and classical cell populations within patient samples assessed by mIF. Log ratio of %
basal and classical cells in each sample (x axis; dotted line at 0 indicates equal percentages of basal and classical cells) versus their % co-expressing cells (y axis).
(E) mIF images and marker detection in single cells from two primary PDAC samples identifies IC cells in primary specimens also. Scale bar represents 10 mm.
 ll
Article OPEN ACCESS

(legend on next page)

 ll
OPEN ACCESS Article

Figure S5. Quality metrics, cell type identification, serial sampling, and media influences on transcriptional state in patient-matched or-
ganoids, related to Figures 3 and 4
(A) Distribution of unique molecules and genes captured in quality cells per organoid sample, median values are indicated for each metric (dotted line) and violin
plots are colored by patient ID (top, Log10(UMIs); bottom, number of genes).
(B) t-SNE visualization of all biopsy and matched organoid cells from iterative passages, colored by patient ID (left). Dotted circles indicate the only two SNN
clusters (4 and 32) with appreciably admixed clusters and low CNV scores, the rest were patient-specific. Bar chart shows number of organoid cells recovered per
matched sample (center). On the right t-SNE visualization, cells are organized in the same manner but organoid cells are colored by passage number.
(C) Relative expression for genes defining cluster 4 (top) and cluster 32 (bottom). Cluster 4 had an ambiguous epithelial identity while cluster 32 cells were defined
by canonical fibroblast genes and low to absent detection of CNVs.
(D) Fraction of cluster 4 cells at each passage, demonstrating that these cells did not survive iterative passaging.
(E) Single-cell inferred copy number alterations for each sample in the biopsy cohort. Tumors are grouped by expression of their dominant subtype based on the
clustering in Figure S3G, P-values comparing presence of each alteration among the groups (scBasal, scClassical, IC) are determined by Fisher’s exact test.
(F) Cell state diagram with marginal density plots for all cells with inferred KRAS amplifications (4 biopsy-model pairs) in biopsy (black) and organoid (red) mi-
croenvironments. P-value compares biopsy versus early passage organoid score distributions (top density) and was determined by Student’s t test.
(G-I) Heatmaps show inferred CNV copy number status for every cell in each of three biopsy/early passage organoid pairs. Cells (in rows) are ordered by hi-
erarchical clustering of their CNV profiles and letters on the far left indicate subclones that have significant statistical evidence for tree-splitting (STAR Methods).
Each cell’s origin is also noted (‘‘Source’’ column; biopsy tissue, gray; early passage organoid, red). Right metadata bars indicate if that cell came from an
admixed, likely non-malignant SNN cluster (4 or 32 in S5B).
(J-K) Matched cell state and genotype evolution at each passage in PANFR0489R (S5J) and PANFR0575 (S5K). Frequencies of individual CNV clones at each time
point (y axis) are tied by colored lines. Fill represents the transcriptional cell state fraction for each CNV clone. In sample PANFR0575 (S5K), clones D and E had
inferred TP63 amplifications which expanded over time.
(L) Bulk RNA-seq expression of basal and classical states in 4 organoid models cultured in 3D or adapted to 2D in complete organoid media.
(M) Inferred CNVs for each cell from the PANFR489R samples cultured in either Minimal (gray) or Complete (red) organoid media conditions in Fig. 4B.
(N) Brightfield images of organoids grown in standard organoid media (‘‘Complete’’) or in media without any growth factors (‘‘Minimal’’) at days 1 and 11 after
seeding. Scale bar represents 200 mm.
(O) Expression of scBasal, IC, and scClassical genes (rows) in organoid cells (columns) from model PANFR0562 cultured for 6 days in Complete medium, Minimal
medium, or OWRNA medium. P-values for group differences were calculated by ANOVA followed by Tukey’s HSD.
(P) Expression of scBasal, IC, and scClassical genes (rows) in organoid cells (columns) from model PANFR0562 cultured for 6 days in Complete organoid medium
or in ‘‘Cell line’’ medium. P-values for group differences were calculated by ANOVA followed by Tukey’s HSD.
(Q) Expression of scBasal, IC, and scClassical genes (rows) in single cells from the PDAC cell line CFPAC1 (columns) cultured for 6 days in standard ‘‘Cell line’’
medium or in Complete organoid medium. P-values for group differences were calculated by ANOVA followed by Tukey’s HSD.
 ll
Article OPEN ACCESS

(legend on next page)

 ll
OPEN ACCESS Article

Figure S6. Identification of T cell, natural killer cell, macrophage, and mesenchymal heterogeneity and transcriptional state associations in
the metastatic microenvironment, related to Figure 5
(A) t-SNE visualization of sub-clustering (SNN) performed on T/NK cells in the metastatic cohort. Cells are colored by their type identity based on shared SNN
cluster membership.
(B) Select cell type marker expression overlaid on the t-SNE visualization from S6A.
(C) PCA identifies 3 major subsets of TAMs in the metastatic niche. PC1 largely separates FCN1+ monocyte-like TAMs from more committed macrophage
phenotypes. PC2 separates SPP1+ from C1QC+ macrophage phenotypes.
(D) Heatmap visualization of the gene expression programs specific to each TAM subset identified by the PCA and SNN clusters in S6C.
(E) Relative expression for select cell type markers and binarized cell cycle program (top bar; black, cycling) with cell type colors (bottom color bar) as in Figure 5A.
(F) t-SNE visualization of sub-clustering (SNN) performed on mesenchymal cells colored by their anatomical site. Cell subsets (1-4) determined by SNN clustering.
(G) Average expression of the indicated mesenchymal/fibroblast and adrenal endocrine marker genes in each of the cell subsets (1-4) identified in S6F. Dot size
indicates fraction of cells expressing a given gene.
(H) PCA over mesenchymal cells in the cohort (excluding Adrenal endocrine cells; subset 4, S6F). Scatterplot of PC2 versus PC3 defines 3 states for mesenchymal
cells in our cohort.
(I) Same visualization as in S6H, but cells are colored by previously identified myCAF or iCAF signature scores. myCAF is evenly distributed across PC2 and iCAF
associates with higher PC3 scores.
(J) Expression for select markers overlaid on the PCA from S6H.
(K) Cross-correlation of mesenchymal signatures in single-cells. Fibroblast-like versus pericyte-like signatures provide non-overlapping information. PC3 in-
flammatory phenotypes are similar to the previously reported iCaf phenotype and our PC3-derived inflammatory mesenchymal signature.
(L) Frequency of mesenchymal cells (y axis, cell count) across PC2 scores, colored by site of biopsy tissue. P-value determined by Student’s t test.
(M) Comparison of the distribution of mesenchymal phenotypes across different sites as groups (top) or individual tumors (bottom). Heat indicates the fraction of
mesenchymal cells in that score bin.
(N) Non-malignant cell types (color fill) and counts per biopsy, included samples organized as in Figure 5G. Five biopsies were excluded from the analysis in
Figure 5E,5G-I because they either had low cell capture or were from a tumor with indeterminant malignant transcriptional subtype due to negligible malignant cell
capture.
(O) Correlation between T cell fraction and malignant IC score.
(P) Cross-TCGA analysis for basal and immune cell type markers in epithelial tumors with known basal subtypes. Clusters were determined by dendrogram
splitting, and disease type for each sample is indicated at bottom.
 ll
Article OPEN ACCESS

(legend on next page)

 ll
OPEN ACCESS Article

Figure S7. State-specific drug responses and TME-associated state shifts in organoid models and patient samples, related to Figures 6
and 7
(A) Independent experiments demonstrating state-specific drug sensitivities in isogenic PANFR0562 organoid model pairs skewed toward scBasal or scClassical
states by altering media formulation (scBasal, TGF-b or Minimal media; scClassical, OWRNA media) for different time periods. Bars represent the difference in
growth rate-corrected Area Over the Curve (AOC) between each scBasal-scClassical model pair for each independent experiment depicted in Figure 6G.
(B) Growth rate-adjusted dose response curves for isogenic PANFR0562 organoid model pairs in different media conditions and at different time points. Points
are mean ± SD of 3 technical replicates. Orange curves represent responses in more scBasal conditions and blue curves represent responses in more scClassical
conditions.
(C) Cell state diagrams with marginal density plots for the organoid model PANFR0489 depicted in Figure 7C in control medium (OWRNA, reduced organoid
medium) or control medium with IFNg. P-values for group differences between B/C commitment (top) and IC scores (right) were calculated by ANOVA followed by
Tukey’s HSD.
(D) Cell state diagrams with marginal density plots for organoid model PANFR0562 at 2 time points with exposure to IFNg. P-values for group differences between
B/C commitment (top) and IC scores (right) were calculated by ANOVA followed by Tukey’s HSD.
(E) t-SNE visualization of single cells from metastatic biopsies collected from the same patient (PANFR0473, Figure 7D) but at two different sites (liver, left; lung,
right). All cells from both samples are shown in both t-SNE plots, but cells for each site are labeled red in the corresponding plot.
(F) Cell state diagrams with marginal density plots (top) and expression heatmaps (bottom) depict cell state differences between malignant cells from two distinct
metastatic sites for the samples (PANFR0473) in Figure 7D and S7E. P-values for group differences between B/C commitment (top) and IC scores (right) were
calculated by ANOVA followed by Tukey’s HSD.
(G) t-SNE visualization of single cells from metastatic biopsies collected from the same site in one patient (PANFR0489, Figure 7E) pre- and post-immunotherapy.
All cells from both samples are shown in both t-SNE plots, but cells for each time point are labeled red in the corresponding plot.
(H) Cell state diagrams with marginal density plots from metastatic samples collected pre- and post-immunotherapy (PANFR0489, Figure 7E,F and S7G). P-
values for group differences between B/C commitment (top) and IC scores (right) were calculated by ANOVA followed by Tukey’s HSD.