Transcriptomic Plasticity Is a Hallmark of Metastatic Pancreatic Cancer

Single-cell transcriptional profiling of primary tumor and metastases from rapid autopsy samples of an individual with pancreatic cancer, combined with probabilistic clonal inference by PICASSO, reveals substantial transcriptomic plasticity in metastatic cells.

Abstract

Metastasis is the leading cause of cancer deaths. To develop strategies for intercepting metastatic progression, a better understanding of how tumor cells adapt to vastly different organ contexts is needed. To investigate this question, a single-cell transcriptomic atlas of primary tumors and diverse metastatic samples (liver, omentum, peritoneum, stomach wall, lymph node, and diaphragm) from a patient with pancreatic ductal adenocarcinoma who underwent rapid autopsy was generated. Using unsupervised archetype analysis, both shared and site-specific gene programs were identified, including lipid metabolism and gastrointestinal programs prevalent in peritoneal and stomach wall lesions, respectively. We developed phylogenetic inference from copy-number alterations in single-cell sequencing observations (PICASSO) as a probabilistic approach for inferring clonal phylogeny from single-cell and matched whole-exome sequencing data. Comparison of PICASSO-generated clonal structure with phenotypic signatures revealed that pancreatic cancer cells adapted to local environments with minimal contribution from clonal genotype. Our results suggest a paradigm whereby strong environmental effects are imposed on highly plastic cancer cells during metastatic dissemination.

Significance:

Single-cell transcriptional profiling of primary tumor and metastases from rapid autopsy samples of an individual with pancreatic cancer, combined with probabilistic clonal inference by PICASSO, reveals substantial transcriptomic plasticity in metastatic cells.

This article is part of a special series: Driving Cancer Discoveries with Computational Research, Data Science, and Machine Learning/AI .

Graphical Abstract

Introduction

Metastasis is a systemic disease responsible for the majority of cancer-related deaths (1), yet our understanding of how tumor cells disseminate and thrive in distant tissues remains limited. To metastasize, cancer cells must overcome many hurdles, including the need to escape from the tissue of origin, migrate, evade immune surveillance, and invade distant tissue (2). The microenvironments of different organs each pose additional adaptive challenges for cancer cell colonization. For some tumor types, selection may act on intratumor genetic heterogeneity to shape these adaptive processes (3), whereas for others, genomic studies have uncovered few recurrent mutations associated with specific metastatic behaviors or organotropism (4). More recently, epigenetic plasticity has emerged as a hallmark of cancer, which confers the ability to reinvent cellular phenotypes and drive phenotypic heterogeneity in the service of adaptation (5). How this plasticity manifests at the molecular level, the extent to which it shapes tumor progression (6), and its relevance to treatment (7) are major open questions.

Pancreatic ductal adenocarcinoma (PDAC) exhibits particularly low heterogeneity in driver mutations, which tend to be shared across primary and metastatic sites (8), underscoring the need to identify alternate adaptive mechanisms. Advanced tumors are not commonly resected, and metastases are rarely biopsied sequentially, making it difficult to reconstruct tumor progression and providing scant metastatic data in some organs. Rapid autopsy offers a critical opportunity for systematically investigating shared and organ-specific metastatic programs in multiple lesions derived from a single germline (9). The ability to collect multiple independent metastases from a single organ also provides an unparalleled approximation of a controlled biological replicate in human cancer. Such postmortem sampling, coupled with genotyping and lineage reconstruction, recently provided insights into modes of evolution and metastatic seeding in PDAC (10).

To gain insights into the molecular mechanisms of adaptation in this patient-centric view, however, requires a combination of clonal lineage information and deep phenotypic profiling. Single-cell gene expression data provide rich phenotypic information at the cellular level, but they are problematic for clonal and phylogenetic reconstruction, whereas simultaneously sequencing single-cell DNA can provide genotype information but does not scale sufficiently. Typical phenotypic analyses are also not designed to find adaptive gene programs. Computational approaches are thus needed to overcome these challenges and enable the comparison of clonal lineages and molecular phenotypes in a single cancer across multiple lesions and organs.

We collected two primary and nine metastatic tumors from a patient with PDAC who underwent a rapid autopsy and subjected the samples to single-nucleus RNA sequencing (snRNA-seq), recovering the transcriptomes of more than 45,000 cancer epithelial cells. Using archetypal analysis, we identified adaptive gene programs that are missed by standard clustering. To investigate the evolutionary dynamics of metastatic PDAC, we developed integrateCNV, an approach to robustly infer copy-number alterations (CNA) from snRNA-seq and matching bulk whole-exome sequencing (WES) data, and phylogenetic inference from CNAs in single-cell sequencing observations (PICASSO), a method to identify cell clones and generate clonal phylogenies using potentially noisy single-cell CNA profiles. We find evidence of strong adaptation to the local organ microenvironment, including metabolic rewiring of peritoneal lesions—a very common but little-studied site of metastasis in PDAC—as well as multiple different shared epithelial–mesenchymal transition (EMT) programs. Our work identifies plasticity as the major force in PDAC metastatic adaptation and provides approaches for deep phenotypic and phylogenetic analysis from single-cell expression data.

Materials and Methods

Biospecimen collection

Patient information

Warm autopsy samples were collected from a 35-year-old female patient with informed consent to the Last Wish Program and approval from the patient’s family. Written informed consent was obtained from all patients whose tissues were used. The study was conducted in accordance with the recognized ethical guidelines of the Declaration of Helsinki and the Belmont Report and was approved by the Institutional Review Board at Memorial Sloan Kettering Cancer Center (MSKCC; IRB protocol 15-021).

The patient was diagnosed with metastatic PDAC, exhibiting macroscopic lesions in the pancreas and liver (detected by CT scan) and upregulated CA19-9 tumor biomarker. The patient was treated with standard modified 5-fluorouracil, leucovorin, irinotecan, and oxaliplatin (mFOLFIRINOX) therapy, and tumors showed a clinical response for approximately 6 months before they stopped responding, at which point, mFOLFIRINOX was halted and a dose of gemcitabine + nab-paclitaxel was given, but no further response was observed. The patient survived for just over 9 months from diagnosis, which is expected in a patient with metastatic PDAC treated with standard chemotherapy.

Both primary and metastatic tumors were readily detectable. The primary tumor appeared as a white–gray mass, whereas liver metastases were white–yellow with extensive necrosis. Multiple peritoneal and omental metastases, along with a single gastric metastasis, were palpable and appeared as white nodules. Prominent diaphragm metastases resembling an “omental cake” were also identified.

Biospecimen collection

Samples were obtained using standard autopsy techniques, specifically the Rokitansky method. Following the removal of all organs from the body, more than 50 samples were collected from macroscopically identifiable tumors in both primary and metastatic sites. Autopsies were initiated within 2 hours of death, and biospecimens were collected within an hour. Multiple lesions collected from the same organ were clearly separate anatomically. The exception is the primary tumor, for which two adjacent sections were processed as pancreas A and B samples (see below for sectioning information) for snRNA-seq. Tumors larger than 1 cm in size were trimmed to 1 cm squares and then divided in half. One half was used to generate a formalin-fixed, paraffin-embedded block for detailed histologic analysis. The other half was cut into 5 to 7 mm pieces, placed in cryotubes, rapidly frozen in liquid nitrogen, and stored at −80°C. For particularly large primary tumors, samples were obtained after slicing. The position of each sampling site within the organ was meticulously documented during the autopsy. Approximately 10 normal tissue samples were taken alongside the tumors.

For WES, a portion of each flash-frozen sample was used to create an optimal cutting temperature (OCT) block. Hematoxylin and eosin (H&E) staining of frozen OCT sections was performed to identify tumor regions and confirm the inclusion of sufficient tumor tissue before macrodissection to extract DNA for bulk WES, typically from 5 to 10 sections. H&E staining was performed by the MSKCC Pathology Core Facility.

For snRNA-seq, a different portion of the frozen tissue was sectioned, and tumor tissue inclusion was confirmed using the frozen H&E slide before proceeding with single-nucleus suspension and sequencing library preparation.

Experimental methods

WES

For bulk WES, genomic DNA was extracted from each tissue sample using the QIAamp DNA Mini Kit (QIAGEN; RRID:SCR_008539). Sequencing was carried out on an Illumina HiSeq 4000 (RRID:SCR_016386) or NovaSeq 6000 (RRID:SCR_016387) platform by the MSKCC Integrated Genomics Operation Core, with a target coverage of 250× for all samples.

snRNA-seq

Generation of nucleus suspensions

Single-nucleus suspensions were generated following the frozen tissue dissociation for snRNA-seq protocol (https://dx.doi.org/10.17504/protocols.io.81wgb13pqvpk/v1). This protocol is optimized for the capture of epithelial cells. Specifically, frozen rapid autopsy specimens were cut into approximately 2 mm³ pieces using a disposable scalpel (Technocut, 10148-882) and transferred to 1 mL of freshly prepared ice-cold lysis solution [250 mmol/L sucrose, 50 mmol/L citric acid, 0.01% diethyl pyrocarbonate (DEPC)]. Next, the entire lysis solution with specimens was transferred to a Dounce homogenizer (Sigma, D8938-1SET). Tissue grinding was performed by gently moving a large-clearance pestle (tube A) up and down 10 to 15 times, followed by a small-clearance pestle 10 times (tube B). After grinding, the homogeneous suspension of minced tissue was strained through a 35-μm snap cap strainer (Fisher Scientific, 352235) and kept on ice for 1 minute. Filtered nucleus suspension was transferred into a 2 mL tube and spun at 4°C in a swinging bucket centrifuge at 500 g for 5 minutes. The supernatant was discarded, leaving ∼20 μL above the nucleus pellet. Next, the pellet was resuspended in 1 mL ice-cold nucleus wash buffer [250 mmol/L sucrose, 50 mmol/L citric acid, 1% (w/v) BSA, 20 mmol/L DTT, and 0.2 U/μL RNase inhibitor (Ambion Inc.; RRID:SCR_008406, AM2682), in DEPC-treated water (Ambion Inc., AM9915G)]. The tube was centrifuged in a swinging bucket at 500 g for 5 minutes at 4°C, and the supernatant was aspirated without disrupting the now-smaller pellet. The pellet was then resuspended in 0.5 mL nucleus resuspension buffer [3× saline-sodium citrate (Invitrogen, AM9770), 20 mmol/L DTT, 1% (w/v) BSA, and 0.2 U/μL RNase inhibitor (Ambion Inc., AM2682), in DEPC-treated water (Ambion Inc., AM9915G)] and passed through a 35 μm snap cap strainer. Nuclei were quantified by staining 10 μL of nucleus suspension with 0.2 μL of 100 μg/mL DAPI and 10 μL of 0.4% Trypan Blue and carefully inspected for quality and separation under bright-field and fluorescence microscopes. The entire procedure took approximately 1 hour to complete and generated 10⁶ to 10⁷ single nuclei per 1 mL.

Single-nucleus enrichment

Prior to snRNA-seq, single nuclei were purified by fluorescence-activated cell sorting (FACS) to remove debris and clumps following our protocol. In a typical scenario, a 50 μL aliquot of the nucleus suspension was added to 250 μL of nucleus resuspension buffer and used as an unstained reference sample for FACS, and the remaining suspension (∼900 to 950 μL) was stained with 10 μL of 100 μg/mL DAPI. Nucleus sorting was performed on a BD FACS Aria II Cell Sorter (RRID:SCR_018934) instrument equipped with a 100 μm nozzle. Sorting was conducted at 5,000 to 10,000 events/second by selecting events based on DAPI signal and particle size. The sorted nuclei were transferred to a 1.5 mL Protein LoBind tube (Eppendorf) and centrifuged in a swinging bucket at 600 g for 5 minutes at 4°C. The nucleus pellet was resuspended in 100 μL of supernatant and manually counted under a bright-field microscope after mixing 10 μL of nucleus suspension with 10 μL of 0.4% Trypan Blue. The suspension concentration was adjusted to obtain ∼2,000 nuclei/μL before proceeding with the v3 chemistry kit on the Chromium instrument (10x Genomics; RRID:SCR_023672).

All samples were split and processed by the sorting protocol above or without it (unsorted). Both unsorted and sorted samples were submitted for snRNA-seq preparation to ensure no systematic biases were experimentally generated.

snRNA-seq library preparation

snRNA library preparation was performed following the Chromium Single Cell 3′ Reagent Kits User Guide, v3.1 Chemistry (10x Genomics), as per our protocol. Library sequencing was performed on Illumina NovaSeq 6000 instruments using a paired-end 2 × 150-bp configuration.

Algorithmic development

IntegrateCNV for copy-number inference

Copy-number inference from single-cell RNA-seq (scRNA-seq) data assumes that changes in gene expression reflect underlying changes in gene dosage. However, epigenetic factors also affect expression and obscure the link between expression and copy number. Furthermore, scRNA-seq data are noisy and sparse, leading to noise in the inferred copy-number profiles. To mitigate noise and sparsity, we restrict single-cell copy-number inference to regions that are known, with high confidence, to harbor CNAs based on bulk WES data, thereby greatly reducing false positive calls. Sparsity is also mitigated by aggregating expression across genes for greater robustness within these regions.

We developed IntegrateCNV to infer per-cell copy-number variation from single-cell or snRNA-seq paired with sample-matched bulk WES data. IntegrateCNV first identifies regions likely to harbor CNAs in WES data and then calculates the likelihood of each of these genomic regions being altered in each single cell. IntegrateCNV accepts as input (i) a cell $\mathrm{\times }$ gene count matrix of scRNA-seq data and “normal” or “tumor” annotation for each cell and (ii) paired copy-number profiles from bulk WES data in matching samples. Using this information, the algorithm determines (i) a set of chromosomal regions that are copy-neutral across all samples and (ii) a set of chromosomal regions of sufficient size that are altered in at least one sample. Finally, integrateCNV outputs (iii) a cell $\mathrm{\times }$ region matrix containing the likelihood of that cell being copy number neutral in that region for each (cell, region) pair.

IntegrateCNV algorithm

The integrateCNV algorithm performs a two-tailed hypothesis test to determine whether each (cell, region) pair has expression levels that differ significantly from the expression levels in known normal cells. The null distribution of expression in each region is Gaussian, with the expression mean and variance taken from matching regions in a set of reference normal cells. The algorithm performs the following steps:

1. Identify chromosomal regions that are copy number neutral across all samples as a normalization factor.

2. Identify chromosomal regions that are copy number–altered in at least one sample based on bulk WES data.

3. Aggregate expression across genes within each altered region.

4. Normalize and log-transform the per-region expression.

5. Determine the null distribution based on annotated nontumor cells.

6. Perform a hypothesis test to indicate the presence or absence of an alteration.

IntegrateCNV allows us to better normalize single-cell expression data against neutral regions without removing the biological signal inherent in library size.

Determining neutral and altered regions

The first input to integrateCNV is a set of copy-number profiles derived from bulk DNA sequencing. For each sample, we use Fraction and Allele-Specific Copy Number Estimates from Tumour Sequencing (FACETS; RRID:SCR_026264; ref. 11) to identify the total copy number in each region. CNAs are centered around 0 so that a neutral region is represented by the copy number “0.” The CNA profiles are saved as BED files, containing, for each region, information about the chromosome, start position, end position, and copy number. BED files from all samples are processed to find intersecting genomic regions using the multi_intersect function from pybedtools (RRID:SCR_021018). The resulting intersections capture the chromosomal regions and CNAs in each sample. Neutral regions are then identified as those with no CNA in any sample. We denote the set of neutral regions by A⁰.

Candidate altered regions are first identified as those in which at least one sample contains an alteration. Of the candidate regions, only those containing sufficient genes (>20 by default) are retained for downstream analysis to provide sufficient coverage to reliably recover copy numbers without the potential effects of a few outlier genes. This set of altered regions, A²⁰⁺, is used as the set of regions within which we will infer CNAs.

Processing count data

We denote the scRNA-seq cell $\mathrm{\times }$ gene count matrix by $X$, where $X_{c\mathrm{,}g\mathrm{\hspace{0.17em}}}$ represents the expression of gene $g$ for cell $c\mathrm{\hspace{0.17em}}\mathrm{=}\mathrm{\hspace{0.17em}}1\mathrm{,}\mathrm{\hspace{0.17em}}\mathit{\cdots }\mathrm{,}\mathrm{\hspace{0.17em}}n\mathrm{.}$ Using the set of candidate regions (A²⁰⁺), we aggregate counts over genes within a given region, indexed by $r$, to determine a cell $\mathrm{\times }$ region matrix, $U\mathrm{.}$

$$U_{c\mathrm{,}r}\mathrm{=}\mathrm{\hspace{0.17em}}{\displaystyle \mathrm{\sum }}_{g\mathrm{\in }{G}_r}{X}_{c\mathrm{,}g}$$

where $g\mathrm{\hspace{0.17em}}\mathit{\in }{G}_r$ are the genes that physically overlap with the genomic region indexed by $r$.

The counts from regions A⁰ that are found to be neutral in all samples are used as a pseudo “spike-in” control in order to normalize count data without removing the biological signal of total library size, which can correlate with copy-number burden. The total counts from genes across all neutral regions are summed for each cell, $c$, and the sum is denoted by the library size normalization factor, $l_c$.

$$l_c\mathrm{\hspace{0.17em}}\mathrm{=}\mathrm{\hspace{0.17em}}{\displaystyle \mathrm{\sum }}_{r\mathit{\in }{A}^0}{\displaystyle \mathrm{\sum }}_{g\mathit{\in }{G}_r}{X}_{c\mathrm{,}g}$$

The cell $\mathrm{\times }$ region matrix, $U$, is then divided by the library size normalization factor, and the $\mathit{log}$ of the resulting normalized expression is computed to give the data matrix, $V$

$$V_{c\mathrm{,}r}\mathrm{=}\mathrm{\hspace{0.17em}}\mathit{log}\frac{1}{l_c}{U}_{c\mathrm{,}r}$$

Inferring CNAs and extracting integer copy-number calls

The log-normalized expression matrix, restricted to normal cells, now defines a null Gaussian distribution on expression levels in unaltered cells for each region. For each cell and region, the $z$-score is computed using this null distribution and is used to define copy number–altered regions.

Finally, a two-tailed hypothesis test is performed for each (cell, region) pair to determine whether the cell has expression values significantly higher or significantly lower than expected in a diploid cell. A P value threshold (default 0.05) is used to determine the critical values for two-tailed hypothesis testing. All regions above or below the upper or lower critical values, respectively, are called alterations. We note that because deleted regions have a small dynamic range (0, 1, or 2), there is less power to detect them, and thus, the procedure results in many false negatives for deleted regions. For all called alterations, we use the copy number from the corresponding bulk WES sample to insert an integer copy number. This denoising procedure ensures that, for a region to be denoted as altered in a cell, it must be supported by evidence from both snRNA-seq and sample-level bulk DNA data. The final output matrix is an integer copy-number profile for each cell and can be used for downstream phylogenetic analysis of clonal relationships.

Comparison of CNA inference methods

To benchmark integrateCNV against existing approaches that infer CNAs from scRNA-seq data, we first determined single-cell copy-number z-score profiles, which are computed without any prior knowledge of which sample (or bulk WES data) each cell belongs to. We then aggregated cells within samples to compare against the “ground truth” bulk WES copy-number profile.

For each sample, we ran inferCNV (RRID:SCR_021140; ref. 12) and CopyKAT (RRID:SCR_024512; ref. 13), which return per-region and per-gene CNA scores, respectively, for each cell in the sample. We also ran Numbat (14) both with and without bulk copy-number profiles per sample as input to the algorithm. Numbat performed best with bulk profiles provided, and thus, these single-cell copy-number profiles were used for comparisons. For integrateCNV, we computed the z-score for each altered region harboring an alteration in each cell. The z-scores are computed per cell in a sample-agnostic manner so that no sample-identifying information is provided to integrateCNV. We then computed the average score across all cells in the sample to determine a pseudo-bulk CNA score for each method. As all methods return continuous-valued predictions of alterations rather than discrete copy-number calls, we computed the correlation between the bulk DNA CNA call and the pseudo-bulked inferred CNA score.

Identifying recurrent CNAs

We use the four gamete test (15) to identify potential violations of the infinite sites assumption that may be due to recurrent alterations. The four gamete test considers mutation states at pairs of sites. We binarize CNAs, representing diploid sites as 0 and aneuploid sites as 1. For any two sites in a sequence, there are four possible combinations of mutation states: (1,1), (1,0), (0,1), and (0,0). If all four combinations are observed in a population, this violates the infinite sites model (which assumes that each mutation only occurs once).

For each pair of regions for which single-cell copy-number profiles were computed by IntegrateCNV, we identify all pairs of mutation states that are observed in our inferred CNA profiles. To account for noise in the copy-number inference, we consider only pairs that are represented in at least 100 cells. If all four mutation state pairs are observed, we denote that region pair as violating the infinite sites assumption, likely due to recurrent CNAs.

Phylogenetic inference from single-cell CNA calls

Most efforts to reconstruct tumor phylogenies rely on single-nucleotide variants (SNV) derived from DNA sequencing data. A few approaches specifically address CNA phylogenies (16, 17), but they are designed for copy-number profiles derived from deconvolved bulk DNA sequencing or single-cell DNA sequencing. These methods typically assume that input copy-number profiles are reliable and accurately specified for contiguous genomic regions, and most do not scale to large numbers of cells. These assumptions do not hold when considering CNA profiles derived from scRNA-seq experiments, as inferred copy-number profiles are very noisy and dataset sizes are significantly larger. Researchers thus often resort to distance-based agglomerative clustering methods such as neighbor joining to reconstruct cell hierarchies.

To overcome these challenges, we developed PICASSO to infer cellular clones and their phylogenetic relationships from CNA calls derived from single-cell expression data. The PICASSO algorithm assumes that observed single-cell copy-number profiles are noisy measurements of true clonal profiles such that cells in the same clone share similar CNA patterns. Phylogenetic relationships are unobserved and result from (potentially recurrent) gain and loss of copy-number variants from an original parent clone. PICASSO thus aims to group single cells based on membership to inferred clones and determine the evolutionary relationships between these clones.

As input, PICASSO accepts a character matrix of cells by regions, with each entry consisting of an integer CNA state for the corresponding region and cell. Using this information, the algorithm generates (i) assignments of cells to clones and (ii) a phylogeny describing the relationship between clones.

PICASSO algorithm

PICASSO is a tree-recursive algorithm, whereby each iteration considers the cells currently assigned to a leaf node of the phylogenetic tree and determines whether to split that leaf into further branches. It comprises the following steps:

1. Encode integer copy numbers into ternary profiles. If the maximum absolute copy number (relative to diploid) is $j$, copy number $k$ is encoded as a vector of length $j$ with $k$ leading 1s so that similar copy-number profiles are similar in the encoded space. In practice, we cap the maximum copy number at $j\mathrm{\hspace{0.17em}}\mathrm{=}\mathrm{\hspace{0.17em}}2$, distinguishing only between amplified and highly amplified copy numbers. Similarly, negative copy number $\mathrm{-}k$ is encoded as a vector of length $j$ with $k$ leading −1s. This allows us to represent the cumulative nature of CNAs, whereby moderate gains or losses may precede more severe alterations and also account for small mistakes when inferring CNA magnitude.

2. Construct an initial phylogeny comprising a single leaf node containing all cells in the dataset.

3. For each leaf node, split the node into two clones based on shared CNAs using expectation–maximization (EM). Cells are partitioned such that (i) CNAs are allowed to recur independently in distinct clones and (ii) cells are grouped based on the global CNA profile, mitigating the outsized effect of noisy or incorrect calls in a few genomic regions.

   More explicitly, for each nonterminal leaf in the phylogeny, perform the following steps:

   1. If sufficient evidence exists to split cells, assign cells to one of two subclones using EM. These subclones are the new children of the original leaf node.

   2. If insufficient evidence exists to split cells, designate this leaf as a terminal node.

   3. Repeat until all leaf nodes are terminal nodes.

4. Cell groupings identified from this iterative process constitute clone assignments, and relationships between groups constitute the phylogenetic relationships between clones. The tree is rerooted so that the clone with the fewest CNAs is the most ancestral, reflecting the fact that CNA burden generally increases during evolutionary progression.

5. As optional postprocessing, we may collapse small subclones containing too few cells to draw meaningful statistical conclusions.

Encoding the character matrix

We denote the cell $\mathrm{\times }$ region matrix of integer CNAs by $B$, where $B_{c\mathrm{,}r}$ represents the inferred copy number of cell $c\mathrm{\hspace{0.17em}}\mathrm{=}\mathrm{\hspace{0.17em}}1\mathit{\cdots }n$ in region $r$ in $A^{20\mathrm{+}}$ and $A^{20\mathrm{+}}$ denotes the set of altered regions. To facilitate further analysis, we transform matrix $B$ into a matrix $M$ using the following encoding scheme:

1. Determine the maximum absolute value.

   For each column (region) $r$ in $B$, determine the maximum absolute value, $p_r$, representing the highest CNA observed in that region. In practice, we cap this value at copy number +2 (two copies more than expected in a diploid cell), as we may not trust or be able to reliably distinguish between very large copy numbers.

2. Encode copy-number values.

   For each cell $c$ and region $r$, encode the copy number $k\mathrm{\hspace{0.17em}}\mathrm{=}{B}_{c\mathrm{,}r}$ into $p_r$ columns in $M$ according to this scheme:

   1. If $k\mathrm{\hspace{0.17em}}\mathit{\ge }\mathrm{\hspace{0.17em}}0$, the encoding is $\left[1\mathrm{,}\mathrm{\hspace{0.17em}}1\mathrm{,}\mathrm{\hspace{0.17em}}\mathit{\dots }\mathrm{,}\mathrm{\hspace{0.17em}}1\mathrm{,}\mathrm{\hspace{0.17em}}0\mathrm{,}\mathrm{\hspace{0.17em}}0\mathrm{,}\mathrm{\hspace{0.17em}}\mathit{\dots }\mathrm{,}\mathrm{\hspace{0.17em}}0\right]$ with $k$ ones followed by $p_r\mathrm{-}k$ zeros.

   2. If $k\mathrm{\hspace{0.17em}}\mathrm{<}\mathrm{\hspace{0.17em}}0$, the encoding is $\left[\mathit{–}1\mathrm{,}\mathrm{\hspace{0.17em}}\mathit{–}1\mathrm{,}\mathrm{\hspace{0.17em}}\mathit{\dots }\mathrm{,}\mathrm{\hspace{0.17em}}\mathit{–}1\mathrm{,}\mathrm{\hspace{0.17em}}0\mathrm{,}\mathrm{\hspace{0.17em}}0\mathrm{,}\mathrm{\hspace{0.17em}}\mathit{\dots }\mathrm{,}\mathrm{\hspace{0.17em}}0\right]$ with $\left|k\right|$ negative ones followed by $p_r\mathrm{-}\mathrm{\hspace{0.17em}}\left|k\right|$ zeros.

3. Construct the matrix $M$.

Replace each column $r$ in $B$ with $p_r$ columns in $M$ according to the above encoding scheme, resulting in a ternary matrix in which each original region is expanded into multiple columns representing CNA magnitude and direction.

This transformation allows us to enforce similar copy-number profiles between CNAs of similar values. The dimension of the resulting matrix $M$ is n$\mathrm{\hspace{0.17em}}\mathrm{\times }{\displaystyle \mathrm{\sum }}_{r\mathit{\in }{A}^{20\mathrm{+}}}{p}_r$.

Top-down phylogeny construction

We use an EM approach to construct a top-down phylogenetic tree based on shared patterns of copy-number breakpoints. The phylogeny is initialized with a single clone containing all the cells in the dataset. At each iteration, the depth of the existing tree may be increased by one as each leaf clone may be split into two further subclones if there is sufficient evidence of differences between them. Sufficient evidence of differences between potential subclones exists when the copy-number patterns observed cannot be reasonably explained by a single population. Using the Bayesian information criterion (BIC), we only create a new branch in the evolutionary tree when the data strongly suggest that two distinct copy-number clone populations exist. Alternatively, any given clone may remain intact as a terminal clone.

Mixture model for clustering CNA clones

The input to PICASSO is the copy-number profile of distinct genomic regions that are likely to harbor CNAs. We therefore assume that CNA occurrences at each genomic region are independent, which allows us to consider each profile as a draw from a multivariate categorical mixture model. We can use an EM algorithm to cluster each existing leaf into two subclones, mimicking the evolutionary process that distinguishes clones by the accumulation of copy-number differences.

For each subclone, we learn a probabilistic profile over CNAs, allowing us to capture several essential features. The learned probability associated with the categorical distribution for a given CNA can be less than 1, permitting CNAs to be present in only a subset of cells in an inferred subclone. Furthermore, the subclonal structure can be disentangled by subclone splitting in subsequent iterations. Additionally, the probabilistic profile allows us to model the high degree of false positives and false negatives in inferred CNA data by tolerating small probabilities of a clone missing or containing a specific alteration. Finally, there may be a positive probability of a particular alteration occurring at the same position in both clones, allowing for the independent recurrence of copy-number changes in multiple clonal lineages, which has been observed extensively in previous CNAs of cancer data (18).

The EM algorithm is a widely used iterative method to find maximum likelihood estimates of parameters in probabilistic models, particularly for clustering problems. PICASSO uses an EM algorithm for clustering categorical data with states $\left\{\mathit{–}1\mathrm{,}0\mathrm{,}1\right\}\mathrm{,}$ which represent different CNAs in cells.

The observed copy-number profiles, $M\mathrm{\hspace{0.17em}}\mathrm{=}\mathrm{\hspace{0.17em}}\left\{M_1\mathrm{,}{M}_2\mathrm{,}\mathrm{\hspace{0.17em}}\mathit{\dots }\mathrm{,}\mathrm{\hspace{0.17em}}{M}_n\right\}$, contain the encoded CNAs for each cell, $c\mathrm{\hspace{0.17em}}\mathrm{=}\mathrm{\hspace{0.17em}}1\mathrm{\hspace{0.17em}}\mathit{\cdots }\mathrm{\hspace{0.17em}}n$, across regions. Each $M_c\mathrm{=}\mathrm{\hspace{0.17em}}\left[m_{c1}\mathrm{,}{m}_{c2}\mathrm{,}\mathrm{\hspace{0.17em}}\mathit{\dots }\mathrm{,}\mathrm{\hspace{0.17em}}{m}_{cd}\right]$ is a vector of $d$ categorical observations for cell $c$. Each observation $m_{cj}$ can take one of three states: −1, 0, or 1, representing different CNAs.

We assume there are two clusters representing an evolutionary split between subclones, and each cluster $k$ is characterized by a set of parameters ${\theta }_k\mathrm{=}\mathrm{\hspace{0.17em}}\left\{{\pi }_k\mathrm{,}\mathrm{\hspace{0.17em}}{\varphi }_k\right\}$, where ${\pi }_k$ parametrizes the prior probability of cluster $k$ and ${\varphi }_k$ the probability distribution over the states for each observation in cluster $k$.

EM algorithm

The EM algorithm iterates between the expectation (E) and maximization (M) steps until convergence. The goal is to assign each cell to one of the subclones in a way that maximizes the likelihood of the observed data.

We let ${\varphi }_k\mathit{\in }{\mathbb{R}}^{3\mathrm{\times }d}$ represent the parameters of the categorical distribution for component $k\mathrm{\hspace{0.17em}}\mathit{\in }\left\{1\mathrm{,}2\right\}$ and ${\pi }_k$ represent the mixture proportions. We also define the latent variable $z_c$, which indicates the membership of the $c$-th observation to one of the two components, where $z_c\mathit{\in }\left\{1\mathrm{,}2\right\}$. The responsibility ${\gamma }_{ck}\mathrm{=}\mathbb{E}\left[z_{ck}\right]$ is the expectation of $z_{ik}\mathrm{.}$

The complete data log-likelihood is as follows:

$$\mathit{log}\mathrm{\hspace{0.17em}}p\left(M\mathrm{,}Z\mathrm{\hspace{0.17em}}\mathrm{|}\mathrm{\hspace{0.17em}}\pi \mathrm{,}\mathrm{\hspace{0.17em}}\varphi \right)\mathrm{=}\mathrm{\hspace{0.17em}}{\displaystyle \mathrm{\sum }}_{c\mathrm{=}1}^n{\displaystyle \mathrm{\sum }}_{k\mathrm{=}1}^2{z}_{ck\mathrm{\hspace{0.17em}}}\left(\mathit{log}{\pi }_k\mathrm{\hspace{0.17em}}\mathrm{+}{\displaystyle \mathrm{\sum }}_{j\mathrm{=}1}^d\mathrm{\hspace{0.17em}}\mathit{log}{\varphi }_{k\mathrm{,}\mathrm{\hspace{0.17em}}{m}_{cj}\mathrm{,}j}\right)$$

The E step updates the prediction of which subclone each cell belongs to based on the likelihood of the observed data under the current model. We calculate the posterior probabilities, ${\gamma }_{ck}$, also known as responsibilities, which represent the probability that each cell $c$ belongs to each cluster $k$.

The M step uses the assignment probabilities calculated in the E step to update the model parameters. Specifically, we adjust the subclone priors ${\pi }_k$ and the categorical distribution parameters ${\varphi }_k\mathit{\in }{\mathbb{R}}^{3\mathrm{\times }d}$ to maximize the expected log-likelihood of the observed data, weighted by the assignment probabilities. The categorical distribution parameters ${\varphi }_k\mathit{\in }{\mathbb{R}}^{3\mathrm{\times }d}$ for clone $k$ represent the probability of observing each (copy-number state, encoded region) pair. This step ensures that the parameters better reflect the observed data given the current cluster assignments.

By iteratively updating the assignment probabilities in the E step and the model parameters in the M step, the EM algorithm gradually converges to a set of parameters that maximize the likelihood of the data. This iterative process allows the algorithm to find the most probable clustering of the cells based on shared CNA patterns.

Initialization. We begin by randomly initializing the subclone assignments, ${\gamma }_{ck}$, of each cell so that cells are distributed randomly between clones. In order to mitigate the effect of local minima when performing this iterative optimization, we perform five random restarts and select the model that has the highest likelihood among the five trials.

E step. To determine the optimal assignment of cells to subclones, we compute the posterior probabilities (responsibilities) that each cell $M_c$ belongs to subclone $k$:

$${\gamma }_{ck}\mathrm{\leftarrow }\frac{{\pi }_k{\displaystyle \mathrm{\prod }}_{j\mathrm{=}1}^d{\varphi }_k\left(m_{cj}\right)}{{{\displaystyle \mathrm{\sum }}}_{l\mathrm{=}1}^2{\pi }_l{\displaystyle \mathrm{\prod }}_{j\mathrm{=}1}^d{\varphi }_l\left(m_{cj}\right)}$$

where ${\varphi }_k\left(m_{cj}\right)$ is the probability of observing $m_{cj}$ in cluster $k$.

M step. To update the probabilistic subclone profiles, we update the parameters ${\pi }_k$ and ${\varphi }_k$ to maximize the expected log-likelihood:

$${\pi }_k\mathrm{\hspace{0.17em}}\mathrm{\leftarrow }\mathrm{\hspace{0.17em}}\frac{1}{n}\mathrm{\hspace{0.17em}}{\displaystyle \mathrm{\sum }}_{c\mathrm{=}1}^n{\gamma }_{ck}$$

$${\varphi }_k\left(z\right)\mathrm{\hspace{0.17em}}\mathrm{\leftarrow }\frac{{{\displaystyle \mathrm{\sum }}}_{c\mathrm{=}1}^n{{\displaystyle \mathrm{\sum }}}_{j\mathrm{=}1}^d{\gamma }_{ck}\delta \left(m_{cj}\mathrm{,}z\right)}{{{\displaystyle \mathrm{\sum }}}_{c\mathrm{=}1}^n{{\displaystyle \mathrm{\sum }}}_{j\mathrm{=}1}^d{\gamma }_{ck}}$$

where $z$ is a copy-number state (−1,0,1) being updated and $\delta \left(a\mathrm{,}b\right)$ is the Kronecker delta function, which is 1 if $a\mathrm{=}b$ and 0 otherwise.

Termination of subclone splitting

We implement two methods to determine whether a clone should be split further. The first (and preferred) option compares the BIC score of a model with one clone to that of a model with two clones and terminates the splitting process if the BIC score does not improve with two clones. Specifically, we calculate as follows:

$$BIC\mathrm{\hspace{0.17em}}\mathrm{=}\mathrm{\hspace{0.17em}}\mathrm{-}2\mathrm{\hspace{0.17em}}\mathit{ln}\left(L\right)\mathrm{\hspace{0.17em}}\mathrm{+}\mathrm{\hspace{0.17em}}k\mathrm{\hspace{0.17em}}\mathrm{\times }\mathit{ln}\left(n\right)$$

where $L$ is the maximum likelihood, $k$ is the number of parameters in the model, and $n$ is the number of cells. When splitting a clone into two subclones, the model gains additional parameters (new probabilistic profiles and mixing proportions), which incurs a penalty term in the BIC calculation. Only when the improvement in likelihood outweighs this complexity penalty do we proceed with the split. This approach rigorously controls model complexity by requiring substantial evidence that observed variations reflect genuine biological differences rather than stochastic noise.

In cases with limited cell numbers, the statistical power needed for BIC to detect meaningful biological differences may be insufficient. The cell assignment confidence approach provides a complementary criterion that can identify biologically relevant subpopulations even when BIC would prematurely terminate splitting, making it particularly valuable for datasets with fewer cells or more subtle clonal differences.

The second method relies on cell assignment confidence. Using the responsibilities matrix from the EM algorithm, we check the proportion of confidently assigned cells. Specifically, if a cell’s responsibility value exceeds a user-defined threshold (e.g., 0.75), it is considered confidently assigned. If the proportion of confidently assigned cells falls below a user-defined threshold (typically 0.6–0.8), the splitting process is terminated. This ensures that further subdivisions are only made when cells show clear membership patterns, avoiding overfitting to noisy data.

Postprocessing subclones

Inference from scRNA-seq data produces inherently noisy copy-number profiles due to technical limitations in the sequencing process. These profiles may contain artifacts and false signals that can lead to the detection of spurious subclones. To ensure the reliability of our phylogenetic analysis, we implement a postprocessing step that retains only those clones with sufficient statistical support and biological plausibility, filtering out clusters that likely arise from technical noise rather than true clonal evolution.

In order to mitigate the occurrence of clones derived from noise in the copy-number inference process, we require clones to (i) be composed of more than 75 cells and (ii) contain at least one CNA at high frequency. We selected a conservative threshold of 75 cells as a minimum clone size to ensure that clones are likely to represent true biological subpopulations rather than technical artifacts arising from the copy-number inference process.

For a given clone, we define high-frequency CNAs as alterations present in at least 80% of the cells in that clone. The requirement for at least one high-frequency CNA provides additional confidence that the identified clone represents a genuine biological subpopulation with shared genomic alterations.

Clones that do not satisfy these conditions are removed from the phylogenetic analysis as we do not have sufficient confidence to draw conclusions about the cells they contain.

PICASSO benchmarking

To evaluate PICASSO’s phylogenetic reconstruction accuracy, we simulated a series of ground truth CNA trees. Existing single-cell phylogenetic algorithms are not well suited to constructing clone trees from noisy, large-scale datasets. For example, CNETML (17), a maximum likelihood algorithm for deriving phylogenies from copy-number profiles, only scales to the low hundreds of cells. We thus compared our ability to recover phylogenetic relationships in these simulations with an agglomerative tree-building algorithm, neighbor joining.

Simulation experiments

We start by generating random binary trees that form the backbone of our CNA simulations, providing a structure on which we can model evolutionary relationships. Each leaf in the tree represents a copy-number clone, and branches depict the divergence of clonal lineages over time. Next, we annotate these trees with regions and alterations using a Dirichlet distribution to generate probability vectors for region selection. This distribution allows us to model the relative likelihood of alterations occurring across different genomic regions and capture the biological reality that some regions are more susceptible to CNAs than others.

Each branch is assigned specific alterations (values of −2, −1, +1, or +2) based on a predefined probability distribution [0.5, 0.3, 0.2] that determines how many alterations will occur per branch (with a 0.5 probability of one alteration, a 0.3 probability of two alterations, and a 0.2 probability of three alterations), reflecting the accumulation of genetic changes as cells evolve. The actual alterations themselves are randomly selected from the set (−2, −1, +1, +2) with equal probability.

To capture the cumulative effect of these alterations, we calculate the aggregated alterations for each leaf node by tracing the path from the root to the leaf. This gives us a comprehensive copy-number profile for each clone, accounting for all the genetic changes that occurred along its lineage. Cells are then attached to the leaves of the tree, with the number of cells per clone partially determined by the distribution of clone sizes observed in the data. Specifically, we leverage real-world PDAC data, using half the actual observed clone sizes to balance computational efficiency with biological fidelity while preserving the relative proportions of clonal populations seen in patient samples.

In order to simulate realistic copy-number profiles inferred from single-cell data, it is essential to introduce realistic noise, including extensive false positives and false negatives:

1. Add noise to neutral regions (false positives). Simulate false-positive inferred CNAs by randomly selecting a proportion of neutral (no copy-number change) regions within the cell profiles and applying random alterations. We perform these simulations across four false-positive-rate parameter regimes: The false-positive rate (0.01, 0.1, 0.2, or 0.3) directly determines the proportion of neutral regions altered—for example, at a rate of 0.1, 10% of neutral regions receive artificial alterations. The magnitude of these alterations follows a distribution derived from observed alterations to ensure realistic noise patterns.

2. Zero out existing alterations (false negatives). Simulate false negatives or loss of signal by zeroing out existing alterations in the cell profiles randomly. The false-negative rate directly determines the probability of removing each existing alteration—for example, at a rate of 0.2, each real alteration has a 20% chance of being removed. This stochastic process simulates scenarios in which genuine copy-number changes go undetected.

3. Perturb existing alterations. Simulate CNAs whose presence is correctly inferred, but whose magnitude is not, by slightly increasing or decreasing copy-number values. We introduce magnitude perturbations with a probability of 0.1 per alteration, randomly adjusting values by +1 or −1 while preserving the direction (gain or loss). This simulates measurement uncertainty in copy-number estimation from sequencing data. These perturbations create a consistent baseline of noise across all experimental conditions, independent of the varying false-positive and false-negative rates being tested, better reflecting the technical challenges in precise CNA quantification.

We conducted simulation experiments with three replicates across multiple parameter configurations. Each simulation maintained 60 leaves and 110 regions, dimensions comparable with the PDAC tree inferred by PICASSO. By systematically varying false-positive and false-negative rates (0.01, 0.1, 0.2, and 0.3), we comprehensively evaluated the robustness of both neighbor joining and PICASSO methods under increasingly challenging conditions of data quality.

Metric for evaluating PICASSO

We evaluated PICASSO and neighbor joining phylogenies using the triplets correct metric (19), which assesses the tree’s ability to reconstruct correct phylogenetic relationships between triplets of cells. For each simulated tree, we sample 10,000 triplets $\left(a\mathrm{,}b\mathrm{,}c\right)$. For each triplet, the ground truth tree induces a phylogenetic ordering on the cells. For example, for triplet $\left(a\mathrm{,}b\mathrm{,}c\right)$, the ground truth phylogenetic relationship of these cells may be $\left(\left(a\mathrm{,}b\right)\mathrm{,}c\right)$, indicating that cells $a$ and $b$ share a more recent common ancestor than $a$ and $c$ or $b$ and $c$. In an inferred tree, the triplet is scored as “correct” if the phylogenetic relationship between these cells is accurately recovered.

As the simulated tree defines leaves as “clones” (groups of cells that cannot be distinguished from each other by CNA profile), some triplets will have no clear phylogenetic relationship; they are siblings in a clone. Unlike PICASSO, neighbor joining computes a fully resolved cell tree. Therefore, when computing the proportion of triplet relationships that are correctly determined, we only consider triplets with clearly defined phylogenetic relationships. By counting the proportion of correctly inferred triplets, the triplets correct metric provides a quantitative measure of the tree’s accuracy, helping to identify discrepancies and assess the overall quality of the inferred phylogenetic tree.

PICASSO runtime and memory comparison

Given the large size of scRNA-seq datasets, runtime complexity is a significant concern. The neighbor-joining algorithm, commonly used in phylogenetic analysis, has a theoretical runtime complexity of O(n³), where n is the number of cells although some implementations of neighbor joining use heuristics to improve performance in practice (20).

We evaluated run times on simulated datasets with 20,000 cells. Given the large size of the datasets, we used a heuristic implementation of neighbor joining, rapidNJ. We measured the runtimes for both neighbor joining and PICASSO on each dataset across all replicates and found that PICASSO is significantly faster and less memory-intensive than neighbor joining.

PICASSO robustness testing

To evaluate the robustness and reproducibility of PICASSO, we ran the method five times on the full PDAC dataset of approximately 40,000 single cells and assessed the consistency of the resulting phylogenetic reconstructions. Pairwise comparisons of the clone assignments across runs were quantified using normalized mutual information (NMI) and adjusted Rand index (ARI), widely used metrics for comparing the similarity between two clustering assignments.

NMI measures the mutual information shared between two clustering assignments, normalized to a 0 to 1 scale:

$$NMI\left(U\mathrm{,}V\right)\mathrm{\hspace{0.17em}}\mathrm{=}\mathrm{\hspace{0.17em}}2\mathrm{\hspace{0.17em}}\mathrm{\times }\frac{MI\left(U\mathrm{,}V\right)}{H\left(U\right)\mathrm{\hspace{0.17em}}\mathrm{+}\mathrm{\hspace{0.17em}}H\left(V\right)}$$

where $MI\left(U\mathrm{,}V\right)$ is the mutual information between clustering assignments $U$ and $V$ and $H\left(U\right)$ and $H\left(V\right)$ are their respective entropies. An NMI score of 1 indicates perfect agreement between clusterings, whereas 0 indicates completely independent clusterings. The NMI scores we observed were consistently high, averaging around 0.85, indicating strong agreement in the overall clustering structure across runs.

ARI measures the similarity between two clustering assignments by counting pairs of elements that are either assigned to the same cluster or different clusters in both assignments, adjusted for chance:

$$ARI\mathrm{\hspace{0.17em}}\mathrm{=}\mathrm{\hspace{0.17em}}\frac{RI\mathrm{\hspace{0.17em}}\mathrm{-}\mathrm{\hspace{0.17em}}E\left(RI\right)}{\mathit{max}\left(RI\right)\mathrm{\hspace{0.17em}}\mathrm{-}\mathrm{\hspace{0.17em}}E\left(RI\right)}$$

where $RI$ is the raw Rand index, $E\left(RI\right)$ is the expected raw RI, and max(RI) represents the theoretical maximum value the RI could achieve for the given clustering problem. The raw RI is defined as follows:

$$RI\mathrm{\hspace{0.17em}}\mathrm{=}\mathrm{\hspace{0.17em}}\frac{TP\mathrm{\hspace{0.17em}}\mathrm{+}\mathrm{\hspace{0.17em}}TN}{TP\mathrm{\hspace{0.17em}}\mathrm{+}\mathrm{\hspace{0.17em}}TN\mathrm{\hspace{0.17em}}\mathrm{+}\mathrm{\hspace{0.17em}}FP\mathrm{\hspace{0.17em}}\mathrm{+}\mathrm{\hspace{0.17em}}FN}$$

where $TP$ is the number of pairs that are in the same cluster in both clusterings, $TN$ is the number of pairs in different clusters in both clusterings, $FP$ is the number of pairs that are in the same cluster in the first clustering but in different clusters in the second, and $FN$ is the number of pairs that are in different clusters in the first clustering but in the same cluster in the second.

ARI ranges from −1 to 1, with 1 indicating perfect agreement, 0 indicating random cluster assignments, and negative values indicating worse-than-random agreement. The observed ARI, which is sensitive to both the number and composition of clusters, averaged around 0.6, reflecting a reasonable level of stability given the complexity of the dataset and the stochastic nature of the method.

To further assess consistency at the phylogenetic level, we computed the proportion of triplets (evolutionary relationships between three cells) recovered in each run that matched those identified in a separate run designated as the ground truth. High concordance of triplets across runs demonstrates that PICASSO reliably reconstructs phylogenetic relationships despite inherent variability in clustering.

To further assess the robustness of PICASSO, we conducted a downsampling analysis by randomly subsampling the dataset to 75%, 80%, 85%, and 90% of the original dataset. For each downsampled dataset, we ran PICASSO and measured the proportion of triplets in the reconstructed phylogenies that matched the triplets identified in the full dataset, which served as the reference. Across all levels of downsampling, the proportion of correctly recovered triplets remained high, demonstrating the method’s robustness.

Computational analysis

Digital histopathology

Whole slide imaging data were obtained with the assistance of the Molecular Cytology Core Facility at MSKCC. H&E-stained slides were scanned using a Pannoramic scanner (3DHISTECH) equipped with a 20×/0.8 NA objective. The resulting data were analyzed using QuPath (version 0.5.1; RRID:SCR_018257; https://qupath.github.io/).

Adipose and fibrous tissues were annotated by a pathologist using QuPath. Following annotation, the Pixel Classification tool in QuPath was applied with default settings to quantify the areas of adipose and fibrous tissues.

For cell type evaluation, QuPath’s Cell Detection tool was used to identify and analyze tumor, stromal, and immune cells. Regions containing these three cell types were annotated, and only tumor-cell-containing areas were included in the analysis. A cell classifier was trained using the Object Classification tool in QuPath with default settings, based on pathologist annotations. Features such as nuclear circularity and eccentricity were calculated to characterize the detected cells. The classifications were validated by the annotating pathologist to ensure accuracy.

WES data analysis

WES data preprocessing

Initial processing began with adapter trimming of FASTQ files using cutadapt (version 1.9.1; RRID:SCR_011841) to remove standard Illumina 5′ and 3′ adapter sequences. The trimmed reads were then mapped to the b37 reference genome from the Broad GATK resource bundle using BWA-MEM (version 0.7.12; RRID:SCR_010910). Postalignment processing included sorting of SAM files and addition of read group tags using PICARD tools (version 1.124; RRID:SCR_006525). The read group information includes sample identifiers, sequencing library identifiers, and Illumina platform information. The sorted BAM files were then processed with Picard MarkDuplicates to identify PCR duplicates (https://github.com/soccin/BIC-variants_pipeline).

CNA calling

CNAs in solid tumors were computed from tumor and matched normal tissue WES data using default settings in the FACETS (version 0.6.2) algorithm (https://github.com/mskcc/facets-suite; ref. 11). FACETS provides allele-specific copy-number estimates at the level of both gene and chromosome arm.

SNV calling

We used the standardized Illumina (HiSeq) Exome Variant Detection Pipeline to detect variants in the output of preprocessed WES data. Following duplicate marking, BAM files are processed according to GATK (version 3.4-0; RRID:SCR_001876) best practices version 3 for tumor–normal pairs. This includes local realignment using Abra (version 2.17; SCR_003277) with default parameters, followed by base quality score recalibration using BaseQRecalibrator with known variants from the Broad GATK B37 resource bundle, including the Single Nucleotide Polymorphism Database (version 138; RRID:SCR_002338).

Somatic variant calling is performed using muTect (version 1.1.7; RRID:SCR_000559) with default parameters for SNV detection, whereas somatic indels are identified using GATK HaplotypeCaller with subsequent custom postprocessing. A final “fill-out” step computes the complete read depth information at each variant position across all samples using the realigned BAMs. This step applies quality filters requiring mapping quality ≥20 and base quality ≥0, with no filtering for proper read pairing.

All analyses were performed using a standardized computational environment managed through Singularity (version 2.6.0). The complete pipeline source code, including all postprocessing scripts, is available at the following links:

• https://github.com/soccin/BIC-variants_pipeline

• https://github.com/soccin/Variant-PostProcess

Additional software versions used in the pipeline include Perl (version 5.22.0; RRID:SCR_018313), Samtools (version 1.2; RRID:SCR_002105), VCF2MAF (version 1.6.21; RRID:SCR_027063), and VEP (version 102; RRID:SCR_007931).

SNV and CNA visualization

To visualize the SNV and CNA status of key cancer genes, as well as tumor mutation burden, we used CoMut (21).

snRNA-seq data preprocessing

After quality control (see next section), snRNA-seq generated a total of 73,142 high-quality transcriptomes from 11 samples (Supplementary Table S1).

Alignment of sequencing reads

All scRNA-seq samples were preprocessed as follows: FASTQ files from the rapid autopsy samples were processed with the SEQC (version 0.2.4) pipeline (https://github.com/dpeerlab/seqc; ref. 22) using the hg38 human genome reference, default parameters, and platform set to 10x Genomics v3 3′ scRNA-seq kit. The SEQC (version 0.2.4) pipeline performs read demultiplexing, alignment, and unique molecular identifier (UMI) and cell barcode correction, producing a preliminary count matrix of cells by unique transcripts. By default, the pipeline will remove putative empty droplets and poor-quality cells based on (i) the total number of transcripts per cell (cell library size), (ii) the average number of reads per molecule (cell coverage), (iii) mitochondrial RNA content, and (iv) the ratio of the number of unique genes to library size (cell library complexity).

Nuclear transcriptomes from human rapid autopsy samples are expected to have lower RNA content and quality than regular single-cell assays (23). To obtain a more comprehensive representation of cancer phenotypes, we included both FACS and nonsorted samples (see snRNA-seq section); however, nonsorted samples carry a greater degree of low-quality nuclei. Therefore, due to the intrinsic lower RNA content and sample quality of flash-frozen snRNA-seq–derived transcriptomes, we performed further quality control steps as described in the following sections.

snRNA-seq data quality control

Ambient RNA removal

During nucleus extraction from flash-frozen tissue, cell-free ambient RNA is liberated into the dissociation solution and becomes encapsulated with nuclei during library construction. Ambient RNA contamination can create undesired technical artifacts in single-cell data, such as ectopic gene expression and the obscuring of real biological differences between distinct cell population transcriptomes.

To address this issue, we corrected for ambient RNA expression using CellBender (version 0.1.0; https://github.com/broadinstitute/CellBender; ref. 24). CellBender is an unsupervised Bayesian model that requires no prior knowledge of cell type-specific gene expression profiles to identify ambient RNA counts. The approach is based on the principle that ambient RNA contamination will have a relatively uniform distribution across all cells, whereas cell-specific RNA will display more variable expression patterns. The procedure for removing ambient RNA using CellBlender involved the following steps with default parameters:

Quality control: Rapid autopsy snRNA-seq samples (particularly nonsorted samples) have more low-quality droplets with debris and ambient RNA than regular scRNA-seq samples (23). To increase the signal-to-noise ratio between ambient RNA and real RNA counts, we first performed a lenient QC by removing nuclei with more than 5% mitochondrial genes and fewer than 127 genes or fewer than 255 reads and by removing genes present in fewer than 10 cells. The estimated cell number of each batch was inferred with SEQC (22). We applied CellBender (RRID:SCR_025990) to this initial lenient-filtered snRNA-seq data as follows.

Estimation of ambient RNA levels: CellBender estimated levels of ambient RNA for each gene across all nuclei by assessing the distribution of expression levels for each gene and identifying genes with a uniform distribution as candidates for ambient RNA contamination.

Subtraction of ambient RNA: Next, CellBender subtracted the estimated ambient RNA contamination from the expression level of each gene in every droplet. This process generated a corrected gene expression matrix with nontransformed integer counts.

Evaluation of ambient RNA correction: We selected 5,000 highly variable genes (HVG) using the variance-stabilizing transformation method (25). To normalize the data, we scaled each cell to 10,000 reads and applied a log₂(X + 1) transformation. Dimensionality reduction was performed using principal component (PC) analysis (PCA), and the top 50 components were utilized for downstream analysis. We constructed a k-nearest neighbor (kNN) graph using k = 30 and applied PhenoGraph (RRID:SCR_016919; ref. 26) to identify distinct coarse cell clusters. Cell type–specific markers were used post hoc to evaluate ambient RNA correction. CellBender successfully retained cell type-specific markers in corresponding clusters while removing unexpected RNA counts, particularly genes from acinar cells that appeared in other cell types.

Filtering low-quality nuclear transcriptomes

Proceeding with the CellBender-corrected count matrix, cells with a low number of detected genes, a low total UMI count (sequencing depth), and a high fraction of mitochondrial counts were designated as low-quality cells, as they can represent dying cells with broken membranes (27). Previous snRNA-seq protocols have also reported that ribosomes can remain attached to the nuclear membrane during nucleus isolation (28); therefore, data were further assessed for library size, total gene counts, and mitochondrial and ribosomal RNA content.

Library size and gene count thresholds: We removed cells with fewer than 500 RNA counts and fewer than 200 genes.

Mitochondrial and ribosomal RNA content thresholds: As our droplets contained nuclear transcriptomes, we reasoned that mitochondrial and ribosomal RNA should be greatly reduced in high-quality transcriptomes. Hence, we checked for cells with high mitochondrial and ribosomal content. Cells with higher levels of mitochondrial and ribosomal genes primarily belonged to nonsorted samples, suggesting that these droplets contained higher levels of debris, as expected. After manual assessment, we removed droplets with more than 1% mitochondrial RNA and/or more than 10% ribosomal RNA fractions.

Doublet detection

Multiplets (droplets containing more than a single nucleus), predominantly doublets, are an undesired byproduct of library production that creates artifactual transcriptomes and confounds real biological signals. Homotypic doublets encapsulate two nuclei from the same cell type, and heterotypic doublets capture two different cell types, leading to cell-type mislabeling (27). Given the challenging task of differentiating single transcriptomes from doublets, using more than one detection approach and comparing results can increase the accuracy of doublet detection (29). We used DoubletDetection (https://zenodo.org/record/2678042) and Scrublet (30), two of the top-performing doublet detection algorithms (31), and further inspected identified doublets to confirm a larger library size compared with singlets, as well as the expression of conflicting gene markers. For each sample independently, we visually compared putative doublet and singlet total count distributions together and their clustering distribution in Uniform Manifold Approximation and Projection (UMAP) for dimension reduction.

DoubletDetection: DoubletDetection is a machine learning algorithm for identifying doublets in scRNA-seq data (https://zenodo.org/record/2678042; https://github.com/JonathanShor/DoubletDetection). It generates synthetic doublets, clusters them together with the original data using PhenoGraph (26), and assigns a score and P value for clusters with enriched synthetic doublets using a hypergeometric test. We used DoubletDetection separately on each sample’s raw snRNA-seq count matrix with default parameters.

Scrublet: Scrublet (RRID:SCR_018098; https://github.com/swolock/scrublet; ref. 30) simulates doublets from the observed data and uses a kNN classifier to calculate a continuous doublet_score (between 0 and 1) for each transcriptome. The score is automatically thresholded to generate predicted_doublets, a boolean array that is true for predicted doublets and false otherwise. We used Scrublet independently on each sample’s raw snRNA-seq count matrix with default parameters.

We found the results from both methods to be complementary and removed cells identified as doublets by either method. Transcriptomes passing library size, mitochondrial, ribosomal, and doublet detection criteria were retained, and the data matrices were concatenated into a single matrix (73,142 cells and 22,318 genes) for downstream analysis.

snRNA-seq data analysis

Feature selection, normalization, and variance stabilization

Following quality control, we selected 5,000 HVGs using “seurat_v3” in Scanpy (version 1.9.8; RRID:SCR_018139; https://github.com/scverse/scanpy; ref. 32), which computes a normalized variance for each gene on the raw counts (25). Other parameters were set to default. To normalize the data, we scaled each cell to 10,000 reads. The normalized counts were then log-transformed (base 2).

Dimensionality reduction and visualization

PCA of the log-normalized matrix was performed using the ARPACK solver on the selected HVGs. We retained the first 50 PCs, which explained 33.5% of the variation in the data, and constructed a kNN graph using k = 30. To visualize the data, UMAP was applied to the PCA-reduced data with a minimum distance of 0.1.

As noncancer cells from different libraries were well integrated, we did not perform any batch correction on our data. Differences between samples from different anatomic locations were regarded as biologically driven.

Gene signature scores

To generate all gene signature scores in our study, we used the Scanpy score_genes function (33), which calculates the mean expression of genes of interest subtracted by the mean expression of a random expression-matched set of reference genes. To control for gene set sizes, we selected the random reference set to be the same size as the gene set of interest. Other parameters were set to default.

Cell type annotation

Cancer cell type annotation: To annotate cell types, we first sought to discern cancer cells from noncancer cells. The tumors harbor a truncal KRAS^G12V mutation, detected both by MSK-IMPACT (34) and WES mutation calling; therefore, we used two independent but complementary KRAS signatures from the literature to generate a KRAS_signaling score per cell:

KRAS_PDAC (35): This signature of 36 genes is based on differential expression between epithelial cells in wild-type KRAS and KRAS-knockout mouse tumors. We used the human orthologs provided in the signature.

KRAS_addiction (36): This signature was generated by comparing human lung and pancreatic cancer lines that require KRAS to maintain viability with those lines that do not; all lines harbored KRAS mutations and were treated with short hairpin RNAs to deplete KRAS. The resulting signature is specific to KRAS-dependent cells and is associated with a well-differentiated epithelial phenotype also observed in primary tumors.

We scored these signatures separately, and although high-scoring cells for the two signatures did not overlap fully, both robustly identified the same clusters; thus, we used the union of KRAS_PDAC and KRAS_addiction to generate the KRAS_signaling signature for cancer cell annotation. Positive clusters were confirmed by CNA profiles inferred from the scRNA-seq data using inferCNV (12), as described in the following section.

Noncancer cell-type annotation: To label noncancer cells, we clustered all cells using PhenoGraph (26) with default parameters on the previously obtained PCs and used literature-curated canonical cell type–specific markers (Supplementary Table S2) to annotate the clusters. For clusters related to smooth muscle cells, MUC1/MUC6 epithelial cells, and adipocytes, no initial cell type identity could be discerned. Therefore, we ranked the genes underlying each cluster using the Scanpy function scanpy.tl.rank_genes_groups with the sparse matrix and default parameters. Reference clusters were set to “rest,” as well as adjacent clusters with known cell type identity, for increased granularity. Genes among the top 20 ranked genes were used to identify the cell identity of those clusters.

Inferring CNAs from snRNA-seq data

To infer chromosomal CNAs in tumor cells, we ran inferCNV (version 1.10.0; https://github.com/broadinstitute/inferCNV; ref. 12) and CopyKAT (version 1.1.0; https://github.com/navinlabcode/copykat; ref. 13) using the Python API of these algorithms implemented in the infercnvpy package (version 0.1.0). We ran both packages using default parameter settings and used noncancer cell types as the diploid reference. InferCNV was run with a window size of 100 genes and a step size of 1 to balance the detection of focal and broad CNA events.

Phylogenetic inference in rapid autopsy data

We used ductal and acinar cells as reference normal cells for IntegrateCNV. The algorithm returned a matrix containing copy numbers for 43,949 cells in 116 genetic regions. We only took the subset of cells annotated as tumor and expanded this matrix to a ternary matrix, as described above, resulting in 177 features. We then removed features that are highly similar across all cells by filtering out features that are modal with a frequency of 99% or higher, reasoning that small variations in copy number (frequencies below 1%) are likely noise, leaving a final input matrix containing 101 features.

We applied PICASSO to this input data and required that each cell have a UMI count greater than 750 and that each clone contain at least 75 cells, generating 66 clones. As a final filtering step to remove noisy clone data from the phylogeny, we required each clone to have at least one CNA at a prevalence greater than 80% to be considered valid. We reasoned that clones without highly prevalent CNAs are not likely to be well supported and may represent “noise” clones with cellular CNA profiles that are inconsistent with more well-defined clones. Removing four such noisy clones left a total of 62 clones (95–1,613 cells per clone, median = 618 cells) containing 40,994 cells in the phylogeny.

We defined a primary clone as containing at least 50% of cells from the primary tumor, yielding four primary clones in the data. As a proxy for the metastatic behavior of each primary clone, we calculated the proportion of nonprimary cells within each clone, with higher values indicating greater dissemination.

AC5 clone assignment

To confirm that primary cells expressing the archetype cluster 5 (AC5) program were strongly associated with advanced clones, we focused on the two advanced AC5 clones with the most primary cells (clones I and J, bearing 7 and 9 cells, respectively). We compared the CNA profiles of these cells with the clone profiles (CNA change probabilities at each site) of their assigned clones, as well as the clone profiles of a representative clone (clone 1-1-0-1-1-0) with a majority of primary cells.

We also computed the log-likelihoods of the primary cell CNA profiles in these clones, compared them with those of all other cells in the clone, and found that they exhibited median levels of clone confidence compared with the other (primarily stomach and liver) cells in the clones.

Pairwise diffusion distances of pancreas primary AC5 cells

To quantitatively evaluate the similarity of pancreas primary AC5 cells with metastatic cells versus other pancreas primary cells, we compared the pairwise diffusion distances from all primary AC5 cells with all metastatic AC5 cells and with all other primary cells separately. We used the “scipy.spatial.distance.cdist” (RRID:SCR_008058) function with the metric = euclidean on the diffusion map coordinates. This computes the distance between each pair of the two collections of inputs.

Archetype analysis

We used archetype analysis to identify optimal phenotypes (representing adaptive processes) among cancer cell transcriptomes, which may be shared or specific to one or more tumor sites. Archetype analysis identifies the vertices of a convex polytope—an approximation of a convex hull that encapsulates the data in phenotypic space (37), which in our case is diffusion space. Archetypes often correspond to the extremes of single diffusion components (DC), which are commonly used to approximate the major axes of variation within the phenotypic manifold. As m, the number of DCs as computed in the section “DCs” below, corresponds to the dimensionality of the data, we selected the number of archetypes we wished to identify as m + 1. To understand the gene programs that cancer cells use to adapt to different metastatic sites, which likely pose unique challenges and stresses, we computed archetypes in each tissue independently, as described below.

Archetype analysis per tumor site

First, we partitioned the data by tumor site (pancreas primary, liver, omentum, peritoneum, diaphragm, stomach, lymph node). Each site was normalized independently by scaling each cell to 10,000 reads and applying a log₂(X + 1) transformation.

The selection of the number of HVGs is crucial for capturing meaningful biological variability while minimizing technical noise. Too few HVGs (<500) risk losing important biological variation, whereas too many HVGs (>5,000) increase noise without adding significant biological variation. In general, our study and others with large datasets (>50,000 cells) and diverse cell types select around 5,000 HVGs. For medium-sized datasets (5,000–50,000 cells) and less cell type diversity, 2,000 to 3,000 HVGs are recommended. To perform archetype analysis per site, which includes only cancer cells from the same organ (470–23,950 cancer cells per organ, median 4,031), we computed 2,000 HVGs using the “seurat_v3” (25) method in Scanpy.

We computed 50 PCs using the svd_solver = “arpack” on the HVGs in each dataset. Sites included PDAC primary (3,479 cells, 40% variance explained by PCA), liver (4,031 cells, 36% variance), peritoneum (23,950 cells, 36% variance), lymph node (470 cells, 44% variance), stomach (4,075 cells, 35% variance), diaphragm (6,137 cells, 37% variance), and omentum (3,305 cells, 34% variance). We then computed the kNN graph with k = 30 neighbors on the PC space representation (X_pca). We chose 30 neighbors to balance between adding noise (<20 neighbors) and losing biological variation (>50 neighbors) in the medium-sized datasets we analyzed.

To visualize each site separately, we computed UMAP (min_dist = 0.1) and force-directed layout (FDL) with default parameters on the kNN graph. We then clustered each dataset using Leiden clustering in Scanpy with default parameters and further assessed cell quality and cancer cell purity in each cluster. We detected some outlier clusters with low library size in lymph node (n = 12 cells) and stomach (n = 226 cells) and noncancer cell contamination in liver (n = 85 cells) data partitions. Given the objective of archetype analysis in detecting extreme data points in the multidimensional space, we removed those cells from each data partition and from the entire dataset.

DCs: Given the presence of different cell-state densities in the data, we used an adaptive anisotropic kernel (38), which adjusts the local bandwidth (sigma) based on local density, to compute diffusion maps. This can provide more flexibility in regions with different densities, improving resolution in sparse areas and reducing oversmoothing in dense areas compared with the fixed anisotropic Gaussian kernel with a predefined scale (sigma) in Scanpy, which is more appropriate for relatively uniform cell-state density datasets.

With the adaptive anisotropic kernel, we computed 10 DCs on the PC projections of the data and calculated their corresponding eigenvalues and the diffusion operator. We used the eigenvalue knee point to determine the number of DCs for each site: pancreas 5 archetypes, liver 6 archetypes, lymph node 5 archetypes, peritoneum 5 archetypes, omentum 6 archetypes, stomach 4 archetypes, and diaphragm 5 archetypes. Archetypes were calculated on the DCs using the Python implementation of the PCHA algorithm with delta = 0. Archetypes were identified independently 10 times to assess robustness, and the nearest real cell to each archetype was identified using Euclidean distance in diffusion space.

Archetype neighborhoods: We next sought to annotate each archetype based on gene expression. As each archetype is identified as a single cell, we enhance statistical power by defining archetypal neighborhoods, consisting of each archetype’s most similar cells in diffusion map space. The neighborhoods are defined such that they include enough cells to enhance the robustness of inference while maintaining the archetypal phenotype and distinction between archetypes. Importantly, different metastatic sites have different numbers of cancer cells and archetypes, and the density of cells in the phenotype space varies. To account for all these differences, for a given archetype A in a given tissue, we calculate the diffusion distance (D) to its nearest archetype and define the neighborhood for A as the set of cells that are within a fraction of D. This ensures no overlap between the archetypal neighborhoods, thereby maintaining their distinctions. Parameters used for each site are as follows: PDAC primary DC fraction distance = 1/3 (91–1,571 cells per neighborhood), liver DC fraction distance = 1/3 (63–696 cells), peritoneum DC fraction distance = 1/4 (311–1,098 cells), lymph node DC fraction distance = 1/2 (36–112 cells), stomach DC fraction distance = 1/3 (315–997 cells), diaphragm DC fraction distance = 1/3 (160–2,037 cells), and omentum DC fraction distance = 1/3 (43–417 cells). To visualize archetype neighborhoods, we colored the selected neighborhood cells on the FDL projections.

Differential gene expression: For each tumor site, differentially expressed genes (DEG) were calculated for each archetype neighborhood versus all other neighborhoods from the same site, using raw counts. Genes expressed in fewer than 5% of cells in each group were filtered out to reduce noise. Differential expression was performed using diffxpy (https://github.com/theislab/diffxpy) with a Wald test, considering DEGs with a log₂ fold change >0.05 and a q value < 0.01.

Robustness analysis of archetype neighbors

We tested the robustness of our archetype analysis and archetype neighborhood selection by downsampling library size to various extents for each organ separately. For this, we downsampled counts from each tumor site’s raw counts data using “sc.pp.downsample_counts.” We set the count_per_cell parameter to be 10% or 20% of the original library size, resulting in a randomly downsampled dataset. For each site and downsampling level, we repeated the analysis 20 times with a different random seed for subsampling.

We repeated the entire archetype analysis process using the same parameters as described above in the subsampled datasets. We then compared the selected archetype neighborhoods using the Jaccard metric, which measures the similarity between two sets of elements by quantifying how many elements (archetype neighbor cells) the sets have in common relative to their total unique elements.

To assess the robustness to higher synthetic dropout rates (10% and 20%), we computed Jaccard similarity among the archetypal neighborhoods across different iterations. We observed a high similarity of >0.75, indicating that the selected archetype neighborhoods are robust.

Cell-state density estimation

To evaluate if archetype neighborhoods were driven by the cell-state density distribution in the high-dimensional space, we estimated the cell-state density of each tumor site partitioned data using Mellon (https://github.com/settylab/Mellon; ref. 39) with default parameters. Mellon is a nonparametric cell-state density estimator based on a nearest-neighbor distance distribution. It estimates cell-state densities from high-dimensional representations of single-cell data using a Gaussian process. We preprocessed and calculated cell-state densities for each tumor site separately, following the basic tutorial (https://github.com/settylab/Mellon/blob/main/notebooks/basic_tutorial.ipynb).

Integrated archetype clusters

To capture possible shared processes, we subsetted the data to include all cells labeled with an archetype and all genes that were included in any DEGs associated with any archetype in any organ. All 14,826 archetype cells and 15,017 genes were combined into a single matrix, which we median-count normalized and log-transformed counts. PCA (56 PCs, 20% variation explained) was followed by kNN graph construction (k = 30 neighbors), Leiden clustering (resolution = 1), PAGA (40), and UMAP visualization (min_dist = 0.1, init_pos = PAGA). The resulting Leiden clusters aggregate together archetypes calculated from the different sites; hence, we defined the Leiden clusters as “integrated archetype clusters.” We reasoned that each integrated archetype cluster could capture specific biological processes shared between different sites (e.g., cell cycle, EMT) or unique to a given site (e.g., lipid metabolism). To identify the underlying biological processes specific to each integrated archetype cluster, we then calculated DEGs for each cluster and identified gene modules as described below:

Level 1: Differentially upregulated genes. Differential expression using diffxpy (https://github.com/theislab/diffxpy; Wald test, DEGs with log₂ fold change >1 and q value < 0.05) was calculated for each integrated archetype versus all other archetypes.

Level 2: Gene modules. Cancer cells are able to express a variety of gene expression programs that may resemble distinct modular processes in a physiologic setting. To disentangle these gene expression programs, we used Hotspot (41). Hotspot identifies informative genes based on gene–gene autocorrelation in local neighborhoods in the phenotypic manifold, using a kNN graph, which we generated with weighted_graph = false, n_neighbors = 30, and FDR <0.05. Gene modules were computed on these informative genes: Informative genes from Hotspot modules were ranked by local correlation z-score. Then, preranked gene set enrichment analysis (GSEA; refs. 42, 43) was performed using GSEApy (RRID:SCR_025803; https://github.com/zqfang/GSEApy; ref. 44) against selected GSEApy-supported gene set libraries (https://maayanlab.cloud/Enrichr/#libraries) and expert-curated gene sets:

GSEApy libraries: GO_Biological_Process_2021, MSigDB_Hallmark_2020, Reactome_2016, KEGG_2021_Human, GO_Cellular_Component_2021, GO_Molecular_Function_2021, WikiPathways_2019_Human, and Azimuth_Cell_Types_2021

Expert-curated gene sets: Azimuth_Pancreas_Cells (https://azimuth.hubmapconsortium.org/references/human_pancreas/), PDAC_Subtypes (classical and basal), PDAC_Signatures, Pancreas_Development, Cancer_Metaprograms, Cell_Cycle, and KRAS_signaling (see Supplemementary Table S10 for references and manually curated gene sets).

The pancreas development gene set was generated by calculating DEGs [using MAST (RRID:SCR_016340; ref. 45) with default parameters] between emergent endodermal pancreas (clusters marked by PRX1) and other emerging endodermal organs, followed by mapping gene orthologs between mouse and human genomes.

Level 3: Archetype genes. DEGs and genes with modules whose mean expression is highest in a given archetype were used to characterize the archetype. This level of annotation ensures that genes are specifically upregulated in the archetype over other archetypes. Level 3 genes in each archetype were manually inspected to confirm GSEA results and to increase the granularity of the archetype descriptions. Archetypes with low normalized enrichment scores (NES) from GSEA were further inspected and labeled according to level 3 genes.

CZ CELLxGENE Discover (RRID:SCR_024894; ref. 46) was used to annotate AC5. Gene expression of AC5 genes was evaluated in CZ CELLxGENE. Higher average expression was observed in intestinal, stomach, and gallbladder tissues. Specifically, epithelial cell types were then evaluated for the expression of AC5 genes.

Intestine: Endocrine cell, columnar/cuboidal epithelial cell, secretory cell, enterocyte, epithelial cell, mesothelial cell, glandular epithelial cell, goblet cell, absorptive cell, brush cell, intestinal crypt stem cell of colon, intestinal epithelial cell, and intestinal enteroendocrine cell

Stomach: Enterocyte, epithelial cell, ciliated epithelial cell, columnar/cuboidal epithelial cell, glandular epithelial cell, secretory cell, enteroendocrine cell, endocrine cell, peptic cell, mucous cell of the stomach, parietal cell, glandular cell of the esophagus, epithelial cell of the esophagus, intestinal epithelial cell, brush cell, type G enteroendocrine cell, mucus-secreting cell, goblet cell, and intestinal goblet cell

Gallbladder: Epithelial cell, secretory cell, and goblet cell

Pancreas: Pancreatic ductal cell and epithelial cell of the pancreas

To annotate AC2 at a more granular level, we also used the Kyoto Encyclopedia of Genes and Genomes (KEGG) database (RRID:SCR_012773; ref. 47). Specifically, we used the KEGG Mapper Search Tool (https://www.genome.jp/kegg/mapper/search.html), which searches various KEGG objects, including genes, KEGG orthologs (KO), Enzyme Commission (EC) numbers, metabolites, and drugs, against KEGG pathway maps and other network entities. Then the top matching KEGG objects found were used to explore and annotate the biology of AC2 modules:

Fatty acid and cholesterol biosynthesis: Metabolic pathways (hsa01100) and fatty acid metabolism (hsa01212)

Oxidative stress and detoxification: Metabolic pathways (hsa01100), ferroptosis (hsa04216), glutathione metabolism (hsa00480), and chemical carcinogenesis–reactive oxygen species (hsa05208)

Archetype cluster annotation

Archetype cluster annotation was performed by first considering NES and the specific archetype genes deemed significant by GSEA. The NES genes were used as an initial general guide. Higher priority was then given to the specific gene modules and genes to annotate clusters in a granular and specific manner. CZ CELLxGENE Discover (46) was used to annotate AC5 as only the PDAC adhesive gene program (48) was significantly enriched.

Comparison with Leiden clustering

To compare archetype clusters and Leiden clusters, we first clustered the cancer data using “sc.tl.leiden” with default parameters. Then, the same level 1 and 2 steps employed for archetypes were used to annotate gene programs associated with Leiden clusters. We compared the archetype and Leiden clusters’ DEGs using Jaccard similarity.

Analysis of RA19_21 peritoneum metastases

Two PDAC peritoneum metastases were harvested from the rapid autopsy RA19_21, and snRNA-seq data were collected following the same protocol described in the snRNA-seq data preprocessing, scRNA-seq data analysis, and archetype analysis sections above for RA19_10. Data preprocessing and quality control were also performed using the same workflow. The same archetype analysis and gene program annotation workflows were performed for these peritoneal metastatic samples to evaluate the expression of the lipid metabolism and oxidative stress programs found in AC2. No integration of archetype clusters was required as only peritoneal metastases were analyzed.

Entropy of archetype distributions

We sought to determine whether each clone exhibits a greater diversity of archetypes than expected by chance, which would indicate phenotypic plasticity across the phylogeny. For each clone, we computed the Shannon entropy of the observed archetype distribution as a measure of phenotypic diversity. Shannon entropy, $\mathrm{\hspace{0.17em}}H$, is calculated as follows:

$$H\mathrm{\hspace{0.17em}}\mathrm{=}\mathrm{\hspace{0.17em}}\mathrm{-}\mathrm{\hspace{0.17em}}\mathit{log}{\displaystyle \mathrm{\sum }}_{i\mathrm{=}1}^k{p}_{i\mathrm{\hspace{0.17em}}}\mathit{log}\left(p_i\right)$$

where $p_i$ is the proportion of cells within the clone assigned to archetype $i$ and $k\mathrm{=}18$ is the number of unique archetypes. This entropy metric allows us to quantify the spread of archetype diversity within clones, with higher entropy values indicating more even and diverse distributions of archetypes.

Null model comparisons

To contextualize the observed entropy and evaluate whether the diversity observed within clones is greater than expected by chance, we compared our results to several null models. Each null model simulates archetype distributions under different assumptions, providing a range of baselines. In order of decreasing expected diversity, they are as follows:

1. Random assignment. Archetypes are assigned to cells randomly across all clones, with probabilities matching the global frequencies of each archetype. This model retains the overall prevalence of each archetype but removes any structure associated with clone or site, simulating a scenario in which cells randomly adopt a phenotype without any constraints.

2. Site-constrained random shuffle. Archetypes are assigned to cells randomly within sites, preserving each site’s archetype frequency distribution. This model retains the overall presence of each archetype and its prevalence within each site but removes any structure associated with clones.

3. High site–archetype concordance assignment. Archetypes are assigned to cells to minimize the dispersion of archetypes across sites. We carry out a greedy assignment of archetype labels to cells within sites in a way that retains the global archetype frequency but not the per-site frequencies. This model shows the expected diversity if cells were insufficiently plastic to adopt the same phenotype in multiple distinct sites.

4. Site entropy within clones. We compute the entropy of site distribution within each clone, ignoring archetype labels, to model the simplistic scenario in which the site drives all phenotypic variation.

Plasticity analysis

The entropy of archetypes within clones provides information about the number of phenotypes a clone can adopt. However, to measure lineage plasticity—which we define as the cells’ inherent capability to flexibly transition between various lineage states or phenotypes—we must examine cellular phenotypes in the context of their phylogenetic relationships. This approach allows us to assess the extent to which cells or cell groups adopt distinct phenotypes compared with their evolutionary ancestors.

We leveraged two complementary data modalities to develop metrics for measuring lineage plasticity: PICASSO, which enables the reconstruction of phylogenetic relationships, and archetype analysis, which characterizes the breadth of phenotypes present in the cells. Existing methods for measuring plasticity have key limitations, including the need to discretize continuous cell states, dependence on fully resolved cell phylogenies, and sensitivity to neighborhood size hyperparameters. Our integrated approach specifically addresses these concerns.

Yang and colleagues (19) defined three metrics for quantifying cellular plasticity. scEffectivePlasticity applies the Fitch–Hartigan algorithm to calculate a normalized parsimony score based on discrete Leiden cluster transitions across a phylogenetic tree, whereas scPlasticityAllelic provides a tree-agnostic alternative by measuring the proportion of cells belonging to Leiden clusters that are not their closest genetic relatives (determined by edit distance). Both of these approaches rely on discretizing phenotypes into Leiden clusters, which makes them sensitive to clustering resolution and is poorly suited for continuously varying phenotypes, in which small changes near cluster boundaries can be misclassified as plasticity. scPlasticityL2 addresses this limitation by using continuous phenotypic measurements, calculating the Euclidean distance in scVI latent space between cells and their tree-defined neighbors. However, both scEffectivePlasticity and scPlasticityL2 require a fully described tree topology and are thus highly dependent on the accuracy of tree inference, and scPlasticityL2 and scPlasticityAllelic further depend on a user-defined neighborhood size—a single predefined value that is difficult to choose optimally across datasets with varying sequencing depth, sampling density, and degrees of phenotypic change.

Schiffman and colleagues (49) introduce phylogenetic correlations to quantify how cellular measurements are distributed across a phylogenetic tree using Moran’s I (a measure of spatial autocorrelation) and its bivariate generalization. This approach measures correlation patterns directly across the phylogeny, facilitating the analysis of both continuous expression patterns and discrete cell states within their evolutionary context. The method transforms pairwise phylogenetic distances into a weighted matrix, using carefully selected weighting functions. The choice of weighting function is critical as phylogenetic correlations depend significantly on the structure of the normalized weight matrix, and the function selected by the authors only includes cells that are each other’s nearest phylogenetic neighbors. This choice of weighting function may not be suitable for larger-scale datasets on the order of tens of thousands of cells.

To address these concerns, we developed plasticity analysis from single-cell transcriptional and evolutionary neighborhood overlap (PLASTRO), a metric for quantifying plasticity from jointly profiled lineage and scRNA-seq information, without relying on the inference of complete and exact tree topologies, fixed neighborhood size hyperparameters, or discretization of cell phenotypes. PLASTRO accepts two distance matrices as input: (i) lineage distance, which reflects how similar clones are to each other in evolutionary space, and (ii) phenotypic distance, which reflects how similar clones are to each other functionally. Given these matrices, we can define a lineage neighborhood and a phenotypic neighborhood of radius $r$ clones for each clone. Each clone’s neighborhoods comprise its $r$ closest cells in lineage and phenotype space, respectively. The key idea behind this approach is that there will be substantial agreement between the lineage neighborhood and the phenotypic neighborhood in nonplastic clones; thus, overlap in these neighborhoods will be high on average. In contrast, highly plastic clones will exhibit phenotypes distinct from other clones in their lineage, and their neighborhoods will overlap very little on average.

Computation of PLASTRO score

PLASTRO accepts a lineage distance matrix and a phenotypic distance matrix as input. Given these matrices, we can define, for each cell, a lineage neighborhood and a phenotypic neighborhood of radius r cells. The choice of radius clearly has a strong effect on the degree of overlap between phylogenetic and phenotypic neighborhoods. At very small radii, even nonplastic cells may exhibit low overlap by random chance. Conversely, at very large radii, plastic cells will exhibit strong overlap as well, given that each neighborhood contains nearly all the cells in the dataset. In addition, different radii provide varying signals that help differentiate plastic and nonplastic cells depending on the parameters of the dataset. To circumvent this issue and avoid reliance on neighborhood size as a parameter of our approach, we measure neighborhood overlap at varying scales and combine the signal present at each scale.

PLASTRO consists of four main steps:

1. Compute the lineage and phenotypic distance matrices.

2. For each cell, rank all other cells in terms of the distance from that cell in both (i) lineage space and (ii) phenotypic space.

3. For a given cell at overlap radius r, compute the overlap in its r closest cells as defined by phenotypic distance and by lineage distance. This is the number of cells that lie in both the phenotypic neighborhood of size r and the lineage radius of size r.

4. Aggregate the signal across radii by computing the area under the overlap versus radius graph.

Phenotypic distance matrix

We calculate the pairwise phenotypic distances between clones using Bray–Curtis dissimilarity, a metric that captures differences in relative abundances and is commonly used in ecologic and compositional analyses. Bray–Curtis is particularly suited to compositional data as it accounts for the proportional structure of the data, measuring dissimilarity on a scale from 0 (identical composition) to 1 (completely dissimilar).

The archetype composition for clone $A$ is denoted by $a\mathrm{\hspace{0.17em}}\mathit{\in }\mathrm{\hspace{0.17em}}{R}^k$, where $k$ is the number of archetypes and satisfies

$${\displaystyle \mathrm{\sum }}_{i\mathrm{=}1}^k{a}_j$$

$$\mathrm{\hspace{0.17em}}0\mathrm{<}\mathrm{\hspace{0.17em}}{a}_i\mathrm{<}1\mathrm{.}$$

The Bray–Curtis distance between two clones $A$ and $B$ is then given by the following:

$$D\left(A\mathrm{,}B\right)\mathrm{\hspace{0.17em}}\mathrm{=}\mathrm{\hspace{0.17em}}\frac{{{\displaystyle \mathrm{\sum }}}_{i\mathrm{=}1}^k\left|a_i\mathrm{-}{b}_i\right|}{{{\displaystyle \mathrm{\sum }}}_{i\mathrm{=}1}^k\left|a_i\mathrm{+}{b}_i\right|}\mathrm{.}$$

The Bray–Curtis distance ranges from 0 to 1, where 0 indicates that the two samples have identical compositions and 1 indicates that the two samples have completely disjoint compositions (no shared components).

Phylogenetic (lineage distance) matrix

We use the phylogeny inferred by PICASSO to construct a pairwise distance matrix between clones; the distance between two clones is given by the number of edges separating them in the phylogeny.

Overlap computation

Given a lineage distance matrix $D_L$ and a phenotypic distance matrix $D_P$ constructed on a set of cells, $X$, we compute the overlap for the cell of interest $c$ at radius $r$ as follows. We denote the distance in lineage space between cell $c$ and its $r^{th}$ nearest neighbor as $D_L\left(c\mathrm{,}r\right)$. Similarly, $D_P\left(c\mathrm{,}r\right)$ is the distance in phenotypic space between cell $c$ and its $r^{th}$ nearest neighbor.

We define the lineage neighborhood of cell $c$ at radius $r$ as follows:

$$N_L\left(c\mathrm{,}r\right)\mathrm{\hspace{0.17em}}\mathrm{=}\mathrm{\hspace{0.17em}}\mathrm{\{}x\mathrm{\hspace{0.17em}}\mathrm{\in }X\mathrm{\hspace{0.17em}}\mathrm{|}\mathrm{\hspace{0.17em}}{D}_L\left(c\mathrm{,}x\right)\mathrm{\hspace{0.17em}}\mathrm{\le }\mathrm{\hspace{0.17em}}{D}_L\left(C\mathrm{,}r\right)\mathrm{\}}$$

and the phenotypic neighborhood of cell $c$ at radius $r$ as follows:

$$N_P\left(c\mathrm{,}r\right)\mathrm{\hspace{0.17em}}\mathrm{=}\mathrm{\hspace{0.17em}}\mathrm{\{}x\mathrm{\hspace{0.17em}}\mathrm{\in }X\mathrm{\hspace{0.17em}}\mathrm{|}\mathrm{\hspace{0.17em}}{D}_P\left(c\mathrm{,}x\right)\mathrm{\hspace{0.17em}}\mathrm{\le }\mathrm{\hspace{0.17em}}{D}_P\left(C\mathrm{,}r\right)\mathrm{\}}$$

The overlap for cell $c$ at radius $r$ is then defined as the Jaccard similarity of its phenotypic neighborhood and its lineage neighborhood:

$$Overlap\left(c\mathrm{,}r\right)\mathrm{\hspace{0.17em}}\mathrm{=}\mathrm{\hspace{0.17em}}\frac{N_L\left(c\mathrm{,}r\right)\mathrm{\hspace{0.17em}}\mathrm{\cap }\mathrm{\hspace{0.17em}}{N}_P\left(c\mathrm{,}r\right)}{r}\mathrm{.}$$

Plastic cells will have lower agreement between lineage and phenotypic neighborhoods, particularly at lower radii, and thus a lower overlap at that radius on average, compared with less plastic cells.

Aggregating signal across radii

To avoid hard coding a radius, which may have a strong effect on the measured plasticity, we aggregate signals across radii by considering overlap size as a function of radius, which is an increasing function bounded by the line $y\mathrm{=}x$. We compute plasticity as the difference between the area under the line $y\mathrm{=}x$ and the area under the overlap–radius curve.

For more plastic clones, the number of cells in the overlap is lower for smaller radii as the phenotypic neighborhood is highly distinct from the phylogenetic neighborhood and grows to include all cells as the neighborhood size increases, resulting in a higher plasticity score. For less plastic cells, the overlap proportion is expected to be higher overall, and the overlap–radius curve more closely resembles the $y\mathrm{=}x$ line and thus yields a lower plasticity score.

Application of PLASTRO to PDAC clones

We apply PLASTRO to compute the plasticity of each clone in our data. The lineage distance matrix is computed based on the topology of the phylogenetic tree, in which clones $A$ and $B$ have a phylogenetic distance $D_L\left(A\mathrm{,}B\right)\mathrm{=}n$ if there are $n$ branches on the shortest tree path between them. The phenotypic distance was computed as described above using the Bray–Curtis dissimilarity between the archetype composition of clones.

Calculation of global plasticity

We used PLASTRO to calculate plasticity at the clonal level, but to assess global plasticity across the entire biological system, we turned to the Mantel test (50), which assesses the correlation between two distance matrices (the phylogenetic and phenotypic compositional distance matrices). The Mantel test is a nonparametric test for assessing matrix correlations and is well suited for evaluating the phylogenetic signal in data without assuming a specific model of evolution. Mathematically, the Mantel test statistic is computed as follows:

$$m\mathrm{\hspace{0.17em}}\mathrm{=}\mathrm{\hspace{0.17em}}\frac{1}{\left(n\mathrm{-}1\right)}\mathrm{\hspace{0.17em}}{\mathrm{\sum }}_{i\mathrm{,}j}\frac{(A_{i\mathrm{,}j}\mathrm{-}\overline{A})}{\sqrt{{{\displaystyle \mathrm{\sum }}}_{k\mathrm{,}l}(A_{k\mathrm{,}l}\mathrm{-}\overline{A})}}\frac{(B_{i\mathrm{,}j}\mathrm{-}\overline{B})}{\sqrt{{{\displaystyle \mathrm{\sum }}}_{k\mathrm{,}l}(B_{k\mathrm{,}l}\mathrm{-}\overline{B})}}$$

where $A\mathrm{,}B\mathrm{\hspace{0.17em}}\mathit{\in }{R}^{n\mathrm{\times }n}$ are the distance matrices being compared and $\overline{A}\mathrm{,}\mathrm{\hspace{0.17em}}\overline{B}$ are their respective means. The statistic ranges from −1 to +1, with +1 indicating a perfect positive correlation (as distances in one matrix increase, distances in the other matrix increase proportionally).

A significant positive correlation between the compositional and phylogenetic distance matrices would indicate that clones with closer evolutionary relationships also have more similar compositions. A value of −1 represents a perfect negative correlation (increasing distances in one matrix correspond to decreasing distances in the other, reflecting a complete inverse relationship). A Mantel test statistic near 0 indicates no correlation between the two matrices such that distances in one matrix do not predict distances in the other, implying that compositional differences are more likely to be driven by factors other than shared ancestry. We used Spearman correlation to measure the association between matrices and performed 1,000 permutations to test the significance of the observed correlation.

Results

A patient-specific atlas of PDAC metastasis

Using rapid autopsy specimens from a single patient with PDAC and optimized specimen dissociation and snRNA-seq protocols, we constructed a comprehensive atlas spanning primary and metastatic sites, enabling the study of how cancer evolves and adapts across diverse tissue environments. We integrated snRNA-seq and matched WES data from each specimen to uncover both clonal architecture and adaptive transcriptional programs driving metastatic progression.

The patient was diagnosed at age 35 with PDAC and extensive synchronous liver metastases, as evidenced by CT, which was used in addition to CA19-9 tumor marker levels to follow disease status over the 9 months that the patient survived (Fig. 1A and B). Despite an initial robust response to standard-of-care mFOLFIRINOX, the rapid emergence of refractory disease, unresponsive to second-line gemcitabine + nab-paclitaxel, highlighted the cancer’s remarkable adaptive capacity within months of treatment.

Figure 1.

Profile of a cancer ecosystem from a single patient with PDAC. A, Top, maximal diameter of primary and liver tumors, based on CT measurements at the indicated time points from diagnosis (day 0). Black bar marks the period of mFOLFIRINOX treatment. Bottom, levels of CA19-9 tumor marker in blood, based on indicated measurement days. Baseline at diagnosis (day 0) is 13,000 U/mL, and the upper physiologic limit is 37 U/mL (red line). B, Representative CT scans. Primary and liver metastatic tumors are overdrawn with colored ellipses. L, left; R, right; A, anterior; P, posterior. C, Anatomic location of collected biospecimens used to generate matched snRNA-seq, WES, and H&E data. Circle diameter indicates relative tumor size. D, FDL of cancer cell transcriptomes (45,134 nuclei), colored by sample (Materials and Methods). Stomach refers to stomach wall metastasis.

We collected 11 tumor specimens representing diverse tissue microenvironments, including the pancreas and six distal organs, by rapid autopsy. The sampling included, where possible, anatomically separate lesions from the same organ (the best approximation of biological replicates in human cancer): two peritoneal and three liver metastatic samples, in addition to two regions of the primary tumor (Fig. 1C). These 11 samples, collected from 7 distinct organ sites, exhibit diverse cell type compositions and tissue morphologies (Supplementary Fig. S1A–S1C).

We recovered 73,142 high-quality snRNA-seq profiles from all samples (Supplementary Fig. S1D–S1F; Supplementary Table S1), organized into 39 clusters by PhenoGraph (26), which we annotated based on known marker genes (Supplementary Fig. S1B and S1C; Supplementary Table S2). To distinguish cancer cells from noncancer, we identified cells with high KRAS signaling (35, 36) and detected clusters with accumulated CNAs using inferCNV (Supplementary Fig. S2A and S2B; ref. 12). We evaluated the expression of genes associated with ductal cells, PDAC, mesenchymal, and EMT gene programs to distinguish normal ductal cells from primary PDAC cells (Supplementary Fig. S2C). In total, we recovered 45,134 cancer epithelial nuclei across all lesions, bearing multiple cancer-related mutations (Fig. 1D; Supplementary Fig. S2D and S2E). From the bulk WES, we identified the expected common PDAC alterations such as KRAS and TP53 missense mutations; copy-number deletions of CDKN2A, SMAD4, and DCC; and copy-number amplifications of MYC, MCL1, and CCNE1 (Supplementary Fig. S2E). In addition, CNA bulk analysis showed that both PDAC primary tumor samples and metastatic samples harbor broad CNAs widespread across the genome (Supplementary Fig. S2F). Together, these analyses underscore that the genomic landscape of this patient with PDAC recapitulates the known alterations and genomic features of metastatic PDAC.

PICASSO resolves single-cell phylogenies

The availability of both primary and metastatic cells from the same patient provides a unique opportunity to study how cancer cells evolve and adapt to different tissue environments. To dissect the relative roles of genetic mutations and epigenetic plasticity in metastatic adaptation, it is essential to reconstruct the evolutionary history of cancer cells and compare their genotypic and phenotypic characteristics within a shared phylogenetic framework. However, current approaches face significant limitations.

Bulk WES offers a coarse view of phylogenetic relationships across lesions; however, it lacks single-cell resolution and cannot link genetic mutations to cellular phenotypes. Combined DNA–RNA single-cell assays (51, 52) are limited by cost and throughput—published studies consist of too few cells (typically <1,000; refs. 51–53) to capture the full phenotypic heterogeneity typically observed within lesions (54). Although copy-number inference from single-nucleus or scRNA-seq data (12, 13) can inform clonal relationships, current methods are extremely noisy and strongly affected by confounding factors such as the influence of tumor cell state and its related gene expression patterns (55–57). In addition, many phylogenetic algorithms assume that mutations occur only once (“perfect phylogeny”), whereas in cancer, CNAs are highly recurrent (58–60). For example, more than 50% of CNA regions violate the perfection assumption in our data, complicating traditional phylogenetic approaches (Supplementary Fig. S3A). Finally, classic algorithms for phylogenetic analysis assume evolutionary characters are reliable, whereas CNAs called from single-cell expression data are uncertain and noisy. Uncovering genotype–phenotype relationships and the role of epigenetic plasticity during cancer progression thus requires new approaches that can (i) reliably infer CNAs from scRNA-seq data and (ii) construct a robust phylogeny of cancer clones, taking into account noise, uncertainty, and possible CNA recurrence, as well as the large scale of single-cell data.

To address these challenges, we instigated a two-step approach. First, we developed IntegrateCNV, a statistical framework that leverages matched bulk WES and snRNA-seq profiles to infer CNAs at single-cell resolution (Supplementary Fig. S3B; Materials and Methods). Unlike existing methods that infer CNAs genome-wide from scRNA-seq alone (13, 61, 62), IntegrateCNV uses bulk WES data to identify regions harboring CNAs before performing targeted inference from scRNA-seq data for individual cells in these candidate regions. This focused strategy increases signal-to-noise ratio by limiting analysis to regions with strong evidence of copy-number variation. Specifically, for each cell and candidate region, it determines whether an alteration is likely to be present based on gene expression relative to a copy-neutral reference (Supplementary Fig. S3B). IntegrateCNV achieves higher or equal correlation with sample-level CNAs derived from bulk WES data compared with widely used tools such as inferCNV (12), CopyKAT (13), and Numbat (Supplementary Fig. S3C; ref. 14), even when bulk copy-number calls are provided to Numbat to guide inference. As IntegrateCNV only calls CNAs for a confident subset of the genome, it is significantly faster than CopyKAT and Numbat, requiring only hours to run on a standard laptop compared with multiple days on high-performance computing clusters.

Although IntegrateCNV improves CNA detection accuracy, the profiles it generates still contain many errors (Supplementary Fig. S3C), especially false negatives. Unfortunately, even phylogenetic reconstruction methods that allow errors in the character matrix typically assume them to be minimal. Moreover, the few algorithms that infer phylogenies from single-cell CNA profiles are designed for small-scale single-cell DNA sequencing experiments and assume error-free input (16, 17). To overcome these challenges and construct robust phylogenies, we developed PICASSO, a maximum-likelihood method tailored to large-scale, noisy CNA profiles (Fig. 2A; Supplementary Fig. S3D; Materials and Methods). PICASSO employs a tree-recursive algorithm that starts with a single leaf node containing all cells and then iteratively decides whether to split each leaf into two subclones. Each decision to split is based on maximizing shared information in consensus CNA patterns, corrected for noise and missing values, using EM. When there is insufficient evidence for further splitting, a leaf is marked as terminal. The output of PICASSO is a probabilistic assignment of cells to clones and a likelihood-optimized final tree describing clonal phylogenetic relationships and associated uncertainties. This top-down recursive approach only reconstructs major evolutionary relationships with good evidential support and is substantially more robust to noisy data than standard bottom-up approaches.

Figure 2.

PICASSO generates a CNA-derived single-cell phylogeny. A, PICASSO takes CNA profiles from scRNA-seq data (inferred by IntegrateCNV, for example) as input and encodes them in a probabilistic manner, then iteratively splits clones into subclones based on clustering shared patterns by EM. The algorithm proceeds in a top-down fashion until it reaches terminal leaves, which lack evidence for further splits. PICASSO output is the probabilistic assignment of cells to subclones and a maximum likelihood–optimized tree. B, Phylogenetic relationships between clones (rows) derived from single-cell CNA profiles for all cancer cells in the rapid autopsy dataset. Each stacked bar plot indicates the clone’s site composition (fraction of cells from each metastatic site), and the heatmap at right shows the modal copy numbers inferred by IntegrateCNV for that clone. The four clones that are predominantly from the primary tumor (stars) are distinguished by whether they also contain cells in metastatic lesions. C, FDL of all cancer cells, colored by the sample of origin for cells from the primary tumor clones (>50% cells from the primary tumor) that also contain metastatic cells. The inset indicates the number of cells from each sample in these two clones. D and E, FDL of PDAC primary cells showing cancer clones colored by the proportion of primary cells within the clones (D) and PDAC phenotypes (E). Gray cells in D were removed from phylogenetic analysis due to low transcript counts.

We validated performance using simulated data, which demonstrated that PICASSO produces more parsimonious phylogenies and outperforms agglomerative clustering in both speed and accuracy under varying levels of noise (Supplementary Fig. S4A and S4B). By providing a probabilistic assignment of cells to clones and a likelihood-optimized tree describing clonal relationships and uncertainties, PICASSO is an effective tool for dissecting the relationship between genotype and phenotype during cancer progression.

Evolutionary reconstruction of metastatic PDAC

We applied IntegrateCNV to cancer cells in our metastatic PDAC dataset and used PICASSO on the resulting 45,134 single-cell copy-number profiles in this large-scale dataset (Fig. 2B; Materials and Methods). Based on CNA calls in 116 candidate regions, PICASSO resolved 62 clones with a clear phylogenetic structure following noise removal. The resulting phylogeny is highly stable; despite the probabilistic nature of the algorithm, most evolutionary relationships are conserved across repeated runs (Supplementary Fig. S4C). Furthermore, bootstrapping analysis reveals that the tree structure remains stable even when removing a fraction of cells from each region (Supplementary Fig. S4D).

We used the inferred phylogeny to investigate patterns of metastasis, first asking which clones in the primary tumor spread and why. We identified four primary clones, defined as containing at least 50% of cells from the primary tumor—two that metastasized and two that did not (Fig. 2B). A subset of tumor cells from liver metastases were found to be closely related to the metastasizing clones from the primary tumor. Notably, liver-dominant clones are the most closely related metastatic clones to those found in the primary tumor, suggesting that the liver was the initial site of metastasis in this patient, consistent with the observation that PDAC typically spreads to the liver first (Fig. 2B and C; ref. 63). We also observed that peritoneal samples, unlike other organ sites, were composed of several clones that seem to be unique or nearly unique to that site (Fig. 2B). This pattern could be due to several reasons, including better sampling (the peritoneum has many more cells than other sites; Supplementary Fig. S2D) or limited intermetastatic seeding due to the large physical distance separating peritoneal lesions from other metastases (Fig. 1C).

Analysis of the metastasizing primary clones revealed distinct genomic and transcriptional features associated with metastatic potential. Metastatic clones from the primary had many more CNAs than their nonmetastatic counterparts. Notably, amplification of the oncogenic KRAS^G12V locus is a hallmark of nearly all metastatic clones. We recently showed that oncogenic KRAS enhances plasticity during PDAC premalignancy, partly by remodeling the communication between cancer cells and their environment (64). Our findings suggest that oncogenic KRAS continues to drive plasticity in advanced disease and that its amplification provides an additional boost that promotes metastatic competence.

In addition to genetic alterations, our dataset provides a rare opportunity to examine the transcriptional states of metastasizing clones. Mapping known PDAC tumor phenotypes revealed that most primary tumor cells from metastatic clones display a mesenchymal phenotype, indicating an EMT process, which has been strongly associated with metastasis (Fig. 2D and E; ref. 65). In contrast, nonmetastatic clones are enriched for epithelial phenotypes, suggesting that metastatic clones are already transcriptionally poised for dissemination while in the primary tumor. The observation of mesenchymal phenotypes in cells from nonmetastatic clones signifies that EMT alone is insufficient for successful metastasis. This level of resolution is uniquely enabled by our approach, as it allows us to connect transcriptional phenotypes to phylogenetic patterns.

The phylogenetic tree provides insights into metastatic seeding and spread. Although some clones map to a dominant metastatic site, most are found in multiple organs, suggesting that metastatic clones can adapt to diverse tissue environments. Conversely, each metastatic site contains cells from multiple clones, some separated by large phylogenetic distances (Fig. 2B), implying that metastatic sites were seeded by multiple clones in independent events. This and similar findings in other contexts (66, 67) support the idea that once tumor cells establish themselves at distal sites, they remodel the local microenvironment to create a favorable “soil” for further seeding by the primary tumor or other metastases (68).

Archetype analysis identifies metastatic gene programs

The observation that most clones metastasized to multiple organs raises the question of how tumor cells adapt to these distinct environments. Although metastatic cells must overcome universal hurdles such as migration and extravasation to establish themselves at distal locations, each organ presents unique challenges requiring site-specific adaptations for successful colonization and growth. We reasoned that these adaptations should manifest as highly optimized transcriptional phenotypes and that examining multiple metastatic sites from the same primary tumor would make it possible to uncover both shared and organ-specific mechanisms.

To systematically identify these adaptive programs, we applied archetype analysis (69, 70), which identifies boundary phenotypes known to represent optimized tasks, using a two-tiered approach (Supplementary Fig. S5A and S5B; Materials and Methods). Our strategy was to first analyze each organ separately, identifying four to six archetypes per tissue in a highly robust manner (Supplementary Figs. S5C and S6A). Archetype neighborhoods did not associate with cell-state density (Supplementary Fig. S5C), suggesting that archetype neighborhoods may represent both major cancer cell state phenotypes (high-density) and rare (low-density) cancer phenotype states (39). Next, to find archetypes and programs that are potentially shared across organs, we integrated all 14,513 archetype-labeled cells (32% of all cancer cells) into a single matrix and applied graph-based clustering, yielding 19 archetype clusters (AC0−AC18; Fig. 3A). Finally, each archetype cluster was annotated using DEGs and Hotspot (41) analysis (Fig. 3B).

Figure 3.

Archetype gene programs of primary and metastatic PDAC. A, UMAP of clustered archetypal cells from primary and metastatic sites, colored by cluster. The gray box encompasses three distinct archetype clusters related to EMT. B, Archetype gene program expression in each tumor sample. ER, endoplasmic reticulum. C, Fraction of cancer cells per archetype cluster, colored by sample. D, Expression of individual markers in AC1, AC3, AC4, and AC16. AC3 corresponds to classical–squamous cells that are more epithelial and is included for comparison. Canonical epithelial, mesenchymal, and EMT markers, as well as targets and modules downstream of MYC signaling, inflammation, and HLA-I antigen processing and presentation are indicated.

Our analysis generated well-defined archetype clusters, including some that are unique to one organ and others that appear in multiple organs (Fig. 3B and C; Supplementary Fig. S6B; Supplementary Tables S3–S5). For example, cells in AC3, AC9, AC14, and AC18 are only found in primary PDAC; AC13 (unfolded protein response: HSP90AA1, HSPH1, HSPD1, DNAJA1) is unique to the liver; and AC8 (development: PBX1, HES1, PDGFB; wound healing: FOS, JUNB, NR4A1, ANGPTL4) is specific to the omentum. Additionally, AC5 (gastrointestinal: MUC13, FABP1, FCGBP) is found mostly in the stomach wall, and AC2 (lipid metabolism: HMGCS1, SQLE, FDPS) is mainly in the peritoneum.

In contrast, we found that archetypes related to core cellular processes, such as the cell cycle, migration, EMT, and cell–environment interactions, such as extracellular matrix (ECM) interactions and inflammation, are typically shared across multiple organs (Fig. 3B and C). To gain insight into biological functions that broadly contribute to metastatic capacity, we focused on AC1, AC4, and AC16, which are present in multiple organs that together comprise all seven organ sites (Fig. 3C). These three clusters express mesenchymal genes and transcription factors related to EMT programs associated with metastatic spread (Fig. 3D; ref. 71). However, further analysis revealed that these apparently similar EMT states are distinguished by distinct gene and regulatory programs (Supplementary Tables S4 and S5).

Cluster 1 is enriched for programs associated with cytokine and chemokine secretion as well as TNFα/NF-κB and IL17 signaling, suggesting inflammatory activation. Cluster 16 is enriched for focal adhesion and ECM interactions. ECM remodeling is required for cancer cell growth and can recruit immune cells (68), suggesting a potential role in establishing the metastatic niche. In contrast, AC4 is enriched for glucagon signaling, a common liver-expressed pathway (72) that may reflect the influence of the liver microenvironment on AC4 cells, most of which originate in this organ (Fig. 3C). We found that AC4 has the highest expression of MYC, MYC target ODC1, and CA9, which can be regulated by MYC under hypoxic conditions (73), as well as genes downstream of MYC signaling (Fig. 3D), and is most enriched for MYC-expressing cells (Supplementary Fig. S6C and S6D). Data from patients with PDAC and mouse models have linked MYC hyperactivation to more aggressive metastatic disease (74) and chemoresistance (75). Moreover, both MYC and EMT pathways are enriched in metastases compared with primary tumors in patients with PDAC (74, 76). Further distinguishing these states, AC1 expresses inflammatory genes, whereas AC16 expresses HLA-I antigen processing genes (Fig. 3D). Together, these findings reveal three distinct EMT phenotypes: AC1 and AC16 have a mesenchymal profile associated with the inflammatory response, whereas AC4 has an EMT program that co-occurs with MYC signaling.

Unlike archetype analysis, traditional clustering approaches are not designed to identify gene programs optimized for specific biological tasks. Rather, they aim to define groups of cells that have more similar average expression than cells in other clusters. Direct comparison of these approaches in our dataset reveals substantial differences in cell groupings, DEGs between groups, and biological annotations (Supplementary Fig. S6E–S6G). Although clustering detects broad processes such as EMT, proliferation, and stress (Supplementary Table S6), more specific adaptations to metastatic sites, such as lipid metabolism and gastrointestinal gene programs, are only identified by archetypes (Supplementary Table S4). The ability to identify adaptive programs in archetype analysis stems from the focus on boundary states that represent specialized cellular functions, rather than average behaviors captured by clustering. The combination of comprehensive sampling across metastatic sites and archetype-based analysis thus provides a powerful framework for discovering key metastatic phenotypes.

Stomach wall metastases express gastrointestinal gene programs

Although liver metastasis in PDAC is well-studied, metastasis to the stomach wall is rare and poorly characterized, despite often leading to severe gastrointestinal complications such as pain, ascites, bowel obstruction, and other morbidities. Our analysis revealed evidence of organ-specific adaptation. Tumor cells from the stomach wall are enriched in AC5, and the vast majority of AC5 cells originate from this site (Figs. 3B, C, and 4A). Hotspot analysis identified three distinct gene modules expressed by AC5 cells that correspond to intestinal, stomach, and gallbladder epithelial cells based on healthy human single-cell reference data (Fig. 4B; ref. 46). These gene modules are minimally expressed in normal pancreatic tissue, indicating that although these metastatic cells are of pancreatic origin, they have acquired transcriptional programs resembling those of other gastrointestinal epithelia (Fig. 4C; Supplementary Tables S7 and S8).

Figure 4.

A gastrointestinal archetype indicates PDAC adaptation to the stomach environment. A, UMAP embedding (left) and distribution by sample (right) of AC5 cells. B, Hotspot modules in all AC5 cells, based on 78 HVGs with significant autocorrelation (FDR <0.05). Highlighted genes were used to annotate intestine, stomach, and gallbladder modules. C, Expression of AC5 genes in normal pancreas and gastrointestinal tissues based on the CZ CELLxGENE Discover database.

AC5 genes reflect diverse functions of the gastrointestinal tract, including digestion, nutrient absorption, protective barrier maintenance, and bile production (a gallbladder function), which are distinct from physiologic pancreatic capabilities. Although the pancreas is a gastrointestinal tissue, it is considered an accessory organ whose primary function is to secrete digestive enzymes and bicarbonate to neutralize stomach acid. We found AC5 gastrointestinal genes related to cell adhesion and structural integrity (CDH17, RHPN2, CLDN7, MYO1A, MYO7B), mucus production and protection (MUC17, MUC13, MUC5B, FCGBP, GCNT3), metabolism and transport (HSD17B2, FABP1, SLC22A18, GDA), and epithelial cell differentiation (PLAC8). As expected, AC5 genes are specific to the archetype cells and are minimally expressed in nonarchetype cells in the tissues present in AC5 (Supplementary Fig. S6H). Our analysis thus demonstrates that PDAC metastatic cells acquire extensive new gastrointestinal features in the stomach, likely as an adaptation or response to its unique signaling milieu.

Interestingly, a small group of cells from the primary tumor also predominantly express the AC5 gene program. Mapping archetype clusters to the primary tumor revealed that AC5 cells correspond to a classical mucin phenotype (Fig. 2E; Supplementary Fig. S7A), which has been observed in human primary PDAC tumors, as well as in primary lung, colorectal, gastric (77), liver, and head and neck cancers (78). Consistent with this phenotype, AC5 cells in the primary tumor express high levels of mucin (MUC13, MUC5AC, MUC5B), mucin production (GCNT3, TFF1–TFF3), and mucus-producing goblet cell differentiation (CREB3L1) genes, compared with other primary tumor cells (Supplementary Fig. S7B). We find that the classical mucin phenotype is more similar to metastatic states than to other primary phenotypes, as classical mucin cells coembed near metastatic AC5 cells and are separated by shorter diffusion distances, reflecting greater transcriptional similarity (Supplementary Fig. S7C).

We examined clonal membership to understand the origins of AC5 classical mucin cells in the primary tumor, finding that they belong to advanced clones composed mainly of metastatic liver and stomach cells (Supplementary Fig. S7D). Although it is difficult to conclusively distinguish between reseeding from stomach metastases and primary spread to the stomach, several observations favor the reseeding hypothesis. These clones are enriched for more advanced classical mucin phenotypes and not the earlier classical mucin–HLA-II phenotypes (Fig. 2E; Supplementary Fig. S7E). Moreover, their copy-number profiles are more similar to cells that metastasized to the stomach and express the AC5 phenotype compared with other primary cells (Supplementary Fig. S7F). To confirm the clone assignments of primary cells expressing AC5, we examined the CNA profiles of cells from the two advanced AC5 clones harboring the most primary cells (clones I and J, Supplementary Fig. S7D) and found that they are more similar to the profiles of their assigned clones than to those of nonmetastasizing primary clones (Supplementary Fig. S7F). The primary cells in these clones exhibit similar assignment confidence values as the other cells (primarily stomach and liver) in their assigned clones (Supplementary Fig. S7F). In addition, these primary cells show a higher copy-number burden than other primary cells, reaching levels comparable with those of metastatic cells (Supplementary Fig. S7G and S7H). Thus, although few primary cells express the AC5 program, the combination of their advanced phenotype, greater similarity to stomach metastatic cells than to other primary cells, and elevated copy-number burden provides evidence consistent with reseeding from stomach metastases.

Another mucus production program, which includes robust expression of the transcription factor SPDEF and its targets AGR2 and ERN2, is highly expressed in precancerous lesions and classical tumor subtypes (79). We found that these genes are enriched in primary AC14, which also expresses high levels of HLA-II molecules, thus fully capturing the PDAC primary classical mucin–HLA-II phenotype (Fig. 3B; Supplementary Fig. S7B). Moreover, the classical mucin–HLA-II cells belong to the earliest clone in the phylogeny (Fig. 2B, D, and E), supporting that this program is indeed related to early PDAC stages, as reported in mouse models and laser capture microdissection epithelium from patients with PDAC (79). In a phase II first-line chemoimmunotherapy clinical trial in patients with advanced gastroesophageal adenocarcinoma, a gene program containing AC5 genes TFF1 and MUC5AC was the program most highly expressed by cancer epithelial cells in fast-progressing patients compared with slow progressors (80). In contrast, the expression of AC14 genes (HLA-II programs) by cancer epithelial cells was significantly higher in slow progressors (80). Our results suggest that PDAC cells can express at least two different mucin production programs—classical mucin–HLA-II captured by AC14, representing earlier-stage primary cells with a less aggressive prognosis, and classical mucin associated with AC5, representing later clones associated with greater metastatic potential or chemotherapy resistance.

Peritoneal metastases rewire lipid metabolism

AC2 consists almost entirely of cells from the peritoneum (Figs. 3B and 5A). The peritoneal cavity is the second most common site of metastasis in pancreatic cancer (63), but the mechanisms of metastatic initiation, progression, and adaptation remain poorly understood. Unlike hematogenous metastases to the liver or lungs, which typically present as discrete nodules or masses, peritoneal dissemination often occurs through trans-coelomic spread, leading to thin, diffuse layers over the omentum that escape detection (81). Peritoneal metastases are typically only diagnosed after reaching an advanced, treatment-refractory state known as peritoneal carcinomatosis, which accelerates cachexia—a syndrome characterized by malabsorption, significant weight loss, malignant ascites, and bowel obstruction—and the subsequent rapid decline limits opportunities for investigation.

Figure 5.

Lipid metabolic rewiring is a prominent feature of peritoneal metastases. A, UMAP embedding, colored by tissue site (left), and sample distribution and composition (right) of all AC2 cells. B, AC2 hotspot analysis, highlighting lipid metabolism and oxidative stress and detoxification modules. C, Digital pathology of H&E-stained primary and peritoneal metastasis tissue, showing the expansion of adipose tissue in the peritoneum. D, Quantification of adipose and fibrotic tissue in sections in C. E, Cancer clone phylogeny, indicating AC2-enriched clones (purple triangles), fractional tumor site composition for each clone (stacked bars), and the proportion of cells in each clone assigned to AC2 (outer circle).

We found that the two peritoneal metastases from opposite flanks of the patient both contribute substantially to AC2 (Fig. 5A) and have very similar transcriptomic profiles (median 33% of a cell’s kNN graph neighbors are from the other site), including strong upregulation of lipid metabolism genes compared with other archetype clusters (Supplementary Fig. S8A; Supplementary Table S3). Hotspot identified multiple gene modules, including one associated with fatty acid and cholesterol biosynthesis and another with oxidative stress and detoxification (Fig. 5B). Genes uniquely upregulated in AC2 include key players in cholesterol (TM7SF2) and fatty acid (ME1, IDH1) biosynthesis, aldo-ketoreductases (AKR1B10, AKR1C2, AKR1C3), prostaglandin regulators (PTGIS, PTGR1), and redox balance genes (GCLM, GCLC, GPX2, GSR, PIR, SLC7A11, TXNRD1, UGDH) that respond to oxidative stress triggered by lipid production and accumulation (Fig. 5B; Supplementary Fig. S8B). Genes involved in lipid droplet turnover (SQSTM1), lipid transport (ABCA10, ABCC3), and adipocyte differentiation (PLAC8) are also differentially upregulated in AC2 cells. As expected, AC2 genes are specific to the archetype cells and minimally expressed in nonarchetype cells in the peritoneum samples (Supplementary Fig. S6H). Lipid metabolic and oxidative stress genes are not expressed appreciably in tumor immune or stromal cells, confirming that their detection in cancer cells is not due to ambient peritoneal RNA (Supplementary Fig. S8B).

The peritoneal cavity is supported by metabolically active adipose tissue that is rich in free fatty acids and signaling molecules, including adipokines and cytokines (82). Digital pathology of peritoneal and primary tumor sections revealed a greater fraction of adipose tissue in peritoneal metastases than in primary samples, which are dominated by fibrotic stroma (Fig. 5C and D). Moreover, whereas cancer cells in primary PDAC tumors typically occur in multiple distinct pockets (83), they are interspersed among adipocytes in the peritoneal samples (Supplementary Fig. S8C). The upregulated genes associated with AC2 constitute many components of the lipogenic pathway (Supplementary Fig. S8D), by which fatty acids are synthesized for energy storage and cell membrane biosynthesis, primarily in the liver and adipose tissue. Thus, in contrast to the catabolic processes and patient-level wasting caused by cachexia, our observations suggest that metastatic PDAC cells respond to the adipocyte-rich peritoneal environment by upregulating lipid anabolism and oxidative stress detoxification. A lipogenic phenotype has been reported previously in PDAC cell lines (84) as well as in preclinical models and patients with PDAC (85), but its robust upregulation has not been previously associated with peritoneal metastasis.

To determine whether the lipid metabolic phenotype generalizes beyond the two independent samples in our patient, we obtained two postmortem peritoneal metastases from a different patient with PDAC and performed snRNA-seq and similar data analysis (Materials and Methods). Importantly, the second subject was 70 years old, succumbed to metastatic disease within 3 months of diagnosis, and did not receive treatment. Despite the markedly different clinical circumstances in these two cases, we found that fatty acid and cholesterol biosynthesis, as well as cholesterol metabolism and homeostasis, are the most significantly enriched gene programs in the second case (Supplementary Fig. S9A; Supplementary Table S9).

Lipid metabolic rewiring is not driven by genotype

We sought to understand whether the highly specialized phenotypes that dominate peritoneal metastases are due to clonal selection of genetically encoded adaptive traits or were acquired by epigenetically plastic cells in response to a novel environment. To help distinguish between these possibilities, we leveraged the two anatomically separate peritoneal metastases and the cancer phylogeny.

We hypothesized that if clonal selection—under the clonal evolution model (86)—drove the lipid anabolism phenotype, distinct clades of AC2 clones would map to each peritoneal site; after passing through the original selection bottleneck, cells at each site would accumulate unique sets of alterations over time due to genetic drift. On the other hand, if the lipid anabolism phenotype was due to plastic cells responding to the lipid-rich peritoneal environment, there would be no association between clone identity and peritoneal site, and diverse peritoneal clones could contain cells from opposite flanks of the body. We assessed which lipid anabolism–enriched clones (defined as >10% of cells with AC2 phenotype) belonged to each peritoneal metastasis in the phylogenetic tree and found 26 clones spread across all three major clades, including early branches with fewer CNAs as well as late branches (Fig. 5E; Supplementary Fig. S9B). Both pure and mixed clones are present in the independent peritoneal sites. For example, early clones enriched for lipid metabolism are derived from both peritoneum A (16%–68%) and B (8%–40%) sites, and late clones are derived from a mix of sites as well (69% to 83% peritoneum A and 8% to 12% peritoneum B). Intermediate clones are pure for either peritoneum site but still share the same clades, suggesting common ancestors in the primary tumor (Supplementary Fig. S9B).

The existence of diverse clones enriched for lipid anabolism over several branches of the tree—some populating both peritoneal sites, some specific to each site but belonging to the same clade—supports the hypothesis that cancer cell plasticity drove the lipid anabolism phenotype through phenotypic convergence to local environmental pressures.

Transcriptomic plasticity is a hallmark of PDAC metastasis

The plasticity we identified in peritoneal tumors involves multiple clones that express a diversity of additional archetypes, motivating a more systematic investigation of whether plasticity is a general feature of PDAC metastasis. Indeed, although each clone represents a shared genetic lineage, clones throughout the phylogeny do not seem to be constrained by their lineage and express a diversity of archetypal phenotypes, corresponding to high per-clone archetype entropy (Fig. 6A; Materials and Methods). This is in line with a lineage tracing study in a PDAC mouse model, which found that cell cycle and EMT cell states are not correlated with cellular phylogeny (49). We asked whether the diversity of archetypes that clones exhibit is greater than expected, which would indicate substantial phenotypic plasticity. The empirical distribution of per-clone archetype entropy (mean μ = 1.42) is shifted higher than expected under simulations in which the site is highly predictive of phenotype (μ = 0.97) but lower than expected under random assignment of archetypes to cells (μ = 2.42, Fig. 6B). This suggests that the variety of archetypes present in each clone is not driven by the diversity of tumor sites within each clone but rather by the ability of cells to acquire a range of phenotypes even within a single site.

Figure 6.

Transcriptomic plasticity is a common feature of metastatic cells. A, PDAC tumor cell clonal phylogeny (center), showing the fraction of cells from each site, the fraction of archetypes, and archetype entropy from inside to outside for each clone (leaf in the phylogeny). B, Entropy distributions for three null models and for data in this study (Materials and Methods). Bars indicate the number of clones for each binned entropy value (n = 62 clones for each distribution), curves represent smooth trends, and dashed vertical lines correspond to mean entropy. Observed clones have lower archetype entropy than clones with randomly assigned archetypes, but more than models based on strong archetype bias for metastatic sites, indicating high cellular plasticity. C, PDAC tumor cell phylogeny showing copy-number profiles of each clone, PLASTRO score (Materials and Methods), and archetype composition. The scatterplot indicates that a higher CNA burden is associated with higher plasticity. r, Pearson correlation.

To quantify plasticity more rigorously at the clonal level, we developed PLASTRO (Supplementary Fig. S10; Materials and Methods). PLASTRO compares evolutionary similarity (lineage distance) and phenotypic similarity (phenotypic distance with respect to archetype composition) between cells, based on the assumption that low cellular plasticity should result in a strong overlap between lineage and phenotype. Specifically, for a given clone, it quantifies the degree of discordance between phenotypic and phylogenetic neighborhoods while remaining insensitive to neighborhood size. Interestingly, we found that cells with few CNAs tend to have low PLASTRO scores, whereas more advanced clones bearing extensive CNAs score high for plasticity (Fig. 6C). Given that CNA burden correlates with metastasis in PDAC and other cancers (4), our finding that CNA burden is strongly associated with plasticity is consistent with a model whereby plasticity enables metastasis. To evaluate this effect more quantitatively, we performed a Mantel test (50), which assesses the correlation between two distance matrices. We observed a Mantel test statistic of 0.13 (p < $1\mathrm{\hspace{0.17em}}\mathrm{\times }10^{\mathrm{-}3}$) for matrices of phenotypic distances within distinct clones, suggesting that cells are plastic (the statistic ranges between −1 and 1, with 0 denoting no correlation), as their phenotypes differ significantly from those of their lineage.

Discussion

Rapid autopsy makes it possible to investigate clonal lineage histories and adaptive phenotypes in the ecosystem of a single cancer patient. In our comprehensive analysis of a patient with PDAC who underwent rapid autopsy, we evaluated the phenotypic landscape that a single cancer can occupy and developed computational approaches that bridge single-cell transcriptomics with phylogenetic reconstruction to dissect the relative contributions of clonal evolution and transcriptomic plasticity to metastatic adaptation. Our analysis reveals that transcriptional plasticity, rather than genetic evolution and selection, is the dominant force shaping metastatic phenotypes.

We note that we tested multiple tools for phylogenetic reconstruction from bulk WES data, but each produced a strikingly different topology. Moreover, a probabilistic approach, CONIPHER, yielded multiple divergent trees with similar likelihoods. CONIPHER is designed to detect subclonal structure from bulk data and consistently revealed extensive clonal mixing within each site and reseeding events to the primary tumor—supporting our own observations—but could not resolve a tree with more than a few branches that are supported across alternative trees, reinforcing the need for single-cell resolution in this context.

Although the patient in this study was diagnosed with metastatic PDAC at an unusually young age (35 years old), our molecular analyses are highly concordant with previously published datasets derived from larger and more typical PDAC cohorts (48, 78). Specifically, the transcriptional phenotypes observed in this patient’s primary tumor closely match subtype programs identified in a comprehensive study of 43 treatment-naive and neoadjuvantly treated PDAC specimens profiled using snRNA-seq and spatial transcriptomics. This molecular overlap suggests that the cellular programs we identified reflect conserved features of PDAC biology rather than patient-specific outliers.

A critical insight from our study is that successful metastatic clones exhibit remarkable phenotypic diversity, even within the same anatomic site. We demonstrate that genetically related metastatic clones can colonize multiple organs while manifesting diverse transcriptional states independent of their anatomic location. Moreover, each organ site harbored multiple phylogenetically distant clones, suggesting extensive parallel seeding. This evidence points to nongenetic plasticity as a key mechanism enabling metastatic cells to transition between different gene programs across metastatic sites, thereby enhancing their adaptability and survival. This plasticity is notably amplified in clones with a higher CNA burden, suggesting that genomic instability may facilitate transcriptional adaptation—not through specific mutations, but by creating a permissive state for phenotypic exploration. This observation aligns with recent findings that chromatin accessibility increases with genomic instability in various cancers, potentially enabling broader transcriptional responses to environmental cues (87, 88).

Our profiling of common but understudied metastatic sites in PDAC revealed distinct organ-specific adaptation programs, providing new insight into how cancer cells respond to diverse tissue environments. The acquisition of gastrointestinal programs by stomach wall metastases demonstrates remarkable cellular plasticity, suggesting that tumor cells can co-opt organ-specific transcriptional modules to enhance colonization and acquire fitness in new environments. Similarly, peritoneal metastases upregulate lipid anabolism and oxidative stress response pathways, suggesting that tumor cells adopt metabolic features of adipocytes and adapt their redox response to counteract reactive oxygen species generated by metabolic stress. This is consistent with prior studies showing that lipid metabolism plays a crucial role in PDAC progression (89) and chemoresistance (90). Both site-specific gene programs suggest that metastatic cells adapt to their microenvironment, possibly in response to stroma-derived signaling and environmental lipid availability (91). The convergent adaptation of these phenotypes across multiple independent clones strongly supports the role of microenvironmental pressures in shaping cellular phenotypes, independent of genetic evolution.

The methodologic advances developed for this study—particularly PICASSO for phylogenetic reconstruction and our approach to archetype analysis—provide a robust framework for similar investigations across cancer types. However, several critical questions emerge from our findings. How do specific tissue environments orchestrate the activation of adaptive programs? Can we target the mechanisms underlying cellular plasticity with therapies? What do genomic markers such as RAS amplification contribute, given the importance of plasticity as an emerging resistance mechanism to RAS therapies? Although we focused on epithelial cells and optimized their capture over other cell types, the role of stromal cells also remains an open question. Future studies combining spatial transcriptomics with single-cell lineage tracing could help address these questions and further illuminate the complex interplay between genetic inheritance and environmental adaptation in cancer progression.

Our findings emphasize the fundamental roles of cellular plasticity and metabolic adaptation in enabling the successful colonization of diverse organ sites. They suggest that effective therapeutic strategies must account for both genetic and nongenetic mechanisms of adaptation, potentially through approaches that constrain cancer cell plasticity or target site-specific vulnerabilities. These insights may guide the development of more effective treatments for metastatic disease, particularly for challenging sites such as peritoneal metastases that currently lack targeted therapeutic options.

Supplementary Material

Figure S1

Representative histopathology images and single-nucleus RNA-seq analysis showing cellular diversity, quality control, and phenotypic composition across primary and metastatic PDAC samples

Figure S2

Analysis of PDAC cells showing KRAS pathway activity, copy number alterations, and mutational profiles across metastatic tumor sites

Figure S3

Overview and benchmarking of IntegrateCNV and PICASSO methods for inferring single-cell copy number alterations and reconstructing subclonal phylogenies in metastatic PDAC

Figure S4

Performance benchmarking of PICASSO showing improved phylogenetic accuracy, computational efficiency, reproducibility, and robustness compared to neighbor-joining approaches

Figure S5

Overview of the archetype analysis framework and its application to single-cell transcriptomes from metastatic organ sites in the rapid autopsy PDAC dataset

Figure S6

Assessment of archetype robustness and biological significance, showing consistent archetype gene programs across downsampling tests, their association with MYC, lipid metabolism, and gastrointestinal programs, and comparison with clustering analysis

Figure S7

Analysis of archetype 5 cells suggesting possible reseeding of the primary pancreas tumor from the stomach metastasis, integrating expression programs, clonal assignments, and copy-number alterations

Figure S8

Expression and histological evidence of lipid metabolism in peritoneal metastases, showing up-regulation of lipid anabolism and oxidative stress pathways in archetype 2 cancer cells

Figure S9

Validation of lipid metabolic rewiring in peritoneal metastases using single-nucleus RNA-seq from an independent rapid-autopsy patient and phylogenetic analysis revealing lipid anabolism up-regulation in peritoneal lesions and archetype 2 clones across distant clades

Figure S10

Comparison of phenotypic and phylogenetic relationships using the PLASTRO metric, quantifying the alignment between archetype composition and clonal structure to assess metastatic plasticity

Table S1

Summary of genomic and sequencing statistics for the single-nucleus RNA-seq samples analyzed in the main rapid autopsy patient

Table S2

List of marker genes used to annotate cell types across primary and metastatic samples in the single-nucleus RNA-seq dataset

Table S3

Differentially expressed and modular genes for each archetype cluster, including upregulated genes (log₂ fold change > 1, p < 0.05) and Hotspot-identified informative genes with significant autocorrelation (FDR < 0.05)

Table S4

Gene set enrichment analysis (GSEA) scores for archetype cluster gene programs that were significantly enriched at FDR < 0.05

Table S5

Comprehensive gene set enrichment analysis (GSEA) results showing full enrichment scores for all archetype cluster gene programs

Table S6

Gene set enrichment analysis (GSEA) scores for Leiden cluster gene programs that were significantly enriched at FDR < 0.05

Table S7

Expression levels of archetype 5 gastrointestinal gene signatures across intestinal, stomach, and gallbladder epithelial cell types from the CELLxGENE reference atlas

Table S8

Percentage of cells expressing archetype 5 gastrointestinal gene signatures across intestinal, stomach, and gallbladder epithelial cell types from the CELLxGENE reference atlas

Table S9

Gene set enrichment analysis (GSEA) scores for archetype neighborhood cells significantly enriched (FDR < 0.05) in peritoneal metastases from the rapid autopsy case RA19_21

Table S10

Expert-curated gene sets and associated references compiled for pathway and enrichment analyses performed in this study

Table S11

Names of samples, whole-exome sequencing EGA file IDs, and snRNA-seq HTAN file IDs

Data Availability

The snRNA-seq data generated in this study are publicly available in the Human Tumor Atlas Network (HTAN; ref. 92) Data Portal (93) at https://data.humantumoratlas.org/publications/hta8_2025_biorxiv_alejandro-jim%C3%A9nez-s%C3%A1nchez. The WES data (BAM files) generated in this study are publicly available through the European Genome-phenome Archive (EGA) as part of the EGAD00001011109 dataset (multiregion sequencing of patients with PDAC) and can be accessed at https://ega-archive.org/datasets/EGAD00001011109. The names of the samples used in this study, snRNA-seq HTAN IDs, and the WES EGA IDs are listed in Supplementary Table S11. Each WES EGA ID relates to two or more snRNA-seq HTAN IDs because different experimental protocols were applied during the nuclei extraction prior to the snRNA-seq encapsulation process (see Materials and Methods). The data analyzed in Fig. 4C and Supplementary Tables S7 and S8 in this study were obtained from the CZ CELLxGENE Discover database (46) at https://cellxgene.cziscience.com/gene-expression. The IntegrateCNV algorithm, along with documentation, notebooks, and tutorials, is available at dpeerlab/integrateCNV. The PICASSO algorithm, as well as documentation and tutorials for inferring CNA phylogenies and visualizing transcriptional and phenotypic information alongside the tree, is available at https://github.com/dpeerlab/picasso. Code for computing the PLASTRO metric, as well as documentation and tutorials, is available at https://github.com/dpeerlab/PLASTRO. All other raw data are available from the corresponding authors upon request.

Authors’ Disclosures

A. Mehta reports personal fees from The Column Group and Third Rock Ventures outside the submitted work. W. Park reports grants from the NIH/NCI, Merck, Lepu Biopharma, Amgen, Break Through Cancer, The Society of MSK, Parker Institute for Cancer Immunotherapy, Society for Immunotherapy of Cancer, and EXACT Therapeutics and grants and personal fees from Astellas, Revolution Medicines, and Astellas outside the submitted work. R. Chaligne reports personal fees from Gerson Lehrman Group and nonfinancial support from LevitasBio and Sanavia Oncology outside the submitted work. E.M. O’Reilly reports grants from Genentech/Roche, BioNTech, AstraZeneca, Arcus Therapeutics, Elicio Therapeutics, Digestive Care, Agenus, Amgen, Revolution Medicines, Break Through Cancer, and the NIH/NCI during the conduct of the study as well as other support from Arcus, Amgen, AstraZeneca, Ability Pharmaceuticals, Alligator Biosciences, Pfizer, Agenus, Ipsen, Ikena, Immuneering, Moma Therapeutics, Novartis, Astellas, BMS, Revolution Medicines, Regeneron, Tango Therapeutics, and Arcus and personal fees and other support from BioNTech outside the submitted work; and American Association of Cancer Research (AACR), American Society of Clinical Oncology (ASCO), Imedex, Research To Practice, and Stand Up To Cancer (SU2C). C.A. Iacobuzio-Donahue reports other support from Bristol Myers Squibb outside the submitted work. D. Pe’er reports ownership/equity interests in Insitro, Inc. No disclosures were reported by the other authors.

Disclaimer

The Editor-in-Chief of Cancer Research is an author on this article. In keeping with AACR editorial policy, a senior member of the Cancer Research editorial team managed the consideration process for this submission and independently rendered the final decision concerning acceptability.

Authors’ Contributions

A. Jiménez-Sánchez: Data curation, software, formal analysis, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing. S. Persad: Software, formal analysis, investigation, visualization, methodology, writing–original draft, writing–review and editing. A. Hayashi: Data curation, investigation. S. Umeda: Formal analysis, validation, investigation, methodology. R. Sharma: Data curation, software, formal analysis, supervision, methodology, writing–original draft, writing–review and editing. Y. Xie: Data curation, software, formal analysis, investigation, visualization, methodology. A. Mehta: Writing–original draft. W. Park: Data curation, methodology. I. Masilionis: Data curation, methodology. T. Chu: Data curation, formal analysis, investigation, methodology. F. Zhu: Software. J. Hong: Formal analysis. R. Chaligne: Data curation, methodology. E.M. O’Reilly: Supervision. L. Mazutis: Data curation, methodology. T. Nawy: Writing–original draft, project administration, writing–review and editing. I. Pe’er: Supervision, methodology. C.A. Iacobuzio-Donahue: Conceptualization, resources, supervision. D. Pe’er: Conceptualization, resources, supervision, funding acquisition, methodology, writing–original draft, project administration, writing–review and editing.