Integrated spatial morpho-transcriptomics predicts functional traits in pancreatic cancer

Abstract

Analyses of patient-derived cell lines have greatly enhanced discovery of molecular biomarkers and therapeutic targets. However, characterization of cellular morphological properties is limited. We studied cell morphologies of human pancreatic adenocarcinoma (PDAC) cell lines and their associations with drug sensitivity, gene expression, and functional properties. By integrating live cell and spatial messenger RNA imaging, we identified KRAS inhibitor–induced morphological changes specific for drug-resistant cells that correlated with gene expression changes. We then categorized a large panel of patient-derived PDAC cell lines into morphological and organizational subtypes and found differences in gene expression, therapeutic targeting potential, and metastatic proclivity. Patterns of cancer cell organization in human PDAC tissues stratified distinct gene expression signatures with clinical significance. In summary, we highlight the potential of cell morphological information in rapid, cost-effective assays to aid precision oncology efforts leveraging patient-derived in vitro models and tissues.

Changes in pancreatic cancer cell shape are correlated with treatment response and are associated with cellular functions.

INTRODUCTION

The size, shape, and organization of cells in culture represent a fundamental characteristic that distinguishes cellular identity, state, and function (1). Thousands of cell line models have been derived over the past century (2), each having distinct structural and organizational features that correspond to phenotype and tissue lineage (3). Cancer cell lines have previously been described to exist on a spectrum between epithelial-like and mesenchymal shapes (4, 5). More aggressive and invasive cancer cells often exhibit mesenchymal morphology (6, 7), a hallmark of the epithelial-mesenchymal transition (EMT) (4). Still, a broad degree of morphologic and organizational variation is not adequately captured by this simple binary schema.

More detailed characterization of cell morphology can provide a high-dimensional “fingerprint” associated with behaviors such as movement, response to perturbation, and cell death. Diverse patterns of motility, blebbing, cell adhesions, axon-like projections, nuclear morphology, and lipid metabolites can all be visualized by phase imaging (8). Cancer cells can change their morphology in response to drug treatment, which may highlight targetable drug resistance mechanisms. For instance, cells treated with DNA damaging agents begin to express DNA damage response proteins and undergo senescence to stop replication (9), taking on a distinctive flat and irregular morphology with increased cell area. Genetic alterations can also lead to changes in cell morphology. For example, mutations in genes regulating the cytoskeleton (e.g., Rho family of guanosine triphosphatases) can alter cell shape and motility (10), and cell morphology can even predict mutational status (7). Microenvironmental ligands such as transforming growth factor–β (11) as well as culturing conditions including mechanical substrate and extracellular matrix (ECM) composition (12, 13) also have potent morphological remodeling effects.

Large-scale image-based screens have recently enabled comprehensive analyses of the effects of various perturbations on cell morphology in single cell lines (14–16). However, the baseline heterogeneity in morphology across different cell lines remains largely unquantified, presenting an opportunity to integrate these measurements with gene expression profiling and dependency resources like the Cancer Cell Line Encyclopedia (CCLE) (17, 18). In addition, the shared and distinct morphological changes in response to perturbations across diverse cell lines are not well understood. An alternative screening strategy, assessing the remodeling effects of a few perturbations across many cell lines, has not yet been conducted at scale (19). Such an approach could reveal associations between morphology and -omics data, such as dependency and gene expression, and how these relationships shift under perturbation.

To address this gap in knowledge, we studied a large panel of human pancreatic adenocarcinoma (PDAC) cell lines in response to clinically deployed compounds including KRAS inhibitors (e.g., MRTX1133 and RMC-6236) and chemotherapy [e.g., 5-fluorouracil (5-FU) and gemcitabine]. We pioneered a single-cell framework, Spatial Morphology and RNA Transcript (SMART) analysis, to integrate cell morphology and transcriptomic state using a combination of holotomography (Nanolive 3D Cell Explorer 96focus), Cell Painting (20), and spatial molecular imaging (SMI; Bruker/NanoString CosMx) (21). Using SMART, we provide evidence of distinct cytoskeletal adaptations in response to KRAS inhibition and chemotherapy that correlate with drug sensitivity.

More broadly, to holistically evaluate the translational importance of cell morphology in patient-derived model systems, we performed analysis of -omics data from the Cancer Dependency Map (DepMap) (22), functional experiments including clonogenicity and invasion assays, and the Cell Painting assay to quantify and categorize the morphological and organizational states present in PDAC cell lines. We identified molecular features of cells associated with their respective morphologies, finding that genetic dependency is predictive of morphology using an XGBoost (23) machine learning model. Using our categorization schema, we performed functional assays to find morphologies and organizational patterns associated with clonogenicity, invasion, and metastasis. Last, we identified molecular correlates to in situ tissue morphology and identified strong associations between the basal-like transcriptional subtype (24), which is associated with worse prognosis and treatment resistance (25), and small cell clusters, as well as the classical transcriptional subtype and large cell clusters. Together, we provide a robust framework for integrating morphological profiling into -omics workflows, offering a powerful approach to uncover biomarkers, characterize drug responses, and elucidate mechanisms of therapeutic resistance.

RESULTS

SMART analysis to characterize treatment-induced remodeling of cellular morphology and transcriptomic state

Cell morphology is a dynamic and cell state–specific biomarker that can be investigated at different length scales and by visualizing different molecular and structural features. To comprehensively capture morphological phenotypes and their molecular correlates, we developed a multimodal framework using multiple independent assays, which we termed SMART analysis (Fig. 1A). The collection of assays in the SMART framework can be integrated to answer a broad set of questions relating to cellular morphology. First, we leveraged an ultrahigh (200 nm) resolution label-free live cell phase imaging technology called holotomography (Nanolive 3D Cell Explorer 96focus) to measure cell shape dynamics in response to perturbations such as drug treatment. In addition, we used a Cell Painting assay that provided higher throughput characterization and quantification of structural features such as actin fibers. To directly correlate morphology and associated transcriptional features, we integrated SMI (Bruker/NanoString CosMx), which enables high-plex spatial transcriptomics in addition to static imaging data, to link morphology and transcriptome. Each of our SMART measurements provides single-cell data, overcoming both inter– and intra–cell line morphologic heterogeneity. Furthermore, in contrast to dissociative single-cell approaches, which perturb cell morphology and transcriptomic state, our in situ SMI method directly enables accurate cell shape and transcriptome measurements in a single assay, while preserving cell-cell interactions that may influence cell state.

Fig. 1. Dynamic morphologic and organizational remodeling in PDAC cell line models in response to KRAS inhibition by phase holotomography.

(A) Schematic of SMART analysis framework. (B) Representative images of cell lines profiled with phase holotomography following 48 hours of KRAS inhibitor (KRASi; RMC-6236) treatment and under baseline conditions. Scale bar, 20 μm. Panc0203 MR refers to the Panc0203 parental cell line grown to resistance to MRTX1133 in increasing drug concentration (37). (C) Quantification of cell object area and density throughout the treatment course with RMC-6236, stratified by RMC-6236 sensitivity. (D) Holotomography and clustering analysis of an additional KRAS inhibitor sensitive cell line, AsPC1, under KRAS inhibitor treated and untreated conditions. (E) Quantification of cell death of AsPC1 cells under treated and untreated conditions. Cycle length is 20 min. (F and G) Quantification of AsPC1 cells in clusters versus singlets under KRAS inhibitor treated and untreated conditions. Statistical significance assessed using Fisher’s exact test.

We applied SMART to study the treatment response of PDAC cell lines to KRAS inhibitors, due to their emerging clinical relevance in this highly treatment-refractory disease. More than 90 to 95% of PDACs harbor KRAS gain-of-function mutations (26), which have recently become targetable using both allele-specific and pan-KRAS mutation small-molecule inhibitors (24, 27). Specifically, we chose RMC-6236, a RAS(ON) multiselective inhibitor of KRAS currently in phase 2 trials (27), and MRTX1133, an allele-specific KRAS^G12D inhibitor previously in phase 2 trials (28) for our studies. We hypothesized that changes in morphology may serve as a predictive biomarker for early cellular response to a drug perturbation and shed insights on potential resistance mechanisms.

First, we performed live cell holotomography to visualize morphologic and organizational changes in patient-derived PDAC cell lines (Fig. 1B and Materials and Methods). In a panel of six cell lines, four of which are resistant [area under the dose-response curve (AUC) > 0.65] and two of which are sensitive (AUC < 0.50) to KRAS inhibition, we identified distinct morphologic changes associated with treatment sensitivity. Resistant cell lines enlarged their cell area without increasing cell mass, while sensitive cell lines clustered together forming tightly aggregated clumps that reduced in size over time, likely due to cell death (Fig. 1, B and C). To validate our findings, we took another sensitive cell line, AsPC1, and treated it with the KRAS^G12D specific inhibitor MRTX1133 before performing live cell holotomography (Fig. 1D, left). Similar to the sensitive cell lines Patu8988S and Panc0203, we observed some cell death following 24 hours of treatment (Fig. 1E), but most notably, there was induction of tight clustering in response to KRAS inhibition (Fig. 1D).

To further investigate the causes of treatment-associated morphologic changes, we harnessed another SMART assay, SMI, to make concurrent spatially resolved morphology and transcriptomic measurements, which can be integrated with the dynamic morphological changes captured by holotomography (Fig. 1B and Materials and Methods). We performed SMI on AsPC1 cells treated with MRTX1133 to quantify clustering over a larger imaging window and to identify gene expression changes in clusters versus singlets (Fig. 1D, right). Globally, there were significantly more cells in clusters of five or more under the treated versus untreated conditions and significantly fewer singlets, defined as cells without any touching neighbors (Fig. 1, F and G). Spatially resolved gene expression analysis identified an increase in transcripts encoding cell junction proteins including various integrins (e.g., ITGB1, ITGA3, and ITGA6) in cell clusters relative to singlets and under the treated condition relative to the untreated condition.

Given the heterogeneity in treatment response across resistant and sensitive cell lines, we sought to characterize and quantify treatment-induced morphological changes at a broader scale. We first assembled a panel of 15 PDAC cell lines with a range of treatment sensitivities to RMC-6236 and MRTX1133. We then performed modified Cell Painting (Materials and Methods) to directly visualize the morphology of individual cells after 72 hours of drug treatment. In addition to KRAS inhibitors, we included additional compounds with clinical relevance to PDAC, including 5-FU, a thymidylate synthase inhibitor that is the backbone of the widely used FOLFIRINOX multiagent chemotherapy regimen (29, 30), and gemcitabine, a nucleoside analog, used as part of an alternate chemotherapy regimen (31). We performed confocal fluorescence imaging with a 20× objective on each cell line in both the untreated [dimethyl sulfoxide (DMSO) carrier control] and drug-treated settings, totaling 75 conditions in arrayed format. Using CellProfiler (32), we segmented 14,488 cells and quantified cell morphology, organization, and staining intensity features for downstream analyses (fig. S1A).

We performed principal components analysis (PCA) and calculated a uniform manifold approximation and projection (UMAP) on normalized morphological measurements (fig. S1B). Despite a limited set of measurements that precluded clustering of cells by their cell line identity, we observed patterns of UMAP proximity that varied according to the drug treatment that cells were subjected to (fig. S1C). Measurements related to cell size as well as compactness/eccentricity, a measure of deviation from a perfectly circular shape, were represented as top features in the first two principal components (fig. S1, D and F).

There was global treatment-induced remodeling of morphology under all treatment conditions, especially within the cell area and eccentricity measurements (Fig. 2, A to D). These changes were consistent across a separate larger cohort of cells (n = 147,318) imaged with a 10× objective (fig. S1G). Of the two KRAS inhibitors, RMC-6236 treatment exhibited the more substantial morphological shifts with highly correlated changes in cell area and actin intensity (Fig. 2, A and E). However, these shifts were heterogeneous across cell lines, with 9/15 cell lines demonstrating significantly increased cell area or actin intensity (Fig. 2B). In contrast, gemcitabine caused statistically significant increases in eccentricity in almost all cell lines tested (13/15) (Fig. 2B). Certain cell lines including Panc1, KP4, MIA PaCa2, and Patu8988T exhibited a marked increase in cell area upon treatment with RMC-6236 while others such as AsPC1, Patu8988S, and Panc0403 showed no change or even featured decreased cell area in response to KRAS inhibition, supporting our phase holotomography findings (Fig. 1B). These patterns of morphological change showed a correlation with sensitivity to RMC-6236 (Fig. 2, F and G), suggesting that treatment resistance can manifest as morphologic changes. Our results indicate that cytoskeletal changes are a critical indicator of response to treatment and motivate additional experiments to determine whether targeting cytoskeletal remodeling via focal adhesion kinases has synergy with KRAS inhibition.

Fig. 2. Treatment-associated remodeling of cell morphology by Cell Painting.

(A) Boxplot of calculated eccentricity, cell area, and actin integrated intensity features stratified by treatment. Statistical significance assessed by Mann-Whitney U test. (B) Change in eccentricity, area, and actin intensity measurements for each cell line. Blue dots represent significantly increased metric, red dots represent significantly decreased metric, and gray denotes lack of significant change. Statistical significance assessed by Mann-Whitney U test. (C) Representative image of MIA PaCa2 cells treated with RMC-6236 or carrier control. Scale bars, 50 μm. (D) Representative image of Psn1 cells treated with gemcitabine or carrier control. Scale bars, 50 μm. (E) Correlation of actin intensity and cell area fold change when treated with RMC-6236 across the panel of cell lines. Each dot represents an average of measurements across all profiled cells within each cell line. (F and G) Correlation of change in actin intensity and cell area (y axis) with RMC-6236 drug sensitivity (x axis) measured in PRISM pooled viability assay (DepMap). Statistical significance assessed using Student’s t test on Pearson correlation coefficient.

Gemcitabine elicited the largest change in eccentricity relative to the untreated condition, suggestive of senescent morphology driven by cell cycle arrest. This was characterized by increased cell area, increased abundance of cell projections, reduced cell-cell contacts, and increased actin staining intensity (Fig. 2, A and D). These changes also occurred in the 5-FU–treated cells, indicating a potential class effect of chemotherapy, but the magnitude of changes was not as substantial as with gemcitabine at the median inhibitory concentration for each drug. Senescence-associated morphological changes have previously been observed in PDAC cell lines and DNA damaging agents, such as gemcitabine, are at least additive with senolytic agents in the preclinical setting (33).

In summary, we have developed and applied morphologic profiling assays to study treatment-dependent changes in cellular morphology in response to KRAS inhibition and chemotherapy. We identified uniform increases in eccentricity under treatment with gemcitabine and drug sensitivity–dependent changes in cell area and actin staining intensity in response to KRAS inhibitors.

SMI enables morpho-transcriptomic measurements of pooled cell lines

Our next goal was to link morphological shifts to transcriptional content across our entire panel of cell lines. To achieve this scale in a cost-effective manner, we applied our SMI assay on pools of cells encompassing all 15 profiled PDAC cell lines under each of our previously defined treatment conditions (Materials and Methods). Briefly, cell lines were stained with nuclear [4′,6-diamidino-2-phenylindole (DAPI)] and cell membrane (CD298/B2M) markers, which were used to generate a segmentation mask and assign imaged transcripts to single-cell profiles (fig. S2, A to C). We used Panc1 cells to optimize our assay and detected ~700 unique genes and a mean of 7000 to 8000 transcripts per cell using a 1000-plex RNA imaging panel (fig. S2D). We tested slide coating with poly-d-lysine to improve cell adherence. Coated slides generally had better cell adherence and nearly equivalent performance metrics (fig. S2, D and E), enabling us to analyze approximately fivefold more cells per assay. In our pooled assay, we were able to achieve a mean of 429 genes and 5482 transcripts per cell after filtering, with similarly strong reproducibility across technical replicates (fig. S2, F and G). The overall lower genes and counts per cell in the pooled assay is likely secondary to relative differences in cell size and transcript content across cell lines (fig. S2H). These metrics were consistent across treatment conditions as measured by correlation of mean transcript counts per gene across technical replicates (fig. S2I).

To analyze this high-dimensional dataset, we first performed Leiden clustering and visualized a UMAP projection with distinct clusters (Fig. 3A) (34, 35). We developed a similarity metric based on ridge regression to map bulk RNA sequencing (RNA-seq) profiles from DepMap onto clusters to determine whether clusters separate by cell lines (fig. S3A and Materials and Methods). Most clusters cleanly mapped to a single cell line (fig. S3B). For example, the group of cell clusters labeled 1, 2, 24, and 21 all cleanly mapped to Suit2, and clusters 3 and 6 cleanly mapped to MIA PaCa2. However, some clusters mapped to multiple cell lines, making definitive annotation more difficult. We ultimately assigned clusters based on their optimal mapping and calculated the composition of the cell line mixture across each treatment condition (Fig. 3B and Materials and Methods). We detected all cell lines under each treatment condition and identified expansion or contraction of cell populations that was consistent with their known sensitivity to certain compounds. For example, RMC-6236–resistant cell lines KP4, Patu8988T, and Panc1 all expanded relative to other cell lines under the RMC-6236 condition, while sensitive lines including MIA PaCa2, Patu8988S, and Psn1 all contracted (Fig. 3B).

Fig. 3. Correlation of morphology and transcriptional state at single-cell resolution.

(A) UMAP projection of single-cell transcriptome profiles labeled by cluster assignment (left), treatment condition (middle), and cell line annotation (right). (B) Stacked barplot depicting cell line distributions across treatment conditions. Bars are represented as proportions and are colored according to cell line. Bolded boxes denote the top five cell lines resistant to the associated drug by AUC. (C to E) Histograms and probability distribution functions (black lines) of the aspect ratio (C), cell area (D), and cluster sizes (E) as defined by DBSCAN (Materials and Methods). (F) Heatmap depicting GSEA enrichment score (color) of cell state signatures (36) in each morphologic comparison (rows). (G) Clustered heatmap of enriched gene sets (columns) in RMC-6236 treated versus untreated conditions in each cell line (rows). Color density correlates with GSEA enrichment magnitude from significance P values (Materials and Methods). (H) Volcano plot of enriched genes under RMC-6236 treated condition across all cell lines. Highlighted blue colored dots are related to inflammatory signaling mentioned in text.

In addition to annotating cell lines, we also sought to stratify cells by their morphological features with the goal of identifying transcriptomic features associated with differences in cellular morphology. We defined three feature sets: cell area, cell aspect ratio, and cell aggregation (fig. S3C). We plotted distributions of these variables (Fig. 3, C to E, and fig. S3, D and E) and stratified cells into groups. We compared cells with aspect ratio greater than 1.5:1 (termed oblong; 26.4% of total cells) with those exhibiting aspect ratios of 1.2:1 or below (termed spherical; 40.7% of cells). For cell area, we divided cells into bottom (<575 μm²) or top (>1440 μm²) quartile. Last, we stratified cell aggregation using the spatial clustering algorithm, density-based spatial clustering of applications with noise (DBSCAN) (Materials and Methods). Cells with minimal contact with neighboring cells (singlets; 20.6% of total cells) were compared to cells that aggregated into clumps of five or more cells (56.5% of total cells).

Notably, distinct expression patterns of cell state signatures, including morphobiotypes (36) derived from in situ microdissected human PDAC specimens, were observed across these strata. Oblong shape was associated with a signature derived from microdissected PDAC cells with undifferentiated appearance [normalized enrichment score (NES): 1.40, P_adj. = 0.024]. Cells with large areas were enriched for EMT signatures including a signature derived from microdissected PDAC cells with structures transitioning away from ductal morphology (NES: 1.58, P_adj. = 3.18 × 10⁻⁴) and the hallmark EMT gene set (NES: 1.68, P_adj. = 1.35 × 10⁻⁵) but depleted for glandular morphology signatures (NES: −1.96, P_adj. = 1.58 × 10⁻¹⁰). The association of EMT signatures with enlarged cell area is consistent with our finding that cells resistant to KRAS inhibition increase their cell area. EMT signatures were previously found to be strongly enriched in mouse models of PDAC treated to resistance with MRTX1133 (37). Last, cells in clusters were enriched for glandular signatures (NES: 1.40, P_adj. = 0.024) and depleted for EMT signatures (hallmark EMT NES: −1.81, P_adj. = 4.24 × 10⁻⁶; transitional morphobiotype NES: −1.73, P_adj. = 7.98 × 10⁻⁵) (Fig. 3F).

To study treatment effects on transcriptional and morphological features, we performed gene set enrichment analysis (GSEA) on the top differentially expressed genes between the treated and untreated cells under each condition (fig. S4, A to D). Notably, we observed several KEGG (Kyoto Encyclopedia of Genes and Genomes) terms enriched under the RMC-6236–treated condition, including focal adhesion (P_adj. = 1.68 × 10⁻⁶), phosphatidylinositol 3-kinase (PI3K)–Akt signaling (P_adj. =6.25 × 10⁻⁵), and ECM-receptor interaction (P_adj. = 3.32 × 10⁻⁴). PI3K pathway activation has been observed in mouse tumors treated with tool compound RMC-7977, an analog of RMC-6236, and in human tumors treated with RMC-6236 (38, 39). KRAS withdrawal models as well as KRAS mutated PDAC lines with primary resistance to KRAS inhibitors have also been reported to use ECM-derived focal adhesion signaling as an acute resistance mechanism (40, 41).

In response to chemotherapy, we observed increased expression of inflammatory factors (fig. S4, A and B). This finding is consistent with several known chemoresistance mechanisms including extracellular release of damage-associated molecular patterns, which trigger secretion of interleukins and chemokines (42), and the senescence-associated secretion of inflammatory cytokines, chemokines, and growth factors (43). Cell adhesion–related gene sets (e.g., regulation of cell-cell adhesion) were enriched across all treatment groups relative to untreated cells. The cell adhesion response has commonly been reported as an ECM-mediated drug resistance mechanism to multiple classes of drugs (44, 45).

Morphologically, we observed that all pooled cell lines increased their cell area in response to gemcitabine (fig. S4E), similar to our arrayed Cell Painting experiments (Fig. 2 and fig. S4, F to H). However, certain shifts in morphology, especially under KRAS inhibitor–treated conditions, occurred in a cell line–dependent manner also similar to observations in the arrayed Cell Painting assay (Fig. 2 and fig. S4, F to H). To investigate heterogeneity in the transcriptional response to KRAS inhibition, we performed differential expression analysis in a cell line–specific manner. Using the top differentially expressed genes under the treated versus untreated condition for each cell line, we performed GSEA and clustered all the enriched terms. We visualized the clusters as a heatmap (Fig. 3G) to identify shared and cell type–specific changes in gene expression. Several modules emerged including gene sets relating to collagen processing, the extracellular matrix, integrin and focal adhesion signaling, immune response signatures for cytokine, interleukin, and interferons, as well as oncogenic growth signaling pathways (e.g., PI3K, Ras, and Rap1) (Fig. 3G).

Consistent with the known immunostimulatory effect of KRAS inhibition (46), major histocompatibility complex (MHC) I expression was significantly increased in all cell lines (Fig. 3H). Other differentially expressed genes enriched in KRAS inhibitor–treated cells included PSAP, which encodes a secreted glycoprotein previously reported to reduce lymphocyte infiltration (47); TNFSF10, which encodes TRAIL, another secreted molecule capable of inducing cancer cell apoptosis; B2M, which encodes an MHC I subunit; and CD74, which encodes a portion of the MHC II complex (Fig. 3H). Interferon response genes including IFI6, ISG15, ITM2B, IFIT1, IFIT3, and IFITM1 were also significantly enriched in a subset of cell lines. In particular, MIA PaCa2, which is sensitive to RMC-6236 but had a marked increase in cell size when treated with RMC-6236 (Fig. 2C), had one of the highest enrichment of interferon gene signature expression as measured by SMI. Thus, our pooled SMART assay recovered both shared and cell line–specific transcriptional and morphologic responses to KRAS inhibition and chemotherapy.

Together, we demonstrate how an image-based spatial transcriptomics platform can identify both morphologic and transcriptomic changes in individual cell lines within a larger pool. Our method uncovered cell type–specific responses to acute drug treatment such as PI3K and focal adhesion pathway signaling and identified specific transcriptional features associated with differences in cellular morphology.

Defining a morphologic and organizational classification of human PDAC models

Our next goal was to holistically evaluate the translational importance of cell morphology in patient-derived cell line avatars as a potential biomarker. We reasoned that it is essential to first define the landscape of morphologic diversity across patient-derived models. In addition to our cohort of 15 human PDAC cell lines used for treatment-induced morphologic analysis, we identified 34 additional cell lines previously profiled and published in DepMap with publicly available phase-contrast images on the DepMap portal and elsewhere (18, 48, 49). These patient-derived cell lines encompass a broad range of age, sex, tumor site, and genetic diversity and thus provided a diverse cohort for morphologic, transcriptomic, and functional characterization in this study.

We categorized cell lines by their predominant morphological appearance (i.e., polygonal, irregular, and spheroid) and organizational pattern with neighboring cells (i.e., aggregated, multilayered, and dispersed) (Fig. 4A and fig. S5). Polygonal cells had well-formed edges typically associated with aggregation into distinct clumps of cells with close contact among neighbors, but without physical contact with other clumps. Irregular cells featured nonsymmetrical shapes, typically with a spindle-like appearance and/or projections emanating from the cell body. Last, spheroid morphology was characterized by spherical cells. There were also distinct organizational patterns. The aggregated pattern was characterized by close contacts among cells in a cluster, such that discernible gaps were not visible. The multilayered pattern was characterized by cells growing on top of each other in the z plane. Last, the dispersed pattern describes a configuration where the cells were more evenly distributed over the plated surface compared to the aggregated and multilayered patterns.

Fig. 4. Transcriptional and dependency correlates of morphology.

(A) Categorization of cell morphologies (polygonal, irregular, and spheroid) and organizational patterns (aggregated, multilayered, and dispersed) of cell lines in our study. Categories were assigned based on microscopic visualization in culture and image analysis (CellProfiler). Scale bars, 100 μm. (B) Volcano plot depicting differentially expressed genes between polygonal and irregular cell lines using bulk RNA-seq data from the CCLE. y-axis P values are assessed by performing Mann-Whitney U test of expression (transcripts per million, TPM) in irregular versus polygonal cell lines. (C) Barplot highlighting top five GSEA terms enriched in polygonal (left) and irregular (right) cell lines. (D) GSEA enrichment plot of cell cycle gene set in polygonal cell lines relative to irregular cell lines. FDR, false discovery rate; Pval, p-value. (E) GSEA enrichment plot of cell junction organization gene set in polygonal cell lines relative to irregular cell lines. (F) Cell Painting of Patu8988S and Patu8988T. Scale bars, 100 μm. (G) Boxplot of CDH1 and VIM transcript expression stratified by polygonal or irregular morphology. Statistical significance assessed by Mann-Whitney U test. (H) Volcano plot depicting dependency differences between polygonal and irregular cell lines. Effect size (x axis) is calculated as the difference in mean dependency. y-axis P values are assessed by performing Mann-Whitney U test of gene dependency (Chronos score) in irregular versus polygonal cell lines. (I) Boxplots of notable dependency differences (TRA2B, DHODH, CDK6, and ARHGEF7) stratified by morphology. Statistical significance assessed by Mann-Whitney U test.

The aggregated organizational pattern was most closely aligned with the polygonal cell morphology, although certain cell lines with irregular and spheroid morphology exhibited heterogeneous organization (e.g., some cells aggregated, others dispersed). Morphologies were generally conserved across passages, freeze-thaw cycles, different standard medium conditions [e.g., RPMI or Dulbecco’s modified Eagle’s medium (DMEM)–based complete medium], and cell confluency. Notably, cellular organization was partially dependent on cell confluency, with ultralow confluences supporting more dispersed organization and high confluency forcing some degree of aggregation in all cell lines. Thus, our organizational and morphologic classifications were performed at intermediate confluence during the log growth phase (i.e., 40 to 70% confluency).

Morphologic classifications have distinct dependency and gene expression profiles

We analyzed transcriptomic correlates of cell morphology by querying RNA expression data associated with each of the 49 PDAC cell lines in our combined cohort. We first focused on differences between polygonal and irregular morphology given the abundance of these morphologies in our cohort (n = 26 polygonal and n = 17 irregular). We performed differential expression analysis (Fig. 4B) and GSEA of the differentially expressed genes (Fig. 4, C to E). Relative to irregular cells, polygonal cell lines were enriched for cell junction genes (NES: 1.716, P_adj. < 0.0001), as well as gene sets for keratinocyte development and proliferation. In contrast, irregular cell lines were broadly enriched for cell cycling gene sets (NES: 2.511, P_adj. < 0.0001) including chromosome segregation, DNA replication, and microtubule cytoskeletal organization. While there were a limited number of cell lines with spheroid morphologies (n = 6), we noted that, compared to both polygonal and irregular cell lines, they had the highest enrichment of cell junction and adhesion formation genes, as well as neuron projection development (fig. S6A).

Polygonal cell lines were notably enriched for glutathione peroxidase activity, which protects against oxidative stress. GPX4, a key enzyme in this family, was previously identified as a selective dependency in mesenchymal cell lines due to their increased reliance on fatty acid metabolism (50, 51). Despite these findings, the expression of GPX4 and other key regulators of fatty acid transport, including CYP4F2, CYP4A11, ACSL1, and ACSL5, also have significantly reduced expression in more mesenchymal appearing irregular cells relative to polygonal cell lines. The increased dependency on GPX4 in mesenchymal cell lines, despite reduced expression of its encoding gene and related pathway components, suggests that mesenchymal cells are unable to buffer against reduction in GPX4 levels and thus have greater sensitivity to GPX4 inhibition (52).

Patu8988S and Patu8988T are two cell lines derived from the same PDAC liver metastasis that exhibit distinct morphological, functional, and transcriptional properties (Fig. 4F). The Patu8988S cell line features polygonal morphology, aggregated organization, and high levels of CDH1 consistent with its overall epithelial phenotype (53) (Fig. 4, B and G). In contrast, Patu8988T has irregular morphology and dispersed organization, expresses higher levels of VIM, and exhibits a markedly higher proliferation rate, consistent with its overall mesenchymal phenotype (54) (Fig. 4, B and G). To ensure that none of the morphologies map to atypical transcriptomes that are not represented in patient tumors, we queried a dataset quantifying the similarity of cell line models to transcriptional profiles from resected tumors in The Cancer Genome Atlas (55). The morphological and organizational patterns identified were each represented among cell lines representative of patient-derived tumors (fig. S7).

Next, we analyzed DepMap dependency data, quantifying differences in Chronos score between polygonal and irregular morphologies. We found distinct dependencies associated with morphologic state, including some therapeutic targets of interest in PDAC (Fig. 4, H and I). CDK6, the cell cycle kinase targeted by palbociclib and other CDK4/6 inhibitors used in hormone receptor–positive (HR⁺) breast cancers, had higher dependency in polygonal PDAC cell lines. Supportively, HR⁺ breast cancers also most commonly exhibit a luminal epithelial morphology (56). Another polygonal specific dependency was ARHGEF7, which encodes a RAC1 guanine nucleotide exchange factor that is involved in cytoskeletal remodeling. This gene had enriched dependency in PDAC, biliary tract, and oral squamous cell carcinomas and was recently found to sensitize PDAC cell lines to CHK1 inhibition (57). DHODH, encoding dihydroorotate dehydrogenase in the pyrimidine biosynthesis pathway, is a dependency specific for the irregular morphology. DHODH is a druggable metabolic dependency in subsets of PDAC cell lines that are resistant to KRAS ablation (58).

The most significant differential dependency between polygonal and irregular morphologies in our cohort was the highly conserved RNA splicing factor TRA2B, which exhibits a mean difference in Chronos dependency score of 0.427 in irregular versus polygonal lines. TRA2B has critical functions in development, including for somitogenesis (59), cortical neurodevelopment (60), and spermatogenesis (61), and has been reported to contribute to oncogenesis in mesenchymal cancers including osteosarcoma (62) and in laryngeal squamous cell carcinoma models (63). TRA2B was expressed significantly higher in primary (P = 0.0001) and metastatic (P = 0.038) PDAC relative to physiologic pancreatic tissue and was associated with worse overall survival in a cohort of 178 PDAC patients (P = 0.032) (fig. S8, A and B).

Last, we applied a tree-based gradient boosted machine learning algorithm (XGBoost) to measure the predictive value of expression or dependency data for classifying polygonal or irregular morphology. Using a leave-one-out cross validation approach, we found that transcript expression measurements were a weaker predictor of morphology (AUC = 0.701) than dependency (AUC = 0.828) (fig. S8, C and D). When the two feature sets were combined into a single model, accuracy for predicting morphology (AUC = 0.845) was only marginally improved over the dependency model alone (fig. S8, D and E). TRA2B dependency was by far the most predictive feature in each of the models using dependency as quantified by improvement in accuracy (gain) when added to the model (fig. S8, D and E).

Overall, we have defined a reproducible schema for interpreting and categorizing the morphologic (i.e., polygonal, irregular, and spheroid) and organizational (i.e., dispersed, multilayered, and aggregated) features of PDAC patient-derived cell lines. In addition, we integrated these classifications with gene expression and dependency datasets from DepMap to identify distinct gene expression programs and dependency profiles associated with each category.

Morphology and cellular organization correlates with clonogenicity, invasion, and metastasis

Our next question was whether distinct cell morphology and organizational patterns were associated with functional differences in phenotype. We started by investigating invasion and colony formation, key properties important for tumor growth and metastatic dissemination.

Using our cohort of cell lines, we performed arrayed colony formation assays (fig. S9, A and B) and Boyden chamber invasion assays (fig. S9, C and D) on each of the cell lines. We measured invasiveness and clonogenicity associated with distinct morphology and organizational patterns. While our measurements were limited by sample size and high intragroup variance, we noted that, in general, polygonal and aggregated cells had both lower colony formation proficiency (colony count and occupied plate area) and transwell invasiveness compared to cells with other morphologies and organizational patterns. In addition, all high outlier measurements in the colony formation area measurement were irregular or spheroid cell lines, which typically fall into the dispersed and multilayered organizational categories. KP4, Suit2, Patu8988T, and MIA PaCa2 featuring irregular/spheroid morphology and dispersed/multilayered organization were the cell lines with the highest colony formation proficiency, whereas Panc1, KP4, and CFPAC-1, all representing irregular and dispersed cell lines, were the most invasive cell lines. These measurements of invasiveness and colony formation were generally correlated: the Pearson correlation coefficients of log-transformed colony formation assay measurements were 0.845 (P = 7.2 × 10⁻⁵) between colony count and colony area covered, 0.561 (P = 0.030) between invasion area and colony area, and 0.408 (P = 0.13) between invasion area and colony count.

A subset of the PDAC cell lines in DepMap (n = 30) has also been profiled in a metastatic tropism mouse xenograft model where barcoded cell lines were pooled and delivered via intracardiac injection (64). Using this model, we assessed the predictive power of cellular morphology and organizational pattern to stratify metastatic proclivity. We found that, consistent with our in vitro data (fig. S9, A to D), irregular/spheroid morphology and multilayered organization were associated with numerically higher rates of metastasis (fig. S9, E and F). Of the multilayered cell lines, Suit2, Psn1, and MIA PaCa2 have each been independently reported to have high metastatic capacity in various in vivo metastasis models (65, 66). In addition, human pancreatic cancer cells with an amoeboid phenotype were recently found to be more proficient at perineural invasion relative to an isogenic line with epithelial morphology (67). It is possible that spheroid and multilayered lines may be better suited for anchorage independent growth, a key property enabling seeding at distant metastatic sites (68). However, we caution that the intracardiac injection model does not recapitulate locoregional mechanisms of metastasis, such as PDAC metastasis to the liver (69). In addition, variability has been reported in the relative metastatic proclivity of the PDAC cell line models described here (70). Nevertheless, our collective data suggest that the mechanisms underlying morphological and organizational patterns may also regulate metastatic potential; however, additional validation with larger cohorts and improved preclinical metastasis models is required to fully dissect this phenotypic association.

Tissue organization of malignant cells correlate with transcriptional state

Last, to augment the clinical relevance of our approach to studying associations between morphology and cell state, we explored the relevance of cell morphology in the tissue context. PDAC has previously been described by two predominant malignant cell subtypes, namely, the classical and basal-like, which display clinically observed differences in chemosensitivity, invasiveness, and prognosis (71). It is known that three-dimensional (3D) organoids have a more classical transcriptional state, whereas 2D cell lines skew toward a more basal-like or mesenchymal cell state (72). We confirmed that this is true even when stratified by our morphological subtypes (Fig. 5, A and B). Given that organoids typically adopt a glandular morphology with a central lumen similar to classical glands in human PDAC specimens while 2D cell lines exhibit mesenchymal sheet-like arrangements and are unable to form glandular structures, we hypothesized that morphological heterogeneity in human tumor specimens may also correlate with transcriptional state.

Fig. 5. In situ tissue morphology in human PDAC correlates with transcriptional subtype.

(A and B) Classical (A) and basal-like (B) single-sample GSEA (ssGSEA) signature scores (81) in 2D cell line models stratified by morphology and in 3D human PDAC organoid lines. Statistical significance assessed by Mann-Whitney U test. (C) Representative comparison of small clusters versus large clusters in DE analysis visualized using immunofluorescence of PanCK (green) from a patient-derived tumor tissue section. Scale bar, 300 µm. (D) Histogram of cluster size distribution plotted on log-log axis. Solid blue line depicts the empirical probability density function, and the dotted red line depicts a power law fitted line. (E) Volcano plot depicting differentially expressed genes in small clusters (right) versus large clusters (left). (F and G) GSEA enrichment plot of classical (F) and basal-like (G) transcriptional subtype signatures (81) in small clusters. (H and I) Barplot of top GO terms enriched in small clusters (H) and large clusters (I).

Using a dataset of human PDAC specimens profiled with SMI at subcellular resolution using a 990-plex mRNA panel (73), we assigned cancer cells into clusters using a spatial clustering algorithm (DBSCAN; Materials and Methods). Cluster sizes followed a power law distribution (likelihood ratio test, P = 1.87 × 10⁻⁵), and accordingly, we classified individual cells according to their membership in small clusters of five cells or fewer versus cells in large clusters of 100 cells or more (Fig. 5, C and D). We performed differential expression analysis of small versus large clusters (Fig. 5E). The most differentially expressed genes in the small clusters were mesenchymal genes including VIM, LGALS1, KRT6, MMP2, and KRT17, while the most differentially expressed genes in large clusters were genes associated with the classical state including CEACAM6, LYZ, EPCAM, AGR2, and ERBB3. Collectively, basal-like and classical signatures were strongly enriched in small and large clusters, respectively (Fig. 5, F and G), and had intermediate expression in an intermediate cluster size defined as a group of 5 to 100 cells (fig. S10).

We performed GSEA on the top overrepresented genes in the small and large clusters (Materials and Methods). Small clusters had enrichment of gene sets related to ECM constituents (P_adj. = 9.10 × 10⁻²⁷), myeloid migration (P_adj. = 3.70 × 10⁻²²), focal adhesion (P_adj. = 2.45 × 10⁻¹⁷), PI3K-Akt signaling (P_adj. = 2.32 × 10⁻¹⁶), and platelet-derived growth factor signaling (P_adj. = 1.88 × 10⁻¹⁴). Large clusters featured enrichment of terms related to cytokine signaling (P_adj. = 7.90 × 10⁻²²), interferon receptor activity (P_adj. = 1.25 × 10⁻¹⁰), and receptor tyrosine kinase pathways (P_adj. = 5.59 × 10⁻¹⁰) (Fig. 5, H and I). Collectively, our results demonstrate that differences in malignant cell organization in situ strongly stratify transcriptional cell state, supporting the use of cancer cell organization as a clinically relevant biomarker.

DISCUSSION

In this study, we systematically evaluated the features associated with cell morphology and organization in patient-derived PDAC cell lines and tissue specimens. Our work provides not only a comprehensive assessment of morphological heterogeneity, but also a methodological framework, SMART analysis, for understanding and classifying the diversity of human in vitro cell lines extensively used by the scientific community. We demonstrated that patterns readily observable in cell culture correlate with functional traits including stemness, invasion, and metastasis, as well as molecular properties such as pathway dependency and gene set expression. Furthermore, we developed an assay to integrate morphologic and transcriptomic measurements in pools of cell lines. We extended our findings to patient tissues to find distinct transcriptional state associations with cancer cell morphology and organization in situ. Diverse treatment-emergent morphologies were shown to arise in response to both emerging and established clinical compounds, suggesting distinct mechanisms of acute treatment response. Using our SMART assays, we identified specific transcriptional and morphological phenotypes associated with treatment using KRAS inhibitors. Our findings support known mechanisms of treatment resistance, involving focal adhesion, integrin, and EMT signaling pathways, while also nominating hypotheses related to interferon signaling. In response to chemotherapy (e.g., 5-FU and gemcitabine), we observed consistent induction of a morphologic state characterized by increased eccentricity and cell area. Overall, we demonstrate how morphology can be leveraged as a phenotypic measurement to better understand cancer biology and biomarker for drug screening.

Our work highlights directions to expand the characterization of morphology as a resurgent surrogate biomarker of cancer cell state. In addition, the functional and morphological differences of the Patu8988S/T cell lines derived from the same patient tumor demonstrate that intratumoral heterogeneity can be modeled ex vivo. While groups like the JUMP Cell Painting Consortium have pioneered the study of many perturbations including small molecules and under- or overexpression of genes in a few model cell lines, we took the complementary approach of profiling a diverse set of cell line backgrounds with a carefully selected set of clinically relevant perturbations. Our approach shares some concepts with the recently reported STAMP (single-cell transcriptomics analysis and multimodal profiling) technology (74) but is unique in its integrated characterization of morphology and cellular organization and application to large pools of cancer cell lines. Using barcoded pools of cell line models and imaging-based barcode identification, one could envision characterizing cell morphology changes in a diverse set of genetic backgrounds at large scale beyond what can be accomplished in a single academic laboratory. Recently, technological improvements such as laboratory automation and microfluidic imaging have also greatly expanded the scale of morphological profiling studies. For example, Recursion Pharmaceuticals has built screening platforms to assay whole genome CRISPR knockout and activation libraries, and large compound libraries to find compounds that can phenocopy genomic perturbations (75). Single-cell flow cytometry–based morphology profiling (DeepCell) is able to perform high-throughput image-based cell sorting (76), which has proved useful as a screening endpoint to accelerate functional profiling.

Image-based profiling is a relatively low-cost measurement and thus can be a resourceful way of adding complementary phenotypic information to already acquired transcriptomic databases of cell line models such as the Cancer Dependency Map. Expanding imaging measurements to include cancer types beyond PDAC would improve our ability to correlate morphologies with relevant -omic and functional features. Certain features such as multilayered organization and spheroid or irregular morphology that we identified to be associated with greater invasive capacity should be validated across multiple tumor lineages and experimental settings. In addition, more widespread measurement of morphology could improve identification of molecular features and distinct dependencies of different morphologies. While GPX4 dependency has been described in the literature as a feature of mesenchymal cells (50), it was identified by single-sample GSEA (ssGSEA) using RNA expression instead of morphologic characteristics. In our data, we identified several other mesenchymal, or irregular morphology, specific dependencies including TRA2B, which was not previously described. Functional and molecular characterization of the role of the TRA2B splicing factor is warranted to assess its potential as a cell morphology–specific therapeutic target.

We recognize several important limitations of our work. Cell lines undergo transcriptional and morphologic drift in culture. While live-cell imaging with holotomography can resolve dynamic changes in morphology, the SMI assay uses a static measurement of spatially resolved cellular transcriptomic state. The SMI technology also segments cells in only two dimensions, limiting the accuracy of transcript assignment in 3D. Our pooled screening approach has several limitations including the lack of barcodes in cell lines to confirm identity, and the limited 1000-plex probe set used to distinguish cells. The use of SMI panels with increased molecular plex such as 6000-plex and whole transcriptome panels that have recently been developed (77) could further improve cell line annotation. While we have demonstrated important clinicopathologic ramifications of cellular and tissue morphology and organization in PDAC, validating our findings in larger cohorts and expanding the application of our method to other cancer types are warranted. In conclusion, our SMART approach highlights the exciting possibility of leveraging integrated morphologic, transcriptional, and functional measurements on multiple patient avatars in response to drug treatment to comprehensively model patient disease in real-time (78).

MATERIALS AND METHODS

SMART framework

We derived cell morphology data from holotomography, Cell Painting, and SMI-based measurements, each providing a distinct advantage. Holotomography provided ultrahigh-resolution dynamic live cell imaging, while Cell Painting provided higher-throughput characterization and quantification of structural features such as actin fibers. Last, SMI enables high-plex transcriptomic feature assignment to static cells in a single pooled assay.

Cell lines

For our in vitro experiments, we used 15 human patient-derived PDAC cell lines that are publicly available through American Type Culture Collection. All cell lines were verified for authenticity before use. The list of cell lines used in this study are AsPC1 (RRID:CVCL_0152), MIA PaCa2 (RRID:CVCL_0428), Psn1 (RRID:CVCL_1644), Patu8988S (RRID:CVCL_1846), Patu8988T (RRID:CVCL_1847), Suit2 (RRID:CVCL_3172), Pacadd119 (RRID:CVCL_1848), Cfpac1 (RRID:CVCL_1119), Hupt4 (RRID:CVCL_1300), HPAC (RRID:CVCL_3517), Panc0203 (RRID:CVCL_1633), KP4 (RRID:CVCL_1338), KP2 (RRID:CVCL_3004), Panc0403 (RRID:CVCL_1636), and Panc1 (RRID:CVCL_0480).

Holotomography

We performed holotomography on cultures of human PDAC cell lines under KRAS inhibitor (RMC-6236 or MRTX1133) treated and untreated conditions. Briefly, 300,000 cells were plated on ibidi 35-mm culture dishes with uncoated polymer coverslips (81156) or 25 mm–by–75 mm microscope cover glasses (Mercedes Scientific, MERR2575) with ibidi removable culture chambers (80841) and allowed to adhere overnight in DMEM with 10% fetal bovine serum and penicillin-streptomycin. The next day, the medium was exchanged with either DMSO or KRAS inhibitor (RMC-6236 100 nM) containing medium, and cells were placed on the Nanolive 3D Cell Explorer 96focus for a 48- or 72-hour acquisition using a 5-by-5 or 10-by-10 grid scan. Videos of data in Fig. 1B are available as supplementary movies. The EVE analytics software platform was used for automatic cell segmentation and calculation of biophysical and shape metrics including cell mass (derived from refractive index) and area.

Cell Painting assay

Cells were seeded at varying density (typically 7500 cells per well, but 3000 cells per well for fast-growing cell types including KP4, Panc1, and MIA PaCa2) to achieve 40 to 80% confluence after 72 hours in Corning 96-Well Black Polystyrene Microplates (category #3650). Drug or DMSO vehicle control (1 μM MRTX1133, 100 nM RMC-6236, 3 μM 5-FU, and 100 nM gemcitabine) was added the following day after seeding. Cells were first fixed with 4% paraformaldehyde for 20 min and then washed with phosphate-buffered saline (PBS) two times for 5 min each. Cells were stained at the manufacturer-recommended dilution with fluorophore conjugated phalloidin (Thermo Fisher Scientific, A12380), fluorophore conjugated concanavalin A (Thermo Fisher Scientific; C11252), and DAPI and imaged at 10× or 20× using the Nikon AXR. For certain experiments, we switched out the concanavalin A conjugate stain for an anti-paxillin primary unconjugated antibody (Thermo Fisher Scientific, MA5-33075; RRID:AB_2810168) and subsequent staining with anti-rabbit secondary antibody (RRID:AB_143165). CellProfiler quantification only used the phalloidin stain and the DAPI nuclear stain. Imaging batches were collected for all cell lines and treatment conditions in the same imaging session with identical microscope settings.

Morphology quantification with CellProfiler

Raw .tiff images corresponding to each channel (DAPI, phalloidin, and concanavalin A) were processed using CellProfiler version 4.2.6. We designed a pipeline to quantify image features by first extracting primary objects using the “IdentifyPrimaryObjects” module and the DAPI channel image as input with typical pixel diameter range of 20 to 70 units, global thresholding with minimum cross-entropy, a threshold smoothing scale of 1.3488, and correction factor of 1.0, and used intensity to distinguish clumped objects. These parameters were established empirically using the guess and check method. We then used the “IdentifySecondaryObjects” module to identify cell bodies propagated from the identified primary objects and the actin channel stain. Here, we also used the global minimum cross-entropy thresholding method with a regularization factor of 0.005. Then, we used the “MeasureObjectIntensity,” “MeasureObjectSizeShape,” and “MeasureObjectNeighbors” to quantify cellular characteristics and exported using “ExportToSpreadsheet.”

The following measurements were exported and used for PCA/UMAP analysis: Area, BoundingBoxArea, Compactness, ConvexArea, Eccentricity, EquivalentDiameter, Extent, FormFactor, MajorAxisLength, MaxFeretDiameter, MaximumRadius, MeanRadius, MedianRadius, MinFeretDiameter, MinorAxisLength, Orientation, Perimeter, Solidity, IntegratedIntensity_Actin, IntegratedIntensity_Paxillin, AngleBetweenNeighbors_Adjacent, FirstClosestDistance_Adjacent, FirstClosestObjectNumber_Adjacent, NumberOfNeighbors_Adjacent, PercentTouching_Adjacent, SecondClosestDistance_Adjacent, and SecondClosestObjectNumber_Adjacent. Descriptions of these variables are available in the CellProfiler manual: https://cellprofiler-manual.s3.amazonaws.com/CellProfiler-3.0.0/modules/measurement.html.

Spatial molecular imaging

We performed SMI (CosMx, Bruker/NanoString) on treated and untreated cultured cells. For monoculture experiments on AsPC1 cells, we treated cells with 1 μM MRTX1133, and for experiments performed on the pooled cohort of cell lines, we treated cells with 100 nM RMC-6236. This is due to MRTX1133 being an allele-specific KRAS^G12D inhibitor and RMC-6236 providing inhibition across the broader set of KRAS alleles present in our cohort. Briefly, we seeded cells in various configurations [either a single chamber per slide or multiple smaller ibidi removable culture chambers (80841) per slide] at a density of 1000 cells per 1 mm² on glass Superfrost Plus slides (Thermo Fisher Scientific, 12-550-15) that were precoated overnight with poly-d-lysine solution (0.1 mg/ml) in PBS (Gibco, A3890401). For pooled experiments, cells were seeded as equicellular mixtures. Cells were allowed to adhere overnight and were treated the following day with drug compounds (1 μM MRTX1133, 100 nM RMC-6236, 3 μM 5-FU, and 100 nM gemcitabine) for 72 hours.

Cells were fixed with 10% neutral buffered formalin (NBF) at room temperature for 30 min, washed two times in PBS for 5 min, and dehydrated with 70% EtOH at 4°C until slide preparation for CosMx SMI. Slides are stable for up to 2 weeks. Slide preparation followed a modified procedure from Bruker MAN-10184-03. Briefly, slides were rehydrated in 5-min incubations in 70% EtOH and 50% EtOH, washed with PBS, permeabilized with PBS-T for 10 min, and washed twice more with PBS. Samples were digested with Endoproteinase GluC (5 μg/ml; New England Biolabs, P8100S) for 15 min at 40°C in a hybridization chamber. After incubation, slides were transferred to a jar with PBS, and a 0.0003% fiducial concentration was applied in SSC-T buffer for 5 min. The slides were washed once in PBS and transferred for postfixation to 10% NBF for 1 min and then two 5-min incubations with NBF Stop Buffer (tris-glycine buffer) before transferring to PBS for 5 min. A freshly prepared 100 mM N-hydroxysuccinimide–acetate mixture (Thermo Fisher Scientific, 26777) was applied to the cells and incubated for 15 min at room temperature. Slides were twice transferred to 2× SSC buffer for 5-min washes and stored at 4°C before in situ hybridization. We used the CosMx Human Universal Cell Characterization RNA Panel (Bruker, CMX-H-USCP-1KP-R), targeting 950 human genes, and a 50-target add-on panel set for all our experiments. These probe sets were individually denatured at 95°C in a preheated thermal cycler for 2 min and crash cooled on ice for 1 min. We added ribonuclease inhibitor, Buffer R, and diethyl pyrocarbonate–treated water according to suggested ratios and incubated slides in a hybridization tray for 16 to 18 hours at 37°C overnight. We then performed two stringent washes with prewarmed 50% deionized formamide in 2× SSC for 25 min each, followed by a transfer to 2× SSC until staining. We stained with DAPI, CD298/B2M, PanCK, and CD45 according to the recommended concentrations in the slide preparation manual (Bruker, CosMx FFPE Slide Preparation RNA Kit), first for 15 min of DAPI and then 1 hour with the protein cocktail. We washed slides three times in PBS for 5 min each and stored at 4°C until being run on the instrument. We used configuration D for cell segmentation and configuration A for prebleaching.

Invasion assays

We performed transwell Boyden invasion assays to assess invasion and migration capabilities in each of our cell lines. Cells were serum-starved for 24 hours and then seeded at 50,000 cells per Matrigel-coated insert (Corning 354480) with 500 μl of serum-containing medium below the insert and 500 μl of serum-free medium containing cells seeded on top of the insert. Cells were incubated and allowed to invade for 48 hours and stained according to Diff Quik staining kit instructions (Electron Microscopy Sciences, catalog #26096-25). Invasiveness was quantified by taking bright-field images of five representative fields of view and quantifying stained areas within the field as well as using the “Analyze Particles” module in ImageJ to count cell bodies.

Colony formation assays

Cells were detached using TrypLE and seeded in 24-well plates at a density of 500 cells per well. After 5 days, cells were stained using crystal violet staining solution composed of 0.125 g in 50 ml of 20% methanol. Colonies were imaged against a bright white background using an iPhone 13 and quantified using the “Analyze Particles” module in ImageJ to measure average colony diameter, covered plate area, and total colony count within each well.

Cancer dependency map analysis

Genome-wide CRISPR dependency effect sizes were obtained from the DepMap 24Q2 Chronos dataset. These data are publicly available through the DepMap portal. These dependency effect sizes are quantified as Chronos scores, a statistical measure inferred from a population dynamics model that corrects for single guide RNA efficacy, screen quality, cell intrinsic growth rates, and bias related to DNA cutting toxicity. Dependency scores are continuous values that range in magnitude but have empirically demonstrated that essential genes commonly have a Chronos score of <−1 and unexpressed genes commonly have a Chronos score equal to 0. RNA expression data were obtained from the DepMap 24Q2 and were quantified as log₂(transcripts per million + 1).

Functional MetMap metastasis data were retrieved from the MetMap 500 dataset (64) located online at (https://depmap.org/metmap/data/index.html). This dataset contains metastatic potential and penetrance of 488 barcoded cell lines to five target organs (bone, brain, kidney, liver, and lung). Metastatic potential is a continuous variable on a log₁₀ scale, ranging from −4 ~ 4 that quantifies the barcode abundance of cell lines within a tissue. Values ≤−4 refer to nonmetastatic lines, −4 ~ −2 are weakly metastatic but with low confidence, and ≥−2 means metastatic with higher confidence (64). Penetrance refers to the percentage of animals that the cell lines were detected via barcode sequencing, ranging from 0 ~ 1.

Drug sensitivity data were obtained from the Broad PRISM screening platform (79), with data published on DepMap and elsewhere (27). Briefly, the PRISM platform uses a DNA molecular barcoding method to screen drugs against cell lines in pools. By sequencing and quantifying barcode abundance, the sensitivity of individual cell lines can be calculated at high throughput.

Cluster analysis in human tissue

We used a primary human PDAC dataset composed of 13 resection specimens profiled with 990-plex spatial molecular RNA imaging as described previously (73). We first classified annotated cancer cells into large or small clusters using DBSCAN, a spatial clustering algorithm. Within each field of view, we applied the DBSCAN algorithm with the eps parameter set to 120 and min_samples to 3. Clusters with more than 100 cells were classified as large clusters and clusters of less than five cells were classified as small clusters. After cluster assignment, we performed differential expression using a Mann-Whitney unpaired U test to identify differentially expressed genes between cells in small clusters versus large clusters.

We used the “score_genes” function implemented in the Scanpy package (80) to calculate single-cell gene signature scores for previously defined classical and basal-like signatures (81). Briefly, for each cell, the function computes the average expression of a predefined set of “signature” genes and subtracts the average expression of a reference gene set matched for expression distribution. This approach normalizes for technical differences across cells and accounts for the baseline activity of genes with similar abundance. The resulting signature score is a z-like metric that reflects the relative enrichment of the specified gene program in individual cells.

Differential expression analysis

A challenge for the 1000-plex SMI technology is a limited probe set for performing differential expression analyses. The limited coverage interferes with the ability to identify significantly enriched gene sets due to dropout and the assumption by rank-based GSEA and commonly used gene sets that the provided gene list represents the entire transcriptome. We therefore supplemented traditional rank-based GSEA using predefined gene sets with an overrepresentation analysis method implemented by g:Profiler (https://biit.cs.ut.ee/gprofiler/gost). Significantly enriched genes under a condition above a Benjamini-Hochberg adjusted P value threshold of 0.01 and a log 2 fold change difference in mean threshold of 0.25 were used to identify significantly enriched gene sets. For all g:Profiler analysis, we capped the maximum gene set size to 500 genes.

Supplementary Materials

The PDF file includes:

Figs. S1 to S10

Legends for movies S1 to S6

References

Other Supplementary Material for this manuscript includes the following:

Movies S1 to S6