Lineage plasticity in prostate cancer depends on JAK/STAT inflammatory signaling

Abstract

Drug resistance in cancer is often linked to changes in tumor cell state or lineage, but the molecular mechanisms driving this plasticity remain unclear. Using murine organoid and genetically engineered mouse models, we investigated the causes of lineage plasticity in prostate cancer and its relationship to antiandrogen resistance. We found that plasticity initiates in an epithelial population defined by mixed luminal-basal phenotype and that it depends on elevated JAK and FGFR activity. Organoid cultures from patients with castration-resistant disease harboring mixed-lineage cells reproduce the dependency observed in mice, by upregulating luminal gene expression upon JAK and FGFR inhibitor treatment. Single-cell analysis confirms the presence of mixed lineage cells with elevated JAK/STAT and FGFR signaling in a subset of patients with metastatic disease, with implications for stratifying patients for clinical trials.

One Sentence Summary:

Lineage plasticity in prostate cancer develops through expansion of epithelial cells of mixed luminal and basal lineage and is dependent on JAK/STAT and FGFR activation.

Resistance to molecularly targeted therapies is often due to mutations in the drug target, but there is growing recognition of other modes of tumor escape, particularly involving next-generation inhibitors designed to circumvent target-based resistance mechanisms. For example, in EGFR- and KRAS-mutant lung cancer and in metastatic prostate cancer, tumor cells can undergo a lineage transition from adenocarcinoma to squamous or neuroendocrine histology following treatment with the respective inhibitors (1-6). Similarly, BRAF-mutant melanomas can transition from a MITF+ differentiated melanocyte phenotype to more mesenchymal or neural crest stem-like cell states in response to BRAF inhibition (7-9). Transcriptomic analyses indicate that these changing cell states share features of normal developmental, regenerative and stem cell signatures, suggesting that tumor cells co-opt these gene expression programs to escape lineage-dependent therapies (10).

The increasing prevalence of this mode of tumor escape has sparked numerous investigations into the underlying mechanism. Analysis of patient specimens, particularly at the single-cell level, has been instructive in defining the complexity of the problem, revealing remarkable heterogeneity within and across patients (11, 12). Because these studies typically capture a single snapshot during tumor evolution, it is difficult to discern precisely when, where and how plasticity arises and its relationship with drug treatment and resistance.

Human prostate cancer provides an ideal model to study the role of plasticity in drug resistance due to its near universal association with prior exposure to androgen receptor (AR) directed therapies. Prostate cancers initially present as AR-positive invasive adenocarcinomas that express luminal lineage markers, including CK8, CK18 and kallikreins such as KLK3 (prostate specific antigen or PSA). Histopathologic criteria for diagnosis classically include a loss of TP63-positive basal epithelial cells, underscoring the luminal-like identity of prostate cancer. Prostate cancers are uniformly sensitive to androgen deprivation therapy (ADT) but eventually relapse as castration resistant prostate cancer (CRPC).

Clinical observations and sequencing data demonstrate that CRPC has heterogeneous phenotypes (13). Two subtypes, adenocarcinoma (CRPC-adeno) and neuroendocrine prostate cancer (NEPC), can be broadly identified by histology and staining for luminal (AR, KLK2/3) and neuroendocrine (SYP, INSM1 CHGA/B, CD56) lineage markers. However, these markers do not capture the full spectrum of CRPC phenotypes, which include mixed luminal and neuroendocrine features, loss of canonical luminal and neuroendocrine markers (‘double negative’), and gain of basal, mesenchymal and stem-like gene expression signatures (4, 14). Phenotypic heterogeneity in CRPC is thought to be a consequence of lineage plasticity within adenocarcinomas, rather than the outgrowth of pre-existing subclones with alternative lineages. This conclusion is supported by the fact that genomic alterations found in early luminal cancers, such as ERG translocations, are retained in NEPC (15). Furthermore, these diverse CRPC phenotypes cannot be clearly distinguished by tumor genotype. Although the loss of TP53 and RB1 is highly enriched in NEPC and double negative CRPC, alterations in these tumor suppressors, as well as PTEN, are also found in CRPC-adeno. A possible explanation is that loss of function in these genes may initiate a plastic cell state, which enables alternative lineage phenotypes to emerge during treatment with antiandrogen therapy.

Another critical question is whether anti-AR therapy, which is highly effective in inhibiting the growth of luminal epithelial tumor cells, may also accelerate the emergence of non-luminal CRPC phenotypes. The relative frequency of NEPC and double negative CRPC cases has increased over the past decade (4), coinciding with increased clinical use of next generation anti-AR therapeutics (such as abiraterone and enzalutamide) that inhibit AR signaling more potently than standard ADT. Although this correlation may be incidental, it begs the question of whether lineage plasticity, which is presumed to play a causal role in the development of antiandrogen resistance, is itself promoted or enhanced by anti-AR therapy.

Inflammatory signaling in adenocarcinoma cells precedes NEPC

To investigate these questions, we turned to two closely related genetically engineered mouse models (GEMMs) of prostate cancer that recapitulate the adenocarcinoma to NEPC transition following prostate–specific deletion of the human–relevant tumor suppressor genes PTEN, RB1 and TP53 using CRE recombinase. scRNA–seq was performed on whole prostates harvested from 29 mice [9 wildtype (WT); 7 Pten^−/−Rb^−/− (PtR); 13 Pten^−/−Rb^−/−Trp53^−/− (PtRP)] at relevant timepoints reflecting adenocarcinoma to NEPC transformation (Fig. 1A, Table S1) (16). Transcriptomes from 67,622 cells were analyzed and visualized by UMAP, using coarse labeling to identify cell types and Gfp to mark cells that underwent recombination (Fig. 1B, Fig. S1, Methods). In addition to the expected adenocarcinoma (Epcam–Gfp) and NEPC (Syp+) populations, we were surprised to see three additional Gfp+ clusters defined by expression of Pou2f3 and Dclk1 (Pou2f3–Gfp), Vim, Twist2 and Ncam1 (Vim–Gfp), and Tff3 (Tff3–Gfp), respectively. Pou2f3 is a transcription factor (TF) expressed by rare chemosensory cells in normal gut and lung tissue (often called tuft cells) and defines a non–neuroendocrine variant subtype of small cell lung cancer with morphologic features distinct from ASCL1 and NEUROD1 subtypes (17, 18). Tff3 encodes a secreted protein expressed by intestinal columnar epithelial cells involved in maintaining mucosal integrity and is a clinical biomarker for Barrett’s esophagus (19).

Figure 1. Inflammatory JAK/STAT Signals Emerge in Adenocarcinoma Prior to Transition to NEPC in GEMMs.

A) Experimental design includes WT (9 mice), PtR (7 mice, PbCre:Rosa26^mT/mGPten^fl/flRb1^fl/fl), and PtRP (13 mice, PbCre: Rosa26^mT/mGPten^fl/flRb1^fl/flTp53^fl/fl) at labeled time points. Timepoints for WT included 8 (1 mouse), 12 (2 mice), 24 (3 mice) and 32 (3 mice) weeks; for PtR were 24 (3 mice), 30 (2 mice), and 47 (2 mice) weeks; and for PtRP were 8 (1 mouse), 9 (1 mouse), 12 (4 mice), and 16 (1 mice) weeks relevant to the adenocarcinoma to neuroendocrine transition (4-6). Furthermore, six PtRP mice were castrated at 8 weeks of age for either a total of 4 weeks (3 mice, ‘Cas’), or 2 weeks followed by 2 weeks of dihydrotestosterone (DHT, ‘Cas/Reg’) addback (3 mice).

B) UMAP of GEMMs (N=67,622 cells) shown for all cell types based on imputed Gfp expression (restricted to non–immune cells) as a marker for mutant cells (Methods). Abbreviations include B (basal), L1 (luminal 1), L2 (luminal 2), SV (seminal vesicles), Adeno (adenocarcinoma), Vim-Gfp (Vimentin-Gfp), NEPC (neuroendocrine prostate cancer).

C) Force–directed layout (FDL) of mutant Gfp–positive and wild–type (WT) cells (N=28,934) colored by cell type. WT includes luminal 1 (L1), luminal 2 (L2), and basal (B). Mutant cell types include adenocarcinoma (Adeno), Tff3-Gfp, Pou2f3-Gfp, Vimentin-Gfp (Vim-Gfp), neuroendocrine prostate cancer (NEPC).

D) FDLs separated by genotype and timepoint, colored by cell type. Top left: WT with N=7,435 cells. Top: PtRP 24/30/47 weeks with N=2,565/662/1,441 cells. Bottom: PtRP 8/9/12/16 weeks with N=981/353/4,984/569 cells.

E) Hematoxylin/eosin and multiplex immunofluorescence (mIF) (40x) of mutant subtypes, including adenocarcinoma to NEPC transition, NEPC, and VIM. mIF channels include GFP (mG, green), DAPI (nuclei, blue), VIM (mesenchymal, red), SYP (synaptophysin, purple), and E–CAD (e–cadherin, white). Also shown are hematoxylin/eosin and immunohistochemistry (IHC) of TFF3– and POU2F3–positive cells in mutant tissue. Scale bars represent 100 μm or 50 μm as noted.

F) Heatmap of significantly enriched gene sets per cell type, restricted to wildtype epithelial (Basal/B, Luminal 1/L1, Luminal 2/L2), adenocarcinoma (Adeno), and neuroendocrine (NEPC) cells (FDR<0.01, abs(NES) > 1 where NES is the normalized enrichment score, see Methods).

G) Dissecting the adenocarcinoma to NEPC transition. Left: FDL of mutant adenocarcinoma (subsetted to adeno–B and adeno–L2) and NEPCs cells from the PtR model (N=16,593 cells) colored by genotype and time. Bottom: FDLs colored by normalized log2(X+1) expression of Cdkn2a or pseudotime scaled from 0 to 1. Right: Gene trends for TF DEGs across the adenocarcinoma to NEPC branch probability were calculated using a generalized additive model, with cubic splines across 500 equally sized bins (solid = mean, dashed = standard deviation) (25). Gene trends were grouped by Phenograph cluster (k=30) (Methods).

H) Transcription factors (TFs) groups include adenocarcinoma (purple), putative transition (red), and NEPC (blue). Heatmap of gene trends of select TF DEGs (Phenograph cluster versus rest within PtRP, refer to Methods: ‘Identifying DEGs’) from each category are ordered by the putative transition from adenocarcinoma to NEPC, with gene labels colored by aforementioned groups (scale gene trends of imputed expression, −0.5 to 1.5). Top panel shows a spline fit of the average Z-score of JAK/STAT and androgen signaling (NELSON_ANDROGEN_SIGNALING_UP) (Methods).

We examined potential routes of plasticity by evaluating cell connectivity in k-nearest-neighbor graphs (kNN), visualized using force–directed layouts (FDLs) for each genotype across different timepoints. We noted the early appearance of Gfp+ adenocarcinoma (Epcam–Gfp) as well as Tff3–Gfp cells, followed by the expansion of Vim–Gfp and NEPC and Pou2f3–Gfp populations, with the transition to NEPC more prevalent in the PtRP genotype, indicative of the additional impact of TP53 loss (Fig. 1C-D). The presence of each of these subpopulations was confirmed by multiplex immunofluorescence (IF) and immunohistochemistry (IHC) (Table S2), which documented SYP+ (NEPC) and VIM+ (VIM) positive cells in a background of ECAD+ (and PANCK+) epithelial cells (top row for ECAD in Fig. 1E, PANCK stain in Fig. S2); larger clusters of NEPC (top row) and VIM (bottom row) and foci of TFF3+ or POU2F3+ cells (bottom rows) (Fig. 1E). While this unexpectedly diverse range of cell types, often seen within a single mouse, is of interest, we focused our analysis here on the adenocarcinoma and NEPC populations because of their established clinical importance. Deeper analysis of the Tff3+ and Pou2f3+ populations will be reported separately.

Due to the large number of independent alleles, definitive lineage tracing experiments cannot be easily performed in this GEMM. Consequently, we relied on various computational approaches to predict the most likely scenario by which lineage plasticity emerges. As a starting point, we reasoned that the early emergence of the adenocarcinoma population, along with its placement between wildtype and other mutant clusters in the kNN-graph, implicate it as the most likely source of plasticity (Figs. S3A-D, Methods). Therefore, we examined differentially expressed genes (DEGs) and pathways associated with populations along this putative transition, using gene set enrichment analysis (GSEA) (Fig. 1F, Tables S3-5). First, we noted enrichment for a stem-like (L2) luminal signature that we and others have previously reported from single cell analysis of the normal mouse prostate (20-22) as well as enrichment for EMT programs and for signatures of early developmental programs (23, 24). Most surprising was strong enrichment for a range of inflammatory response gene signatures including Ifn-a, Ifn-g, Jak/Stat and Tnfa (Table S4). Given the indications implicating the adenocarcinoma population as a potential source of plasticity and its clinical relevance, we used Palantir trajectory analysis (25) to consider potential routes between adenocarcinoma and NEPC (Fig. 1G, Methods). We noted three major trends in TF expression: (i) those active in adenocarcinoma alone, (ii) those expressed during the putative transition to NEPC and (iii) NEPC–specific regulators (Fig. 1G-H, Fig. S3E-F for robustness analysis, Table S5). Strikingly, four inflammatory response TFs (Stat1, Stat2, Irf7, and Irf1), along with JAK/STAT signaling are specifically enriched in adenocarcinoma cells during the transition phase to NEPC (Fig. 1G-H).

Cell autonomous induction of plasticity in normal prostate organoids

Having documented the onset and evolution of lineage plasticity in vivo in the mouse prostate, we next asked whether plasticity is dependent on the tumor microenvironment or whether it can occur in a cell autonomous manner. For this purpose, we turned to organoid culture because prostate epithelial cells can be grown in isolation without accessory stromal or immune cell populations from the prostate microenvironment (26). Because this system supports the growth of normal as well as malignant epithelial cells, it also provides an opportunity to examine the immediate and longer-term consequences of tumor suppressor gene deletion with precise kinetics.

Focusing initially on Rb1 and Trp53 deletion (to model the alterations in human NEPCs in a minimalist system), we infected Trp53^loxP/loxP, Rb1^loxP/loxP organoids with Cre-expressing lentivirus. We also leveraged a mTmG reporter cassette to identify Trp53^−/−; Rb^−/− cells using the red-to-green color change initiated by Cre expression (Movie S1). In contrast to the cystic appearance of uninfected organoids, those with Trp53/Rb1 co-deletion lost their basal/luminal polarity and cystic lumens by ~4 weeks (hereafter called hyperplastic), followed by the appearance of migratory cells and cellular protrusions outside the normally sharp exterior organoid borders at ~8-10 weeks (hereafter called “slithering,” based on similarity to the migratory phenotype originally reported for lung neuroepithelial bodies) (27) (Fig 2A-B, Fig S4A-C, Fig S5, Movie S2). This phenotype was also seen following CRISPR/Cas9-mediated co-deletion of Trp53 and Rb1 in wild-type organoids (Fig S4B) and required deletion of both tumor suppressors. No morphologic changes were observed after deletion of Trp53 or Rb1 alone (Fig S4B), consistent with earlier work in human prostate cancer cell lines (28).

Figure 2. Establishment of an organoid model of spontaneous lineage plasticity.

A) Upper: Representative brightfield pictures of the various organoid phenotypes prior to and 4,6,8 and 10 weeks post Tp53 and Rb1 loss. Lower: Schematic representation of lineage plasticity development in Tp53^Δ/Δ Rb1^Δ/Δ organoids. Scale bar represents 50 μm.

B) Left: Representative H&E staining of LentiCre Tp53^Δ/Δ Rb1^Δ/Δ organoids ~10 weeks post deletion with cystic (A, blue), Hyperplastic (B, red) and slithering (C, silver) phenotypes. Right: Bargraph with percentage of organoids with given phenotypes during the time course. Scale bar represents 100 μm.

C) Westernblot verification of differentially expressed genes identified in RNA-seq data in wild-type (Tp53^loxP/loxP Rb1^loxP/loxP organoids, left) and ~10 weeks post deletion (LentiCre Tp53^Δ/Δ Rb1^Δ/Δ organoids, right). Proteins as marked. Actin was used as loading control. Fold change in protein level determined with ImageJ is given on the right.

D) Representative IHC of basal markers (p63, Ck5), luminal marker (Ck8) and Ar in wild-type (Tp53^loxP/loxP Rb1^loxP/loxP organoids, left) and ~10 weeks post deletion (LentiCre Tp53^Δ/Δ Rb1^Δ/Δ organoids, right). Scale bars represent 100 μm.

E) Volcano plot showing differentially expressed genes (DEGs) in the mutant organoid compared to wild-type. Blue lines indicate thresholds for significant DEGs (Bonferroni-adjusted p-value < 0.001, abs(log2 fold change) > 1). Red labels: significant DEGs of interest.

F) Lollipop plot showing significantly enriched pathways in RB1/TP53-deleted organoids vs wild-type using bulk RNA-seq based on GSEA (Benjamini-Hochberg adjusted p-value < 0.15; abs(NES) > 1)), where NES is normalized enrichment score (red = enriched in mutant; blue = enriched in wildtype). Lollipop size corresponds to significance, or −log2(p-value).

G) Westernblot of total Stat1 and Stat 3 levels and phosphorylated (p-) Stat1 and p-Stat3 in LentiCre Tp53^Δ/Δ Rb1^Δ/Δ organoids treated with increasing doses of the JAK1/JAK2 inhibitor Ruxolitinib for 48 hours. Actin was used as loading control.

H) Schematic representation of single cell cloning experiment of Basal (CD49f⁺) and Luminal (CD24⁺) cells. Organoids harboring a Cre recombinase inducible Cas9 (Rosa26 lsl-Cas9 (32)) were transduced with guide RNA’s targeting Tp53 and Rb1 (LentiGuide Puro, guide sequences, Table S23). Basal and luminal cells were sorted based on CD49f⁺ (Basal) and CD24⁺ (Luminal) expression and subsequently Cas9 was activated by transduction with adenovirus expressing Cre. Approximately 1000 single cells were seeded and grown out. Nutlin-3 and Palbociclib were added 3 days post Cas9 activation to select for Tp53 and Rb1 mutant cells. 24 basal and 24 luminal clones were randomly chosen and expanded for 4 weeks.

I) Quantification whole culture phenotypes (Cystic or Hyperplastic) of single-cell derived (CD49f⁺ or CD24⁺ FACS sorted) organoids after 4 weeks of culture. Left bargraph absolute number of cultures, right bargraph percentage of single-cell derived cultures in cystic or hyperplastic state.

To further characterize the changes induced by Trp53/Rb1 co-deletion, we performed bulk RNA-seq, western blot analysis, and IF or IHC analysis of various lineage markers. The most dramatic findings were reduced expression of luminal lineage genes such as Nkx3.1, Dpp4 (Cd26), Krt18 and Krt8 and increased expression of mesenchymal genes such as Snai1, Zeb2 and Vim (Fig 2C-E, S4C-E, Table S6). GSEA analysis of the bulk RNA-seq data confirmed enrichment of EMT (Fig 2F, Table S7) as seen in PtRP mice, further supported by elevated protein levels of Slit3 (also in the leading edge of the EMT gene set) and the migratory slithering phenotype observed morphologically.

Most surprising, however, was strong enrichment for the same inflammatory response pathways (e.g. Jak/Stat) that we saw earlier in PtRP GEMMs, as well as overexpression of Fgfr1 (5.0 log2 fold change; FDR<7x10⁻¹⁹⁹) and the ligands Fgf1/2/7/10 (all fold change >4; FDR<5x10⁻⁸) (Fig 2E-F). We confirmed this activation biochemically, with robust pStat1 and pStat3 activation in Trp53^−/−;Rb1^−/− organoids that was reversed by 48-hour treatment with the Jak kinase inhibitor ruxolitinib (Fig 2F-G, Fig S4F). Thus, the early manifestations of plasticity seen in GEMMs can be recapitulated in normal prostate epithelial cells in culture in the absence of any accessory cells, indicative of a cell-autonomous process.

However, unlike the later NEPC phenotype observed in PtRP mice and previous reports of Trp53^−/−;Rb1^−/− GEMMs (29), we failed to detect significant expression of Syp or other neuroendocrine markers (Chga, Chgb, Ascl1 or Neurod1), even in Trp53^−/−;Rb1^−/− organoids passaged for >12 months. Postulating this might be due to absence of critical prostate microenvironmental factors in organoid culture media, we asked if the organoid phenotype evolved further after orthotopic (OT) implantation of Trp53^−/−/Rb1^−/− organoids into Trp53^+/+; Rb1^+/+ hosts. After confirming highly efficient engraftment of hyperplastic/slithering stage organoids in a pilot experiment (by detection of GFP+ cells at 4 weeks in the dorsal lobe of hormonally intact NSG mice), we scored lineage phenotype and tumorigenicity, as well as the effect of androgen blockade, in a larger cohort by treating half the mice with castration and daily enzalutamide (Enz) 4 weeks after OT injection. In the hormonally intact mice, engrafted cells remained Ar+, Syp− (as in organoid culture); however, large foci of Ar-, Syp+ engrafted cells were seen in all mice in the castration + Enz group, indicating progression to a neuroendocrine-like lineage, a consistent finding across four independent Trp53/Rb1 co-deletion experiments (Fig S6A-D). To determine if this effect of in vivo androgen blockade on lineage plasticity is also observed in culture, we returned to the organoid model to determine the effect of Enz on the slithering phenotype following Trp53/Rb1 co-deletion. The percentage of organoids scored as slithering increased from ~10% to >30% (~3-fold) after 7 days of Enz treatment (Fig S6E).

Collectively, the above experiments demonstrate that changes in lineage marker expression and inflammatory response pathway activation seen in the adenocarcinoma population in PtRP mice can be recapitulated in mouse prostate organoids within weeks of Trp53 and Rb1 co-deletion. Furthermore, anti-AR therapy (castration and/or enzalutamide) can accelerate the rate at which these changes evolve, consistent with recent data using a patient-derived xenograft model (30). To address which cells can give rise to the plasticity phenotypes described above, we took advantage of prior work showing that organoids with normal basal and luminal bilayer architecture can be generated from single luminal or basal cells (26). Basal (CD49f+) and luminal (CD24+) cells from established Rosa26-lsl-Cas9 organoids stably expressing guide RNAs targeting Trp53 and Rb1 were isolated by FACS, transduced with Adenoviral Cre, seeded as single cells, expanded and then scored after 4 weeks using the morphologic criteria for plasticity (cystic versus hyperplastic, described in Fig 2B). Approximately 80% of plated basal and luminal cells developed hyperplastic changes, indicating that either cell type has lineage plasticity potential in organoid culture (Fig 2H, Fig S4H,I).

Lineage plasticity initiates with mixing of basal-luminal identity

Our organoids constitute a tractable experimental model that initiates and elaborates lineage plasticity in a reproducible manner. To gain more molecular insight into this plasticity, we characterized emerging transcriptional cell states and gene programs by collecting single-cell RNA-sequencing (scRNA-seq) data, beginning 2 weeks after Trp53/Rb1 co-deletion (prior to any morphologic evidence of plasticity), as well as 4 and 8 weeks after deletion. A parallel set of cultures were treated with Enz at 2 weeks and harvested at 4 and 8 weeks (Fig 3A; Table S1) to characterize the effects of androgen blockade on plasticity (see Fig S6).

Figure 3. Single-cell analysis of organoids reveals basal-luminal mixing, used as a proxy measure for plasticity to identify associated inflammation, JAK/STAT, and FGFR pathway activation.

A) Experimental design of organoid sequencing time-course. Samples collected at weeks 0, 2, 4, and 8 are enzymatically digested and subjected to scRNA-seq. Samples from weeks 4 and 8 are collected with or without enzalutamide (ENZ) treatment.

B) Force-directed layout (FDL) of wild-type basal and luminal SEACell metacells (N=142), labeled by wild-type subpopulations determined by Phenograph clusters. Edges between metacells indicate the k-nearest neighbors (k = 6) (Methods).

C) Mean log2(X+1) expression of differentially expressed genes in each basal and luminal subpopulation in the wild-type organoid.

D) FDL (k=10) of metacells in the prostate organoid before and after RB1/TP53 deletion (N=884), annotated by either wild-type subpopulations or by time and treatment following mutation.

E) Contour plots of basal and luminal cell densities at each timepoint, depicting a convergence of cell identities after mutation (Methods). Single cells are plotted using mean Z-scores of the set of basal genes that are gained in mutant luminal cells (y-axis), and the set of luminal genes that are gained in mutant basal cells (x-axis). Basal cell, red; luminal cells, blue; wild-type (WT), light colors; RB1/TP53-mutant (MUT), dark colors. Samples were collected at Week 0 (N=2,629), 2 (N=1,904), 4 (N=2,554), 8 (N=2,690), 4+ENZ (N=2,850), and 8+ENZ (N=3,050). Gene imputation is performed using MAGIC (k=30; t=3) prior to Z-score. See Fig S9A for a corresponding version without gene imputation.

F) Mean Euclidean distance between matched basal and luminal cells, based on the Linear Sum Assignment Problem (LSAP). Mean distances within each timepoint are an inverse measure of plasticity, which increases with time and treatment.

G) Top pathways significantly enriched for plasticity across timepoints (using GSEA, Bonferroni-adjusted p<0.01, NES > 1; see Fig S10C, Table S14). Each pathway score is measured as the average Z-score of gene expression in each pathway among metacells. Rows (pathways) are ordered by increasing pathway score from bottom to top in the early timepoint at week 2. Red asterisks indicate pathways also enriched in ENZ-treated compared to untreated samples (Methods).

H) Cellular plasticity increases as a function of JAK/STAT signaling. Plasticity is measured as the average entropy of cell-type classification probabilities per metacell. JAK/STAT score is the average Z-score of the leading-edge gene subset of KEGG_JAK_STAT_SIGNALING_PATHWAY and HALLMARK_IL6_JAK_STAT3_SIGNALING. The entropy of classification probabilities was first calculated for each single cell and then averaged per metacell for visualization. A linear fit is shown with Pearson’s correlation of 0.76.

I) Top scoring ligand-receptor (L-R) interactions known to activate JAK/STAT signaling, based on adjusted R² (R_adj²) of the regression JAK_STAT ~ L + R + L:R, where JAK_STAT is the JAK/STAT signaling score, and L and R represent scaled imputed ligand and receptor expression (Methods). Only L-R pairs with non-zero R_adj² are shown.

J) Mean expression (rows) by sample timepoint (columns) for candidate receptors activating downstream JAK/STAT signaling in metacells. Receptor genes are ordered by increasing expression in the early timepoint at week 2 from bottom to top.

We observed great transcriptional heterogeneity both within and between each time point and condition. To systematically catalog the observed cell states, we started with the wildtype phenotypic landscape then followed how it remodels following perturbation. To gain a concise description of the observed phenotypic diversity, we used our SEACells algorithm (31) (see Methods) to calculate metacells separately within each sample. The metacell approach (24) aggregates similar single cells into a set of distinct, highly granular cell states such that any differences between cells within a metacell are likely technical rather than biological. Aggregating counts within each metacell generates a robust (Fig S7), full transcriptomic quantification for each cell state, thus mitigating noise and sparsity in scRNA-seq data. An added benefit of metacells is the enumeration of observed cell states that are more amenable to comparison.

To establish a baseline for assessing plasticity changes in mutant organoids, we first focused on the cell states in wild-type organoids. Phenograph clustering (25) identified 5 distinct subpopulations, which we classified as luminal (L-Org1, L-Org2) or basal (B-Org1, B-Org2, B-Org3) based on published transcriptomes (see Methods) (Fig 3B-C, S8A-B, Tables S8-9). L-Org1 cells share expression of stem-like genes (e.g. Ly6d, Tacstd2/Trop2) with L-Org2 cells but also express other cytokeratins (e.g. Krt6a, Krt10) not typically associated with luminal cells. L-Org2 cells resemble stem-like (Ly6a/Sca1+, Psca+) luminal cells seen in vivo (called L2 in Karthaus et al (20); LumP in Crowley et al (21)) as well as canonical luminal cytokeratins (Krt8/Krt18) (20, 21). B-Org1 cells express canonical basal markers (Krt5, Krt14, Trp63), B-Org2 cells express secretory proteins such calcitonin and inhibin A, whereas B-Org3 are largely distinguished by expression of proliferation markers (Mki67, Aurka).

Analyzing the entire time-course following Trp53/Rb1 co-deletion, we found no overlap between cell states in wild-type and mutant organoids, even at two weeks, when organoids were still morphologically normal. When visualized in a force-directed layout (FDL) (see Methods), these metacells organize along the axis of time with a distinct gap between the wild-type and 2-week timepoint. Basal and luminal metacells identified by correlation to basal and luminal gene signatures are phenotypically well-separated in the wild-type setting but converge together on the FDL following mutation (Fig 3D, S8C-D; see Fig S8E for similar trends at the single-cell level). To understand these phenotypic changes, we characterized the basal and luminal features gained and lost following mutation (Tables S10-12, see Methods). Over time, mutant organoids gain a mixed phenotype, in which mutant basal metacells acquire typical luminal features (KRT8, KRT18, PROM1) and lose typical basal features (KRT5, KRT14). In parallel, mutant luminal metacells gain typical basal features (KRT17) but lose luminal markers (PSCA, LY6D) (Fig. S8F-G). At the same time, both basal and luminal mutant cells concurrently gain EMT-like features (VIM, FN1, COL4A1/2, ZEB1/2). Plasticity changes in the mutant organoid can thus be defined as a gain in mixed basal-luminal phenotype with EMT-like features. Furthermore, all of these changes are escalated by antiandrogen (Enz) treatment, consistent with the accelerated morphologic phenotypes observed in culture and in vivo.

To ensure that the observed mixed basal-luminal phenotype was not merely a loss of original basal/luminal identity, nor a product of the metacell approach, we calculated basal and luminal scores for each individual cell based on basal and luminal features gained following mutation (Table S12; see Methods). Corresponding density plots showed that mutant luminal cells become more basal-like over time, and vice versa, resulting in a convergence between luminal and basal phenotype over time and with treatment (Fig 3E, S9A).

The apparent infidelity in basal/luminal lineage suggested that the mixed phenotype could serve as a useful proxy for quantifying plasticity. We therefore developed alternative measures of basal-luminal mixing at each timepoint as a proxy for plasticity (see Methods). These per-sample metrics confirmed that plasticity increases over time following tumor suppressor deletion and with Enz treatment (Figs 3F, S9). As a per-cell measure that captures plasticity changes both within and across sample timepoints, we used cell-type classification-based approach (25), to associate each mutant cell with a vector of classification probabilities for each wild-type organoid phenotype, where classification uncertainty corresponds to increased mixing and plasticity (Methods). We visualized these per-cell probabilities as a ternary plot (Fig S10A-B). As expected, wild-type cells favor a single assignment to their respective phenotype, whereas mutant cells converge to maximal uncertainty at the center of the ternary plot over time. We quantified this time-dependent increase in basal-luminal mixing using entropy of cell-type probabilities, which robustly agrees with our per-sample measures of plasticity while capturing a range of plasticity within each timepoint/condition (Fig S10C-D, see Methods). Interestingly, Enz-treated samples display even fewer cells favoring the L-Org2 phenotype, consistent with our transcriptional and morphological data showing that AR inhibition accelerates the loss of luminal identity (Fig S6D, S10E). Notably, we observed similar basal-luminal mixing in the PtRP mice, further validating use of the organoid platform for interrogating plasticity phenotypes (Fig S10F-J). Thus, several orthogonal analytical approaches consistently detect increasing plasticity over time, in both organoids and GEMMs.

Jak/Stat and Fgfr upregulation coincides with plasticity onset and progression

To uncover the gene programs that drive plasticity following tumor suppressor gene deletion, we leveraged our per-cell measure of plasticity to identify pathways activated in highly plastic cells. After correlating genes with our plasticity measure, we used GSEA to identify gene programs significantly associated with plasticity (Fig S11A, Table S13-14; see Methods). Among these plasticity-associated pathways, we noted upregulated EMT and NEPC signatures (Fig S11B-C) even though morphological evidence of neuroendocrine transformation was not observed until these organoids were studied in vivo (Fig S6A-D). To enrich for causal genes, we narrowed our list of candidate driver programs to pathways that meet two additional criteria: 1) activated early in the time course, as plasticity emerges prior to morphological changes (week 2 following Trp53/Rb1 deletion) and 2) further enriched with Enz treatment, based on the premise that plasticity increases with therapeutic pressure (Tables S15-16, see Methods). The top resulting pathways include JAK-STAT (also LIF, IFNg, IFNa), FGFR, NOTCH, and EMT (also SMAD2) (Fig 3G, S11D-F).

Because our prostate organoids are cultured using defined, serum-free conditions that contain no ligands that might activate IFN response genes, we searched for potential mechanisms to explain the IFN response. Elevated IFN signaling is often a consequence of cytoplasmic sensing of double-stranded DNA or RNA, due to ruptured micronuclei or reactivation of endogenous retroviruses (32). However, CRISPR deletion of several genes involved in this sensing pathway (Tmem173 (Sting), Mavs, Tbk1, Ifih1) had no effect on the acquisition of the hyperplastic morphology following Trp53/Rb1 co-deletion. We therefore explored the possibility of an autocrine loop, reasoning that if a ligand-receptor (L-R) pair activates JAK-STAT, then high abundance of the L-R should correlate with increased JAK-STAT signaling in the mutant organoid cells. A key advantage of our organoid model over the in vivo PtRP GEMM model is that it only needs to consider autocrine L-R interactions; thus, we were able to exclude cell type as a variable in L-R analysis. We evaluated a set of 74 L-R pairs known to activate JAK-STAT signaling and scored each candidate based on their correlation with a summary statistic of 33 JAK-STAT associated genes enriched following mutation (Tables S17-18; see Methods).

The top ligand-receptor pairs identified by this analysis were LIF and its heterodimeric receptor (LIFR/IL6ST), FGF1/FGFR, IL15/IL15RA, the chemokines CCL2 and CCL5 and their cognate receptors CCR2 and CCR5, and HGF/MET (Fig 3I, S12A). To prioritize L-R pairs that cause an increase in plasticity, we applied similar criteria as for our candidate driver programs: (1) upregulation of receptor expression early in the time course (week 2), and (2) further upregulation with antiandrogen treatment. FGFR and LIF (co)receptors remained as top candidate drivers after filtering, based on increased expression of FGFR1, LIFR and IL6ST following ENZ treatment as well as early expression of FGFR1-3 and LIFR following Trp53/Rb1 co-deletion (Fig 3J, S12B-C). Thus, multiple lines of evidence from PtRP GEMM mice and from Trp53/Rb1-deleted organoids implicate Jak/Stat pathway activation as an early event that precedes the acquisition of various plasticity phenotypes such as basal-luminal lineage mixing, EMT and transition to NEPC, with FGFR and LIF prioritized as top candidates for functional interrogation based on L-R analysis.

Jak/Stat and FGFR inhibition restores luminal lineage identity

To test these predictions at a functional level, we added recombinant FGF1 or LIF to organoids following Trp53/Rb1 co-deletion, reasoning that plasticity might be further enhanced by an artificial increase in ligand concentration. We also included HGF, which was among the top candidates in our L-R analysis but deprioritized due to its delayed activation relative to the onset of the plasticity phenotype, and thereby acted as a validation control for our prioritization schema. Recombinant FGF1, but not LIF or HGF, enhanced the magnitude of slithering (from 6% to 50% of Matrigel surface) within 3 days (Fig S13A, arrows indicate slithering), providing functional support for Fgfr signaling as a driver of plasticity.

To address the functional role of JAK/STAT signaling, we derived Trp53^loxP/loxP, Rb1^loxP/loxP organoids in which either Jak1 or Jak2 alone or in combination were disrupted using CRISPR/Cas9-directed sgRNAs. We then deleted Trp53 and Rb1 by lentiviral Cre infection and scored the percentage of organoids with cystic, hyperplastic or slithering phenotypes at 8 weeks, as in Fig 2B. Disruption of Jak1 or Jak2 alone blunted upregulation of pStat1 and pStat3, but only modestly impacted development of the plasticity phenotype (~6% to ~15% increase in cystic organoids (p=0.01, Jak1 guides; p=0.09 Jak2 guides)), with most organoids progressing to the hyperplastic morphology (Fig S13B,C). In contrast, organoids with combined Jak1/Jak2 disruption (which more effectively blocked Stat1/3 phosphorylation) had substantial preservation of cystic morphology (>3-fold increase, S13B,C). We confirmed this effect pharmacologically by treatment with the dual Jak1/2 kinase inhibitor Ruxolitinib (Rux), which restored cystic morphology (from ~5% to ~30%) in a dose-dependent manner (Fig 4A, S13D,E)). As further evidence of reversion to a more luminal lineage, these morphological changes were accompanied by a modest reduction in Vim+ cells and an increase in Ar expression (Fig 4C,D), commensurate with the reduction of pStat1 and pStat3 levels back to baseline.

Figure 4. Pharmacological inhibition of JAK-STAT signaling and FGFR signaling resensitizes prostate cancer organoids to ARSI.

A) Representative brightfield pictures, H&E staining and IHC of CK8, Ck5 and Vimentin of Tp53^Δ/Δ Rb1^Δ/Δ organoids treated for 14 days with indicated drugs. Scale bars represent 100 μm.

B) Quantification of phenotypes cystic (Blue), Hyperplastic (Red) and slithering (Silver) of Tp53^Δ/Δ Rb1^Δ/Δ organoids treated for 14 days with indicated drugs, see methods for exact medium composition.

C) Quantification of vimentin positivity of organoids in A. Total number of organoids positive for marker were quantified and normalized.

D) Left: Western blot of lineage markers and JAK-STAT signaling components in Tp53^Δ/Δ Rb1^Δ/Δ organoids treated for 14 days with indicated drugs, see methods for exact medium composition. Proteins probed as indicated. Actin was used as a loading control. Right: Protein fold change or Ar, Ck8 and Vimentin as determined by ImageJ analysis in Tp53^Δ/Δ Rb1^Δ/Δ organoids treated for 14 days with indicated drugs.

E) Top: Schematic overview of resensitization drug experiments. Organoids are treated with Erdafitinib 100 nM and Ruxolitinib 10 μM for 14 days in low Egf organoid medium (ENR+A83+DHT), control organoids were cultured in low EGF organoid medium for 14 days. Subsequently organoids (10000 cells per well, triplicate) are reseeded in organoid culture medium without EGF (NR+A83), containing an AR agonist (DHT 1 nM) or antagonist (Enzalutamide 10 μM). Viability was measured by CellTiterGlo after 7 days of Enzalutamide treatment of Tp53^Δ/Δ Rb1^Δ/Δ organoids treated for 14 days with indicated drugs (Lower). Results were normalized to control culture condition. See Methods for exact medium composition.

Having demonstrated that perturbation of either FGFR or JAK/STAT signaling can impact the plasticity phenotypes in our organoid model, we used the clinical grade FGFR inhibitor erdafitinib (Erda) to examine the functional consequences of pharmacological FGFR inhibition on plasticity, alone and in combination with JAK inhibition. Although treatment with Erda also reduced the percentage of organoids with the slithering phenotype, we observed a striking ~12-fold increase (5% to 60%) in cystic morphology when fully plastic organoids were co-treated with Erda + Rux (Fig 4A,B).

Consistent with the morphologic change, we noted >3-fold reduction in cells expressing the mesenchymal lineage marker Vim by IHC (Fig 4C), confirmed by western blot, and increased levels (1.5-fold) of the luminal markers Ck8 and Ar (Fig 4D, Fig S13F). Importantly these changes in morphology and lineage marker expression occurred in the absence of any effects on organoid proliferation (Fig 4E, Fig S13G), suggesting that the primary consequence of combined Jak/Fgfr inhibition is on lineage reprogramming rather than selective outgrowth of rare organoids that had retained cystic morphology after Rb1/Trp53 deletion. The restoration of cystic luminal organoids following combined Fgfr + Jak kinase inhibition led us to ask if sensitivity to antiandrogen therapy is also restored. Indeed, Enz sensitivity was restored to previously Enz-resistant organoids (~50% decrease in proliferation, comparable to wild type organoids), but only after cystic morphology was restored by 14 days of Rux/Erda pre-treatment (Fig 4E, S13H).

JAK/STAT and FGFR activation in human CRPC

To determine if the transcriptional programs linked to plasticity in our murine prostate models are utilized in human prostate cancer, we analyzed scRNA-seq data from 12 CRPC patient samples representing both CRPC-adeno and transformed NEPC histologies (Fig. S14, Table S19), obtained by radiologically guided biopsies of metastatic index lesions visualized by metabolic imaging. As expected, these samples were enriched for genomic alterations in TP53, RB1 and PTEN, or in other alterations linked to these pathways such as CDKN2A/CDKN2B (RB1) and AKT1 (PTEN) (13, 33, 34) (Fig 5A). Transcriptomes from 51,094 cells were analyzed (see Methods) and the tumor cells (n=27,338) revealed remarkable heterogeneity across samples with a clear distinction based on CRPC-adeno and NEPC histology (Fig 5B, Fig. S15A-C). Cells were scored for JAK/STAT, FGFR, AR, EMT and NEPC signatures (see Methods), derived from published GEMM and organoid plasticity models (Table S18) and a mean value was determined for each Phenograph cluster (k=30). Scatterplots of these clusters revealed a strong positive correlation between JAK/STAT+FGFR signaling and EMT (r=0.70), but a negative correlation with AR signaling (r=−0.47) (Fig. 5D). One particular CRPC adenocarcinoma tumor (HMP13) uniquely displayed intratumoral heterogeneity, harboring a subpopulation marked by JAK/STAT and FGFR upregulation, EMT, and downregulation of AR signaling (Fig. S15D, Table S20). Mapping of TFs in this sample across pseudotime (25) revealed a similar activation of transcriptional regulators, namely STAT2/3, IRF1, and IRF7 (Fig. S15E-F for robustness analysis). Overall, clusters with JAK/STAT+FGFR and EMT high programs are found in 5/9 (or 56%) of CRPC adenocarcinoma samples (Fig. 5E, defined by having at least >20 cells in a cluster). These results are highly consistent with our findings from the GEMM and organoid models where JAK/STAT and FGFR are strongly correlated with EMT programs and lower AR signaling, but absent in NEPC cells.

Figure 5. Activation of JAK/STAT and FGFR signaling arises in CRPC patient biopsies and patient-derived organoids.

A) Treatment–refractory metastatic castrate resistant prostate cancer (CRPC) patients underwent IR–guided biopsies. High-risk index lesions were guided by advanced molecular imaging. Mutational oncoprint labeled by sample ID (HMP; human metastatic prostate), histology (adenocarcinoma, orange; and neuroendocrine prostate cancer, yellow), and mutational status (deletion, blue; amplification, red; missense, green; stop or frameshift, black).

B) UMAPs of tumor cells from patient-derived metastatic CRPC (N=27,338 cells) colored by tumor ID and grouped by tumor type.

C) UMAPs of tumor cells from patient-derived metastatic CRPC, annotated by Phenograph clusters (k=30), divided into three groups defined by Z–score of JAK/STAT and FGFR, or NEPC signatures: 1) AR high and JAK/STAT+FGFR low adenocarcinoma (JAK and FGFR signature, Z-score ≤ 1, green), 2) AR low and JAK/STAT+FGFR high adenocarcinoma (JAK and FGFR signature, Z-score>1, blue), and 3) NEPC (neuroendocrine signature, Z-score>1) (Methods).

D) Scatterplot of Phenograph clusters (points) based on Z-score of a combined JAK/STAT+FGFR signature (x-axis) versus EMT, AR, and NEPC signatures (y-axis), respectively. Linear fits were calculated for adenocarcinoma clusters only, with corresponding Pearson’s correlation denoted. Clusters annotated by groups definIin (C): NEPC (red), AR high and JAK/FGFR low adeno (green), and AR low and JAK/FGFR high adeno (blue)

E) Proposed classification of metastatic CRPC samples based on JAK/STAT+FGFR and NEPC signatures. HMP13 is assigned to the JAK/STAT+FGFR high group, as this sample harbors a well-demarcated JAK/FGFR and EMT-high subpopulation, described in Figure S14D.

F) Schema of baseline attributes and functional change in human organoids with null to low AR expression following treatment with combined Rux/Erda treatment, with hold-out sample MSKPCA2 used as reference level for AR-high subtype. 1) Baseline attributes including an oncoprint of TP53/RB1/PTEN genotype, transcriptional subtypes based on Tang, et al. (36), and low versus null AR protein expression. 2) Heatmap of baseline gene signatures in the human organoids using publicly available bulk RNA sequencing based on mean Z-scored expression of JAK/STAT signaling, FGFR signaling, and EMT gene sets (see Methods). 3) Functional changes in human organoids with null to low AR expression following treatment with Rux/Erda, including cell viability, and log2 fold change (FC) in AR and VIM protein expression.

G) Western blot of lineage markers and JAK-STAT signaling components in MSKPCA3 organoid after 14 days treatment with Erdafitinib 100 nM, Ruxolitinib 10 μM or with Erdafitinib 100 nM & Ruxolitinib 10 μM combination. Proteins probed as indicated. Actin was used as loading control.

H) Representative histology and IHC (CK8, CK5 and Vimentin) of MSKPCA3 organoids after 14 days treatment with Erdafitinib (100 nM) & Ruxolitinib (10 μM) combination in full organoid medium (ENRFFPN, A83-01, Nicotinamide, DHT). See methods for exact medium composition. Scale bars represent 100 μm.

I) Upper: Representative staining or AR in MSKPCA11 in control treated and treatment with Erdafitinib 100 nM & Ruxolitinib 10 μM. Lower: Bar graph of AR IHC staining patterns in patient derived organoids MSKPCA3, LuCAP176, MSKPCA8, MSKPCA11 and MSKPCA12. Organoids were control treated or treated with Erdafitinib 100 nM & Ruxolitinib 10 μM. AR staining was classified as negative (Grey), Low intensity (Blue) and high intensity (Red).

J) Proposed model system of lineage plasticity. JAK-STAT signaling activation and FGFR signaling activation leads to a AR^low, ARSI insensitive state that can be reprogrammed back to an ARSI sensitive state. Potentially the JAK-STAT^high, AR^low, ARSI insensitive state is a cellular state preceding AR negative ARSI insensitive CRPC. Mechanisms leading to AR-negative CRPC are unknown.

To address whether the JAK/STAT and FGFR activation we observed in human CRPC samples is functionally linked to lineage plasticity phenotypes such as EMT and NEPC, we turned to a panel of human prostate cancer organoids derived from patients with CRPC, all of which have been characterized by whole exome sequencing and bulk RNA-seq, and represent the CRPC-adeno and NEPC histology phenotypes described above. As expected, these patient-derived organoids (PDOs) were enriched for alterations in TP53, RB1 and PTEN. For the CRPC-adeno subgroup, we have previously shown that those with high AR expression (e.g. MSKPCa2) express luminal lineage markers and remain sensitive to Enz (35). We therefore focused on CRPC-adeno organoids with low or absent AR expression, as well as PDOs with NEPC histology, to determine if Rux + Erda treatment might restore expression of luminal genes as we saw in Trp53/Rb1-deficient mouse organoids. Transcriptomic analysis revealed elevated JAK/STAT activation and EMT signatures in 5 PDOs (MSKPCa3, 8, 11, 12 and LuCap176), all of which have low AR expression and therefore reflect the JAK/STAT high, mixed AR low subgroup seen in the patient biopsies (Fig 5F, S16A, Table S21). Unlike our genetically defined mouse organoid model, the human PDOs contain a complex array of genomic alterations beyond the tumor suppressor gene annotated in Fig 5F, and have growth properties that, in general, are much slower (4-7 day doubling times) than mouse organoids (1-2 days). In addition, individual PDOs often have heterogeneous expression of lineage markers, exemplified by variations in the level and percentage of cells with AR expression when examined on a cell-by-cell basis (Fig 5G-I).

Despite these differences between the mouse and human organoid models, we observed increased AR expression after 14 days of Rux + Erda treatment in 4 of the 5 AR low/JAK-STAT high PDOs. This change in AR expression was accompanied by increased expression of the luminal cytokeratin CK8 and decreased expression of the mesenchymal marker (VIM) (Fig 5F-I, Fig S16B, S17). As with the Trp53/Rb1-deleted mouse organoid, Rux+Erda treatment did not impact the proliferation rate of the AR low/JAK-STAT high PDOs (Fig S18A, B). One important distinction from the mouse system, however, is that a substantial percentage of cells within each of these Rux + Erda sensitive organoids remained AR-negative (Fig 5I). Whether this is a consequence of the greater genomic complexity of the human versus mouse models and/or the need for longer periods of drug treatment to reprogram human organoids fully (due to their slower growth rate) requires further study.

In addition to the luminal reprogramming phenotype seen in the AR low, JAK/STAT high organoids, we noted a distinct phenotype of substantial growth inhibition in two AR-negative PDOs with CRPC-adeno histology (MSKPCa1, 16) (Fig S18A, B). These PDOs are noteworthy for signatures of highly activated FGFR signaling but modest/absent JAK/STAT activation and likely reflect a dependency on FGFR for proliferation. Indeed, evidence of FGFR-dependent CRPC has been previously reported in the context of AR-negative, SYP-negative (double negative) prostate cancer (4). Growth of one of the 3 NEPC PDOs (LuCaP49) was also notably inhibited by Rux + Erda despite a low FGFR and JAK/STAT activation states across all three lines. The mechanism underlying this sensitivity requires further investigation.

Finally, we mapped the Rux+Erda sensitivity profiles to 4 subtypes of human CRPC (AR-positive, Stem-like, Wnt and NEPC) recently reported based on integrated chromatin accessibility and transcriptome analysis (36). The 4 PDOs that display luminal reprogramming all fall within the stem-like subtype (also defined by elevated JAK/STAT activation in Tang et al (36), whereas the 2 CRPC-adeno PDOs with growth inhibition fall with the Wnt subtype (Fig 5F), providing evidence of unique and exploitable therapeutic vulnerabilities for molecularly distinct subtypes of human CRPC.

Discussion

Across multiple cancers, marked changes in tumor cell state are implicated in the resistance to targeted therapies. This lineage plasticity is particularly relevant in the context of next-generation drugs designed to overcome conventional resistance arising from drug target mutations. However, lineage transformation remains poorly characterized, and we even lack a basic quantitative description of phenotypic plasticity. To address this mechanism of resistance therapeutically, it is critical to understand when, where and how cell-state changes arise—ideally, by profiling tumor cells as resistance arises. We performed single-cell RNA-seq time-course experiments on prostate cancer GEMMs and organoids, enabling an in-depth study of the adenocarcinoma to NEPC transition upon deletion of tumor suppressor gene combinations relevant to human CRPC (Trp53, Rb1, Pten). Our analysis identified an adenocarcinoma population with mixed luminal and basal identity that emerges within 2 weeks of tumor suppressor loss and progresses to EMT or NEPC phenotypes. We developed approaches to quantify plasticity, showing that it arises in a cell autonomous fashion in both normal luminal and basal cells, and increases as migratory behavior and EMT-like transcriptional signatures are acquired. However, progression to an NEPC phenotype only occurs in vivo, presumably reflecting a need for stromal or immune cell signals that are missing from organoid culture.

Prior work (4, 15, 16, 28) established that the onset of lineage plasticity is accompanied by resistance to Ar pathway inhibition, which we hypothesize to be a consequence of acquiring novel Ar-independent cell states, including basal or mesenchymal lineages. Importantly, we also observed the converse—plasticity is initially driven by loss of tumor suppressor genes, and subsequently accelerated by Ar pathway inhibition. The fact that lineage plasticity, initially responsible for drug resistance, is exacerbated by further drug treatment, has important medical implications.

We searched for molecular drivers of plasticity by identifying therapy-induced gene programs associated with early acquisition of our newly quantifiable high-plasticity state, and discovered an unexpected, robust signature of JAK/STAT activation in both organoids and GEMMs. Other manifestations of plasticity such as EMT and NEPC likely evolve later, based on the temporal sequence observed in organoids, as well as absence of JAK/STAT activation in the NEPC component of GEMM tumors and human NEPC samples. This observation suggests an early “critical period” following the onset of plasticity, during which therapies might prevent, delay, or even reverse plasticity. Supporting this hypothesis, JAK (and FGFR) inhibitors restored expression of luminal markers and decreased mesenchymal markers in mouse organoids and human PDOs, but only in those with elevated JAK/STAT activation. These changes evolved over several weeks of treatment in the absence of growth inhibition, consistent with a lineage reprogramming mechanism akin to differentiation therapy used in subclasses of acute myeloid leukemia (37, 38).

By searching for autocrine loops associated with JAK/STAT activation, we discovered that FGFR signaling increases with the onset of plasticity in organoids, GEMMs and human CRPC samples. Using a highly selective pan-FGFR inhibitor (Erda), we found (1) potent luminal lineage reprogramming in combination with JAK inhibition (Rux) in mouse organoids and PDOs, perhaps due to crosstalk between these pathways and reprogramming transcription factors or chromatin modifiers, and (2) consistently reduced mesenchymal gene (VIM) expression in organoids following FGFR inhibition, independent of JAK/STAT activation. This is consistent with FGFR’s role in tumors with EMT phenotype, as proposed in AR-neg, SYP-neg prostate cancer and in breast cancer (39). Importantly, the primary consequence of FGFR inhibition in PDOs with low JAK/STAT activation is growth inhibition, suggesting a dependency on FGFR for survival—distinct from the lineage reprogramming phenotype in PDOs with high JAK/STAT activation. Collectively, these findings suggest a therapeutic strategy by which CRPC patients with JAK/STAT-high, AR-low signaling (defined as stem-like by Tang et al (36)) would be candidates for combined JAK + FGFR inhibition in order to reprogram tumor cells to an AR-high luminal state and restore sensitivity to AR-targeted therapy. Those with elevated FGFR signaling and no JAK/STAT activation (defined as Wnt subtype by Tang et al (36)) would also be candidates for FGFR inhibition to exploit FGFR dependency for survival. Therapeutic timing is critical, as shown by cell-state heterogeneity within a single CRPC patient (or GEMM tumor) and absence of JAK/STAT activation in NEPC. These findings provide a rationale for molecular diagnostic assays that could aid patient selection.

Supplementary Material

Table S1

Table S1. QC statistics single cell RNA-seq in the organoid

Table S2

Table S2. GEMM, organoid, human detailed antibody information.

Table S8

Table S8. Differentially expressed genes in the wild-type organoid

Table S3

Table S3. Pathways used in GSEA analysis (GMT format)

Table S4

Table S4. DEGs/GSEA Enrichment Pertaining to Fig 1F

Table S6

Table S6. Differentially expressed genes in mutant versus wild-type organoid using bulk RNA-sequencing

Table S7

Table S7. Pathway enrichment in mutant versus wild-type organoid using bulk RNA-sequencing

Table S9

Table S9. Curated list of wild-type differentially expressed genes

Table S10

Table S10. Differentially expressed genes in mutant versus wild-type organoid

Table S11

Table S11. Pathway enrichment in mutant versus wild-type organoid

Table S12

Table S12. Basal and luminal markers gained and lost in mutant organoid

Table S5

Table S5. DEGs For All Epithelial Populations in GEMMs and adeno versus NEPC TFs

Table S13

Table S13. Gene correlation to plasticity in the organoid

Table S14

Table S14. Pathway enrichment for plasticity in the organoid

Table S15

Table S15. Differentially expressed genes in Enz-treated versus untreated organoid

Table S16

Table S16. Pathway enrichment in Enz-treated versus untreated organoid

Table S17

Table S17. Curated list of JAK-STAT activating Ligand-Receptor pairs

Table S19

Table S19. Clinical and molecular features of human CRPC tumor biopsies

Table S18

Table S18. JAK/STAT, JAK/STAT+FGFR, FGFR, AR, EMT, and NEPC Gene Sets Used for Organoid, Human, PDO sc-RNA Analyses

Table S21

Table S21 Baseline clinical and molecular features of human prostate organoids.

Table S22

Table S22 Organoid culture conditions

Table S23

Table S23 guide RNA’s used for CRIPSR/Cas9 mediated knockout

Table S20

Table S20. Differentially expressed genes in clusters in HMP13

Table S24

Table S24 List of abbreviations used in manuscript

Movie S1

Movie S1 Organoid 7 days post deletion

Movie S2

Movie S2 Organoid 10 weeks post deletion