Single-cell analysis of treatment-resistant prostate cancer: Implications of cell state changes for cell surface antigen–targeted therapies

Significance

Treatment of prostate cancer is rapidly evolving with several promising drugs targeting different cell surface antigens. Selection of patients most likely to benefit from these therapies requires an understanding of how expression of these cell surface antigens varies across patients and how they change during disease progression, particularly in tumors that undergo lineage plasticity. Using immunohistochemistry and single-cell mRNA sequencing, we reveal heterogeneity of cell states across a cohort of advanced disease prostate cancer patients; this heterogeneity is not captured by conventional histology-based designations of adenocarcinoma and neuroendocrine prostate cancer. We show that these cell states can be identified by gene regulatory networks that could provide additional diagnostic precision based on their correlation with clinically relevant cell surface antigen expression.

Abstract

Targeting cell surface molecules using radioligand and antibody-based therapies has yielded considerable success across cancers. However, it remains unclear how the expression of putative lineage markers, particularly cell surface molecules, varies in the process of lineage plasticity, wherein tumor cells alter their identity and acquire new oncogenic properties. A notable example of lineage plasticity is the transformation of prostate adenocarcinoma (PRAD) to neuroendocrine prostate cancer (NEPC)—a growing resistance mechanism that results in the loss of responsiveness to androgen blockade and portends dismal patient survival. To understand how lineage markers vary across the evolution of lineage plasticity in prostate cancer, we applied single-cell analyses to 21 human prostate tumor biopsies and two genetically engineered mouse models, together with tissue microarray analysis on 131 tumor samples. Not only did we observe a higher degree of phenotypic heterogeneity in castrate-resistant PRAD and NEPC than previously anticipated but also found that the expression of molecules targeted therapeutically, namely PSMA, STEAP1, STEAP2, TROP2, CEACAM5, and DLL3, varied within a subset of gene-regulatory networks (GRNs). We also noted that NEPC and small cell lung cancer subtypes shared a set of GRNs, indicative of conserved biologic pathways that may be exploited therapeutically across tumor types. While this extreme level of transcriptional heterogeneity, particularly in cell surface marker expression, may mitigate the durability of clinical responses to current and future antigen-directed therapies, its delineation may yield signatures for patient selection in clinical trials, potentially across distinct cancer types.

In recent years, there has been a remarkable increase in the clinical development of antibody drug conjugates (ADCs), radioligand therapies (RLTs), bispecific T cell engagers, and chimeric antigen receptor-expressing T cells (CAR–Ts), all of which are designed to target cell surface antigens expressed on cancer cells (1–5). ADCs selectively deliver potent chemotherapeutic toxins and RLTs deliver lethal doses of radiation, whereas bispecifics and CAR–Ts leverage the immune system for tumor killing. All four approaches require expression of the target antigen on cancer cells (to ensure tumor reduction/elimination), and the level of expression often must be greater than that in normal tissue to achieve an acceptable therapeutic index (6, 7). Consequently, clinical trials must be designed in a manner that ensures the selection of patients that meet these criteria, often through a companion diagnostic. The development of a radiotheranostic for castrate-resistant prostate cancer (CRPC) is a particularly noteworthy example, wherein a small molecule (PSMA-617) targeting prostate-specific membrane antigen (PSMA) is combined with a therapeutic radioisotope (¹⁷⁷Lutetium) to specifically target prostate cancer cells (1).

Acquired resistance to conventional molecularly targeted therapies is often due to mutations within the drug target in rare clones that emerge during treatment (8). Many next-generation inhibitors have been designed to overcome this form of “on target” resistance, with durable long-term remissions achieved in multiple tumor types including lung cancer (Osimertinib) and chronic myeloid leukemia (Asciminib) (9). However, a growing mode of resistance, commonly referred to as “lineage plasticity,” results from tumor cells adapting to environmental stresses, such as those associated with tumor invasion and metastases, as well as the selective pressure of drug therapy (9–12). The transition of adenocarcinoma to neuroendocrine cancer typifies this process and can be seen individually in up to 20% of prostate, lung, and gastric cancer patients who relapse on primary therapy. Unfortunately, cancer patients who harbor these plasticity-associated tumors have dismally short survival (11, 13).

To understand the repertoire of lineage states and, in that context, assess for cell surface marker expression across treatment-resistant prostate cancer, we utilized an integrated experimental and computational approach to analyze single-cell RNA sequencing (scRNA-seq) from 21 human treatment-resistant prostate tumor biopsies and two genetically engineered mouse models (GEMMs), together with human tissue microarray analysis (TMA) comprising 131 CRPCs with prostate adenocarcinoma (PRAD) and neuroendocrine carcinoma (NEPC) histologies. The latter allowed spatial analysis at the protein level of lineage marker expression in tumors, including an assessment of inter- and intrapatient heterogeneity.

Through these comprehensive datasets, we find that PRAD and NEPC tumors display a high degree of phenotypic heterogeneity with an array of androgen receptor (AR) positive and negative and NEPC gene regulatory networks (GRNs). Furthermore, through a comparative analysis of human small cell lung cancer (SCLC) and NEPC subtypes, we find a shared set of transcription factors (TFs) and cell surface antigens, indicative of conserved plasticity-associated gene programs. Finally, by evaluating the expression of cell surface proteins that have been or are being targeted therapeutically, namely PSMA, STEAP1/2, TROP2, CECAM5, and DLL3, we find a high degree of heterogeneity within and across CRPC and NEPC patients and across different GRNs.

The degree of heterogeneity in the expression of cell surface markers in metastatic CRPC revealed by our analysis raises challenges in maximizing the clinical utility of cell surface targeted therapeutics in plasticity-associated states, underscoring the need to intervene prior to their emergence. Furthermore, the TF-specific signatures identified here could prove useful to more comprehensively classify patients, possibly across tumor types, based on evidence of shared regulatory networks.

Results

Lineage Markers in Human Treatment-Resistant Prostate Cancer.

To evaluate the fidelity of reported cell type and surface markers in treatment-resistant prostate cancer, we first performed immunohistochemistry on prostate cancer tissue microarrays (TMAs) (14–16) constructed from rapid autopsy samples of patients with advanced CRPC. The included cases span the clinical disease spectrum from adenocarcinoma to NEPC (SI Appendix, Methods). The TMA consisted of 131 tumors from 16 patients, including primary prostate tissue and distant metastases, with 2 to 21 anatomically distinct tissue samples per patient. Samples were annotated by histology as PRAD, high-grade carcinoma (HGC) and NEPC (SI Appendix, Methods for definitions), as well as by tumor site. Human TMAs were stained for the following markers: luminal or basal epithelial (AR, NKX3.1, CK8, and P63), neuroendocrine (SYP, INSM1, ASCL1, NEUROD1 and FOXA2), cell surface (TROP2 and DLL3), proliferation (Ki-67), as well as other markers of interest from scRNA-seq analyses of prostate cancer GEMMs (YAP1, POU2F3, MYC, SOX2, TFF3, and EZH2) (17–19) (Fig. 1A and Dataset S1). Levels of protein expression within histologies (i.e., NEPC) were not significantly affected by ischemic time postmortem, except for FOXA2 (Dataset S2).

Fig. 1.

Tissue microarray of lineage and cell surface markers in human CRPC–adenocarcinoma and NEPC. (A) Heatmap of human CRPC tissue microarray–based immunohistochemical expression studies of patients from the rapid autopsy program at University of Washington. H-scores (immunohistochemical score, scale 0 to 200, and red gradient) are shown for select markers, namely luminal or basal (AR, NKX3.1, CK8, and P63), neuroendocrine prostate cancer (NEPC) (SYP, INSM1, ASCL1, NEUROD1, FOXA2), other single cell RNA-sequencing candidates from GEMMs (YAP1, POU2F3, CMYC, SOX2, EZH2, and TFF3), cell surface markers (CSM) (TROP2 and DLL3), and proliferative score (KI67, scale 0 to 100, and black gradient). Corresponding de-identified patient IDs (top row), site (bone, yellow; liver/lung, light purple; prostate, dark purple; lymph node, purple; other viscera, green), and histology (PRAD or prostate adenocarcinoma, light blue; HGC or high-grade carcinoma, orange; and NEPC or high-grade neuroendocrine, red) are labeled. Dark gray boxes are substituted in place of H-score for tumors with no immunohistochemical information. (B–D) Boxplot of H-scores of NKX3.1, YAP1, and ASCL1 grouped by histology (PRAD, HGC, and NEPC). Significance of H-score distribution was assessed by Wilcoxon signed-ranked test. (E) Scatter plot of H-scores of EZH2 (y axis) and proliferative index of Ki67 (x axis). Linear fit was calculated between two markers; the corresponding Pearson’s correlation is noted. (F and G) INSM1 or SYP (y axis) and ASCL1 (x axis) are shown with the color of the dot representing histology (PRAD, HGC, and NEPC) with corresponding Lin’s concordance correlation coefficient noted (95% CI). (H and I) Boxplot of H-scores of cell surface markers, TROP2 and DLL3 grouped by histology (PRAD, HGC, and NEPC). Of note, TROP2 and DLL3 expression has been assessed in a larger TMA (inclusive of these data) separated by categories: AR+/NE−, AR−/NE+, AR+/NE+, and AR−/NE− by our groups in Ajkunic et al. (20). Significance of H-score distribution was assessed by Wilcoxon signed–ranked test. Abbreviations include: not significant (ns), * (<0.05), **(<0.01), ***(<0.001), ****(1 × 10⁻⁴).

As expected, PRADs showed high immunohistochemical scores for CK8, along with the prostate luminal markers AR (SI Appendix, Fig. S1A) and NKX3.1 (Fig. 1B) (21). This pattern was also noted in HGC but not in histologically classified NEPC tumors (example shown in SI Appendix, Fig. S2 A and D). The basal lineage marker P63 was absent in all tumors (Fig. 1A, positive control shown in SI Appendix, Fig. S1B). YAP1, a downstream nuclear effector of the Hippo pathway, has been implicated in the stem-cell-like subsets of human tumoroids (22). YAP1 and ASCL1 H-scores showed an inverse relationship. Specifically, YAP1 was high in PRAD/HGC but not NEPC (Fig. 1C, P < 1 × 10⁻⁴), while ASCL1 and other neuroendocrine-associated TFs were high in NEPC histologies but not PRAD/HGC (Fig. 1D and SI Appendix, Fig. S1C). This profile mirrors that of YAP1 expression in lung adenocarcinoma vs. small-cell lung cancer (SCLC) (23, 24). Finally, EZH2, a subunit of the polycomb repressive complex 2, which has been implicated in NEPC transformation (25, 26), showed higher expression in HGC and NEPC compared with PRAD (highest in NEPC) (SI Appendix, Fig. S1A) and showed a positive correlation with the Ki-67 index (R = 0.73, Fig. 1E).

We next focused on additional neuroendocrine markers, beyond ASCL1, that have been previously described in SCLC and NEPC (17, 24, 27). Histologically defined NEPC tumors were enriched for expression of INSM1, SYP, FOXA2, and SOX2 (SI Appendix, Fig. S1A, example of NEPC IHC stains shown in SI Appendix, Fig. S2 A and B). The H-score for INSM1 showed strong concordance with ASCL1 (CCC = 0.86, 95% CI 0.81 to 0.9, Fig. 1F) whereas SYP, often considered a canonical NEPC marker (28), was less concordant with ASCL1 (CCC = 0.52, 95% CI 0.39 to 0.63) (Fig. 1G) and other NEPC TFs (CCC = 0.46, 95% CI 0.34 to 0.56) (SI Appendix, Fig. S1D). This is likely because several SYP-positive tumors were negative for ASCL1 and INSM1 but positive for luminal markers such as AR, NKX3.1, and CK8. These AR-positive, SYP-positive tumors (often referred to as amphicrine, example shown SI Appendix, Fig. S2C) have adenocarcinoma histology and clinically behave differently than bona fide NEPC tumors (29). These examples of SYP-positive, ASCL1-negative tumors suggest that SYP expression alone may not be sufficient to diagnose NEPC (29, 30).

Nuclear staining of NEUROD1, which marks a distinct small-cell subtype in SCLC (31), was not detected in this human TMA cohort (Fig. 1A, positive control stain shown in SI Appendix, Fig. S1B). However, a NEUROD1-expressing NEPC subset has been implicated previously in ATAC-seq analysis of prostate cancer PDX models (32) and was detected by scRNA-seq in at least one NEPC sample not represented on the TMA (discussed below). This may be due to the overall low incidence of NEUROD1 expression in NEPC. POU2F3 expression was also neither detectable in the TMA by IHC (positive control stain shown in SI Appendix, Fig. S1B) nor in the scRNA-seq cohort discussed below. Although clearly detectable in prostate GEMMs with NEPC histology, POU2F3 expression may be rare in human prostate cancers (17, 18). In contrast, TFF3, a mucosal-associated protein that is expressed in a subset of SCLCs and marks a nonneuroendocrine prostate population in prostate GEMMs (10, 33), was readily detected in human TMA specimens within subsets of PRAD, HGC, and NEPC samples (SI Appendix, Fig. S1A).

We next focused on the cell surface antigens TROP2 and DLL3, which are targets of various therapeutic agents currently in clinical trials. TROP2, the target of an FDA-approved antibody-drug conjugate in breast and bladder cancers (34, 35), was expressed in all PRADs and HGCs but not in the majority of NEPC samples (Fig. 1H) nor in cells with expression of NEPC TFs (SI Appendix, Fig. S1E). Conversely, expression of DLL3, the target of multiple agents currently under clinical investigation in SCLC (ADCs, T cell engagers, CAR–T cells) (36, 37) was restricted to NEPC tumors (Fig. 1I) and showed strong concordance with NEPC TFs (CCC = 0.9, 95% CI 0.87 to 0.93) (SI Appendix, Fig. S1F). This is consistent with a prior study from our groups documenting the expression of TROP2 and DLL3 in rapid autopsy samples (20).

Finally, the availability within the TMA to interrogate multiple independent metastatic sites from the same patient allowed us to detect intra-patient lineage heterogeneity. The most striking example from this analysis was the expression of luminal epithelial markers (AR, NKX3.1, and CK8) within individual bone or soft tissue metastases of three patients (patients 6, 7, and 10) with a clinical diagnosis of NEPC based on analysis of other tissue sites. These site-specific lineage differences are consistent with the notion that tissue microenvironmental signals may influence lineage conversion (10) (Fig. 1A and Dataset S1 for patient-specific TMA H-scores).

Taken together, profiling of late-stage CRPC with a broad panel of lineage markers documents that a) YAP1 loss generally occurs in ASCL1-positive NEPC tumors, b) TROP2 is predominantly expressed in PRAD/HGC, whereas DLL3 is almost exclusively present in NEPC tumors, c) in comparison to ASCL1, the expression of other transcription factors linked with neuroendocrine phenotypes, such as NEUROD1 and POU2F3, is less common, and d) SYP expression alone has limitations as a diagnostic marker for NEPC. Furthermore, our TMA demonstrates that a single-site biopsy is insufficient to adequately capture the intra-tumoral heterogeneity in late-stage prostate cancer patients.

Diverse Transcriptional Networks in Human CRPC.

To extend our analysis of lineage heterogeneity in human CRPC beyond in situ methods, we studied gene expression networks in a set of human tumor biopsies through single-cell RNA sequencing. We previously reported the transcriptomic architecture of 12 CRPC biopsies to identify JAK–STAT and FGFR as signaling pathways required for plasticity (10). We now report GRNs on an expanded cohort of 23 tumors (from 21 unique patients), including 9 naive or castration-sensitive prostate cancer (38, 39) and 14 late-stage metastatic CRPC tumors (119,083 profiled cells) (10, 40, 41). All tumors were reviewed by a genitourinary pathologist and classified histologically as CRPC–adenocarcinoma or NEPC. Furthermore, all CRPC–adenocarcinoma or NEPC tumors had been treated with more than two lines of therapy at the time of biopsy, with the majority having received taxanes (refer to Dataset S3 for details on histology, tissue site, tumor genomics and prior treatment, and Fig. 2A).

Fig. 2.

Diverse gene-regulatory networks in castration-resistant prostate cancer. (A) UMAP of tumor cells (N = 35,696 cells), colored by patient ID (large panel on Left), category (Top Right panel), treatments (Middle Right panel; categories include untreated, androgen-receptor signaling inhibitor/ARSI, and ARSI plus taxane–based chemotherapy) or TP53/RB1 genomic status (Bottom Right panel). Also detailed in Dataset S3. (B) UMAPs showing expression [log(X +1)] of lineage genes, namely AR, YAP1, and CHGA. (C) Boxplot of inter-patient heterogeneity measured by Shannon entropy based of patient frequencies. To control for cell sampling, 100 cells were subsampled from each Phenograph cluster (k = 30) within tumor compartments 100 times with replacement (Wilcoxon signed-rank test, SI Appendix, Methods). Immune and mesenchymal inclusion shown in SI Appendix, Fig. S3H. Abbreviations: * (<0.05), **(<0.01), ***(<0.001), ****(1 × 10⁻⁴). (D) Heatmap of CRPC-adenocarcinoma and NEPC cells (x axis) and per cell scaled regulon activity scores (z-score: −2 to 2) is shown for select TFs (paratheses denotes number of genes within regulon, extended heatmap in SI Appendix, Fig. S4). A dendrogram cutoff of 15 based on adjusted Rand index was used to unbiasedly define the number of gene-regulatory networks (GRNs), yielding 10 and 3 CRPC–adeno and NEPC GRNs, respectively. Regulons were assigned to GRNs based on regulon specificity score (RSS) and ranked by significance (Dataset S6). Adenocarcinoma GRNs were labeled based on AR activity (light blue on top panel of heatmap; bracketed by AR-positive GRNs) and without or having low AR activity (dark blue on top panel of heatmap; bracketed by AR-negative GRNs). NEPC regulons are shown (red on Top panel of heatmap; bracketed by NEPC GRNs). AR(12 g), NEUROD1(59 g), and ASCL1(34 g) regulons are bolded for reference.

Unsupervised clustering was used to iteratively label coarse cell types into lineage-defined groups using canonical markers (SI Appendix, Methods) (SI Appendix, Fig. S3 A–E and Dataset S4). A total of 35,696 primary naive or CSPC and metastatic CRPC tumor cells were labeled using select tumor markers (Fig. 2B and SI Appendix, Fig. S3 B and C) and copy number detection (SI Appendix, Fig. S3D). Given the observation of lineage marker heterogeneity in the CRPC TMA, we assessed the degree of interpatient heterogeneity by calculating the Shannon diversity of different patient phenotypes (SI Appendix, Methods). Clusters of cells associated with CRPC PRAD and NEPC were significantly more heterogenous (patient-specific, lower entropy) than those from CSPC tumors (Fig. 2C and SI Appendix, Fig. S3F), consistent with the notion that plasticity arises after androgen deprivation therapy (11). Furthermore, compared with tumor cells, higher entropy (lower phenotypic diversity or multipatient phenotypes) was noted in stromal, myeloid, and lymphoid cell populations (SI Appendix, Fig. S3 G and H), as has been described in single-cell analyses of other cancer types (33, 42).

To study tumor cell heterogeneity specifically in CRPC PRAD and NEPC samples and reasoning that lineage plasticity is likely driven by transcription factor (TF) networks, we focused on shared and unique gene-regulatory networks (GRNs) across samples using single-cell regulatory network inference (SCENIC) (Fig. 2D, SI Appendix, Fig. S4 A–C and Dataset S5). SCENIC has been utilized effectively to identify GRNs and cell types from single-cell RNA-sequencing data with improved accuracy when integrated with chromatin accessibility data (43, 44). We thus used hierarchical clustering of regulon activity within tumor cells, which unbiasedly identified 10 distinct putative GRNs in CRPC PRAD and 3 GRNs in NEPC tumor cells (SI Appendix, Methods, refer to cell- and patient-based robustness analyses for recurrent GRNs in SI Appendix, Figs. S5A/5B and S6, respectively). We further ranked regulons for differential activity within each GRN (Dataset S6) (SI Appendix, Methods). The 10 CRPC adenocarcinoma GRNs broadly separated based on the activity of the AR regulon, the dominant oncogenic pathway in CRPC. There were five GRN groups with AR regulon activity. Two of the identified AR+ GRNs displayed high HOXB13 activity, showing either higher levels of FOXA1 (labeled as AR+HOXB13+FOXA1+) or CREB3 (labeled as AR+HOXB13+CREB3+). One AR+ regulon showed lower HOXB13 and higher GATA2 and HOXA13 activity (labeled as AR+HOXB13–), largely derived from one sample (MSK–HP13) that had lost FOLH1 (which encodes for PSMA) expression. The lower expression of HOXB13 is consistent with recent reports implicating HOXB13 as a potential regulator of PSMA expression (45). We identified two further AR+ GRNs: “inflammatory” displaying high activities for IRF7/9 and STAT1/2 (10) (labeled as AR+ IRF7+STAT1+ Inflam) and “GI–lineage” with high activity of HNF4G and RELA (AR+HNF4G+ GI). The GI-lineage regulon showed enrichment of a mid-to-hindgut differentiation pathway, consistent with prior studies describing a role for HNF4G in promoting castration resistance (Dataset S7) (46).

The other five CRPC PRAD GRNs identified by hierarchical clustering lacked or had low activity of the AR regulon. One of these included activity for SOX2 and SOX4, along with FOXA2, TCF7L1, and TWIST2 (Fig. 2D and Dataset S6) (labeled as SOX2/4+ Embryo/EMT). These genes are highly expressed in the developing embryo and are enriched in the epithelial-to-mesenchymal transition and WNT signaling (TCF4) (Dataset S7). Another GRN had high activity for IRF2, along with NFATC1/2 and EGR2, consistent with our recent report of tumor-intrinsic inflammatory JAK/STAT signaling and inflammatory programs driving lineage plasticity (labeled as IRF2+ Inflam) (10). BATF and FOSL1/2 marked another non-AR GRN, concordant with a stem-cell-like group identified from patient-derived tumoroids and xenografts (FOSL1/2 + AP–1) (22) (Fig. 2D and Dataset S6). These latter cells were enriched for stem cell programs and the AP–1 pathway based on GSEA (Dataset S7). Two GRNs, albeit comprising a smaller number of tumor cells, showed high activity for the TCF7L2 regulon (along with KLF8, FOXK1, FOXP2, and BACH1) (labeled TCF7L2+) and CTCF (along with MAFG and NR1H) (labeled CTCF+), respectively (Fig. 2D and Dataset S6). TCF7L2 was previously identified as the top transcription factor (TF) candidate that marks a WNT-dominant CRPC phenotype (22).

Finally, we noted three putative NEPC GRNs largely distinguished by ASCL1 and NEUROD1 (32). NEPC–ASCL1+ cells (NEPC–A) showed high expression of E2F and neuronal targets (Dataset S7), along with ONECUT2 and NKX2–1 (Fig. 2D). Within a population of cells with lower ASCL1 activity, there was a subgroup that showed higher activity of HOXD11 and SOX6 (NEPC–H/S) (Dataset S6) that was also enriched for NOTCH and β-catenin signaling (Dataset S7). Within the NEUROD1 GRN (NEPC–N), there was activity for NEUROD2, ONECUT1, and SOX11 (Fig. 2D). This group is akin to SCLC–N and showed strong overrepresentation of BMP signaling (SOX11 and ZNF423) (Dataset S7).

Convergence between CRPC, SCLC, and GEMM Regulons.

Given that the aforementioned analysis of human CRPC tumors yielded snapshots into an array of CRPC PRAD and NEPC transcriptional states, we reanalyzed previously published single-cell sequencing data from GEMMs across multiple time points during the adenocarcinoma-to-neuroendocrine transition (10). This allowed us not only to identify overlapping TFs between mouse and human tumors but, importantly, to also detect potential intermediate cell populations that may not be captured in snapshots of human tumors. In our GEMMs of prostate-specific deletion of Pten, Rb1, and/or Tp53, we focused on two genotypes at varying time points of tumorigenesis. PtR mice (Pb-Cre;Pten^flox/flox;Rb1^flox/flox) were studied at 24, 30, and 47 wk and PtRP mice (Pb-Cre;Pten^flox/flox;Rb1^flox/flox;Tp53^flox/flox) at 8, 9, 12, and 16 wk (10, 40).

By implementing SCENIC, we found nine tumor-associated GRNs within the PtR and PtRP GEMMs including one defined by Stat1/2 and Irf2/7/9, validating our recent findings on the critical role of JAK/STAT signaling in initiating plasticity (Fig. 3 A and B and SI Appendix, Fig. S7 and Datasets S8 and S9). Furthermore, certain human CRPC and putative GEMM GRNs showed an overlap of specific cell populations, including NEPC–A and the above-mentioned inflammatory GRN with high activity of IRF7/9 and STAT1/2 (SI Appendix, Fig. S8 A/B).

Fig. 3.

GEMM GRNs and NEPC and SCLC overlap. (A) Heatmap of GEMM tumor cells (N = 21,499) (x axis) and per cell scaled regulon activity scores (z-score: −2 to 2) is shown for select TFs (paratheses denotes number of genes within regulon). A dendrogram cutoff of 12 based on adjusted Rand index yielded 9 GRNs with regulons assigned to GRNs based on regulon specificity score (RSS) and ranked by significance (SI Appendix, Methods, Dataset S9). Ar– extended (14 g) and Ascl1 (150 g) are shown in the top, bolded, and boxed in red for reference. (B) UMAP of GEMMs mutant Gfp-positive cells are colored by annotated GRN (color scheme corresponds to in A), or by regulon activity (z-score) of Ar_extended (14 g), Ascl1 (150 g), Twist1 (164 g), Pou2f3 (471 g), Tff3 (62 g), Trp63 (131 g), and Stat2 (94 g). (C) NEPC-N vs. NEPC-A (shown on x axis) or SCLC-N vs. SCLC-A (shown on y axis) were compared using MAST and the log₂FC for each gene is shown on the scatter plot. Genes with log₂FC > 0.4 (and P_adj < 0.05) are labeled with TFs noted in red or purple for being enriched in both NEPC and SCLC ASCL1 and NEUROD1 subsets, respectively. (D) Venn diagram shows the overlap of top DEGs (average log₂FC > 0.4, adjusted P-value < 0.05) shared between NEPC-A and SCLC-A (red) or NEPC-N and SCLC-N (purple). A Fisher’s exact test was used for significance of overlap.

This comparative analysis also allowed the detection of GRNs unique to GEMMs that may represent intermediate states. Specifically, 2 GRNs showed activity of the basal cell lineage factor Trp63, together with the coexpression of Hes1, Bach1, and Fosl1. One of the Trp63-marked clusters also displayed higher levels of Ar activity and high regulon activity for Sox4, Sox6, and Cux1; the latter TFs have been implicated in dendritogenesis and neuronal differentiation (47) (Fig. 3 A and B and Dataset S10). Given the lack of P63 in the human TMA and human scRNA-sequencing dataset, P63-positive tumors likely represent a rare entity, or an intermediary state not readily captured in human tumors. This is in comparison to our detection of P63-negative basal-like populations in human tumors, consistent with prior reports and as shown in our single-cell dataset (SI Appendix, Fig. S3E). In addition, we previously reported a unique nonneuroendocrine population marked by the tuft cell marker Pou2f3 (Fig. 3 A and B); however, we have been unable to find convincing evidence of POU2F3 expression in human CRPC at the RNA or protein level, as discussed earlier (17). Finally, one smaller subset with Ascl1 expression also displayed Ar and Stat5a/5b activity (Ar/Ascl1), suggesting that Ar and its regulon may be expressed within a subset of Ascl1-positive NEPC GEMM cells (Fig. 3 A and B). Note that this GRN is distinct from the AR-positive, SYP-positive (amphicrine) but ASCL1-negative human CRPC tumors discussed earlier (Fig. 1A).

Given the consistent presence of ASCL1 GRNs in both human CRPC and GEMM NEPCs, we utilized SCLC data to determine whether there are common transcriptional networks across prostate and lung histologies (33). We found significant enrichment of NEPC–A and NEPC–N regulons in corresponding SCLC–A and SCLC–N subsets, respectively (Fig. 3 C and D and SI Appendix, Fig. S9A and Dataset S11). Examples of such shared TFs between NEPC and SCLC subtypes included ASCL1, HOXB5, ETS2, ELF3, XBP1, and PROX1 (ASCL1 subtype), and NEUROD1, HES6, TCF4, NFIA, and JARID2 (NEUROD1 subtype). Furthermore, while we did not detect the GEMM Pou2f3 subset in our human dataset, we compared this GRN with SCLC–P and noted that both the GEMM Pou2f3 and inflammatory GRNs showed enrichment in SCLC–P. There were TFs, namely POU2F3, SMARCC1, and MYB, which were enriched in both Pou2f3 and SCLC–P populations (SI Appendix, Fig. S9 B–D).

Targeting Lineage Plasticity States.

The recent approval of PSMA-targeted therapies has directed attention into the degree of inter- and intrapatient heterogeneity, particularly as it may impact therapeutic response (15). Given our identification of putative GRNs in both murine and human treatment-resistant tumors, we explored the expression of several prostate cancer targets in our dataset, notably FOLH1 (PSMA), STEAP1/STEAP2, TACSTD2 ((TROP2)), CEACAM5, and DLL3, all of which have drug candidates in various stages of clinical development (1–5).

We first focused on PSMA given the expanding clinical usage of Lu¹⁷⁷–PSMA–617 (Pluvicto) for advanced CRPC (1). Upon scoring each regulon for its average expression of FOLH1/PSMA, we noted a positive association with most AR-positive GRN within the CRPC PRAD samples (CCC = 0.71) (Fig. 4A). The highest ranking PSMA^high/AR^high GRNs were associated with the luminal HOXB13+ or inflammatory IRF7/9+ GRNs (Fig. 4 A and B). In contrast, the AR+ GI GRN with HNF4G activity had lower expression of both PSMA and AR. There was also a PSMA^low/AR^high GRN, from patient MSK–HP13, which had lower HOXB13, KLF15, NFIL3 activity, but showed the highest activity for AR, and was enriched for GATA2, HOXA6, HOXA13, and RELB (Fig. 4B and SI Appendix, Fig. S10 A and B). There were several PSMA^low/AR^low CRPC GRNs that were enriched for genes and pathways related to embryonic, epithelial-to-mesenchymal, and/or WNT pathways (Dataset S12). Furthermore, NEPC samples within our cohort did not express PSMA and, as expected, displayed minimal AR signaling activity (Fig. 4A); however, it is possible that aberrant PSMA expression may be present in NEPC given reports of its coexpression with HOXB13 (45). Finally, we analyzed each tumor sample for intratumoral PSMA heterogeneity. Patient MSK–HP13 showed a cluster of AR-positive, PSMA-positive cells, while the remaining clusters were AR-positive, PSMA-negative (Fig. 4C). Certain TFs followed this pattern of negative PSMA expression with AR-positivity, including HOXB13, SOX4, and GATA2 (SI Appendix, Fig. S10C). Whether the high level of PSMA expression in a subset of cells from this lesion is sufficient to score as PSMA-positive on a PET scan is unknown, but cases such as this underscore the potential of heterogeneous PSMA expression even within single tumor foci (15, 48).

Fig. 4.

Expression of cell surface markers in CRPC and NEPC GRNs. (A) Scatter plot with scaled FOLH1/PSMA expression (y axis) and AR module score (x axis) (SI Appendix, Methods) for each GRN as colored in Fig. 2D (Lin’s concordance correlation coefficient = 0.71). (B) Heatmap of top 10 differentially active regulons in AR^highFOLH1/PSMA^high, AR^highFOLH1/PSMA^low (from MSK–HP13), AR^lowFOLH1/PSMA^low, and NEPC/FOLH1/PSMA^low. Per cell regulon activity scores are shown (scale: −2 to 2) (SI Appendix, Methods). (C) UMAP of AR and FOLH1/PSMA expression in tumor cells of MSK–HP13. Dotted circles denote region of FOLH1–positivity in otherwise largely FOLH1–negative MSK–HP13 biopsy. Heatmap of scaled expression (scale 0 to 1) is shown below with a blue box marking FOLH1/PSMA–positive cell population. (D) Scatter plots are shown of scaled expression of respective cell surface antigen (STEAP1, STEAP2, CEACAM5, and TACSTD2/TROP2, y axis) and AR module score (x axis) with each dot representing a GRN. Colors of GRNs correspond to GRN annotation on right separated by AR-positive, AR-negative and NEPC groups. Linear fit was calculated between two markers for only CRPC–adeno GRNs; the corresponding Pearson’s correlation is noted only for CRPC–adenocarcinoma GRNs or AR-positive and AR-negative GRNs alone. (E) A boxplot for DLL3 imputed expression (MAGIC, k = 20, t = 1) is shown for NEPC-A, NEPC-H/S and NEPC-N regulons. Significance was assessed by Wilcoxon-signed rank test. Abbreviations: ****(P < 1 × 10⁻⁴). (F) Immunohistochemistry of a liver with multiple metastases (PMID 3459916) shows distinct ASCL1–dominant (green dotted line) and NEUROD1-dominant (pink dotted line) foci prospectively stained for DLL3 expression. Zoomed images of two regions with DLL3+ and DLL3-negative foci are shown for DLL3, ASCL1, and NEUROD1 expression. (Scale bar, 50 µM.) (G) Dot plot of DLL3 expression [non-imputed, log(X + 1)] in CRPC tumor biopsies in single cell human RNA–sequencing data. This analysis suggests that a subset of CRPC adenocarcinoma cells are DLL3 expressors. On the right, representative immunohistochemistry is shown of a biopsy with interspersed ASCL1/DLL3 cells among AR positive cells (Patient 4). (Scale bar, 50 µM.)

We next studied STEAP1 and STEAP2, both of which showed a positive correlation with AR signaling in CRPC samples (R = 0.92, P = 1.8 × 10⁻⁴ and R = 0.94, P = 5.7 × 10⁻⁵, respectively). Of note, the AR+ HOXB13-negative GRN in the AR^high/PSMA^low sample showed robust STEAP1 and STEAP2 expression, suggesting that cotargeting of STEAP and PSMA in AR-positive disease may be an effective strategy to achieve broader tumor cell coverage (4) (Fig. 4D). In this context, we also unbiasedly identified other cell surface markers within our GRNs that could be utilized in combination with known cell surface antigens, such as PSMA (e.g., CEACAM5, FGFR1, PMEPA1, and others in SI Appendix, Fig. S10D).

Turning next to TACSTD2 (TROP2), the target of a clinically approved ADC for triple-negative breast and bladder cancers (with additional clinical trials underway in lung adenocarcinoma and prostate cancer), we noted TROP2 expression in most CRPC–adenocarcinoma clusters but with no correlation with AR expression (Fig. 4D). This finding is consistent with our immunohistochemistry data where TROP2 was expressed in all ADCs and nearly all HGCs (Fig. 1H) (20). In contrast to TROP2, CEACAM5 displayed a negative correlation with AR in CRPC–adenocarcinoma and was expressed in all NEPC clusters, suggesting that CEACAM5 is an actionable target for non-AR-driven disease (Fig. 4D) (49).

Given that DLL3 is a therapeutic target for both SCLC and NEPC and is downstream of ASCL1 (50), we next explored the expression of DLL3 in our NEPC regulons. While DLL3 was expressed in all NEPC regulons, NEPC–N regulons displayed lower expression compared with NEPC–A (Fig. 4E and SI Appendix, Fig. S10 E and F). As there were no NEPC–N tumors represented in our TMA cohort, which consists of punch biopsies from tumor blocks (Fig. 1), we further studied full face sections of a liver metastasis from a patient with ASCL1-positive NEPC, a case previously identified in a study of neuroendocrine chromatin landscapes (32). IHC analysis revealed divergent ASCL1 and NEUROD1 expression in discrete tumor foci. While DLL3 was abundantly expressed in the ASCL1-positive foci, we observed little or no expression in the NEUROD1-positive foci (Fig. 4F). While this spatial analysis is from a single patient, the collective single-cell sequencing data reveal differing levels of DLL3 expression across the NEPC spectrum. This heterogeneity could become an important variable in interpreting clinical response data in NEPC patients receiving DLL3-targeted therapy.

Because NEPC typically arises as a consequence of lineage transformation from PRAD, we next looked at DLL3 expression in our adenocarcinoma cohort and found clear evidence of DLL3 expression within subsets of cells within CRPC–adenocarcinoma tumors (Fig. 4G and SI Appendix, Fig. S10G). Within these CRPC PRAD samples, a subset of the DLL3 expressors were also positive for CHGB and ASCL1 expression and scored highly for the NEPC gene signature (SI Appendix, Fig. S10H). Similarly, the TMA identified rare HGC tumors with mixed lineage marker expression (AR and ASCL1) that also expressed DLL3 (example shown in Fig. 4G). Collectively, these data raise the possibility of early therapeutic targeting of rare NEPC cells in tumors with high-grade morphology or plasticity-associated genotypes, such as TP53 and/or RB1 loss.

Discussion

Multiple cancer types, after treatment with next-generation targeted inhibitors, can evolve to develop an array of heterogenous lineage states—a process often referred to as lineage plasticity (9). Prostate cancer serves as an archetype for the emergence of such plastic drug-resistant cell states, typified by the transformation from adenocarcinoma to neuroendocrine cancer (10). These cell states in prostate cancer as well as other tumor types are generally associated with poor responses to signaling inhibitors, current cell surface-based therapeutics, or chemotherapeutics (11). While there has been growing insight into the different cell states that may emerge in both mouse and human prostate tumors, an understanding of the diversity of transcriptional networks underlying these cell state changes and their associated lineage and cell surface marker expression in plastic prostate tumors remain limited.

To enhance our understanding of these lineage states and how they relate to cell surface marker expression, we pursued two parallel approaches: i) annotation of individual marker gene expression across an extensive TMA panel of late-stage PRAD and NEPC samples and ii) single-cell transcriptome analysis to identify distinct cell states, as well as the putative GRNs associated with those states, with subsequent linkage back to individual marker gene expression. The single marker gene approach confirmed that both YAP1 and TROP2 are robust markers for PRAD and HGC histology, whereas DLL3 is an exclusively NEPC-specific cell surface antigen. However, deeper analysis of the NEPC state with additional markers (SYP, ASCL1, NEUROD1 and INSM1) revealed important caveats that could help refine and sharpen the clinical diagnosis of NEPC. For example, both ASCL1 and INSM1 are highly specific markers of NEPC and tend to lose expression of YAP1, providing a strong dichotomy between PRAD and NEPC states. SYP is robustly expressed in the majority of NEPC as well (and is correlated with ASCL1 expression) but is also robustly expressed in a subset of AR-positive PRAD and HGCs that do not express ASCL1 or INSM1 (often referred to as “amphicrine”). Thus, SYP expression alone is neither sufficiently sensitive nor specific for defining bona fide NEPC states.

Our analysis of NEUROD1 expression through TMA-based protein expression and scRNA-seq analysis is similarly revealing due to the rarity of NEUROD1-positive vs. ASCL1-positive NEPC, particularly considering the relative frequencies of NEUROD1- vs. ASCL1-positive SCLC. These differences could simply be a consequence of tissue/cell of origin (e.g., prostate adenocarcinoma cells vs. lung neuroendocrine cells). However, recent studies in SCLC have established that NEUROD1-positive clones can emerge from ASCL1-positive cells, particularly in response to bottlenecks imposed by selective pressure from chemotherapy (51, 52). This plasticity between ASCL1- and NEUROD1-positive states may explain our detection of both populations by scRNA-seq in two NEPC samples. The fact that DLL3 expression is significantly lower in NEUROD1-positive NEPC cells may have implications for the clinical success of DLL3-targeted therapies. Taken together, these insights into ASCL1 vs. NEUROD1 expression within NEPC and SYP expression in PRAD argue for a standardized IHC panel-based approach using ASCL1, NEUROD1, INSM1, and SYP, in conjunction with histomorphological assessment, to add greater precision to the diagnosis and treatment of NEPC.

While IHC panel–based approaches may yield improved insight into cell states, our study has demonstrated they do not capture the heterogeneity of late-stage prostate cancer, both across patients, but equally important within a single patient. Our analysis of single-cell transcriptomes has thus provided further insight into the heterogeneity of cell states that underlie PRAD and NEPC. First, we observed markedly increased transcriptional diversification in CRPC and NEPC when compared to naive/CSPC tumors. This increase may be a consequence of the expanded number of putative GRNs. In addition to well-established AR-positive state (HOXB13+ and FOXA1+ GRNs), we identified inflammatory and GI lineage states (10, 46) that have previously been implicated in ARSI resistance. The AR-negative GRNs included epithelial-mesenchymal and embryonic/stem (TWIST2, SOX2/4, FOSL1/2), inflammatory (STAT1/2), and WNT signaling (TCF7L2). Many of these patient-derived transcriptional states are also present in GEMMs as well as human tumoroids (10, 22), providing further validation of the clinical relevance of these models for preclinical studies. We also note specific populations unique to murine models (e.g., Pou2f3+ cells in the setting of prostate-specific Trp53, Rb1, Pten deletion) that we failed to detect in our human TMA or single-cell data.

We analyzed how the expression of common cell antigens used for antibody–drug conjugates, T cell engagers, and theranostics varies as a function of transcriptional states. As expected, AR-positive GRNs were correlated with PSMA expression, with a clear exception in MSK–HP13 which demonstrated an AR-positive, HOXB13-negative regulon lacking PSMA expression. The latter is consistent with recent evidence implicating HOXB13 as a direct regulator of PSMA (45). However, because comparable levels of HOXB13 activity can be present in PSMA+ and PSMA– cells (for example, see MSK–HP13 SI Appendix, Fig. S10), HOXB13 expression alone is not sufficient. Towards this, AR^high/PSMA^high and AR^high/PSMA^low networks could be useful in identifying these additional PSMA regulators. Another clinically relevant finding driven by our analysis is the tight correlation of STEAP1 and STEAP2 expression with all AR^high regulons regardless of PSMA status, raising the potential for STEAP-targeted ADC therapy alone or in combination with PSMA-directed RLT (4). This is further supported by immunohistochemical analysis of the rapid autopsy tissues, which demonstrated a lower proportion of PSMA-high (45%) vs. STEAP1-high (70%) tumors (53). TROP2 expression was enriched across all adenocarcinomas regardless of AR status, indicative of a broader lineage profile (20).

In summary, our findings provide a comprehensive atlas of progressive heterogeneity in late-stage prostate cancer, including the identification of putative transcriptional networks and their association with lineage and cell surface markers. As far as potential limitations (site of biopsy, time of processing, batch correction, etc.) are concerned, identification of such GRNs requires additional validation in larger cohorts, ideally linked with chromatin accessibility data (e.g., multiome). With recent advances in comprehensive molecular diagnostic liquid assays, one can envision incorporation of GRN-based classification as an additional tool to refine patient selection and therapy decisions.

Supplementary Material

Appendix 01 (PDF)

Dataset S01 (XLSX)

Dataset S02 (XLSX)

Dataset S03 (XLSX)

Dataset S04 (XLSX)

Dataset S05 (XLSX)

Dataset S06 (XLSX)

Dataset S07 (XLSX)

Dataset S08 (XLSX)

Dataset S09 (XLSX)

Dataset S10 (XLSX)

Dataset S11 (XLSX)

Dataset S12 (XLSX)

Dataset S13 (XLSX)

Dataset S14 (XLSX)

Dataset S15 (XLSX)

Dataset S16 (XLSX)

Data, Materials, and Software Availability

Human raw data for a subset of Naive and CSPC samples, as per Karthaus et al. (38) are available at the Data Use and Oversight System controlled access repository https://duos.broadinstitute.org/ [Accession No. DUOS-000115, samples: HP95 (MSK–HP01), HP96 (MSK–HP02), HP97 (MSK–HP03), HP99 (MSK–HP04), HP100 (MSK–HP05), and HP101 (MSK–HP06)] (39). Human raw data and 10X formatted files for CRPC samples are available at Gene Expression Omnibus repository [GSE210358, (10)] (40). For previously unpublished samples, HMP22 (MSK–HP16), HMP23A/B (MSK–HP07), HMP24 (MSK–HP08), FASTQ, and 10X files have been uploaded to Gene Expression Omnibus repository, along with two RDS files that contain all and tumor cells, respectively (Accession ID GSE264573) (41). GEMM raw data and 10X formatted files for WT, PtR, and PtRP are available at Gene Expression Omnibus repository [GSE210358, (10)] (40). Code for notebooks to reproduce figures will be available at Githhub: https://github.com/zaidisamir/ (54). All other data are available in the manuscript or supporting information.