Increasing Stemness Drives Prostate Cancer Progression, Plasticity, Therapy Resistance and Poor Patient Survival

Abstract

Cancer progression involves loss of differentiation and acquisition of stem cell-like traits, broadly referred to as “stemness”. Here, we test whether the level of stemness, assessed by a transcriptome-derived Stemness score, can quantitatively track prostate cancer (PCa) development, progression, therapy resistance, metastasis, plasticity, and patient survival. Integrative analysis of transcriptomic data from 87,183 samples across 26 datasets reveals a progressive increase in Stemness and decline in pro-differentiation androgen receptor activity (AR-A) along the PCa continuum, with metastatic castration-resistant PCa (mCRPC) exhibiting the highest Stemness and lowest AR-A. Both the general Stemness score and a newly developed 12-gene “PCa-Stem Signature” correlate with and predict poor clinical outcomes. Mechanistically, increased AR-A may promote Stemness in early-stage PCa while MYC amplification and bi-allelic RB1 loss likely drive greatly elevated Stemness in mCRPC where AR-A is suppressed. Our findings establish Stemness as a robust quantitative measure of PCa aggressiveness and offer a scalable framework for PCa risk stratification.

INTRODUCTION

Malignant transformation and cancer progression are characterized by progressive loss of differentiation (or oncogenic dedifferentiation) and gain of stem cell traits (i.e., stemness), and poorly differentiated tumors are enriched in gene expression profiles of normal embryonic and adult stem cells and in cancer stem cells^{1, 2, 3, 4, 5}. However, the oncogenic dedifferentiation-based stemness concept⁶ that prostate cancer (PCa) represents abnormal organogenesis^7,8 has not been systemically addressed and there lack quantitative measurements of the overall stemness-associated aggressiveness in PCa.

PCa is projected to claim the lives of >35,000 American men in 2025⁹. PCa incidence rates increased 3% per year from 2014 – 2019 driven by advanced PCa diagnosis and the proportion of men diagnosed with metastases has doubled from 2011 to 2019 when the next generation of androgen receptor (AR) pathway inhibitors (ARPIs) such as abiraterone (Zytiga; FDA approval in 2011) and enzalutamide (Xtandi; FDA approval in 2012) were introduced to the clinic. Significant increases in diagnosis of aggressive and advanced disease and continued high PCa mortality have been linked to therapy resistance driven by intrinsic PCa cell heterogeneity (e.g., in AR expression) and treatment-induced plasticity⁸.

The level of malignancy of primary PCa (Pri-PCa) has been traditionally assessed by pathologists using a combined Gleason Score (GS), which gauges histopathological differentiation pattern of prostatic glands. The higher the GS, the more undifferentiated and more malignant the Pri-PCa is. The GS system, although clinically the ‘gold’ standard in pinpointing PCa diagnosis, is only semi-quantitative, can be subjective (to individual pathologist’s assessment), and is generally not applicable to treatment-failed tumors and distant metastases (which have mostly lost differentiated glandular structures). Notably, whether more advanced Pri-PCa (GS8-10), treatment failure and metastatic dissemination are quantitatively associated with increased cancer stemness remains unknown. On the other hand, AR is the master regulator of normal prostatic epithelium differentiation and represents the therapeutic target of PCa treatments including androgen-deprivation therapy (ADT) and ARPIs. Inevitably, tumors develop resistance to ADT/ARPIs leading to the state of castration-resistant PCa (CRPC), which involves many mechanisms including genetic alterations of AR and key genes in AR pathway¹⁰, lineage plasticity^{8, 11, 12}, and increased stemness^{8, 13}. It is unclear how the global pro-differentiation AR signaling activity changes and evolves during the spectrum of PCa initiation and progression and during development of treatment resistance and metastasis. These discussions raise the need to quantify the AR signaling and stemness and closely track their dynamic evolution and inter-relationship during PCa progression.

Herein, we adapted and developed transcriptome-based gene signature scores to quantitatively assess the pro-differentiation (canonical) AR activity (AR-A) and oncogenic stemness levels (Stemness Index or Score) and their inter-relationship in the PCa spectrum. Specifically, we used gene expression data collected from 87,183 samples in 26 distinct clinical cohorts encompassing the PCa continuum, i.e., from normal/benign prostate to early stage, advanced and metastatic PCa, to pre- and post-therapy PCa, and, finally, mCRPC, to track the dynamic evolution of and inter-relationship between AR-A and the Stemness Index. The results indicate that increased Stemness quantitatively tracks PCa aggressiveness, plasticity, treatment resistance and progression, and represents a poor prognostic factor for patient outcomes. Our study offers a trajectory-integrated, pan-cohort framework that defines and validates cancer Stemness as a quantifiable clinical determinant of PCa aggressiveness, plasticity, progression and therapy resistance.

RESULTS

Adapting and developing Stemness and AR-A metrics for PCa analysis.

To determine the global and dynamic changes of and inter-relationship between degree of malignancy (oncogenic dedifferentiation) and level of (luminal) differentiation during PCa development and progression (Supplementary Fig. S1A–B), we employed two transcriptome-based algorithms to annotate 26 datasets of 3,102 bulk RNA-seq and 84,081 microarray samples encompassing the PCa continuum (Supplementary Fig. S1B; Supplementary Table S1). In our studies, we focused on and compared three progressive stages: from normal prostate to localized Pri-PCa, from low-grade to high-grade Pri-PCa, and from treatment-naïve Pri-PCa to CRPC/mCRPC (Supplementary Fig. S1A). We defined “progression” as stages beyond Pri-PCa and as disease entities that are more aggressive in a comparative manner. Specifically, we adapted the Stemness Index⁶ (mRNAsi; Stemness for short) (Supplementary Fig. S1C) to gauge the degree of malignancy (or aggressiveness) and employed the Z-score normalized canonical (pro-differentiation) AR activity or AR-A (Supplementary Fig. S1D) to measure the level of differentiation. Canonical AR-A can be determined by quantifying AR-regulated transcripts under intact androgen/AR signaling conditions, and there are several such AR-regulated gene lists^{14, 15, 16, 17, 18} (Supplementary Fig. S2A–B). We validated the AR-A list based on their enrichment in prostatic epithelium differentiation, correlation of individual gene expression with AR-A signatures in clinical datasets, and overall consistency in correlation among the AR-A scores (Supplementary Fig. S2B–D). Given that Bluemn’s AR-A score¹⁴ correlated well with other scores (Supplementary Fig. S2E–G) and the 10 genes in this signature consistently represented AR targets in multiple clinical datasets (Supplementary Fig. S2D; Supplementary Table S2), we decided to use this AR-regulated gene list for our AR-A quantification.

Normal prostate epithelium shows a negative correlation between Stemness and AR-A.

The normal human prostate (NHP) comprises 2 major epithelial types: basal and luminal cells (Supplementary Fig. S3A). We annotated two studies that profiled FACS-purified NHP luminal and basal cell populations^{19, 20} and observed that the luminal cell population (Trop2⁺CD49f^lo) possessed higher AR mRNA levels and AR-A but lower Stemness compared to basal cells (Trop2⁺CD49f^hi) and, in both datasets, AR-A negatively correlated with Stemness (Supplementary Fig. S3B–C). These results support that the NHP basal cell compartment harbors primitive stem cells^{7, 19, 20, 21, 22} and AR-A measures the level of prostate epithelial differentiation.

Pri-PCa exhibits concordantly increased Stemness and AR-A, and neoadjuvant ADT (nADT) decreases both Stemness and AR-A.

In analyzing AR-A and Stemness in matched pairs of prostate tumor (T) and normal (N; tumor-adjacent benign) tissues in the TCGA-PRAD¹⁸ (n=51 pairs), Wyatt 2014²³ (n=12 pairs), and Long 2020²⁴ (n=10 pairs) cohorts, we observed concordantly increased AR-A and Stemness (Fig. 1A–C). Accompanying elevated AR-A, AR mRNA levels were also increased in Pri-PCa compared to benign tissue (Fig. 1A–C; note that the increase in AR mRNA levels in the Long 2020 cohort approached statistical significance). The paired T/N comparisons (Fig. 1A–C) suggest that early PCa initiation and growth are characterized by concordantly increased Stemness and AR-A. To test whether AR-A might causally drive increased Stemness early during prostate tumorigenesis, we investigated 4 PCa nADT (i.e., ~2–6 months of ADT before prostatectomy or radiation) cohorts^{24, 25, 26}. In the first 3 cohorts with matched pairs of pre- and post-nADT samples (from the same patients), nADT concordantly downregulated AR-A (without decreasing AR mRNA levels) as well as the Stemness (Fig. 1D–F). In the Nastiuk cohort with non-paired post-nADT and no-nADT samples (i.e., from different patients but matched for ages and tumor grades and stages), nADT increased AR mRNA levels (Supplementary Fig. S4A) but reduced both AR-A and Stemness (Supplementary Fig. S4B–C).

Figure 1. Pri-PCa shows concordant increases of AR-A and Stemness and both are decreased by nADT.

A-C, PCa has increased AR-A as well as the Stemness compared to matched benign tissues. Pairwise comparisons of AR mRNA levels, canonical AR-A, and Stemness scores showing significant increases in primary tumors (T) as compared to the matched adjacent benign prostate tissues (N) from the same individual in TCGA-PRAD cohort (n=51; A), and Wyatt 2014 cohort (n=12; B). In the Long 2020 cohort (n=10; C), increases in AR-A and Stemness reached statistical significance whereas the increase in AR mRNA levels approached statistical significance (P = 0.06).

D-F, nADT decreases both AR-A and Stemness. Pairwise comparison of AR mRNA levels, AR-A, and Stemness showing significant decrease of both AR-A and Stemness after nADT (Post-nADT) as compared to the matched samples before nADT (Pre-nADT) from the same individual (D, Rajan 2014 cohort; E, Sharma 2018 cohort; F, Long 2020 cohort).

In all plots, sample sizes are indicated (see also Supplementary Table S1 for cohort and RNA-seq data access information). Within the plots, results are shown as the mean ± SD and pairs are linked with a grey line. Significance was calculated by two-tailed paired Student’s t-test (ns, not significant; *, P < 0.05; **, P < 0.01; ***, P < 0.001; ****, P < 0.0001).

Pri-PCa progression is accompanied by reduced AR-A but continually increasing Stemness.

We next assessed the dynamic changes of Stemness and AR-A during spontaneous (i.e., therapy-free) Pri-PCa progression (Fig. 2). To this end, we sorted out treatment-naïve tumors in TCGA-PRAD by excluding samples from patients who received adjuvant (post-surgery) hormone therapy (i.e., Pri-PCa/ADT or Tx; n=70) (Fig. 2A) and grouped the Pri-PCa samples according to increasing GS (Fig. 2B–D). When compared with the 52 benign (N) samples, all the four GS groups from Pri-PCa (GS6, GS7, GS8, and GS9/10 combined) showed higher AR-A and Stemness (Fig. 2C–D), consistent with the results in pairwise T/N comparisons (Fig. 1A–C). However, when compared to GS6 PCa, the three higher GS groups showed increasing AR mRNA levels (Fig. 2B) but decreasing AR-A (Fig. 2C) and increasing Stemness (Fig. 2D). Jonckheere-Terpstra’s (J-T) trend test validated a gradual reduction in canonical (pro-differentiation) AR signaling but an increasing trend of Stemness with Pri-PCa progression (Fig. 2C–D). Notably, although we observed a modest positive correlation between AR-A and Stemness in pooled Pri-PCa samples (Pearson’s r=0.27; Fig. 2E), this positive correlation gradually diminished with PCa progression when we categorized the samples by tumor grade (Fig. 2F). The linear regression analysis confirmed that the interaction between Stemness and AR-A significantly depended on tumor grade (P=0.0003), with a linear decrease in the slope of the correlation as GS increased (Fig. 2G).

Figure 2. Continual increases in Stemness with declining AR-A during spontaneous PCa progression.

A, The pie chart showing sample sizes in different Gleason grades in the TCGA untreated Pri-PCa cohort.

B-D, Progression of treatment-naïve Pri-PCa is accompanied by decreasing AR-A but increasing Stemness. Data is presented as mean ± SD. Shown on top are the Jonckheere-Terpstra’s trend (J-T) test results (wedges).

E-F, Progression in Pri-PCa is accompanied by a gradual loss of the positive correlation between canonical AR-A and Stemness. Scatter plots illustrating the correlation between AR-A and Stemness in (E) combined Pri-PCa cases in TCGA with an overall linear regression model and (F) individual GS grades with corresponding linear regressions showing decreasing slopes from low to high GS grades by J-T test. The slopes of best-fit regression line are also shown.

G, Merged scatter plot showing the correlation between AR-A and Stemness across the 4 GS PCa with each GS represented by a distinct color. The correlation becomes progressively weaker from GS6 to GS9/10, transitioning from positive to nearly zero. The correlation coefficients (r) and p-values for each GS are as follows: GS6 (r = 0.56, P < 0.0001), GS7 (r = 0.37, P < 0.0001), GS8 (r = 0.48, P = 0.022), and GS9/10 (r = −0.02, P = 0.89). Linear regression lines are overlaid for each GS, demonstrating the decline in the correlation co-efficiency as the GS increases.

All statistical analyses were performed using R and statistical significance was assessed using one-way analysis of variance followed by Tukey’s multiple-comparison test (B-D). J-T trend test in B-D and F was used to calculate the statistical significance of the trend across different groups. *, P < 0.05; **, P < 0.01; ***, P < 0.001; ****, P < 0.0001; ns, not significant.

Collectively, these results indicate that, in contrast to early tumorigenesis, spontaneous PCa progression (i.e., with increasing GS and tumor grade) is marked by a divergence between oncogenic Stemness and pro-differentiation AR-A.

Therapies and metastatic progression drive further divergence between attenuated AR-A and increased Stemness.

Next, we asked how clinical treatment and metastatic progression may impact the Stemness and AR-A by querying untreated Pri-PCa and Pri-PCa/ADT from TCGA and mCRPC in SU2C¹⁰ (Fig. 3A–C). AR-A was elevated in Pri-PCa compared to adjacent benign tissue (N) but reduced in more aggressive Pri-PCa/ADT samples, while the Stemness, in contrast, continually and persistently increased from N → Pri-PCa → Pri-PCa/ADT (Fig. 3B–C). Strikingly, mCRPC, despite markedly upregulating AR mRNA (Fig. 3A), showed the lowest AR-A but highest Stemness (Fig. 3B–C), indicating a shift from a positive correlation between AR-A and Stemness in Pri-PCa to a negative association in mCRPC (Fig. 3D).

Figure 3. Therapies and metastatic progression drive further divergence between Stemness and canonical AR-A with mCRPC showing the highest Stemness but the lowest AR-A.

A-D, Hormonal treatment further drives down AR-A but significantly upregulates Stemness. A-C, mCRPC displays the lowest pro-differentiation AR-A but highest Stemness. Shown are AR mRNA levels (A), the canonical AR-A (B) and Stemness (C) in tumor-adjacent benign tissues (N), treatment-naïve Pri-PCa and advanced Pri-PCa with post-surgery adjuvant ADT (Tx) in TCGA, and mCRPC (SU2C 2015). Sample sizes are indicated in parentheses. Shown on top are the Jonckheere-Terpstra’s (J-T) trend test results (wedges). D, Stemness inversely correlates with AR-A in mCRPC. Shown are scatter plots with best-fit linear regression lines illustrating the correlation between AR-A and Stemness with Pearson correlation coefficient (r), the slopes of lines and J-T test results of slopes indicated below.

E-F, PCa progression is accompanied by loss of AR-A (E) but gain of Stemness (F) across PCa states. Shown are samples representing the spectrum of PCa development and progression, from normal (organ donors in GTEx) to (tumor-adjacent) benign tissues, to (treatment-naïve) Pri-PCa, aggressive Pri-PCa treated with adjuvant ADT (Pri-PCa/ADT), mCRPC, and, finally, to (clonally-derived) PCa models including PDX, xenografts and cells lines. From Left to right: Normal (n=116): Normal prostates from GTEx; Benign (n=52): tumor-adjacent benign tissue from TCGA-PRAD; Pri-PCa (n=422): treatment-naïve localized PCa from TCGA-PRAD; Pri-PCa/ADT (n=70): advanced Pri-PCa treated with adjuvant hormone therapy from TCGA-PRAD; mCRPC (n=59): mCRPC from GSE126078; mCRPC (n=135): mCRPC from SU2C 2015 cohort; mCRPC-Adeno (n=34): adenocarcinoma mCRPC from Beltran 2016 cohort; mCRPC-NE (n=15): neuroendocrine mCRPC from Beltran 2016 cohort; mCRPC (n=42): mCRPC from Westbrook 2022; mCRPC (n=25): mCRPC from Alumkal 2020; PDX-LuCaP (n=39): PCa PDXs from GSE126078; PCa cell lines (n=8): PCa cell lines from CCLE; XG-LAPC9 (n=10) and XG-LNCaP (n=12): PCa xenografts from GSE88752. Statistical significances between groups were shown in inverted pyramid grid plot as indicated by different colors. Student’s t-test was used to compare each group and adjusted p-value for multiple comparisons was done by Tukey’s range test. See Supplementary Table S4 for the summary of statistical comparisons. J-T trend test was used to assess the statistical significance of the trend across groups. See also Supplementary Table S1 for cohort and RNA-seq data access information.

G, Scatter plots in the Bolis 2021 integrated cohort (normal n=174, Pri-PCa n=714, mCRPC n=335) showing the changing correlation between AR-A and Stemness across disease stages: positive in normal to weaker correlations in Pri-PCa to negative correlations in mCRPC. Regression lines are color-coded, and the slope of the best-fit regression line and Pearson’s r are provided: yellow for normal (Pearson’s r=0.42, P <0.0001; slope: 43.87, 95%CI, 29.57 to 58.18); blue for Pri-PCa (r=0.32, P <0.0001; slope: 15.04, 95%CI, 11.76 to 18.32); and green for mCRPC (r=−0.15, P = 0.005; slope: −17.25, 95%CI, −29.24 to −5.26).

H, Violin plot showing increasing Stemness among the NCCN Risk Groups containing 82470 microarray samples with localized disease in GRID. J-T trend test was used to assess the statistical significance of the trend across groups.

I-M, Increasing Stemness during PCa progression in the Taylor 2010 cohort consisting of tumor-adjacent benign tissue (Normal; n=29), Pri-PCa (n=131) and metastatic mPCa (n=19). The AR-A significantly increased in Pri-PCa vs. Normal and significantly decreased in mPCa compared to Pri-PCa (I) whereas the Stemness continued to increase from Benign to Pri-PCa to mPCa (J). K-M, Scatter plots illustrating the correlations between AR-A and Stemness in different contexts with corresponding linear regression lines: K, Positive correlation in normal (Pearson’s r=0.69, P <0.0001); L, a gradual loss of positive correlation between Stemness and AR-A during PCa progression (Pearson’s r=0.18, P = 0.043); M, a loss of positive correlation between the Stemness and AR-A at more advanced mPCa stage (Pearson’s r=0.18, P =0.47).

All data in dot plots and violin plots are shown as mean ± SD. Statistical significance was determined using one-way analysis of variance (ANOVA) followed by Tukey’s multiple-comparison test. Jonckheere-Terpstra’s trend (J-T) test was used to calculate the statistical significance of the trend across different groups. Pearson correlation coefficients (Pearson’s r) were used to calculate the correlations between the Stemness and AR-A. Note: ns, not significant; *, P < 0.05; **, P < 0.01; ***, P < 0.001; ****, P < 0.0001.

By employing normalized RNAseq data from the PCa transcriptome atlas (Bolis 2021), which included normal prostate specimens (n = 174), Pri-PCa (n = 714) and mCRPC (n = 335)²⁷, we tracked the Stemness and AR-A across the PCa spectrum (Fig. 3E). AR-A increased from benign to Pri-PCa but declined in Pri-PCa/ADT and continued to decrease in mCRPC, reaching its lowest in mCRPC-NE²⁸ as supported by J-T test (Fig. 3E). In contrast, the Stemness exhibited a continually increasing trajectory across the PCa spectrum encompassing Normal/Benign → treatment-naïve Pri-PCa → Pri-PCa/ADT → (m)CRPC (P<0.0001, J-T Trend test; Fig. 3F; Supplementary Fig. S5A). Notably, the 3 subtypes of mCRPC, i.e., AR-positive (ARPC), CRPC-NEPC and AR⁻/NE⁻ double-negative PCa (DNPC) all showed higher Stemness than Pri-PCa and also displayed increasing Stemness among the 3 subtypes (Supplementary Fig. S5B). Unexpectedly, tumor-adjacent benign prostate tissues showed much higher Stemness than normal prostate in the Genotype-Tissue Expression Project (GTEx)²⁹ (Fig. 3F; Supplementary Fig. S5A), which were all obtained from cancer-free organ donors. This surprising finding suggests that the transcriptomes of adjacent benign tissues have apparently skewed towards those of cancer samples, consistent with previous observations that adjacent benign tissues represent an intermediate state between healthy and tumor tissues³⁰. Also of interest, the Cancer Cell Line Encyclopedia (CCLE)^{31, 32} cell lines including PCa cell lines, and PCa-derived PDX³³ and xenografts³⁴ exhibited the highest Stemness (Fig. 3F; Supplementary Fig. S5A), supporting that cancer cell lines and xenografts are clonally derived from and highly enriched in cancer stem cells (CSCs).

As in our early dissection of the relationship between AR-A and Stemness during spontaneous Pri-PCa progression (Fig. 2F–G), analysis of the Bolis 2021 PCa transcriptome atlas²⁷ revealed a similar change from a positive correlation between AR-A and Stemness in Pri-PCa (blue line; Pearson’s r=0.32, P<0.0001; Fig. 3G) to a negative correlation between the two in mCRPC (green line; Pearson’s r=−0.15, P=0.005; Fig. 3G). We further extended our Stemness and AR-A interrogation to include microarray-derived transcriptomic datasets^{35, 36}, allowing us to access data from PCa cohorts generated before the Next-generation sequencing (NGS) era. Analyzing the GRID database of over 82,000 prospectively collected biopsy samples of localized PCa from clinical use of the Veracyte Decipher test³⁵, we systematically quantified the Stemness across the NCCN Risk Groups and observed increasing Stemness along the NCCN Risk trajectory (Fig. 3H). Likewise, in the Taylor dataset³⁶, we observed that AR-A was the highest in Pri-PCa but significantly decreased in metastatic PCa (mPCa) (Fig. 3I) while the Stemness increased progressively from benign to Pri-PCa and reached the highest levels in mPCa (Fig. 3J). Linear regression analysis confirmed a gradual loss of positive correlation between AR-A and Stemness during PCa progression, as illustrated by distinct regression lines for each stage (Fig. 3K–M).

Finally, we analyzed the dynamic changes of Stemness and AR in genetically engineered mouse models³⁷ (GEMMs) of PCa with increasing aggressiveness due to individual or combined deletion of 3 tumor suppressor genes, Pten, Rb1 and Trp53. Briefly, Pten^−/− single knockout (SKO) prostate tumors develop around 9 weeks, and mice rarely develop metastasis with a median lifespan of 48 weeks. In contrast, the Pten^−/−;Rb1^−/− double knockout (DKO) mice develop highly metastatic PCa that shortens median survival to ~38 weeks. When Trp53 is further deleted in DKO background, the triple KO (TKO; Pten^−/−;Rb1^−/−;Trp53^−/−) tumors are exclusively AR⁻ NEPC and castration-resistant de novo with high metastatic rate and lifespan of ~16 weeks³⁷. We found that the AR-A was the highest in SKO tumors but much reduced in aggressive DKO and TKO tumors (Supplementary Fig. S5C). In contrast, Stemness was the lowest in SKO but significantly increased in DKO and TKO (Supplementary Fig. S5D). Correlation analyses revealed a positive correlation between AR-A and Stemness in SKO that turned negative in both DKO and TKO tumors (Supplementary Fig. S5E–H).

The above results, collectively, indicate that hormone therapy and metastatic progression further drive marked increases in Stemness with concomitant decreases in pro-differentiation AR-A.

Stemness is prognostic of poor survival in PCa patients.

We investigated global Stemness and its prognostic significance, and observed that high Stemness consistently correlated with worse patient survival across various cohorts including the Spratt 2017³⁸ (P=0.001, log-rank; Fig. 4A), Tosoian 2020³⁹ (P=0.014, log-rank; Fig. 4B), CHAARTED 2021⁴⁰ (P=0.006, log-rank; Fig. 4C; P=0.149, log-rank; Supplementary Fig. S6A), and the mCRPC (SU2C 2019⁴¹; P=0.06, log-rank; Supplementary Fig. S6B–C) as well as Taylor 2010³⁶ (P=0.028, log-rank; Supplementary Fig. S6D–E). Notably, the CHAARTED cohorts displayed the highest mean Stemness compared to Tosoian and Spratt cohorts, and the CHAARTED cohort showed the highest median Decipher Prostate Genomic Classifier^{38, 39, 40} among the three cohorts (Fig. 4D). Interestingly, like high Stemness, low AR-A was correlated with poor patient survival in the Taylor dataset (P=0.040, log-rank; Supplementary Fig. S6F).

Figure 4. High Stemness is associated with poor patient overall survival.μ.

A-C, Kaplan-Meier plots showing that high Stemness associates with worse overall survival in PCa patients from Spratt 2017 (A), Tosoian 2020 (B), and CHAARTED 2021 (C) cohorts. P value was determined using the log-rank test.

D, Stemness increases across patient groups with higher Decipher genomic classifier scores. The three cohorts were arranged by their median Decipher genomic classifier scores, and the J-T trend test was used to assess statistical significance of this trend.

Stemness-high PCa are associated with aggressive molecular subtypes and a 12-gene PCa-Stem signature prognosticates poor patient survival.

To identify key determinants of Stemness and explore differences between tumors with high versus low Stemness, we stratified Pri-PCa (TCGA-PRAD) and mCRPC (SU2C 2019) samples separately based on Stemness scores (Fig. 5A). Differential gene expression (DEG) analyses revealed significant differences between the top 33% (Stemness-high) and the bottom 33% (Stemness-low) PCa samples in both treatment-naïve Pri-PCa (Fig. 5B) and treatment-failed mCRPC (Fig. 5C), resulting in a 12-gene “Pca-Stem signature” (Fig. 5D; Supplementary Table S2; see Methods).

Figure 5. Stemness-high PCa are associated with aggressive molecular subtypes and a 12-gene PCa-stem signature prognosticates poor patient survival.

A, Schematic illustrating the profiling strategy used to assess associations between Stemness and molecular features in PCa. Patients were ranked by Stemness score and stratified into Stemness-high (top 33%) and Stemness-low (bottom 33%) groups for downstream analyses. Genome-wide functional analyses were performed using transcriptomic, genomic, and clinical data.

B-C, Volcano plots showing differentially expressed genes (DEGs) between Stemness-high and Stemness-low groups in Pri-PCa (B) and mCRPC (C) cohorts. Red and blue indicate upregulated and downregulated DEGs, respectively. Genes with FDR > 0.05 are shown in gray. Sample sizes (n) of each group are indicated. See Supplementary Table S3 for complete DEG lists.

D Venn diagram showing the 12-gene “PCa-Stem signature” derived from overlapping upregulated DEGs (FC ≥ 2, FDR < 0.05) in both Pri-PCa and mCRPC.

E-F, GSEA showing low Stemness is associated with immune signaling and inflammatory response (E) while high Stemness in both cohorts is enriched for embryonic stem cell traits, aggressiveness, DNA repair, MYC activation, and mTORC1 signaling (F). All enrichments are significant (P < 0.05, FDR < 0.05).

G-H, PAM50 subtyping showing a higher frequency of LumB subtype in Stemness-high versus Stemness-low samples in Pri-PCa (G) and mCRPC (H). (****, P < 0.0001, χ² test).

I-L, GSEA showing enrichment of the aggressive PCS1 gene expression signature in Stemness-high groups and the less aggressive PCS3 signature in Stemness-low groups in Pri-PCa (I, K) and mCRPC (J, L). The Prostate Cancer Subtype (PCS) classification system (45,46) stratifies tumors into three molecular subtypes based on transcriptomic features: PCS1 (most aggressive), PCS2 (intermediate), and PCS3 (least aggressive). See Supplementary Table S2 for PCS gene signature.

M-R, GSEA showing AR signature enrichment in Pri-PCa Stemness-high (M), NEPC signature enrichment in mCRPC Stemness-high (N), and shared enrichment of cell cycle (O, P) and lineage plasticity (Q, R) programs in both cohorts. See Supplementary Table S5 for gene sets.

S-V, Kaplan-Meier analyses showing that the 12-gene PCa-Stem signature predicts worse progression-free survival (PFS) in TCGA Pri-PCa (S) and worse overall survival (OS) in SU2C mCRPC (U). Multivariable Cox models adjusting for clinicopathologic variables confirm the PCa-Stem-High subgroup as independently prognostic for PFS (T) and OS (V).

Gene set enrichment analysis (GSEA) was performed to delineate the molecular differences between Stemness-low and Stemness-high PCa samples (Fig. 5E–F). GSEA revealed that Stemness-high PCa samples were enriched in gene sets associated with embryonic stem cells (ESC), as well as aggressive cancer phenotypes, such as undifferentiated cancer and metastasis, in both Pri-PCa and mCRPC (Fig. 5F). In contrast, Stemness-low PCa samples were enriched in gene signatures linked to lower aggressiveness and inflammatory pathways (Fig. 5E). Additionally, hallmark pathways such as E2F targets, MYC targets, mTORC1 signaling, and DNA repair were upregulated in Stemness-high PCa (Fig. 5F), while TGF-β signaling, TNF-α signaling, and IFN-γ responses were prominent in Stemness-low PCa (Fig. 5E).

The aggressive PAM50-LumB molecular subtype^{42, 43, 44} was significantly more prevalent in Stemness-high PCa, accounting for 74% in Pri-PCa (compared to 15% in Stemness-low PCa; Fig. 5G) and 43% in mCRPC (compared to 4% in Stemness-low PCa; Fig. 5H). The PCS1 subtype^{45, 46}, known for its association with ADT resistance and the highest risk of progression to advanced disease in comparison with PCS2 or PCS3, was predominantly enriched in both Stemness-high Pri-PCa and mCRPC (Fig. 5I–L). Notably, the PAM50-Basal subtype^{42, 43} increased in Stemness-high mCRPC but not in Pri-PCa (Fig. 5G–H), aligning with the characteristics of mCRPC, which involves lineage plasticity and acquisition of basal, mesenchymal, neural and stem-like phenotypes when developing ARPI resistance^{7, 44, 47}. Consequently, Stemness-high samples remained largely AR-dependent⁴⁸ in Pri-PCa (Fig. 5M) but transitioned to AR-independent NE subtype⁴⁴ in mCRPC (Fig. 5N). Stemness-high PCa was also characterized by high proliferation as evidenced by their enrichment in a 31-gene cell-cycle progression signature⁴⁴ (Fig. 5O–P). Finally, we examined the association between Stemness and a signature of lineage plasticity risk after Enza treatment for mCRPC¹² and found that Stemness-high tumors may be at risk for this virulent form of treatment resistance (Fig. 5Q–R).

To further assess the clinical relevance of Stemness, we analyzed the prognostic impact of our newly developed 12-gene PCa-Stem signature (Fig. 5D). We observed that higher PCa-Stem signature correlated with worse patient survival in both Pri-PCa (P<0.0001, log-rank; Fig. 5S–T) and mCRPC (P<0.0001, log-rank; Fig. 5U–V) patients. Multivariate analysis, adjusted for clinicopathologic parameters such as age, GS and stage, showed that patients with high PCa-Stem signature had significantly worse independent prognosis compared to those with low PCa-Stem signature. This was demonstrated by worse progression-free survival (PFS) in Pri-PCa (multivariable-adjusted HR=1.636 (1.014 – 2.64), P=0.0438; Fig. 5T) and overall survival (OS) in mCRPC (multivariable-adjusted HR=4.712 (2.011 – 11.04), P<0.001; Fig. 5V), respectively. Furthermore, the PCa-Stem signature steadily and continually increased across the PCa evolutionary spectrum and correlated with PCa disease progression scores as supported by trajectory inference analysis of the Bolis PCa transcriptome atlas²⁷ (Pearson’s r=0.72; Supplementary Fig. S7A–B). Strikingly, gene expression analysis revealed that 11 of the 12 genes (except KLK12) in the PCa-Stem signature significantly correlated with the PCa disease progression (Supplementary Fig. S7C). In contrast, among the genes commonly downregulated in Stemness-high PCa samples in both Pri-PCa and mCRPC (Supplementary Table S3), some (e.g., PTGDS, SPARCL1, CLU, GJA1 and S100A4) were involved in regulating prostate-specific glandular structures and luminal functions while others (e.g., FBLN1, SFRP1, IGFBP6, GAS1 and TIMP2) played general roles in epithelial differentiation (Supplementary Fig. S7D).

Overall, these findings indicate that PCa with high Stemness exhibit greater aggressiveness, a higher likelihood of developing lineage plasticity and therapy resistance, and poor survival.

Stemness-high PCa have high tumor mutation burden (TMB) and unique genomic features.

We determined the potential differences in genomic features between Stemness-high vs. Stemness-low PCa and found that mCRPC had a higher fraction of altered genome (Fig. 6A) and higher TMB (Fig. 6B) than Pri-PCa. In both Pri-PCa and mCRPC, Stemness-high PCa displayed a higher fraction of altered genome (Fig. 6A) and TMB (Fig. 6B) than Stemness-low PCa although there was no difference in patient ages (Fig. 6C). In Pri-PCa, Stemness-high samples had a higher prevalence of PTEN deletion (DEL), SPOP and FOXA1 mutations (MUT) and MYC amplification (AMP) whereas in mCRPC, Stemness-high PCa had more prevalent RB1 DEL and AR MUT (Fig. 6D–F).

Figure 6. Stemness-high PCa is associated with higher genomic instability.

A-C, Box plots showing that Stemness-high subgroup in both Pri-PCa (n = 141) and mCRPC (n = 89) displays a higher fraction of altered genome (A), higher TMB (B), and no age difference compared to Stemness-low cases (C). Wilcoxon rank-sum test was used to determine significance (****, P < 0.0001; ns, not significant). Statistical significance was determined using the two-sided Wilcoxon rank-sum test. Box plots represent the median, interquartile range (IQR), and whiskers extending to 1.5× the IQR.

D-F, Comparisons of genomic alteration frequencies (Alt. Freq) in cancer-associated genes between Stemness-high and Stemness-low subgroups in TCGA Pri-PCa (D, F) and SU2C mCRPC (E, F) cohorts. Statistical significance was determined by two-sided Fisher’s Exact test (*, P < 0.05).

Globally, both Pri-PCa and mCRPC exhibited mutations in more than a dozen genes although most of these alterations were <10% (Supplementary Fig. S8A; green symbols), consistent with reports that PCa is genomically heterogeneous with a broad spectrum of genomic alternations of low penetrance^{49, 50}. On the other hand, homozygous deletion (HOMDEL) events, particularly, loss of PTEN and RB1, occurred frequently in both Pri-PCa and mCRPC (Supplementary Fig. S8A; blue symbols) whereas AMP events in oncogenes such as MYC were more prevalent in mCRPC (Supplementary Fig. S8A; red symbols). Nevertheless, PTEN and RB1 DEL represented among the most prevalent alterations in early-stage (GS6) PCa (Supplementary Fig. S8B–C). Analysis of genetic alterations across the spectrum of PCa progression revealed that among the 3 tumor suppressors (RB1, PTEN and TP53), while homozygous RB1 DEL remained relatively constant at 13%, 6%, 10% and 10% in low-grade (GS6), high-grade (GS9/10), treated Pri-PCa and mCRPC, respectively, PTEN HOMDEL (9%, 23%, 30% and 26%) and TP53 MUT (0%, 21%, 27% and 37%) continued to increase along this progression (Supplementary Fig. S8B–C). Among the 3 oncogenic events analyzed, while high levels of TMPRSS2 fusions (~40%) remained unchanged from low-grade to high-grade Pri-PCa to mCRPC, MYC (2%, 12%, 19% and 24%) and AR (0%, 1%, 6% and 49%) AMP continued to increase along the progression trajectory (Supplementary Fig. S8B–C). In fact, prevalent AR MUT (11%) were observed only in the most aggressive and Stemness-highest mCRPC (Supplementary Fig. S8B–C), suggesting potentially high noncanonical AR-A in mCRPC. Analyzing these major genomic alterations in association with PCa Stemness, we observed that RB1 DEL, FOXA1 (together with SPOP, IDH1, and CHD1) MUT, and especially MYC AMP were significantly associated with increased Stemness in Pri-PCa (Supplementary Fig. S8D). In mCRPC, only RB1 DEL exhibited a statistically significant association with higher Stemness (Supplementary Fig. S8E).

MYC further drives Stemness during spontaneous PCa progression.

AR AMP and MUT are exclusively treatment-induced as there were virtually no AR genomic alterations in treatment-naïve Pri-PCa (Supplementary Fig. S8B–C). We showed that early during PCa development increased AR-A may drive elevated Stemness (Fig. 1D–F); however, treatment-naïve Pri-PCa displayed continually increasing Stemness but decreasing AR-A (Fig. 2). What could be driving persistently high Stemness during Pri-PCa progression with decreasing AR-A and in the absence of AR genomic alterations? We found that Stemness-high Pri-PCa, compared to Stemness-low Pri-PCa, were enriched in MSigDB Hallmark MYC_TARGETS (MYC_Targets_V1 and MYC_Targets_V2; Fig. 5F) and had more prevalent MYC AMP (Fig. 6D and 6F). High-grade (GS9/10) Pri-PCa also had more MYC AMP events than low-grade (GS6) Pri-PCa (12% vs. 2%; P=0.006, χ² test; Supplementary Fig. S8C). Importantly, MYC AMP was most prominently associated with increased Stemness in Pri-PCa (Supplementary Fig. S8D). These observations led us to hypothesize that MYC may represent a critical Stemness driver during spontaneous Pri-PCa progression (in absence of therapeutic pressure). To test this hypothesis, we quantified MYC signaling activity⁵¹ (MYC-sig; See Methods and Supplementary Table S2 for gene signature information) across the PCa progression spectrum. We observed that: 1) both MYC mRNA and MYC-sig increased in Pri-PCa compared to matched benign samples (Fig. 7A–B); 2) MYC-sig continued to go up during Pri-PCa progression (Fig. 7C); 3) unlike nADT-induced concomitant decline in both AR-A and Stemness (Fig. 1D–F; Supplementary Fig. S4), nADT showed no discernible effect on MYC-sig (Supplementary Fig. S9A–D); and 4) MYC-sig continued to increase across the disease spectrum with mCRPC exhibiting the highest MYC activity (Fig. 7D–E) and with MYC-sig positively correlated with the Stemness (Fig. 7F–H).

Figure 7. MYC activation contributes to increased Stemness during spontaneous PCa progression.

A-B, Pairwise comparisons showing increased MYC mRNA levels and MYC activity (MYC-sig) in primary tumors (T) versus matched adjacent benign tissues (N) in TCGA-PRAD cohort (n=51; A) and Wyatt (n=12; B) cohorts. Data is shown as mean ± SD; paired samples are linked by grey lines. Significance was determined using two-tailed paired Student’s t-test (****, P < 0.0001).

C-H, MYC activity increases during PCa progression, peaking in mCRPC. In treatment-naïve Pri-PCa (C), MYC-sig is significantly increased in all-grade (GS) tumors as compared to tumor-adjacent benign tissues (N) with the most advanced GS9/10 tumors exhibiting the highest MYC-sig. In D, MYC-sig increases from normal to treatment-naïve Pri-PCa and further in advanced Pri-PCa treated with adjuvant ADT (Tx) in TCGA-PRAD, while mCRPC shows the highest MYC-sig in both SU2C 2015 (D) and Taylor 2010 (E) cohorts. Data is presented as mean ± SD. Statistical significance was assessed by one-way ANOVA with Tukey’s multiple-comparison test (C-E). J-T trend test was used in (C, D) to assess the significance of progressive increase. **, P < 0.01; ***, P < 0.001; ****, P < 0.0001; ns, not significant. F-H, Scatter plots showing positive correlations between MYC activity and Stemness in the Taylor 2010 cohort across different stages. Pearson’s r and P-values are indicated.

These results, together, support MYC as a Stemness driver during PCa progression.

DISCUSSION

The current project was undertaken to quantitatively co-analyze the global changes of and the interrelationship between cancer stemness and pro-differentiation AR signaling activity during PCa evolution. By employing gene expression-based general Stemness and AR-A signature scores and through developing a new 12-gene PCa-Stem signature, we show that while the differentiation-regulating AR-A steadily declines in the PCa continuum, pervasively increasing Stemness quantitatively measures PCa aggressiveness, plasticity and progression, and represents a poor prognosticator for patient outcomes. We report that Stemness-high PCa possess unique genomic features with high TMBs and are associated with aggressive molecular subtypes such as PAM50-LumB^42–44 and PCS1^45,46. We further present evidence that elevated Stemness in early PCa might be driven by increased AR-A but increased MYC-sig activity and other genomic alterations likely propagate the exponentially increased Stemness in mCRPC where AR-A is suppressed.

Early studies reported that a cohort of AR-regulated genes declines in high-grade versus low-grade Pri-PCa^{52, 53} and that loss of AR expression or inhibition of AR-A promotes a stem-like phenotype^{8, 34, 54, 55}. However, the global pro-differentiation AR-A and cancer stemness have not been systematically and quantitatively assessed and co-analyzed in PCa, and it’s unclear whether PCa progression is accompanied by increased stemness as in some other cancers such as glioblastoma and breast cancer⁶. Starting with systematic, comprehensive and intentionally “progression timeline-informed” collection of clinical datasets, here we integrate 26 transcriptomic datasets (covering 87,183 samples) encompassing normal prostate, benign, primary tumors, ADT-treated tumors, CRPC subtypes, metastases, and model systems. This enables a “progression-conscious” and clinically anchored stratification framework with clearly annotated clinical information across these 26 cohorts. Our pan-cohort approach enables signature validation in rare archival biopsy samples and cross-platform integration — a translational advantage not addressed by other studies. With these novel and unique dataset collection and curation and data analysis pipelines, our study demonstrates that PCa progression is accompanied by a continual increase in Stemness (score) but loss of canonical AR-A, with the two inversely correlated with each other (Supplementary Fig. S10). Importantly, the Stemness tracks treatment-induced plasticity and PCa aggressiveness, and both high global Stemness and high PCa-Stem signature prognosticate poor patient survival. Consequently, the stemness, a previously qualitative and somewhat ambiguous term that now permeates oncology field and cancer research publications, can be quantitatively measured and, in the case of PCa, the Stemness score quantitatively and positively tracks the entire spectrum of PCa development, lineage plasticity, therapy resistance and metastatic progression. To our knowledge, this is also the first comprehensive study to simultaneously quantify the oncogenic stemness and pro-differentiation AR-A across the PCa continuum.

The normal prostatic acini and ducts are lined by AR⁺ secretory luminal cells and AR⁻ basal cells together with rare neuroendocrine cells. We find that the NHP luminal cells have higher AR-A compared to basal cells, confirming AR-A as a metric for quantifying canonical AR activity and prostatic epithelial differentiation. On the other hand, basal cells have higher Stemness, consistent with earlier findings from our lab¹⁹ and others^{20, 21, 22, 56} showing the basal cell compartment harbors stem cells. Oncogenic transformation is known to confer stem-like features in the initiated cells. Indeed, compared to adjacent benign tissue, Pri-PCa exhibit elevated Stemness, which is associated with and likely driven by increased AR-A. Loss of PTEN and RB1 (9% and 13%, respectively, in GS6 tumors) could also potentially contribute to increased Stemness in early prostate tumors.

During our studies, we made a surprising finding that tumor-adjacent benign tissues, the so-called ‘normal’ tissues used in most comparative studies, actually display much higher Stemness than the true normal prostate obtained from cancer-free organ donors, suggesting that these benign tissues are not normal and their transcriptomes have apparently skewed towards those of cancer samples. This important new revelation prompted us to use, in relevant subsequent studies, the GTEx-derived normal prostate tissue as the baseline comparator, which obviously strengthens the normalization standard across cohorts, improves cross-stage comparation and interpretation, and enhances the scientific rigor of our data.

In treatment-naïve Pri-PCa, tumor progression is accompanied by decreasing AR-A but persistently high Stemness (Supplementary Fig. S10). The decreasing AR-A is driven by progression-related loss of differentiation while increasing Stemness is likely driven by MYC amplification and increased MYC activity with contributions from other oncogenic evens such as loss of PTEN and RB1. Long-term ADT/ARPI treatment, as expected, further attenuates AR-A but substantially upregulates the Stemness such that mCRPC have the highest Stemness, which is probably driven by continually increasing genomic alterations such as MYC amplification and loss of tumor suppressors. Notably, as AR alterations (AMP/MUT) are exclusively induced by ARPIs, noncanonical AR-A (i.e., castration-resistant AR activity) likely make significant contributions to the increased Stemness in mCRPC.

What biological correlates do our Stemness metrics measure? First, it measures transcriptional features and relative abundance of stem cells and CSCs. Therefore, the NHP basal/stem cells display higher Stemness than luminal cells (consistent with literature^{19, 20, 21, 22}); ARPI treatment, while shutting down AR signaling, enriches CSCs^{7, 34, 54, 55} and upregulates Stemness (this study); and PCa cell lines, xenografts and PDX, known to be clonally derived from cancer stem/progenitor cells, demonstrate the highest Stemness. Second, it measures cellular and lineage plasticity induced by treatment and genetic alterations. Thus, mCRPC, which harbor more abundant treatment-reprogrammed AR^−/lo cells and PCa cells with stem-like features or in an AR-indifferent state^{8, 12, 13, 14, 16, 33, 47, 48}, manifests the highest Stemness. Our analysis further substantiates the aggressive nature of (AR^−/lo) CRPC-NEPC and DNPC compared to ARPC, as evidenced by their elevated PCa-Stem signature scores. This suggests a shift in the phenotypic landscape of mCRPC towards more stem-like and aggressive forms, particularly under the pressure of advanced therapies. Likewise, aggressive DKO and TKO murine prostate tumors have undergone lineage transformation and also exhibit higher Stemness than indolent SKO tumors. Third, the Stemness metric likely reports the high proliferative status of advanced PCa and mCRPC, as Stemness-high tumors in both Pri-PCa and mCRPC are enriched in “Cell-cycle progression” signature and most genes in the 12-gene PCa-Stem Signature are involved in mitosis and cell-cycle progression. Finally, the Stemness score measures PCa aggressiveness as high Stemness correlates not only with more advanced and aggressive stages of PCa but also with aggressive PAM50-LumB and PCS1 molecular subtypes.

Of clinical significance, both the 12-gene PCa-Stem signature and the global Stemness score correlate with poor patient survival, suggesting the possibility of developing the Stemness metric into a predictive biomarker to distinguish aggressive from indolent primary tumors and a prognostic indicator of unfavorable clinical outcomes in mCRPC patients. Notably, 11 of the 12 genes (except KLK12) in the newly developed PCa-Stem signature are significantly correlated with PCa progression (Supplementary Fig. S7C) implicating their individual contributions to the increased Stemness in advanced Pri-PCa and mCRPC. Analysis using data from The Human Protein Atlas (www.proteinatlas.org) reveals that these 11 genes are also recognized as unfavorable prognostic markers in PCa and several other cancers (Liu X and Tang, unpublished), raising the potential utility of the PCa-Stem signature as a prognostic marker not only for PCa but also for other malignancies. Coupled with earlier studies linking increased Stemness with aggressive breast and brain cancer subtypes⁶, our findings herein underscore the increased Stemness as a ‘universal’ and fundamental characteristic of cancer progression and aggressiveness.

In recent years, several studies^27,57–60 have proposed stemness-based classifications or prognostic models using either single-cell RNA-seq (scRNA-seq), bulk RNA-seq or integrative analyses. These studies reinforce the biological relevance of stemness and highlight oncogenic pathways such as MYC activation, RB1/PTEN loss, and lineage plasticity in PCa progression. However, to our knowledge, none has established a quantitative, clinically validated stemness metric across a pan-cohort framework as comprehensive as ours. Of note, our 12-gene PCa-Stem signature, derived from bulk transcriptomic stratification, shows strong biological overlap with the lineage plasticity signature (LPSig) reported by Zhao et al. using scRNA-seq⁵⁹. Among the shared genes, HMMR emerges as a critical node of convergence as it has been shown that HMMR promotes NE differentiation and aggressive behavior in CRPC models⁵⁹. The presence of HMMR, along with DLGAP5, AURKB, and BIRC5 in our PCa-Stem signature, highlights the translational potential of this gene set in targeting NEPC and other plasticity-driven CRPC subtypes. Supporting this, Zheng et al. used multi-omics integration and machine learning to define three stemness-based PCa subtypes and constructed a nine-gene prognostic signature⁵⁷ that also included HMMR and DLGAP5. KLK12, another PCa-Stem gene, was recently identified as a unique marker for SOX9^High stem-like luminal epithelial cells enriched following androgen deprivation⁶¹. Additionally, our findings that MYC amplification, RB1 deletion, and PTEN loss are enriched in Stemness-high tumors are consistent with the previously reported chromatin accessibility (ATAC-seq) patterns⁴⁸ and multi-omics spatial data⁶¹. It is worth noting that although our PCa-Stem signature was initially designed for risk stratification in mCRPC, the analytical framework is inherently flexible as the same platform can be extended to derive subtype-specific signatures (e.g., PCa-Stem-v2) tailored for distinct clinical phenotypes such as DNPC, ARPC, or NEPC. This scalability provides a foundation for personalized prognostication and therapy selection in PCa.

We recognize that the Stemness algorithm utilized in our studies was not developed specifically for PCa and therefore, its sensitivity to distinguish subtle intra-group differences remains to be fully validated. The bulk RNA-seq data used to derive Stemness and AR-A was not from pure epithelial cancer cells, a technical issue that can be potentially mitigated by scRNA-seq analysis and other Stemness pipelines such as the CytoTRACE^{62, 63} and by using cancer-specific genomic features derived, for example, from cell-free DNA⁶⁴. Finally, it will be important to determine whether and how noncanonical (castration-associated) AR signaling (driven by AR AMP/MUT) may contribute to the highest Stemness in mCRPC.

In conclusion, we have developed transcriptome-based gene signature scores to assess pro-differentiation AR activity (AR-A), MYC activity (MYC-sig) and oncogenic stem-cell features (general Stemness score and the 12-gene PCa-Stem signature). These scores enable us to globally profile differentiation and aggressiveness status across the PCa progression spectrum. This work advances our understanding of dynamic evolution of AR-A and cancer stemness during PCa progression and therapy resistance, and the signature scores reported herein have the potential to serve as clinical classifiers and aid in the identification of patients who may benefit from stemness-targeting therapies.

METHODS

PCa clinical datasets and data collection

In this study, we utilized a total of 26 published and in-house prostate cancer (PCa) related datasets, including various cohorts encompassing the evolutionary spectrum of PCa (Supplementary Fig. S1A–B): normal prostate, tumor-adjacent benign tissue, treatment-naïve primary PCa (Pri-PCa) with increasing GS, PCa treated with neo-adjuvant ADT (nADT) or long-term ADT, metastatic castration-resistant PCa (mCRPC), and PCa cell lines, PCa patient-derived xenografts (PDX) and xenografts. Clinical information, genomic data and gene expression profiling data were downloaded and detailed information on each dataset is listed in Supplementary Table S1. Dataset descriptions include dataset source references, PubMed IDs (PMIDs; RRID:SCR_004846), DOI links, sample type and source (e.g., normal prostate tissue, adjacent benign prostate tissue or primary culture from PCa patient, primary or metastatic prostate tumor, xenografts, cell lines), sample sizes, data types used in the current study (clinicopathologic data, RNAseq, microarray, whole exome sequencing [WES]), and public data access portals.

Data Access and Repository Information:

Data for this study were retrieved from the following public repositories and databases (persistent identifiers where applicable):

1. GTEx Portal (RRID:SCR_013042; https://gtexportal.org/)

2. Gene Expression Omnibus (GEO; RRID:SCR_005012; https://www.ncbi.nlm.nih.gov/geo/)

3. cBioPortal for Cancer Genomics (RRID:SCR_014555; https://www.cbioportal.org/)

4. Zenodo Repository (RRID:SCR_004129; https://zenodo.org/)

5. Decipher GRID Database (RRID:SCR_006552; ClinicalTrials.gov identifier: NCT02609269)

6. European Genome-phenome Archive (EGA; RRID:SCR_004944; https://ega-archive.org/)

7. European Nucleotide Archive (ENA; RRID:SCR_006515; https://www.ebi.ac.uk/ena/browser/home)

8. UCSC Xena Functional Genomics Portal (RRID:SCR_018938; https://xenabrowser.net/)

Accession numbers and portal links for the 26 datasets described below are summarized in Supplementary Table S1.

The 26 datasets are briefly described below:

• 1GTEx 2013 (PMID: 23715323)²⁹: The Genotype-Tissue Expression (GTEx) Project repository collected high-throughput and clinical data of normal tissue from many organs. RNAseq data of 245 normal prostate tissue samples was downloaded from the GTEx portal (RRID:SCR_013042; ref. 29; DOI: 10.1038/ng.2653; https://www.gtexportal.org/, version: v8).
• 2Smith 2015 (PMID: 26460041)²⁰: RNAseq data of 5 prostate basal cell samples (Trop2⁺CD49f^hi) and corresponding 5 luminal cell samples (Trop2⁺CD49f^lo) FACS-purified from benign human prostate tissue in 5 PCa patients (n=5) was retrieved from the GEO database GSE82071 (ref. 20; DOI: 10.1073/pnas.1518007112).
• 3Liu 2016 (PMID: 27926864)⁵⁶: RNAseq data of 3 FACS-purified human prostate basal cell samples (CD45⁻EpCAM⁺CD49f^hiCD38^lo) with corresponding luminal (CD45⁻EpCAM⁺CD49f^loCD38^hi) and luminal progenitor (CD45⁻EpCAM⁺CD49f^loCD38^lo) cell populations purified from the benign prostate tissues of 3 PCa patients (n=3) was retrieved from the supplementary data from Liu et al., 2016 (ref. 56). The microarray-based transcriptomic profiles of the same samples were retrieved from the GEO database GSE89050 (ref. 56; DOI: 10.1016/j.celrep.2016.11.010).
• 4Zhang 2016 (PMID: 26924072)¹⁹: RNAseq data of 3 human prostate basal cell (Trop2⁺CD49f^hi) with the corresponding 3 luminal cell (Trop2⁺CD49f^lo) samples FACS-purified from benign prostate tissues of 3 PCa patients (n=3) was retrieved from the GEO database GSE67070 (ref. 19; DOI: 10.1038/ncomms10798).
• 5TCGA PRAD 2015 and Pan-Cancer Atlas 2018: The Cancer Genome Atlas (TCGA) (https://www.cancer.gov/tcga) repository collected high-throughput molecular and clinical data from primary cancer and matched normal/benign samples spanning 33 human cancer types, including prostate adenocarcinoma (PRAD). For each cancer type, TCGA published an initial “marker papers” summarizing analyses performed. Supplemental and associated data files supporting these “marker papers” are available through the NCI Genomic Data Commons (GDC) website (https://gdc.cancer.gov/about-data/publications). For prostate cancer, TCGA PRAD 2015 (PMID: 26544944; ref. 18; DOI: 10.1016/j.cell.2015.10.025) is the marker paper on PCa published in 2015, which includes the dataset of 333 primary prostate adenocarcinomas (281 treatment-naïve Pri-PCa, 49 PCa with post-surgery adjuvant ADT, and 52 matched normal/benign prostate tissue). Genomic data were obtained from cBioPortal for Cancer Genomics (RRID:SCR_014555; https://www.cbioportal.org), specifically the TCGA-PRAD Firehose Legacy study (https://www.cbioportal.org/study/summary?id=prad_tcga). RNAseq expression data was obtained from Zhang et al., 2020 (PMID: 32350277; ref. 16; DOI: 10.1038/s41467-020-15815-7).

  The expanded TCGA PRAD 2018 (Pan-Cancer Altas) cohort includes 494 patients with Pri-PCa (422 treatment-naïve Pri-PCa, 70 PCa with post-surgery adjuvant hormone treatment, 2 PCa with chemotherapy) and 52 matched adjacent normal/benign prostate tissue. Clinicopathologic features (age, Gleason scores, PSA levels, tumor stages) and genomic alterations data (mutation frequency, copy number alterations [CNAs], structural variants [SVs], tumor mutation burden [TMB], fraction genome altered) were retrieved from cBioPortal for Cancer Genomics PRAD PanCancer Atlas study (https://www.cbioportal.org/study/summary?id=prad_tcga_pan_can_atlas_2018) and from the Pan-Cancer Atlas companion site (welcome to the Pan-Cancer Atlas webpage: https://www.cell.com/pb-assets/consortium/pancanceratlas/pancani3/index.html). Survival data were obtained from Liu et al., 2018⁶⁵ (PMID: 29625055; DOI: 10.1016/j.cell.2018.02.052). Normalized RNAseq data were downloaded via NCI GDC hub at UCSC Xena Functional Genomics Portal (https://gdc.xenahubs.net, version 2019-07-20; DOI: 10.1038/s41587-020-0546-8).

• 6Rajan 2014 (PMID: 24054872)²⁵: Clinicopathologic (Gleason scores) and RNAseq data of 7 matched pre- and post-nADT PCa samples were obtained, respectively, from Table 1 in Rajan et al., 2014 (ref. 25; DOI: 10.1016/j.eururo.2013.08.011) and the GEO database GSE48403.
• 7Sharma 2018 (PMID: 30314329)²⁶: Clinicopathologic (Gleason scores) and RNAseq data of 7 matched pre- and post-nADT PCa samples from the responder group (also defined as “Low Impact Group” in the source reference) was obtained, respectively, from supplementary Table 1 in Sharma et al., 2018 (ref. 26; DOI: 10.3390/cancers10100379) and the GEO database GSE111177.
• 8Long 2020 (PMID: 32951005; DOI: 10.1016/j.eururo.2013.08.011)²⁴: RNAseq data of 6 matched pre- and post-nADT PCa samples was obtained from the GEO database GSE82071. Additionally, RNAseq data from 10 locally advanced PCa samples along with their paired adjacent benign prostate tissues were obtained from the GEO database GSE82071.
• 9Nastiuk (RPCI) 2020 (submitted)⁶⁶: Clinicopathologic data (Gleason scores) and RNAseq data of 43 Pri-PCa samples from post-nADT PCa patients and 43 tumor samples from age-, stage-, clinically-matched PCa patients without nADT was obtained from Jamroze et al., 2024 (work led by Drs. K. Nastiuk and G. Chatta at the Roswell Park Comprehensive Cancer Center). RNAseq data and clinicopathologic information from the Nastiuk 2020 cohort (unpublished) were provided in-house by Dr. K. Nastiuk and are available from the corresponding author upon reasonable request.
• 10Gerhauser 2018 (PMID: 30537516; DOI: 10.1016/j.ccell.2018.10.016)⁶⁷: RNAseq data of 118 primary tumor samples and 9 adjacent benign prostate tissue from early-onset PCa patients was obtained from the European Genome-Phenome Archive (EGA) (https://ega-archive.org), accession number EGAS00001002923.
• 11Wyatt 2014 (PMID: 25155515; DOI: 10.1186/s13059-014-0426-y)²³: Clinicopathologic (Gleason scores) and RNAseq data of 12 primary PCa samples from treatment-naïve PCa patients and 12 adjacent benign prostate tissue were obtained from the European Nucleotide Archive (ENA) (https://www.ebi.ac.uk/ena/browser/home), accession number PRJEB6530.
• 12Spratt 2017 (PMID: 28358655; DOI: 10.1200/JCO.2016.70.2811)³⁸: Transcriptomic profiles (Microarray) and clinicopathologic data of 855 radical prostatectomy specimens from a multi-institutional study of intermediate- and high-risk localized PCa patients were obtained from Veracyte GRID (https://decipherbio.com/grid/).
• 13Tosoian 2020 (PMID: 32231245; DOI: 10.1038/s41391-020-0226-2)³⁹: Transcriptomic profiles (Microarray) and clinicopathologic data of 405 radical prostatectomy and biopsy specimens from high-risk localized PCa patients were obtained from Veracyte GRID (https://decipherbio.com/grid/).
• 14CHAARTED correlatives 2021 (PMID: 34129855; DOI: 10.1016/j.annonc.2021.06.003)⁴⁰: Transcriptomic profiles (Microarray) and clinicopathologic data of 160 biopsy specimens from a Phase 3 trial of metastatic hormone-sensitive PCa patients were obtained from the NCTN Data Archive (https://nctn-data-archive.nci.nih.gov), accession number NCT00309985-D14.
• 15Decipher GRID Biopsy (PMID: 37060201; DOI: 10.1002/cncr.34790)³⁵: Transcriptomic profiles from 82,470 prospectively collected prostate biopsy samples were retrieved from the Decipher GRID database (RRID: SCR_006552; https://decipherbio.com/grid; ClinicalTrials.gov identifier: NCT02609269; https://clinicaltrials.gov/ct2/show/NCT02609269). Patient data were de-identified from clinical use of the Decipher prostate genomic classifier in accordance with the Safe Harbor method described in the HIPAA Privacy Rule 45 CFR 164.514(b) and (c) (Veracyte, San Diego, CA). The samples, comprising a large cohort of localized prostate cancer (PCa), were utilized to compare the distribution of Stemness Index values across National Comprehensive Cancer Network (NCCN) risk groups.

Clinicopathologic Breakdown and Risk Group Analysis:

Detailed clinicopathologic review revealed that the majority of the cohort presents with low-grade localized PCa. The TNM staging breakdown includes:

• T1 stage: Approximately 50% of samples, indicating early-stage cancer localized within the prostate.

• T2 stage: Approximately 8%, representing cancer confined within the prostate but more extensive than T1.

• T3 and T4 stages: Less than 1%, suggesting advanced local spread.

• N0: Approximately 98%, indicating no regional lymph node involvement.

• N1: Approximately 2%, suggesting regional lymph node involvement.

The Decipher GRID database (RRID:SCR_006552) categorizes patients into NCCN risk groups based on TNM stage, PSA level, and Gleason score:

• Very Low and Low Risk: T1–T2a, N0, M0, PSA <10 ng/mL, and Gleason score ≤6; often managed with active surveillance or less aggressive treatment.

• Intermediate Risk – Favorable: Typically T2b–T2c, N0, M0, PSA 10–20 ng/mL, and Gleason score 7; managed with less aggressive treatment but closer monitoring.

• Intermediate Risk – Unfavorable: Similar clinical features to favorable but with higher disease burden or adverse factors requiring careful monitoring.

• High Risk: Usually involves T3a, N0, M0, PSA >20 ng/mL, or Gleason score 8–10; often managed aggressively with surgery, radiation, and/or hormonal therapy.

• Very High Risk: Typically includes T3b–T4, any N, M0, and high Gleason scores or PSA levels (PSA >20 ng/mL); patients are at significant risk for metastatic progression and are treated aggressively, often with multimodal therapies.

Observations on Disease Severity and Treatment Naïvety:

Analysis of the cohort indicates that only approximately 13% of cases fall into the High and Very High risk categories (8,376 + 2,399 = 10,775 samples), suggesting that the majority (approximately 87%) are below Intermediate Risk (≤GS7). This distribution implies that most patients were treatment-naïve and likely androgen deprivation therapy (ADT)-free due to less aggressive disease status.

• 16Bolis 2021 (PMID: 34857732; DOI: 10.1038/s41467-021-26840-5)²⁷: RNAseq and clinicopathologic data of 1,223 clinical samples from an integrated cohort consisting of normal prostate specimens (n = 174), Pri-PCa (n = 714) and mCRPC (n = 335) was obtained from the Zenodo repository (Record ID: 5546618; https://zenodo.org/records/5546618). This resource of Prostate Cancer Transcriptome Atlas (https://prostatecanceratlas.org) was built and integrated from the following studies/datasets: (1) Genotype-Tissue Expression Database (GTEx; PMID: 23715323; ref. 29); (2) The Cancer Genome Atlas (TCGA, TCGA-PRAD; PMID: 26544944; ref. 18); (3) Atlas of RNA-sequencing profiles of normal human tissues⁶⁸ (GSE120795; PMID: 31015567; ref. 68); (4) Integrative epigenetic taxonomy of PNPCa⁶⁹ (GSE120741; PMID: 30464211; ref. 69); (5) Prognostic markers in locally advanced lymph node-negative prostate cancer (PRJNA477449); (6) The long noncoding RNA landscape of NEPC and its clinical implications⁷⁰ (PRJEB21092; PMID: 29757368; ref. 70); (7) Integrative clinical sequencing analysis of metastatic CRPC reveals a high frequency of clinical actionability (PRJNA283922; dbGaP: phs000915; PMID: 26000489; ref. 10); (8) CSER—exploring precision cancer medicine for sarcoma and rare cancers⁷¹ (PRJNA223419; dbGaP: phs000673; PMID: 28783718; ref. 71); (9) Molecular basis of NEPC (Beltran 2016; PRJNA282856; dbGaP: phs000909; PMID: 26855148; ref. 28); (10) Heterogeneity of androgen receptor splice variant-7 (AR-V7) protein expression and response to therapy in CRPC⁷² (GSE118435; PMID: 30334814; ref. 72); (11) RNAseq of human PCa cell lines and mCRPC tumor^{73, 74, 75} (GSE14750; PMIDs: 32460015; ref. 73; 33658518; ref. 74; 34244513; ref. 75 | GSE171729; PMID: 34244513; ref. 75); and (12) Molecular profiling stratifies diverse phenotypes of treatment-refractory metastatic CRPC (PRJNA520923; GEO: GSE126078; PMID: 31361600; ref. 33).
• 17Taylor 2010 (PMID: 20579941; DOI: 10.1016/j.ccr.2010.05.026)³⁶: Transcriptomic profiles (Microarray) and clinicopathologic (recurrence-free survival) data of 29 adjacent benign prostate tissue, 131 Pri-PCa, and 19 mPCa were obtained from the GEO database (GSE21034) and the cBioPortal for Cancer Genomics PRAD MSKCC 2010 study (https://www.cbioportal.org/study/summary?id=prad_mskcc).
• 18SU2C 2015 (PMID: 26000489; DOI: 10.1016/j.cell.2015.05.001)¹⁰: Clinical and genomic data of 150 mCRPC samples were obtained from the cBioPortal for Cancer Genomics PRAD SU2C 2015 study (https://www.cbioportal.org/study/summary?id=prad_su2c_2015). RNAseq data of 98 mCRPC samples was retrieved from Zhang et al., 2020¹⁶ (PMID: 32350277; DOI: 10.1038/s41467-020-15815-7; ref. 16).
• 19SU2C 2019 (PMID: 31061129; DOI: 10.1073/pnas.1902651116)⁴¹: This dataset includes clinicopathologic and genomic data for 444 mCRPC samples. RNAseq data of 266 mCRPC samples (prepared by using poly(A) enrichment for RNAseq library construction) were obtained from the cBioPortal for Cancer Genomics PRAD SU2C 2019 study (https://www.cbioportal.org/study/summary?id=prad_su2c_2019). In parallel with TCGA-PRAD, clinicopathologic features (age, Gleason scores, PSA levels), genomic alterations (mutation frequency, CNAs, SVs, TMB, fraction genome altered), and patient outcome data (overall survival) were also retrieved from cBioPortal (RRID:SCR_014555). These datasets were used for comparative analysis between Stemness-high and Stemness-low groups, and to explore genomic drivers and clinical associations as described in the Methods.
• 20Labrecque 2019 (PMID: 31361600; DOI: 10.1172/JCI128212)³³: RNAseq data of 98 mCRPC samples and 39 PDX-LuCaP was obtained from the GEO database GSE126078.
• 21Alumkal 2020 (PMID: 32424106; DOI: 10.1073/pnas.1922207117)¹³: RNAseq data of 25 mCRPC samples before treatment with enzalutamide (ENZA) was obtained from Supplementary dataset S01 of Alumkal et al., 2020 (ref. 13).
• 22Westbrook 2022 (PMID: 36109521; DOI: 10.1038/s41467-022-32701-6)¹²: RNAseq data of 42 mCRPC samples consisting of 21 matched pre/post Enzalutamide treatment samples was retrieved from Westbrook et al., 2022 (ref. 12).
• 23Beltran 2016 (PMID: 26855148; DOI: 10.1038/nm.4045)²⁸: RNAseq data of mCRPC samples, including 34 castration-resistant adenocarcinoma (mCRPC-Adeno) and 15 neuroendocrine histologies (mCRPC-NEPC) samples, was obtained from cBioPortal for the Cancer Genomics (https://www.cbioportal.org/study/summary?id=nepc_wcm_2016).
• 24Goodrich 2017 (PMID: 28059767; DOI: 10.1126/science.aah4199)³⁷: RNAseq data of murine prostate tumors from genetically engineering mouse models (GEMMs) of PCa of different genotypes, including 4 SKO (single knockout: PBCre4:Pten^f/f), 13 DKO (double knockout PBCre4:Pten^f/f:Rb1^f/f), and 6 TKO (PBCre4:Pten^f/f:Rb1^f/f:Trp53^f/f) was obtained from the GEO database GSE90891 (f = a floxed allele of the indicated genes).
• 25CCLE 2018 (PMID: 22460905; DOI: 10.1038/nature11003)³¹: RNAseq data of 1,019 human cell lines were obtained from the Xena Functional Genomics Portal⁷⁶ (https://xenabrowser.net, version: 2018-05-30; ref. 76). Among them, eight prostate cancer cell lines — DU145 (RRID:CVCL_0105), LNCaP clone FGC (RRID:CVCL_1379), MDA PCa 2b (RRID:CVCL_4745), NCI-H660 (RRID:CVCL_0459), PC-3 (RRID:CVCL_0035), VCaP (RRID:CVCL_2235), 22Rv1 (RRID:CVCL_1045), and PRECLH (PrEC LH; RRID:CVCL_V626) — were included for analysis.
• 26XG-LNCaP and XG-LAPC9 (PMID: 30190514; DOI: 10.1038/s41467-018-06067-7)³⁴: RNAseq data of 12 LNCaP xenografts (XG-LNCaP) and 10 LAPC9 xenografts (XG-LAPC9) was obtained from the GEO database GSE88752.

Sex as a biological variable

All clinical datasets analyzed in this study consisted exclusively of male patients, consistent with the male-specific nature of PCa. Preclinical models, including xenografts, patient-derived xenografts, and genetically engineered mouse models (GEMMs), were also male-derived. All PCa cell lines utilized in the study were of male origin, with the exception of Supplementary Figure S5A, where a broader cohort of 1,019 cell lines from the Cancer Cell Line Encyclopedia (CCLE), comprising both male- and female-derived samples across multiple cancer types, was analyzed. Analyses of CCLE data in Supplementary Figure S5A were performed without stratification by sex.

Transcriptomic data normalization and transcriptome-based signature scores

To ensure consistency and compatibility across diverse datasets, we employed normalization strategies tailored to the specific data types and analytical approaches.

Canonical AR activity score (AR-A) and MYC activity score (MYC-sig):

Canonical pro-differentiation AR-A can be determined by RNA-seq based measurements of AR-regulated transcripts under intact androgen/AR signaling conditions. Similarly, MYC-sig can be determined by MYC-regulated transcripts. We employed Z-score normalization method⁷⁷ to determine AR-A and MYC-sig. Briefly, the numeric AR or MYC activity score was calculated based on a linear combination of expression values (Z-scores) of experimentally validated AR targets¹⁴ or MYC targets⁵¹ (Supplementary Table S2, GSEA Molecular Signature Database 2022.1 version, MSigDB). The target gene lists were detailed in Supplementary Table S2. The expression value for each AR or MYC target genes (Supplementary Table S2) was converted to Z-score by $\mathrm{Z}=(\mathrm{x}-\mathrm{\mu })/\mathrm{\sigma }$, where $\mathrm{x}$ is the expression value of the target gene in a specific sample, $\mathrm{\mu }$ is the mean and σ is the standard deviation across all samples of a gene. Finally, the combined Z-scores were summed across all genes to represent the expression score of canonical AR or MYC target genes. Analysis of gene expression signatures using Z-score normalization (AR-A and MYC-sig) was confined within individual datasets, obviating the need for cross-dataset comparisons.

To further minimize batch effects and ensure data integrity and comparability across PCa stages, normalized RNAseq data from a published integrated data framework, the Prostate Cancer Transcriptome Atlas²⁷ (https://prostatecanceratlas.org), was used. In this harmonized resource, raw RNAseq data with high-quality sequencing reads from different datasets were re-mapped to the human reference genome GRCh38, adjusted for library size, and normalized using the variance stabilizing transformation pipeline using DESeq2 (version 1.28.1; RRID:SCR_015687)²⁷.

Transcriptome-based quantitative measurement of oncogenic stemness (the Stemness Index):

We adapted mRNA-based Stemness Index (mRNAsi) to determine the degree of oncogenic dedifferentiation (Stemness)⁶. Briefly, we computed Spearman correlations between the stemness model’s weight vector and the transcriptome expression profile of the queried samples, assuming paired observations, at least ordinal scale data, and a monotonic relationship without significant outliers. Spearman correlation calculations were performed using the ‘stats’ package (version 4.3.3) in R (version 4.3.3; RRID:SCR_001905). The resulting correlation coefficients were then linearly transformed to scale the mRNAsi scores between 0 and 1 using the following formula: Stemness Index = (Spearman correlation + 0.2619)/0.4151. Spearman correlation is advocated for calculating the Stemness Index because it is more robust with respect to potential cross-dataset batch effects that may arise⁶.

Transcriptome-based determination of “PCa-Stem Signature”:

To derive the PCa-Stem Signature, we performed single-sample gene set enrichment analysis (ssGSEA) using the Gene Set Variation Analysis (GSVA)⁷⁸ package (version 1.50.5; RRID:SCR_021058) in R (version 4.3.3; RRID:SCR_001905). Briefly, to identify robust stemness-associated genes in PCa, we performed differential gene expression (DEG) analysis between the top 33% (Stemness-high) and bottom 33% (Stemness-low) of samples, stratified by Stemness scores in both treatment-naïve primary PCa (TCGA-PRAD) and metastatic CRPC (SU2C 2019) cohorts. Genes significantly upregulated in both cohorts (fold change ≥ 2, FDR < 0.05) were intersected to derive a 12-gene consensus signature—termed the PCa-Stem Signature (Supplementary Table S2). ssGSEA was then performed using the GSVA R package to calculate PCa-Stem scores across PCa datasets. This signature was validated in independent cohorts and assessed for clinical relevance and transcriptomic subtype enrichment, including associations with aggressive PCa subtypes and patient survival outcomes.

Integrative analysis of genomic alterations in primary PCa (TCGA) and mCRPC (SU2C)

To explore molecular and clinical features associated with Stemness, we performed integrative analysis of genomic, transcriptomic, clinicopathologic, and outcome data using cBioPortal for Cancer Genomics (RRID:SCR_014555), which provides interactive access to multidimensional cancer genomics datasets and integrated visualization tools⁷⁹. Specifically, we analyzed data from primary prostate cancer (Pri-PCa; TCGA-PRAD 2018, Pan-Cancer Atlas) and metastatic castration-resistant PCa (mCRPC; SU2C 2019) cohorts. Clinicopathologic variables (e.g., age, Gleason scores, PSA levels, tumor stage), genomic alterations (e.g., mutation frequency, TMB, fraction genome altered, CNAs, and SVs), and patient outcome data (e.g., overall survival, progression-free survival) were retrieved from cBioPortal (RRID:SCR_014555) as described in the “PCa clinical datasets and data collection” section. Transcriptomic data used for Stemness score calculation were obtained as described in the same section. For comparative analysis, mCRPC samples were stratified into top and bottom 33% by Stemness score. In TCGA, only treatment-naïve Pri-PCa samples were used for a similar 33% stratification, excluding 70 patients who had received adjuvant (post-surgery) hormone therapy to reflect spontaneous disease progression. Differential expression analysis was conducted using DESeq2⁸⁰ (version 1.42.1; RRID:SCR_015687), and gene-level and genome-wide alteration frequencies were assessed across stratified groups and disease stages (e.g., Pri-PCa vs. mCRPC) using OncoPrint visualizations, bar plots, and correlation analysis. Associations with clinical outcomes were evaluated as described in the “Statistical analysis” section.

Statistical analysis

Statistical analyses were performed using GraphPad Prism software (version 10.4.1 (532); RRID:SCR_002798) and R (version 4.3.3; RRID:SCR_001095).

For comparisons between two groups, independent sample or paired t-tests were used to compare group means, assuming normality and equal variance. For multi-group comparisons, Tukey’s range test was employed following one-way ANOVA (adjusted p-values listed in Supplementary Table S4). The Wilcoxon rank-sum test was applied for comparisons of altered genome fraction, TMB, or age between two groups. For differential expression analyses and comparisons involving multiple hypotheses (e.g., RNAseq-based gene comparisons), p-values were further adjusted using the false discovery rate (FDR) method to control for multiple testing.

Pearson’s correlation coefficient was used to evaluate linear associations between two continuous variables, assuming normally distributed paired observations. Linear regression modeling was applied to examine the relationship between two continuous variables, with the slope of the best-fit regression line reported where applicable. Spearman’s rank correlation was employed to assess monotonic relationships between continuous or ordinal variables, assuming paired observations and monotonic trends without significant outliers.

The Jonckheere-Terpstra (J-T) trend test was used to assess monotonic association between an ordinal variable and a continuous variable. Chi-square test or Fisher’s exact test was performed to evaluate the associations between categorical variables.

For the analysis of patient progression-free survival (PFS), recurrence-free survival (RFS) and overall survival (OS), the standard Kaplan-Meier method was used for the estimation of survival fractions and log-rank tests were used for group comparisons. Cox proportional hazards models were used for multivariable analysis and for assessing the prognostic significance of gene signatures (Stemness, Pca-Stem signature, etc). For the mRNA-based gene signatures (Stemness, PCa-Stem signature, AR-A), patients were stratified into high and low signature groups based on median split in cohorts including Spratt 2017 (Fig. 4C), Tosoian 2020 (Fig. 4C), CHAARTED 2021 (Supplementary Fig. S6C), Taylor 2010 (Supplementary Fig. S6C), TCGA-PRAD 2018 (Supplementary Fig. S6C), and SU2C 2019 (Supplementary Fig. S6C). For certain analyses, including SU2C 2019 (Supplementary Fig. S6C) and CHAARTED 2021 (Fig. 4C), patients were alternatively stratified by comparing the top 25% versus the bottom 75% of signature scores to assess whether patients with very high Stemness levels exhibited a higher risk of disease progression.

All statistical tests were two-sided, and P < 0.05 was considered statistically significant.

Data Availability:

All datasets analyzed in this study are publicly available from repositories including GEO (RRID:SCR_005012), cBioPortal (RRID:SCR_014555), the NCI GDC Portal, Xena Functional Genomics Portal (RRID:SCR_018938), EGA (RRID:SCR_004944), ENA (RRID:SCR_006515), and Zenodo (RRID:SCR_004129), as detailed in Supplementary Table S1.