Identification and validation of recurrence-associated endothelial cells using a 54-gene panelAs the first step in exploring the prognostic value of TECs in PCa, we estimated the abundance of TECs in primary tumors by calculating the expression levels of three classic endothelial markers, including PECAM1, ENG, and VWF, and investigated whether the TEC abundance is associated with tumor recurrence. In the TCGA-PRAD cohort, Kaplan–Meier analyzes demonstrated that tumors with higher expression levels of these markers had significantly higher recurrence rates (all log-rank P < 0.05; Fig. 1a). This observation is further supported by Cox regression analyzes, as tumors with a higher expression level for each of the three genes were at a significantly higher risk of developing recurrence (PECAM1 HR: 1.663, 95% CI: 1.098-2.518; ENG HR: 1.706, 95% CI: 1.126–2.584; VWF HR: 1.555, 95% CI: 1.023–2.363). A significant association of higher ENG or VWF level with PCa recurrence was also detected in the DKFZ-PRAD cohort (Fig. S1). These results suggested that increased TECs are associated with tumor recurrence in primary PCa.Fig. 1: Recurrence-associated endothelial cells were identified and validated using an established 54-gene panel in primary prostate cancer.a Higher mRNA levels of PECAM1, ENG, and VWF are associated with worse prognosis in prostate cancer, as determined by the Kaplan–Meier analysis in the TCGA-PRAD cohort (top tertile versus the low and median tertiles). P values were calculated using the log-rank test. b UMAP plot of 7 major cell types using scRNA-seq data of 12 primary prostate cancer samples from the Chen study [17]. c Dotplot of marker genes’ expression levels of the major cell types displayed in panel b. The size of a dot indicates the percentage of cells that express corresponding genes, whereas the color of a dot reflects a gene’s expression level. d Scissor algorithm-inferred endothelial cells (ECs) associated with recurrence in the TCGA-PRAD cohort. Blue dots represent recurrence-associated ECs (RAECs), associated with a worse prognosis, whereas yellow dots represent ECs associated with a better prognosis. Gray dots mark ECs that do not show an association with prognosis. e, f Vocalnol plots of differentially expressed genes between RAECs and non-RAECs (i.e., Scissor-ECs and background ECs) and those between ECs and non-ECs (i.e., other major cell types as shown in panel b) f Red dots indicate upregulated genes, whereas blue dots indicate downregulated genes. g Venn diagrams of 54 RAEC-related genes, as identified by their differential expression between RAECs and non-RAECs, higher expression levels in ECs than non-ECs, and lower expression levels in non-ECs than ECs. h A higher score of the RAEC 54-gene panel (top tertile versus the rest) is correlated with a shorter recurrence time in the TCGA-PRAD cohort as determined by Kaplan–Meier analysis. The P value was derived from the Log-rank test.TECs are highly heterogeneous, so it is necessary to identify which subpopulation(s) of TECs is responsible for TEC’s association with tumor recurrence in PCa. To address this question, we first analyzed the Chen scRNA-seq dataset, which contains 12 untreated primary PCa samples from radical prostatectomy. After quality control, 29,457 cells were analyzed, and all major cell types were annotated based on specific markers (Fig. 1b, c). A total of 2457 ECs were annotated according to the expression of PECAM1, VWF, and ENG. We then leveraged the Scissor algorithm to integrate these ECs’ transcriptomics to the TCGA-PRAD cohort and reliably identified (P < 0.05) a subset of ECs (n = 347) whose expression profiles are associated with PCa recurrence. This subpopulation was hereafter referred to as recurrence-associated ECs (RAECs) (Fig. 1d).We then applied differential gene analyzes to identify the genes whose expression profiles define RAECs and thus can be used to estimate RAEC abundance in bulk sequencing datasets. In total, 165 genes were upregulated and 122 downregulated in RAECs compared to the remaining ECs (non-RAECs) (Fig. 1e; Table S5); and 801 genes were upregulated in ECs compared to non-ECs (Fig. 1f; Table S6). Of those 801 genes, 163 were upregulated in both RAECs and ECs. To enhance the specificity of these genes to ECs, we used expression cutoffs to remove 109 genes that were also highly expressed in non-ECs, which left 54 RAEC-related genes (Fig. 1g).In the TCGA-PRAD cohort, expression levels of the 54 RAEC-related genes were summed, and Kaplan–Meier and Cox regression analyzes were performed. PCa with higher scores of the 54 genes showed a significantly worse disease-free survival (log-rank P = 0.008, HR: 1.741, 95% CI:1.150–2.635; Fig. 1h). These findings confirm the robustness of the Scissor selection.RAECs are characteristic of tip ECs and increased angiogenic activitiesTo further characterize RAECs, we performed sub-clustering and fine annotation of ECs (Fig. S2A, b) in the Chen dataset, in which ECs were classified based on functional states instead of biological cell types [17]. We identified 7 biological subsets of ECs, including arterial defined by FBLN5 and ENPP2; postcapillary vein (PCV) by ACKR1, and SELP; activated PCV by POSTN and CCL14; intermediate; immature by APLNR; and tip by ESM1 and APLN (Fig. S2C). Arteries, PCVs, and capillaries are different EC subtypes belonging to traditional vascular beds. Activated PCV is previously identified in lung cancer and choroid neovascularization, and is considered to be the EC subtype from which neovessels originate [19]. Immature cells resemble stalk-like cells, which elongate vessel sprouts whereas Tip cells guide and navigate vessel sprouts during neovascularization [18, 19]. Intermediate cells are considered a plastic phenotype possibly transitioning from activated PCV to angiogenic cells [18, 19]. RAECs consisted of all annotated EC subtypes (Fig. S2D), tip cells were most abundant (n = 78), followed by immature cells (n = 60) and intermediate cells (n = 58). Compared to non-RAECs, RAECs contained more tip cells (22.5% vs. 14%), immature cells (17.3% vs. 8.7%), and intermediate cells (16.7% vs. 5.8%, Fig. S2E).We then evaluated whether different EC subtypes have distinct gene expression profiles between RAECs and non-RAECs. Among the 5 subtypes of ECs, tip cells had the most DEGs (Fig. S2F). Specifically, tip cells in RAECs expressed higher levels of gene signatures associated with tip cell markers, migration, and extracellular matrix (ECM) modeling than their counterparts in non-RAECs (Fig. S2G). This finding suggests that a more differentiated state of tip cells plays the most important role in RAEC-associated PCa recurrence. GSEA demonstrated that RAECs expressed higher levels of tip cell markers, collagen, and VEGFRs (Fig. S2H) and were enriched in angiogenesis, migratory, and ECM modeling pathways compared to non-RAECs (Fig. S2I). For potential regulators of RAECs, the SCENIC analysis demonstrated that SOX4 and ZEB1 were the most specific regulons of RAECs (Fig. S2J). Consistently, compared to other subtypes of RAECs, tip cells had the highest mRNA levels of SOX4 and ZEB1 (Fig. S2F). Taken together, RAECs primarily contain tip cells and immature cells, and such tip cells are highly differentiated with pronounced angiogenic and ECM modeling activities.Development of a robust 18-gene recurrence-associated endothelial cell signature (RACEsig)Since RAECs were prognostic for PCa recurrence, we sought to construct a robust gene signature that represents RAECs and can predict tumor recurrence. Univariate Cox regression analysis demonstrated that 18 of the 54 RAEC-specific genes were associated with disease-free survival (P < 0.1) in the TCGA-PRAD cohort (Fig. 2a and Table S7). Consistent with the RAEC characteristics described above, marker genes of tip cells (ESM1, FSCN1) and immature cells (APLNR) were among the 18 genes.Fig. 2: Development of an 18-gene signature of recurrence-associated endothelial cells.a Forest plot of 18 prognostic genes identified by univariate Cox regression analysis of 54 RAEC-related genes presented in Fig. 1g. P < 0.1 for all 18 genes. b Heat map of relative mRNA levels of the 18 prognostic RAEC-related genes across major cell types in the Chen dataset. c RAECs had the highest score when the 18 genes’ z-scores were summed for different types of cells. d Heat map of the 18 genes’ mRNA expression pattern in TCGA-PRAD. e C-index and IBS in outer 10-fold validation using the 11 hyperparameter-tuned models. Dots indicate each model’s mean of 10 C-indices or integrated Brier scores (IBS). The model with the highest average C-index and lowest average IBS was selected and termed recurrence-associated endothelial cell signature (RAECsig). f Bar plot showing feature importance of the 18 prognostic RAEC-related genes inferred by the eXtreme Gradient Boosting (XGBoost). Greater importance suggests more contributions to the XGBoost model when predicting PCa progression. g RAECsig scores among major cell types in the Chen dataset. SVM, support vector machine; Enet, elastic network; Lasso, Least Absolute Shrinkage and Selection Operator; plsRcox, partial least squares regression for Cox; StepCox, stepwise Cox regression; GBM, generalized boosted regression modeling; RSF, random survival forest; SuperPC, supervised principal components.Each of the 18 RAEC genes was highly expressed in ECs, with 6 genes downregulated and 12 upregulated in RAECs (Fig. 2b). The z-scores for these 18 genes were summed across other major cell types, and the sum scores were the highest in the RAECs (P < 0.001, Fig. 2c), further indicating the power of these genes in distinguishing RAECs from other types of cells. The expression patterns of the 18 RAEC genes were confirmed in the Ge dataset (Fig. S3A), in which almost all RAEC genes showed higher mRNA levels in ECs than non-ECs, and the 18 genes’ scores were the highest in ECs (Fig. S3B, C).Importantly, these prognostic genes derived from scRNA-seq data can be applied in the bulk RNA-seq data. By measuring gene coexpression and hierarchically clustering these genes’ coexpression patterns, we found that the mRNA expression levels of 12 upregulated RAEC genes were highly correlated with but separated from those of the 6 downregulated RAEC genes (Fig. 2d and S4). Therefore, the scRNA-seq-derived gene markers could be leveraged to infer RAEC’s abundance in bulk RNA-seq data.Based on the 18 genes, we benchmarked 11 survival-related machine-learning algorithms through nested CV in the TCGA-PRAD. As shown in Fig. 2e and Table S8, the XGBoost survival model achieved the best performances with the highest mean C-index (0.692) and lowest mean IBS (0.151). This model with tuned hyper parameters was then fitted on the entire TCGA-PRAD dataset and termed RAECsig hereafter. The feature importance of the 18 genes is shown in Fig. 2f, where the top 5 features included FAM107A, FSCN1, TMEM255B, GABRD, and DOCK6. Using the RAECsig, we then calculated risk scores of different cell types in the Chen dataset and found that RAECs ranked first among all major cell types (P < 0.001, Fig. 2g), demonstrating that the RAECsig is indicative of RAECs. Furthermore, the RAECsig score was the highest in tip cells of RAECs (Fig. S5A) and could significantly discriminate them from other ECs (AUC = 0.844; Fig. S5B).RAECsig is an independent risk factor for primary PCa recurrenceThe prognostic value of the RAECsig was validated using 5 bulk sequencing datasets of primary PCa. The RAECsig risk score was calculated for each case, and the “survminer” package was used to define the optimal risk score threshold across various datasets. Using the cut off 0.58 derived from the GSE21034 cohort, we assigned patients to the high- and low-risk groups in each dataset. A higher RAECsig score was significantly associated with shorter disease-free survival in each cohort (log-rank P ≤ 0.01, Fig. 3a–f and S6A). Univariate Cox regression analyzes showed that the RAECsig was a risk factor for PCa recurrence in all cohorts (Fig. 3g). After adjusting for age, Gleason score, serum PSA, and TNM stage, RAECsig remained a statistically significant prognostic factor in the TCGA-PRAD (HR: 5.321, 95% CI: 3.347–8.459, P < 0.001), DKFZ-PRAD (HR: 2.897, 95% CI: 1.578–5.321, P < 0.001), GSE70768 (HR: 6.382, 95% CI: 1.722-23.649, P = 0.006), GSE70769 (HR: 2.742, 95% CI: 1.173–6.407, P = 0.02), GSE94767 (HR: 2.893, 95% CI: 1.338–6.254, P = 0.007), and Meta-Cohort (HR: 2.663, 95% CI: 1.707–4.155, P < 0.001) except in GSE21034 (HR: 2.499, 95% CI: 0.817–7.642, P = 0.108). Furthermore, time-dependent ROC analysis revealed a robust discrimination power of RAECsig in PCa recurrence. Specifically, the AUC values at 1, 3, and 5 years were 0.82, 0.79, and 0.77 in TCGA-PRAD; 0.88, 0.91, and 0.83 in DKFZ-PRAD; 0.84, 0.75, and 0.80 in GSE70768; 0.71, 0.75, and 0.74 in GSE70769; 0.60, 0.67, and 0.65 in GSE94767; 0.82, 0.69, and 0.66 in GSE21034; and 0.70, 0.70, and 0.68 in Meta-cohort (Fig. 3h–m and S6B). Taken together, we found that a universal RAECsig cut off can classify patients into either high- or low-risk groups and that RAECsig can be an independent risk factor for PCa recurrence.Fig. 3: Independent validation of the 18-gene RAECsig in multiple PCa cohorts.a–f Recurrence-free survival of high- and low-risk groups stratified by a universe RAECsig value (0.58), including TCGA-PRAD (a), DKFZ-PRAD (b), GSE70768 (c), GSE70769 (d), GSE94767 (e), and GSE21034 (f). P values were derived from log-rank tests and were < 0.05. g Forest plot showing hazard ratio (HR) at 95% confidence interval (CI) and the corresponding P values of RAECsig and clinical and pathological characteristics using both the univariate (above the dashed lines) and the multivariate Cox regression analyzes (below the dashed lines) in 6 PCa cohorts. Only variables with a P value < 0.05 in univariate analyzes were included in multivariate analyzes. The Meta-Cohort consists of 4 PCa cohorts (panels c–f) in which gene expression was detected using the microarray platform. h–m AUROC curve analysis of RAECsig for predicting recurrence at 1, 3, and 5 years in the cohorts of TCGA-PRAD (h), DKFZ-PRAD (i), GSE70768(J), GSE70769(K), GSE94767(l), and GSE21034 (m). pT, pathological tumor stage; PSA, prostate-specific antigen; pN, pathological lymph node stage; AUROC, area under the receiver operating characteristic curve.Comparison of RAECsig with clinicopathological features and commercial genetic assaysClinicopathological variables such as age, Gleason score, and TNM stage are commonly used to assess PCa prognosis in clinical practice. We, therefore, used the C-index to compare the predictive power of RAECsig to those of common clinicopathological features across datasets (Table S9). The C-index [95% CI] of RAECsig was calculated with 0.767 [0.714–0.819] in TCGA-PRAD, 0.836 [0.765–0.907] in DKFZ-PRAD, 0.769 [0.671–0.868] in GSE70768, 0.679 [0.608-0.750] in GSE70769, 0.641 [0.546–0.736] in GSE94767, and 0.669 [0.564–0.774] in GSE21034 (Fig. 4a–f). Overall, RAECsig performed better than age, Gleason score, and TNM stage in TCGA-PRAD and GSE70768 but was statistically non-inferior in other datasets.Fig. 4: Comparison of RAECsig with clinical features and published models in predicting recurrence across PCa cohorts.a–f C-indices of the RAECsig, age, pathological tumor stage (pT), pathological lymph node stage (pN), prostate-specific antigen (PSA), and Gleason scores in different PCa cohorts, including TCGA-PRAD (a), DKFZ-PRAD (b), GSE70768 (c), GSE70769 (d), GSE94767 (e), and GSE21034 (f). g Univariate Cox regression analysis of RAECsig and published models across PCa cohorts. h C-indices of the RAECsig and published models in various PCa cohorts. Data in (a–h) are presented as mean ± 95% confidence interval. *P < 0.05; **P < 0.01; ***P < 0.001; ****P < 0.0001.We also compared RAECsig with OncotypeDX, Prolaris, and Decipher. There were no overlapping genes between RAECsig and each of these 3 assays. We calculated OncotypeDX-like, Prolaris-like, and Decipher-like scores to samples in all datasets. Similar to the RAECsig scores, the OncotypexDX-like, Prolaris-like, and Decipher-like scores showed consistent positive associations with recurrence in all datasets (Fig. 4g). Notably, the univariate HRs of RAECsig were greater than those from the three assays in all datasets. Furthermore, we calculated C-indices for these signatures and found that RAECsig significantly outperformed the other three in TCGA-PRAD (OncotypeDX-like: 0.661 [0.592–0.729]; Prolaris-like: 0.654 [0.586–0.721]; Decipher-like: 0.588 [0.518–0.659]). In other datasets, RAECsig was comparable to the other three assays (Fig. 4h; Table S10). Additionally, we confirmed that RAECsig performed better than the endothelial signature in predicting PCa recurrence (Fig. S7). Altogether, RAECsig displayed robust predictive powers compared to clinical features and gene panels in different cohorts and, therefore, had better extrapolation potential.Key RAEC genes enhance angiogenic activity and PCa cell proliferationOf the 12 upregulated RAEC-associated genes, FSCN1 was ranked the top one by XGBoost (Fig. 2f), followed by TMEM255B and GABRD. Additionally, these 3 genes’ expression levels were significantly associated with PCa recurrence and higher Gleason scores (Fig. S6). We thus chose FSCN1, TMEM255B, and GABRD to evaluate the RAECsig in PCa progression. FSCN1 is a marker of tip cells and was significantly upregulated in ECs from samples with higher Gleason scores (Fig. 5a). The expressions of TMEM255B and GABRD in ECs were also confirmed at the protein level (Fig. S7).Fig. 5: RAECs are associated with enhanced angiogenic activity and PCa tumor growth.a Representative immunostaining micrographs of FSCN1 and CD31 in human PCa samples with higher and lower Gleason scores. Nuclei were stained with DAPI. Images to the right are magnified boxed areas. Scale bars, 20 μm. Data are presented as means± s.e (n = 28, shown in the right panel). P value was calculated by the Wilcoxon rank sum test. b Gene set enrichment analysis revealed top dysregulated HALLMARK pathways in PCa patients with higher RAECsig scores. c Validation of RNAi-mediated knockdown of FSCN1, TMEM255B, and GABRD in HUVECs, as measured by real-time qPCR. Data are presented as means ± s.d (n = 3). Statistical significance was determined using two-tailed Student’s t-tests. d Tube formation of HUVECs was affected by silencing GABRD and TMEM255B. Data are presented as means ± s.d (n = 3, shown in right panels). e Knockdown of GABRD by 2 siRNAs in HUVEC, as measured by real-time qPCR. Data are presented as means ± s.d (n = 3). f GABRD knockdown inhibited HUVEC migration and invasion, as determined by the transwell assay. Data are denoted as means ± s.d (n = 3). g GABRD and TMEM255B knockdown induced HUVEC apoptosis. Data are denoted as means ± s.d (n = 3). h The number of PCa cells was decreased by conditioned media from HUVECs with the knockdown of FSCN1, TMEM255B, and GABRD, as determined by the CCK-8 assay. Data are shown as means ± s.d (n = 3). Statistical significance was determined by one-way ANOVA with Dunnett’s test for (d–h). *P < 0.05; **P < 0.01; ***P < 0.001.As a higher RAECsig score suggested more abundant RAECs, particularly more tip cells, we reasoned that PCa with higher RAECsig scores could manifest increased angiogenic activities to enhance tumor growth by supplying more oxygen and nutrients. Consistent with this idea, PCa with higher RAECsig scores were enriched in pathways of angiogenesis and cell cycle progression based on GSEA of consensus DEGs from the bulk RNA-seq datasets (Fig. 5b and S8; Table S11). Therefore, we assessed whether FSCN1, TMEM255B, and GABRD functionally enhance angiogenic activities. In HUVECs, silencing GABRD or TMEM255B significantly inhibited their tube formation, an indicator of angiogenesis (Fig. 5c, d). GABRD silencing was more potent than other genes in the tube formation assay. Consistently, GABRD silencing significantly reduced the migration and invasion of HUVECs (Fig. 5e, f). Also, we found that silencing GABRD or TMEM255B significantly induced apoptosis in HUVEC cells, which may also explain the dramatic reduction in their tube formation (Fig. 5g).Accumulating evidence indicates that ECs can also regulate PCa progression via interacting with tumor cells [44,45,46]. We thus used collected media from HUVECs with the silencing of FSCN1, TMEM255B, and GABRD to treat human PCa cells and performed cell proliferation analysis. We found that silencing FSCN1, TMEM255B, or GABRD in HUVECs significantly inhibited the proliferation of PC-3 and 22Rv1 cells (Fig. 5h). These findings suggest that RAECs could also promote PCa progression by directly modulating tumor cell proliferation.The RAECsig also predicts castration resistanceCell cycle progression (CCP) is inversely correlated with AR activities and can predict primary abiraterone resistance in metastatic PCa [47]. We noticed that the downregulation of the androgen response pathway was accompanied by the enrichment of cell cycle-related pathways in RAECsig-high tumors (Fig. 5b). We thus investigated whether RAECs also play a role in castration resistance in PCa. We calculated the CCP scores, AR activities, and CRPCsig51 scores for tumor samples in the TCGA-PRAD, DKFZ-PRAD, and Meta-Cohort and found that higher RAECsig scores were significantly correlated with both CCP and CRPCsig51 scores while weakly correlated with negative androgen activities (Fig. 6a–c and Fig. S9A, B). To further test whether an increase in RAECs is associated with the development of castration resistance in PCa, we curated scRNA-seq data of ECs from both primary PCa and mCRPC based on the expression of PECAM1 and PROX1 (Fig. S10A–M). After removing samples with less than 50 ECs (Fig. S10N) and correcting batch effects (Fig. S11A–D), a total of 16645 ECs from 28 samples were used to generate a transcriptional atlas of ECs across PCa stages (Fig. 6d). All EC subtypes described above were identified based on canonical markers (Fig. 6e and S10F–11E).Fig. 6: Higher RAECsig scores are associated with castration resistance in PCa.a–c Associations of higher RAECsig values with higher cell cycle progression (CCP) scores (a), higher values of a CRPC signature (CRPCsig51, b), and lower AR activities (c). The correlation coefficient R and corresponding P values were determined using Spearman’s rank correlation analysis. d UMAP plot of ECs collected from 4 PCa scRNA-seq datasets of primary PCa (blue) and metastatic CRPC (mCRPC, red). e UMAP plot of EC subtypes. EC (LS) represents cells with lower sequencing depth. f Featureplot showing RAECsig scores of ECs. RAECsig scores were calculated using scaled data as input for the RAECsig model. g UMAP plot of ECs stratified by RAECsig scores. The cut off RAECsig value for discretion was determined using the Youden index that predicts tip cells from RAECs with maximized sensitivity and specificity. h, i Boxplot showing proportions of RAECsig-high and RAECsig-low cells (h) and EC subtypes (i) in primary PCa and mCRPC. Dots represent the cell proportion from a sample. Statistical significance was determined using the Wilcoxon rank sum test. ns, not significant; *, P < 0.05; **, P < 0.01; ***, P < 0.001; ****, P < 0.0001. (j) Kaplan–Meier analysis showing overall survival with a first-line ARSI between RAECsig-high (red) and RAECsig-low (blue) mCRPC. k–n Boxplots showing baseline RAECsig (k), AR activity (l), CCP score (m), and CRPCsig51 score (n) between enzalutamide responders and non-responders in primary PCa. Statistical significance was determined using the Wilcoxon rank sum test. (o) ROC curves of RAECsig, AR activity, CCP score, and CRPCsig51 score in predicting the probability of primary resistance to enzalutamide in the GSE197780 dataset. CCP, cell cycle progression; AR, androgen receptor; PCV, postcapillary vein; LS, lower sequencing depth; mCRPC, metastatic castration-resistant prostate cancer; ARSI, AR-signaling inhibitors; ROC, receiver operating characteristic.We then calculated the RAECsig score for each subtype of ECs. Again, ECs with higher RAECsig scores were mostly tip cells (Fig. 6f). RAECsig-high ECs were significantly enriched in mCRPC (Fig. 6g) whereas RAECsig-low ECs were more enriched in primary tumors (Fig. 6h). Consistent with the enrichment of tip and immature cells in RAECs, both tip and immature cells were also enriched in mCRPC (Fig. 6i). In addition, mCRPCs with higher RAECsig scores had significantly worse prognoses after treatments with first-line AR-signaling inhibitors (ARSI) (Fig. 6j).We also tested whether higher RAECsig scores can predict primary resistance to ARSI in primary PCa using the GSE197780 dataset, in which patients underwent bulk RNA sequencing before 3 months of neoadjuvant enzalutamide treatment [35]. Non-responders had significantly higher RAECsig scores than responders in these primary tumors (Fig. 6k). However, we did not detect significant differences in the primary tumors’ AR activities, CCP scores, and CRPCsig51 scores (Fig. 6l–n). ROC analysis showed that RAECsig could markedly discriminate non-responders from responders of enzalutamide (AUC = 0.704; Fig. 6o). These findings suggest that RAECsig can identify patients who may not respond to ADT.
