5074 Background: Metastatic castration-resistant prostate cancer (mCRPC) is characterized by profound clinical and biological heterogeneity, including variable dependence on androgen receptor signaling, frequent genomic instability, lineage plasticity, and highly divergent clinical outcomes. Despite the widespread use of genomic profiling and expression-based signatures, robust patient-level risk stratification and biologically grounded disease subtyping remain limited. We hypothesized that inference of transcriptional regulatory programs underlying tumor state, integrated with genomic and clinical features, would improve prognostication in mCRPC. Methods: We analyzed genomic, transcriptomic, and clinical data from mCRPC tumors in the Stand Up To Cancer (SU2C) cohort. A transcriptional regulatory network was reverse-engineered using the latest version of the Algorithm for the Reconstruction of Accurate Cellular Networks (ARACNe3) and transcriptional regulator activity was inferred using Nonparametric analytical Rank-based Enrichment Analysis (NaRnEA). We constructed a novel prognostic machine learning model using a random survival forest integrating regulatory activity features, canonical genomic alterations, and clinical variables. Results: Samples were stratified by RNA-sequencing library preparation method into polyA-purified (training, n=150) and exon-capture (testing, n=100). Our model demonstrated strong and statistically significant performance in both the training cohort (out-of-bag Cox PH p=4.63×10⁻⁸, AUROC = 0.7711) and the independent test cohort (Cox PH p=2.92×10⁻⁵, AUROC = 0.7413). The most informative features included inferred activities of key transcriptional regulators implicated in proliferation, lineage programs, and neuroendocrine differentiation (E2F1, MYC, SOX2, AR, REST), canonical genomic alterations relevant to mCRPC biology (AR, TP53, RB1, FOXA1, APC, PI3K pathway), and select clinical variables (prior therapy, Gleason score, age, metastatic site). Our model also outperformed several widely used gene expression signatures (AR, NEPC, Cell Cycle Progression, RB1 Loss). Regulatory activity–based analyses also revealed substantially greater biological structure and cluster coherence than gene expression–based analyses in subsequent analyses, highlighting distinct and clinically aggressive regulatory states in this dataset. Conclusions: Integrating regulatory network–inferred transcriptional programs with genomic and clinical features enables robust, independently validated prognostication in mCRPC that outperforms widely used expression-based signatures. This approach exposes clinically relevant biological heterogeneity underlying aggressive and treatment-resistant disease states and provides a framework for improved risk stratification in advanced prostate cancer.
Griffin et al. (Wed,) studied this question.