Background: Although the survival outcomes for patients with mCSPC has improved over the last 5 years, disease remains universally fatal even with improved therapies. Currently, genomic information from the tumor is not taken into account for treatment selection and prognostication. AI is increasingly being used in clinical cancer genomics research. Probabilistic Graphical Models (PGMs) are AI algorithms that capture multivariate, multi-level dependencies in complex patterns in large datasets while retaining human interpretability. We hypothesize that PGMs can establish correlation of baseline somatic genomic alteration with poor survival outcomes in mCSPC. Methods: Eligible men had new mCSPC starting systemic therapy and had tumor genomic profiling done through a CLIA certified lab. Gene alterations with known pathogenicity were grouped into canonical pathways. Multilevel associations between survival, clinical variables (including baseline PSA, Gleason ≥ 8, and visceral metastasis), and genomic signatures (PI3K/AKT/mTOR, HRR, G1/S Cell Cycle, SPOP, TP53, WNT, and MYC) were assessed using a Bayesian Network (BN), and confidence intervals were estimated by bootstrapping. A Kaplan-Meier (KM) survival analysis was performed independently to support the results generated by the BN. Results: Among all variables, only genomic alterations in TP53 and the G1/S pathway were significantly associated with poor overall survival by BN analysis. KM analysis showed concordant results for TP53 (median OS, altered 50 mos vswild-type 84 mos; HR=2.79, 95% CI 1.63 – 4.80; P=0.0002) and G1/S (median OS altered 23 mos vswild-type 73 mos; HR=8.21, 95% CI 3.40 – 19.86; P<0.0001). Conclusions: These hypothesis-generating data reveal genomic signatures associated with poor survival in mCSPC patients. Our results, after external validation, may aid with counseling and treatment selection, as well as patient stratification in future trials in mCSPC.