Explainable machine learning for continuous transcriptomic modeling of metastatic risk in prostate cancer (MET-score).

Abstract

5129 Background: Metastatic progression is the leading cause of mortality in prostate cancer, yet clinicopathological stratification incompletely identifies primary tumors with latent metastatic potential. Emerging evidence supports metastatic competence as a gradual transcriptional reprogramming, motivating molecular tools that capture continuous risk states. Methods: Transcriptomes from primary and metastatic prostate tumors (n = 766; PRIMARY = 507, MET = 259) spanning 16,234 gene features were integrated to train an interpretable XGBoost model. SHapley Additive exPlanations were used to derive a minimal 30-gene signature and define a continuous metastasis-likeness transcriptomic score (MET-score). MET-score was applied to primary tumors to quantify metastasis-likeness and characterize associated biology. External validation was performed in an independent primary prostate cohort profiled with exon microarrays (GSE21034) without model retraining. Results: The minimal 30-gene signature captured metastasis-associated signal with high discriminative performance (ROC AUC ≈ 0.98). In primary tumors (n = 507), MET-score exhibited a right-skewed continuous distribution (mean 0.0079, SD 0.0640; range 0.0023–0.9953), consistent with heterogeneous degrees of metastasis-like transcriptional activity within primary disease. Using a prespecified high-risk threshold at the 0.85 quantile (cutoff 0.002603), 81/507 (16.0%) primary tumors were classified as HIGH MET-score and 426/507 as LOW, identifying a subset with transcriptomic profiles most closely resembling metastatic disease. MET-score generalized robustly to GSE21034 while preserving its continuous structure and enabling molecular stratification despite platform differences and partial gene loss: 26/30 signature genes were present in the microarray cohort (missing C15orf40, ADIRF, CTC1, NTPCR). However, prognostic validation against clinical endpoints could not be performed because the publicly available metadata did not include clearly annotated time-to-event and event-status variables. Conclusions: An explainable 30-gene MET-score was developed to quantify metastasis-likeness as a continuous transcriptomic trait in primary prostate cancer. The score identifies a high-risk primary subset and transfers to an independent external cohort (GSE21034) without model retraining, supporting cross-platform robustness and prioritizing MET-score for clinical outcome validation in RNA-seq–annotated cohorts.