Integrative single-cell and spatial transcriptomics with explainable AI reveal lethal prognostic axis in prostate cancer.

Abstract

Prostate cancer (PCa) remains clinically heterogeneous. We integrated single-cell and spatial transcriptomics with explainable machine learning to define a lethal tumor axis and establish an interpretable prognostic model. From 141,986 high-quality single cells spanning localized, hormone-sensitive, and castration-resistant PCa, we identified a malignant C4 epithelial subpopulation characterized by high chromosomal instability, androgen receptor and cell-cycle activation, and stemness potential. Spatial mapping further revealed immune-enriched yet suppressive niches, where fibroblasts and myeloid cells coexisted with exhausted lymphocytes, reflecting functional immune imbalance. We benchmarked 101 machine learning pipelines, selecting a Lasso plus PLS-Cox model that achieved strong concordance across independent cohorts. The C4-based risk score independently predicted recurrence-free survival after adjustment for age, Gleason score and T stage, and a nomogram combining this score with clinical variables showed good discrimination. SHAP interpretation highlighted MT1M, PCSK1N, and ACSL3 as major risk-driving features. PCSK1N was progressively upregulated from normal prostate to castration-resistant disease and promoted proliferation, clonogenicity, migration and enzalutamide resistance, while its inhibition sensitized organoids and xenografts to AR-targeted therapy. These findings define a C4-centered lethal tumor axis and provide an explainable, experimentally supported framework for prognostic stratification in PCa.