PROMISE-Surv: Proteomics-guided survival representation transfer for deployable transcriptomic cancer modeling

Deep proteomic and phosphoproteomic profiling captures functional tumor signaling, but these assays are unavailable in most large survival cohorts. Existing multi-omics approaches often require complete molecular profiles or reconstruction of missing modalities, leaving unclear whether survival-relevant protein-layer information can be transferred directly to routinely available data. We therefore investigated whether proteomics-informed survival structure could be transferred to an RNA-plus-clinical model without reconstructing missing protein measurements.

PROMISE-Surv was developed as a teacher-student framework using complete-case CPTAC/PDC tumors from lung adenocarcinoma (LUAD), kidney renal clear cell carcinoma (KIRC), and head and neck squamous cell carcinoma (HNSC) as the multimodal discovery cohort (n = 314). The student model was then deployed in matched TCGA cohorts from the same cancer types (n = 1571), with independent GEO cohorts used as endpoint-specific external stress tests.

Across ten repeated held-out splits, the blended student achieved a mean C-index of 0.687 (95% CI, 0.638-0.730); matched discovery scores preserved teacher-derived risk ordering across all three cancers (Spearman ρ=0.981-0.989). In TCGA, transferred risk stratified survival in KIRC and LUAD and recovered cancer-specific RNA programs linked to adverse tumor biology, with concordant pathway-level support from RPPA. Cross-cancer analysis identified eight directionally conserved RNA-risk markers, while enrichment of the broader candidate ranking implicated RNA processing and ribosome biogenesis. These findings support biological coherence of the transferred risk representation.

PROMISE-Surv provides a biological-transfer framework for extending proteomics-informed survival structure to RNA-plus-clinical cohorts without protein-layer measurements.