Cross-cohort projection of clinically anchored latent risk enables multi-omics interpretation without refitting.

Linking clinically derived risk signals to reproducible molecular states across independent cohorts remains a major challenge in translational bioinformatics. Existing approaches often rely on cohort-specific model fitting, limiting cross-dataset comparability and downstream biological interpretation. We developed a cross-cohort projection framework that maps baseline clinical variables to a clinically anchored latent risk coordinate, $\mu$, enabling application across external datasets without refitting. The fixed projector was trained in a local imaging cohort and applied unchanged to independent cohorts. Projected $\mu$ was evaluated across multiple molecular layers, including bulk transcriptomics, single-cell-guided deconvolution, spatial transcriptomics, and circulating cell-free DNA (cfDNA). In an independent external cohort, projected $\mu$ preserved separation of time to castration resistance across predefined strata (P = .002), with 30-month risk increasing from 0.13 to 0.86 across ordered $\mu$ bins. In bulk transcriptomics, higher projected $\mu$ was associated with increased proliferation-related signaling and reduced androgen receptor/lineage programs ($\rho$ = 0.40 and -0.26; both P < .001). Deconvolution analyses linked higher projected $\mu$ to reduced AR-high epithelial cell fractions ($\rho$ = -0.17, P = .001). Spatial transcriptomics demonstrated organized tissue-level structure of prespecified molecular programs. In cfDNA, higher projected $\mu$ was associated with a more negative RB1 copy-number signal in the detectable subset ($\rho$ = -0.49, P = .0278). This study presents a projection-based framework for cross-cohort translation of clinically anchored latent risk into interpretable multi-omics context. By enabling reuse of a fixed coordinate without refitting, the approach provides a practical strategy for linking clinical risk to molecular programs and blood-based readouts across datasets.