Multi-task deep learning for risk stratification and shared molecular architecture of cardiometabolic multimorbidity under fragmented multi-omics

Cardiometabolic multimorbidity (CMM) reflects a systemic failure of interconnected physiological networks, yet current risk stratification approaches based on macroscopic clinical phenotypes leave substantial molecular residual risk unresolved. Although multi-omics integration offers a means to capture this latent vulnerability, its population-scale application is limited by pervasive fragmentation and non-overlapping availability of proteomic and metabolomic data. Here, we develop the Multi-Omics Perceiver for Survival (MOP-Surv), a deep learning framework designed to accommodate incomplete multi-omics profiles without imputation through dynamic attention masking and multi-task survival learning. Applied to 297,067 UK Biobank participants, MOP-Surv achieved consistent risk stratification across six cardiometabolic endpoints and provided modest incremental predictive value. Beyond prediction, MOP-Surv identified a hierarchical structure of cross-endpoint prognostic associations and highlighted a parsimonious set of biomarkers including GDF15, EDA2R, and WFDC2 with consistent prognostic relevance across diverse disease trajectories. Therefore MOP-Surv provides a practical approach for integrating fragmented multi-omics data to characterize shared prognostic patterns across cardiometabolic outcomes.