Population Science Seminar - Huang Lin, PhD | DiAGMI: Disentangling Outcome-Predictive and Subtype-Associated Signals in Multi-omics Cancer Data

Multi-omics cancer cohorts often contain far more molecular features than patient samples, making subtype discovery and outcome prediction difficult to separate. Existing approaches frequently combine subtype-related and outcome-predictive signals into a single representation, limiting interpretability and introducing confounding. We present DiAGMI, a two-module framework designed to disentangle these signals while improving performance in small multi-omics datasets. DiAGMI-Gen separates outcome-predictive and subtype-associated latent representations across multiple omic layers and generates high-fidelity synthetic patient samples to augment training data. We identify key modifications that prevent clustering collapse and introduce a fidelity gate to detect synthetic data drift. DiAGMI-Factor then integrates real and accepted synthetic samples through interpretable non-negative multi-view factorization to support outcome prediction, subtype-aware clustering, and cross-omics feature ranking. Across five benchmark scenarios spanning varying signal strengths, class imbalance, and multiple outcome types, DiAGMI generated realistic synthetic samples and improved prediction stability while preserving biologically meaningful feature associations. Applied to a real breast cancer multi-omics cohort, DiAGMI disentangled subtype-associated and survival-predictive variation, identifying candidate cross-omics biomarkers linked to outcome independently of subtype for future biological validation.