B1165
Title: Joint additive factor analysis for multi-omics data integration
Authors: Niccolo Anceschi - Duke University (United States) [presenting]
Federico Ferrari - Duke University (United States)
Himel Mallick - Cornell University (United States)
David Dunson - Duke University (United States)
Abstract: In precision medicine, it is common to gather data from multiple modalities to characterize different aspects of a patient across biological layers. Such data can lead to more accurate prediction of health responses, motivating principled approaches to integrate modalities. With multi-omics data, the signal-to-noise ratio can vary substantially across modalities, which requires more structured statistical tools beyond standard late and early fusion. This challenge comes with the need to preserve interpretability, allowing the identification of relevant biomarkers and proper uncertainty quantification for the predicted outcomes. While these properties are naturally accounted for within a Bayesian framework, state-of-the-art factor analysis (FA) formulations for multi-omics data rely on loose modeling assumptions. A novel joint FA model having a structured additive design is proposed, accounting for shared and view-specific components and allowing for flexible covariate and outcome distributions. A fast implementation is provided via MCMC, and the approach is extended to account for interactions among latent factors and deviations from normality.
