Multimodal deep learning enhances genomic risk prediction for cardiometabolic diseases.

Abstract

Cardiometabolic diseases (CMDs) are multifactorial disorders influenced by numerous genetic variants and their complex interactions. Although recent studies have advanced the understanding of genetic risk prediction, current approaches predominantly rely on linear models that may not fully capture the complex, nonlinear relationships between genetic factors. Here, we present DeepGP (Deep learning-based Genome-wide Predictor), a novel multimodal deep learning framework that incorporates bidirectional state space models to predict CMD risk using genome-wide variants and demographic data. We conducted extensive experiments to evaluate DeepGP's performance. First, in simulation studies incorporating joint genetic and environmental interactions, we demonstrated DeepGP's superior prediction performance across varying levels of heritability. When evaluated on eight CMDs in cohorts of European ancestry from the UK Biobank (UKB), DeepGP achieved significantly higher accuracy compared with conventional polygenic risk scores and machine learning methods. Model interpretability analysis identified both well-established genes and potential new signals contributing to the risk of the disease. Evaluation within UKB subpopulations of African and Caribbean ancestry provided preliminary evidence of within-resource transferability for type 2 diabetes. These results demonstrate the potential of deep learning technologies to enhance genetic risk stratification for complex diseases, while underscoring the need for external validation in more diverse cohorts.