Disease-specific variant pathogenicity prediction using multimodal biomedical language models

Missense variants play a key role in the diagnosis of genetic disorders and in disease risk prediction. Existing methods focus primarily on the prediction of variant effects in terms of their deleteriousness, without taking into account the disease-specific context, and are therefore limited in terms of their utility in real-world diagnosis and decision making. Here, we introduce di sease-specific va riant pathogenicity prediction (DIVA), a novel deep learning framework that directly predicts specific disease types alongside the probability of deleteriousness for missense variants. Our approach integrates information from two different modalities – protein sequence and disease-related textual annotations – encoded using two pre-trained language models and optimized within a contrastive learning paradigm designed to align variants with relevant diseases in the learned representation space. Our results demonstrate that DIVA outperforms baselines and provides accurate disease predictions with high relevance to clinically curated disease annotations for missense variants. Variant deleteriousness prediction is enhanced by incorporating AlphaMissense scores through learnable weights derived from protein function annotations, which additionally boosts DIVA ’ s ability to accurately classify deleterious variants. Our work provides new insights into variant pathogenicity prediction with awareness of disease specificity, addressing a hitherto unmet need in relation to clinical variant interpretation.