Disentangled multimodal evolutionary representations for cross-virus predictive modeling of antigenic change

Abstract

Antigenic evolution drives immune escape and complicates early assessment of emerging pathogens, especially when genomic surveillance is sparse and functional measurements lag behind real-world spread. Early prediction and generalization to unknown viruses require faithful and information-rich representations of viral evolution, a need that is not fully addressed by existing protein language model-based approaches. Here we introduce DERIVE, a flow-based generative framework that learns a disentangled latent representation of multimodal evolutionary history by integrating sequence homology with physicochemical and structural features. Using only pre-pandemic SARS-CoV-2 sequences, DERIVE prioritizes high-risk mutations, reconstructs strain-level evolutionary trajectories, forecasts immune escape and prevalence trends, and produces mutation-level effect maps. The learned representation transfers robustly across viral families, with strong concordance to functional data for influenza virus, HIV, rabies lyssavirus and chikungunya virus. Together, these results position DERIVE as a generalizable and interpretable framework for anticipating antigenic evolution and supporting early, actionable assessment of future "Disease X" threats.