A structure-informed evolutionary model for predicting viral immune escape and evolution

Abstract

Persistent emergence of viral variants capable of evading host immunity constitutes a significant threat to public health. This antigenic evolution frequently outpaces the development of vaccines and therapeutics, highlighting the necessity of predictive models surveilling the immune escape potential of emerging variants. However, existing models suffer from two key limitations: they inadequately incorporate protein structural information and neglect the importance of distinguishing large-impact mutations from neutral ones given multiple mutations. To address these gaps, we presented KEScape, a deep learning model designed to predict viral immune escape and evolution. KEScape integrates evolutionary context with protein structural information, introduces a novel top-K L2-differential pooling mechanism to prioritize mutations with large functional effects, and incorporates a supervised L2 margin loss to facilitate the L2-distance-based ranking of high-impact mutations. We demonstrated that KEScape significantly outperformed state-of-the-art models on the most comprehensive benchmark to date, comprising eleven deep mutational scanning experiments spanning diverse viruses. Furthermore, KEScape exhibited outstanding performance in the practical applications of identifying immune escape hotspots and variants in emerging lineages and real-time surveillance of lineages associated with WHO-designated variants of SARS-CoV-2. These results show that KEScape is an effective model to predict viral immune escape and evolution. Its capacity for early warning can directly inform public health interventions and guide the development of countermeasures, thereby mitigating the threat of future viral pandemics.