A deep mutational scanning-informed protein language model predicts SARS-CoV-2 evolution dynamics with spatiotemporal resolution.

Early identification of emerging dominant variants of pathogens such as SARS-CoV-2 is important for effective public health responses, yet existing approaches are not feasible for real-time surveillance. Here we introduce DeepCoV (DMS-Empowered Evolution Prediction of CoronaVirus), a deep-learning framework for the dynamic identification of emerging variants with high potential to become prevalent at spatiotemporal resolution. It integrates deep mutational scanning (DMS)-derived mutation phenotypes, evolutionary sequence data and epidemiological surveillance data reflecting human immune pressures. Benchmarked against logistic regression-based methods and representative deep-learning approaches in simulated retrospective surveillance scenarios, DeepCoV accurately forecasts the dominance of recently circulating lineages a month in advance, achieving a 90% reduction in false discovery rate while capturing temporal and geographic dynamics of variant spread and reconstructing their regional prevalence trajectories. It also identified mutational hotspots of Omicron-derived backbones in silico, revealing convergent evolution trends. This provides a scalable framework for timely identification of immune-evasive variants and critical mutations, providing actionable insights.