A Novel Coupled Epistatic Algorithm to Predict Host-Adapted Hidden Genotypes of Influenza A Virus Hemagglutinin

Predicting adaptive genotypes of influenza A virus (IAV) hemagglutinin (HA) is challenged by epistatic mutations, often missed in single-site analyses. We developed a novel Coupled Epistatic Algorithm (CEA) and CEA-refined Language Models (CLM) to identify epistatic mutations driving human adaptation and pandemic risk in IAV HA. The CEA integrated evolutionary constraints and penalty metrics, outperforming existing methods in capturing residue constraints validated by deep mutational scanning. The CLM combined fine-tuned ESM-2 and vBERT-based classifiers, and revealed that key adaptive mutations in receptor binding domains (RBD) epistatically coupled with overlooked mutations beyond RBD. This framework predicted 100% of experimentally validated high-frequency binding sites and identified natural cross-species transmission mutations. Specifically, CLM predicted elevated human-adaptive risk in 2010-2020 H1N1 and post-2020 H5N1 IAVs, linked to the 2009 pandemic and recent North American bovine spillover events. This work established a novel predictive framework for identifying hidden IAV genotypes potentially important for pandemics.