CIST: A clade-informed sequence transformer framework for predicting influenza virus antigenicity

Rapid antigenic evolution of influenza viruses complicates timely vaccine strain selection and renders prospective interpretation of serological assays-including neutralization (NT) and hemagglutination inhibition (HI)-increasingly challenging in the face of a continuously shifting antigenic landscape. Here, we propose CIST, a clade-informed deep-learning-based framework for predicting influenza antigenicity directly from paired virus-vaccine or virus-antiserum hemagglutinin (HA) sequences. CIST leverages masked-language-model pre-training on 70,425 influenza A HA sequences to learn context-aware residue representations and integrates explicit evolutionary clade embeddings reinforced by a contrastive feature enhancement objective. The resulting sequence representations are processed by a local-global dual-attention encoder to regress both NT values and HI titers. Across two H3N2 antigenicity prediction tasks, CIST outperformed traditional machine-learning baselines and a standard Transformer, achieving a mean squared error (MSE)/mean absolute error (MAE)/root mean squared error (RMSE) of 0.0014/0.0141/0.0374 for NT prediction compared with 0.0029/0.0210/0.0538 for the Transformer, and 0.0059/0.0378/0.0768 for HI prediction, corresponding to error reductions of up to 51.7% in MSE for NT and 11.9% for HI over the Transformer baseline. Pearson correlation coefficients of 0.97 (NT) and 0.82 (HI) further confirm the quantitative agreement between predicted and observed titers. Attention-based interpretability analyses further revealed that high-contribution residues were enriched in known antigenic regions of HA, including classical antigenic sites A-E, supporting the biological plausibility of the learned features. Additional zero-shot temporal external validation on emerging K subclade viruses (n = 36 virus-antiserum pairs, absent from training) showed that CIST achieved the lowest prediction error among the evaluated methods (MSE = 0.092, Pearson r = 0.62), whereas RF showed limited correlation with observed titers (r = 0.25), supporting the potential utility of clade-informed deep learning for antigenic surveillance of emerging strains. Together, these results demonstrate that clade-informed deep learning enables accurate and interpretable computational assessment of influenza antigenicity, providing a practical complement to laboratory assays for antigenic surveillance and vaccine evaluation.