FluEvoFormer: A Structure-Guided Generative Foundation Model for Prospective Influenza Antigenic Evolution and Vaccine Strain Selection

Seasonal influenza vaccine strain selection remains challenging because circulating viruses may drift after vaccine recommendations are made. This study presents FluEvoFormer, a structure-guided generative foundation model for prospective influenza antigenic evolution forecasting and vaccine strain ranking. The framework jointly encodes hemagglutinin and neuraminidase sequences and incorporates residue-contact graphs. Separate prediction heads estimate future viral dominance and the vaccine–virus antigenic match. Controlled future-like variant stress testing, uncertainty adjustments, and clade balancing are then used to rank vaccine candidates. A rolling retrospective evaluation was performed for target seasons involving the influenza A(H1N1)pdm09 virus and influenza A(H3N2) virus. The evaluation used cutoff-restricted sequence records, hemagglutination inhibition data, vaccine-composition records, protein-structure resources, and vaccine-effectiveness indicators. The historical training corpus for the influenza A(H1N1) virus also contained pre-2009 seasonal records. FluEvoFormer achieved the lowest held-out antigenicity prediction error, with mean absolute error (MAE) values of 0.389 for the combined historical influenza A(H1N1) virus corpus and 0.456 for the influenza A(H3N2) virus corpus. It also improved the future dominance prediction, with Kullback–Leibler (KL) divergence values of 0.255 and 0.289, respectively. The model selected candidates with higher empirical normalized coverage scores in seven out of 10 influenza A(H1N1)pdm09 virus seasons and nine out of 10 influenza A(H3N2) virus seasons. The predicted coverage score showed a strong positive correlation with external vaccine-effectiveness estimates. These findings support FluEvoFormer as a computational decision-support framework for prioritizing influenza vaccine candidates before downstream laboratory and public health evaluations.