VirPLM: Antigenic prediction of influenza A/H3N2 viruses with a fine-tuned protein language model

Human influenza A/H3N2 viruses undergo rapid antigenic evolution primarily driven by the hemagglutinin subunit 1 (HA1). Within HA1, amino acid substitutions under immune pressure cause antigenic drift, necessitating frequent updates to vaccine strains. While hemagglutination inhibition (HI) assays remain the gold standard for assessing antigenic relationships, their labor-intensive and low-throughput nature limits scalability. Fortunately, the rapid accumulation of HA1 sequences enables sequence-based antigenic prediction, yet effectively extracting informative representations from these viral sequences remains challenging.

In this study, we present VirPLM, a two-stage framework that adapts the ESM-2 protein language model to H3N2 HA1 sequences for antigenic prediction. VirPLM significantly outperforms representative methods and maintains robust performance under both cross-validation and retrospective time-split evaluations. Moreover, VirPLM identifies highly critical sites enriched in known regions related to antigenic evolution. In the season-specific coverage analysis, VirPLM-prioritized strains achieve higher estimated coverage rates than the corresponding historical strains recommended by the World Health Organization in most evaluated seasons, suggesting that VirPLM can provide complementary sequence-based evidence for candidate strain prioritization.

The source code is available at https://github.com/xingyili/VirPLM, and the version used in this study is archived in Zenodo (DOI: 10.5281/zenodo.21650323).

Supplementary information is available at Bioinformatics online.