Conserved influenza A epitope candidate regions and a benchmark of ESM-2 sequence features

Influenza A virus antigenic drift drives the need for annual vaccine reformulation, leading researchers to seek out conserved epitope candidates that may support broadly protective vaccines. A study screened influenza A virus sequences from H1N1, H3N2, and H5N1, across nine viral proteins, to identify 98 conserved candidate regions.

Thirty-eight of these regions were identical across the H1N1, H3N2, and H5N1 consensus sequences, all located in the polymerase complex and nucleoprotein (PB2, PB1, PA, NP). The remaining ten surface-glycoprotein (HA/NA) candidates were specific to each subtype.

Two protein-language-model (ESM-2) features were benchmarked against alignment conservation. The group-masked log-probability showed a moderate correlation with MSA conservation (Spearman rho = 0.25 to 0.39 for HA), but did not provide any additional value for T-cell epitope discrimination. The change in AUROC was only +0.004, with a p-value of 0.46, indicating no significant improvement. Attention-derived contact-density was also found to be an unreliable proxy for solvent-accessibility.

A curated antibody-epitope benchmark, consisting of 22 clusters and 5 neutralization-supported entries, was found to be underpowered for a high-confidence B-cell test. The study highlights data-quality and reproducibility issues, such as length heterogeneity, coordinate mapping, and pseudoreplication, which could affect the reliability of the results. To address these concerns, the researchers have released an auditable benchmark for further investigation.

The findings reveal that ESM-2 sequence scores did not enhance epitope prioritization beyond alignment-derived conservation in the evaluated benchmarks. This study provides an auditable candidate resource for researchers and demonstrates that, in the examined benchmarks, ESM-2 sequence scores offer no significant advantage in epitope prioritization compared to alignment-derived conservation.