A Unified Benchmark and Conditional Sequence Modeling Framework for Future Influenza Evolution Prediction

Vaccination remains the most effective strategy for seasonal influenza prevention, but vaccine strain selection depends on anticipating future circulating variants before they are fully observed. This is challenging because influenza viruses evolve rapidly and existing studies often use heterogeneous datasets, targets, and evaluation settings. We formulate future influenza evolution prediction as a conditional sequence modeling problem and present a unified benchmark, with an initial release on influenza A/H1N1 hemagglutinin (HA). The benchmark distinguishes raw and aligned sequence layers, separates hard and soft target definitions, and adopts leakage-safe rolling evaluation. We instantiate the benchmark with a representative conditional seq2seq framework that combines pretrained protein representations, prevalence signals, and a compact multi-objective training objective. The resulting study provides a reproducible basis for systematic comparison in future influenza forecasting research.