Abstract
Anticipating how mutations change viral fitness is central to genomic surveillance and vaccine design, yet the supervised phenotype data behind the most accurate variant-effect predictors are unavailable for most emerging pathogens. We ask how far label-free scoring can go using only sequences, their evolutionary history, and structure. We assemble a modular, fully unsupervised pipeline that estimates a few interpretable terms (intrinsic replicative fitness, antigenic escape, and realized growth), and that lets each term be produced by more than one estimator, so the estimator itself becomes a testable modeling choice. Benchmarking the intrinsic term on 21 viral deep-mutational-scanning assays from ProteinGym, we find that a 650-million-parameter single-sequence protein language model predicts viral mutational fitness weakly and heterogeneously (mean Spearman
0.15
), whereas a trivial site-independent alignment model more than doubles it (
0.39
, better on 17 of 21 assays), with the largest gains on the antigenic surface proteins where the language model fails. Yet the ordering reverses across 186 non-viral ProteinGym assays, where the language model instead exceeds the alignment model, localizing the weakness to viral families under-represented in the model's training data. Alignment-conditioned language models (MSA Transformer, Tranception) recover this accuracy but do not clearly exceed the simple alignment, so the decisive feature is the family alignment, not model scale or architecture. Our central result is evolutionary. Using dated samples of SARS-CoV-2 spike and influenza H3N2 hemagglutinin, we show that the epoch of the alignment is itself a leading, virus-specific determinant of accuracy. This traces to non-stationary selection: the site-specific amino-acid preferences drift over time, abruptly for spike at the emergence of Omicron and gradually for H3N2 hemagglutinin. A phylogenetic mutation-selection estimator does not match the far cheaper alignment model, falling significantly below it on matched data. Unsupervised viral fitness prediction is, then, as much an evolutionary problem as a modeling one.