A moving target: non-stationary selection governs unsupervised prediction of viral fitness

Anticipating how mutations change viral fitness is central to genomic surveillance and vaccine design. However, the phenotypic data behind the most accurate supervised variant-effect predictors is unavailable for most emerging pathogens. It is therefore unclear how a label-free, unsupervised approach would perform using only sequences, their evolutionary history, and their protein structure. To address this outstanding question, we assemble a modular, fully unsupervised pipeline that estimates key interpretable terms: intrinsic replicative fitness, antigenic escape, and realized growth. Each term is produced by more than one estimator, so that the fitness predictor becomes a testable model 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 (PLM) is a weak predictor of viral fitness, whereas a simple site-independent alignment model more than doubles performance, with the largest gains obtained for the antigenic surface proteins not included in the training of the PLM. On the other hand, PMLs based on protein alignment (MSA Transformer, Tranception) regain the accuracy lost with the single-sequence PML, but do not significantly exceed the simple alignment. Therefore, the most important feature to predict viral fitness is the protein alignment, not model scale or architecture. Our central result is evolutionary: using dated sequences of SARS-CoV-2 spike and influenza H3N2 hemagglutinin, we find that the epoch of the alignment is itself a virus-specific determinant of accuracy. This demonstrates that selection is a non-stationary process, in a manner that is itself virus-specific. A phylogenetic mutation-selection estimator of fitness does not match the far simpler and less computationally expensive alignment model. Altogether, this work shows that unsupervised viral fitness prediction is as much an evolutionary problem as a modeling one.