Abstract

The tempo of viral adaptation is usually read indirectly from the composition of mutations, through measures such as dN/dS. Here we measure it directly from the dynamics of variant frequencies, where we use multinomial logistic regression to estimate a fitness for each co-circulating variant. We aggregate these estimates to derive the rate of change of mean population fitness, referred to as fitness flux. Tracing SARS-CoV-2 from 2020 to 2025 and comparing against seasonal influenza A/H3N2, we find that SARS-CoV-2 adapted rapidly with a 6.8-fold increase in fitness from 2020 to 2023, before slowing to a 2.2-fold increase from 2023 to 2025. Influenza H3N2 sustains a slower, steadier pace roughly threefold below recent SARS-CoV-2. In both, the rate of fitness gain closely tracks the variance in fitness, matching the 1:1 expectation of Fisher’s fundamental theorem. Phylogenetic contrasts between parent and child lineages localize most fitness gain to spike, and within spike to the receptor-binding domain, where a simple count of spike S1 substitutions predicts lineage fitness about as well as deep-learning escape and protein-language-model scores. Measuring fitness directly thus offers a transparent, frequency-based alternative to mutational proxies for tracking and anticipating viral adaptation.

The website blab.github.io/fitness-flux/ is the intended reading experience of this paper, providing responsive layout and interactive figures.