Speaking the Language of the Spike: Predicting SARS-CoV-2 Viral Fitness Using Protein Language Models

Abstract

The rapid evolution of SARS-CoV-2 has highlighted the critical need for computational tools capable of forecasting viral fitness and variant emergence. New viral lineages often appear and spread faster than our ability to respond, preventing timely implementation of updated vaccines, therapeutics, or policy interventions. While the pandemic has advanced global genomic surveillance capabilities, current variant prediction methods face substantial limitations, creating a need for scalable models that can rapidly assess viral fitness without relying on slow experimental pipelines. Protein language models (PLMs) offer a promising approach for predicting mutational consequences without structural information or labor-intensive experiments. Prior work demonstrated that PLM embedding distances can predict immune escape for individual single-residue mutations, validated against experimental deep mutational scanning data. In contrast, we investigate whether embedding distances can predict relative fitness change between whole lineages using viral genome tracking data. We analyzed ESM-2 embedding distances and compared them to Levenshtein distance as predictors of growth advantage between parent and child SARS-CoV-2 lineages across 44 time windows spanning the duration of the pandemic. Both methods identified identical significant windows during early 2021 B.1 lineage evolution and showed significantly correlated temporal trends, but showed no significant global effects throughout all periods. Our findings suggest that PLMs are capable of capturing fitness-relevant signals comparable to established metrics during periods of simpler evolutionary dynamics, while revealing fundamental limitations of distance-based strategies for complex viral evolution. These results provide insights for developing more sophisticated forecasting frameworks for rapidly evolving pathogens in the future.