Protein Fitness Prediction

The field predicts how mutations change protein properties, using measurements from deep mutational scanning and directed evolution to guide experiments such as enzyme optimization and viral variant assessment. Common approaches score sequences with pretrained protein language models, train assay-specific predictors, or combine sequence information with structure, dynamics, and evolutionary evidence. Researchers evaluate these methods on benchmarks such as ProteinGym and FLIP, using both held-out variant prediction and budgeted searches for high-fitness candidates. A central challenge is that strong performance on random splits or familiar assays may not transfer to future experimental rounds, deeper mutations, or new proteins, where data are sparse and epistasis matters.