(a). Loss-trajectories over the course of evotuning for MS2 and PETase. (b). Improvement over random sampling for ability to predict the relative fitness rankings of test-set mutants; datasets include MS2 single-mutants, MS2 double-mutants, TEM-1 β-lactamase single-mutants, and PETase single-mutants. (c). Comparison of top-model predictions of MS2 double-mutant fitness when trained on MS2 single-mutant dataset. Top models are trained using global UniRep, eUniRep, and one-hot sequence representations. Predicted fitness values are compared to experimentally measured fitness, and additive fitness-change predictions based on single-mutant differences from wild-type fitness.