Fig. 5. Combinatorial space between F1*U and F1*U^m contains many neutral mutants predictable by deep learning.

a Fraction of neutral mutants in the combinatorial library between the F1*U and F1*U^m (WT/Mut) and in generation 1 according to different RA thresholds of neutrality. A total of 10⁶ unique mutational paths were randomly sampled that could transform the WT sequence to F1*U^m in 16 direct mutational steps. Paths were considered neutral if none of the steps resulted in a mutant with RA lower than the neutrality threshold. b Fraction of neutral mutants at each Hamming distance in the WT/Mut library as measured by experimental assay (yellow). The blue area indicates the fraction of the neutral mutants that was also identified by MLP trained on data from generations 1–7. c Performance metrics of MLP when used to classify the WT/Mut library.