Fig. 1. Design of protein sequences using structural information alone and assessment of designed sequences using surrogate fitness functions trained on experimental observations.

a Structural environment predicts residue mutation preferences and can be used for designing combinatorial protein variants. b Learning supervised fitness functions from experimental high-throughput variant measurements. The functional form of the fitness function, f(.), can be learned by fitting to observed data, and enables predicting the function of unobserved sequence variants. c Assessment of designed sequences with the surrogate fitness function enables comparing different sequence design strategies. Source data are provided as a Source Data file.