Statistical Models of the Protein Fitness Landscape: Applications to Protein Evolution and Engineering

Abstract

Understanding the protein fitness landscape is important for describing how natural proteins evolve and for engineering new proteins with useful properties. This mapping from protein sequence to protein function involves an extraordinarily complex balance of numerous physical interactions, many of which are still not well understood. Directed evolution circumvents our ignorance of how a protein’s sequence encodes its function by using iterative rounds of random mutation and artificial selection. The selection criteria is based on experimental measurements, which permits the optimization of protein sequence properties that are not understood. While directed evolution has been useful for exploring protein fitness landscapes, these searches have been relatively local in comparison to the vast space of possible protein sequences. Here, we present several classes of statistical models that map protein sequence space on a larger scale. We use these simple models to interpret data from SCHEMA recombination libraries, understand the evolutionary benefit of intragenic recombination, and design optimized protein sequences. By training on directly on experimental data, these models implicitly capture the numerous and possibly unknown factors that shape the protein fitness landscape. This provides an unrivaled quantitative accuracy across a massive number of protein sequences.