Efficient searches in protein sequence space through AI-driven iterative learning

Protein sequence space is vast. This fact, together with the prevalence of epistasis, hampers the engineering of novel enzymes through library screening and is a major obstacle to any attempt to predict natural protein evolution. Recently, specialized methodologies have been used to determine fitness data on ∼260000 sequences for the gene of the enzyme dihydrofolate reductase and antibody affinity data for all combinations of the mutations present in the receptor binding domain (RBD) of the Omicron strain of SARS-CoV-2 (∼30000 variants). We show that, upon iterative training on a total of just a few hundred variants, various state-of-the-art AI tools (multi-layer perceptron, random forest and XGBoost algorithms) find very high fitness variants of the enzyme and predict the antibody-evasion patterns of the RBD. This work provides a basis for efficient, widely applicable, low-throughput experimental approaches to assess viral protein evolution and to engineer enzymes for biotechnological applications.