Amino acid clustering and epistasis guide sequence-based prediction of phenotype modulation

Protein sequences encode the molecular information necessary for their biological function. As functional adaptation often emerges from combined effects of multiple positions, understanding how sequence variation gives rise to functional differences remains a central challenge. Here, we developed an improved strategy to identify phenotype modulating positions and use them to predict phenotypes, by relating amino acid identity to experimentally measured quantitative properties across families. Using family-specific sequence-phenotype datasets, our results show that epistasis plays a crucial role in phenotype modulation. Position specific amino acid clustering further improved performance indicating that key functional information is often captured by broad physicochemical amino acid classes, with hydrophobicity playing a crucial role in phenotype modulation. Our identified phenotype modulating positions involve both proximal and distal sites relative to functional sites. While the former is captured by individual positions, the latter is captured when epistasis is taken into account. Comparison with experimentally identified positions showed partial recovery of known functionally relevant residues, while at the same time revealing candidate positions that may contribute to phenotypic variation. Taken together, these results suggest that quantitative phenotypes are often governed by a network of interacting amino acid residues and may be engineered by a limited amount of sequence modifications. This finding highlights the potential of our framework to deepen our understanding of protein function, evolution, and adaptation, providing new targets for the rational modulation and engineering of protein phenotypes.