A structure-aware generative AI framework for revealing functional relationships in protein families

Abstract

Significance Protein databases contain millions of rapidly accumulating sequences, while increasingly accurate 3D structure predictions provide unprecedented data for analysis. Predicted structures can also be encoded using a structure-focused alphabet alongside standard amino acid (AA) representations. We introduce a framework that treats the AA and 3D interaction (3Di) alphabets as parallel views of a protein family, jointly navigated using a quantitative measure of coevolutionary fitness. By learning sequence landscapes defined by coevolutionary energies and quantifying their correspondence, we predict when AA and 3Di views overlap and when they provide independent information. Together, these representations yield a unified view of structural conservation and functional groupings across diverse protein families.