Local structural preference maps encode transferable protein-interface energetics

Structural preference maps derived from protein interfaces can effectively encode transferable molecular energetics, even without direct thermodynamic labels. Researchers have developed a machine-learning model that learns energetically relevant interaction preferences directly from native structures, without the need for affinity or mutation data.

By decomposing interfaces into local structural motifs and assessing the compatibility of each motif with its surrounding molecular environment, these models generate target preference maps (TPMs). These TPMs are spatial fields of local interaction compatibility that can be applied across a protein surface.

When trained exclusively on native antibody-antigen structures, the TPMs successfully recover charged, aromatic, and backbone-mediated recognition patterns across peptide-protein and other non-antibody interfaces. In an out-of-domain SKEMPI benchmark, wild-type TPM scores demonstrated the ability to discriminate mutation-sensitive positions, achieving an AUC of 0.643 from a total of 1,323 data points.

Furthermore, TPM-based substitution scores ranked alternative amino acids comparably to more established methods like FoldX and Rosetta Flex {Delta}{Delta}G. These findings indicate that native interface structures can effectively supervise the learning of transferable, energetically relevant molecular preferences, all without direct thermodynamic labels.