Dimitris Kalafatis

Predicting protein-metal interactions using HoloFold
With the advent of novel protein folding prediction software like AlphaFold3, it is becoming increasingly easier to guide experiments based on predicted protein domains, interacting partners, or ligands. However, predicted ligand binding has high false positive rates, and models have difficulty distinguishing different metal binding sites in vivo. To solve this problem, we developed HoloFold, a machine-learning model that incorporates evolutionary, structural, and ligand-specific information to accurately and quickly predict protein-ligand interactions. Holofold has been shown to have extremely high recall even with rare metals, greatly outperforming AlphaFold3 in predicting de novo metalloproteins. By applying HoloFold to the human and yeast proteomes for all metals common in biology, we plan to curate a high-confidence metalloproteome with important implications for disease and new biology, as well as providing similar predictions in a simpler model system to confirm evolutionarily conserved interactions. Specifically, by using HoloFold to predict mitochondrial protein interactions with copper, we have identified novel proteins that likely bind copper in vivo, changing what we know about mitochondrial biology.