Machine-Learning Prediction\nof Protein Function from\nthe Portrait of Its Intramolecular Electric Field

We introduce a machine learning framework designed to\npredict enzyme\nfunctionality directly from the heterogeneous electric fields inherent\nto protein active sites. We apply this method to a curated data set\nof heme-iron oxidoreductases, spanning three enzyme classes: monooxygenases,\nperoxidases, and catalases. Conventional analysis, focused on simplistic,\npoint electric fields along the Fe–O bond, is shown to be inadequate\nfor accurate activity prediction. Our model demonstrates that the\nenzyme’s heterogeneous 3-D electric field, alone, can accurately\npredict its function, without relying on additional protein-specific\ninformation. Through feature selection, we uncover key electric field\ncomponents that not only validate previous studies but also underscore\nthe crucial role of multiple components beyond the traditionally emphasized\nelectric field along the Fe–O bond in heme enzymes. Furthermore,\nby integrating protein dynamics, principal component analysis, clustering,\nand QM/MM calculations, we reveal that while dynamic complexities\nin protein structures can obscure predictions, the model still retains\nits accuracy. This research significantly advances our understanding\nof how protein scaffolds possess signature electric fields tailored\nto their functions at the active site. Moreover, it presents a novel\nelectrostatics-based tool to harness these signature electric fields\nfor predicting enzyme function.