Identifying structural and dynamic features of proteins for machine learning models that predict rates of energy transfer

Abstract

The flow of vibrational energy in proteins is mediated both by protein structure and structural fluctuations. However, detailed information about contributions of physical properties that mediate energy flow is still lacking. In this report, we present analyses of more than 105 rates of energy transfer across non-covalent contacts of a G-protein coupled receptor (GPCR) and several mutants using models that account for specific structural and dynamic features. Two models are considered, one including only structural features and a second including structural and dynamic properties. While dynamic properties are found to play a substantial role in energy transfer across non-covalent contacts, a structural model with appropriately chosen physical features can predict rates of energy transfer about as well as models incorporating both structural and dynamic features, apparently due to correlations between structural fluctuations and structure, which are also examined. The work presented here indicates that energy flow in GPCRs and likely other proteins may be predicted using relatively simple models trained with structural information.