Integration of Machine Learning Improves the Prediction Accuracy of Molecular Modelling for M. jannaschii Tyrosyl-tRNA Synthetase Substrate Specificity

Abstract

AbstractDesign of enzyme binding pocket to accommodate substrates with different chemical structure is a great challenge. Traditionally, thousands even millions of mutants have to be screened in wet-lab experiment to find a ligand-specific mutant and large amount of time and resources is consumed. To accelerate the screening process, here we propose a novel workflow through integration of molecular modeling and data-driven machine learning method to generate mutant libraries with high enrichment ratio for recognition of specific substrate. M. jannaschii tyrosyl-tRNA synthetase (Mj. TyrRS) is used as an example system to give a proof of concept since the sequence and structure of many unnatural amino acid specific Mj. TyrRS mutants have been reported. Based on the crystal structures of different Mj. TyrRS mutants and Rosetta modeling result, we find D158G/P is the critical residue which influences the backbone disruption of helix with residue 158-163. Our results show that compared with random mutation, Rosetta modeling and score function calculation can elevate the enrichment ratio of desired mutants by 2-fold in a test library having 687 mutants, while after calibration by machine learning model trained using known data of Mj. TyrRS mutants and ligand, the enrichment ratio can be elevated by 11-fold. This molecular modeling and machine learning-integrated workflow is anticipated to significantly benefit to the Mj. tyrRS mutant screening and substantially reduce the time and cost of web-lab experiment. Besides, this novel process will have broad application in the field of computational protein design.CCS Concepts• Applied computing • Life and medical sciences • Computational biology • Molecular structural biology