(254a) Machine Learning Approaches for Enzyme Engineering

Enzyme engineering aims to improve a variety of different properties, such as stability, activity, and binding affinity. A variety of experimental and computational approaches have been used to improve such properties, including directed evolution and structure-guided rational design using various tools (e.g., Rosetta). We have recently developed a machine learning approach (called MLProScape) to model the protein fitness landscape and to guide enzyme engineering efforts. MLProScape numerically captures amino acid based physio-chemical properties as features, builds a machine learning model based on these features, and then identifies synthetic designs with improved enzyme properties. We have applied MLProScape to successfully design novel variants of glycosyl hydrolases with improved catalytic activity for hydrolyzing both cellulose and xylan, two major components found in lignocellulosic biomass. In this case, MLProScape made highly accurate predictions for glycosyl hydrolase activity under 5 different conditions, with pearson correlation coefficients varying between ~0.72 and ~0.90 and root mean squared errors varying between ~0.4 and ~16.2 mmol/min/mg. Examples of other MLProScape protein engineering applications will also be presented. MLProScape is capable of identifying residue changes outside of the active site and can handle complex design criteria that aim to improve multiple enzyme properties, simultaneously. As such, this machine learning based approach will facilitate enzyme engineering efforts aimed at improving various properties, including activity and thermostability.