MECE: a method for enhancing the catalytic efficiency of glycoside hydrolase based on deep neural networks and molecular evolution

Abstract

Abstract High efficiency glycoside hydrolases (GH) are in high demand for numerous industrial applications. This study demonstrates the use of a deep neural network and molecular evolution (MECE) platform for predicting catalysis-enhancing mutations in GHs. The MECE platform integrates a deep learning model (DeepGH), trained with 119 GH family protein sequences from the CAZy database. Ten-fold cross-validated DeepGH models showed 96.73% predictive accuracy. MECE also includes a quantitative mutation design component that uses Grad-CAM with homologous protein sequences to identify key features for mutation in the target GH. Validation of the MECE platform with chitosanase CHIS1754 and glucoamylase GA51, resulted in generation of CHIS1754-MUT7, harboring seven amino acid conversions, and GA51-MUT5, carrying five residue conversions. The kcat/Km of CHIS1754-MUT7 was 18.08-fold higher than CHIS1754, while GA51-MUT5 was 7.64-fold greater than that of GA51. This resource can facilitate the rational design of catalytically efficient enzymes for a broad range of applications.