Binding prediction of multi-domain cellulases with a dual-CNN

Abstract

Cellulases hold great promise for the production of biofuels and biochemicals. However, they are modular enzymes acting on a complex heterogeneous substrate. Because of this complexity, the computational prediction of their catalytic properties remains scarce, which restricts both enzyme discovery and enzyme design. Here, we present a dual-input convolutional neural network to predict the binding of multi-domain enzymes. This regression model outperformed previous molecular dynamics-based methods for binding prediction for cellulases in a fraction of the time. Also, we show that when changed to a classification problem, the same network can be back-propagated to suggest mutations to improve enzyme binding. A similar approach could increase our understanding of the structure-activity relationship of enzymes, and suggest new promising mutations for enzyme design using explainable artificial intelligence.

Recent Advances in β-Glucosidase Sequence and Structure Engineering: A Brief Review

“…A data-driven approach of ML combined with statistics overcame this disadvantage by inferring the numerous and possibly unknown factors which map from sequence to function according to the above data and provided superior predictive accuracy for predicting mutation sites in BGL [ 66 ]. For example, a regression model based on a dual-input convolutional neural network was used to predict the binding affinity of cellulase to the substrate to improve enzyme activity [ 67 ]. Due to the lack of negative sequence examples in the DMS dataset, and the inability to learn directly from the large-scale sequence function DMS dataset with the ML-supervised method, a method was developed to classify the DMS dataset as positive unlabeled data and successfully applied to design thermally stable BGLs [ 68 ].…”