BioStructNet: Structure-Based\nNetwork with Transfer\nLearning for Predicting Biocatalyst Functions

Enzyme–substrate interactions are essential to\nboth biological\nprocesses and industrial applications. Advanced machine learning techniques\nhave significantly accelerated biocatalysis research, revolutionizing\nthe prediction of biocatalytic activities and facilitating the discovery\nof novel biocatalysts. However, the limited availability of data for\nspecific enzyme functions, such as conversion efficiency and stereoselectivity,\npresents challenges for prediction accuracy. In this study, we developed\nBioStructNet, a structure-based deep learning network that integrates\nboth protein and ligand structural data to capture the complexity\nof enzyme–substrate interactions. Benchmarking studies with\ndifferent algorithms showed the enhanced predictive accuracy of BioStructNet.\nTo further optimize the prediction accuracy for the small data set,\nwe implemented transfer learning in the framework, training a source\nmodel on a large data set and fine-tuning it on a small, function-specific\ndata set, using the CalB data set as a case study. The model performance\nwas validated by comparing the attention heat maps generated by the\nBioStructNet interaction module with the enzyme–substrate interactions\nrevealed from molecular dynamics simulations of enzyme–substrate\ncomplexes. BioStructNet would accelerate the discovery of functional\nenzymes for industrial use, particularly in cases where the training\ndata sets for machine learning are small.