AI-Driven Prediction of Enzyme Structure and Function and Its Applications

Abstract

Enzymes are chemical catalysts that are essential in all the biochemical changes that occur in the biological world. Their catalytic activities are closely connected with the three-dimensional (3D) organization, which terms active sites, substrate-binding pockets, and other mechanically relevant conformational states. Protein sequences have been sequenced affordably at an unprecedented rate using high-throughput methods; but experimentally measured structures and high-quality functional annotation are relative rarities, resulting in an ongoing gap in the so-called sequence to structure to functional gap. In recent years (in the last five years), the deep-learning branch of artificial intelligence (AI) brought disruptive improvements that created a significant undercut in this gap. The structure predictors of deep learning structure predictors such as AlphaFold2 have demonstrated high-confidence structural models of a large number of proteins, including enzymes, leading to lower time and resource costs associated with structure determination. Simultaneously, prediction of enzyme functions has moved away both homology-driven transfer prediction and manually hand-crafted features, to representation learning with sequences and structures, such as protein language models, graph neural networks, and contrastive learning models, but also integrative (multimodal) prediction with the combination of complementary signals. The review presents the latest developments in AI-based enzyme structure prediction and AI-based enzyme function prediction and describes its scientific and practical importance introduced by such developments. We emphasize such representative uses as enzyme engineering, synthetic biology, and drug discovery and development. We also specify outstanding constraints like the prediction of multimeric assemblies, conformational heterogeneity and dynamics, and realistic enzyme-ligand interfaces, and the outlook of future directions through additional training information, combining physical constraints, and close integration of AI prediction with experimental justification.