Deep Learning in Enzyme Function Prediction and Novel Enzyme Discovery
TL;DR
This review summarizes recent progress with an emphasis on benchmark design rather than reported scores alone, and compares representative methods by input modality, training data, split strategy, leakage control, evaluation metric, and validation evidence.
Abstract
The rapid growth of genomic and metagenomic data has created a large gap between protein sequence discovery and enzyme functional characterization. Deep learning has become an important tool for enzyme-related prediction tasks, including EC and GO annotation, substrate specificity, catalytic-residue identification, kinetic-parameter estimation, thermostability, pH optima, and candidate discovery. This review summarizes recent progress with an emphasis on benchmark design rather than reported scores alone. We compare representative methods by input modality, training data, split strategy, leakage control, evaluation metric, and validation evidence. Across tasks, apparent performance gains can arise from dataset scale, annotation quality, homolog overlap, negative sampling, or endpoint definition rather than architecture alone. Current models are valuable for annotation and candidate prioritization, but robust discovery of remote homologs, rare functions, or new catalytic activities still requires low-identity or out-of-distribution evaluation and experimental validation.