Boltz2ESI: Accurate Enzyme Specificity Prediction with Co-folding Foundation Model

Xiwei Cheng, Seonghwan Seo, Jihang Chen, Songlin Jiang, Pengkang Guo, Wengong Jin

Keywords: Enzyme–substrate specificity prediction, Co-folding foundation models

Abstract: Predicting enzyme-substrate specificity is a fundamental challenge in biocatalysis and enzyme engineering. However, the scarcity of experimentally labeled data often limits the ability of existing methods to generalize across diverse enzyme families. Here, we present Boltz2ESI, an end-to-end framework based on the co-folding foundation model Boltz-2 for accurate enzyme-substrate interaction prediction. Our approach integrates a local spatial representation of the enzyme-substrate interface with corresponding enzyme evolutionary context provided by ESM3. On ESIBank benchmark, Boltz2ESI achieves state-of-the-art performance on four out of six evaluation sets, reaching an average AUROC of 0.7309. In particular, our approach shows significant improvements in data-scarce enzyme families, demonstrating robust generalization capabilities. Finally, we conduct extensive ablation studies to analyze the impact of each component on the enzyme-substrate interaction prediction.

Submission Number: 33