EPIC: A Machine-Learning Framework for Product-Dependent Behavior in β‑Glucosidases

Enzyme activity is often modulated by interactions with small molecules. As molecules that can accumulate in the vicinity of enzymes, products can directly influence enzymatic activity and give rise to diverse enzyme response behaviors. Despite this biological importance, relatively few computational tools explicitly model enzyme-product interactions and characterizing such interactions typically relies on resource intensive experimental assays, thereby limiting enzyme mining and engineering efforts. Here, we present Enzyme-Product Interaction Classifier and Curve Predictor (EPIC), a machine learning framework that predicts glucose-dependent relative activity profiles of β-glucosidases (BGLs) from information on amino acid sequence and assay conditions. We curated a dataset of 105 unique BGL sequences for glucose response prediction. We formulate this problem as both a classification task, assigning enzymes to different product response classes, and a regression task, modeling full relative activity-glucose concentration profiles. Across extensive cross-validation, EPIC demonstrates more balanced classification performance across all glucose-response classes than a naïve sequence-identity-based baseline, attaining an F ₁-score of 0.577 ± 0.022 and an accuracy of 0.594 ± 0.024. EPIC substantially outperforms this baseline in predicting glucose concentration-dependent activity changes, capturing both monotonic and nonmonotonic response trends (Spearman's ρ = 0.659 ± 0.025). Using ancestral and other enzymes beyond the original dataset, EPIC generalizes to unseen sequences and a range of response behaviors. Together, these results establish EPIC as a scalable framework for modeling relative enzyme activity in the context of enzyme-product interactions. Our study also demonstrates that classification and regression offer complementary perspectives, capturing the complex glucose-response behaviors of BGLs.