Related Experiment Video
Updated: Aug 6, 2026
14:37
Modeling an Enzyme Active Site using Molecular Visualization Freeware
Published on: December 25, 2021
EPIC: A Machine-Learning Framework for Product-Dependent Behavior in β‑Glucosidases
Ali Malli1, Denys Vasyutyn1, Anuj Majumder1
1Department of Chemical and Biomolecular Engineering, New York University, 6 MetroTech Center, Brooklyn, New York 11201, United States.
ACS Omega
|July 24, 2026
Summary
A new machine learning framework, EPIC, predicts enzyme activity influenced by product interactions. This tool aids in enzyme mining and engineering by modeling glucose-dependent activity profiles for beta-glucosidases (BGLs).
Area of Science:
- Biochemistry and Molecular Biology
- Computational Biology and Bioinformatics
- Enzyme Kinetics and Engineering
Background:
- Enzyme activity is significantly modulated by interactions with small molecules, particularly reaction products.
- Product accumulation near enzymes can lead to diverse and complex enzyme response behaviors.
- Current computational tools for modeling enzyme-product interactions are limited, and experimental characterization is resource-intensive.
Purpose of the Study:
- To develop a machine learning framework, Enzyme-Product Interaction Classifier and Curve Predictor (EPIC), for predicting enzyme activity profiles.
- To model glucose-dependent relative activity of beta-glucosidases (BGLs) based on amino acid sequence and assay conditions.
- To establish a scalable computational tool for enzyme mining and engineering by characterizing enzyme-product interactions.
Main Methods:
- Curated a dataset of 105 unique BGL sequences for training and validation.
- Formulated the prediction problem as both a classification task (response classes) and a regression task (activity-concentration profiles).
- Employed machine learning algorithms to predict enzyme response behaviors from sequence and assay data.
Main Results:
- EPIC achieved a balanced classification performance with an F1-score of 0.577 ± 0.022 and accuracy of 0.594 ± 0.024, outperforming a sequence-identity baseline.
- EPIC significantly improved the prediction of glucose concentration-dependent activity changes, capturing monotonic and nonmonotonic trends (Spearman's ρ = 0.659 ± 0.025).
- The framework demonstrated generalization capabilities to unseen sequences and diverse response behaviors, including ancestral enzymes.
Conclusions:
- EPIC provides a scalable computational framework for modeling relative enzyme activity influenced by enzyme-product interactions.
- The study highlights the complementary nature of classification and regression approaches in capturing complex enzyme response behaviors.
- EPIC facilitates enzyme mining and engineering efforts by offering a predictive tool for enzyme-product interaction dynamics.
