Enzyme Promiscuity Prediction Using Hierarchy-Informed Multi-Label Classification.

Area of Science:

• Biochemistry
• Computational Biology
• Machine Learning

Background:

• Experimental characterization of enzyme capabilities is costly and time-consuming.
• Predicting enzyme-substrate interactions is crucial for understanding enzyme function and drug discovery.
• Enzyme promiscuity, the ability to interact with non-natural substrates, is a significant factor in enzyme activity.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting enzyme promiscuity.
• To identify which of 983 distinct enzymes (by Enzyme Commission numbers) are likely to interact with a query molecule.
• To assess the impact of inhibitor data and hierarchical enzyme relationships on prediction accuracy.

Main Methods:

• Framed enzyme promiscuity prediction as a multi-label classification task.
• Utilized data from the BRENDA database on enzyme-substrate interactions.
• Developed and compared several machine learning models, including a hierarchical multi-label neural network (EPP-HMCNF), k-nearest neighbors, and other ML models.
• Incorporated inhibitor and unlabeled data, along with hierarchical enzyme class relationships, into model training.

Main Results:

• The hierarchical multi-label neural network, EPP-HMCNF, demonstrated superior performance compared to other evaluated models.
• Inhibitor information consistently enhanced predictive power, especially for EPP-HMCNF.
• Model performance decreased when evaluated on realistic data splits and non-natural substrates compared to random splits and natural substrates.

Conclusions:

• EPP-HMCNF is the most effective model for enzyme promiscuity prediction.
• Leveraging inhibitor data and enzyme hierarchy significantly improves prediction accuracy.
• The study highlights the challenges and importance of predicting enzyme interactions with non-natural substrates.