A general model to predict small molecule substrates of enzymes based on machine and deep learning

Area of Science:

• Biochemistry
• Bioinformatics
• Computational Biology

Background:

• Identifying enzyme-substrate pairs is crucial for understanding biological pathways.
• Experimental methods for substrate characterization are often inefficient and expensive.
• Existing machine learning models are limited by training data scarcity, particularly for non-substrates.

Purpose of the Study:

• To develop a general machine learning model for predicting enzyme-substrate pairs.
• To overcome limitations of existing models by incorporating non-substrate data.
• To provide an efficient in silico tool for substrate prediction.

Main Methods:

• Developed ESP, a machine learning model using a modified transformer architecture for enzyme representation.
• Augmented training data with randomly sampled small molecules designated as non-substrates.
• Evaluated model performance on independent and diverse test datasets.

Main Results:

• Achieved over 91% accuracy in predicting enzyme-substrate pairs on independent test data.
• Demonstrated successful application across diverse enzyme families and metabolites.
• Outperformed models specialized for individual enzyme families.

Conclusions:

• ESP provides an accurate and generalizable method for predicting enzyme-substrate relationships.
• The ESP web server facilitates in silico substrate testing, supporting basic and applied scientific research.
• This approach addresses the critical need for efficient enzyme function prediction.