Descriptor-augmented machine learning for enzyme-chemical interaction predictions.

Area of Science:

• Computational Biology and Cheminformatics
• Enzyme Engineering and Design

Background:

• Descriptors are crucial for characterizing enzymes and chemicals in enzyme design for sustainable biochemical synthesis.
• Understanding enzyme-chemical relationships is key to advancing biochemical production.
• Machine learning models require effective descriptors for accurate prediction.

Purpose of the Study:

• To evaluate the impact of various protein and chemical descriptors on Random Forest model performance for enzyme-chemical relationship prediction.
• To assess descriptor influence across three scenarios: new relationship, new enzyme, and new chemical evaluations.
• To provide guidance for developing machine learning models tailored to specific enzyme families.

Main Methods:

• Curated enzyme activity data for seven enzyme families from existing literature.
• Developed a machine learning pipeline using 10-fold cross-validation to evaluate model performance.
• Assessed protein (sequence-based and structure-based, including esm-2) and chemical descriptors in predicting enzyme activity for known/novel enzymes and chemicals.

Main Results:

• Protein descriptors significantly improved model performance for new enzyme evaluation in most datasets, with esm-2 descriptors yielding the best results.
• Chemical descriptors showed no effect on new enzyme evaluation but significantly improved new chemical evaluation in several datasets.
• Dataset characteristics, such as sequence similarity and size, correlated with model performance and descriptor utility.

Conclusions:

• The choice of descriptors is critical and depends on the specific prediction task (new enzyme vs. new chemical).
• Protein descriptors, particularly esm-2, are valuable for predicting novel enzyme activities.
• Chemical descriptors are more effective for predicting activities of new chemicals on known enzymes, guiding future enzyme design and discovery.