Exploration and Evaluation of Machine Learning-Based Models for Predicting Enzymatic Reactions.

Area of Science:

• Bioinformatics
• Computational Biology
• Enzymology

Background:

• Increasing volume of unannotated gene sequences necessitates advanced computational tools for functional prediction.
• Novel enzyme discovery is crucial for metabolic engineering and requires accurate sequence annotation.

Purpose of the Study:

• To develop and evaluate computational models for predicting enzyme functions, including Enzyme Commission (EC) numbers and reaction specifics.
• To assess the effectiveness of machine learning algorithms in enzyme sequence annotation and reaction prediction.

Main Methods:

• Enzyme-models (E-models) utilized amino acid sequences to predict Enzyme Commission (EC) numbers.
• Substrate-Enzyme models (SE-models) and Substrate-Enzyme-Product models (SEP-models) incorporated sequence and chemical structure information to predict EC numbers, substrates, and products.
• Various machine learning algorithms were tested, including Random Forests.

Main Results:

• E-models showed suboptimal accuracy in predicting EC numbers.
• SE-models and SEP-models demonstrated high accuracy in predicting EC numbers and enzymatic reactions across tested machine learning methods.
• A Random Forests-based SEP-model achieved an Average AUC score over 0.94 for predicting EC first digits.

Conclusions:

• Combining sequence and chemical structure information significantly enhances the accuracy of enzyme reaction prediction.
• Machine learning approaches, particularly SE-models and SEP-models, are effective for annotating unannotated genes and discovering novel enzymes.