A general model for predicting enzyme functions based on enzymatic reactions.

Area of Science:

• Biochemistry and Bioinformatics
• Enzymology
• Machine Learning in Chemistry

Background:

• Accurate enzyme commission (EC) number prediction is crucial for enzyme function understanding, biocatalysis, and biosynthetic planning.
• Existing machine learning (ML) models often have limited accuracy when trained without specific enzymatic catalyst information.
• Experimental verification of enzyme functions is time-consuming and costly.

Purpose of the Study:

• To introduce BEC-Pred, a novel BERT-based multiclassification model for predicting EC numbers of chemical reactions.
• To leverage transfer learning for precise EC number forecasting using only substrate and product SMILES sequences.
• To demonstrate the model's capability in classifying enzymatic reactions without explicit catalyst data.

Main Methods:

• Development of BEC-Pred, a BERT-based multiclassification model.
• Utilizing transfer learning on SMILES sequences of substrates and products.
• Comparative analysis against existing sequence-based and graph-based ML methods.

Main Results:

• BEC-Pred achieved a prediction accuracy of 91.6%, outperforming other ML methods by 5.5%.
• The model demonstrated superior F1 scores, with improvements of 6.6% and 6.0% over existing approaches.
• Accurate classification examples were shown for Novozym 435-induced hydrolysis and lipase-catalyzed synthesis.

Conclusions:

• BEC-Pred offers a reliable and accurate method for predicting EC numbers solely from reaction data.
• The model's performance highlights its potential to accelerate research in synthetic biology and drug metabolism.
• BEC-Pred is poised to significantly impact the progression of enzymatic research and biocatalysis.