Development of Reaction-Centered Encoders and Benchmarking of Enzyme-Reaction Pair Models

Area of Science:

• Biochemistry and computational biology
• Enzyme function prediction
• Metabolic engineering

Background:

• Uncharacterized enzymes offer significant potential for therapeutics, sustainable materials, and understanding metabolic networks.
• Current high-throughput screening methods for enzyme activity are technically complex, necessitating *in silico* predictive models.
• Protein language models and computer-aided synthesis generate vast numbers of uncharacterized enzymes and reactions.

Purpose of the Study:

• To develop and evaluate high-performing predictive models for enzyme-reaction pairs.
• To address the need for *in silico* prescreening of enzyme-reaction pairs due to the complexity of experimental screening.
• To create models capable of predicting catalysis for uncharacterized enzymes and reactions.

Main Methods:

• Compiled a high-quality dataset of enzyme-reaction pairs.
• Developed and rigorously evaluated a suite of predictive models, including Reaction-Center Graph Neural Networks (RC-GNNs).
• Employed varied data splitting and negative sampling strategies for robust model assessment.

Main Results:

• RC-GNNs demonstrated high performance in predicting enzyme-reaction catalysis, achieving 0.88 and 0.84 ROC-AUC under challenging conditions with dissimilar query reactions.
• On a time-based split, an RC-GNN model reached 0.91 ROC-AUC.
• Models showed strong predictive power on enzymes and reactions not seen during training.

Conclusions:

• RC-GNNs provide a powerful tool for predicting enzyme function, particularly for uncharacterized enzymes and reactions.
• These models can significantly aid metabolic engineers and evolutionary biologists in exploring novel enzymatic capabilities.
• The developed models represent a breakthrough in *in silico* enzyme function prediction, facilitating discovery in various biological and synthetic applications.