Accurately predicting enzyme functions through geometric graph learning on ESMFold-predicted structures.

Area of Science:

• Biochemistry and structural biology
• Computational biology and bioinformatics

Background:

• Enzymes are vital biological catalysts, with Enzyme Commission (EC) numbers classifying their function.
• Existing EC number prediction methods often overlook crucial structural and active site information.
• Accurate prediction of enzyme characteristics is essential for various biological and biotechnological applications.

Purpose of the Study:

• To develop GraphEC, a novel geometric graph learning-based predictor for enzyme active sites and EC numbers.
• To enhance EC number prediction by integrating structural data and homology information.
• To predict the optimal pH of enzymes to better understand their catalytic activity.

Main Methods:

• Utilized ESMFold-predicted protein structures and a pre-trained protein language model.
• Developed a model for predicting enzyme active sites, feeding into EC number prediction.
• Employed a label diffusion algorithm to incorporate homology information for improved EC number prediction.
• Integrated optimal pH prediction to complement functional analysis.

Main Results:

• GraphEC demonstrated superior performance in predicting enzyme active sites, EC numbers, and optimal pH compared to existing methods.
• The model effectively extracts functional information directly from protein structures using geometric graph learning.
• Validation confirmed the model's capability in identifying unannotated enzyme functions.

Conclusions:

• GraphEC offers a powerful approach for predicting enzyme function, active sites, and optimal pH.
• Geometric graph learning on protein structures is effective for enzyme characterization.
• This technology has significant potential to accelerate research in synthetic biology, genomics, and enzyme discovery.