EnzyNet: enzyme classification using 3D convolutional neural networks on spatial representation.

Area of Science:

• Computational biology
• Biochemistry
• Machine learning

Background:

• Deep learning excels in computer vision, driven by increased computational power and data availability.
• The Protein Data Bank (PDB) has grown significantly, enabling advanced protein function prediction models.
• Protein structure is more conserved than amino acid sequence, making it a potentially more reliable predictor of function.

Purpose of the Study:

• To introduce EnzyNet, a novel 3D convolutional neural network classifier.
• To predict Enzyme Commission (EC) numbers using only the voxel-based spatial structure of enzymes.
• To evaluate the effectiveness of protein shape in predicting enzymatic function.

Main Methods:

• Developed a two-layer 3D convolutional neural network architecture (EnzyNet).
• Trained and tested EnzyNet on a large dataset of 63,558 enzymes from the PDB.
• Utilized a binary representation of protein shape as input, with spatial distribution of biochemical properties as complementary information.

Main Results:

• EnzyNet achieved an accuracy of 78.4% in predicting enzyme function based solely on protein structure.
• The model demonstrated the efficacy of using 3D structural information for enzyme classification.
• Analysis of biochemical property distribution provided complementary insights.

Conclusions:

• EnzyNet offers a novel and effective method for enzyme function prediction using deep learning on 3D structural data.
• Predicting enzymatic function from protein shape alone is feasible and accurate.
• The study highlights the potential of structural bioinformatics and machine learning in advancing enzymology.