BionoiNet: ligand-binding site classification with off-the-shelf deep neural network

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning in Biochemistry

Background:

• Accurate classification of protein ligand-binding sites is crucial for functional annotation, protein engineering, and drug development.
• Deep learning (DL) models show promise for classifying ligand-binding pockets due to their ability to handle complex data patterns.

Purpose of the Study:

• To demonstrate that readily available DL models can efficiently recognize nucleotide- and heme-binding sites with high accuracy.
• To introduce BionoiNet, a novel DL framework for classifying ligand-binding sites.

Main Methods:

• Developed BionoiNet, a DL framework utilizing a ResNet model for image classification.
• Transformed protein ligand-binding site structures into 2D Voronoi diagrams as input for the ResNet classifier.
• Implemented BionoiNet in Python, making the source code publicly available.

Main Results:

• BionoiNet achieved high classification accuracies: 85.6% for nucleotide-binding sites and 91.3% for heme-binding sites.
• The framework demonstrated good generalization to unseen data.
• Introduced BionoiScores for assessing the significance of pocket atoms in ligand interactions.

Conclusions:

• BionoiNet provides an accurate and computationally efficient method for classifying ligand-binding sites.
• The framework serves as a lightweight alternative to complex 3D deep learning architectures.
• BionoiNet facilitates automatic functional annotation and aids in drug discovery research.