BionoiNet

BionoiNet classifies ligand-binding sites in proteins using a ResNet-based convolutional neural network on 2D Voronoi diagram representations to identify and characterize nucleotide- and heme-binding pockets.

Key Features:

• Deep Learning Integration: BionoiNet employs a pretrained ResNet convolutional neural network for classification of ligand-binding pockets.
• 2D Voronoi Diagrams: The tool converts molecular structures of ligand-binding sites into 2D Voronoi diagrams to serve as inputs for the CNN classifier.
• High Accuracy: Reported accuracies are 85.6% for nucleotide-binding pockets and 91.3% for heme-binding pockets.
• BionoiScores: BionoiNet computes per-atom significance scores for pocket atoms, termed BionoiScores, to quantify interactions between ligand molecules and binding sites.
• Computational Efficiency: The framework uses a lightweight 2D approach as an alternative to computationally intensive 3D architectures to reduce computational cost.

Scientific Applications:

• Protein Functional Annotation: Classification of ligand-binding sites supports automatic functional annotation of protein structures.
• Protein Evolution Studies: Binding-site classification provides data for comparative analyses in protein evolution research.
• Protein Engineering and Drug Development: Precise pocket classification can inform design and optimization efforts in protein engineering and drug development.

Methodology:

Molecular structures of ligand-binding sites are transformed into 2D Voronoi diagrams and processed by a pretrained ResNet-based convolutional neural network as a lightweight alternative to 3D architectures.