VoroCNN

VoroCNN predicts local residue-level quality of 3D protein structures using a deep convolutional neural network built on a 3D Voronoi tessellation of molecular structures.

Key Features:

• 3D Voronoi tessellation representation: Represents molecular structures using a Voronoi tessellation to capture spatial relationships within protein structures.
• Voronoi-based CNN architecture: Employs a deep convolutional neural network architecture constructed on the Voronoi tessellation that integrates convolution and pooling operations adapted to an irregular data domain.
• Local residue-level quality prediction: Predicts local qualities of 3D protein folds at the residue level.
• Evolutionary information integration: Leverages evolutionary information to inform predictions of 3D protein structures and their complexes.
• Binding interface recognition: Can be applied to recognize protein binding interfaces.
• Performance relative to prior 3D CNNs: Provides performance competitive with state-of-the-art methods and reported improvements over earlier 3D CNN architectures for similar tasks.
• Training on Voronoi-represented datasets: The network is trained on datasets represented through Voronoi tessellation of molecular structures.

Scientific Applications:

• Structure quality assessment: Validates predicted protein structures against native configurations by providing residue-level quality scores.
• Protein–protein and protein–ligand interaction analysis: Identifies binding interfaces to support studies of molecular interactions and function.
• Computational structural biology for drug discovery: Supports evaluation of structural models used in drug discovery and enzyme design workflows.
• Molecular disease mechanism studies: Aids investigation of structure-related disease mechanisms by assessing local fold accuracy in structural models.

Methodology:

A deep convolutional neural network is trained on datasets represented by 3D Voronoi tessellation of molecular structures, integrating convolution and pooling operations adapted to the irregular Voronoi domain and incorporating evolutionary information to learn patterns associated with local structural quality and binding interfaces.