DOVE

DOVE evaluates protein-protein docking models using voxel-based convolutional deep neural networks to identify near-native docking decoys and support prediction of protein quaternary structures.

Key Features:

• Convolutional deep neural network (CNN): Uses a CNN architecture to score docking decoys.
• Ensemble networks: Employs eight distinct deep learning networks for comprehensive evaluation.
• 3D voxel-based scanning: Scans the protein-protein interface in three-dimensional voxels to represent spatial atomic arrangements.
• Atomic and energetic input features: Incorporates atomic interactions and energetic contributions as input features for the networks.
• Training and validation datasets: Models were trained and validated using ZDock and DockGround datasets.
• Probabilistic scoring: Produces probabilities that indicate the likelihood of a docking model being near-native.
• Comparative performance: Demonstrated superior performance relative to existing scoring functions across various feature combinations.

Scientific Applications:

• Protein quaternary structure prediction: Supports identification of near-native complex conformations for quaternary structure modeling.
• Docking decoy ranking: Ranks docking decoys to prioritize candidate near-native docking models.
• Complementing experimental methods: Provides probabilistic assessments to complement and guide experimental structure determination.
• Elucidating molecular mechanisms: Aids studies of molecular interactions and cellular processes by improving complex-structure identification.

Methodology:

Uses 3D voxel-based scanning of protein-protein interfaces as input to convolutional deep neural networks (an ensemble of eight networks), incorporates atomic interactions and energetic contributions as features, and trains/validates models on ZDock and DockGround datasets.