Gnina

Gnina predicts ligand binding poses and ranks sampled ligand–receptor conformations using an ensemble of convolutional neural networks as the core scoring function for molecular docking.

Key Features:

• Deep Learning-Enhanced Scoring: Uses an ensemble of convolutional neural networks (CNNs) as the core scoring function to evaluate the fitness of sampled ligand poses.
• Integration with smina/AutoDock Vina: Extends the smina/AutoDock Vina docking workflow by replacing or augmenting the traditional scoring function with CNN-based scoring.
• Performance Optimization: The 1.0 release was evaluated and optimized across parameter settings to balance docking accuracy and computational efficiency.
• Superior Docking Accuracy: When binding pockets are defined, Top1 (≤2 Å RMSD) improves from 58% to 73% for redocking and from 27% to 37% for cross-docking; for whole-protein docking, Top1 improves from 31% to 38% (redocking) and from 12% to 16% (cross-docking).
• Generalization Capability: The CNN ensemble generalizes to unseen proteins and ligands, enabling application across diverse molecular targets.

Scientific Applications:

• Drug discovery: Improves pose prediction for structure-based drug discovery and supports rational design of therapeutics by providing more accurate ligand binding conformations.
• Structural biology: Aids interpretation of protein–ligand atomic interactions by predicting binding poses that correlate with RMSD to known complexes.
• Computational chemistry: Provides a CNN-scored docking approach for exploring protein–ligand dynamics and for benchmarking docking accuracy in redocking and cross-docking tasks.

Methodology:

Incorporates an ensemble of convolutional neural networks as the core scoring function to evaluate sampled ligand poses; was evaluated and optimized across parameter settings, with performance assessed by RMSD and redocking/cross-docking Top1 percentages.