GNINA
CNN-based molecular docking combining traditional physics with deep learning for accurate binding predictions
What is GNINA?
GNINA (pronounced "guh-NINA") is a molecular docking tool that combines traditional AutoDock Vina-style physics-based docking with deep learning CNN scoring functions. It achieves 73% pose prediction accuracy compared to Vina's 58% on redocking benchmarks, while maintaining practical computation times.
The hybrid approach uses Vina's search algorithm to explore binding poses, then applies convolutional neural networks to score and rank results. This gives you the reliability of physics-based sampling with the pattern recognition capabilities of deep learning.
For comprehensive virtual screening workflows, consider combining GNINA with ADMET-AI for pharmacokinetic predictions or Lipinski's Rule of Five for drug-likeness assessment.
How does GNINA work?
GNINA uses a two-stage docking pipeline that leverages both classical optimization and modern deep learning.
Stage 1: Pose sampling
The docking engine inherits Vina's Iterated Local Search algorithm with BFGS quasi-Newton optimization. It explores the binding space through random mutations of ligand position, orientation, and torsion angles, generating diverse candidate poses.
Stage 2: CNN scoring
Each candidate pose passes through an ensemble of convolutional neural networks trained to recognize protein-ligand binding patterns. The CNNs operate on 3D grids of Gaussian atom-type densities, learning spatial features that distinguish native binding modes from decoys.
The networks are trained on two objectives simultaneously:
- Pose classification: Predicting whether a pose is within 2Å RMSD of the experimental structure
- Affinity prediction: Estimating binding free energy in kcal/mol
CNN architectures
GNINA 1.3 includes models trained on the CrossDocked2020 v1.3 dataset:
Default ensemble uses five independent CNN models and averages their predictions. This provides the most robust scoring at the cost of increased computation.
Dense architecture uses twelve convolutional layers organized into three densely connected blocks. Each layer connects to all subsequent layers, improving gradient flow and feature reuse.
Knowledge-distilled models compress ensemble performance into a single faster model, making high-throughput screening more practical without significant accuracy loss.
Input requirements
Protein
GNINA accepts PDB files or RCSB PDB IDs. The protein should be prepared with hydrogens added and missing residues resolved. Use PDB Fixer for automated preparation if your structure has issues.
Ligand
You can provide ligands as SMILES strings, SDF files, or MOL2 files. GNINA automatically generates 3D coordinates from SMILES using OpenBabel. For complex molecules like macrocycles or peptides, pre-prepared 3D structures typically give better results.
Docking parameters
Exhaustiveness
Controls search thoroughness by setting the number of independent Monte Carlo runs. Higher values explore more of the binding landscape but increase computation time linearly. Values of 8-16 work well for most targets.
Number of poses
Specifies how many binding poses to return. GNINA uses RMSD-based filtering to ensure structural diversity, removing poses that are too similar to each other.
CNN model
Default ensemble averages predictions from five independently trained models. We recommend this for production use when accuracy matters more than speed.
Dense uses a single deep CNN with densely connected layers. Choose this when you need faster turnaround for large screening campaigns.
CrossDock 2018 is an older model trained on the original CrossDocked dataset. It remains available for reproducibility with earlier work.
Search box padding
Extra space added around the ligand when auto-generating the search box. Larger padding allows the ligand to explore more positions but increases the search space.
Understanding the results
CNN score
The CNN pose score indicates confidence that a predicted pose is close to the true binding mode. Scores closer to 1.0 indicate higher confidence, while scores near 0 suggest the pose may be incorrect. GNINA ranks poses primarily by this score.
The CNN is trained to classify poses as "good" (within 2Å RMSD of the crystal structure) or "bad". The output probability directly reflects this binary classification confidence.
Affinity
Predicted binding affinity in kcal/mol. More negative values indicate stronger predicted binding:
| Range | Interpretation |
|---|---|
| -4 to -6 | Weak binding |
| -6 to -8 | Moderate binding |
| -8 to -10 | Strong binding |
| < -10 | Very strong binding |
The affinity prediction comes from a separate CNN head trained on experimental binding data. It's useful for comparing compounds but shouldn't be interpreted as precise thermodynamic measurements.
Interpreting results
Examine the top 3-5 ranked poses rather than relying solely on the highest-ranked prediction. Alternative binding modes within the top results may represent valid conformations. Visual inspection in PDB Viewer helps verify that predicted poses make chemical sense.
Comparison to other docking tools
| Feature | GNINA | AutoDock Vina | Smina | DiffDock |
|---|---|---|---|---|
| Method | Vina + CNN scoring | Physics-based + ML | Vina fork | Diffusion model |
| Pose accuracy | ~73% top-1 | ~58% top-1 | ~60% top-1 | ~43% top-1 |
| Speed | 2-3 min | 1-2 min | 1-2 min | 5-10 min |
| Best for | Pose accuracy | General docking | Custom scoring | Blind docking |
GNINA provides the best pose prediction accuracy among Vina-family tools. DiffDock uses a completely different diffusion-based approach that excels at blind docking when the binding site is unknown.
Best practices
Prepare your protein structure before docking. Use PDB Fixer to add hydrogens, resolve missing residues, and remove water molecules unless they're structurally important.
Start with default settings (exhaustiveness 8, 9 poses, default ensemble). These work well for most targets. Increase exhaustiveness to 16-32 only for difficult cases with large binding sites or highly flexible ligands.
Validate important predictions experimentally. Computational docking provides hypotheses about binding modes, not definitive answers. Use the results to guide further investigation rather than as final conclusions.
Consider the binding site when interpreting results. GNINA performs best when the approximate binding region is known. For truly blind docking with no prior knowledge, DiffDock may be more appropriate.
Common use cases
Virtual screening campaigns benefit from GNINA's combination of accuracy and throughput. The knowledge-distilled models enable screening large compound libraries while maintaining CNN-level pose quality.
Lead optimization projects use GNINA to compare binding modes of structurally similar compounds. Small modifications can shift binding poses, and GNINA's CNN scoring helps identify which changes improve complementarity with the target.
Binding mode analysis helps medicinal chemists understand how compounds interact with their targets. The 3D poses reveal hydrogen bonding patterns, hydrophobic contacts, and potential steric clashes that inform design decisions.
Related tools
- AutoDock Vina - The foundation that GNINA builds upon
- Smina - Vina fork with custom scoring and faster minimization
- DiffDock - Diffusion-based blind docking
- LightDock - Protein-protein docking
- PDB Fixer - Prepare proteins for docking
- PocketFlow - Structure-based ligand design
- Molecular Descriptors - Calculate ligand properties
- ADMET-AI - Predict pharmacokinetic properties
Based on: McNutt, A. T., et al. (2025). GNINA 1.3: the next increment in molecular docking with deep learning. Journal of Cheminformatics, 17:22. https://doi.org/10.1186/s13321-025-00973-x
Original method: Sunseri, J., & Koes, D. R. (2021). GNINA 1.0: molecular docking with deep learning. Journal of Cheminformatics, 13:43. https://doi.org/10.1186/s13321-021-00522-2