ProteinIQ.io

GNINA

CNN-based molecular docking combining traditional physics with deep learning for accurate binding predictions

proteinsmall moleculemolecular dockingcnn scoringvirtual screeningdeep learning
Loading...

Model overview: GNINA

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 provides more accurate binding predictions than traditional force-field methods while being much faster than pure AI approaches like DiffDock.

How GNINA Works

GNINA uses a two-stage approach:

  1. Traditional docking: Uses AutoDock Vina's search algorithm to explore the binding space and generate candidate poses
  2. CNN scoring: Applies an ensemble of convolutional neural networks to score poses based on learned protein-ligand interaction patterns

This hybrid approach combines the speed and reliability of traditional docking with the accuracy of deep learning.

Input requirements

Protein

  • Format: PDB file or RCSB PDB ID
  • Should be prepared (hydrogens added, missing residues fixed)
  • Use PDB Fixer tool if needed

Ligand

  • Formats: SMILES string, SDF file, or MOL2 file
  • 3D coordinates generated automatically from SMILES
  • Multiple ligands can be screened sequentially

Output

GNINA generates:

  • Multiple binding poses (default: 9 poses)
  • CNN scores for each pose (more negative = higher confidence)
  • Binding affinity estimates (kcal/mol)
  • SDF files with 3D coordinates

Poses are ranked by CNN score, with the top-ranked pose typically having the highest confidence.

Parameters

Exhaustiveness (1-32, default: 8)

Controls search thoroughness. Higher values = more comprehensive search but longer computation time. Default of 8 is sufficient for most applications. Values above 16 rarely improve results.

Number of poses (1-20, default: 9)

How many binding poses to generate. More poses give more diverse predictions but take longer.

Note on Pose Filtering: GNINA uses RMSD-based filtering (min_rmsd_filter, default 1.0 Å) to remove redundant poses that are too similar. Unlike AutoDock Vina which filters by energy range, GNINA focuses on structural diversity.

CNN model

  • Default ensemble (recommended): Uses 3 models for best accuracy
  • Dense: Faster single model
  • CrossDock 2018: Alternative single model

Search box padding (2-12 Å, default: 4)

Extra space added to automatic search box around ligand. Larger values allow more flexible docking but increase computation time.

References

  • McNutt AT, et al. (2025). "GNINA 1.3: the next increment in molecular docking with deep learning." Journal of Cheminformatics, 17:22.
  • Sunseri J, Koes DR (2021). "GNINA 1.0: molecular docking with deep learning." Journal of Cheminformatics, 13:43.

Tips for Best Results

  1. Prepare protein structure: Use PDB Fixer to add hydrogens and fix missing residues
  2. Use default settings first: Exhaustiveness of 8 and 9 poses work well for most cases
  3. Check multiple poses: Don't rely solely on the top-ranked pose - examine top 3-5
  4. Validate with experiments: Computational predictions should be validated experimentally
  5. Consider binding site: GNINA works best when approximate binding site is known

Common Use Cases

  • Virtual screening: Screen large compound libraries quickly
  • Lead optimization: Compare binding of similar compounds
  • Binding mode analysis: Understand how ligands interact with targets
  • Structure-based drug design: Guide medicinal chemistry efforts