GNINA Docking Web App

A FastAPI-based web application for structure-based molecular docking using GNINA, with real-time progress tracking, optional post-processing analyses, and PyMOL session generation.

Features

• Flexible ligand input: SMILES strings or SDF files (2D or 3D)
• Automatic 2D→3D conversion: OpenBabel-based 3D coordinate generation and protonation at target pH
• Dual-GPU docking: Load-balanced across two GPUs via CUDA_VISIBLE_DEVICES
• Real-time progress: WebSocket-based live updates during docking
• Optional post-processing (per pose, written as SDF fields):
  • MCS_RMSD — MCS-aligned RMSD vs reference ligand (RDKit, heavy atoms only)
  • Shape_Sim — 3D shape Tanimoto similarity vs reference ligand (RDKit)
  • Ref_Sim — 2D Morgan ECFP4 Tanimoto similarity vs reference ligand (RDKit)
  • PLIF_Sim — Protein-Ligand Interaction Fingerprint Tanimoto similarity vs reference ligand (ODDT)
  • PB_Flags — PoseBusters failure count (config='mol')
• Protein preparation: Optional integrated pipeline — fetch by PDB ID or upload, select chains and reference ligand, auto-populates docking inputs. Pipeline: PDBFixer repair → PDBFixer protonation → ASN/GLN/HIS rotamer optimisation → two-stage OpenMM minimization (stage 1: H positions optimised to convergence with all heavy atoms restrained; stage 2: sidechain relaxation with backbone-only restraints)
• PyMOL session: Headless PyMOL generates a .pse file with protein rainbow cartoon, binding site surface, reference ligand (green sticks), and docked poses
• Named sessions: User-defined session name propagated to output SDF, PSE, and ZIP filenames
• DataWarrior-compatible SDF output: Correct protonation states (COO⁻, NH₃⁺), proper block formatting, DockingRank field

Requirements

External tools

• GNINA ≥ 1.3
• OpenBabel ≥ 3.1 (with tetrazole phmodel fix — see below)
• PyMOL (open-source or commercial, headless-capable)

Python environments

Main app (gnina_webapp env):

conda create -n gnina_webapp python=3.10
conda activate gnina_webapp
conda install -c conda-forge rdkit openbabel
pip install -r requirements.txt
pip install posebusters  # optional, for PB_Flags

Protein preparation (openmmdl env — only needed for the Protein Preparation feature):

conda create -n openmmdl python=3.10
conda activate openmmdl
conda install -c conda-forge biopython pdbfixer openmm

The protein preparation pipeline (protprep.py) is called as a subprocess using the openmmdl Python interpreter. Set the OPENMMDL_PYTHON environment variable if it is installed at a non-default path:

export OPENMMDL_PYTHON=/path/to/envs/openmmdl/bin/python

Usage

conda activate gnina_webapp
python gnina_webapp.py

Then open http://localhost:9000 in your browser.

Running as a persistent service

To keep the app running across sessions and restart it automatically on reboot, use a systemd service. Create /etc/systemd/system/gnina-webapp.service:

[Unit]
Description=GNINA Docking Web App
After=network.target

[Service]
Type=simple
User=YOUR_USERNAME
WorkingDirectory=/path/to/gnina/webapp
ExecStart=/path/to/conda/envs/gnina_webapp/bin/python gnina_webapp.py
Restart=on-failure
RestartSec=5

[Install]
WantedBy=multi-user.target

Then enable and start it:

sudo systemctl daemon-reload
sudo systemctl enable gnina-webapp
sudo systemctl start gnina-webapp
sudo systemctl status gnina-webapp   # check it's running

Logs are accessible via:

journalctl -u gnina-webapp -f

Inputs

Field	Description
Protein Preparation	Optional: upload or fetch PDB, select chains/reference ligand, runs PDBFixer + OpenMM minimization, auto-fills receptor and reference inputs
Receptor (PDB)	Protein structure for docking
Reference ligand (SDF)	Defines the binding site (autobox); used as reference for post-processing metrics
Ligands (SDF or SMILES)	Molecules to dock
Session name	Prefix for output filenames
LigPrep pH	Target pH for OpenBabel protonation
Poses per ligand	Number of docked poses to generate
Exhaustiveness	GNINA search exhaustiveness (1–64)
CNN scoring	GNINA CNN scoring mode

Outputs

A ZIP file containing:

• {session_name}.sdf — all docked poses with scoring and optional post-processing fields, ready for DataWarrior
• {session_name}.pse — PyMOL session (if requested)

SDF Output Fields

Field	Description
minimizedAffinity	Vina-style binding affinity (kcal/mol)
CNNscore	GNINA CNN pose quality score
CNNaffinity	GNINA CNN predicted affinity
CNN_VS	GNINA CNN virtual screening score
DockingRank	Pose rank within each ligand
MCS_RMSD	MCS-aligned RMSD to reference (Å)
Shape_Sim	3D shape Tanimoto similarity to reference (0–1)
Ref_Sim	2D ECFP4 Tanimoto similarity to reference (0–1)
PLIF_Sim	Protein-Ligand Interaction Fingerprint Tanimoto similarity to reference (0–1)
PB_Flags	Number of failed PoseBusters checks

Configuration

Key settings near the top of gnina_webapp.py:

GNINA_PATH = "/opt/gnina/gnina.1.3.2"
PYMOL_PATH = "/path/to/pymol"
DOCK_GPU_ID = 1        # CUDA device index for docking
N_GPU = 1              # Number of GPUs to use

Notes

OpenBabel tetrazole fix

OpenBabel ≥ 3.1 can double-protonate tetrazoles. Apply this fix to /usr/local/share/openbabel/3.1.1/phmodel.txt (adjust path for your install):

TRANSFORM c1[nH:1]nnn1 >> c1[n-:1]nnn1  4.89
TRANSFORM c1n[nH:1]nn1 >> c1n[n-:1]nn1  4.89

GNINA GPU selection

The --device flag is ignored in GNINA 1.3.2. GPU selection is done via CUDA_VISIBLE_DEVICES on the subprocess instead.