GNINA Docking Tutorial

This tutorial is part of the course Molecular Modeling of Biological Systems offered by the Department of Physics, University of Cagliari. It has been adapted for use on the Aula I (Block B) computer systems, which may result in minor compromises in performance and accuracy.

This repository provides a practical tutorial for common GNINA docking workflows:

• Redocking
• Docking from a random conformer
• Blind docking
• Flexible docking
• Scoring / rescoring
• Basic virtual screening (VS)

It follows the GNINA software (Roes lab).
Reference: Roes et al., Journal of Cheminformatics (2021)

---

📦 Prerequisites

• GNINA installed and available in your PATH (or use full binary path, e.g. /usr/local/gnina/gnina.1.3.2)
• OpenBabel (obabel, obrms)
• wget, gzip, awk, python3 with matplotlib and pandas
• A molecular viewer (VMD, PyMOL, ChimeraX)

---

📂 Repository Layout

Repository layout used in this tutorial (created on the fly):

• re-docking/ — reproduce bound pose (redocking)
• docking/ — docking from random conformer
• rescoring/ — scoring-only experiments
• flexible/ — flexible docking experiments
• VS/ — virtual screening example
• run_dock_exhaustiveness.sh — helper to test exhaustiveness
• plot_exhaustiveness.py — plot RMSD vs exhaustiveness

Quick notes about common commands used here

• gnina -r -l [options] -o
• --autobox_ligand : center the docking box on a ligand (e.g. the bound ligand)
• --exhaustiveness N : increase search thoroughness (default: 8)
• --score_only : run scoring only (no pose generation)
• --flexres A:52,A:103 : specify flexible residues
• --flexdist N and --flexdist_ligand : set distance-based flexible region
• obrms -firstonly ref.pdb poses.pdb : compute RMSD of the first (top) pose vs reference

📦 1) Redocking: recover the exact bound conformation

---

Purpose: show GNINA can re-create the experimental bound pose.

1. Make a directory and download the PDB

   mkdir re-docking
   cd re-docking
   wget https://files.rcsb.org/download/3ERK.pdb

3. Extract protein atoms only (ATOM records)

   grep "^ATOM" 3ERK.pdb > rec.pdb
   
   

4. Fix connectivity (OpenBabel will add bond info)

   obabel rec.pdb -O rec-fix.pdb

6. Extract ligand (SB4 is the ligand code in 3ERK)

   grep SB4 3ERK.pdb > lig.pdb

7. Perform a simple docking (autobox_ligand centers the box on the bound ligand)

   gnina -r rec.pdb -l lig.pdb --autobox_ligand lig.pdb --seed 0 -o docked.pdb --cpu 8 --cnn crossdock_default2018

8. Compute RMSD of top pose vs experimental ligand

   obrms --firstonly lig.pdb docked.pdb

Open docked.pdb in VMD / PyMOL and compare poses visually.

 pymol 3ERK.pdb rec.pdb rec-fix.pdb docked.pdb lig.pdb

Assignment 1 (suggested)

• Investigate the effect of --exhaustiveness on the docking RMSD.
• Use the helper script run_dock_exhaustiveness.sh included in this repo to automate runs and produce a CSV of RMSD vs exhaustiveness values.
• Plot RMSD vs exhaustiveness with plot_exhaustiveness.py and determine the lowest exhaustiveness that yields converged (stable/acceptable) RMSD. Default is 8.

bash run_dock_exhaustiveness.sh

for i in 8 16 32 64; do (cd "$i" && obrms --firstonly lig.pdb docked.pdb > rmsd.dat); done

python plot_exhaustiveness.py

📦 2) Docking from a random conformer (same system, new starting ligand)

---

Purpose: check robustness to ligand starting conformation.

1. Go up one directory and start a new folder

   cd ..
   mkdir docking
   cd docking

2. Copy receptor and reference ligand for autoboxing

cp ../re-docking/rec.pdb ../re-docking/lig.pdb .

3. Now let's dock with a randomly-generated conformation of the ligand. From the pdb page of the 3ERK system, scroll until you find the "Small molecules" section. You will find 1 unique ligand, namely "4-(4-FLUOROPHENYL)-1-(4-PIPERIDINYL)-5-(2-AMINO-4-PYRIMIDINYL)-IMIDAZOLE". Copy that string into pubchem (https://pubchem.ncbi.nlm.nih.gov/), and copy the SMILES string and generate a random conformation with openbabel

obabel -:'C1CNCCC1N2C=NC(=C2C3=NC(=NC=C3)N)C4=CC=C(C=C4)F' -O lig-random.sdf --gen3D

4. Dock the random conformer (example uses exhaustiveness 8)

gnina -r rec.pdb -l lig-random.sdf --autobox_ligand ../re-docking/lig.pdb --seed 0 -o docked_random.pdb --exhaustiveness 8 --cpu 8 --cnn crossdock_default2018

5. RMSD of the best pose vs experimental ligand

obrms --firstonly ../re-docking/lig.pdb docked_random.pdb

Increase the exhaustiveness parameter to 512, and investigate how rmsd and scoring changes

Also assess which starting conformation gives better RMSD and CNN score (GNINA prints CNN scoring information). Compare affinity (CNN score, Vina/Vinardo scores) as reported in the output SDF/PDB.

📦 3) Scoring-only / rescoring

---

Purpose: evaluate different scoring functions on fixed poses.

cd ..
mkdir rescoring
cd rescoring

• Score only with default scoring:

gnina --score_only -r ../re-docking/rec.pdb -l ../re-docking/lig.pdb --verbosity=2

• List available scoring functions:

gnina --help | grep scoring | head -3

• Score with classical scoring (example: ad4_scoring):

gnina --score_only -r ../re-docking/rec.pdb -l ../re-docking/lig.pdb --verbosity=2 --scoring ad4_scoring

• Print scoring terms:

gnina --score_only -r ../re-docking/rec.pdb -l ../re-docking/lig.pdb --verbosity=2 --scoring ad4_scoring --print_terms

• Custom scoring terms file:

gnina --help | grep scoring | head -3

gnina --print_terms | grep -v atom_type > scoring-terms-tmp

Try the convolutional neural network (CNN) scoring family:

• See available CNN models:

  gnina --help | grep "cnn arg" -A 12

• Score with CNN (CNN score is between 0 and 1; higher is better)

gnina --score_only -r ../rec.pdb -l ../lig.pdb | grep CNN

for instance,

gnina --score_only -r ../re-docking/rec.pdb -l ../re-docking/lig.pdb --cnn redock_default2018_3

📦 4) Blind docking

---

Purpose: search the whole receptor surface for possible binding sites.

cd ..
mkdir blind
cd blind
cp ../re-docking/rec.pdb ../re-docking/lig.pdb .

• Blind docking by centering autobox on receptor (autobox_ligand rec.pdb)

gnina -r rec.pdb -l lig.pdb --autobox_ligand rec.pdb -o docked-blind.pdb --seed 0 --cpu 8 --cnn crossdock_default2018

• For a random ligand conformer:

obrms --firstonly lig.pdb docked-blind.pdb | head -n 10

Notes:

• Blind docking is slower and may require larger exhaustiveness.
• Evaluate top clusters/poses and their CNN scores to prioritize predicted sites.

📦 5) Flexible docking

---

Purpose: allow receptor residues (sidechains / local backbone) to move during docking.

cd ..
mkdir flexible
cd flexible

1. Dock ligand from 3ERK into a similar protein (4ERK) without flexibility:

wget http://files.rcsb.org/download/4ERK.pdb

grep "^ATOM" 4ERK.pdb > rec2.pdb
obabel rec2.pdb -O rec2-fix.pdb
grep OLO 4ERK.pdb > lig2.pdb

gnina -r rec2-fix.pdb -l ../re-docking/lig.pdb --autobox_ligand lig2.pdb --seed 0 -o docked-lig-onto-4ERK.pdb --cpu 8 --cnn crossdock_default2018

obrms --firstonly ../re-docking/lig.pdb docked-lig-onto-4ERK.pdb

2. Try higher exhaustiveness if initial docking is poor:

gnina -r rec2-fix.pdb -l ../re-docking/lig.pdb --autobox_ligand lig2.pdb --seed 0 -o docked-lig-onto-4ERK.pdb --exhaustiveness 512 --cpu 8 --cnn crossdock_default2018

3. Enable flexible residues by distance from the autobox ligand:

gnina -r rec2-fix.pdb -l ../re-docking/lig.pdb --autobox_ligand lig2.pdb --seed 0 -o docked-lig-onto-4ERK-flex.pdb --flexdist 4 --flexdist_ligand lig2.pdb --out_flex flexout.pdb --cpu 8 --cnn crossdock_default2018

4. Or specify exact residues to be flexible:

gnina -r rec2-fix.pdb -l ../re-docking/lig.pdb --autobox_ligand lig2.pdb --seed 0 -o docked-lig-onto-4ERK-flex2.pdb --exhaustiveness 16 --flexres A:52,A:103 --out_flex flexout2.pdb --cpu 8 --cnn crossdock_default2018

Assignment 2 (suggested)

• Investigate convergence of results vs. exhaustiveness in the flexible and rigid cases and compare RMSD / CNN score.
• Which flexible setup (distance-based vs explicit residues) gives the best recovery for the 4ERK system?

📦 6) Virtual screening (VS) example

---

Purpose: score a ligand database against a receptor and assess enrichment.

cd ..
mkdir VS
cd VS

1. Example files used below:

wget http://files.rcsb.org/download/4PPS.pdb

wget http://bits.csb.pitt.edu/files/workshop_minimized_results.sdf.gz

   grep ^ATOM 4PPS.pdb > errec.pdb

2. A simple VS run with VINARDO scoring:

gnina -r errec.pdb -l workshop_minimized_results.sdf.gz --minimize -o gnina_scored_vinardo.sdf.gz --scoring vinardo --cpu 8 --cnn crossdock_default2018

3. Compute ROC / AUC:
   • The file how-to-run-VS1.txt contains instructions and pointers to VS1.py (script included in this repo or workshop). Use that script to get ROC/AUC and typical enrichment metrics.

python BS1.py

Assignment 3 (suggested)

• Try other scoring functions (cnn family, classic scorings) and compare AUC/enrichment.
• Try rescoring GNINA poses with CNN to see improvement vs classical scoring.

Helper scripts included

• run_dock_exhaustiveness.sh : run GNINA for a list of exhaustiveness values and collect RMSD of the top pose into a CSV.
• plot_exhaustiveness.py : read CSV and produce a plot (matplotlib).

Final notes and tips

• Use --seed for reproducibility.
• Increase --exhaustiveness for more thorough search; the runtime increases roughly linearly with it.
• Use --cnn_only if you only want CNN-based pose ranking (check your GNINA build options).
• When comparing RMSDs, make sure you align atoms consistently (obrms -firstonly is useful) and be aware of symmetric ligands.

References

• GNINA paper: Roes et al., Journal of Cheminformatics (2021). https://jcheminf.biomedcentral.com/articles/10.1186/s13321-021-00522-2