virtual_screening_vina

Virtual Screening with AutoDock VINA

Virtual screening is a computational drug discovery technique that searches large libraries of small molecules to find those most likely to bind a protein target (e.g., a viral enzyme or cancer receptor). By predicting binding affinity computationally, researchers narrow millions of candidates down to a few hundred for lab testing, cutting cost and time in early-stage drug discovery.

This job bundle uses AutoDock VINA, an open-source docking engine. It splits a compound library into chunks and docks them in parallel across a fleet of workers.

    Protein Target              Compound Library (millions)         Top Hits
    ┌─────────┐                ┌─┬─┬─┬─┬─┬─┬─┬─┬─┬─┐            ┌─────┐
    │  ╭───╮  │   AutoDock     │ │ │ │ │ │ │ │ │ │ │   Ranked    │ ★ 1 │ -9.2 kcal/mol
    │  │   │  │ ──── VINA ───→ │ │ │ │ │ │ │ │ │ │ │ ─── by ──→ │ ★ 2 │ -8.7 kcal/mol
    │  ╰───╯  │   (parallel)   │ │ │ │ │ │ │ │ │ │ │  affinity   │ ★ 3 │ -8.4 kcal/mol
    └─────────┘                └─┴─┴─┴─┴─┴─┴─┴─┴─┴─┘            └─────┘
    (binding site)              (SDF/SMILES format)                (CSV output)

Example Output

Most compounds bind weakly (-4 to -6 kcal/mol), but 44 compounds cross the -7.0 threshold into "worth testing in the lab" territory. Those are the drug candidates you'd synthesize and validate experimentally.

How It Works

Step 1: PrepareReceptor  (1 task)
  Convert protein PDB → PDBQT format via Open Babel

Step 2: SplitLibrary  (1 task, depends on Step 1)
  Split SDF compound library into N chunks

Step 3: DockCompounds  (N tasks, parallel, depends on Steps 1+2)
  ┌────────┐ ┌────────┐ ┌────────┐     ┌────────┐
  │Chunk 0 │ │Chunk 1 │ │Chunk 2 │ ... │Chunk N │
  └───┬────┘ └───┬────┘ └───┬────┘     └───┬────┘
      │          │          │               │
      ▼          ▼          ▼               ▼
  obabel → VINA dock → extract scores (per chunk, idempotent)

Step 4: ScoreAndRank  (1 task, depends on Step 3)
  Aggregate all chunk results → ranked CSV of top hits by binding affinity

Each docking task is idempotent (safe for Spot preemption, since it skips if results already exist).

Prerequisites

1. Deadline Cloud farm with a Linux SMF fleet (x86_64, Spot recommended).

2. Software dependencies:

   • OpenBabel: Installed automatically via the queue's Conda environment from conda-forge (default CondaPackages includes openbabel).
   • AutoDock VINA: Not available on conda-forge. Install it one of these ways:
     • (Recommended) Build the conda recipe at conda_recipes/autodock-vina-1.2.5/ into your S3 conda channel, then add autodock-vina to CondaPackages and your S3 channel to CondaChannels.
     • (Quick start) Use a fleet host configuration script to install the VINA binary at worker boot:

       curl -sL "https://github.com/ccsb-scripps/AutoDock-Vina/releases/download/v1.2.5/vina_1.2.5_linux_x86_64" -o /usr/local/bin/vina
       chmod 755 /usr/local/bin/vina

3. Deadline CLI:

   pip install deadline

Sample Data

No pre-downloaded data required. The template can automatically download and filter compounds from ChEMBL.

Receptor: download any protein from the RCSB Protein Data Bank. Example with COVID-19 Main Protease:

curl -LO https://files.rcsb.org/download/6LU7.pdb
grep "^ATOM" 6LU7.pdb > receptor.pdb  # strip to protein atoms only

Compound library: set CompoundLibrary=chembl (default) and the template will:

1. Download ChEMBL chemical representations from EBI FTP
2. Filter to drug-like molecules (configurable SMILES length, no salts/mixtures)
3. Convert to SDF format for docking

Or provide your own SDF/SDF.GZ file via the CompoundLibrary parameter.

Data Attribution

Source	License	Used for
RCSB Protein Data Bank	CC0 1.0 (Public Domain)	Receptor PDB structures
ChEMBL (Zdrazil et al., 2024, Nucleic Acids Research)	CC BY-SA 3.0	Compound library (auto-downloaded)

Usage

# Using ChEMBL (auto-download 100k drug-like compounds):
deadline bundle submit path/to/virtual_screening_vina \
  -p "ReceptorPdb=receptor.pdb" \
  -p "CompoundLibrary=chembl" \
  -p "ChEMBLVersion=37" \
  -p "MaxCompounds=100000" \
  -p "CompoundsPerChunk=100" \
  -p "MaxChunkIndex=999" \
  -p "CenterX=-10.7" \
  -p "CenterY=12.4" \
  -p "CenterZ=68.8" \
  -p "Exhaustiveness=8"

# Using your own compound library:
deadline bundle submit path/to/virtual_screening_vina \
  -p "ReceptorPdb=receptor.pdb" \
  -p "CompoundLibrary=my_compounds.sdf.gz" \
  -p "CompoundsPerChunk=100" \
  -p "MaxChunkIndex=499" \
  -p "CenterX=-10.7" \
  -p "CenterY=12.4" \
  -p "CenterZ=68.8"

Key Parameters

Parameter	Description	Default
CompoundLibrary	SDF file path, or chembl to auto-download	chembl
ChEMBLVersion	ChEMBL release version (when using auto-download)	37
MaxCompounds	Max compounds to extract from ChEMBL	100000
CompoundsPerChunk	Molecules per parallel task	1000
MaxChunkIndex	Last chunk index (num_chunks - 1)	999
CenterX/Y/Z	Docking box center (Angstroms)	0.0
SizeX/Y/Z	Docking box dimensions (Angstroms)	20.0
Exhaustiveness	Search thoroughness (1-64)	8
TopN	Number of top hits to report	500

Performance

Tested with 100,000 ChEMBL compounds against COVID-19 Main Protease:

• 1,000 parallel tasks across 10 Spot workers
• ~5-8 hours wall clock at exhaustiveness=4
• Best hit: -14.53 kcal/mol
• 44 compounds with affinity < -7.0 kcal/mol (strong binders)

Software Setup

Tool	Source	Install method
OpenBabel	conda-forge	Queue Conda environment (automatic)
AutoDock VINA	GitHub releases	Conda recipe (build into S3 channel) or fleet host config
Python	conda-forge	Queue Conda environment (automatic)

See Prerequisites for setup details.