autodock-gpu¶

GPU-accelerated molecular docking. AutoDock-GPU (Vina family) as a stateless HTTP service.

Pre-reqs¶

NVIDIA GPU (any modern card, L4 / A10G / A100 / L40S / H100)
Docker + NVIDIA Container Toolkit
CUDA 12.x drivers on the host
~4 GB GPU memory per concurrent docking job

Deploy¶

Via docker-compose (default):

Uncomment the autodock-gpu block in docker-compose.yml, then:

docker compose up autodock-gpu

Standalone (remote GPU box):

docker run -d \
  --name novomcp-autodock-gpu \
  --gpus all \
  -p 8022:8022 \
  --restart unless-stopped \
  ghcr.io/novomcp/autodock-gpu:latest

Verify GPU is visible inside the container:

docker exec novomcp-autodock-gpu nvidia-smi

Wire into the engine¶

# Local
export AUTODOCK_GPU_URL=http://localhost:8022

# Remote GPU box
export AUTODOCK_GPU_URL=http://gpu-box.local:8022

Verify¶

curl -s http://localhost:8022/health
# {"status":"healthy","service":"autodock-gpu","gpu_available":true}

End-to-end via the engine:

curl -s -X POST http://localhost:8018/mcp/tools/dock_molecules \
  -H 'Authorization: Bearer x' \
  -H 'Content-Type: application/json' \
  -d '{
    "arguments": {
      "ligand_smiles": "CC(=O)Oc1ccccc1C(=O)O",
      "protein_pdb_id": "1CX2"
    }
  }'

Tools that light up¶

dock_molecules, synchronous docking against a PDB ID
dock_with_strain, pairs with novomcp-qm for post-dock strain filtering

Env vars¶

Var	Default	Purpose
`PORT`	`8022`	HTTP listen port
`MAX_CONCURRENT`	`1`	Docking jobs per GPU (raise if you have headroom)
`TIMEOUT_SECONDS`	`600`	Per-job hard timeout

Cost + performance¶

Single docking: ~30–60 seconds on L4/A10G
Small library screen (100 ligands): ~10 minutes
Cheapest cloud spot for occasional use: g4dn.xlarge (AWS) or n1-standard-4 + T4 (GCP) at roughly $0.30/hour spot

For batch screens, run the container on a spot instance and burst through the queue rather than paying for a dedicated GPU.