AutoDock-GPU Service

GPU-accelerated molecular docking service, part of the NovoMCP computational chemistry engine. It wraps the AutoDock-GPU and AutoGrid binaries behind a FastAPI HTTP interface, with automated binding-site detection, box optimization, reference-ligand co-docking, and optional ADMET-based pose reranking.

Tech Stack

Component	Purpose
CUDA 11.8	GPU acceleration
AutoDock-GPU (built from source)	GPU molecular docking (Lamarckian genetic algorithm)
AutoGrid4 (built from source)	Grid map generation
OpenBabel	PDBQT conversion (receptor + ligand fallback)
Meeko	Ligand PDBQT preparation (primary, with torsion tree)
RDKit	SMILES parsing, 3D conformer generation, molecular descriptors
BioPython	PDB structure parsing, binding site detection
scikit-learn	DBSCAN clustering for pocket detection
PLIP	Protein-ligand interaction profiling (contacts per pose)
FastAPI	HTTP API framework

Features

Docking Pipeline

1. Protein preparation: PDB fetched from RCSB or provided directly, cleaned to protein-only ATOM records, converted to PDBQT via OpenBabel with Gasteiger charges at pH 7.4.
2. Ligand preparation: SMILES converted to a 3D conformer (RDKit), then to PDBQT with a torsion tree (Meeko, OpenBabel fallback).
3. Binding site detection: Automated cavity detection using a grid-based approach + DBSCAN clustering. Scores pockets by volume, depth, and buriedness. Returns the top sites.
4. Box optimization: Docking box sized to 2-3x ligand dimensions, adjusted by pocket volume.
5. Docking execution: AutoGrid4 grid maps + AutoDock-GPU. Number of LGA runs configurable.
6. Reference co-docking: Optionally docks a co-crystallized (auto-extracted) or user-provided reference ligand with identical box parameters and reports delta_vs_reference_kcal on each pose.
7. Interaction profiling: PLIP extracts H-bonds, hydrophobic contacts, salt bridges, pi-stacking, halogen bonds, water bridges, and metal coordination for each pose.
8. ADMET reranking (optional): If an ADDIE ADMET service is configured, poses are reranked using a combined score of docking energy, Lipinski drug-likeness, and ADMET properties.

Strict Error Policy

All docking results are from real computation. The service raises explicit errors instead of returning fallback/synthetic data:

• Invalid SMILES: 400 error
• PDBQT conversion failure: 500 error with diagnostic message
• No binding pockets detected: 422 error suggesting explicit coordinates
• Docking produces no poses: 500 error with root cause explanation
• Docking timeout: 500 error

API Endpoints

Health & Diagnostics

Method	Path	Auth	Description
GET	/health	No	Service health, GPU status
GET	/about	No	Service capabilities
GET	/diagnostics/binaries	No	Check installed docking binaries

Docking Operations

Method	Path	Auth	Description
POST	/dock	API-Key	Single ligand docking
POST	/batch-dock	API-Key	Batch docking (serial background execution)
GET	/results/{docking_id}	API-Key	Get docking results
GET	/batch-results/{batch_id}	API-Key	Get batch results
DELETE	/results/{docking_id}	API-Key	Delete stored results

All mutating endpoints require an X-API-Key header. Set the expected value via the API_KEY environment variable.

See docs/API_REFERENCE.md for request/response detail.

Environment Variables

Variable	Default	Description
PORT	8022	Service port
API_KEY	autodock-gpu-api-key-2024	API authentication key
MAX_POSES	20	Maximum poses per docking
NRUNS	100	Default number of AutoDock-GPU LGA runs
MAX_CONCURRENT_DOCKING	2	Semaphore limit for parallel docking
ADDIE_SERVICE_URL	(unset)	Optional ADDIE ADMET service for pose reranking; reranking is skipped if unset
ADDIE_API_KEY	(unset)	API key for the ADDIE service

Running

docker run -p 8022:8022 ghcr.io/novomcp/autodock-gpu:latest

Point the NovoMCP engine at this service by setting AUTODOCK_GPU_URL to its URL.

License

Apache-2.0. See LICENSE. Third-party components built or installed at container-build time are listed in NOTICE.