DeepRank-Ab is a geometric deep learning scoring function for ranking antibody-antigen docking models and predicting DockQ scores. Given a raw PDB, it runs chain detection, feature generation (ESM-2 embeddings, CDR annotation, atom-level graphs), and EGNN inference to produce a predicted DockQ score plus structural quality flags.

Publication: https://www.nature.com/articles/s42003-026-10408-4

pip install deeprank-ab

Working on the code itself instead? See DEVELOPMENT.md.

Two image variants are published to GHCR: *-cpu and *-gpu, tagged <release-tag>-cpu/ <release-tag>-gpu plus latest-cpu/latest-gpu on every release. PRs also get a pr-<number>-cpu/pr-<number>-gpu build for reviewing that PR's changes — those are work-in-progress images, not for general use.

Mount your data directory to /data, set it as the working directory, and pass --user "$(id -u):$(id -g)" — without it, the container runs as root and every file it creates (workspace, *.hdf5, the ~2.5GB downloaded ESM-2 weights) ends up root-owned on your host:

docker run --rm \
  --user "$(id -u):$(id -g)" \
  -v "$PWD":/data \
  -w /data \
  ghcr.io/haddocking/deeprank-ab:latest-cpu \
  test.pdb

Chain-override flags work the same way, appended after the PDB file:

docker run --rm --user "$(id -u):$(id -g)" -v "$PWD":/data -w /data \
  ghcr.io/haddocking/deeprank-ab:latest-cpu \
  test.pdb --heavy_chain_id H --light_chain_id L --antigen_chain_id A

Your results will be at <pdb_stem>-deeprank_ab_pred_<...>/.

The *-gpu image (built on nvidia/cuda) needs the NVIDIA Container Toolkit installed on the host, plus --gpus all:

docker run --rm --gpus all --user "$(id -u):$(id -g)" -v "$PWD":/data -w /data \
  ghcr.io/haddocking/deeprank-ab:latest-gpu \
  test.pdb

Build locally instead of pulling:

docker build --platform linux/amd64 -f Dockerfile.cpu -t deeprank-ab:cpu .
docker build --platform linux/amd64 -f Dockerfile.gpu -t deeprank-ab:gpu .

--platform linux/amd64 matters even on Apple Silicon: the vendored hmmscan/voronota binaries are x86-64 Linux ELF and only run correctly (via Rosetta emulation) on that platform — an arm64 build would hit "exec format error" on them regardless of host.

deeprank-ab-predict <pdb_file>

Example:

deeprank-ab-predict example/test.pdb

Input requirements:

• PDB file (single model or ensemble supported)
• Optional chain overrides: --heavy_chain_id, --light_chain_id, --antigen_chain_id

If not provided, chains are auto-detected via ANARCI.

flowchart TD
    A[Workspace creation] --> B[PDB splitting]
    B --> C[Chain detection]
    C --> D[Antigen merging]
    D --> E[FASTA generation]
    E --> F[ESM embeddings]
    F --> G[CDR annotation]
    G --> H[Graph construction]
    H --> I[VdW clash filtering]
    I --> J[Clustering]
    J --> K[DockQ prediction]
    K --> L[CSV output]

*_predictions.hdf5 and a final *.csv with columns:

• pdb_id
• predicted_dockq
• HL_contact_flag: ok / low_HL_contacts / not_applicable
• vdw_clash_flag: ok / potential_clash

Open a GitHub issue for help.