AbFlow

Adaptive Multi-Task Training for Antibody Sequence-Structure Generation via Flow Matching

AbFlow is a generative model for joint antibody sequence and structure design. It can sample CDR loop sequences and all-atom conformations, conditioned on structural or sequence context and achieves state-of-the-art results on in silico and experimental benchmarks.

---

Installation

1. Clone the repository

git clone https://github.com/AstraZeneca/AbFlow.git
cd AbFlow

2. Create the conda environment

To create the environment:

conda env create -f environment.yml

3. Activate the environment

conda activate abflow

4. Install the abflow package

pip install ./

Now, you are ready to run AbFlow.

---

Main Dependencies

AbFlow is built primarily on:

• PyTorch (>=2.4)
• Lightning (==2.2.1)
• BioPython (==1.84)
• pandas (==2.2.2)
• numpy (==1.26.4)

See the environment.yml file for a complete list of packages and pinned versions.

---

Usage

Below we show an example for redesigning CDR loops and generating a new PDB structure. Whether you redesign individual loops, all six loops or the entire antibody depends on your model checkpoint and its configuration.

For more details, use the --help option with any script.

Example: Redesigning CDR Loops

python scripts/design/design_pdb.py \
    --config /path/to/config.yaml \
    --checkpoint /path/to/checkpoint.ckpt \
    --input_pdb /path/to/antibody_complex.pdb \
    --heavy_chain H \
    --light_chain L \
    --antigen_chains A B \
    --scheme chothia \
    --output_dir results \
    --seed 2025 \
    --device cuda:0

• --config: Path to the configuration file associated with the model checkpoint.
• --checkpoint: Path to your trained model checkpoint.
• --input_pdb: Path to the input PDB file of the antibody complex.
• --heavy_chain and --light_chain: Chain identifiers for the antibody.
• --antigen_chains: Chain identifiers for any antigens present (optional).
• --scheme: Numbering scheme (commonly "chothia").
• --output_dir: Directory to store results.
• --seed: Random seed for reproducibility.
• --device: Device to run the computations on (e.g., cuda:0 or cpu).

The output directory will include the redesigned PDB file.

---

Contact

For any issues, suggestions, or questions, please reach out to:

• hz362@cam.ac.uk
• ucabtuc@gmail.com

---

Reference

If you find this code useful, please cite:

@misc{our_paper,
  title={AbFlow: Adaptive Multi-Task Training for Antibody Sequence-Structure Generation},
  author={...},
  year={...},
  archivePrefix={...},
  primaryClass={...}
}