abfold-extended 0.2.0

Extended AbFold package for AbDiff inference workflows

AbFold Extended

AbFold Extended packages the abfold Python modules used by the AbDiff pipeline. The PyPI distribution name is abfold-extended, while the Python import package remains abfold for compatibility with existing AbDiff code.

This repository is based on the original HK-GSAS/AbFold implementation, with engineering changes for packaging and AbDiff integration.

Installation

Install the package from PyPI:

pip install abfold-extended==0.2.0

Code should continue to import abfold:

from abfold.config import config
from abfold.model import AbFold

For the AbDiff H3-mask stage, install the optional ANARCI extra:

pip install "abfold-extended[h3]==0.2.0"

Optional extras:

pip install "abfold-extended[diffusion]==0.2.0"
pip install "abfold-extended[train]==0.2.0"
pip install "abfold-extended[all]==0.2.0"

Package Scope

This MVP package is intended to make AbDiff's AbFold-facing imports stable:

from abfold.config import config
from abfold.model import AbFold
from abfold.data.data_process import process_repr, process_fasta
from abfold.data.data_process import get_CDRs_mask_with_anarci
from abfold.np import protein
from abfold.np.residue_constants import str_sequence_to_aatype
from abfold.train_ema import tensor_dict_to_device
from abfold.training_config import config as training_config

The package does not include pretrained checkpoints, benchmark data, AF2 representations, IgFold embeddings, or a standalone predict.py entry point. Those assets must be supplied by the calling workflow.

AbDiff Usage

AbDiff can depend on this package by installing abfold-extended while keeping its existing source imports as from abfold ....

For a full AbDiff pipeline environment, use the optional extras needed by the stages you run. In particular, Stage 4 requires ANARCI:

pip install "abfold-extended[h3]==0.2.0"

Development Checks

Run the lightweight smoke tests:

python -m unittest discover -s tests -v

Run the wheel build and install smoke test:

ABFOLD_WHEEL_SMOKE=1 python -m unittest discover -s tests -v

On Windows PowerShell:

$env:ABFOLD_WHEEL_SMOKE = "1"
python -m unittest discover -s tests -v

Notes

• The distribution name and import name are intentionally different: pip install abfold-extended, then import abfold.
• Optional dependencies such as diffusers, deepspeed, and anarci are loaded only by the features that need them.
• Published PyPI files are immutable. If a release is flawed, publish a new version and yank the flawed release instead of trying to replace it.

Citation

@article{abfold,
    title = {AbFold -- an AlphaFold Based Transfer Learning Model for Accurate Antibody Structure Prediction},
    author = {Peng, Chao and Wang, Zelong and Zhao, Peize and Ge, Weifeng and Huang, Charles},
    journal = {bioRxiv},
    year = {2023}
}