An unofficial re-implementation of DiffAb, a diffusion-based generative model for antigen-specific antibody design and optimization.

pip install diffab-pytorch

from diffab_pytorch import DiffAb

model = DiffAb()

To train DiffAb, we need to extract local patches of antibody-antigen structures, which is defined as a set of K residues surrounding CDR anchor residues (Here, we use K=128). This repo provides handy script and pipeline to preprocess all the SAbDab structures. For now, the pipeline will work on Chothia-numbered PDB files only. As a prerequisite, download SAbDab all_structures.zip files and extract it under data directory. If it is done correctly, you will have a directory under the path data/all_structures/chothia with 7,599 PDB files (as of August 30, 2023).

Then, run the pipeline using snakemake command below. Replace [CORES] with appropriate number of CPU cores in your system.

In case you don't have it already, install it with pip install snakemake or conda install -c bioconda snakemake

$ snakemake -s pipelines/preprocess.smk -pr -j[CORES] --keep-going

--keep-going flag lets you ignore PDB files that are not able to preprocess for some reason (in most case PDB parser doesn't work with those files). Let's see if dropping them will affect the reproduction performance later.

WIP

$ pytest

@article{luo2022antigen,
  title={Antigen-specific antibody design and optimization with diffusion-based generative models for protein structures},
  author={Luo, Shitong and Su, Yufeng and Peng, Xingang and Wang, Sheng and Peng, Jian and Ma, Jianzhu},
  journal={Advances in Neural Information Processing Systems},
  volume={35},
  pages={9754--9767},
  year={2022}
}