roysh/abodybuilder3

ABodyBuilder3

Code for the paper ABodyBuilder3: Improved and scalable antibody structure predictions.

Code

Download data from zenodo

Data and model weights are hosted at https://zenodo.org/records/11354577.

The bash script download.sh will download and extract data and model weights into appropriate directories.

If you only require model weights for inference, these can be downloaded and extracted with the following commands.

mkdir -p output/ zenodo/
wget -P zenodo/ https://zenodo.org/records/11354577/files/output.tar.gz
tar -xzvf zenodo/output.tar.gz -C output/

Installation

To create a conda environment with all required dependencies, you can use

./init_conda_venv.sh

After installation, the environment can be activated with

conda activate ./.venv

Notebook example

A simple example of using the model is given in notebooks/example.ipynb.

Filter and split data

The repo comes with pre-specified data filtering (specified in data/filters.csv) and splits (specified in data/split.csv). If you want to reproduce these steps then run

1. python src/abodybuilder3/stages/data/combine_data_dfs.py
2. python src/abodybuilder3/stages/data/filter_data.py
3. python src/abodybuilder3/stages/data/split_data.py

Embed sequences using language model

Pre-computed language model embeddings are provided in data/structures/structures_plm after running download.sh. If you wish to regenerate then run

python src/abodybuilder3/stages/data/language_model_embeddings.py

Train model

The model can be trained using

1. python src/abodybuilder3/stages/train.py
2. python src/abodybuilder3/stages/finetune.py

Inference and evaluation

The model can be used to predict structures from the validation and test set using

python src/abodybuilder3/stages/inference.py

For general sequences inputs can be prepared following the examples given in notebooks/example.ipynb.

DVC

Our code is built using dvc pipelines, an alternative way to run the code is via dvc exp run. See experiment_scripts for the configurations we used for the experiments in the manuscript.

Citation

If this code is useful to you please cite our paper using the following bibtex entry.

@article{kenlay2024abodybuilder3,
  title={ABodyBuilder3: Improved and scalable antibody structure predictions},
  author={Kenlay, Henry and Dreyer, Fr{\'e}d{\'e}ric A and Cutting, Daniel and Nissley, Daniel and Deane, Charlotte M},
  journal={arXiv preprint arXiv:2405.20863},
  year={2024}
}