GDA Extraction

This repository contains the source code to train and test Biomedical Relation Extraction (BioRE) models on the TBGA dataset. TBGA is a large-scale, semi-automatically annotated dataset for Gene-Disease Association (GDA) extraction. In addition, the repository contains scripts to compute dataset statistics and to convert other BioRE datasets in the required format.
TBGA dataset is available at: https://doi.org/10.5281/zenodo.5911097.
TBGA paper can be found at: https://rdcu.be/cKkY2.

Usage

Clone this repository

git clone https://github.com/GDAMining/gda-extraction.git

Then install all the requirements:

pip install -r requirements.txt

Note: Please choose appropriate PyTorch version based on your machine (related to your CUDA version).
For details, refer to https://pytorch.org/.

Then install the OpenNRE package with

cd ./OpenNRE
python setup.py install 

If users also want to modify the code, run this instead:

cd ./OpenNRE
python setup.py install
python setup.py develop

Datasets

Users can go into the benchmark folder and download TBGA using the script download_TBGA_dataset.sh.

If interested in running models on BioRel and DTI, users can download and store these datasets as follows.

BioRel:
Download in /benchmark/biorel/ the train.json, dev.json, test.json, and relation2id.json files from https://bit.ly/biorel_dataset.
Then, run the convert_biorel2opennre.sh file in /convert2opennre.

DTI:
Download in /benchmark/dti/ the train.json, valid.json, and test.json files from https://cloud.tsinghua.edu.cn/d/c9651d22d3f94fb7a4f8/.
Then, run the convert_dti2opennre.sh file in /convert2opennre.

Dataset Statistics

Users can compute dataset statistics to understand the differences between datasets. For instance, if a user wants to compute statistics for TBGA, they can run

python data_stats.py --benchmark_fpath ./benchmark/TBGA/

Pretrain

Pretrained embeddings can be downloaded by running scripts in the pretrain folder. For instance, if a user wants to download BioWordVec embeddings, they can run

cd ./pretrain
bash download_biowordvec.sh

Once downloaded, pretrained embeddings need to be tailored to the considered dataset. For instance, if a user wants to experiment with TBGA, they have to run

python prepare_embeddings.py --embs_fpath ./pretrain/biowordvec/ --benchmark_fpath ./benchmark/TBGA/

Training

Users can train RE models on the provided datasets using train_model.py, where model can be CNN, PCNN, BiGRU, BiGRU-ATT, or BERE. For instance, a user can run the following script to train and test the CNN (AVE) bag-level model on the TBGA dataset:

python train_cnn.py \
    --metric auc \
    --dataset TBGA \
    --bag_strategy ave \
    --hidden_size 250 \
    --optim sgd \
    --lr 0.2 \
    --batch_size 64 \
    --max_epoch 20

BioWordVec pretrained embeddings (obtained from download_biowordvec.sh) are used to train RE models on TBGA and BioRel. Biomedical Word2Vec pretrained embeddings (obtained from download_biow2v.sh) are required to train RE models on DTI. Results are reported in terms of Area Under the Precision-Recall Curve (AUPRC) and (micro) F1 score.

Cite

If you use or extend our work, please cite the following:

@article{marchesin-silvello-2022,
  title = "TBGA: a large-scale Gene-Disease Association dataset for Biomedical Relation Extraction",
  author = "S. Marchesin and G. Silvello",
  journal = "BMC Bioinformatics",
  year = "2022",
  url = "https://bmcbioinformatics.biomedcentral.com/articles/10.1186/s12859-022-04646-6",
  doi = "10.1186/s12859-022-04646-6",
  volume = "23",
  number = "1",
  pages = "111"
}

@dataset{marchesin-silvello-2022-gda,
  title = "TBGA: A Large-Scale Gene-Disease Association Dataset for Biomedical Relation Extraction",
  author = "S. Marchesin and G. Silvello",
  publisher = "Zenodo",
  year = "2022",
  version = "1.0",
  url = "https://doi.org/10.5281/zenodo.5911097",
  doi = "10.5281/zenodo.5911097"
}

@inproceedings{han-etal-2019-opennre,
  title = "{O}pen{NRE}: An Open and Extensible Toolkit for Neural Relation Extraction",
  author = "X. Han and T. Gao and Y. Yao and D. Ye and Z. Liu and M. Sun",
  booktitle = "Proceedings of EMNLP-IJCNLP: System Demonstrations",
  year = "2019",
  url = "https://www.aclweb.org/anthology/D19-3029",
  doi = "10.18653/v1/D19-3029",
  pages = "169--174"
}

If you use the BERE RE model or the DTI dataset, please cite the following:

@article{hong-etal-2020-bere,
  title = "A novel machine learning framework for automated biomedical relation extraction from large-scale literature repositories",
  author = "L. Hong and J. Lin and S. Li and F. Wan and H. Yang and T. Jiang and D. Zhao and J. Zeng",
  journal = "Nature Machine Intelligence",
  year = "2020",
  url = "https://www.nature.com/articles/s42256-020-0189-y",
  doi = "10.1038/s42256-020-0189-y",
  volume = "2",
  pages = "347--355"	
}

If you use the BioRel dataset, please cite the following:

@article{xing-etal-2020-biorel,
  title     = "BioRel: towards large-scale biomedical relation extraction",
  author    = "R. Xing and J. Luo and T. Song",
  journal   = "BMC Bioinformatics",
  year      = "2020",
  url = "https://bmcbioinformatics.biomedcentral.com/articles/10.1186/s12859-020-03889-5",
  doi = "10.1186/s12859-020-03889-5",
  volume    = "21-S",
  number    = "16",
  pages     = "543"
}