Drug Repositioning via Text Augmented KG Embeddings

Code repository to NeurIPS 2021 AI for Science Workshop

Repository Setup

1. Install required libraries with

pip install pykeen==1.5.0 wandb

Refer to link for more instructions on WANDB result tracker. You will be able to see various metrics during training and the evaluation after training.

2. Download files into the corresponding directory under root

• Data and Trained Models

Baseline Training

1. Train and save trained baseline embedding model X (transe, distmult, proje, rotate, simple, tucker)

python baseline/save_model_X.py

2. Example hyperparameter optimization

python baseline/hpo_search.py

Text Augmented Model Training

1. (optional) Scrape text from sources

The needed json files of scraped text are already in the directory. If you are interested in doing from scratch, first download the hetionet json dataset

python text_scrape.py -n X

where X must be one of the available entity type('Anatomy', 'Biological Process', 'Cellular Component', 'Compound', 'Disease', 'Gene'(local), 'Molecular Function', 'Pathway', 'Pharmacologic Class')

2. Text Embedding We provide the texual embedding generated with BioBert V1.1 in the aforementioned polybox folder. If you are interested in generate this texual embedding yourself, please refer to

text_process/get_embedding_for_hetionet_drugs.py

You will need to install Hugging Face Transformers library.

3. Find your pykeen library installation path and replace corresponding files with the ones in /pykeen-extension/ This step is to include the texual interaction etc. functionality.

4. Train Text Augmented KG Run the model you desire to train by
   python model_with_text/X.py The file name should be self-explanatory.

5. To calculate % of Disease @10 and Unique Entities @1: please refer to

evaluation/test_evaluate.py