Implementation of our paper titled "Drug-Drug Interaction Prediction Based on Knowledge Graph Embeddings and Convolutional-LSTM Network", proc. of The 10th ACM Conference on Bioinformatics, Computational Biology, and Health Informatics(ACM BCB), 2019.
In this paper, we proposed a new method for predicting potential DDIs by encompassing over 12,000 drug features from DrugBank, PharmGKB, and KEGG drugs with the help of knowledge graph(KGs).
In our pipeline, we extract feature vector representation of drugs from the KGs, using various embedding techniques such as RDF2Vec, TransE, KGloVe, SimplE, CrossE, and PyTorch-BigGraph(PBG). The embedded vectors are then used to train different prediction models.
- Python 3
- PySpark
- Scikit-learn
- Keras
- TensorFlow.
- First, collect the DrugBank, KEGG drug, OFFSIDES, and PharmGKB datasets from their website.
- Then convert them into RDF using 5* linked open data principal e.g. convert the data into n-triple or n-quad format. For example, you can use the PHP script from Bio2RDF project (see https://github.com/bio2rdf/bio2rdf-scripts/wiki/Setting-up-the-developer-environment)
- Then generate the embeddings, which should provide the feature vector for each drug in the knowledge graphs e.g. https://github.com/rezacsedu/DDI-prediction-KG-embeddings-Conv-LSTM/blob/master/RDF2Vec.py can be used to generate embeddings with RDF2Vec method.
- Once you have the feature vectors generated, run the prediction algorithm e.g. https://github.com/rezacsedu/DDI-prediction-KG-embeddings-Conv-LSTM/blob/master/DDI_prediction.py can be used to train ML base lines and the Conv-LSTM (functionality will be added soon) using the embeddings generated by the RDF2Vec method. Below we provie a more concrete example.
An example of using the graph embeddings generated by RDF2Vec can be found in https://github.com/rezacsedu/DDI-prediction-KG-embeddings-Conv-LSTM/blob/master/Sample_DDI_Prediction_RDF2Vec.ipynb, which shows DDI prediction about 2500 drugs.
Due to data sharing restrictions from DrugBank, KEGG, and PharmGKB, created RDF graphs are not publicly accessible. However, soon the SPARQL endpoint will be made public for querying. Please check back to https://cloud39.dbis.rwth-aachen.de:9999/blazegraph/#splash.
Already prepared embeddings can be downloaded from the following links with a password of '123':
- KGlove: https://data.fit.fraunhofer.de/index.php/s/mh2EWlY21lpH4hy
- RDF2Vec (skip-gram): https://data.fit.fraunhofer.de/index.php/s/SSPWORQEbnGJK04
- CrossE: https://data.fit.fraunhofer.de/index.php/s/lSgXQ2JujmkqOUK
- TransE: https://data.fit.fraunhofer.de/index.php/s/aSGfTubvFiR
- SimpleIE: size is large so not available for direct download (please contact author)
- PBG: size is large so not available for direct download (please contact author)
Some concepts are based on https://github.com/rcelebi/GraphEmbedding4DDI by Remzi Celebi et al.
If you use the code of this repository for your reserch, please consider citing the following paper:
@inproceedings{karim2019ddiconvlstm,
title={Drug-Drug Interaction Prediction Based on Knowledge Graph Embeddings and Convolutional-LSTM Network},
author={Md. Rezaul Karim, Michael Cochez, Joao Bosco Jares, Mamtaz Uddin, Stefan Decker, and Oya Beyan},
booktitle={Proceedings of ACM BCB, ACM, New York, NY, USA, 10 pages},
year={2019}
}
For any questions, feel free to open an issue or contact at [email protected]