This is the code for our paper ``SumGNN: Multi-typed Drug Interaction Prediction via Efficient Knowledge Graph Summarization'' (published in Bioinformatics'21) [link].

git clone git@github.com:yueyu1030/SumGNN.git
cd SumGNN
pip install -r requirements.txt

python train.py 
    -d drugbank         # task
    -e ddi_hop3         # the name for the log for experiments
    --gpu=0             # ID of GPU
    --hop=3             # size of the hops for subgraph
    --batch=256         # batch size for samples
    --emb_dim=32        # size of embedding for GNN layers
    -b=10               # size of basis for relation kernel

You can also change the d to BioSNAP. Please change the e accordingly. The trained model and the logs are stored in experiments folder. Note that to ensure a fair comparison, we test all models on the same negative triplets.

We provide the dataset in the data folder.

We provide the mapping file between ids in our pre-processed data and their original name/drugbank id as well as a copy of hetionet data and their mapping file on this link.

We train the knowledge graph embedding based on the framework in OpenKE.

To obtain the embedding on your own, you need to first feed the triples in train.txt (edges in dataset) and relations_2hop.txt (edges in KG) as edges into their toolkit and obtain the embeddings for each node. Then, you can incorporate this embedding into our framework by modifying the line 44-45 in model/dgl/rgcn_model.py.

Please kindly cite this paper if you find it useful for your research. Thanks!

@article{yu2021sumgnn,
  title={Sumgnn: Multi-typed drug interaction prediction via efficient knowledge graph summarization},
  author={Yu, Yue and Huang, Kexin and Zhang, Chao and Glass, Lucas M and Sun, Jimeng and Xiao, Cao},
  journal={Bioinformatics},
  year={2021}
}

The code framework is based on GraIL.