Heterogeneous Graph Neural Networks for Link Prediction in Biomedical Networks

Our aim is to evaluate the performance of recent publicly available heterogeneous graph neural networks in link prediction on biomedical networks. At the same time, comparing these universal heterogeneous graph neural networks with state-of-the-art biomedical methods, the results show that these advanced heterogeneous graph neural networks can be applied to prediction tasks in the biomedical field, and show encouraging results in complex biomedical networks, which are worth further exploration.

Dataset

• DeepViral-Net, ProGO-Net, NeoDTI-Net and deepDR-Net: four complex biomedical networks are downloaded from https://github.com/pengsl-lab/BioERP.git.
• CTD-DDA, NDFRT-DDA, DrugBank-DDI and STRING-PPI: four single biomedical networks are downloaded from https://github.com/xiangyue9607/BioNEV.git.

Code

The heterogeneous graph neural networks for RGCN, CompGCN, RGAT, HAN, HPN, ieHGCN, HetSANN, HGT, SimpleHGN uses the code from OpenHGNN. To ensure the successful execution of their code, we need to install the environment according to their requirements Get Started.

Link Prediction

After configuring the environment, you can run the existing models in the biomedical dataset. Run the command as follows

cd code
python main.py -m model_name -d dataset_name -t task_name -g 0 --use_best_config --load_from_pretrained

optional arguments:
-m name of model, -d name of dataset, -t name of task, -g which gpu to use, if not set -1
--use_best_config use the best config in the dataset with the model
--load_from_pretrained load model from default checkpoint

for example:

python main.py -m RGCN -d NDFRT_DDA -t link_prediction -g 0 --use_best_config

we are using RGCN model to test the DDA prediction performance in the NDFRT-DDA dataset.

Parameter Analysis

If you want to set different hyper-parameters, manually modify config.ini. The best config will override the parameter in config.ini.