MilGNet

Codes for "Meta-Path-Based Deep Multiple Instance Learning with Heterogeneous Graph Neural Network for Drug-disease Association Prediction"

Reference

If you make advantage of the MilGNet model or its modules proposed in our paper, please cite the following in your manuscript:

@inproceedings{gu2022milgnet,
title={Milgnet: a multi-instance learning-based heterogeneous graph network for drug repositioning},
author={Gu, Yaowen and Zheng, Si and Zhang, Bowen and Kang, Hongyu and Li, Jiao},
booktitle={2022 IEEE International Conference on Bioinformatics and Biomedicine (BIBM)},
pages={430--437},
year={2022},
organization={IEEE}
}
@article{gu2025milgnet,
   title = {Deep multiple instance learning on heterogeneous graph for drug–disease association prediction},
   author = {Gu, Yaowen and Zheng, Si and Zhang, Bowen and Kang, Hongyu and Jiang, Rui and Li, Jiao},
   journal = {Computers in Biology and Medicine},
   volume = {184},
   pages = {109403},
   ISSN = {0010-4825},
   DOI = {https://doi.org/10.1016/j.compbiomed.2024.109403},
   url = {https://www.sciencedirect.com/science/article/pii/S0010482524014884},
   year = {2025}
}

Overview

Environment Requirement

• torch==1.8.0
• dgl==0.5.2

k-fold Cross Validation

python main.py -da {DATASET} -sp {SAVED PATH}
Main arguments:
    -da: B-dataset C-dataset F-dataset R-dataset
    -ag: Aggregation method for bag embedding [sum, mean, Linear, BiTrans]
    -nl: The number of HeteroGCN layer
    -tk: The topk similarities in heterogeneous network construction
    -k : The topk filtering in instance predictor
    -hf: The dimension of hidden feature
    -ep: The number of epoches
    -bs: Batch size
    -lr: Learning rate
    -dp: Dropout rate
For more arguments, please see args.py

Note: please see the optimal hyperparameter settings for each benchmark dataset, and other support information in 'supplementary materials.docx'.

Model Intepretebility

Use the model_intepret.ipynb to easily generate topk most important meta-path instances for given drug-disease pair (require pre-trained model first).

Baselines

DDA-SKF, SCPMF, NIMCGCN, DRWBNCF, REDDA, PSGCN, HAN, and MHGNN.

Contact

We welcome you to contact us (email: [email protected]) for any questions and cooperations.