Deep multiple instance learning on heterogeneous graph for drug-disease association prediction.

Area of Science:

• Computational biology
• Pharmacology
• Artificial intelligence

Background:

• Drug repositioning accelerates drug discovery by identifying new uses for existing drugs.
• Current methods for predicting drug-disease associations (DDAs) often lack end-to-end frameworks for path instance-level learning.
• Leveraging topological information in path instances can lead to more precise and interpretable DDA predictions.

Purpose of the Study:

• To introduce MilGNet, a novel deep multiple instance learning framework for drug repositioning.
• To develop an end-to-end approach for learning from path instances in drug-disease heterogeneous networks.
• To improve the accuracy and interpretability of DDA predictions.

Main Methods:

• MilGNet employs a heterogeneous graph neural network (HGNN) encoder for drug and disease node embeddings.
• A pseudo meta-path generator creates multiple meta-path instances from drug-disease pairs (bags).
• A bidirectional instance encoder and multi-scale attention-based predictor refine and aggregate instance representations for prediction.

Main Results:

• MilGNet significantly outperformed ten advanced methods across five benchmark datasets.
• The method achieved accurate and explainable predictions at both bag and instance levels.
• Case studies demonstrated MilGNet's potential for identifying new therapeutic indications.

Conclusions:

• MilGNet offers a powerful and interpretable approach to drug repositioning.
• The framework effectively utilizes topological information for enhanced DDA prediction.
• MilGNet has the potential to accelerate the discovery of novel drug therapies.