Predicting drug-disease associations through layer attention graph convolutional network.

Area of Science:

• Computational biology
• Bioinformatics
• Drug discovery

Background:

• Identifying drug-disease associations is vital for drug development but experimentally challenging.
• Existing computational methods for predicting these associations require enhancement in efficiency and accuracy.

Purpose of the Study:

• To propose a novel computational method, Layer Attention Graph Convolutional Network (LAGCN), for predicting drug-disease associations.
• To improve the accuracy and efficiency of identifying potential therapeutic relationships between drugs and diseases.

Main Methods:

• Constructed a heterogeneous network integrating known drug-disease associations, drug-drug similarities, and disease-disease similarities.
• Applied graph convolution operations to learn drug and disease embeddings, utilizing an attention mechanism to combine embeddings from multiple layers.
• Scored potential drug-disease associations based on integrated embeddings.

Main Results:

• LAGCN achieved an Area Under the Precision-Recall Curve (AUPRC) of 0.3168 and an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.8750 in 5-fold cross-validations.
• The method outperformed existing state-of-the-art and baseline prediction approaches.
• A case study demonstrated LAGCN's capability to uncover novel drug-disease associations not present in the initial dataset.

Conclusions:

• LAGCN serves as an effective tool for predicting drug-disease associations.
• Embeddings from different convolutional layers capture varying orders of proximity.
• The attention mechanism enhances prediction performance by effectively combining these multi-level embeddings.