Multi-Classification of Drug-Drug interaction based on a complete graph convolutional neural network and explainable...

Area of Science:

• Computational chemistry
• Pharmacology
• Artificial intelligence in medicine

Background:

• Multi-drug therapy is increasingly common, particularly in older adults with multiple comorbidities.
• Unanticipated drug-drug interactions (DDIs) pose significant risks, leading to adverse reactions and toxicity.
• Computational models can predict DDIs, improving drug design and reducing research costs.

Purpose of the Study:

• To develop and evaluate a novel computational model for predicting drug-drug interactions (DDIs).
• To enhance the accuracy and efficiency of DDI prediction using a graph convolutional neural network (GCN).
• To improve the interpretability of DDI prediction models through explainable artificial intelligence (XAI).

Main Methods:

• A complete graph convolutional neural (GCN) network was constructed using publicly available DDI data from DrugBank.
• The model processed 37,264 samples with three optimal features: chemical, target, and enzyme.
• The multi-classification model involved drug preprocessing, three GCN layers, and a fully connected network.

Main Results:

• The proposed GCN model achieved a high accuracy of 95.12% in DDI prediction, outperforming previous methods on the same dataset.
• The model demonstrated improved computational time and classification evaluation metrics, even with imbalanced data.
• Explainable artificial intelligence (XAI) techniques, specifically SHapley Additive exPlanations (SHAP), were applied for model interpretability.

Conclusions:

• The developed GCN model effectively predicts DDIs with high accuracy and improved efficiency.
• The integration of XAI enhances model transparency, aiding in the understanding of potential drug hazards.
• This model offers a valuable tool for intelligent pharmaceutical management and mitigating risks associated with multi-drug therapy.