Predicting drug-drug interactions: A deep learning approach with GCN-based collaborative filtering.

Area of Science:

• Pharmacology
• Artificial Intelligence
• Computational Biology

Background:

• Combination drug use is rising, increasing the risk of drug-drug interactions (DDIs) and patient safety concerns.
• Most DDIs stem from enzyme modulation (e.g., cytochrome P450) during drug metabolism, suggesting drugs interacting with similar enzymes are prone to interaction.
• Traditional DDI prediction models face challenges with negative sampling and data imbalance.

Purpose of the Study:

• To develop an AI recommendation model for predicting potential drug-drug interactions (DDIs).
• To analyze drug connectivity using graph convolutional networks (GCN) and collaborative filtering, bypassing chemical structure analysis.
• To overcome limitations of traditional models by avoiding negative sample selection and addressing data imbalance.

Main Methods:

• An AI model integrating graph convolutional network (GCN) and collaborative filtering was developed.
• The model analyzes drug connectivity patterns, inspired by user interest recommendation techniques.
• Utilized the DrugBank database (v5.1.9) with 4,072 drugs and 1,391,790 interaction pairs; validated with TWOSIDES data.

Main Results:

• The GCN-based model accurately predicts potential drug-drug interactions by focusing on connectivity.
• The approach successfully circumvents challenges related to negative interaction sampling and data imbalance.
• Model robustness confirmed via 5-fold cross-validation and external data validation.

Conclusions:

• The developed AI model offers a versatile and accurate framework for predicting drug-drug interactions.
• This method leverages existing DDI reports and connectivity analysis for robust predictions.
• The approach provides a novel solution for identifying potential DDIs, enhancing patient safety in combination drug therapy.