A  substructure-aware graph neural network incorporating relation features for drug-drug interaction prediction

Area of Science:

• Computational chemistry
• Pharmacology
• Artificial intelligence in drug discovery

Background:

• Identifying drug-drug interactions (DDIs) is critical for drug safety and efficacy.
• Current substructure-based DDI prediction methods have limitations in exploiting graph structure and relation embeddings.

Purpose of the Study:

• To propose a novel substructure-aware graph neural network incorporating relation features (RFSA-DDI) for enhanced DDI prediction.
• To address limitations in substructure extraction and drug representation construction in existing DDI prediction models.

Main Methods:

• Developed a directed message passing neural network with substructure attention based on graph self-adaptive pooling (GSP-DMPNN).
• Introduced a substructure-aware interaction module incorporating relation features (RSAM) to enhance drug representations.
• Integrated GSP-DMPNN and RSAM into the RFSA-DDI model for DDI prediction.

Main Results:

• RFSA-DDI demonstrated superior performance in both transductive and inductive settings on two real-world datasets.
• The model effectively predicted DDIs for unseen drugs, indicating good generalization capability.
• RFSA-DDI accurately captured valuable drug structural information, improving DDI prediction accuracy.

Conclusions:

• RFSA-DDI offers a more accurate and reliable approach to predicting potential drug-drug interactions.
• The proposed method enhances drug development and treatment safety by providing better DDI detection.
• The integration of substructure attention and relation features significantly improves DDI prediction models.