DDI-GPT: Explainable Prediction of Drug-Drug Interactions using Large Language Models enhanced with Knowledge Graphs

Area of Science:

• Pharmacology and Toxicology
• Artificial Intelligence in Medicine
• Bioinformatics

Background:

• Drug-drug interactions (DDIs) pose a significant risk to patient safety and present a challenge in drug development.
• Early identification of potential DDIs is crucial for mitigating adverse events and improving therapeutic outcomes.

Purpose of the Study:

• To introduce DDI-GPT, a deep learning framework for predicting drug-drug interactions (DDIs).
• To enhance early detection of potential drug interactions using knowledge graphs and large language models.
• To provide an explainable deep learning tool for drug safety assessment.

Main Methods:

• Developed DDI-GPT, a framework combining knowledge graphs (KGs) and pre-trained large language models (LLMs).
• Utilized feature attribution methods for explainable deep learning (DL) models on pathway and interactome networks.
• Validated performance on the TwoSIDES benchmark dataset and applied to FDA Adverse Event Reporting System data for zero-shot prediction.

Main Results:

• DDI-GPT achieved an AUROC of 0.964 on the TwoSIDES dataset, outperforming existing DL methods.
• Demonstrated 0.84 AUROC in zero-shot prediction on 9,480 DDI records, a 14% improvement over the best prior method.
• Uncovered CYP3A-enriched signals for Bruton's tyrosine kinase (BTK) inhibitor toxicities, providing mechanistic insights.

Conclusions:

• DDI-GPT effectively predicts DDIs, offering a significant advancement in computational drug safety.
• The framework provides explainable insights into DDI mechanisms, aiding in the understanding of drug toxicities.
• DDI-GPT is available as a web server and software package, supporting drug development and clinical safety.