Robust Prediction of Drug Interactions using Chemical Descriptors

Area of Science:

• Pharmacology
• Computational Chemistry
• Artificial Intelligence

Background:

• Polypharmacy necessitates reliable drug-drug interaction (DDI) prediction to prevent adverse drug events.
• Current DDI prediction models often struggle with reliability and lack explainability.
• Severe class imbalance in DDI datasets poses a significant challenge for model training.

Purpose of the Study:

• To develop a novel deep learning framework (T-DDI) for accurate multi-class drug-drug interaction prediction.
• To address the challenges of reliability, explainability, and class imbalance in DDI prediction.
• To provide a tool that supports enhanced drug safety monitoring.

Main Methods:

• Utilized a descriptor-based deep learning framework (T-DDI) employing explicit physicochemical descriptors.
• Implemented an uncertainty-aware estimator to effectively handle severe class imbalance in DDI data.
• Evaluated T-DDI on a large dataset of 868,069 drug pairs across 178 interaction types.

Main Results:

• T-DDI achieved a Macro F1 score of 0.8452 on the held-out test set.
• Performance improved to 0.8992 within the high-confidence subset, outperforming existing baseline models.
• Demonstrated T-DDI's capability to generate plausible DDI hypotheses for novel compounds in a prospective case study.

Conclusions:

• T-DDI offers a reliable and explainable approach to multi-class drug-drug interaction prediction.
• The framework effectively manages class imbalance and provides confidence-stratified predictions.
• T-DDI, coupled with LIME explanations and a web application, enhances drug safety monitoring capabilities.