Predicting the toxic side effects of drug interactions using chemical structures and protein sequences

Area of Science:

• Pharmacology
• Computational Chemistry
• Bioinformatics

Background:

• Drug combinations can cause severe toxic side effects, leading to clinical complications and market withdrawal.
• Accurate prediction of drug-drug interaction (DDI) induced toxicity is crucial for patient safety and drug development.

Purpose of the Study:

• To develop and validate a model, TSEDDI, for identifying drug pairs with a high risk of inducing toxic side effects.
• To improve the early detection of adverse reactions in clinical settings.

Main Methods:

• Utilized convolutional neural networks (CNNs) to extract features from drug chemical structures and protein data (enzymes, transporters, targets).
• Implemented a weighted binary cross entropy loss function to address class imbalance.
• Integrated multi-head attention mechanisms with residual connections to boost model performance.

Main Results:

• The TSEDDI model achieved a high accuracy of 0.9059 (±0.0010) in predicting DDI side effects, outperforming baseline models.
• Demonstrated consistent superior performance across various evaluation metrics.
• A case study validated the model's ability to predict potential mechanisms of drug-induced toxicity for specific drug pairs.

Conclusions:

• TSEDDI effectively predicts toxic side effects of drug-drug interactions, leveraging advanced deep learning techniques.
• The model's architecture, combining attention mechanisms and residual connections, enhances the detection of adverse DDI effects.
• TSEDDI serves as a valuable tool for early identification of adverse reactions in clinical trials, potentially improving drug safety.