Multi-feature machine learning for enhanced drug-drug interaction prediction.

Area of Science:

• Pharmacology
• Computational Biology
• Artificial Intelligence

Background:

• Drug-drug interactions (DDIs) pose significant healthcare risks, leading to severe adverse effects.
• Current machine learning models for DDI prediction often struggle with imbalanced datasets and lack directionality analysis, limiting their clinical applicability.

Purpose of the Study:

• To develop a robust and accurate machine learning framework for predicting drug-drug interactions (DDIs).
• To address limitations in existing methods, specifically data imbalance and the directionality of interactions.

Main Methods:

• Utilized GPT-4o Large Language Model to transform free-text DDI descriptions into structured triplets for directionality analysis.
• Applied SMOTE (Synthetic Minority Over-sampling Technique) to mitigate class imbalance issues.
• Employed Deep Neural Networks (DNNs) incorporating four drug features: molecular fingerprints, enzymes, pathways, and targets.

Main Results:

• Achieved 88.9% accuracy in DDI prediction using the developed DNN model.
• Demonstrated an average AUPR (Area Under the Precision-Recall Curve) gain of 0.68 for minority classes due to SMOTE.
• Attention-based feature importance analysis confirmed that the most influential feature in the DNN model is supported by pharmacological evidence.

Conclusions:

• The proposed framework effectively enhances the accuracy and robustness of drug-drug interaction prediction.
• The integration of LLMs and advanced machine learning techniques offers a promising direction for improving medication safety and clinical decision-making.