An interpretable machine learning framework for adverse drug reaction prediction from drug-target interactions

Area of Science:

• Pharmacology
• Computational Biology
• Machine Learning

Background:

• Adverse drug reactions (ADRs) pose significant risks to patient safety and healthcare.
• Current pharmacovigilance methods, like the Yellow Card Scheme (YCS), lack mechanistic insights into ADRs.
• Interpretable machine learning (ML) combined with drug-target interaction data offers a novel approach to predict and understand ADRs.

Purpose of the Study:

• To develop an interpretable ML framework for predicting significant ADRs using drug-target interaction data.
• To identify key pharmacological relationships underlying ADRs.
• To enhance drug safety through predictive pharmacovigilance (PPV).

Main Methods:

• Integrated drug-target interaction data (STITCH) with ADR reports (YCS).
• Employed disproportionality analysis to identify ADR signals for training Random Forest classifiers across System Organ Classes (SOCs).
• Utilized SMOTE, Tomek, and Bayesian optimization for data balancing and hyperparameter tuning, with feature importance for interpretability and DisGeNET for validation.

Main Results:

• Achieved high prediction performance across SOC categories, with ROC AUC scores up to 0.94.
• Identified pharmacologically relevant drug targets through feature importance analysis.
• Validated findings using DisGeNET and demonstrated the value of real-world data compared to SIDER.

Conclusions:

• The developed interpretable ML framework effectively links drug-target interactions to ADRs.
• This approach shows significant promise for predictive pharmacovigilance (PPV).
• The framework supports safer drug development by providing mechanistic insights into ADRs.