Enhancing Machine Learning-based Drug-Drug Interaction Prediction via Feature Selection

This study explores how feature selection methods can improve machine learning models for predicting the positive and negative effects of drug-drug interactions (DDIs) using detailed chemicalbased representations.A dataset of 83,972 drug pairs was constructed by integrating the Kaggle opensource database and the Drug-Drug Interaction Database, ensuring both data diversity and reliability.Chemical structure features, including Morgan fingerprints, 2D molecular descriptors, and MACCS keys, were extracted using RDKit and PyBioMed, enabling the quantitative encoding of molecular substructures and physicochemical properties.Feature selection was performed using two approaches: a filter method based on Mutual Information to measure the statistical relevance between features and labels, and an embedded method employing XGBoost to evaluate feature importance during model training.Machine learning models such as XGBoost, Random Forest (RF), Multilayer Perceptron (MLP), LightGBM, and CatBoost were trained using the selected features for DDI prediction.Model performance was evaluated using metrics such as accuracy, F1-score, and recall to assess the impact of feature selection.The results demonstrated that both feature selection strategies significantly improved the performance of DDI prediction and enhanced the interpretability and stability of the models.These findings indicate that appropriate feature selection in high-dimensional chemical structure data can substantially aid in early identification of potentially risky drug combinations, detection of adverse drug reactions, and improving the safety of co-prescription in clinical practice.Furthermore, this feature selection-based framework holds promise for future extensions to drug-food and drug-protein interaction prediction, offering broader applicability in pharmacological safety assessment.