A Baseline SMILES-Based Machine Learning Framework for Binary Drug-Drug Interaction Prediction

Abstract

Drug-drug interactions (DDIs) pose a critical threat to patient safety in polypharmacy settings, where the combinatorial space of potential drug pairs far exceeds the capacity of traditional experimental screening methods. Although deep learning and graph-based approaches have demonstrated promising results in DDI prediction, they typically rely on complex molecular fingerprints, biological network data, or three-dimensional structural information, limiting their applicability in low-resource or early-stage drug discovery scenarios. This gap motivates the need for computationally lightweight yet effective baseline frameworks that leverage only primary chemical structure information. In this study, a SMILES-based machine learning framework is proposed for binary DDI prediction, in which character-level TF-IDF encoding is applied to SMILES strings to extract structural features without requiring external biological annotations. A balanced dataset of 265,149 positive interaction pairs and an equal number of negative samples was constructed from the DrugBank database, and three classifiers - Logistic Regression, k-NN, and XGBoost - were systematically evaluated. Logistic Regression achieved the highest performance, with 95.23% accuracy and a ROC-AUC of 0.9848, demonstrating that the DDI problem is largely linearly separable in high-dimensional, sparse TF-IDF representations. These results establish a strong and interpretable baseline for structure-only DDI prediction and provide a reproducible reference point for future multimodal and deep learning studies. © 2026 IEEE.