Optimizing Drug-Drug Interaction Prediction Through Different Models

Abstract

Drug–Drug Interaction (DDI) prediction plays a critical role in computational pharmacology, where the accurate identification of potential adverse interactions can prevent harmful clinical outcomes and improve patient safety. Despite notable progress in the field, many existing approaches still face key challenges such as limited interpretability, dependence on narrow datasets, and suboptimal generaliza- tion across different biomedical domains. To address these challenges, this research proposes a hybrid deep learning framework that integrates Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term Memory (BiLSTM) networks for automated DDI classification. The study further explores how different optimiza- tion strategies—Stochastic Gradient Descent (SGD), Adam, and a custom-designed optimizer affect convergence stability and predictive performance. The experiments were conducted using the Balanced Drug–Drug Interaction Dataset obtained from Kaggle, which contains 171,542 labeled interaction records categorized into three classes: Increase, Decrease, and Neutral. Two experimental setups were employed to evaluate model robustness: a 70:30 train–test split and an 80:20 train–test split. Each record represents a drug pair supported by textual biomedical context, with rigorous preprocessing including tokenization, cleaning, and embedding generation applied before model training. Under the 70:30 split, the proposed CNN–BiLSTM hybrid model achieved an overall accuracy of 95% using the custom optimizer, outperforming SGD (94%) and Adam (89%). When evaluated under the 80:20 split, the model maintained high performance with an accuracy of 88.0%, again surpassing both baseline optimizers. These consistent outcomes across both experimental ratios validate the robustness and generalizability of the proposed hybrid architecture and optimization strategy. Beyond numerical evaluation, the study presents a comparative analysis of optimizers across multiple performance metrics—Accuracy, Precision, Recall, and F1-score—highlighting that optimizer choice significantly affects model learning ef- ficiency and generalization behavior. Overall, this research contributes, a clearly defined and reproducible DDI dataset description, a hybrid CNN–BiLSTM archi- tecture optimized for biomedical text classification, and a comprehensive dual-split experimental analysis under 70:30 and 80:20 configurations. These contributions en- hance both the interpretability and reliability of automated DDI prediction systems, promoting safer and more data-driven pharmacological decision-making.