Integrated Similarity-Based Drug Interaction Prediction Using Semi-Supervised Techniques

Abstract

Drug–drug interactions (DDIs) represent a critical challenge in modern healthcare because the concurrent use of multiple medications can lead to adverse drug reactions, reduced therapeutic efficacy, and serious health risks. Traditional experimental methods for identifying potential DDIs are often time-consuming, costly, and limited in scalability. Therefore, computational approaches have become increasingly important for predicting unknown interactions between drugs. This study proposes a predictive framework for drug–drug interaction identification based on integrated similarity measures and semi-supervised learning techniques. The proposed method combines multiple drug similarity features, including chemical structure similarity, therapeutic similarity, target protein similarity, and side-effect similarity, to construct a comprehensive drug similarity network. A semi-supervised learning model is then applied to effectively utilize both labeled and unlabeled data, enabling improved prediction performance even when labeled interaction data is limited. By propagating information through the similarity network, the model captures complex relationships between drugs and identifies potential interactions that have not yet been experimentally validated. Experimental evaluation demonstrates that the proposed approach improves prediction accuracy, robustness, and generalization compared with traditional supervised learning methods. The results highlight the potential of integrated similarity and semi-supervised learning frameworks to support pharmacovigilance, assist clinicians in safe prescription practices, and accelerate drug discovery and drug safety research