Researchers led by Associate Professor Hilal Tayara at Jeonbuk National University in South Korea have developed DDINet, a novel deep learning model designed to accurately predict drug-drug interactions, even for new and unseen medications. The study, published in the journal Knowledge-Based Systems, addresses critical limitations of existing artificial intelligence tools used in polypharmacy, a practice where patients take multiple drugs simultaneously that can lead to dangerous adverse reactions or reduced therapeutic efficacy. Traditional deep learning models for drug interaction prediction often rely on randomly split datasets, which fail to mimic real-world clinical environments where doctors frequently prescribe drugs that have not been extensively studied together. Consequently, many current models experience significant performance drops when confronted with truly unseen drugs. Furthermore, existing solutions often demand substantial computational resources, hindering their practical deployment. DDINet overcomes these challenges through a streamlined architecture comprising five fully connected layers that utilize molecular fingerprints as input. This design minimizes overfitting, allowing the model to generalize effectively to new data. Unlike complex graph-based models, DDINet is lightweight and scalable while simultaneously predicting whether an interaction will occur and identifying its specific biological effects. The system handles both binary classification, determining the likelihood of an interaction, and multi-classification, identifying the mechanism of the interaction. To rigorously evaluate the model, the research team constructed a large-scale dataset from DrugBank and implemented a strict data-splitting protocol to simulate realistic clinical scenarios. They tested the model across three specific conditions. The first scenario involved random splitting of drug pairs. The second scenario included interactions where one drug was known and the other was unseen. The most challenging third scenario involved pairs where both drugs were completely unseen, a situation reflective of actual medical practice. Among five different molecular fingerprinting techniques tested, Morgan fingerprints proved to be the most effective and were selected for the final implementation. Across all evaluation scenarios, DDINet demonstrated performance equal to or superior to existing models, particularly excelling in the difficult unseen drug scenario. Its stable performance across various metrics confirms its reliability for real-world application. Associate Professor Tayara notes that the model's compact and efficient design enables large-scale deployment in hospital settings, drug discovery pipelines, and pharmacovigilance systems. The development of DDINet represents a significant step forward in patient safety and drug development efficiency. By accurately predicting interactions for new drugs without requiring heavy computational power, the technology offers a practical solution to the risks associated with polypharmacy. As Tayara concludes, this innovation has the potential to accelerate drug development processes while significantly improving safety outcomes for patients relying on complex medication regimens.