Managing complex medical conditions often requires the simultaneous use of several different drugs, a process known as polypharmacy. Although necessary, this can significantly increase the risk of drug-drug interactions (DDIs), which can enhance or reduce treatment efficacy or cause adverse drug reactions (ADRs), which can lead to longer hospital stays and potentially life-threatening outcomes.
In recent years, researchers have increasingly focused on deep learning models to predict DDI. Although these models often outperform traditional methods, they are typically tested under ideal conditions where training and test data are randomly split, and do not reflect real-world clinical settings. As a result, many existing models exhibit a rapid drop in performance when evaluating truly unknown drugs. Some require large amounts of computational resources, limiting their usability in the real world.
To overcome these limitations, a research team led by Associate Professor Hilal Tayara from the School of International Engineering Sciences at Jeonbuk National University (JBNU) in South Korea developed DDINet, a lightweight and scalable model specifically designed to predict new drug interactions that have never been seen before. “DDINet can simultaneously predict whether an interaction will occur and determine its biological impact, while requiring significantly less computational power than complex graph-based models.” explains Dr. Tayara. Their research was made available online in November 2025 and was published in volume 333 of the journal. knowledge based system January 2026.
DDINet utilizes a streamlined architecture with five fully connected layers and uses a drug’s molecular fingerprint as input. This approach avoids overfitting to the training data. This is a common reason why many models have difficulty generalizing to unidentified drugs. Importantly, it is designed to handle binary classification tasks that involve predicting the likelihood that particular drug pairs will interact or not, as well as multiclassification tasks that aim to predict the biological effects or mechanisms of known DDIs.
The researchers trained and evaluated DDINet using a large dataset built from DrugBank. They also tested five different molecular fingerprinting techniques. To enhance generalizability, the researchers employed a rigorous data partitioning protocol during the evaluation. Specifically, we created three scenarios for model evaluation. In scenario 1 (S1), drug pairs were randomly split into training and testing datasets. Furthermore, we utilized a DDI-based split where 10% of all DDI pairs formed an independent test set and the rest were used for training.
Scenario 2 includes a DDI with one drug known and the other unconfirmed, and scenario 3 includes a DDI with both drugs unknown, representing a realistic clinical setting. To classify drugs as invisible and visible, the team applied a rigorous drug-based partitioning protocol based on DrugBank annotations.
Morgan’s fingerprint was identified as the best performing and used for the final implementation. In all evaluation scenarios, DDINet performed as well as or better than existing models, especially on the most difficult S3. We demonstrated stable performance across a variety of metrics on both binary and multiple classification tasks.
“DDINet’s compact and efficient architecture enables large-scale deployment in hospitals, drug discovery pipelines, and pharmacovigilance systems,” concludes Dr. Tayara. “Ultimately, this technology will help accelerate drug development while improving safety for patients dependent on multiple drugs.”
reference: Ali S, Alam W, Chong KT et al.DDINet: Multitask neural network for accurate drug-drug interaction prediction and efficacy analysis. Knowledge-based system. 2026;333:114981. Doi: 10.1016/j.knosys.2025.114981
This article has been reprinted from the following material: Note: Materials may be edited for length and content. Please contact the citation source for details. You can access our press release publishing policy here.
