Drug-Drug Interaction Prediction in Polypharmacy Using Graph Learning and Transformer Models

Abstract

Drug-drug interactions are a significant concern in mod¬ern medicine, particularly because of the rising incidence of polypharmacy. While numerous computational models have been developed for DDI prediction and extraction, many con-straints remain, such as inadequate molecular representation, limited multi-drug context handling, insufficient model inter-pretability, and insufficient annotated training data. We address these research gaps in the current study and present an inte-grated framework using graph neural networks, transformer-based NLP models, and data fusion strategies to overcome re¬search challenges. In addition to a critical assessment and theo¬retical framework of the proposed approach, we focused on the need for interpretability and clinical sensitivity for the devel¬opment of real-time decision-making or warning-based support systems. Open questions remain around fine-tuning large lan¬guage models on molecular graph data, building DDI pipelines that slot into live EHR systems without disrupting clinical work¬flows, and establishing evaluation standards that hold across heterogeneous datasets. Until those gaps close, even a well-designed system will struggle to earn the kind of trust that clin¬ical deployment actually requires.

Introduction

The document focuses on the growing clinical problem of drug–drug interactions (DDIs) in patients receiving polypharmacy, where multiple medications are prescribed simultaneously. While combining drugs can improve treatment outcomes, it also significantly increases the risk of harmful interactions, which may lead to hospitalization or death. Existing computational methods for predicting DDIs face major limitations, including weak molecular representations, poor interpretability, data imbalance, lack of scalability, and minimal integration into real clinical workflows such as Electronic Health Records (EHRs).

The study highlights a gap between high-performing research models and clinically usable systems. Current clinical decision support systems rely on static drug databases, which are incomplete and often generate excessive or poorly prioritized alerts, leading to alert fatigue. Another major issue is the lack of explainability—most AI models can predict interactions but cannot clearly explain the pharmacological reason behind them in a way that clinicians can trust or act upon.

The problem is formally defined as predicting interacting drug pairs within a prescription, classifying the interaction mechanism (pharmacokinetic or pharmacodynamic), assigning severity levels, and generating meaningful natural-language explanations. However, existing systems fail to handle multi-drug interactions, rely on simplified molecular encodings like SMILES strings, produce only binary outputs, and cannot effectively integrate heterogeneous biomedical data sources such as chemical databases, genomic pathways, and clinical records.

The proposed solution introduces a unified framework that addresses these limitations. It uses graph-based molecular representations to preserve structural information, transformer models to capture multi-drug context, and multimodal fusion of chemical, genomic, and clinical data. The system also includes a hierarchical classifier for interaction type and severity, along with an explanation module designed to generate clinically meaningful rationales. Importantly, it is designed for real-time integration into EHR systems to support clinical decision-making at the point of care.

Conclusion

Looking across the body of work reviewed here, a consistent picture emerges: the DDI field has made genuine progress on prediction accuracy but has largely left the clinical usabil¬ity problem unsolved. Molecular encodings remain limited, polypharmacy contexts are routinely ignored, interaction gran¬ularity is poor, and explanations—where they exist at all— rarely tell a clinician something they can act on. The frame¬work proposed here addresses all six of these failure modes in a single, modular pipeline: GNN-based encoding, a multi-drug contextual parser, a PK/PD severity classifier, cross-modal data fusion, a rationale generator, and an EHR-connected alert layer [?, ?, 13]. Whether or not every module gets implemented as described, the architecture at least maps out what a clinically deployable DDI system would need to look like.

References

[1] M. Rastegar-Mojarad et al., “A new method for prioritizing drug repositioning candidates extracted by literature-based discov-ery,” in Proc. IEEE BIBM, pp. 1030–1035, 2015. [2] Z. Xiong et al., “Pushing the boundaries of molecular represen¬tation for drug discovery with a graph attention mechanism,” J. Med. Chem., vol. 63, no. 16, pp. 8749–8760, 2019. [3] Y. Zhang, Y. Wang, and X. Liu, “Multi-drug interaction extrac¬tion via hierarchical transformers,” J. Biomed. Inform., vol. 145, p. 104319, 2024. [4] X. Sun, Y. Zheng, and W. Zhou, “Addressing class imbalance in drug interaction datasets,” Comput. Struct. Biotechnol. J., vol. 21, pp. 3502–3510, 2023. [5] K. Huang, T. Fu, W. Gao, and Y. Zhao, “Clinical pharmacology-based classification of DDI,” J. Med. Syst., vol. 48, no. 1, p. 17, 2024. [6] W. Zhao, H. Li, and Y. Wu, “Fusion of heterogeneous data for robust DDI prediction,” J. Cheminform., vol. 14, no. 1, p. 47, 2022. [7] A. Rajkomar, E. Oren, and K. Chen, “Explainable AI for drug safety: challenges and opportunities,” NPJ Digit. Med., vol. 5, no. 1, p. 15, 2022. [8] B. Wang, Y. Dai, and Y. Zhang, “Evaluation metrics for medical NLP systems: limitations and proposals,” J. Biomed. Semantics, vol. 12, no. 1, p. 6, 2021. [9] Q. Liu, E. Chen, and Z. Wu, “BioBERT: a pre-trained biomed¬ical language representation model for biomedical text mining,” Bioinformatics, vol. 39, no. 7, p. btad245, 2023. [10] S. Huang et al., “Drug–drug interaction analysis based on infor¬mation and graph neural networks,” Medicine, 2025. [11] G. Geng et al., “MGDDI: A multi-scale graph neural network for drug–drug interaction prediction,” Artif. Intell. Med., 2024. [12] X. Liu et al., “Recent developments in GNNs for drug discovery,” arXiv preprint, 2025. [13] J. Jiang et al., “A review of transformer models in drug discovery and beyond,” Artif. Intell. Rev., 2025. [14] M. Garcia et al., “Computational intelligence for multimodal biomedical data fusion,” Front. Artif. Intell., 2023. [15] J. R. Teoh et al., “Advancing healthcare through multimodal data fusion,” PeerJ Comput. Sci., 2024.

Copyright

Copyright © 2026 Mayank Tiwari, Varun Jha. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.