Graphrxinsight: GNN-Enhanced Ensemble Framework for Detecting Risky Polypharmacy and Drug Side Effects

Abstract

Artificial Intelligence (AI), Machine Learning (ML), and Deep Learning (DL) are widely used in healthcare for analyzing molecular and clinical data to support decision-making. AI builds intelligent systems, ML enables data-driven learning, and DL, especially Graph Attention Networks (GAT) and Deep Neural Networks (DNN), extract complex features for accurate drug–drug interaction (DDI) prediction[6]. In existing systems to predict DDI, the models using Deep Neural Networks (DNN), Random Forest, and XGBoost with DrugBank’s structural similarity profiles are applied. They use the Synthetic Minority Over-sampling Technique to handle class imbalance in clinical datasets[5]. These approaches aim to improve the prediction of rare but critical adverse drug interactions and have demonstrated accuracy around 93.80%[8]. Even though current models provide useful insights, they have limitations in fully capturing complex drug interactions and real-world usage[3][1]. They struggle with handling imbalanced data and do not fully use Graph Neural Networks (GNN) to model multi-drug networks. We propose a hybrid model, CLINENSEMBLE, combining Graph Attention Networks (GAT), Deep Neural Networks (DNN), and CatBoost.Using chemical structure data and graph features along with clinical information, it improves rare event detection by over 5% and overall accuracy. This helps doctors and researchers better assess drug interaction risks and improve patient safety.

Introduction

Drug–drug interactions (DDIs) occur when one drug alters the effect of another, potentially causing adverse reactions, reduced therapeutic efficacy, toxicity, or serious health complications. The growing prevalence of polypharmacy, especially among elderly patients and those with chronic diseases, has significantly increased the risk of DDIs. Traditional DDI detection methods rely on static databases, expert rules, and clinical reports, which are effective for known interactions but struggle to identify rare, unknown, or emerging interactions and require continuous manual updates.

Recent advances in Artificial Intelligence (AI), Machine Learning (ML), Deep Learning, and Graph Neural Networks (GNNs) have greatly improved DDI prediction by learning complex relationships from biomedical data. ML models such as Random Forest, Support Vector Machine, and XGBoost effectively analyze high-dimensional datasets, while GNNs represent drugs as graph structures to capture structural and biological relationships between compounds. Ensemble learning further enhances prediction accuracy and reduces overfitting by combining multiple models.

To address existing limitations, the proposed GraphRxInsight framework integrates Graph Neural Networks with ensemble learning for detecting risky polypharmacy and predicting adverse drug side effects. The system combines chemical descriptors, biological targets, therapeutic classifications, and side-effect information into a unified representation of drugs. Dynamic graph-based learning enables the framework to identify hidden interaction patterns and predict unknown or rare adverse reactions more effectively than traditional rule-based systems.

The methodology consists of multiple stages, including data acquisition from biomedical sources such as DrugBank and SIDER, data preprocessing, feature engineering, graph construction, GNN-based representation learning, ensemble model training, side-effect prediction, model evaluation, and deployment. Feature engineering integrates chemical, biological, therapeutic, and side-effect information, while graph construction models drugs as nodes and interactions as edges to capture complex dependencies. The ensemble combines GNNs, Random Forest, and XGBoost models, using techniques such as hard-negative sampling, hyperparameter tuning, and cross-validation to improve robustness and generalization.

Beyond interaction prediction, GraphRxInsight dynamically predicts adverse side effects by analyzing graph-based relationships and adverse reaction datasets, providing risk probabilities, severity levels, and frequently co-occurring side effects. The system was evaluated using accuracy, precision, recall, and F1-score, achieving approximately 93–94% accuracy with an F1-score of 0.96, outperforming many traditional machine learning methods.

The final model is deployed as a real-time web application using Flask for backend services and React for the frontend, allowing healthcare professionals to input drug combinations and instantly receive interaction risk assessments, probability scores, side-effect predictions, and severity indications. This makes GraphRxInsight an effective clinical decision support system that improves prescription safety, supports healthcare professionals in managing polypharmacy, and enhances patient care.

Conclusion

In this project, GraphRxInsight was developed as a GNN-enhanced ensemble framework for detecting risky polypharmacy and predicting adverse drug side effects using multi-source biomedical data. The proposed system integrates chemical, biological, therapeutic, and side-effect features to generate comprehensive drug representations capable of capturing complex interaction patterns between compounds. Unlike traditional rule-based systems and static machine learning approaches, the proposed framework dynamically analyzes drug relationships using Graph Neural Networks and ensemble learning techniques. The integration of graph-based representations with ensemble classifiers significantly improved prediction accuracy, robustness, and model generalization. The framework also provides dynamic side-effect analysis, enabling the identification of possible adverse reactions associated with risky drug combinations. Experimental evaluation demonstrated that the proposed system achieved approximately 93–94% accuracy with an F1-score of 0.96, outperforming several baseline machine learning approaches. The use of multi-feature integration, hard-negative sampling, dimensionality reduction, and ensemble optimization contributed to improved predictive performance and efficient handling of high-dimensional biomedical datasets. The final system was successfully deployed as a web-based application using Flask and React, enabling real-time prediction of drug interactions and side-effect risks. The proposed framework can serve as an intelligent clinical decision support system to assist healthcare professionals in safer prescription management and improved patient safety. Future enhancements of the system include incorporating real-world clinical datasets, improving interpretability using explainable AI techniques, integrating advanced Graph Attention Networks (GATs), and targeting prediction accuracies above 98%. The framework can also be extended to support personalized medicine and large-scale healthcare analytics.

References

[1] Breiman, L., “Random Forests,” Machine Learning, vol. 45, no. 1, pp. 5–32, 2001. [2] Chen, T. and Guestrin, C., “XGBoost: A Scalable Tree Boosting System,” Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2016. [3] Deng, Y., et al., “A multimodal deep learning framework for predicting drug–drug interaction events,” Bioinformatics, 2020. [4] Hamilton, W., Ying, Z., and Leskovec, J., “Inductive Representation Learning on Large Graphs,” Advances in Neural Information Processing Systems (NeurIPS), 2017. [5] Kipf, T.N. and Welling, M., “Semi-Supervised Classification with Graph Convolutional Networks,” International Conference on Learning Representations (ICLR), 2017. [6] Kuhn, M., et al., “SIDER: The Side Effect Resource,” Nucleic Acids Research, 2016. [7] Pedregosa, F., et al., “Scikit-learn: Machine Learning in Python,” Journal of Machine Learning Research, 2011. [8] Vilar, S., et al., “Drug-drug interaction through molecular structure similarity analysis,” Journal of the American Medical Informatics Association, 2012. [9] Wishart, D.S., et al., “DrugBank: A Comprehensive Resource for In Silico Drug Discovery and Exploration,” Nucleic Acids Research, 2018. [10] Zhang, W., et al., “Predicting drug-drug interactions using graph-based approaches,” Bioinformatics, 2017.

Copyright

Copyright © 2026 Ganesh Venkat. 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.