DrugPred: an EdgeConv-GNN and Bio_ClinicalBERT based polypharmacy ADR prediction and specialist recommendation model

Abstract

Adverse drug reactions (ADRs) are caused by medication and are considered a serious issue in healthcare when there is simultaneous use of different medications resulting in drug-drug interaction (DDI). Traditional approaches mostly focus on the effects caused by a single drug, and they fail to capture the side effects from drug combinations. In this research work, a deep learning-based DrugPred framework is proposed to predict ADR risks by integrating individual drug effects, interaction statistics of the drugs, and structural relationships between drug-interaction graphs. In the first phase, baseline association scores between individual drugs and their corresponding ADRs, such as nausea, liver toxicity, or cardiovascular effects, are learned using a multi-layer perceptron (MLP) trained on the OFFSIDES dataset, providing probabilistic values for individual drug-ADR pairs. In the second phase, a Graph Neural Network (GNN) based on the EdgeConv architecture is utilized to model drug-drug interactions using the TwoSIDES dataset, capturing relational dependencies by utilizing pairwise statistical values such as co-occurrence frequency and interaction statistics derived from pharmacovigilance signals, including transformed PRR-based signals. In the third phase, Bio_ClinicalBERT, which is a pre-trained transformer-based language model specifically trained on clinical notes and biomedical text, is used to encode drug pair representations and combine them with association and interaction scores through an attention-guided mechanism, which enables the model to adaptively weight heterogeneous features and finally performs multi-label ADR prediction. The proposed model achieves an accuracy of 95.73%, F1-score of 0.94, ROC-AUC of 0.993, and PR-AUC of 0.988. These results indicate that our proposed DrugPred framework is effective in the prediction of ADR risks with high precision and recall. Overall, the proposed framework provides a scalable approach for predicting ADRs in multi-drug settings. We further extend our framework with a retrieval-based guidance system that maps the risk levels of the predicted ADRs to appropriate System Organ Classes (SOC) and medical specialists to provide recommendations using real-world clinical data sources like PubMed and OpenFDA.