Beyond Structural Representations: A Semantic-Relational Graph Neural Network for Polypharmacy Side Effect Prediction

The concurrent use of several medications, known as polypharmacy, has increased in prevalence, especially among older adults with complex health conditions. Despite its effectiveness, this therapeutic strategy increases the risk of Polypharmacy Side Effects (PSEs), leading to substantial clinical and economic burden. Recent deep learning methods have integrated graph neural networks with pretrained language models to predict PSEs. However, these approaches rely on relation-agnostic decoders and fail to consider the pharmacological context of each PSE type. To address these issues, we propose SIREN, a semantic-relational graph neural network that integrates semantic embeddings from ChemBERTa and BioBERT with a multirelational biomedical graph structure. The model employs a relation-specific decoder to adapt the score of each drug pair based on the target side effect type. Experimental evaluation on the TWOSIDES dataset demonstrated high AUROC and AUPRC, outperforming current state-of-the-art models under the transductive setting and remaining competitive under inductive settings. Ablation studies further validated the importance of the relation-specific decoding mechanism and pre-trained node embeddings. By integrating semantic representations with graph-based reasoning, this study establishes an innovative framework for data-driven pharmacovigilance, thereby enhancing patient safety, reducing healthcare costs, and supporting regulatory compliance.