A Graph-Based Transformer Neural Network for Multi-Label ADR Prediction

Abstract

Adverse drug reactions (ADRs) pose substantial health hazards and financial burdens on patients. Accurate prediction of these reactions has become crucial within the clinical domain to guarantee prompt intervention. Many techniques have been presented to predict ADRs based on the drug’s molecular structure. However, these techniques are limited to the transformation of a multi-label classification problem into multiple binary problems and the formation of distinct classifiers for individual drug reactions. Such techniques can be computationally expensive and time-consuming when dealing with a large no. of ADRs. Moreover, the multi-label classifier can learn associations between multiple related ADRs more effectively. Therefore, the objective of this research is the multi-label classification of adverse drug reactions by incorporating transformers-based graph neural networks (GNNs). This paper presents a new model called GTransfNN (graph-based transformer neural network) that leverages graphs with transformers to analyze the molecular structure of drugs. It aims to predict 27 ADR categories based on the system organ class. The proposed model introduces three key characteristics as its main components: First, it considers an attention mechanism that operates on the interconnectivity among neighboring nodes within the graph. Second, it incorporates edge features together with node features while calculating the attention weight for each node. Finally, it replaces layer normalization with batch normalization. The results indicate that the proposed model outperforms the other state-of-the-art models, such as neural fingerprint and Attentive_FP model, with notable increases of 10% and 18% in AUC, respectively. It achieves an AUC of 0.82 and an accuracy of 0.83 on the SIDER dataset. Similarly, it showcases steady performance enhancements on the ADRECS dataset, attaining an accuracy of 0.84 and an AUC of 0.82 by showcasing a 5%, 16%, and 25% increase in AUC as compared to iADRGSE, BERT_Smile, and Attentive_FP methods. These results show the model’s robustness and reliability across different datasets, thereby contributing to more effective drug safety assessments and health-care decision-making processes.