Enhancing Pharmacological Interaction Prediction in Otolaryngology Using Computational Methods

Abstract

Drug-Drug Interactions (DDIs) cause major clinical challenges leading to adverse drug reactions, lower the therapeutic value of drugs, and increase the overall healthcare burden, especially in the otolaryngology domain where patients are frequently prescribed multiple medications such as antibiotics, antihistamines, analgesics, and NSAIDs. The traditional methods for identifying DDIs based on clinical observation and experimental validation are time consuming and expensive. To solve these challenges, computational deep learning based models are proposed, ENT Multimodal Convolutional Neural Network for Drug-Drug Interactions (ENTMCNN- DDI ) and ENT Graph Neural Network (ENT-GNN-DDI) for Drug-Drug Interactions to enhance the prediction of drug-drug interactions in the otolaryngology domain. These frameworks use the drug features, chemical substructures, targets, enzymes, and pathways from the DrugBank dataset and side effects from TwoSIDES. Filtering based on the domain is specific to otolaryngology, and it is by the extraction of the keywords from the indication field of DrugBank that able to use them to determine the relevance. The ENT-MCNN-DDI model gets these multimodal features through independent 1D CNN layers, merges the resultant embeddings for interaction prediction. The ENT-GNN-DDI model uses graph-based learning where drugs are considered as nodes and interactions as edges thereby enabling graph convolutional networks (GCNs) to conduct relational reasoning and to capture the relations and the neighborhood information. Both models were trained and tested with balanced positive and negative sampling using a subset of drugs specifically for otolaryngology. Based on a stratified train validation split (80 : 20) experiment, the model performance is quite stable. The ENT-MCNN-DDI achieved an Accuracy of 0.94, Precision of 0.94, Recall of 0.93, F1-score of 0.94, AUC of 0.986, and AUPR of 0.984. The ENT-GNN-DDI slightly outperformed well by achieving an Accuracy of 0.95, Precision of 0.95, Recall of 0.94, F1-score of 0.95, AUC of 0.987, and AUPR of 0.985. These results shows the effectiveness of the ENT-MCNN-DDI and ENT-GNN-DDI models in accurately predicting domain specific DDIs, offering a valuable tools for clinicians and researchers to minimize the occurrence of adverse drug events in otolaryngology.