PGCN-DDI: Drug-Drug Interaction Prediction Based on Multidimensional Drug Features and Neighborhood Overlap Similarity

Wangli

Keywords: drug–drug interaction prediction, multidimensional features, neighborhood overlap similarity

TL;DR: We propose a PGCN-DDI model, which enhances the message passing function by neighborhood overlap similarity and improves the aggregator using Pearson correlation coefficient, finally enhancing the model’s capability to represent drug node features.

Abstract: The co-administration of drugs may lead to adverse drug interactions, posing risks to the organism. Therefore, predicting potential drug interactions is crucial. Compared to in vitro experiments and clinical trials for DDI prediction, computational methods are widely used due to their efficiency and other advantages. We propose a deep learning technique, namely the PGCN-DDI model, which enhances the traditional GCN’s message passing function through a neighborhood overlap similarity algorithm and improves the aggregator using the Pearson correlation coefficient, ultimately enhancing the model’s capability to represent drug node features. We utilize multidimensional features of drugs, including chemical substructure, metabolic pathways, targets, drug ingredients, and drug categories, as model inputs for DDI prediction. We conduct experiments on different feature sets to assess the amount of information contained in different features, leveraging PGCN-DDI to learn drug node features. The results indicate that the combination of drug features (i.e., chemical substructure, targets, and metabolic pathways) outperforms other features in DDI prediction. The experimental results demonstrate that the PGCN-DDI model achieves an accuracy of 0.8921, an AUC of 0.9458, and an AUPR of 0.9208, all of which show improvements over several baseline models.

Submission Number: 19