Microbe-drug association prediction model based on graph convolution and attention networks

Abstract

The human microbiome plays a key role in drug development and precision medicine, but understanding its complex interactions with drugs remains a challenge. Identifying microbe-drug associations not only enhances our understanding of their mechanisms but also aids in drug discovery and repurposing. Traditional experiments are expensive and time-consuming, making computational methods for predicting microbe-drug associations a new trend. Currently, computational methods specifically designed for this task are still scarce. Therefore, to address the shortcomings of traditional experimental methods in predicting potential microbe-drug associations, this paper proposes a new prediction model named GCNATMDA. The model combines two deep learning models, Graph Convolutional Network and Graph Attention Network, and aims to reveal potential relationships between microbes and drugs by learning related features. Thus improve the efficiency and accuracy of prediction. We first integrated the microbe-drug association matrix from the existing dataset, and then combined the calculated microbe-drug characteristic matrix as the model input. The GCN module is used to dig deeper into the potential characterization of microbes and drugs, while the GAT module further learns the more complex interactions between them and generates the corresponding score matrix. The experimental results show that the GCNATMDA model achieves 96.59% and 93.01% in AUC and AUPR evaluation indexes, respectively, which is significantly better than the existing prediction models. In addition, the reliability of the prediction results is verified by a series of experiments.