Abstract

Microorganisms play a crucial role in human health and disease. Identifying potential microbe- drug associations is essential for drug discovery and clinical treatment. In this manuscript, we proposed a novel prediction model named GVNNVAE by combining an Improved Graph Neural Network (GNN) and the Variational Auto-Encoder (VAE) to infer potential microbe-drug associations. In GVNNVAE, we first established a heterogeneous microbe-drug network N by integrating multiple similarity metrics of microbes, drugs, and diseases. Subsequently, we introduced an improved GNN and the VAE to extract topological and attribute representations for nodes in N respectively. Finally, through incorporating various original attributes of microbes and drugs with above two kinds of newly obtained topological and attribute representations, predicted scores of potential microbe-drug associations would be calculated. Furthermore, To evaluate the prediction performance of GVNNVAE, intensive experiments were done and comparative results showed that GVNNVAE could achieve a satisfactory AUC value of 0.9688, which outperformed existing competitive state-of-the-art methods. And moreover, case studies of known microbes and drugs confirmed the effectiveness of GVNNVAE as well, which highlighted its potential for predicting latent microbe-drug associations.

Keywords: Microbe-Drug associations prediction, Heterogeneous network, Graph Neural Network, computational models, Variational Auto-Encoder, Topological and attribute representations.