BANSMDA: a computational model for predicting potential microbe-disease associations based on bilinear attention networks and sparse autoencoders

Introduction Predicting the relationship between diseases and microbes can significantly enhance disease diagnosis and treatment, while providing crucial scientific support for public health, ecological health, and drug development. Methods In this manuscript, we introduce an innovative computational model named BANSMDA, which integrates Bilinear Attention Networks with sparse autoencoder to uncover hidden connections between microbes and diseases. In BANSMDA, we first constructed a heterogeneous microbe-disease network by integrating multiple Gaussian similarity measures for diseases and microbes, along with known microbe-disease associations. And then, we employed a BAN-based autoencoder and a sparse autoencoder module to learn node representations within this newly constructed heterogeneous network. Finally, we evaluated the prediction performance of BANSMDA using a 5-fold cross-validation framework. Conclusion Experiments results showed that BANSMDA achieved superior performance compared to other cutting-edge methods. To further assess its effectiveness, we carried out case studies on two common diseases (including Asthma and Colorectal carcinoma) and two important microbial genera (including Escherichia and Bacteroides ), and in the top 20 predicted microbes, there were 19 and 20 having been confirmed by published literature respectively. Besides, in the top 20 predicted diseases, there were 19 and 19 having been confirmed by published literature separately. Therefore, it is easy to conclude that BANSMDA can achieve satisfactory prediction ability.