Integrating microbiome and metabolome data for host disease prediction via deep neural networks

The concurrent profiles of the gut microbiome and metabolome can be used in the diagnosis of complex diseases. However, the establishment of robust predictive models is challenging due to the high dimensionality of data and complex interactions among microbiome, metabolites, and host. Using deep neural networks consisting of an autoencoder for extracting latent representations and a multilayer neural network for disease prediction, we show that gut metabolome is more predictive of inflammatory bowel disease (IBD) than gut microbiome. In addition, we design a new multi-task autoencoder to extract the latent profiles from the combined microbiome and metabolome data. We further demonstrate that the combined latent profiles can further improve the performance of prediction. In summary, our work shows that autoencoders are useful apparatuses in generating low dimensional profiles that contribute to the improved performance and robustness for IBD prediction.