Enhancing Metagenome-based Disease Prediction by Unsupervised Binning Approaches

Metagenomic data from human microbiome is a novel data source to improve diagnosis and prognosis for human diseases. Nevertheless, since the number of considered features is much higher than the number of samples, we meet numerous challenges to perform a prediction task based on individual bacteria data. In addition, we face difficulties related to the very high complexity of different diseases. Deep Learning (DL) has been obtaining great success on major metagenomics problems related to Operational Taxonomic Unit (OTU)- clustering, and gene prediction, comparative metagenomics, assignment and binning of taxonomic. In this study, we introduce one-dimensional (1D) representations based on the unsupervised binning approaches and scaling algorithms to enhance the prediction performance for metagenome-based diseases using artificial neural networks. The proposed method is evaluated on seven microbial datasets related to six different diseases including Liver Cirrhosis, Colorectal Cancer, Inflammatory Bowel Disease (IBD), Type 2 Diabetes, Obesity and HIV with 2 types of data consisting of species abundance and read counts at the genus level. As shown from the results, the proposed method can improve the performance of Metagenome-based Disease Prediction.