A phylogeny-free microbiome dysbiosis detection pipeline for non-invasive disease diagnosis

The microbial diversity and taxonomic profiles of human microbiome carries indicative signals associated with several complex human diseases. Methods quantifying and differentiating profiles belonging to healthy/disease microbiomes have a potential to be used as non-invasive diagnostic tools. 16S rRNA sequencing is a currently popular and feasible technology to generate the required data for such diagnostic. However, current methodology to analyze 16S rRNA data requires biologically motivated approaches including molecular database search and phylogenetic tree construction, that results in losing valuable information related to dysbiotic signals. In this study, information-theoretic measures for microbiome similarity calculation that avoid conventional projections are introduced. A computational pipeline joining the information of three novel similarity measures in an implicit feature space and consequently filtering redundant features is defined. The extracted features are finally exposed to classifiers predicting disease/healthy states. Experiments conducted on previously defined Crohn's Disease, Inflammatory Bowel Disease, Parkinson's Disease, Autistic Children, Colorectal Cancer datasets revealed that the proposed framework can achieve improved detection accuracy over conventional microbiome analysis approaches.