Clustering on Human Microbiome Sequencing Data: A Distance-Based Unsupervised Learning Model

Modeling and analyzing human microbiome allows the assessment of the microbial community and its impacts on human health. Microbiome composition can be quantified using 16S rRNA technology into sequencing data, which are usually skewed and heavy-tailed with excess zeros. Clustering methods are useful in personalized medicine by identifying subgroups for patients stratification. However, there is currently a lack of standardized clustering method for the complex microbiome sequencing data. We propose a clustering algorithm with a specific beta diversity measure that can address the presence-absence bias encountered for sparse count data and effectively measure the sample distances for sample stratification. Our distance measure used for clustering is derived from a parametric based mixture model producing sample-specific distributions conditional on the observed operational taxonomic unit (OTU) counts and estimated mixture weights. The method can provide accurate estimates of the true zero proportions and thus construct a precise beta diversity measure. Extensive simulation studies have been conducted and suggest that the proposed method achieves substantial clustering improvement compared with some widely used distance measures when a large proportion of zeros is presented. The proposed algorithm was implemented to a human gut microbiome study on Parkinson's diseases to identify distinct microbiome states with biological interpretations.