Application of two-part statistics for comparison of sequence variant counts

Investigation of microbial communities, particularly human associated communities, is significantly enhanced by the vast amounts of sequence data produced by high throughput sequencing technologies. However, these data create high-dimensional complex data sets that consist of a large proportion of zeros, non-negative skewed counts, and frequently, limited number of samples. These features distinguish sequence data from other forms of high-dimensional data, and are not adequately addressed by statistical approaches in common use. Ultimately, medical studies may identify targeted interventions or treatments, but lack of analytic tools for feature selection and identification of taxa responsible for differences between groups, is hindering advancement. The objective of this paper is to examine the application of a two-part statistic to identify taxa that differ between two groups. The advantages of the two-part statistic over common statistical tests applied to sequence count datasets are discussed. Results from the t-test, the Wilcoxon test, and the two-part test are compared using sequence counts from microbial ecology studies in cystic fibrosis and from cenote samples. We show superior performance of the two-part statistic for analysis of sequence data. The improved performance in microbial ecology studies was independent of study type and sequence technology used.