A0510
Title: BMDD: A bimodal Dirichlet distribution for microbiome data imputation
Authors: Huijuan Zhou - Shanghai University of Finance and Economics (China) [presenting]
Jun Chen - Mayo Clinic (United States)
Xianyang Zhang - Texas A\&M University (United States)
Abstract: Microbiome sequencing data contain excessive zeros due to the low abundance or physical absence of most taxa. Zeros are problematic when analyzing the data on the logarithmic scale, which is common practice when analyzing abundance data. To tackle this challenge, a probabilistic modeling approach is developed based on a new type of informative prior, termed bimodal Dirichlet distribution (BMDD), to capture the essential distributional characteristics of the compositional data. Multiple posterior samples are obtained as the imputed compositional data. Through extensive numerical studies, it is demonstrated that the new approach generates more accurate composition estimates that reflect the underlying true compositions. It is further shown that the multiple imputation based on the posterior samples from the method produces more robust results in differential abundance analysis of microbiome data.
