A New Association Analysis Method for Longitudinally Measured Microbial Compositional Data Using Latent Dirichlet Allocation Model

Abstract

In recent years, analysis methods of microbiome data are developing rapidly, and many methods for the microbial compositional data which uses the 16S ribosomal RNA gene (16S rRNA data) are proposed. But, methods of association analysis for longitudinally measured 16S rRNA data are not studied well. Latent dirichlet allocation model (LDA) which is used mainly in natural language processing and has high expansion possibilities came to be applied to 16S rRNA data analysis in the past few years. Then, we propose an association analysis method by modifying existing LDA: topic tracking model for longitudinal 16S rRNA data. As the result of predictive performance evaluation, proposed method showed superior performance compared with topic tracking model with regard to perplexity. We applied this method to microbial data of rural Japanese people and identified topics associated with obesity.