Identification of microbial interaction network: zero-inflated latent Ising model based approach

摘要

The human microbiome, the collection of trillions of microbial organisms that live in our body spaces, belong to one of thousands of different species [1, 2]. The organisms that inhabit the human gut are an additional source of genetic diversity that can influence metabolism and modulate drug interactions [3]. Recent advances in genomic technologies enable production of thousands of 16S rRNA sequences per sample [4] and are powerful tools to explore the basic biology about human microbiome. Nevertheless, analyzing microbiome data and converting them into meaningful biological insights are still challenging tasks. First, the observed absolute abundance in sequencing experiment cannot inform the real absolute abundance of molecules in the sample which can be attributed to the sequence depth associated with the experiment. Multiple normalization methods have been proposed in literature to solve this problem among which total sum scaling (TSS) has been widely used in practice [5–9]. TSS scales each sample by the total read count and yields the relative abundance. However, the statistical analysis based on relative abundance can easily lead to spurious association due to the unit-sum constraint [10–14]. Further complicating the analysis of microbiome data is the zero-inflated distribution of read count [3]. As for the dataset in “Restults from the relative abundance of gut microbiota” section, among the 134 taxa, there are only 6 taxa for which the proportions of nonzero observations are greater than 80%. Zero inflation stems from the fact that the majority of the amplicon sequence variants (ASVs) either physically do not exist in the subject or are below the detection threshold for the given sample [2]. Another hurdle for analyzing the microbiome data is its high-dimensionality which usually involves hundreds of microbes; consequently, models equipped for this modeling task should be employed.