The human microbiome consists of trillions of microorganisms, which form an intriguing biological network that influences host health. Recently, metagenomic sequencing has enabled the quantitative measurement of microbial abundances. However, the nature of microbiome data prohibits the application of existing network models.
We developed a novel hybrid method for inferring microbial networks. It consists a Bayesian zero-inflated negative binomial model for data normalization, and a likelihood-based graphical model to infer the microbial network. This hybrid approach achieves high computational efficiency and robustness. We extensively evaluated this model using datasets from both simulations and two real studies. We measured the correctly inferred microbial relationship using AUROC statistics.
In a cancer immune checkpoint blockade study, we discovered a pair of cooperative genera: Faecalibacterium and Holdemania, both enriched in therapy responders and confirmed by literature. In a colorectal cancer study, our method inferred a cluster consisting of Fusobacterium, Parvimonas, Prevotella and Gemella, all confirmed in the literature to be enriched in colorectal cancer patients.