[Paper Review] A Bayesian Zero-Inflated Negative Binomial Regression Model for the Integrative Analysis of Microbiome Data

Shuang Jiang, Guanghua Xiao+3 more|arXiv (Cornell University)|Dec 23, 2018

Gut microbiota and health4 citations

TL;DR

This paper proposes a novel Bayesian zero-inflated negative binomial (ZINB) regression model that jointly identifies differentially abundant microbial taxa across multiple groups and quantifies their associations with covariates such as metabolites and host factors. By integrating overdispersion, zero-inflation, and covariate effects within a hierarchical mixture model using spike-and-slab priors, the method improves detection power and biological interpretability in microbiome-integrative studies, as demonstrated in real datasets linking Bifidobacterium to improved immunotherapy outcomes.

ABSTRACT

Microbiome `omics approaches can reveal intriguing relationships between the human microbiome and certain disease states. Along with the identification of specific bacteria taxa associated with diseases, recent scientific advancements provide mounting evidence that metabolism, genetics and environmental factors can all modulate these microbial effects. However, the current methods for integrating microbiome data and other covariates are severely lacking. Hence, we present an integrative Bayesian zero-inflated negative binomial regression model that can both distinguish differentially abundant taxa with distinct phenotypes and quantify covariate-taxa effects. Our model demonstrates good performance using simulated data. Furthermore, we successfully integrated microbiome taxonomies and metabolomics in two real microbiome datasets to provide biologically interpretable findings. In all, we proposed a novel integrative Bayesian regression model that features bacterial differential abundance analysis and microbiome-covariate effects quantifications, which makes it suitable for general microbiome studies.

Motivation & Objective

• To address the lack of statistical methods that simultaneously identify differentially abundant taxa and quantify microbiome-covariate associations in high-throughput sequencing data.
• To model the overdispersion and excess zeros common in microbiome count data using a zero-inflated negative binomial framework.
• To incorporate multiple covariates—such as metabolites, antibiotics, and host genetics—into a unified regression model for integrative analysis.
• To improve statistical power and reduce false discovery rates in microbiome studies through Bayesian feature selection with spike-and-slab priors.
• To enable biologically interpretable, systems-level insights into host-microbiome interactions in complex diseases like melanoma and inflammatory bowel disease.

Proposed method

• The model uses a hierarchical mixture model combining a point-mass at zero and a negative binomial distribution to account for excess zeros and overdispersion in microbiome count data.
• It employs a log-linear regression structure to model the mean of the ZINB distribution, incorporating group effects, covariate effects, and sequencing depth via offset terms.
• Spike-and-slab priors are applied to features (taxa) to perform Bayesian feature selection, enabling identification of both differentially abundant taxa and significant covariate-taxa associations.
• Posterior inference is conducted using a Gibbs sampling-based MCMC algorithm to estimate model parameters and compute posterior probabilities for each feature.
• The method computes Bayesian false discovery rates (FDR) to control for multiple testing in high-dimensional microbiome data.
• The model is extended to handle multiple phenotype groups by allowing group-specific parameters while maintaining consistent normalization and inference procedures.

Experimental results

Research questions

• RQ1 Can a unified statistical model jointly detect differentially abundant microbial taxa across multiple phenotypic groups and quantify their associations with host covariates such as metabolites and genetics?
• RQ2 How does the inclusion of zero-inflation and overdispersion improve the accuracy of differential abundance detection in microbiome count data compared to standard Poisson or negative binomial models?
• RQ3 To what extent does the Bayesian feature selection with spike-and-slab priors enhance statistical power and reduce false discovery rates in integrative microbiome analysis?
• RQ4 Can the model uncover biologically meaningful associations between specific taxa and clinical or molecular covariates in real-world microbiome datasets?
• RQ5 How robust is the model in detecting known biological relationships, such as the role of Bifidobacterium in melanoma immunotherapy response, when integrated with metabolomic data?

Key findings

• The model successfully identified Bifidobacterium as a responder-enriched taxon in a metastatic melanoma cohort, with strong negative correlations to oncogenic metabolites like 2-oxoarginine and 2-hydroxypalmitate, supporting prior biological findings.
• In simulated data, the model demonstrated superior performance in detecting differentially abundant taxa and covariate-taxa associations compared to existing methods, particularly under high zero-inflation and overdispersion.
• The Bayesian posterior probabilities and FDR control enabled reliable identification of biologically relevant features without imposing phylogenetic constraints, yielding interpretable results clustered within the same phylogenetic branches.
• The method achieved fast posterior inference, with analysis of two real MSS datasets completed in minutes using an efficient MCMC algorithm.
• The model revealed novel associations between microbial taxa and metabolites, such as the suppression of fatty-acid metabolites linked to oncogenic signaling, suggesting potential mechanisms for immunotherapy response.
• The framework is extensible to multiple phenotype groups and can be adapted to model microbiome effects on host omics, such as chromatin accessibility and gene regulation, enabling bidirectional systems biology analysis.