Bayesian Graphical Compositional Regression for Microbiome Data

Jialiang Mao, Yuhan Chen, Li MaJournal of the American Statistical Association 2017 · arXiv

TL;DR

This paper introduces a novel Bayesian framework, BGCR, for analyzing microbiome data to detect differences between groups. It leverages phylogenetic relationships and graphical models to improve statistical power and incorporates covariates to account for confounding factors, offering a more robust and interpretable approach to microbiome comparisons.

What Problem Does This Paper Address?

Microbiome studies often aim to identify differences in microbial communities between different groups of individuals (e.g., healthy vs. diseased, different diets). However, analyzing microbiome data presents significant challenges:

  • High Dimensionality and Sparsity: The number of microbial taxa (often represented as Operational Taxonomic Units or OTUs) can be very large, and many OTUs are present in very low or zero counts across samples.
  • Within-Group Heterogeneity: Samples within the same group can exhibit substantial variation in their microbial composition.
  • Confounding Variables: Observational studies, common in microbiome research, are prone to confounding factors (e.g., diet, age, sex) that can lead to false discoveries or mask true differences if not properly accounted for.
  • Complex Relationships: The evolutionary and functional relationships between different microbial taxa are complex and not always captured by simple statistical tests.

Existing methods often struggle with these challenges, leading to underpowered tests, increased false positives, and difficulty in interpreting results. This paper proposes a method to address these issues by integrating phylogenetic information, graphical models, and a Bayesian framework that handles within-group variation and covariates.

Methods

The core of the proposed method, Bayesian Graphical Compositional Regression (BGCR), is built upon three key ideas:

  1. Phylogenetic Tree-Based Decomposition: The problem of comparing microbiome composition across all taxa is broken down into a series of smaller, "local" tests performed at each internal node of a phylogenetic tree. This tree represents the evolutionary relationships between OTUs.
  2. Graphical Model for Information Sharing: A graphical model is used to link these local tests. This allows for "borrowing information" across related taxa, meaning that the evidence from one test can inform others, particularly those on neighboring nodes in the phylogenetic tree. This is implemented using a "bottom-up" autoregressive structure.
  3. Bayesian Testing Strategy: A Bayesian approach is employed to:
  4. Integrate out Within-Group Variation: Instead of estimating the often unstable within-group variation, BGCR integrates it out, leading to more robust inference.
  5. Incorporate Covariates: A regression adjustment is incorporated at each node to account for potential confounding variables, reducing false positives and improving power.
  6. Probabilistic Characterization: The method provides a fully probabilistic assessment of the evidence for differences, quantified by Posterior Marginal Alternative Probabilities (PMAPs).

The inference is performed using an efficient message-passing algorithm on a clique tree derived from the graphical model, avoiding computationally intensive Monte Carlo simulations.

Datasets

The paper uses the American Gut Project dataset for its application. This dataset contains microbiome samples from a large number of participants, along with extensive covariate information (demographics, dietary habits, etc.) and a phylogenetic tree for the OTUs. For simulations, synthetic datasets are generated based on this data.

Key Insights and Ideas

  • Phylogeny as a Structure for Inference: The evolutionary relationships between microbes are not just descriptive but can be actively used to improve statistical inference by structuring the testing process and enabling information sharing.
  • "Chaining" of Signals: Cross-group differences in microbiome composition often manifest as "chains" along the phylogenetic tree, rather than isolated differences at single taxa. BGCR's graphical structure is designed to capture this phenomenon.
  • Bayesian Integration for Robustness: Integrating out nuisance parameters like dispersion, rather than estimating them, leads to more stable and reliable results, especially in sparse and heterogeneous microbiome data.
  • Covariates are Crucial for Observational Data: For microbiome data from observational studies, failing to adjust for confounders can drastically alter conclusions and lead to false discoveries. BGCR's ability to incorporate covariates is a significant advantage.
  • Visualizing Differences: The PMAPs generated by BGCR can be visualized on the phylogenetic tree, providing an intuitive way to understand where and how the microbial communities differ between groups.

Key Results

  • Improved Power: BGCR consistently outperforms existing methods (DM, DTM-1, DTM-3) in terms of statistical power across various simulation scenarios, especially as the number of OTUs increases and the data becomes sparser.
  • Effective Covariate Adjustment: In simulations involving confounding variables (e.g., gender), BGCR effectively controls false positives when covariates are included, while other methods fail to do so.
  • Meaningful Biological Insights: Application to the American Gut Project data revealed significant associations between dietary habits (fruit, seafood, vegetable, whole grain) and gut microbiome composition. The analysis also highlighted how adjusting for different sets of covariates can dramatically change the conclusions, underscoring the importance of careful confounder analysis.
  • Scalability: The proposed message-passing algorithm is scalable to the size of the phylogenetic tree, allowing for analysis of large microbiome datasets.

Limitations and Future Work

  • Covariate Selection Caution: While BGCR can incorporate covariates, the paper cautions against using standard variable selection methods, as they might inadvertently remove confounders, leading to biased results. The authors suggest careful consideration and provide discussion in supplementary materials.
  • Generalization to Multiple Groups: The current framework focuses on two-group comparisons but is noted to be generalizable to multiple group settings with minimal effort.
  • Model Assumptions: The paper assumes a common dispersion parameter for both groups and a "common slope" for covariates. While these simplify computation, deviations might require extensions.
  • Future Work: The authors suggest further exploration of covariate selection strategies and generalization to multi-group comparisons. The framework's ability to provide a probabilistic characterization of differences also opens avenues for more nuanced biological interpretation.

Summary updated

Abstract

An important task in microbiome studies is to test the existence of and give characterization to differences in the microbiome composition across groups of samples. Important challenges of this problem include the large within-group heterogeneities among samples and the existence of potential confounding variables that, when ignored, increase the chance of false discoveries and reduce the power for identifying true differences. We propose a probabilistic framework to overcome these issues by combining three ideas: (i) a phylogenetic tree-based decomposition of the cross-group comparison problem into a series of local tests, (ii) a graphical model that links the local tests to allow information sharing across taxa, and (iii) a Bayesian testing strategy that incorporates covariates and integrates out the within-group variation, avoiding potentially unstable point estimates. We derive an efficient inference algorithm based on numerical integration and junction-tree message passing, conduct extensive simulation studies to investigate the performance of our approach, and compare it to state-of-the-art methods in a number of representative settings. We then apply our method to the American Gut data to analyze the association of dietary habits and human's gut microbiome composition in the presence of covariates, and illustrate the importance of incorporating covariates in microbiome cross-group comparison.
BibTeX
@article{mao2017bayesian,
  author = {Jialiang Mao and Yuhan Chen and Li Ma},
  title = {{Bayesian Graphical Compositional Regression for Microbiome Data}},
  journal = {Journal of the American Statistical Association},
  year = {2017},
}
Figure