Dirichlet-tree multinomial mixtures for clustering microbiome compositions

Jialiang Mao, Li MaThe Annals of Applied Statistics 2022 · DOI

TL;DR

This paper introduces Dirichlet-tree multinomial mixtures (DTMM), a novel Bayesian model for clustering microbiome data. DTMM leverages the phylogenetic relationships between microbial taxa to better capture complex variations within and between sample groups, outperforming existing methods in simulations and real-world microbiome datasets. This approach can reveal subtle microbial community structures that are missed by traditional methods, potentially leading to more accurate diagnoses and personalized treatments.

What Problem Does This Paper Address?

Analyzing the human microbiome involves understanding the vast community of microbes living within us. A key task is to group samples (e.g., from different individuals or time points) into distinct "enterotypes" or subtypes based on their microbial composition. However, existing clustering methods struggle with microbiome data due to its unique characteristics:

  • High dimensionality and sparsity: Microbiome data often involves thousands of microbial taxa (Operational Taxonomic Units or Amplicon Sequence Variants), with many present in very low abundance.
  • Compositional nature: The data represents proportions of different taxa, meaning the sum of counts for all taxa in a sample is fixed (related to sequencing depth).
  • Complex variability: The relationships between taxa are not independent; they are influenced by evolutionary history (phylogeny) and ecological interactions. Standard distance-based methods (like k-means) and even existing model-based methods (like Dirichlet-multinomial mixtures or DMM) often fail to account for this complex, tree-structured variability, leading to overly dispersed or inaccurate clusters.

This paper aims to develop a more robust and flexible clustering framework that can better handle these challenges and uncover more biologically meaningful microbial community structures.

Methods

The core contribution is the Dirichlet-tree multinomial mixtures (DTMM) model. Here's how it works and what's new:

  • Bayesian Generative Model: DTMM is a probabilistic model that assumes microbiome compositions are generated from a mixture of underlying distributions.
  • Dirichlet-Tree (DT) Distribution: Instead of the standard Dirichlet distribution used in DMM, DTMM employs a Dirichlet-tree distribution. This distribution naturally incorporates the phylogenetic tree structure of the microbial taxa. The tree structure allows for a more flexible modeling of how variations in different taxa relate to each other, capturing dependencies that are missed by simpler models.
  • Hierarchical Structure: The DT distribution models the probabilities of taxa counts hierarchically along the phylogenetic tree. This allows for multiple dispersion parameters, one for each internal node of the tree, providing a richer covariance structure than the single dispersion parameter in the standard Dirichlet distribution.
  • Model Selection for Taxa: DTMM includes a mechanism to identify a subset of "signature taxa" that are most important for distinguishing between clusters. This is crucial because enterotypes are not always defined by the most abundant species but can be driven by specific combinations of less abundant taxa. This feature enhances interpretability and sensitivity.
  • Efficient Inference: The authors develop a blocked Gibbs sampler for posterior inference, which is designed to be computationally efficient by marginalizing out high-dimensional parameters.

Datasets

The paper evaluates DTMM using:

  • Extensive Simulation Studies: The authors generate synthetic microbiome datasets under various scenarios to rigorously test DTMM's performance against other methods. These simulations vary the signal strength and the underlying data-generating distribution (including Dirichlet-tree, Dirichlet, and logistic-normal distributions).
  • American Gut Project (AGP) Data: A large, real-world dataset of fecal microbiome samples from individuals diagnosed with Inflammatory Bowel Disease (IBD) and diabetes. This dataset is used for a case study to identify enterotypes associated with these conditions.
  • Dethlefsen and Relman (2011) Longitudinal Data: A publicly available dataset studying the response of gut microbiota to antibiotic treatment in three patients. This dataset is used for validation to confirm the biological relevance of the clusters found by DTMM.

Key Insights and Ideas

  • Phylogeny Matters for Clustering: The paper strongly suggests that incorporating the evolutionary relationships (phylogenetic tree) of microbes is crucial for accurately clustering microbiome compositions. Ignoring this structure leads to less reliable results.
  • Enterotypes are Complex: Contrary to some previous assumptions, enterotypes are not always defined by a few dominant species. DTMM can reveal that clusters are often formed by a sophisticated interplay of many taxa, including less abundant ones, which are missed by other methods.
  • Model Selection Enhances Interpretability: The ability of DTMM to identify "signature taxa" that define clusters is a significant advantage, making the results more interpretable and biologically relevant.
  • DTMM's Flexibility: The Dirichlet-tree distribution provides a more flexible covariance structure, allowing DTMM to better model the complex variations observed in microbiome data, especially when signals are localized within specific branches of the phylogenetic tree.

Key Results

  • Simulation Performance: In simulation studies, DTMM consistently performed as one of the top two methods, often outperforming others, especially when the signal was weak or localized within specific parts of the phylogenetic tree. Even when the true model was a simpler Dirichlet distribution, DTMM could outperform DMM, particularly with smaller sample sizes.
  • AGP Case Study: DTMM identified enterotypes in IBD and diabetes patients that were characterized by a large number of OTUs jointly, rather than just a few dominant ones. This highlights the nuanced nature of these microbial communities.
  • Validation Study: DTMM successfully identified biologically meaningful clusters in the longitudinal antibiotic treatment data, reflecting distinct states of the gut microbiota (stable, sterile, recover).
  • Comparison with DMM: In case studies, DTMM often found more refined clusters than DMM, suggesting that DMM's simpler model struggles to capture the finer-grained variations that DTMM can resolve.

Limitations and Future Work

  • Computational Complexity: While the authors developed an efficient Gibbs sampler, Bayesian inference for complex models can still be computationally intensive, especially with very large datasets.
  • Generalizing Covariance Structure: The paper notes that while DT offers richer covariance than Dirichlet, it's still limited compared to other families like the logistic-normal (LN). Future work could explore generalizing the covariance structure further without overly complicating inference.
  • Cluster-Dependent Node Selection: Currently, DTMM selects nodes relevant to any cluster. An interesting future direction is to allow nodes to be selected in a cluster-dependent manner, meaning different clusters might rely on different sets of nodes for their definition.
  • Supervised Learning: The paper briefly mentions the potential for DTMM in supervised classification, which could be further explored.

Summary updated

Abstract

Studying the human microbiome has gained substantial interest in recent years, and a common task in the analysis of these data is to cluster microbiome compositions into subtypes. This subdivision of samples into subgroups serves as an intermediary step in achieving personalized diagnosis and treatment. In applying existing clustering methods to modern microbiome studies including the American Gut Project (AGP) data, we found that this seemingly standard task, however, is very challenging in the microbiome composition context due to several key features of such data. Standard distance-based clustering algorithms generally do not produce reliable results as they do not take into account the heterogeneity of the cross-sample variability among the bacterial taxa, while existing model-based approaches do not allow sufficient flexibility for the identification of complex within-cluster variation from cross-cluster variation. Direct applications of such methods generally lead to overly dispersed clusters in the AGP data and such phenomenon is common for other microbiome data. To overcome these challenges, we introduce Dirichlet-tree multinomial mixtures (DTMM) as a Bayesian generative model for clustering amplicon sequencing data in microbiome studies. DTMM models the microbiome population with a mixture of Dirichlet-tree kernels that utilizes the phylogenetic tree to offer a more flexible covariance structure in characterizing within-cluster variation, and it provides a means for identifying a subset of signature taxa that distinguish the clusters. We perform extensive simulation studies to evaluate the performance of DTMM and compare it to state-of-the-art model-based and distance-based clustering methods in the microbiome context, and carry out a validation study on a publicly available longitudinal data set to confirm the biological relevance of the clusters. Finally, we report a case study on the fecal data from the AGP to identify compositional clusters among individuals with inflammatory bowel disease and diabetes. Among our most interesting findings is that enterotypes (i.e., gut microbiome clusters) are not always defined by the most dominant species as previous analyses had assumed, but can involve a number of less abundant OTUs, which cannot be identified with existing distance-based and method-based approaches.
BibTeX
@article{mao2022dirichlet,
  author = {Jialiang Mao and Li Ma},
  title = {{Dirichlet-tree multinomial mixtures for clustering microbiome compositions}},
  journal = {The Annals of Applied Statistics},
  year = {2022},
  doi = {10.1214/21-aoas1552},
}
Figure