A tree-based model for addressing sparsity and taxa covariance in microbiome compositional count data

Zhuoqun Wang, Jialiang Mao, Li MaarXiv 2021 · DOI · arXiv

Abstract

Microbiome compositional data are often high-dimensional, sparse, and exhibit pervasive cross-sample heterogeneity. Generative modeling is a popular approach to analyze such data, and effective generative models must accurately characterize these key features. While high-dimensionality and abundance of zeros have received much attention, existing models often lack flexibility in capturing complex cross-sample variability. This limitation can affect statistical efficiency and lead to misleading conclusions in tasks like differential abundance analysis, clustering, and network analysis. We introduce a generative model, the "logistic-tree normal" (LTN) model, which addresses this issue and effectively captures key characteristics of microbiome data, including abundance of zeros. LTN employs a tree-based decomposition to aggregate sparse taxa counts and uses a (multivariate) logistic-normal distribution at tree splits, allowing for flexible covariance adjustments among taxa as needed. The latent Gaussian structure of LTN enables the incorporation of multivariate analysis tools that enforce sparsity or low-rank covariance assumptions. As a versatile, fully generative model, LTN supports a wide range of applications and offers efficient Bayesian inference computational recipes through conjugate blocked Gibbs sampling with Pólya-Gamma augmentation. We demonstrate application of LTN in a compositional mixed-effects model for differential abundance analysis using numerical experiments and a reanalysis of the infant cohort in the DIABIMMUNE study. Our findings illustrate that LTN, by adequately accounting for cross-sample heterogeneity, appropriately generates the proportion of zeros without requiring an explicit zero-inflation component, confirming a recent viewpoint that "zero-inflation" in count-based sequencing data are often results of unaccounted cross-sample variation.

Introduction

Microbiome research relies on compositional count data—measurements of microbial abundance relative to a total sequencing depth. This data type presents three primary challenges: high dimensionality (thousands of taxa), complex covariance (microbial interactions), and extreme sparsity (prevalence of zeros).

Existing frameworks often force a compromise. The Dirichlet-Multinomial (DM) model is computationally efficient but lacks the flexibility to capture complex correlations. Conversely, Log-Ratio Normal (LN) models capture these correlations but scale poorly as the number of taxa increases. Furthermore, the common practice of adding "zero-inflation" components to handle sparsity can lead to biased results if the underlying biological variation is not correctly modeled.

The Logistic-Tree Normal (LTN) Model

The Logistic-Tree Normal (LTN) model addresses these challenges by utilizing a tree structure (e.g., phylogenetic or taxonomic) to decompose the multinomial distribution of taxa into a series of binary splits.

Tree-based decomposition and transformation from compositional counts $p$ to log-odds $\theta$.

By aggregating counts up the tree, the model gains statistical power even when individual taxa counts are sparse. At each node, the model employs a multivariate logistic-normal distribution to represent branching probabilities. This latent Gaussian structure allows the model to capture correlations across the entire tree. To ensure scalability, the authors utilize Pólya-Gamma (PG) augmentation, a mathematical technique that restores conjugacy and enables efficient Bayesian inference via Gibbs sampling.

Key Insights: Rethinking Zero-Inflation

A significant finding of the LTN framework is that "zero-inflation" may often be an artifact of unaccounted sample variation. By accurately modeling cross-sample heterogeneity, LTN generates the expected proportion of zeros naturally, without requiring a dedicated zero-inflation parameter. This aligns with emerging views that many zeros in sequencing data are biological or sampling-related rather than purely technical noise.

Additionally, the tree structure facilitates "energy concentration," similar to Haar wavelets in signal processing. Biological signals are concentrated at higher nodes, making it easier to detect significant changes that might be obscured at the individual species level. The model also demonstrates robustness to tree misspecification; its flexible covariance structure can compensate for sub-optimal tree topologies.

Experimental Results

The researchers evaluated LTN using both synthetic datasets and real-world data from the DIABIMMUNE study, which tracks the gut microbiome of infants to identify risk factors for Type 1 Diabetes (T1D).

  • Higher Power and FDR Control: In simulations, LTN significantly outperformed standard tools like MaAsLin2 and DirFactor in sensitivity and specificity (ROC curves). Unlike univariate methods that may suffer from inflated False Discovery Rates (FDR) in complex scenarios, LTN’s joint modeling approach provides more reliable discoveries.
  • Biological Discovery: In the DIABIMMUNE re-analysis, LTN identified specific shifts in Bacteroides and Parabacteroides associated with seroconversion (a precursor to T1D). It also demonstrated how dietary changes, such as introducing solid food, shift the Firmicutes/Bacteroidetes ratio at specific tree branches.

Conclusion and Future Directions

The LTN model represents a versatile generative framework for high-dimensional microbiome data. While highly effective, future work may address sensitivity to sparsity-inducing priors and explore the development of "functional trees" based on microbial metabolic roles to further enhance discovery power. The tree-based log-odds decomposition also holds potential for application in other high-dimensional probability distributions beyond microbiome research.

Summary updated

Abstract

Microbiome compositional data are often high-dimensional, sparse, and exhibit pervasive cross-sample heterogeneity. Generative modeling is a popular approach to analyze such data, and effective generative models must accurately characterize these key features. While high-dimensionality and abundance of zeros have received much attention, existing models often lack flexibility in capturing complex cross-sample variability. This limitation can affect statistical efficiency and lead to misleading conclusions in tasks like differential abundance analysis, clustering, and network analysis. We introduce a generative model, the "logistic-tree normal" (LTN) model, which addresses this issue and effectively captures key characteristics of microbiome data, including abundance of zeros. LTN employs a tree-based decomposition to aggregate sparse taxa counts and uses a (multivariate) logistic-normal distribution at tree splits, allowing for flexible covariance adjustments among taxa as needed. The latent Gaussian structure of LTN enables the incorporation of multivariate analysis tools that enforce sparsity or low-rank covariance assumptions. As a versatile, fully generative model, LTN supports a wide range of applications and offers efficient Bayesian inference computational recipes through conjugate blocked Gibbs sampling with Pólya-Gamma augmentation. We demonstrate application of LTN in a compositional mixed-effects model for differential abundance analysis using numerical experiments and a reanalysis of the infant cohort in the DIABIMMUNE study. Our findings illustrate that LTN, by adequately accounting for cross-sample heterogeneity, appropriately generates the proportion of zeros without requiring an explicit zero-inflation component, confirming a recent viewpoint that "zero-inflation" in count-based sequencing data are often results of unaccounted cross-sample variation.
BibTeX
@misc{wang2021tree,
  author = {Zhuoqun Wang and Jialiang Mao and Li Ma},
  title = {{A tree-based model for addressing sparsity and taxa covariance in microbiome compositional count data}},
  year = {2021},
  eprint = {2106.15051},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2106.15051},
}
Figure