Microbiome subcommunity learning with logistic-tree normal latent Dirichlet allocation.

Area of Science:

• Microbiome research
• Computational biology
• Statistical modeling

Background:

• Mixed-membership (MM) models like latent Dirichlet allocation (LDA) are used for microbiome data to find microbial subcommunities.
• These subcommunities aid in understanding microbial interactions and predicting health outcomes.
• Existing LDA models fail to account for significant cross-sample variability in subcommunity compositions.

Purpose of the Study:

• To develop a novel MM model that accounts for cross-sample heterogeneity in microbiome subcommunity compositions.
• To improve the robustness and accuracy of microbial subcommunity identification in microbiome studies.

Main Methods:

• Incorporation of the logistic-tree normal (LTN) model into LDA to create a new MM model.
• The LTN-LDA model allows for variation in subcommunity compositions around a central 'centroid' composition.
• Utilized auxiliary Pólya-Gamma variables for efficient Bayesian inference via a collapsed blocked Gibbs sampler.

Main Results:

• The proposed LTN-LDA model successfully accounts for cross-sample heterogeneity in microbiome data.
• This accounting for heterogeneity restores robustness to the inference of the number of subcommunities.
• The model enables the identification of more biologically meaningful microbial subcommunities.

Conclusions:

• The novel LTN-LDA model offers a significant advancement in analyzing microbiome compositional data.
• By addressing cross-sample variability, the model provides more reliable identification of microbial subcommunities.
• This improved identification can lead to deeper insights into microbial ecology and host-microbe interactions.