A latent allocation model for the analysis of microbial composition and disease

Area of Science:

• Microbiome research
• Statistical modeling
• Computational biology

Background:

• Microbiome data analysis is challenged by a high prevalence of zero counts.
• Existing methods like pseudo-counting or standard distributions can introduce bias.
• Accurate statistical approaches are crucial for linking microbiota to diseases.

Purpose of the Study:

• To develop a new probabilistic model, BERMUDA, for microbiome data.
• To address the challenges posed by excess zeros in microbiome datasets.
• To enable the description of bacteria composition differences associated with diseases.

Main Methods:

• Developed the Bernoulli and Multinomial Distribution-based latent Allocation (BERMUDA) model.
• Employed an annealing Expectation-Maximization (EM) algorithm for efficient model learning.
• Utilized simulation studies and real-world data for performance evaluation.

Main Results:

• The BERMUDA model effectively describes differences in bacteria composition.
• The model addresses the issue of excess zeros in microbiome data.
• The annealing EM algorithm provides a simple and efficient learning procedure.

Conclusions:

• BERMUDA demonstrates strong performance in both simulated and real data analyses.
• The proposed model offers a robust solution for microbiome case-control studies.
• BERMUDA is implemented in R and publicly available on GitHub.