A negative binomial latent factor model for paired microbiome sequencing data.

Area of Science:

• Microbiology
• Statistical Modeling
• Bioinformatics

Background:

• Microbiome compositional data frequently originate from multiple body sites, exhibiting inherent dependencies.
• Joint analysis of microbial compositions across sites leverages cross-site correlations for enhanced statistical power, particularly beneficial with limited sample sizes.

Purpose of the Study:

• To introduce a novel joint statistical model for analyzing microbiome compositions from multiple body sites within the same subjects.
• To address the challenge of cross-site dependency in microbiome data and improve analytical outcomes.

Main Methods:

• Development of a joint model incorporating shared latent factors to capture common subject effects and inter-site correlations.
• Inclusion of mixtures of latent factors to account for heterogeneity in cross-site associations among samples.
• Application of the model to synthetic data and a case study involving urinary and vaginal microbiome samples.

Main Results:

• Simulation studies demonstrated that the proposed joint model with latent factors effectively mitigates bias caused by strong site associations, outperforming models without latent factors.
• The model identified statistically significant covariate associations in the female urogenital microbiome (e.g., UMICRO study) that were missed by separate site analyses.
• The model facilitates prediction of microbial abundance between sites and enables subject clustering based on association strengths.

Conclusions:

• The joint modeling approach provides a robust framework for analyzing multi-site microbiome data, improving statistical power and revealing complex associations.
• The model's ability to handle varying degrees of cross-site correlation and enable predictions enhances its utility in microbiome research.