Bayesian biclustering for microbial metagenomic sequencing data via multinomial matrix factorization.

Area of Science:

• Microbiome Research
• Statistical Bioinformatics
• Human Health

Background:

• High-throughput sequencing enables gut microbiome analysis but presents statistical challenges due to data characteristics (compositional, sparse, noisy, heterogeneous).
• Understanding the gut microbiome's role in diseases like inflammatory bowel disease (IBD) is crucial for developing targeted therapies.

Purpose of the Study:

• To propose an identifiable Bayesian multinomial matrix factorization model for inferring overlapping microbial and host clusters.
• To automatically determine the number of clusters and incorporate taxonomic information for improved interpretability.

Main Methods:

• Developed a Bayesian Dirichlet-multinomial mixture model for over-dispersed, zero-inflated count data.
• Hierarchically built latent cluster structures, integrating microbial taxonomic information.
• Validated the model through simulations and applied it to a human gut microbiome dataset from IBD patients.

Main Results:

• The model successfully identified distinct bacterial clusters in IBD patients.
• Identified clusters included bacteria families (e.g., Bacteroidaceae, Lachnospiraceae) previously associated with IBD and its subtypes.
• Demonstrated superior performance compared to alternative methods in simulation studies.

Conclusions:

• The proposed Bayesian model effectively analyzes complex gut microbiome data, revealing disease-associated microbial communities.
• Findings provide a foundation for generating hypotheses regarding gut microbiome heterogeneity in inflammatory bowel disease.
• The method enhances interpretability by leveraging microbial taxonomy and automatic cluster number determination.