Unsupervised two-way clustering of metagenomic sequences

Area of Science:

• Computational Biology
• Metagenomics
• Bioinformatics

Background:

• Metagenomics faces challenges in characterizing functional and taxonomic content from short reads.
• Clustering method efficacy is influenced by dataset size, read length, and genome abundance.

Purpose of the Study:

• To develop an unsupervised naive Bayes multispecies, multidimensional mixture model for metagenomic reads.
• To cluster reads by species of origin and characterize species abundance.

Main Methods:

• Formulated an unsupervised naive Bayes multispecies, multidimensional mixture model.
• Modeled word count distributions using Gaussian (short words) and Poisson (long words) distributions.
• Employed Gaussian or Poisson mixtures for reads within bins, handling high-dimensionality via word grouping.

Main Results:

• The proposed method accurately clusters metagenomic reads, even those as short as 100 bps.
• Demonstrated accuracy and applicability on both simulated and real metagenomic datasets.
• The method is robust to variations in abundance, divergence, and read length.

Conclusions:

• The developed naive Bayes mixture model effectively addresses challenges in metagenomic read characterization.
• This tool enhances the ability to determine taxonomic and abundance profiles from complex microbial communities.