Unsupervised statistical clustering of environmental shotgun sequences

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Metagenomic data analysis faces challenges in environmental shotgun sequence binning.
• Existing methods often rely on supervised learning with extrinsic data.
• A first-principles statistical model for unsupervised binning was lacking.

Purpose of the Study:

• To develop an unsupervised, first-principles statistical model for metagenomic sequence binning.
• To create a self-training fitting method for clustering sequences by taxonomic origin.
• To implement and evaluate a novel binning algorithm.

Main Methods:

• Derived an unsupervised maximum-likelihood formalism for sequence clustering based on k-mer distributions.
• Implemented the formalism using Markov Chain Monte Carlo in k-mer feature space.
• Introduced dimensionality reduction and a genomic fragment divergence measure.

Main Results:

• The method, LikelyBin, accurately bins short sequences (≥400 nt) with >90% accuracy in low-complexity simulations.
• Genomic fragment divergence strongly correlates with binning performance.
• Over 1000 genomes analyzed confirmed suitability for binning with this formalism.

Conclusions:

• Unsupervised binning using statistical signatures is viable for low-complexity metagenomic samples.
• The method can be integrated into iterative approaches for higher complexity samples.
• LikelyBin offers an open-source solution for environmental sequence binning.