DeLUCS: Deep learning for unsupervised clustering of DNA sequences

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Accurate DNA sequence clustering is crucial for genomic analysis.
• Existing methods often rely on sequence alignment or taxonomic identifiers, limiting their applicability.
• Unsupervised methods are needed for large, unannotated datasets.

Purpose of the Study:

• To introduce a novel Deep Learning method for Unsupervised Clustering of DNA Sequences (DeLUCS).
• To enable accurate DNA sequence clustering without reliance on sequence alignment, homology, or identifiers.
• To overcome limitations of traditional clustering methods for large and complex genomic datasets.

Main Methods:

• Utilized Frequency Chaos Game Representations (FCGR) of DNA sequences.
• Employed multiple neural networks to self-learn genomic signatures from "mimic" sequence FCGRs.
• Applied a majority voting scheme for final cluster assignment.

Main Results:

• Achieved high accuracy (77%-100%) in clustering diverse datasets (vertebrate, bacterial, viral).
• DeLUCS clusters matched true taxonomic groups across various taxonomic levels.
• Outperformed K-means++ and Gaussian Mixture Models by up to 47% on unlabelled data.

Conclusions:

• DeLUCS provides a fast and accurate solution for unsupervised DNA sequence clustering.
• The method effectively handles large datasets (over 1 billion bp) and bypasses common classification limitations.
• DeLUCS is a powerful tool for analyzing previously intractable genomic data.