Performance determinants of unsupervised clustering methods for microbiome data.

Area of Science:

• Microbiome research
• Computational biology
• Data analysis

Background:

• Unsupervised clustering is vital for identifying microbial community structures in microbiome data.
• Existing beta diversity and clustering methods are commonly applied but have limitations.
• Systematic comparison of these methods is needed for robust microbiome analysis.

Purpose of the Study:

• To systematically compare commonly used beta diversity and clustering methods in microbiome analysis.
• To identify scenarios where specific methods underperform.
• To develop an improved clustering metric for microbiome data.

Main Methods:

• Applied multiple beta diversity metrics and clustering methods to five diverse microbiome datasets.
• Systematically modified dataset properties to investigate metric performance.
• Developed and tested a novel combined metric integrating Bray Curtis and unweighted UniFrac.

Main Results:

• No single method consistently outperformed others across all datasets.
• Bray Curtis metric showed poor clustering with rare high-abundance OTUs; unweighted UniFrac struggled with abundant low-abundance OTUs.
• The novel combined metric demonstrated high performance across all tested datasets.

Conclusions:

• Existing clustering methods do not universally perform best for microbiome data.
• A novel combined metric, integrating Bray Curtis and unweighted UniFrac, offers improved and robust clustering performance.
• This metric capitalizes on the complementary strengths of individual metrics for enhanced microbiome analysis.