Interpretable metric learning in comparative metagenomics: The adaptive Haar-like distance

Area of Science:

• Microbiology
• Bioinformatics
• Computational Biology

Background:

• Random forests are effective for predicting environmental characteristics from microbial composition in comparative metagenomics.
• Existing methods often lack biological insight, hindering scientific advancement.
• Phylogenetic β-diversity metrics are crucial for understanding microbial community structure.

Purpose of the Study:

• To introduce a novel, data-driven phylogenetic β-diversity metric for enhanced biological interpretability in metagenomics.
• To develop a weighted nearest-neighbors classifier as a proxy for random forests.
• To improve the visualization of high-dimensional metagenomic data.

Main Methods:

• Leveraging a geometric characterization of random forests.
• Introducing the adaptive Haar-like distance, a new phylogenetic β-diversity metric.
• Developing a weighted nearest-neighbors classifier using the adaptive metric.

Main Results:

• The adaptive Haar-like distance assigns weights to phylogenetic nodes based on their importance in discerning environmental samples.
• The weighted nearest-neighbors classifier achieves accuracy comparable to random forests and CoDaCoRe.
• The new metric and classifier significantly enhance biological interpretability and visualization of metagenomic samples.

Conclusions:

• The adaptive Haar-like distance provides greater biological insight into metagenomic data than traditional methods.
• This approach offers a powerful tool for comparative metagenomics, bridging prediction accuracy with biological understanding.
• The developed metric and classifier facilitate more meaningful exploration of microbial community structures and their environmental associations.