MeLSI: Metric Learning for Statistical Inference in Microbiome Community Composition Analysis.

Area of Science:

• Microbiome research
• Machine learning
• Statistical inference

Background:

• Current microbiome beta diversity analysis uses fixed distance metrics, treating all taxa equally.
• This uniform approach may obscure subtle, biologically significant patterns driven by specific microbes.
• Identifying key microbial drivers is crucial for discovering therapeutic targets and diagnostic biomarkers.

Purpose of the Study:

• To introduce MeLSI (Metric Learning for Statistical Inference), a novel machine learning framework for microbiome analysis.
• To develop data-adaptive distance metrics that optimize detection of community composition differences.
• To provide statistically rigorous and interpretable insights into microbiome data.

Main Methods:

• MeLSI employs an ensemble of weak learners with bootstrap and feature subsampling.
• Gradient-based optimization is used to learn optimal feature weights for distance metrics.
• Rigorous permutation testing ensures statistical inference, with learned metrics usable in PERMANOVA and PCoA.

Main Results:

• MeLSI maintains Type I error control and achieves competitive or superior F-statistics compared to traditional methods.
• The framework provides interpretable feature-weight profiles, clarifying taxa driving group separation.
• On the Atlas1006 dataset, MeLSI demonstrated stronger effect sizes and offered biological insights beyond fixed metrics.

Conclusions:

• MeLSI offers a statistically sound alternative to fixed metrics in microbiome beta diversity analysis.
• The data-driven, interpretable nature of MeLSI enhances biological understanding and hypothesis generation.
• MeLSI accelerates the translation of microbiome data into clinical applications and testable hypotheses.