MeLSI: Metric Learning for Statistical Inference in microbiome community composition analysis

Area of Science:

• Microbiome research
• Computational biology
• Statistical inference

Background:

• Current microbiome beta diversity analysis uses fixed distance metrics (e.g., Bray-Curtis, Jaccard) that treat all taxa equally.
• This "one-size-fits-all" approach may overlook subtle, biologically significant patterns in complex microbial communities.
• Identifying key microbial taxa driving community differences is crucial for biomarker discovery and understanding disease states.

Purpose of the Study:

• To introduce Metric Learning for Statistical Inference (MeLSI), a novel machine learning framework for data-adaptive distance metric learning in microbiome analysis.
• To develop a statistically rigorous method that optimizes detection of community composition differences.
• To provide interpretable feature-weight profiles for identifying key taxa driving group separation.

Main Methods:

• MeLSI employs an ensemble of weak learners with bootstrap sampling and feature subsampling.
• Gradient-based optimization is used to learn optimal feature weights for distance metrics.
• Learned metrics are integrated with permutational multivariate analysis of variance (PERMANOVA) for hypothesis testing and principal coordinates analysis (PCoA) for visualization.

Main Results:

• MeLSI demonstrated proper type I error control and superior statistical power in detecting subtle community shifts across synthetic and real datasets.
• On the DietSwap dataset, MeLSI uniquely identified significant diet-induced community shifts missed by fixed metrics.
• Learned feature weights successfully identified biologically relevant taxa, offering interpretable insights into group separation drivers.

Conclusions:

• MeLSI offers a statistically rigorous and data-driven approach to augment beta diversity analysis with enhanced interpretability.
• The framework effectively detects subtle community shifts and identifies key microbial drivers, surpassing the limitations of fixed distance metrics.
• MeLSI accelerates the translation of microbiome data into testable biological hypotheses and potential clinical applications by providing actionable insights.