MeLSI

MeLSI learns data-adaptive distance metrics to improve statistical power and biological interpretability in detecting group differences in multivariate microbiome beta diversity analyses.

Key Features:

• Data-adaptive metric learning: Learns optimal distance metrics that adapt to specific dataset characteristics rather than applying fixed ecological distance metrics uniformly to all microbial taxa.
• Ensemble of weak learners: Employs bootstrap sampling, feature subsampling, and gradient-based optimization to learn optimal feature weights.
• Feature importance weights: Provides interpretable feature-weight profiles that identify which microbial taxa drive group separation.
• Integration with standard workflows: Learned metrics can be used with PERMANOVA for hypothesis testing and with Principal Coordinates Analysis (PCoA) for ordination visualization.
• Statistical rigor: Maintains proper Type I error control through rigorous permutation testing for statistical inference.

Scientific Applications:

• Microbiome beta diversity analysis: Detects community composition differences between groups with improved statistical power compared to fixed metrics such as Bray-Curtis, Euclidean, Jaccard, and UniFrac.
• Biomarker discovery: Identifies specific bacterial families and microbial taxa that drive group differences, supporting the discovery of therapeutic targets and diagnostic biomarkers.
• Comparative microbiome studies: Applied to sex-based comparisons of gut microbiomes and validated on synthetic benchmarks and real datasets including the Atlas1006 dataset.

Methodology:

MeLSI employs an ensemble of weak learners using bootstrap sampling, feature subsampling, and gradient-based optimization to learn optimal feature weights, combined with rigorous permutation testing for statistical inference.