KERNEL-PENALIZED REGRESSION FOR ANALYSIS OF MICROBIOME DATA.

Area of Science:

• Microbiology
• Bioinformatics
• Statistical Modeling

Background:

• Human microbiome data analysis commonly uses dimension-reduced graphical displays and clusterings.
• Ordination methods often rely on biologically motivated similarity definitions and ecologically defined distances.
• Principal coordinate analysis is frequently employed for its ability to handle non-Euclidean structures.

Purpose of the Study:

• To present a high-dimensional regression framework extending traditional distance-based microbiome analysis methods.
• To incorporate extrinsic information, such as phylogeny, into penalized regression models.
• To address the compositional nature of microbiome data (relative abundances).

Main Methods:

• Development of a high-dimensional regression framework.
• Utilizing kernel-based methods to integrate extrinsic data like phylogeny.
• Application of penalized regression models for estimating taxon-specific associations.
• Addressing compositional data using relative abundance vectors.

Main Results:

• Demonstrated a framework that goes beyond dimension-reduced ordination.
• Showcased the incorporation of phylogeny and other extrinsic information.
• Successfully applied the regression framework to compositional microbiome data.
• Illustrated the approach with simulations and real-world microbiome datasets (gut and vaginal).

Conclusions:

• The proposed high-dimensional regression framework offers an extension to existing distance-based microbiome analysis methods.
• This approach enables the incorporation of diverse extrinsic information, such as phylogeny, into the analysis.
• The framework effectively handles the compositional nature of microbiome relative abundance data.
• The study provides a robust method for identifying microbial associations with phenotypes or clinical outcomes, including a significance test.