Multivariable association discovery in population-scale meta-omics studies.

Area of Science:

• Microbiome research
• Statistical genetics
• Computational biology

Background:

• Microbiome multi-omics data present unique challenges including noise, sparsity, high dimensionality, and non-normality.
• Associating microbial features with metadata like human health outcomes is difficult due to data properties.

Purpose of the Study:

• To introduce MaAsLin 2 (Microbiome Multivariable Associations with Linear Models), a robust methodology for multivariable association analysis of microbial communities.
• To accommodate diverse epidemiological study designs (cross-sectional, longitudinal) and data types (counts, relative abundances) with complex metadata.

Main Methods:

• Utilizes generalized linear and mixed models to handle various data types and study designs.
• Incorporates novel and established statistical techniques for robust association analysis.
• Evaluated through extensive simulations and application to the Integrative Human Microbiome Project (HMP1) dataset.

Main Results:

• MaAsLin 2 preserves statistical power with repeated measures and multiple covariates while controlling false discovery rates.
• The method successfully accounts for the specific characteristics of microbiome multi-omics data.
• Application to HMP2 data reproduced known findings and revealed novel associations with inflammatory bowel diseases (IBD).

Conclusions:

• MaAsLin 2 provides a powerful and flexible tool for analyzing associations between microbial communities and complex metadata in large-scale studies.
• The methodology enhances our ability to understand the interplay between the microbiome and host health.
• This approach facilitates integrated analysis of multi-omics data across different time points and disease states.