Ensemble test for microbiome data.

Area of Science:

• Microbiome research
• Statistical bioinformatics
• Computational biology

Background:

• Human microbiome composition is linked to various diseases.
• Analyzing high-dimensional, sparse microbiome data presents statistical challenges.
• Permutational multivariate analysis of variance (PERMANOVA) is widely used but has limitations.

Purpose of the Study:

• To develop a robust and powerful statistical method for microbiome data analysis.
• To address the limitations of traditional PERMANOVA, including sensitivity to distance metrics.
• To improve the detection of associations between microbiome composition and biological features.

Main Methods:

• Introduced E-MANOVA (Ensemble multivariate analysis of variance using distance matrices), an ensemble learning approach.
• Constructed base tests by powering similarity matrices and combined them for a final test statistic.
• Utilized direct moment approximation and Pearson type III distribution to approximate null distributions, avoiding permutations.
• Employed Cauchy combination method to aggregate p-values across multiple distance metrics.

Main Results:

• The proposed E-MANOVA method demonstrated superior power and robustness compared to existing methods in simulations.
• E-MANOVA effectively identified a higher number of significant associations in real-world microbiome datasets.
• The method avoids computationally intensive permutation tests by using distribution approximations.

Conclusions:

• E-MANOVA significantly outperforms current methods for microbiome association studies.
• The ensemble approach and novel p-value aggregation enhance the detection of biological signals.
• This method provides a more reliable tool for exploring microbiome-disease relationships.