Effects of data transformation and model selection on feature importance in microbiome classification data

Area of Science:

• Microbiome research
• Computational biology
• Machine learning applications

Background:

• Accurate host phenotype classification from microbiome data is vital for developing microbiome-based therapies.
• Gut microbiome data presents challenges like sparsity, compositionality, and population-specificity.
• The impact of microbiome data transformations on machine learning tasks remains largely unexplored.

Purpose of the Study:

• To investigate the effect of microbiome data transformations on machine learning classification accuracy and feature selection.
• To assess the robustness of machine learning models across different data transformations.
• To evaluate the implications for machine learning-based biomarker identification in microbiome research.

Main Methods:

• Analysis of over 8500 samples from 24 shotgun metagenomic datasets.
• Comparison of classification performance using various microbiome data transformations (e.g., presence-absence vs. abundance-based).
• Evaluation of feature importance variation across different transformations.

Main Results:

• Host health classification (healthy vs. diseased) is achievable with minimal dependence on specific algorithms or data transformations.
• Presence-absence transformations perform comparably to abundance-based transformations.
• Accurate classification requires only a small subset of predictors, but feature importance varies significantly across transformations.

Conclusions:

• Microbiome data transformations impact feature selection more than classification accuracy.
• Classification performance is robust across different transformations, but feature selection variability requires caution in biomarker identification.
• This study offers insights into applying machine learning to microbiome data and highlights future research directions.