SuRF: A new method for sparse variable selection, with application in microbiome data analysis.

Area of Science:

• Bioinformatics
• Computational Biology
• Statistical Genetics

Background:

• Variable selection is crucial for analyzing complex biological data, especially in microbiome studies.
• Existing methods often lack sparsity and struggle with the inherent correlation structure of microbiome data.
• Identifying key microbial biomarkers requires robust and accurate variable selection techniques.

Purpose of the Study:

• To introduce a novel variable selection method, subsampling ranking forward selection (SuRF), for regression and classification in microbiome analysis.
• To address limitations of existing methods regarding model sparsity and inference.
• To develop a method capable of identifying biomarkers at the appropriate taxonomic level without a priori assumptions.

Main Methods:

• SuRF combines LASSO penalized regression, subsampling, and forward-selection strategies.
• The method is implemented in an R package for generalized linear models.
• A novel agglomeration approach is used to handle the tree-like correlation structure of microbiome variables.

Main Results:

• Simulations demonstrate SuRF's superior performance in recovering true variables compared to popular existing approaches.
• Application to microbiome datasets for pouchitis prediction and individual identification showed comparable or better prediction accuracy.
• SuRF effectively controls the false positive rate in variable selection.

Conclusions:

• SuRF offers significant advantages in model sparsity and inference for microbiome data analysis.
• The method successfully identifies key biomarkers at data-driven taxonomic levels.
• SuRF provides a powerful tool for robust variable selection in complex biological datasets.