Group Variable Selection with Compositional Covariates

• Activity Number: 237 - Feature Selection and Statistical Learning in Genomics
• Type: Contributed
• Date/Time: Monday, July 31, 2017 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics in Genomics and Genetics
• Author(s): Anna Plantinga* and Michael C. Wu
• Companies: University of Washington and Fred Hutchinson Cancer Research Center
• Keywords: Compositional ; Microbiome ; Group lasso ; ADMM

Abstract

Feature selection methods for microbiome compositional data have recently been proposed as an alternative to taxon level analyses or distance-based methods comparing entire microbial communities. Such models can effectively handle the high dimensionality of the covariates while enforcing the unit sum constraint of compositional data. However, existing compositional feature selection models do not take full advantage of the multi-level structure of microbiome data. For high dimensional regression models with multi-level compositional covariates, we propose an L1/L2 regularized linear log-contrast model that provides group- and taxon-level sparsity. We express the model as a constrained convex optimization problem and propose an alternating direction method of multipliers algorithm, and we demonstrate selection consistency and bounded loss. The selection and estimation accuracy of our method is evaluated using simulation studies; we also demonstrate its efficacy by applying it to a study relating host gene expression to gut microbiome composition.