A0670
Title: Supervised dimensionality reduction for visualization and prediction of microbiome data
Authors: Rebecca Deek - University of Pittsburgh (United States) [presenting]
Abstract: Advances in high-throughput sequencing technology have enabled researchers to directly measure microbial compositions. The resulting data are high-dimensional, with many microbes being observed in a single sample. Accordingly, dimensionality reduction algorithms, such as principal coordinates analysis, are commonly applied to microbiome data for visualization and feature extraction. Such algorithms are unsupervised and therefore can result in visualizations that fail to differentiate between distinct outcome groups and limit their utility beyond data visualization or exploration. As such, a supervised and covariate-adjusted principal coordinates analysis algorithm is proposed that incorporates similarities in the outcome, as well as the microbial compositions, into the reduced dimension data while simultaneously removing unwanted nuisance covariate effects. The method provides enhanced visualization and can be used for downstream modeling. The performance of the proposed method is illustrated using simulations and real microbiome data sets.
