Prediction with Microbiome Sequencing Data via Multi-Kernel Learning

• Activity Number: 310 - SPEED:Statistical Methods for GWAs, Genetics, Genomics, and Other Omics Studies, Part 2
• Type: Contributed
• Date/Time: Tuesday, July 30, 2019 : 9:25 AM to 10:10 AM
• Sponsor: Biometrics Section
• Author(s): Bing Li* and Huilin Li and Shuang Wang
• Companies: Brown University and NYU School of Medicine and Columbia University
• Keywords: Microbiome; Covariates; Prediction; Distance-based methods; Generalized additive models; Multi-kernel learning

Abstract

Studies have established associations between microbiome profilings and different health outcomes. Many statistical models developed for associations utilize the special characteristics of microbiome taxonomic data including the phylogenetic tree structure, the large number of rare taxa, and the complicated associations between taxa and health outcomes. However, existing studies that used microbiome data to predict health outcomes applied methods such as random forest which ignore the phylogenetic information. To fill in this gap, here we developed a multi-kernel generalized additive model (mK-GAM) that takes the unique characteristics of microbiome data into account. This is achieved through extracting various aspects of information into multiple kernels and then learning a conic combination of these kernels to best predict the outcome. In addition, the mK-GAM model allows covariates to improve the prediction performance. We demonstrated the superior performance of our new method over several existing ones through simulation studies and applications to two real microbiome datasets. The proposed multi-kernel method mK-GAM consistently predicts outcomes of interest most accurately.