Identification of microbial features in multivariate regression under false discovery rate control

In many microbiome studies, researchers often aim at detecting statistical associations between microbial taxa and multiple disease-related secondary phenotypes of interest, which are further investigated in downstream functional studies. Most existing approaches tackle this aim by analyzing one taxon at a time and then followed by multiple testing correction. However, the large number of microbial taxa poses a heavy multiple correction burden which often limits the power of discovery of the aforementioned individual taxon-based analyses. Moreover, complicated correlation structures among taxa poses grand challenges for multiple testing correction procedures to achieve a satisfactory performance (e.g., false discovery rate control). To address these potential limitations, a new approach is proposed to detect statistical associations between multiple responses and microbial features in a multivariate regression model, which models the correlations among responses to boost power of association discovery. By utilizing the knockoff filter technique, the proposed procedure also enjoys the property of finite-sample false discovery rate control. It is demonstrated through a comprehensive simulation study to show the validity and usefulness of our new method and apply the methodology to a data set collected from microbiome studies to gain additional biological insights.