科学研究

Biography: Yijuan Hu, a professor in the Department of Biostatistics at Peking University, and an awardee of a national talent program. She earned her Bachelor's degree in Statistics from the School of Mathematical Sciences at Peking University and her Ph.D. in Biostatistics from the University of North Carolina at Chapel Hill. She worked at Emory University from 2011 to 2024 and returned to Peking University in July 2024. Her research focuses on developing statistical theories and methods for high-dimensional omics data in the field of biostatistics, with a particular focus on high-dimensional hypothesis testing, robust inference, and issues related to missingness or experimental bias in microbiome and genetic data. She has published numerous articles as the first author or corresponding author in journals such as the Journal of the American Statistical Association (JASA), Proceedings of the National Academy of Sciences (PNAS), Microbiome, and American Journal of Human Genetics (AJHG). She has been invited multiple times to serve as a review panel member for NIH grants and is an associate editor for BMC Bioinformatics and Statistics in Biosciences.

Inferring microbial networks is complicated by the compositional nature of microbiome sequencing data, which are also sparse, high-dimensional, highly overdispersed, and occasionally derived from clustered samples. Most existing methods, such as SparCC, CCLasso, and COAT, provide only point estimates of Pearson's correlations. We demonstrate for the first time that the correlation estimates tend to be biased downward in the presence of overdispersion. In this article, we introduce a novel testing method called TestNet, which produces well calibrated results by controlling the false discovery rate (FDR). TestNet is based on Pearson's covariance and distance covariance of the centered-log-ratio data to capture linear and nonlinear dependencies, respectively. Because the correlations of interest are unidentifiable, we focus on testing the deviations of the covariances from the null hypotheses. We developed a permutation-based procedure for generating valid null replicates that account for compositional effects and extensive zeros in microbiome data. Our extensive simulation studies indicate that TestNet is the only method that effectively controls the FDR while achieving high power across a wide range of scenarios. Applying TestNet to two real microbiome datasets uncovered scientifically plausible networks.