Abstract
MOTIVATION: Meta-analysis methods widely-used for combining metabolomics data do not account for correlation between metabolites or missing values. Within- and between-study variability are also often overlooked. These can give results with inferior statistical properties, leading to misidentification of biomarkers. RESULTS: We propose a multivariate meta-analysis model for high-dimensional metabolomics data (MetaHD), which accommodates the correlation between metabolites, within- and between-study variances, and missing values. MetaHD can be used for integrating and collectively analysing individual-level metabolomics data generated from multiple studies as well as for combining summary estimates. We show that MetaHD leads to lower root mean square error compared to the existing approaches. Furthermore, we demonstrate that MetaHD, which exploits the borrowing strength between metabolites, could be particularly useful in the presence of missing data compared to univariate meta-analysis methods, which can return biased estimates in the presence of data missing at random. AVAILABILITY AND IMPLEMENTATION: The MetaHD R package can be downloaded through Comprehensive R Archive Network (CRAN) repository. A detailed vignette with example datasets and code to prepare data and analyses are available on https://bookdown.org/a2delivera/MetaHD/. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.