Conducting gene set tests in meta‐analyses of transcriptome expression data

Abstract

Research synthesis, eg, by meta‐analysis, is more and more considered in the area of high‐dimensional data from molecular research such as gene and protein expression data, especially because most studies and experiments are performed with very small sample sizes. In contrast to most clinical and epidemiological trials, raw data are often available for high‐dimensional expression data. Therefore, direct data merging followed by a joint analysis of selected studies can be an alternative to meta‐analysis by P value or effect‐size merging or, more generally spoken, the merging of results. While several methods for meta‐analysis of differential expression studies have been proposed, meta‐analysis of gene set tests has very rarely been considered, although gene set tests are standard in the analysis of individual gene expression studies. We compare in this work the different strategies of research synthesis of gene set tests, in particularly the “early merging” of data cleaned from batch effects versus the “late merging” of individual results. In simulation studies and in examples of manipulated real‐world data, we found that in most scenarios, the early merging has a higher sensitivity of detecting a gene set enrichment than the late merging. However, in scenarios with few studies, large batch effect, moderate and large sample sizes of late merging are more sensitive than early merging.

Batch normalization followed by merging is powerful for phenotype prediction integrating multiple heterogeneous studies

“…With the increasing availability of large collections of omic data, the reproducibility of machine learning prediction models has raised great concerns when conducting cross-study predictions with the impact of study heterogeneity. Previous studies have addressed this issue and developed many statistical methods to overcome study heterogeneity, including merging with batch effect removal [ 36 ] and ensemble learning methods [ 9 ]. In this study, we performed a comprehensive analysis of different methods on the phenotype prediction by integrating heterogeneous omic studies.…”

Large-scale comparative transcriptomic analysis of temperature-responsive genes in Arabidopsis thaliana

“…Similar approach has also been taken by Shen and coworkers (Shen et al 2017 ), who combined microarray abiotic stress datasets and successfully identified common and specific gene modules to different abiotic stresses. Such comparative transcriptomic analyses have several advantages over conventional meta-analysis studies, which normally analyzed the DEGs from different studies separately and compared the gene lists at the end, especially in terms of higher sensitivity of detecting differentially expressed gene sets enrichment (Kosch and Jung 2019 ).…”

Batch Normalization Followed by Merging Is Powerful for Phenotype Prediction Integrating Multiple Heterogeneous Studies

“…With the increasing availability of large collections of omic data, the reproducibility of machine learning prediction models has raised great concerns when conducting cross-study predictions with the impact of study heterogeneity. Previous studies have addressed this issue and developed many statistical methods to overcome study heterogeneity, including merging with batch effect removal [37] and ensemble learning methods [9]. In this study, we performed a comprehensive analysis of different methods on the phenotype prediction by integrating heterogeneous omic studies.…”