Removing unwanted variation from large-scale cancer RNA-sequencing data

The accurate identification and effective removal of unwanted variation are essential to derive meaningful biological results from RNA-seq data, especially when the data come from large and complex studies. We have used The Cancer Genome Atlas (TCGA) RNA-seq data to show that library size, batch effects, and tumor purity are major sources of unwanted variation across all TCGA RNA-seq datasets and that existing gold standard approaches to normalizations fail to remove this unwanted variation. Additionally, we illustrate how different sources of unwanted variation can compromise downstream analyses, including gene co-expression, association between gene expression and survival outcomes, and cancer subtype identifications. Here, we propose the use of a novel strategy, pseudo-replicates of pseudo-samples (PRPS), to deploy the Removing Unwanted Variation III (RUV-III) method to remove different sources of unwanted variation from large and complex gene expression studies. Our approach requires at least one roughly known biologically homogenous subclass of samples shared across sources of unwanted variation. To create PRPS, we first need to identify the sources of unwanted variation, which we will call batches in the data. Then the gene expression measurements of biologically homogeneous sets of samples are averaged within batches, and the results called pseudo-samples. Pseudo-samples with the same biology and different batches are then defined to be pseudo-replicates and used in RUV-III as replicates. The variation between pseudo-samples of a set pseudo-replicates is mainly unwanted variation. We illustrate the value of our approach by comparing it to the TCGA normalizations on several TCGA RNA-seq datasets. RUV-III with PRPS can be used for any large genomics project involving multiple labs, technicians, or platforms.