A scaling normalization method for differential expression analysis of RNA-seq data

Abstract

The fine detail provided by sequencing-based transcriptome surveys suggests that RNA-seq is likely to become the platform of choice for interrogating steady state RNA. In order to discover biologically important changes in expression, we show that normalization continues to be an essential step in the analysis. We outline a simple and effective method for performing normalization and show dramatically improved results for inferring differential expression in simulated and publicly available data sets.

Figure 1

Normalization is required for RNA-seq data. Data from [6] comparing log ratios of (a) technical replicates and (b) liver versus kidney expression levels, after adjusting for the total number of reads in each sample. The green line shows the smoothed distribution of log-fold-changes of the housekeeping genes. (c) An M versus A plot comparing liver and kidney shows a clear offset from zero. Green points indicate 545 housekeeping genes, while the green line signifies the median log-ratio of the housekeeping genes. The red line shows the estimated TMM normalization factor. The smear of orange points highlights the genes that were observed in only one of the liver or kidney tissues. The black arrow highlights the set of prominent genes that are largely attributable for the overall bias in log-fold-changes.

Figure 2

Simulations show TMM normalization is robust and outperforms library size normalization. (a) An example of the simulation results showing the need for normalization due to genes expressed uniquely in one sample (orange dots) and asymmetric DE (blue dots). (b) A lower false positive rate is observed using TMM normalization compared with standard normalization.

Figure 3

False discovery plots comparing several published methods. The red line depicts the length-normalized moderated t-statistic analysis. The solid and dashed lines show the library size normalized and TMM normalized Poisson model analysis, respectively. The blue and black lines represent the LR test and exact test, respectively. It can be seen that the use of TMM normalization results in a much lower false discovery rate.