There are several ways, or aspects, in which an RNAseq data set can (or should) be normalized. One is the between-sample normalization, that accounts for the varying library sizes and potential differences in RNA composition. This is done by the TMM method (or similar). Another normalization type is the between-gene normalization, in order to be able to compare or combine expression measurements from different genes (this can not be done with the raw counts, among other things since long genes typically have more reads than short genes). This type of normalization is not always performed if the only goal is gene-wise differential expression (since then, you don't actually compare different genes). However, since Qlucore also does multivariate analysis like PCA, we settled for applying also the gene length normalization. It does not affect the linear modeling since it is just an additive constant for each gene (constant across samples). So to summarize, the transformation performed by voom is (essentially) applying TMM normalization and log-transforming the normalized counts, while in Qlucore, we also normalize with respect to gene length.
All RNAseq differential expression methods have their pros and cons. One big advantage with the "transformation followed by linear modeling" approach is that many of the typical "microarray"/"normal distribution-assuming" methods (some of which are integral to Qlucore's functionality) are now applicable to the data, while they are not directly applicable to the count-level data. Also, transformation-based methods are good at controlling the false discovery rate (however, at a certain loss of power for small sample sizes).
See also the article by C Soneson and M Delorenzi: A comparison of methods for differential expression analysis of RNA-seq data. BMC Bioinformatics 2013, 14:91.
Note: From version 3.3 are more normalization options available in Qlucore Omics Explorer |
