Differential gene expression analysis vs disease-gene expression association

Question

[Sara]

Hi All! Can anybody help me to understand what is the difference between differential gene expression analysis vs disease-gene expression association analysis? Any help is appreciated!

Answer 1

[Kevin Blighe]

Hello Sara, please, can you tell me where you have seen these terms? Thank you!

[Sara]

Hi Kevin, for example in this paper they used gene-expression association analysis using a regression model. It was confusing for me that they named the associated genes as the differentially expressed genes while differential gene expression analysis has its own methods and packages like Deseq2 and EdgR.

[Kevin Blighe]

Hey, it says that they performed the following:

We tested for (1) differential gene expression between individuals with T2D and normal glucose tolerance (NGT) and (2) association between gene expression and three T2D-related traits (fasting glucose, fasting insulin and body mass index (BMI)) (Fig. 1b), adjusting for covariates (see Methods)

Point (1) is the same as any 'common' differential expression analysis. Point (2) is going to be a regression, like this, for example:

lm(gene ~ BMI + cov1 + cov2 + ... + covn)

[Sara]

Thanks Kevin, but if you check the methods they used the same approach for both analysis. Indeed, they did not use any common approach of differential expression analysis like Deseq2 and they did association analysis. So, my main question is, did they misuse the "differential gene expression" phrase?

[Kevin Blighe]

I see what you mean. Their approach is unorthodox but they are still using some well-reputed programs in their analyses, including DEXseq and PEER, and appear to generally be paying close attention to details. At the end of the day, the more common differential expression analysis tools, like limma, EdgeR, and DESeq2, are also based on independently fitted [per gene] regression models, and then deriving p-values from these models. In the case of EdgeR and DESeq2, the model is a negative binomial regression, while limma has hung onto a linear regression from microarray days right through to RNA-seq (via limma/voom).

If I were to criticise the authors, it would be that they are basing everything on FPKMs - FPKMs are not suitable for any analysis where cross-sample differences are being gauged due to the fact that there is no library size normalisation used when deriving FPKMs. Also, their abundance method is THIS, which is, at this stage, old, and is based on a Poisson model of RNA-seq data; however, both EdgeR and DESeq2 showed that it is better to model RNA-seq data as a negative binomial, not a Poisson.

[Sara]

So many thanks kevin!