Hi edgeR Users,

I have employed edgeR to analyze RNAseq data before. Those data had two groups, one is control and one is treatment. However, now I need to investigate effect of chemical exposure on gene expression, so chemical exposure is a continuous variable, for example, from 30 ng/ml to 1000 ng/ml. I have several questions and hope to get help from you!

1.  for the code, y <- DGEList(counts=x,group=group), I can just use y <- DGEList(counts=x) since no group in my data set, right?

2. there are 320 samples sequenced in four different time since I submitted those samples in four different month. after filtering lowly expressed genes, doing TMM normalization, I will use PCA to check batch effect related to the four different sample submission ; if obvious batch effect observed, I will use combat in sva package to adjust batch effect.  according to my experience, there are some negative values for counted values after batch removal, do I need to add the absolute value of the minimal negative values to the whole data set before running edgeR?

3.  For design matrix, I plan to use:  design <- model.matrix(~ chemical_Concentration + age + race) , both chemical concentration and age are continuous variable, race is a categorical variable.  is this design code correct?

4. then I plan to use robust =TRUE in disp <- estimateDisp(y, design, robust = TRUE) and fit <- glmFit(disp, design, robust = TRUE) as I am analyzing miRNA data, some miRNA would be highly expressed as outliers, so robust =TRUE could minimize the predominant influence from those outlying expression, is my understanding correct?

5. lastly, I would use lrt <- glmLRT(fit, coef = 2) and topTags(lrt, n=Inf, adjust.method = "BH", p.value = 0.05).

in my design, I am interested in the coefficient for chemical concentration (log2 transformed), so coef=2 is correct, right?

6. also I am not very sure about the workflow, so first filtering lowly expressed miRNA, then doing TMM, then adjust batch effect;  or I need to do batch effect before doing filtering and TMM?

Thank you very much in advance!!

Ding

1. Yes.
2. You should be supplying the raw counts to edgeR, and these should always be positive. If there are unknown factors of variation, sva (or RUVseq, or friends) will report these for inclusion into the design matrix. You should not be using any form of corrected values as input into edgeR functions.
3. That seems sensible, assuming no interaction between concentration and the other two variables.
4. Yes, but use glmQLFit with robust=TRUE instead of glmFit. This uses the QL framework, which provides more accurate type I error control. See the workflow here for an example.
5. Use glmQLFTest instead of glmLRT. Your coefficient choice is correct.
6. Filter, TMM normalize, give normalized expression values (from cpm) to sva or RUVseq or whatever, and if it's necessary, cbind the resulting factors to your design matrix for fitting with edgeR.

Question 2: It is not correct or necessary to use the sva package here. Not correct because sva is not designed for use with counts. Not necessary because you know what the batches are already. If you find that there is a batch effect, just include batch as a factor in the design matrix, then edgeR will make the correction as part of the linear model fit.

You might re-visit what you have done for similar analyses in the past, because your remark about getting negative counts is worrying. You should not be applying any batch correction to the counts and you should never have negative counts.