Hello, everyone,

I'm recently meeting this problem with my analysis, which i've done a lots of research and asked people around but their answers are quite confusing, so if I can get more opinions, that'd be terrific and thanks at advance.

So I'm doing an analysis of DEGs using Deseq2 and edgeR, the problem here is that I get those two variable, each has two conditions: grippe virus and no, Carcinoma and other. But we really want to figure out is, the impact of virus, but the histological type is quite important as well. So I'm wondering if there is any way that we can skip ou reduce the impact of histological type by designing our module.

It's not an interaction condition, so I built my design code like this:

Code should be placed in three backticks as shown below

DESeq.ds <- DESeqDataSetFromTximport(txi.rsem,sampleTable , ~ condition+Histological_type)

resultsNames(DESeq_new)
[1] "Intercept"                                 "condition_NEG_vs_VIRUS"                     
[3] "Histological_type_Other_vs_Carcinoma"

DGE.results_histo=results(DESeq_new,pAdjustMethod = "BH",contrast=c("Histological_type","Other","Carcinoma"))

DGE.results_virus=results(DESeq_new,pAdjustMethod = "BH",contrast=c("condition","VIRUS,"NEG"))

So I calculated two conditions, now I've to decide which one should I keep it, Some biologist suggest to find the commun part and get rid of it because what we interested in is virus part. Some of bioinformatist suggest that I need to just build one condition in the first place, like that:

DESeq.ds <- DESeqDataSetFromTximport(txi.rsem,sampleTable , ~ condition)

but in this case I have a lots of outliers cells.

compare to my design with two variables:

I also done the analysis with edgeR, but the part of construct the model makes me less sure, because I don't quite understand what I'm doing, if someone could clarify this part for me would be a great help.

> desigN <- model.matrix(~0+histology_type_edger+Groups_edgeR)
> desigN
    histology_type_edgerCarcinoma  histology_type_edgerOther  Groups_edgeRVIRUS
1                                  1                         0               1
2                                  1                         0               1
3                                  1                         0               1
4                                  1                         0               1
5                                  0                         1               1
6                                  0                         1               1
7                                  1                         0               0
8                                  1                         0               0
9                                  1                         0               0
10                                 1                         0               0
11                                 0                         1               0
12                                 1                         0               0
13                                 1                         0               0
14                                 1                         0               0
15                                 0                         1               0
16                                 1                         0               0

Thanks for your help!!!

As covered in the DESeq2 vignette, this is a clear case for the multi-factor design -- you want to control for one while assessing the impact of the other. I don't know any statisticians who would recommend not including histological type in the design, if it impacts expression.

Linear model construction is the same in edgeR and DESeq2 so I don't understand why you are using completely different variables for edgeR.

Our clear advice in the edgeR guide for most datasets like this is to make a single Group variable with four levels, corresponding to the four combinations of virus state and histology. Then any comparison you want to make becomes quite simple. In particular, you can test for the virus effect for each histological condition separately, which allows you to separate the virus effect from histology completely and transparently.

Trying to fit an additive model, as you have, is bound to give poor results if virus and carcinoma have interacting effects. In my opinion, your idea of subtracting common DEGs in some way is neither necessary nor useful.

Michael Love Gordon Smyth

Hello again Mike and Gordon, Thanks very much for your previous help for the construction of my analysis model, and sorry for bothering you again because my result is quite stranger compared to other bibliographies I read, please kindly help me again.

So I make a ggplot to see the correlation of two log fold-change result.

My objectif is to observe the impact of virus in different histological type of cancer in some patients, here I separated into Carcinoma and Other as histological type, the infection condition is Virus, which has this type of virus, Neg which haven't.

In this circumstance, I made Neg and Carcinoma as reference group

My code for Deseq2 is:

 DESeq.ds <- DESeqDataSetFromTximport(txi.rsem,sampleTable , ~ infection+Histological_type)
DESeq_new <- DESeq(DESeq.ds)
DGE.results=results(DESeq_new,pAdjustMethod = "BH",contrast=c("infection","VIRUS","Neg"))

My edgeR code is :

 y <- DGEList(normCts_final)
list_calculate <- scaleOffset(y, matrix_normM)

#construction of model
Group <- factor(paste(sampleTable$infection,sampleTable$Histological_type,sep="."))
desigN <- model.matrix(~0+Group)
colnames(desigN) <- levels(Group)

keep <- filterByExpr(list_calculate,desigN)
list_pret <- list_calculate[keep, ]

#estimate the dispersion of the data
estimeDisp <- estimateDisp(list_pret, design=desigN,robust=T)
fit <- glmQLFit(estimeDisp, contrast=desigN,robust=T)

contrasts.matrix<-makeContrasts(VIRUScarVSother=VIRUS.Other-VIRUS.carcinome,
                                NegcarVSother=Neg.Other-Neg.carcinome,levels=desigN)
qlf<-glmQLFTest(fit,contrast=contrasts.matrix[,"VIRUScarVSother"])

When I change only my code Deseq2 to :

DGE.results=results(DESeq_new,pAdjustMethod = "BH",contrast=c("Histological_type","Other","carcinome"))

I got this, which seems like what we want, but It would not be correct to explain, because I want to know the imapct of virus rather than histological type. Any thoughts about this?

I did not change my code edger because there is no result otherwise, I tried the estimateGLMRobustDisp, did not work as well.

I send this question for both of you, because I can't be sure for any of those my code, so maybe I did't fully understand your previous suggestion. Thanks a lot for your comments about my result and solutions maybe?

Xin