Hi

I am new user of DESeq2 and have few questions regarding the design formula, results, and their inferences:

I am having four cell types (A,B,C,D) from old and young subject detailed as below:

table(pdata$Age)

  Old Young

   47    35

table(pdata$Age, pdata$CellType)

        A    B   C   D

  Old   11  12  12  12

  Young  8   9   9   9

I used the following designs with the respective objectives in mind

1)  To know the age specific effect over celltype

dds2 <- DESeqDataSetFromMatrix(countData = count2,colData = traitdata2,design = ~Age + CellType + Age:CellType)

2) To know the differential expression for age and various cell type without interaction

dds2 <- DESeqDataSetFromMatrix(countData = count2,colData = traitdata2,design = ~Age + CellType)

3) To know how individual cell type behave in old and young subjects

dds2 <- DESeqDataSetFromMatrix(countData = count2,colData = traitdata2,design = ~Age + CellType)

dds2$group <- factor(paste0(dds2$Age, dds2$CellType))

design(dds2) <- ~group

4) To get one pvalue for the differential expression across all the cell types (similar to ANOVA, without interaction with age)

dds2 <- DESeqDataSetFromMatrix(countData = count2,colData = traitdata2,design = ~CellType )

dds2LTR_ACTI <- DESeq(dds2,test="LRT", reduced=~1)

5) To get one pvalue for the differential expression across all the Age (similar to ANOVA, without interaction with cell type)

dds2 <- DESeqDataSetFromMatrix(countData = count2,colData = traitdata2,design = ~Age )

dds2LTR_ACTI <- DESeq(dds2,test="LRT", reduced=~1)

6) To get one pvalue for the differential expression across all the Age and CellType(similar to ANOVA, with Interaction)

dds2 <- DESeqDataSetFromMatrix(countData = count2,colData = traitdata2,design = ~Age + CellType + Age:CellType)

dds2LTR_ACTI <- DESeq(dds2,test="LRT", reduced=~Age + CellType)

Kindly answer these questions

1. Am I using the correct design?
2. With Contrast “results(dds2,contrast=c("Age","Old","Young"))” I am getting different results. with design 1 (up: 348 & down: 139) and design 2 (up:5 & down9). That is expected, (I guess??) but with design 3 (where in I am trying to know how the individual cells are behaving in young and old conditions), I am not getting any significant difference. Also with design 5 I am not getting any significant difference with respect to age. Any explanation?
3. After using design 4 I am getting around 3 times more differentially expressed genes as compared to design 1 and 2. For example between cell A and B I am getting (up:516 & down:360) with design 1 with design 2 I am getting (up: 689 & down : 844) but when I use design 4 I am getting (up: 1617 and down:2025). Any explanation?

I understand I might not be able to explain my problem completely. Please let me know if you need any other information.

Thanks

Ram

hi,

I would recommend coming up with a single design which you think is appropriate for your samples and sticking to that. You can perform numerous contrasts and tests using one design.

The risk when you try out many different designs is that you can (and will) lose Type I error control if you choose the design based on number of significantly DEG.

Your questions:

"1) Am I using the correct design?"

It seems like you should go with the design with the interaction term, as you are interested in finding genes which have age-specific effects.

"2) With Contrast “results(dds2,contrast=c("Age","Old","Young"))” I am getting different results. with design 1 (up: 348 & down: 139) and design 2 (up:5 & down9). That is expected, (I guess??) but with design 3 (where in I am trying to know how the individual cells are behaving in young and old conditions), I am not getting any significant difference. Also with design 5 I am not getting any significant difference with respect to age. Any explanation?"

The same contrast does not represent the same thing across the different designs. Here, I'd suggest that you speak to a local statistician at your institute about this, because this is quite difficult to explain on a support site. Also, you can take a look at the interactions section of the DESeq2 vignette which has a diagram. 

Again the best approach is to stick to a design (here I recommend the one with interaction term), and then pull out the comparisons of interest from this design. 

"3) After using design 4 I am getting around 3 times more differentially expressed genes as compared to design 1 and 2. For example between cell A and B I am getting (up:516 & down:360) with design 1 with design 2 I am getting (up: 689 & down : 844) but when I use design 4 I am getting (up: 1617 and down:2025). Any explanation?"

Same answer as with 2. I highly recommend meeting with a local statistician, as there are numerous null hypotheses to test here, and you need to ensure that your results tables correspond to the ones you wish to test. You can also look at the examples in ?results to see how to produce various contrasts from one design.