Question: DESeq2 Design with Three Interactions
2
3.6 years ago by
dustar1986300
USA
dustar1986300 wrote:

Hi DESeq2 Community,

I'm currently working on a comparison of drug effects on tumour and wild-type cells. I have tumour cells from 4 patients (2 males and 2 females) and normal cells from 4 normal people (2 males and 2 females). Each cell sample was treated under three different conditions: plain control, drug A and drug B. The data looks like:

``````Sample  Phenotype   Gender  Treatment
Patient1    Tumour  M   0
Patient1    Tumour  M   A
Patient1    Tumour  M   B
Patient2    Tumour  M   0
Patient2    Tumour  M   A
Patient2    Tumour  M   B
Patient3    Tumour  F   0
Patient3    Tumour  F   A
Patient3    Tumour  F   B
Patient4    Tumour  F   0
Patient4    Tumour  F   A
Patient4    Tumour  F   B
Control1    Normal  M   0
Control1    Normal  M   A
Control1    Normal  M   B
Control2    Normal  M   0
Control2    Normal  M   A
Control2    Normal  M   B
Control3    Normal  F   0
Control3    Normal  F   A
Control3    Normal  F   B
Control4    Normal  F   0
Control4    Normal  F   A
Control4    Normal  F   B
``````

In the PCA plot, the data was divided into 4 groups: male control, female control, male patient and female patient were located at top left, bottom left, top right and bottom right of the plot, respectively. Within each of these groups, individuals were separated vertically. For each individual, treatments effect were separated horizontally.

According to the PCA plot, I have the a design to include all the interactions and set Tumour, Male, Treatment0 as base level:

``````design(dds) <- ~Phenotype*Gender*Treatment
dds\$Phenotype <- relevel(dds\$Phenotype, "Tumour")
dds\$Gender <- relevel(dds\$Gender, "M")
dds\$Treatment <- relevel(dds\$Treatment, "0")
dds <- DESeq(dds)
``````

Here is my first question: from what I understand, this design is accounting for patient difference. So I think I don't need to build a design like: ~Sample+PhenotypeGenderTreatment (actually this will lead to matrix full rank error). Am I right?

My biggest problem is I'm not sure how to get the main effect of Treatment on the cell, no matter normal or tumour and male or female (if I wanna keep the current design instead of using a simple comparison to combine every factor into one). Is it like the following?

``````results(dds,contrast=list(c("Treatment_A_vs_0", "PhenotypeNormal.GenderF", "PhenotypeNormal.GenderF.TreatmentA")))
``````

And for the interaction of Phenotype and Treatment B in male:

``````results(dds,contrast=list(c("PhenotypeNormal.TreatmentB")))
``````

for interaction of Phenotype and Treatment B in female:

``````results(dds,contrast=list(c("PhenotypeNormal.TreatmentB", "PhenotypeNormal.GenderF.TreatmentB")))
``````

for interaction of Gender and Treatment A in tumour:

``````results(dds,contrast=list(c("GenderF.TreatmentA")))
``````

for interaction of Gender and Treatment A in normal cell:

``````results(dds,contrast=list(c("GenderF.TreatmentA", "PhenotypeNormal.GenderF.TreatmentA")))
``````

Am I doing the right things for the above 4 interactions?

And in this design, is that possible to estimate the interaction treatment A and B on the tumour males, all males and all samples respectively?

rna-seq R • 2.4k views
modified 3.6 years ago • written 3.6 years ago by dustar1986300

I'd use a much more simple model to begin with. I probably wouldn't include the interaction between sex and either phenotype or treatment unless I had a priori reasons why the treatment should differ between the sexes (eg, if an oestrogen receptor agonist was being used).

Thanks russhh. From the code above if I've done them correctly, there is a small interaction of phenotype and treatment in both male and female (influence about 10 genes), including the key gene lead to tumour. There is also a small interaction of sex and phenotype (affected 50 genes). No interaction of gender and treatment was found. Can I ignore these interactions to build a simple model? The interaction of phenotype and treatment really made me hesitate to use a simple model, as the key gene was involved.

You can always ignore things and build a simpler model; whether it is appropriate to do so is a different matter... This would be so much easier if you could just compare nested models.

Without putting in a fair bit of work, I'm not sure I could tell you if your model & contrasts are correctly set up, though I'd be surprised if your contrasts are correct: you're looking straight at the design coefficients