This is a test version of Biostars. For the public version, visit https://www.biostars.org.
DEseq2 design with multiple assays, conditions, and replicates

I want to perform a differential expression analysis with DEseq2. I have 2 conditions (input (whole cell) and a cell fraction) and 2 treatments (treated and wild type) in replicates (for simplicity, here 2).

I'm interested in the differential expression in the fraction following treatment. I think my design should be something like (Treated_fraction / Treated_input) / (WT_fraction / WT_input).

This is my countData

> head(countData)
       Geneid Length Treated_fraction_1 Treated_fraction_2 Treated_input_1 Treated_input_2 WT_fraction_1 WT_fraction_2 WT_input_1 WT_input_2
1 ENSG00000223972   1756                  0                  0               0               0             0             0          0          0
2 ENSG00000227232   2073                 29                 22              31              47            24            12         23         13
3 ENSG00000243485   1021                  0                  0               0               0             0             1          0          0
4 ENSG00000237613   1219                  0                  0               0               0             0             0          0          0
5 ENSG00000268020    947                  0                  0               0               0             0             0          0          0
6 ENSG00000240361    940                  0                  0               0               0             0             0          0          0

This is my colData

> colData
                         assays conditions replicates
Treated_fraction_1 Treated_fraction    Treated          1
Treated_fraction_2 Treated_fraction    Treated          2
Treated_input_1       Treated_input    Treated          1
Treated_input_2       Treated_input    Treated          2
WT_fraction_1           WT_fraction         WT          1
WT_fraction_2           WT_fraction         WT          2
WT_input_1               WT_input_1         WT          1
WT_input_2               WT_input_1         WT          2

So far my command is

dds <- DESeqDataSetFromMatrix(countData = subset(countData, select = -Length), 
                          colData = colData, 
                          design = ~ assays + conditions + assays:conditions,
                          tidy=TRUE)

but this give me the following error 

Error in checkFullRank(modelMatrix)

Which appears to derive from the replicates column.

What would be the correct colData and design to use in this case?

Following this, I usually do 

deseq.results <- results(dds, contrast=c("conditions", A, B))

What would be the correct results command for this analysis?

This is a simplified version with 2 conditions, can it be generalized to more conditions (i.e. input and multiple fractions).

Thanks!

deseq2 rna-seq r
• 3,144 views • link
updated by swbarnes2 swbarnes2 • written by testtube • testtube

Hi, I have the exact same problem, so I would be interested in knowing your updates about this. We have a fraction and the cell as assay and treated and control as conditions. So for colData, I have one column "assay" with fraction and cell, and another column "condition" with treatment and control. I was said that the design would be (fraction_treat/cell_treat) / (fraction_ctl/cell_ctl), but I used the same formula as you:

dds <- DESeqDataSetFromMatrix(countData = countData,
                          colData = colData,
                          design = ~ assay + condition + assay:condition)

So I think it's a design like yours (i.e. (Treated_fraction / Treated_input) / (WT_fraction / WT_input)).

I followed this post from Michael Love and used this command:

dds <- DESeq(dds, test="LRT", reduced= ~ assay + condition)
results(dds)

I don't know if a good way to do it, because if I understood well, here I use the Likelyhood ratio test (LRT) which allows to compare 2 models. One which takes into account all the possible effects and another one in which you use "reduced model" (the one here) that take off the effect of the interaction and thus allows to quantify this effect. If you think I didn't understand well, feel free to tell me.

Anyway, if you have any news about this, I would be happy to share it with you.

• 0 views • link
written by lumal29 lumal29

1 answer

Which appears to derive from the replicates column.

No it doesn't. Your replicates column is useless, but it's not the issue. The issue is that every single WT_fraction is also in the WT condition.

Make another column that just says fraction and input. To get your (A/B) / C/D) interaction, you want thatcolumn + treatment + thatcolumn:treatment as your design. Use resultNames to get the exact naming to use in the results command.

This will give you the interaction; for example, which genes changed 2 fold between treated and not treated in input,but changed 4 fold between treated and not treated in the fraction.

If you mostly care about the changes in the fraction samples only due to treatment, just use your combo column in the design, and you an specify what to compare to what with contrast.

• 0 views • link
written by swbarnes2 swbarnes2

What did you mean by combo column! ....assays column?

Thanks,

• 0 views • link
written by Nelislam4 • Nelislam4

Log in to answer this question.

More posts like this