Dear experts!

The design formula of DESeq2 has been extensively discussed, however, I could not translate any example/explanation to my specific situation.

I have two different transcriptomic experiments and I want to DESeq2 them together to obtain a single log2FoldChange and padjusted value for the overall class induction (Class) even if the inducers are different compounds for different experiments (treatments). I think technically it is possible, but is it conceptually and/or statistically correct? The two experiments have different design especially the ratio of treated/untreated samples and high batch variability (eg row read counts of untreated control samples for gene 5).

Treatment Gene_1  Gene_2  Gene_3   Gene_4  Gene_5    Class   Experiment
treat1      3      156      1       227     662       CI       1
treat1      3      524      9       1191    1045      CI       1
treat1      2      143      1       295     687       CI       1
control     3      449      2       434     987       Ctr      1
control     4      282      10      550     887       Ctr      1
control    10      440      16      1057    1706      Ctr      1
treat2     1       211      0       396     891       CI       2
treat2     4       194      0       664     848       CI       2
treat2     4       315      7       598     1252      CI       2
control    8       1513     25      1379    10529     Ctr      2
control    7       1524     10      1464    10441     Ctr      2
control    5       1383     7       1202    7460      Ctr      2
control    11      1148     15      1269    8440      Ctr      2
control    7       976      8       886     9633      Ctr      2
control    4       1202     15      1293    6111      Ctr      2

I would use the code below to :

dds <- DESeqDataSetFromMatrix(countData = data,
                              colData = meta,
                              design = ~ Class+ Experiment

and retrieve results by pairwise comparison e.g for CI

results(dds, contrast = c("Class", CI, control))

I have the impression that this approach will dilute or increase the actual fold change and relative padj of genes when compared for instance with the ones obtained by running DESeq2 for a single experiment separately.

Alternatively, how could I answer my question?

Many thanks for any clarification.

A design of ~x + y, followed by testing of fold change due to y, will account for changes in counts due to x, and give you the marginal effect of y across strata of x. Sounds like this is what you want from the first paragraph.

Hi Micheal, thank you for taking the time to have a look at it.

I'm glad to hear that it is actually possible to run a DESeq2 analysis to combine experiments with different sampling size and at the same time correct for such high variability between groups of the same entity.

Yes you got the point, it seems to be just that simple.

Thanks!