I want to compare publicly available RNA-Seq datasets. My end goal is to obtain a set of genes that are collectively up/down regulated across all datasets or in a subset of Treatments. Can I use a master sample table containing all datasets and use Interactions to direct desired comparisons, if the datasets contain a differing number of replicates and different treatments?

These datasets contain a differing number of biological replicates (2 to 4) and different treatment conditions (>5). About half of them have treatment time points, but the time points between datasets are different. I have quantified the Gene-Counts with Salmon.

I am not fluent in R, and so I originally intended to create a new sample table for each comparison that I wanted to make. However this quickly became unfeasible as I tried to manually rename each variable for the new comparison.

Next I tried creating a "master" sample table with the intention of subsetting the samples for comparison (a shortened version of the table is below). I intended to use Interactions, similar to what is describe for genotype comparisons in the Vignette. However, I either get the incorrect comparisons or the Model matrix not full rank error.

Essentially, for each dataset I want to compare the treated condition(s) to the control. Is this possible when the datasets have varying treatment conditions and a varying number of replicates? Or will I have to create independent sample tables for each comparison?

Thanks for any help!

#First I make a dds against all groups (grouped by biological replicates only)
ddsTxi <- DESeqDataSetFromTximport(txi,
                                   colData = samples,
                                   design = ~ Condition)

#Then I try to define Interactions
ddsTxi$group <- factor(paste0(ddsTxi$GSE, ddsTxi$Group))
design(ddsTxi) <- ~ group
ddsTxi <- DESeq(ddsTxi)
results(ddsTxi, contrast=c("group", "Control", "Treatment"))

This code results in comparison to other datasets, but doesn't actually compare Control to Treatment within a dataset (using my full sample table)

> resultsNames(ddsTxi)
 [1] "Intercept"                                    "group_GSE127851Treatment_vs_GSE127851Control"
 [3] "group_GSE135568Control_vs_GSE127851Control"   "group_GSE135568Treatment_vs_GSE127851Control"
 [5] "group_GSE140504Control_vs_GSE127851Control"   "group_GSE140504Treatment_vs_GSE127851Control"
 [7] "group_GSE148166Control_vs_GSE127851Control"   "group_GSE148166Treatment_vs_GSE127851Control"
 [9] "group_GSE152916Control_vs_GSE127851Control"   "group_GSE152916Treatment_vs_GSE127851Control"
[11] "group_GSE80512Control_vs_GSE127851Control"    "group_GSE80512Treatment_vs_GSE127851Control" 
[13] "group_GSE94495Control_vs_GSE127851Control"    "group_GSE94495Treatment_vs_GSE127851Control" 
[15] "group_GSE98352Control_vs_GSE127851Control"    "group_GSE98352Treatment_vs_GSE127851Control"

Example Sample Table

Just to be sure, you obtained raw data and then quantified read counts using Salmon? Batch effects will likely still exist between the datasets, I think. You may want to consider adding dataset to your design formula so that any effect of dataset is adjusted, i.e., using ~ Condition + dataset. There are likely greater / 'grander' issues pertaining to merging datasets in this way, which are as yet not interpretable to me from where I am sitting.

Generally, though, why not just perform pairwise comparisons, as I show here: https://www.biostars.org/p/325009/#325036

I see no immediate need for an interaction design formula.

Essentially, for each dataset I want to compare the treated condition(s) to the control.

If this is all that you want to do, then I would process each dataset independently and then meta-analyse the p-values.

Kevin