Hi,

I have the following set up: 3 replicates of control and treatment in cell type A and 3 replicates of control and treatment in cell type B. 2 of the A samples were prepared and sequenced last fall and 1 of the A samples and 3 B samples were prepared and sequenced this spring. I did PCA analysis and unsupervised clustering with normalized counts in DESeq2 with design ~ 1, and all of the appropriate replicates cluster appropriately. However, I want to make sure that I control for any batch effect. If I make my model design = ~ date + type + condition + type:condition, rather than just design = ~ type + condition + type:condition, I find many more significant hits for the interaction term. I wasn't sure if this was the best approach or if I should run something like svaseq to pull out any batch effects?

If I understand your batches correctly, then I'm not sure you will be able to control for batch, since your B samples are all confounded with batch 2.

In other words, when you add 'date' to your model, you are trying to make the samples from two batches comparable by adjusting for the mean expression at the two different dates. But you don't have any B samples from the first batch, and only one A sample from the second batch. So the mean value for batch 1 is based only on A samples, and the mean value for batch 2 is based on one A and three B samples. So it's hard to say if the differences between the batches are due to the cell types or the batch.

If the samples group like you expect in the PCA plot, then you might be able to claim that as 'proof' that there isn't a big batch effect, in which case you can then claim it is OK to ignore the batch issue.