Adding experiments 1 & 2 should be fine without modification on the computer side of things, as long as your sample preparation is consistent. To be ultra-cautious, you could add in the experiment as an additional covariate for the differential expression model.
I'm less confident about adding in TCGA data; it's probably going to look better than your own samples, which might cause problems in calculating differential expression from raw reads. You can do a separate normalisation/correction of the TCGA data, but that will limit the inferences that can be made about differential expression -- without raw count information, the shot noise associated with read sampling cannot be included in the statistical model. I notice that TCGA has already done their own filtering / expression, and that would be preferable to use (mRNA sequencing, data subtype 'Gene'):
However, even with data that has been properly normalised and filtered to exclude batch effects, there will still be population differences that influence results. My recommendation would be to do some subsampling to reduce the population effects, something like at least the following comparisons:
- 6 Vs 100 + 5 normals
- 6 Vs 100 normals
- 6 Vs 50 + 3 normals [randomly sampled from the two normal groups]
And only consider differential expression that is consistently observed in all three of these comparisons. Note that the third "subsampled" dataset can be generated more than once, taking different subgroups each time.
If you are going to use separately normalised data, then DESeq2 may not be appropiate (because it expects count-level data). Something like Limma could work for DE analysis, see chapter 15 of the user's guide for RNASeq analysis. Unfortunately I'm not very familiar with using Limma, so can't provide any advice on things like what to watch out for in the input data and/or results.