Combining GTEx and TCGA data

Question

[bbibi.msbi20rcms]

Hey,

I want to analyze Differential Transcript Usage (DTU) for TCGA isoform/transcript expression data while taking GTEx as "normal" reference. In other words, TCGA "tumor" and GTEx "normal" samples will be compared.

In order to remove variation caused by batch effects, I have to perform some batch correction. Here is a subset of my design matrix:

X sample condition batch

1. s1 tumor TCGA
2. s2 tumor TCGA
3. s3 tumor TCGA
4. s4 normal GTEx
5. s5 normal GTEx
6. s6 normal GTEx

Following error is received when trying to run DEXSeq: "The supplied design matrix will result in a model matrix that is not full rank"

I know that the error is received due to redundancy in my design matrix. But I want help in tackling this issue? Is there any way for me to modify my design matrix so as to avoid this error? Can I use TCGA and GTEx data without batch correction?

Your help will be much appreciated.

[Jeremy Leipzig]

why not just use the normal TCGA sample from the same individual?

[bbibi.msbi20rcms]

Firstly, the samples are that of "adjacent" tissues (of cancer patients) and not exactly "normal". Secondly, I want to see the differences at 3 levels (NAT, normal, tumor).

Answer 1

[ATpoint]

Batch is fully confounded with treatment, nothing you can do about it. This has been asked dozens of times before, please search google for related threads.

[bbibi.msbi20rcms]

Thanks for your reply.

I know that batch effect can lead to masking/misinterpretation of biological effect. But this is DTU we're talking about. As DTU works on proportions (isoform fractions), so in my opinion, I can go for comparison of TCGA and GTEx data without correcting batch effects.

Some expert's opinion will be appreciated. Thanks.

[ATpoint]

It is still counts that you're comparing. There is no guarantee that the proportions can be directly compared. Any difference you see can be purely technical. On the other hand, if you are absolutely forced to use these data then go ahead but interpret results with care and validate important findings. In any case you cannot correct any batch effects, there is no magic that will do that on a nested design like this.