1
$\begingroup$

I want to play some differentially expressed genes for gliomas cancer. I collected cancer tissue profiles from the TCGA database and normal tissue profiles from the GTEx database. I straightforward use ComBat from R package "sva", and the result shows little differentially expressed genes (DEGs) were found. How can I remove batch effects correctly?

Thanks

$\endgroup$
2
  • 3
    $\begingroup$ You can't because batch is confounded with cancer/normal. All cancer are from batch=TCGA and all normals from batch=GTEx. $\endgroup$
    – ATpoint
    Nov 16 '20 at 13:10
  • $\begingroup$ Yup. I'm very sorry, but RNASeq doesn't work like that. It is far too sensitive to batch effects. $\endgroup$
    – swbarnes2
    Nov 17 '20 at 3:30
0
$\begingroup$

This has been asked and answered here and Biostars. My take home message from many similar posts is "not to bother with correcting batch effects when performing differential expression analysis but to include 'batch' as a covariate". The latter accounts for the batch effects.

EDIT: This does not really answer OP's question, as @swbarnes2 and @ATpoint pointed out, they have a completely confounded design and the linear regression model would not accept such a model matrix.

$\endgroup$
4
  • 1
    $\begingroup$ This cannot possibly work if batch is totally confounded with sample type. $\endgroup$
    – swbarnes2
    Nov 17 '20 at 3:30
  • $\begingroup$ Thanks for your answer. Is it OK to ignore the batch effects? $\endgroup$
    – Kai He
    Nov 17 '20 at 12:12
  • $\begingroup$ Personally I don't think it will be OK to ignore the batch effects. If you still perform the DEA, you should be extremely cautious about the results, you will not be able to tell if the differential expression is the result of true biology or batch effect. You might want to turn to the bench to validate your results if possible. $\endgroup$
    – haci
    Nov 17 '20 at 12:24
  • $\begingroup$ I would at least check the documentation of TCGA and GTEx whether they at least used the same library prep kits. If not it is basically just guessing if you ask me regardless which statistics or machine/deep learning you throw at it. Then go for the genes that in your DE are really highly signiicant, and try to validate findings with independent analysis or experiments. This DE analysis here is basically just hypothesis generation, nothing more. $\endgroup$
    – ATpoint
    Nov 17 '20 at 19:14
0
$\begingroup$

There is a paper that did what you are asking: Unifying cancer and normal RNA sequencing data from different sources.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.