How To Call Loss / Gain On Cgh Data When Tumor And Normal Are Not Analysed Together?

Question

[Michi]

Hello Biostars

I wanted to tackle the TCGA CGH (comparative genomic hybridization) data from their Glioblastoma Project. I was surprised by the fact, that they would compare their tumour and normal samples separately to a reference genome! Nevertheless, for a few cases they actually put the tumour and the normal sample on the same chip, what allows a direct comparison of gains and losses between normal and tumor.

So I can get something like this:

barcode chromosome      start   stop    num.mark        seg.mean (log2)
tumour:TCGA-06-0238-01A-02D-0311-04    3       85466920        85652956        24      -0.7886
normal:TCGA-06-0238-10A-01D-0311-04    3       85458029        85652956        25      -0.7479

Given their high similarity i can assume that their is no loss in the tumour compared to the normal CGH

But I am no quite sure how i should treat the cases where segments are very different in size, the difference between the values are bigger, there is a loss/gain in the normal, but not in the tumour, etc.

I am tempted to just map the segments to genes, and make a substraction, but since they are two different experiments, I doubt this is correct.. any suggestions / thoughts?

[edtuer]

could you specify which paper explains the reason? I'm also interested to know.

"Actually, reading more carefully supplementary material, I found this paper, which explains why they did not match the normal and tumour samples, and how they analysed it accounting for noise (as Jan Oosting said)"

[lilingjoyo]

How do you get calculate gains or losses from these TCGA data? I do not know much about SNP array data analysis, your help will be appreciated.

Answer 1

[Jan Oosting]

The reason to use a panel of normal samples for reference is noise.

The effect of noise is additive also when subtracting the normal signal from the tumor signal. Taking an average of all normal samples will reduce the noise level in your reference.

Quantitively the noise will be reduced approximately to the inverse of the square root of the number of samples in your reference pool. If you have paired samples you can distinguish tumor specific from germline effects or cnv in the way you show.

[Michi]

just a late request of clarification: by "the way I show" you mean substraction? Thanks

Answer 2

[Chris Miller]

Just take the ratio of tumor to normal at each locus. That won't help you identify regions that are altered in both, but presumably, what you're looking for are tumor-specific events.

Answer 3

[Michi]

Thanks for your answers!

Actually, reading more carefully supplementary material, I found this paper, which explains why they did not match the normal and tumour samples, and how they analysed it accounting for noise (as Jan Oosting said)

[Ömer An]

which paper? can you update the link for the paper please? i need to read on this also. thanks.

[Michi]

sorry, i dont remember right now. somehow the link got lost. i guess during the migration of the platform.