question about co-expression in TCGA

Question

[tujuchuanli]

Hi, all

I am working on gene co-expression analysis by using BRCA dataset in TCGA. In order to get a backgroud dataset, I randomly pick two genes and calculate the Pearson correlation coefficient for 4000 times.

Each time I randomly pick two genes and get a data matrix. There are two columns which represent two genes. There are many rows which represent samples. Each row have two values which represent expression value of gene A and B in one specific sample (RNA-seq data, RPKM). I calculate the Pearson correlation coefficient in R using cor.test. I do above calculations for 4000 times.

From my understanding, the frequency distribution of Pearson correlation coefficient should be half negative and half positve. However, most of Pearson correlation coefficients are positive. The percentage of negative Pearson correlation coefficient is less than 10%.

What is wrong to my calculation? Could you please give me some suggestions?

Thanks

[tujuchuanli]

Thanks, Kevin. I download the RNA-seq data of hg38 directly from TCGA website which do not have RSEM version to measure gene expression level. I am planing to download hg19 version to test what you say. Thanks again

[GenoMax]

Please use ADD COMMENT/ADD REPLY when responding to existing posts to keep threads logically organized. This comment belongs under @Kevin's answer.

Answer 1

[Kevin Blighe]

What you find is likely explained by the fact that you're using RPKM data. As mentioned in my other comment ( C: question about identifying differential expressed genes in TCGA ), it would be better to obtain RSEM counts (now available for majority of, if not all, TCGA datasets) and to process these with a 'modern' normalisation strategy, and to then conduct Pearson correlation on the transformed counts.

Kevin