TCGA subtype problem

[eacho111]

Hi,
it is very useful package! I download fpkm gene expression data from TCGA by TCGAbiolinks package, and i try to get the immune subtypes of the LUAD samples by using this package.when i compare this subtypes result with the subtypes of this paper(five_signature_mclust_ensemble_results.tsv.gz),their subtypes are different.
But when i use the gene expression of paper(ebppSubset.tsv.bz2) to classify these patients,the subtype result can match with the paper result(five_signature_mclust_ensemble_results.tsv.gz).So i want to know if there is any problem to classify immune subtype by using fpkm gene expression data from TCGA by TCGAbiolinks.if i need to do any thing with the gene expression download from TCGA?

Thanks!
here are some codes and results!
code and result .pptx

[Gibbsdavidl]

Hey there!

Sorry I was traveling.

So I'm not sure, but something I've seen happen is that data is occasionally already normalized; where each gene is normalized across samples. That completely breaks the classification. It's important to make sure the data is unnormalized across samples. Otherwise, fpkm is fine as all other summarizations (tpm, rpkm, etc.)

Hopefully that helps!
-dave

[KaWingLee9]

@Gibbsdavidl Hi. As you mentioned, FPKM is suitable for the algorithm. So how to define the length of the genes? Exon lengths or maximum transcript length?

[Gibbsdavidl]

Hi there,
For the most part, features are produced per sample by comparing the expression levels between two genes. As long as you're consistent across samples, I have the feeling it is not an issue. But FPKM is a standard format, and I'm sure there's a 'correct' selection on determining gene length. This page from novogene says that one should use the sum of exon lengths.