Can we perform cross-platform differential gene expression analysis?

Question

[duankangdk]

Say hello to everyone first. I am a newbie here and have just started basic research. There is a problem that has just started to bother me recently.

Example 1: Cbioportal not only has TCGA data, but also some other metastatic cancer RNA data, but they are all types of FPKM. Can I select the genes that I want to compare, and then convert FPKM to TPM and analyze their differences(logFC)? Because some databases contain 30,000 genes, but other databases only have 17,000 genes. My question is, after TPM standardization, can we compare with each other?

Example 2. The types and quantities of genes tested on different platforms on GEO are different. Can I perform analysis across platforms? Or is it that the test results from the same platform can be compared?

Thanks to anyone who answered.

Answer 1

[Ali T. A.]

There is no general answer to this question, I guess. A general comparison, by repeating experiments on two platforms for example, is possible. But I guess you mean performing differential expression across platforms. Here are some thoughts:

1. Mixing platforms for one experiment could be generally considered as bad design. We may integrate platforms but not perform something like differential expression across platforms because the confounding variables and the intrinsic differences between the platforms could make interpreting results impossible. For instance differences between NGS and Microarrays regarding dynamic range are substantial.
2. However, if we are talking about very similar platforms, then it is liketly to be able to filter out batch effects. Here it would be helpful to have ground truth genes, for example, to compare results or a high number of replicates. Also meta data about the performed experiments may play a role in correction for bias.
3. Considering your first example, if you are using the counts to compute TPM, then you can use this of course to compare samples. These are not different platforms but different databases. Here, I would only compare genes present in both databases. However, I would go for one of the established DE-Algorithms based mainly on counts (DESeq2 or EdgeR). Using TPM would help you to see the direction but should not be used for DE Analysis. It may be also needed to correct for batch effects.