Which one of the expression data is proper for machine learning?

Q&A
, ,
1

Hi, thank you for your effort in launching the newest dataset version 2024Q2.
But I’m a little confused about the two kinds of gene expression data: original and batch normalized.
If I want to use the gene expression data for the machine learning model training, which one should I use?
I checked that the original has some 0 values and the batch normalized dataset doesn’t have 0 but negative values.
Can anyone recommend the dataset?

2

Hi,

I’d recommend using the batch-corrected data, as we are planning to phase out the non-batch corrected data in the future. For details, please see the pdf attached at the end of the 24Q2 release announcement.

Thanks,
Simone

3

Dear Simone,

Hi, thank you for your kind reply.
Although I already selected your answer as the solution to this question, I have one more question about the dataset; not expression, but copy number.
To my knowledge, the previous version of copy number(CN) values was provided with the log2-transformed version.
But from the 24Q2, the relative copy number matrix is no longer log2 transformed.
If I want to use CN data with the gene expression data which is log2-transformed TPM value for the machine learning data, should I use PortalOmicsCNGeneLog2.csv for the consistent range of value or use OmicsCNGene.csv to keep the original data?

Thank you for reading this question.
I look forward to your reply.

Sincerely,
Songyeon

4

Hi Songyeon,

PortalOmicsCNGeneLog2 is the log2-transformed version of OmicsCNGene. They are essentially the same data, so as far as I know, you may use either depending on what you need for your particular analysis.

Best,
Simone

1 Like
5

Thank you for your kind reply.
I’ll think about it a little more. :slight_smile: