ROC for biomarker discovery

Question

[Whoknows]

Hi

I have done classification analysis by ROC for finding novel miRNA biomarker,

Can I use mean expression of 10 different miRNAs for ROC analysis or not?

Thanks

Answer 1

[Samuel Brady]

Yes, you can use the mean expression of 10 different miRNAs for ROC analysis.

However: If you have tested (or will test) many different combinations of 10 miRNA sets to find one that classifies your samples well, you will want to have an external validation dataset to see whether (1) your biomarker holds up elsewhere and is thus likely to be valid, or (2) is just a random result of performing many statistical tests.

As a side note, miRNAs (and genes in general) can have vastly different basal expression levels. If, for example, miRNA-A is expressed at a level 10-fold higher than miRNA-B, then the mean of miRNA-A will drown out miRNA-B in your calculation of average. This may be so even if miRNA-B is a better classifier of your samples than miRNA-A. So you might want to try rank sums or an approach that doesn't vary drastically based on basal expression.