Comparing log2-fold changes to find elements with virtually the same log2-fold changes across different conditions

Question

[robrob]

Hi,

I recently did a metabolomics experiment where I analysed the intracellular metabolite composition of 2 cell lines (wild-type and knock out for my target gene) grown in 3 different media. Firstly, I calculated for each metabolite the log2-fold change between knock-out (KO) and wild-type (WT) in each different media. In other words, after my analysis I had 3 log2-fold changes for each metabolite:

1. Log2-fold change KO/WT in the first medium
2. Log2-fold change KO/WT in the second medium
3. Log2-fold change KO/WT in the third medium.

Because of the nature of metabolomics, all these log2-fold changes differ between each other. However, I would like to find metabolites which have virtually the same log2-fold change. In other words, I am interested in those metabolites whose relative concentration in KO vs WT stays the same regardless of the medium (for example, UMP has a log2-fold change KO/WT of 5.35 in the first medium, 5.72 in the second medium and 5.54 in the third medium). How can I find those metabolites? Which type of statistical test should I use? And which threshold?

Thanks in advance! 😊

Answer 1

[ATpoint]

One simple way would be to filter for genes where the standard deviation or median absolute deviation is not more than x away from the mean.

A statistical test could be the use of interaction terms measuring 'fold changes of fold changes' and then going for those with strong evidence against differential ones (so interaction fold changes are not different) so large p-values, filtered again for metabolites that in the initial analysis had good evidence for differential events. limma could provide such an analysis.

Answer 2

[i.sudbery]

Was this metabolomics experiment conducted using mass-spec? If so, is the raw data effective counts? If so, you could probably use DESeq2. Similar to what @ATPoint was saying about looking at the 'fold-changes of fold changes'. However, rather than look for comparisons that had a large p-values, DESeq2 allows you to specifcy a different alternative hypothesis than Log2FoldChange != 0, like, for example |lfc| < 0.5. This way metabolites with a adjusted p-value less than 0.05 would be ones you could be confident weren't different between media.