Can A Set Of Related Metabolites Be Significant Even If Individual Metabolites Aren'T?

Question

[jobinv]

I'm asking this on behalf of a colleague who works with metabolomics, who asked me this question that I was unable to help him with. He has an ongoing study where he is comparing treated and untreated rats, and in particular looking at metabolites related to the TCA cycle by using carbon-13. Now, what he finds is that close to every individual metabolite within the TCA cycle is at lower levels in the treated cases, but none are statistically significant. However, the question is as follows: is it possible to say something about the fact that so many of these are pointing in the same direction? Specifically, are there any statistical tests which can allow us to make some statement of the sort "there are/are not more metabolites that are lower in one condition, than would be expected by sheer chance"?

[Devon Ryan]

Is your colleague just doing a T-test between the groups for the individual metabolites? If so, then I expect using a linear model (concentration ~ metabolite+treatment) might work (not having seen what this sort of data looks like).

Edit: As Istvan noted below (and I should have added), remind your colleague not to cherry-pick the metabolites for inclusion into the group!

[jobinv]

Thanks, that does sound like a good idea!

Answer 1

[Istvan Albert]

If the group of genes were selected independently of the results then you can pool the measurements for these genes and create new groups then apply the statistical test for these groups...

[jobinv]

Well, it depends on how strictly we define "independent of the results". The metabolites involved in the TCA cycle is a group in and of themselves, regardless of what the results show. We are also particularly interested in this particular metabolic pathway. However, the reason we want to look at this in detail now, is precisely because of the apparent results where so many within this group seem to be downregulated. Would you call this an independent selection, or not?

[Istvan Albert]

As long as you create the groups independent of the measurements the statistical test are reliable. What you should though account for multiple testing, how many groups did you investigate/look at that did not turn out to be significant.

[jobinv]

I see. Ok, thanks!

[jobinv]

Sorry, one quick followup question with regard to this: how would such a pooling be done? If I merely take the sum of the metabolite values for treated and the sum of the metabolite values for control, then I lose the information that is in the standard deviations for each metabolite. However, if I were to sum all of the standard deviations, I feel that that would result in a huge resulting standard deviation, and very stringent, no?

[Devon Ryan]

I expect the various metabolites have dissimilar concentration ranges, so just use a linear model with the individual metabolites as factors.

[jobinv]

I will try this, thanks. Would this take into account the individual standard deviations? The setup you wrote in the comment above doesn't seem to include it as a variable...