Hi Luis,

the normalization strategy depends a little on the downstream statistics you aim to apply to your metabolomics data set. I assume you initially plan to measure n metabolites in both wt and oe strains and want to make a statement which metabolite levels are significantly different between strains, correct? Ideally you subject the same amount of material to your analysis to avoid any bias. This is problematic (differences in growth performance) which is why you want to normalize.

I would not recommend to use any external factor (protein, DNA content). The reason is that they seem to be reliable but come with variance/error of their own. Also, you don't avoid introducing FPs to your data with such a strategy because you neglect the difference in linear range/analytical response between all your different metabolites. Extreme example: (1) You measure metabolite i at the upper limit of detection in both samples (2) you find/assume that in sample 1 you did use only 50% material of sample 2 and therefore multiply sample 1 with factor 2 to correct for this. (3) You obtain an excellent difference between sample 1 and 2 for metabolite i. However, as you measured it at the upper limit of detection you would *not* have been able to observe this difference had you measured twice the amount of sample 1 in reality.

I would recomend two potential strategies. Both start with measuring your samples using your best guess for comparable amounts. Than you either normalize your samples based on the robust median peak height of each sample (described for instance in DOI: 10.1007/978-1-4939-7819-9_20). This is as good as the 'external' approach, which means it bears the same risks without the additional work. Alternatively you use the information about systematic difference in median peak height and repeat your experiment with adjusted sampling amounts. This way you minimize necessary normalization (good) but may run into trouble explaining this strategy later on to reviewers (bad).

I am afraid there is no optimal solution on a metabolomics scale where it is hard to have spiked standards for a larger number of analytes (but will be curious to learn about potential solutions in other replies).

Regards, Jan

Hello Luis,

Rightly so- before I could even harp on it, the expert Jan Lisec has given beautiful thoughts and logics, and I can not top it up at all, but trying to fill in my thoughts, rather. As he said, really depends on the question you want to ask.

1. I have strong reservations against using macromolecules for normalizing metabolites!, i.e., total DNA, or total protein for normalization (contrary to how reviewers and some from other -omics field suggest or defend) because given there are so many levels of regulation and unknown interactions before a gene/ DNA converts to protein and then into a metabolite, it is useless. Thanks to post-transcriptional, and post-translational and post-catalytic regulations/ modifications as well as say quantification errors even for total DNA/ RNA (salts) or protein (BSA over reading due to extra aromatic AA residues!). Moreover, interference from these metabolites themselves in the macromolecule assays and esp. when DNA/RNA are measured from "aqueous extracts" and then "proteins from organic/ aqueous ones" how and why to normalize a methanolic (where proteins are not there!) or chloroform extract (where DNA/RNA are absent!) for metabolites? If "rough" is better, then better leave without these biological normalizations. Worth checking this "failure" of DNA, RNA, Protein for metabolite normalization methods from Jason Papin's recent work on malaria parasite: https://www.biorxiv.org/content/early/2018/02/22/190421 : )

2. Quoting a famed metabolomics researcher- that variance in a given metabolite just due to biological variation (i.e., two plants of same growth conditions in a lab!) is so HUGE (can range from 40-80% for a given AA) there is no point in normalizing for 'total metabolite TIC/ spectra/ sum' ; also add to the fact that variability in measurement and chemical class are very different as well.

3. I would like to believe that "dry weight" would be far more reliable than "wet/ fresh weight" as that water amount and nutrition/ hydration status can be very different for clones even! Instead of liquid state pelleting, it is better to "filter, wash quickly, and harvest to lyophilize/ freeze drying" would be the best way to go- and if done in larger scales of "grams" then this weight based variability would be negligible- and metabolites during freeze drying are stable.

4. Another 'innovative' way would be to "feed them a labelled (stable isotope) substrate or something that would be transformed by your strains say, a labelled steroid that would be hydroxylated or methylated!" (may not be a metabolic substrate like glucose or acetate!) at the beginning of the experiment for both strains (and their replicates!) and see that it does not interfere with the question you are addressing, and that itself can serve as an internal standard for your experiments without interfering in your pathways/ desirable metabolites etc.

5. Can not agree more to the above that 'median' works fairly well for metabolite data for normalization post-hoc.

Metabolomics is "challenging" for a reason- and each step throws up it's own surprises for every biological question/ model system used! : ) Something beyond just automation or that of a core facility or a company giving you data!

Thanks a lot for starting this discussion and bringing in Jan into it! : )

Best wishes,

Biswa

Dear Luis

For metabolic study you need to target some steps of its metabolic cycle and then you do a blank profile with various sampling time points and then do fragment pattern study and targets few metaboliets from that in HRMS. Do simple pre sample preparation befor injecting in HRMS.

My experience with the yeast strain at aston uni where we were interested in investigating change in lipidomic composition in response to addition of antifoaming agents to bioreactors for membrane protein production, we

Observed that none of normalisation methods was able to account for all unwanted variation and ratiometric analysis based on known precursor product relstionship worked best for the study