I have been applying pre-processing steps to metabolomics data prior to statistical analysis (PCA, heat maps and clustering based on correlations). In general:
- I have log-transformed the data to reduce heterocedasticity
- Then I pareto-scaled the data to balance high abundant and low abundant metabolites
I have been reading some papers and speaking with colleagues about data pre-processing because some people argue that scaling the data should be sufficient, others prefer to transform and then scale prior to analysis. So I was wondering what is your experience and suggestions regarding data pre-processing? Off course it also depends on the assumptions of the statistical method but what's your general view on this?
I think the protocol for targeted and untargeted metabolomics data processing will be different.
Hi Diana,
you didn't mention it yet but normalization for non-biological artifacts (run order effects, batch effects) may be a step to include in the workflow. You will get a good overview which steps are generally applied by processing example data in online workflows like Metaboanalyst (http://www.metaboanalyst.ca/). If you are familiar with R you may want to check out the steps/software described in this paper (DOI: 10.1007/978-1-4939-7819-9_20). As a general comment and recommendation: I always find it useful to check the preformance of any preprocessing steps using a non-targeted, unsupervised method like PCA, i.e. I visualize my data by PCA after individual preprocessing steps to check if the amount of biological variance relative to total variance was increased.
Basically, the scores plot of the PCA (your samples) should reflect your biological/experimental expectation better after a preprocessing step than without this step (e.g. because replicate groups are better clustered together or because technical artifacts are gone etc.)
Bests, Jan
Hi, I belive normalisation Jan Lisec is talking about is different than the one you said. I will try to explain, since I'm not a staticistian and this is just by what I have seen... What happens is that analisys as PCA are highly susceptible for differences in ranges. So if a variable values range from 5 to 10 and another from 50 to 100 the second usually get higher importance in analisys like PCA. To handle with this you can do a normalisation of the variables which forces all variable ranges to be similar.
Some softwares as Statistica do this by default when performing PCA but for instance Primer don't...
Just to add an article, I belive is interesting, about when normalize or standardize:
https://medium.com/@rrfd/standardize-or-normalize-examples-in-python-e3f174b65dfc
- Badges
- Science topic
- Similar topics
- Social Science
- Sociology