Hello,
Setting: We have a rather large dataset consisting of somewhat 500 samples containing metabolomics data. The data was obtained by analysing dry capillary blood using HPLC methods, yielding relative spectral intensity of approximately 650 different metabolites.
Since the relevant databases rarely contain dry blood-derived metabolomics datasets we are currently not entirely sure on how to normalise our dataset. This is as the common suspects, such as normalisation by weight, median, mean either seem very counterintuitive or hardly change the data at all.
Fortunately, we are in possession of various laboratory blood-derived values (Creatinine etc.), sort of as a ground truth value for most of our samples, which were obtained at the same timepoints to the dry blood. We were wondering if those would be useful in finding an adequate parameter to normalise our data with. For instance, Creatinine could be correlated to the amount of "water" and thus dilution of blood samples. Since we measure dry blood, the water would evaporate and therefore, using Creatinine to normalise our data would help us dampen this effect.
It would be much appreciated if anyone could shed some light on their usual practice when it comes to normalising metabolomics data for which there is no significant standard procedure?
What would you like to use this data for? Relative quantities are usually transformed by logs in order to obtain a distribution closer to normal. But it depends on your data and on the task you want to solve using the data available.
The first steps may be to 1. derive descriptive statistics on the analyte quantitations, 2. to quantify technical and biological variability and 3. to quantify/plot association with other available covariates (e.g., creatine). The information collected will help you define a normalisation strategy.
Thank you all for your answers already, I will consider them over the next couple of days. Also Andrey Davydenko let me explain the reason for our desire to normalise the data.
There are currently quite well-established metabolomics datasets from healthy and non-healthy subjects available from whole blood analysis that are used to perform pathway enrichment analysis. They can be used as a "reference" to compare data against, in order to identify potentially altered metabolic substeps ( by means of graph theory, and protein-protein interaction database searches etc). The program Metaboanalyst already automates a lot of these steps.
However, our issue remains that we cannot be certain that our dry blood metabolomics data is adequate for comparison to those references. Therefore, we are looking to find a specific parameter to normalise our data, in order to approach the expected relative concentrations (of our healthy patient samples) of the "reference" datasets.
Does this clear things up?
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics