|
In the context of metabolomics, after we use Multivariate Curve Resolution and obtaining the "concentrations" of each metabolite in a table, we may need to preprocess (normalize) the data (prior to use multivariate methods, like PCA, PLS, PLS-DA). What is the most common type of normalization and what effect it has on the data? |
|
There are many types of preprocessing available. The most used in the context of metabolomics (GCMS, LCMS) are Centering, and above all Unit Variance and Pareto scaling. - Centering: the average of a variable is subtracted from all the values of that variable. It is probably the least used, as variables may have very different orders of magnitude, thus giving more importance to some variables than to others. It is more used with spectroscopic data (NMR, FTIR...) - Unit Variance: the variables are in general centered first, and then each of its values is divided by the standard deviation of the respective variable. The advantage is that all variables get the same variance (1). The disadvantage is that the noise is increased, as irrelevant noisy variables get the same importance as the interesting ones. - Pareto scaling: the variables are also centered first and then each of its values is divided by the square root of the standard deviation of the respective variable. This has an intermediate effect between centering and Unit Variance scaling. It gives each variable a variance equal to its initial standard deviation and give less importance to noise. . |
