Bradley Worley about
Phase-Scatter Correction Simultaneously phase and normalize NMR spectra


While most data sources for chemometrics only suffer from random dilution factors, correctable by normalizing each observation to a common scale, NMR data also contains phase error that can interfere with obtaining correct normalization coefficients during analysis. Because of that, I developed an algorithm (pscorr) to simultaneously phase and normalize 1D NMR spectra for chemometrics applications.

There’s really not much to say about PSC. It’s based on a nonlinear least squares minimization of the L2 norm of the differences between each scaled, phased spectrum and the mean pseudospectrum of the dataset. It runs in GNU Octave and, with some subtle tweaks, would run without issue in MATLAB.

Selected results

To test the performance of PSC relative to other commonly used normalization methods, I ran a Monte Carlo simulation. For various degrees of random phase error introduced into a simulated 1D NMR metabolomics dataset, I calculated normalization factors and modeled the datasets using PCA. The spectral intensities in the data were randomly generated with certain mean values such that six distinct classes should exist in the scores space. Closeness to the true normalization factor (RMSE(alpha)) and quality of the clusters in scores space (J2) were then calculated over all phase errors.

Long story short, PSC copes with more phase error than any other method of normalization, producing better resolved clusters in PCA analyses. Provided the spectra can be perfectly phased, however, MSC provides tighter clusters and HM/PQ provide better normalization factor estimates.

The code

The pscorr algorithm is standalone code, but is available as part of the MVAPACK toolbox. You can get it on the MVAPACK page. You can read more about PSC in the following papers: