Distribution-based classification method for baseline correction of metabolomic 1D proton nuclear magnetic resonance spectra.

Baseline distortion in 1D (1)H NMR data complicates the quantification of individual components of biofluids in metabolomic experiments. Current 1D (1)H NMR baseline correction methods usually require manual parameter and filter tuning by experienced users to obtain desirable results from complex metabolomic spectra, thus becoming prone to correction variation and biased quantification. We present a novel alternative method, BaselineCorrector, for automatically estimating the baselines of 1D (1)H NMR metabolomic data. By collecting the standard deviations of spectral intensities, using a moving window to slide through a spectrum, BaselineCorrector can model the distribution of noise standard deviation as a derived chi-squared distribution in each window and then determine optimal parameters for least-error classification of signal and noise. Due to the universal property of noise distributions, BaselineCorrector can robustly recognize the baseline segments in various spectra. In addition to the commonly used 1D NOESY and CPMG pulse sequences, BaselineCorrector also provides an algorithm for correcting diffusion-edited NMR spectra. Using its classification model, BaselineCorrector is able to preserve low signal peaks and correctly handle wide, overlapping peaks in complex metabolomic spectra.