Data analysis
To eliminate the variability in water resonance presaturation, the chemical shift region between δ 4.66 and 4.88 was removed from all spectra before statistical analysis, except for liver where to avoid bias due to baseline distortion, the region between δ 4.77 and 5.38 was removed. As previously described (Cloarec et al, 2005), all data were analyzed on full-resolution spectra (35 600 data points for liver and 36 500 data points for all other tissue extracts), normalized to the total peak area and models were constructed using O-PLS-DA with unit variance scaling on Matlab 7.0.1 software (The MathWorks Inc.). Despite the use of phosphate buffer, many urine spectra still displayed subtle pH-dependent shifts; therefore, the O-PLS-DA was performed on larger bins of 0.005 p.p.m. (1750 bucketed points) to minimize minor frequency changes in spatial components.
To aid interpretation, the O-PLS coefficients were plotted into a spectral domain using the back-scaling method (Cloarec et al, 2005). Using this method, the weights of each variable are back-scaled to their initial metric of the data and then the shape of NMR spectra and the sign of the coefficients are preserved. However, the weights of the variables can still be compared using a colour code corresponding to the square of the actual O-PLS coefficients. By construction, the O-PLS coefficients are directly proportional to the correlation coefficients between the discriminant axis and the NMR data. For this reason, the square of the coefficients can be represented in terms of correlation after applying the same corrective factor to all coefficients, allowing by this way an estimation of the amount of variance of each NMR variable involved in the discrimination.