Weighted least-squares deconvolution method for discovery of group differences between complex biofluid 1H NMR spectra

Biomarker discovery through analysis of high-throughput NMR data is a challenging, time-consuming process due to the requirement of sophisticated, dataset specific preprocessing techniques and the inherent complexity of the data. Here, we demonstrate the use of weighted, constrained least-squares for fitting a linear mixture of reference standard data to complex urine NMR spectra as an automated way of utilizing current assignment knowledge and the ability to deconvolve confounded spectral regions. Following the least-squares fit, univariate statistics were used to identify metabolites associated with group differences. This method was evaluated through applications on simulated datasets and a murine diabetes dataset. Furthermore, we examined the differential ability of various weighting metrics to correctly identify discriminative markers. Our findings suggest that the weighted least-squares approach is effective for identifying biochemical discriminators of varying physiological states. Additionally, the superiority of specific weighting metrics is demonstrated in particular datasets. An additional strength of this methodology is the ability for individual investigators to couple this analysis with laboratory specific preprocessing techniques.