Quantitative Investigation of Probabilistic Spectral Processing Methods Using Simulated NMR Data

Abstract

We have investigated the performance of two approaches for Bayesian processing of NMR data: Memsys5—a maximum entropy algorithm—and Massive Inference (MassInf). Spectra were simulated at two different noise levels to assess the algorithms' reconstruction of signals with close doublets, linewidth variation, high dynamic range, variable line shapes, and nonuniform baselines. The resulting reconstructions were analyzed in terms of efficacy of deconvolution, reconstruction of mock data, and accuracy of line positions and integrals. In the majority of the tests performed, the residuals between the simulated input and reconstruction were below the input noise. MassInf showed greater robustness than Memsys5 at rejecting noise peaks in regions where there was genuinely no signal, thus producing more visually impressive noise-suppressed spectra. Doublets with splittings down to 0.7 of the line width were resolved, even at relatively low signal to root mean square (rms) noise ratios (∼10), and large relative intensities (e.g., 10:1). Where multiplets were correctly resolved, both algorithms were accurate in their inferred line positions with errors seldom above 0.2 linewidths. At high signal to rms noise ratios (e.g., 100:1), line integrals were comparable with those obtained by directly integrating the input spectrum. However, the relative performance of the Bayesian algorithms improved as the noise level was increased. Finally, it was found that any curvature of the baseline significantly decreased both of the algorithms' noise suppression abilities as well as increasing their processing time requirements.