Theory and application of the maximum likelihood principle to NMR parameter estimation of multidimensional NMR data

Abstract

A general theory has been developed for the application of the maximum likelihood (ML) principle to the estimation of NMR parameters (frequency and amplitudes) from multidimensional time-domain NMR data. A computer program (ChiFit) has been written that carries out ML parameter estimation in the D-1 indirectly detected dimensions of a D-dimensional NMR data set. The performance of this algorithm has been tested with experimental three-dimensional (HNCO) and four-dimensional (HN(CO)-CAHA) data from a small protein labeled with 13C and 15N. These data sets, with different levels of digital resolution, were processed using ChiFit for ML analysis and employing conventional Fourier transform methods with prior extrapolation of the time-domain dimensions by linear prediction. Comparison of the results indicates that the ML approach provides superior frequency resolution compared to conventional methods, particularly under conditions of limited digital resolution in the time-domain input data, as is characteristic of D-dimensional NMR data of biomolecules. Close correspondence is demonstrated between the results of analyzing multidimensional time-domain NMR data by Fourier transformation, Bayesian probability theory [Chylla, R.A. and Markley, J.L. (1993) J. Biomol. NMR, 3, 515-533], and the ML principle.

Databases and Software for NMR-Based Metabolomics

“…The quantification portion of targeted profiling consists of measurement of the intensity of each signal and its profile across the ensemble. Although the "identify" task is usually thought of as a prerequisite to the "quantify" task, a modeling approach such as deconvolution [17][18][19] may quantify signals and modulations of their intensity prior to the signals being assigned to an NMR transition. Fig.…”

A six-dimensional alpha proton detection-based APSY experiment for backbone assignment of intrinsically disordered proteins

“…In addition, 13 C-detection is a very powerful approach to increase signal dispersion and reduce line broadening, which is either caused by solvent exchange or conformational exchange (Felli and Brutscher 2009;Skora et al 2010). 13 C-detected NMR experiments (Bermel et al 2006a(Bermel et al , b, 2012aBertini et al 2004;Pervushin and Eletsky 2003;Shimba et al 2004;Takeuchi et al 2010), in particular when combined with non-uniform sampling techniques (Barna et al 1987;Bermel et al 2012bBermel et al , 2013Chylla and Markley 1995;Coggins and Zhou 2006;Holland et al 2011;Kazimierczuk et al 2006;Kazimierczuk and Orekhov 2011;Korzhneva et al 2001;Novacek et al 2011;Orekhov et al 2001), can therefore enable the resonance assignment and structural characterization of large IDPs (Csizmok et al 2008;Narayanan et al 2010;Novacek et al 2013;Zawadzka-Kazimierczuk et al 2012a).…”

Fast multi-dimensional NMR acquisition and processing using the sparse FFT

“…The latter requirement excludes powerful non-parametric NUS processing algorithms designed to reconstruct the full spectrum, such as Maximum Entropy (ME) (Barna et al 1987;Hoch et al 2014), Projection Reconstruction (PR) (Freeman and Kupce 2003), Spectroscopy by Integration of Frequency and Time Domain Information (SIFT) (Frey et al 2013;Matsuki et al 2009), Signal Separation Algorithm (SSA) (Stanek et al 2012), Compressed Sensing (Holland et al 2011;Kazimierczuk and Orekhov 2011), and Low Rank reconstruction (Qu et al 2014). The parametric methods such as Bayesian (Bretthorst 1990), maximum likelihood (Chylla and Markley 1995), and multidimensional decomposition (MDD) approximate the spectrum using a relatively small number of adjustable parameters, and thus are not limited in spectral dimensionality and resolution. However, due to the intrinsic problems of choosing the right model and convergence, the parametric algorithms cannot guaranty the detection of all significant signals in a large spectrum.…”