Limits...
Deconvolution of Complex 1D NMR Spectra Using Objective Model Selection.

Hughes TS, Wilson HD, de Vera IM, Kojetin DJ - PLoS ONE (2015)

Bottom Line: Fluorine (19F) NMR has emerged as a useful tool for characterization of slow dynamics in 19F-labeled proteins.In current methods, determination of the deconvolution model best supported by the data is done manually through comparison of residual error values, which can be time consuming and requires model selection by the user.In contrast, the BIC method used by decond1d provides a quantitative method for model comparison that penalizes for model complexity helping to prevent over-fitting of the data and allows identification of the most parsimonious model.

View Article: PubMed Central - PubMed

Affiliation: Department of Molecular Therapeutics, The Scripps Research Institute, Scripps Florida, Jupiter, Florida, 33458, United States of America.

ABSTRACT
Fluorine (19F) NMR has emerged as a useful tool for characterization of slow dynamics in 19F-labeled proteins. One-dimensional (1D) 19F NMR spectra of proteins can be broad, irregular and complex, due to exchange of probe nuclei between distinct electrostatic environments; and therefore cannot be deconvoluted and analyzed in an objective way using currently available software. We have developed a Python-based deconvolution program, decon1d, which uses Bayesian information criteria (BIC) to objectively determine which model (number of peaks) would most likely produce the experimentally obtained data. The method also allows for fitting of intermediate exchange spectra, which is not supported by current software in the absence of a specific kinetic model. In current methods, determination of the deconvolution model best supported by the data is done manually through comparison of residual error values, which can be time consuming and requires model selection by the user. In contrast, the BIC method used by decond1d provides a quantitative method for model comparison that penalizes for model complexity helping to prevent over-fitting of the data and allows identification of the most parsimonious model. The decon1d program is freely available as a downloadable Python script at the project website (https://github.com/hughests/decon1d/).

No MeSH data available.


Related in: MedlinePlus

Fitting of simulated spectra with decon1d.a-d) Non-exchange broadened spectra of varying signal-to-noise ratio and number, width, frequency and height of component peaks were simulated (top row). decon1d was then used to fit these simulated spectra allowing for either fixed phase (middle row) or variable phase (bottom row). The color of the component peaks identified in each fit serves as a visual aid for comparisons between fits as it identifies the approximate chemical shift of the peak center, indicated by the colored bar on the bottom. The difference between the data and the fit (residual error) is shown in grey and the sum of individual fitted peaks is shown in green. An alternate deconvolution of the variable phase fit for column c is displayed in S5 Fig.
© Copyright Policy
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC4524620&req=5

pone.0134474.g002: Fitting of simulated spectra with decon1d.a-d) Non-exchange broadened spectra of varying signal-to-noise ratio and number, width, frequency and height of component peaks were simulated (top row). decon1d was then used to fit these simulated spectra allowing for either fixed phase (middle row) or variable phase (bottom row). The color of the component peaks identified in each fit serves as a visual aid for comparisons between fits as it identifies the approximate chemical shift of the peak center, indicated by the colored bar on the bottom. The difference between the data and the fit (residual error) is shown in grey and the sum of individual fitted peaks is shown in green. An alternate deconvolution of the variable phase fit for column c is displayed in S5 Fig.

Mentions: To test the utility and robustness of the decon1d program, we generated simulated Lorentzian peak NMR spectra using MATLAB with randomly chosen number of peaks (generated from a uniform distribution in the interval [0 12]), width (generated from a normal distribution where μ = 0.5 ppm, σ = 5 ppm), center (generated from a normal distribution where μ = 0 ppm, σ = 5 ppm), and intensity (generated from a normal distribution where μ = 1, σ = 0.6, then normalized). These simulated spectra were fit using decon1d with the phase of each peak limited to a range within ± π/50 radians. The results, which are a series of spectral predictions, were compared to the starting simulated spectra. In all cases the best model contained fewer or the same number of peaks that comprised the simulated spectrum (Fig 2 and S2–S4 Figs). In cases where an isolated, well-resolved peak can be detected by eye, the fit is excellent (e.g. Fig 2A and 2B). In cases where peaks are overlapping, the program consistently underestimates the number of peaks that a given spectral area comprises, combining several underlying peaks into one fitted peak (e.g. Fig 2C and 2D); thus the decon1d fitting procedure provides a lower bound on the true number of peaks. In decon1d, limiting the phase of the component peaks in each model to ≤ ± π/50 radians is considered for statistical purposes as ‘fixed phase’. Importantly, allowing the phase to vary over a wider interval is necessary for fitting intermediate exchange and mis-phased data (vide infra).


Deconvolution of Complex 1D NMR Spectra Using Objective Model Selection.

Hughes TS, Wilson HD, de Vera IM, Kojetin DJ - PLoS ONE (2015)

Fitting of simulated spectra with decon1d.a-d) Non-exchange broadened spectra of varying signal-to-noise ratio and number, width, frequency and height of component peaks were simulated (top row). decon1d was then used to fit these simulated spectra allowing for either fixed phase (middle row) or variable phase (bottom row). The color of the component peaks identified in each fit serves as a visual aid for comparisons between fits as it identifies the approximate chemical shift of the peak center, indicated by the colored bar on the bottom. The difference between the data and the fit (residual error) is shown in grey and the sum of individual fitted peaks is shown in green. An alternate deconvolution of the variable phase fit for column c is displayed in S5 Fig.
© Copyright Policy
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC4524620&req=5

pone.0134474.g002: Fitting of simulated spectra with decon1d.a-d) Non-exchange broadened spectra of varying signal-to-noise ratio and number, width, frequency and height of component peaks were simulated (top row). decon1d was then used to fit these simulated spectra allowing for either fixed phase (middle row) or variable phase (bottom row). The color of the component peaks identified in each fit serves as a visual aid for comparisons between fits as it identifies the approximate chemical shift of the peak center, indicated by the colored bar on the bottom. The difference between the data and the fit (residual error) is shown in grey and the sum of individual fitted peaks is shown in green. An alternate deconvolution of the variable phase fit for column c is displayed in S5 Fig.
Mentions: To test the utility and robustness of the decon1d program, we generated simulated Lorentzian peak NMR spectra using MATLAB with randomly chosen number of peaks (generated from a uniform distribution in the interval [0 12]), width (generated from a normal distribution where μ = 0.5 ppm, σ = 5 ppm), center (generated from a normal distribution where μ = 0 ppm, σ = 5 ppm), and intensity (generated from a normal distribution where μ = 1, σ = 0.6, then normalized). These simulated spectra were fit using decon1d with the phase of each peak limited to a range within ± π/50 radians. The results, which are a series of spectral predictions, were compared to the starting simulated spectra. In all cases the best model contained fewer or the same number of peaks that comprised the simulated spectrum (Fig 2 and S2–S4 Figs). In cases where an isolated, well-resolved peak can be detected by eye, the fit is excellent (e.g. Fig 2A and 2B). In cases where peaks are overlapping, the program consistently underestimates the number of peaks that a given spectral area comprises, combining several underlying peaks into one fitted peak (e.g. Fig 2C and 2D); thus the decon1d fitting procedure provides a lower bound on the true number of peaks. In decon1d, limiting the phase of the component peaks in each model to ≤ ± π/50 radians is considered for statistical purposes as ‘fixed phase’. Importantly, allowing the phase to vary over a wider interval is necessary for fitting intermediate exchange and mis-phased data (vide infra).

Bottom Line: Fluorine (19F) NMR has emerged as a useful tool for characterization of slow dynamics in 19F-labeled proteins.In current methods, determination of the deconvolution model best supported by the data is done manually through comparison of residual error values, which can be time consuming and requires model selection by the user.In contrast, the BIC method used by decond1d provides a quantitative method for model comparison that penalizes for model complexity helping to prevent over-fitting of the data and allows identification of the most parsimonious model.

View Article: PubMed Central - PubMed

Affiliation: Department of Molecular Therapeutics, The Scripps Research Institute, Scripps Florida, Jupiter, Florida, 33458, United States of America.

ABSTRACT
Fluorine (19F) NMR has emerged as a useful tool for characterization of slow dynamics in 19F-labeled proteins. One-dimensional (1D) 19F NMR spectra of proteins can be broad, irregular and complex, due to exchange of probe nuclei between distinct electrostatic environments; and therefore cannot be deconvoluted and analyzed in an objective way using currently available software. We have developed a Python-based deconvolution program, decon1d, which uses Bayesian information criteria (BIC) to objectively determine which model (number of peaks) would most likely produce the experimentally obtained data. The method also allows for fitting of intermediate exchange spectra, which is not supported by current software in the absence of a specific kinetic model. In current methods, determination of the deconvolution model best supported by the data is done manually through comparison of residual error values, which can be time consuming and requires model selection by the user. In contrast, the BIC method used by decond1d provides a quantitative method for model comparison that penalizes for model complexity helping to prevent over-fitting of the data and allows identification of the most parsimonious model. The decon1d program is freely available as a downloadable Python script at the project website (https://github.com/hughests/decon1d/).

No MeSH data available.


Related in: MedlinePlus