Limits...
Bayesian weighting of statistical potentials in NMR structure calculation.

Mechelke M, Habeck M - PLoS ONE (2014)

Bottom Line: We demonstrate that an optimally weighted potential leads to an improvement in the accuracy and quality of the final structure, especially with sparse and noisy data.Our findings suggest that no universally optimal weight exists, and that the weight should be determined based on the experimental data.Other knowledge-based potentials can be incorporated using the same approach.

View Article: PubMed Central - PubMed

Affiliation: Institute for Mathematical Stochastics, Georg August University Göttingen, Göttingen, Germany; Department of Protein Evolution, Max Planck Institute for Developmental Biology, Tübingen, Germany.

ABSTRACT
The use of statistical potentials in NMR structure calculation improves the accuracy of the final structure but also raises issues of double counting and possible bias. Because statistical potentials are averaged over a large set of structures, they may not reflect the preferences of a particular structure or data set. We propose a Bayesian method to incorporate a knowledge-based backbone dihedral angle potential into an NMR structure calculation. To avoid bias exerted through the backbone potential, we adjust its weight by inferring it from the experimental data. We demonstrate that an optimally weighted potential leads to an improvement in the accuracy and quality of the final structure, especially with sparse and noisy data. Our findings suggest that no universally optimal weight exists, and that the weight should be determined based on the experimental data. Other knowledge-based potentials can be incorporated using the same approach.

Show MeSH
Impact on structure ensembles from sparse and noisy NMR data.Panels A, C show the results for the sparse Fyn-SH3 data set. Panels B, D show the results for the solid-state data. The top row displays the RMSD distributions with  (white),  (black) and optimal  (grey). The grey distribution shown in the bottom panels is the model evidence as a function of the weight .
© Copyright Policy
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC4067304&req=5

pone-0100197-g004: Impact on structure ensembles from sparse and noisy NMR data.Panels A, C show the results for the sparse Fyn-SH3 data set. Panels B, D show the results for the solid-state data. The top row displays the RMSD distributions with (white), (black) and optimal (grey). The grey distribution shown in the bottom panels is the model evidence as a function of the weight .

Mentions: Incorporation of the backbone potential can significantly improve the accuracy of the ensemble as measured by the RMSD to the crystal structure (Figure 4). But we also observe that an overly strong backbone potential can do more harm than good. For large values, the RMSD distribution deteriorates and even shows multiple peaks in case of the sparse data set. With an optimally weighted backbone potential also the accuracy of the mean structure is consistently higher than the accuracy of the individual members of the structure ensemble (see Table S1 in File S1). This indicates that the structure ensembles are better defined when using the backbone potential. Figure 4 shows that the model evidence peaks where the average RMSD to the crystal structure shows a minimum.


Bayesian weighting of statistical potentials in NMR structure calculation.

Mechelke M, Habeck M - PLoS ONE (2014)

Impact on structure ensembles from sparse and noisy NMR data.Panels A, C show the results for the sparse Fyn-SH3 data set. Panels B, D show the results for the solid-state data. The top row displays the RMSD distributions with  (white),  (black) and optimal  (grey). The grey distribution shown in the bottom panels is the model evidence as a function of the weight .
© Copyright Policy
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC4067304&req=5

pone-0100197-g004: Impact on structure ensembles from sparse and noisy NMR data.Panels A, C show the results for the sparse Fyn-SH3 data set. Panels B, D show the results for the solid-state data. The top row displays the RMSD distributions with (white), (black) and optimal (grey). The grey distribution shown in the bottom panels is the model evidence as a function of the weight .
Mentions: Incorporation of the backbone potential can significantly improve the accuracy of the ensemble as measured by the RMSD to the crystal structure (Figure 4). But we also observe that an overly strong backbone potential can do more harm than good. For large values, the RMSD distribution deteriorates and even shows multiple peaks in case of the sparse data set. With an optimally weighted backbone potential also the accuracy of the mean structure is consistently higher than the accuracy of the individual members of the structure ensemble (see Table S1 in File S1). This indicates that the structure ensembles are better defined when using the backbone potential. Figure 4 shows that the model evidence peaks where the average RMSD to the crystal structure shows a minimum.

Bottom Line: We demonstrate that an optimally weighted potential leads to an improvement in the accuracy and quality of the final structure, especially with sparse and noisy data.Our findings suggest that no universally optimal weight exists, and that the weight should be determined based on the experimental data.Other knowledge-based potentials can be incorporated using the same approach.

View Article: PubMed Central - PubMed

Affiliation: Institute for Mathematical Stochastics, Georg August University Göttingen, Göttingen, Germany; Department of Protein Evolution, Max Planck Institute for Developmental Biology, Tübingen, Germany.

ABSTRACT
The use of statistical potentials in NMR structure calculation improves the accuracy of the final structure but also raises issues of double counting and possible bias. Because statistical potentials are averaged over a large set of structures, they may not reflect the preferences of a particular structure or data set. We propose a Bayesian method to incorporate a knowledge-based backbone dihedral angle potential into an NMR structure calculation. To avoid bias exerted through the backbone potential, we adjust its weight by inferring it from the experimental data. We demonstrate that an optimally weighted potential leads to an improvement in the accuracy and quality of the final structure, especially with sparse and noisy data. Our findings suggest that no universally optimal weight exists, and that the weight should be determined based on the experimental data. Other knowledge-based potentials can be incorporated using the same approach.

Show MeSH