Abstract
Finding the dynamics of an entire macromolecule is a complex problem as the modelfree parameter values are intricately linked to the Brownian rotational diffusion of the molecule, mathematically through the autocorrelation function of the motion and statistically through model selection. The solution to this problem was formulated using set theory as an element of the universal set \({\mathfrak{U}}\)—the union of all modelfree spaces (d’Auvergne EJ and Gooley PR (2007) Mol BioSyst 3(7), 483–494). The current procedure commonly used to find the universal solution is to initially estimate the diffusion tensor parameters, to optimise the modelfree parameters of numerous models, and then to choose the best model via model selection. The global model is then optimised and the procedure repeated until convergence. In this paper a new methodology is presented which takes a different approach to this diffusion seeded modelfree paradigm. Rather than starting with the diffusion tensor this iterative protocol begins by optimising the modelfree parameters in the absence of any global model parameters, selecting between all the modelfree models, and finally optimising the diffusion tensor. The new modelfree optimisation protocol will be validated using synthetic data from Schurr JM et al. (1994) J Magn Reson B 105(3), 211–224 and the relaxation data of the bacteriorhodopsin (1–36)BR fragment from Orekhov VY (1999) J Biomol NMR 14(4), 345–356. To demonstrate the importance of this new procedure the NMR relaxation data of the Olfactory Marker Protein (OMP) of Gitti R et al. (2005) Biochem 44(28), 9673–9679 is reanalysed. The result is that the dynamics for certain secondary structural elements is very different from those originally reported.
Similar content being viewed by others
Introduction
NMR is a powerful tool for probing the fast internal motions of macromolecules on the picosecond to nanosecond timescales. By collecting NMR relaxation data, specifically the R_{1} and R_{2} relaxation rates together with the steadystate NOE, information about the motions of individual bond vectors within the molecule can be gathered. Interpreting these raw numbers by themselves to create a cohesive dynamic description of the molecule is difficult. Therefore a number of theories exist to interpret these data. The most commonly used tool is modelfree analysis (Lipari and Szabo 1982a, b; Clore et al. 1990a).
By parametric restriction of the original modelfree equations of Lipari and Szabo (1982a, b) and the extension by Clore et al. (1990b) a large number of modelfree mathematical models were constructed in the preceding paper (d’Auvergne and Gooley 2007a) which, henceforth, shall be referred to as Paper I. These models were labelled from m0 to m9 (Models 1.0–1.9 of Paper I). By assuming each spin system tumbles independently the overall rotational diffusion of each bond vector can be approximated by a separate correlation time, the local τ_{ m } (Barbato et al. 1992; Schurr et al. 1994). The addition of this parameter creates a new set of modelfree models which were labelled tm0 to tm9 in Paper I. NMR relaxation is influenced not by the correlation function C(τ) of the motions of the XH bond but by the power spectral density function J(ω), a quantity which is related to the correlation function via Fourier transform. Numerically stabilised forms of both the original and extended modelfree spectral density functions are presented in Equations (2) and (3) of Paper I.
In this paper the optimisation of the global model \({\mathfrak{S}},\) which consists of both the Brownian rotational diffusion tensor of the molecule and the internal modelfree motions of individual bond vectors, will be studied. The entirety of the complex modelfree problem, in which the motions of each spin system are both mathematically and statistically dependent on the diffusion tensor and vice versa, can be formulated using set theory (d’Auvergne and Gooley 2007b). Its solution can be derived as an element of the universal set \({\mathfrak{U}},\) the union of the diverse modelfree parameter spaces \({\mathfrak{S}}.\) Each set \({\mathfrak{S}}\) is constructed from the union of the modelfree models \({\mathfrak{F}}\) for all spin systems and the diffusion parameter set \({\mathfrak{D}}.\) A single parameter gain or loss on a single spin system shifts optimisation to a different space \({\mathfrak{S}}.\) The solution within the universal set \({\mathfrak{U}},\) which for simplicity will be referenced as the universal solution \(\widehat{{\mathfrak{U}}},\) can be formulated as (d’Auvergne and Gooley 2007b)
where \(\hat{\theta}\) is the optimised parameter vector of the space \({\mathfrak{S}}, \Updelta_{\rm KL}\) is the Kullback–Leibler discrepancy (Kullback and Leibler 1951), and χ^{2}(θ) is the chisquared function which is minimised. The equation consists of two parts, the first component belongs to the statistical field of model selection (Akaike 1973; Schwarz 1978; Linhart and Zucchini 1986; Burnham and Anderson 1998; Zucchini 2000; d’Auvergne and Gooley 2003) whereas the second belongs to the mathematical field of optimisation (Nocedal and Wright 1999; d’Auvergne and Gooley 2007a).
Ever since the original modelfree publications (Lipari and Szabo 1982a, b) the modelfree problem has been tackled by first finding an initial estimate of the diffusion tensor and then determining the modelfree dynamics of the system. This concept, which for brevity will be called the diffusion seeded modelfree paradigm, is now highly evolved and much theory has emerged to improve this path to the solution \(\widehat{{\mathfrak{U}}}.\) The technique can, at times, suffer from its rigidity assumption (Orekhov et al. 1995, 1999a, b; Korzhnev et al. 1997; d’Auvergne and Gooley 2007b). Here a different approach is proposed for finding the universal solution \(\widehat{{\mathfrak{U}}}\) of the extremely complex, convoluted modelfree optimisation and modelling problem. This new modelfree optimisation protocol incorporates the ideas of the local τ_{ m } modelfree model (Barbato et al. 1992; Schurr et al. 1994) and the optimisation of the diffusion tensor using information from these models, analogously to the linear leastsquares fitting of the quadric model (Brüschweiler et al. 1995; Lee et al. 1997). The quadric model is a methodology for determining the diffusion tensor from the local τ_{ m } parameter together with the orientation of the XH bond represented by the unit vector μ_{ i }. A local τ_{ m } value is obtained for each spin i by optimising tm2 and then the τ_{ m,i } values are approximated using the quadric model
where the eigenvalues of the matrix Q are defined as \(Q_x = ({\mathfrak{D}}_y + {\mathfrak{D}}_z)/2, Q_y = ({\mathfrak{D}}_x + {\mathfrak{D}}_z)/2,\) and \(Q_z = ({\mathfrak{D}}_x + {\mathfrak{D}}_y)/2.\) The diffusion tensor is then found by linear leastsquares fitting.
The new protocol follows the lead of Butterwick et al. (2004) whereby the diffusion seeded modelfree paradigm was reversed. Rather than starting with an initial estimate of the global diffusion tensor from the set \({\mathfrak{D}}\) the protocol starts with the modelfree parameters from \({\mathfrak{T}}.\) The first step of the protocol is the reduced spectral density mapping of Farrow et al. (1995). As R _{ ex } has been eliminated from the analysis, three modelfree models corresponding to tm1, tm2, and tm5 are employed. The modelfree parameters are optimised using the reduced spectral density values and the best model is selected using Ftests. The spherical, spheroidal, and ellipsoidal diffusion tensors are obtained by linear leastsquares fitting of the quadric model of Eq. 2 using the local τ_{ m } values (Brüschweiler et al. 1995; Lee et al. 1997). The best diffusion model is selected via Ftests and refined by iterative elimination of spin systems with high chisquared values. This tensor is used to calculate local τ_{ m } values for each spin system, approximating the multiexponential sum of the Brownian rotational diffusion correlation function with a single exponential (Woessner 1962; d’Auvergne 2006), using the quadric model of Eq. 2. In the final step of the protocol these τ_{ m } values are fixed and m1, m2, and m5 (Models 1.1, 1.2, and 1.5 of Paper I) are optimised and the best modelfree model selected using Ftests.
The new modelfree optimisation protocol utilises the core foundation of the Butterwick et al. (2004) protocol yet its divergent implementation is designed to solve Eq. 1 to find \(\widehat{{\mathfrak{U}}}.\) Models tm0 to tm9 in which no global diffusion parameters exist are employed to significantly collapse the complexity of the problem. Modelfree minimisation (Paper I), model elimination (d’Auvergne and Gooley 2006), and then AIC model selection (Akaike 1973; d’Auvergne and Gooley 2003) can be carried out in the absence of the influence of global parameters. By removing the local τ_{ m } parameter and holding the modelfree parameter values constant these models can then be used to optimise the diffusion parameters of \({\mathfrak{D}}.\) Modelfree optimisation, model elimination, AIC model selection, and optimisation of the global model \({\mathfrak{S}}\) is iterated until convergence. The iterations allow for sliding between different universes \({\mathfrak{S}}\) to enable the collapse of model complexity, to refine the diffusion tensor, and to find the solution within the universal set \({\mathfrak{U}}.\) The last step is the AIC model selection between the different diffusion models. Because the AIC criterion approximates the Kullback–Leibler discrepancy which is central to the universal solution in Eq. 1 it was chosen for all three model selection steps over BIC model selection (Schwarz 1978; d’Auvergne and Gooley 2003; Chen et al. 2004). The new protocol avoids the problem of underfitting whereby artificial motions appear (Schurr et al. 1994; Tjandra et al. 1996; Mandel et al. 1996; Luginbühl et al. 1997; Gagné. 1998; d’Auvergne and Gooley 2007b), avoids the problems involved in finding the initial diffusion tensor within \({\mathfrak{D}}\) including the decision of which bond vectors to utilise for the initial analysis using deviations from the average R_{2}/R_{1} ratio and low NOE values (Kay et al. 1989; Clore et al. 1990a; Stone et al. 1992; Barbato et al. 1992; Tjandra et al. 1995a; d’Auvergne and Gooley 2007b), and avoids the problem of hidden internal nanosecond motions and the inability to slide between universes to get to \(\widehat{{\mathfrak{U}}}\) (Orekhov et al. 1995, 1999a, b; Korzhnev et al. 1997; d’Auvergne and Gooley 2007b).
Methods
A new modelfree optimisation protocol
The five diffusion models
Rather than pursuing the elemental idea whereby the universal solution \(\widehat{{\mathfrak{U}}}\) is sought by initially estimating the optimal parameters \(\hat{\theta}_{\mathfrak{D}}\) of the diffusion set \({\mathfrak{D}}\) and then using these estimates to determine the optimal parameter values \(\hat{\theta}_{\mathfrak{F}}\) and models \({\mathfrak{F}}\) of the modelfree dynamics of the molecule (d’Auvergne and Gooley 2007b), the universal solution \(\widehat{{\mathfrak{U}}}\) can also be found by applying the reverse of this logic. Initially the modelfree parameter values \(\hat{\theta}_{\mathfrak{F}}\) and models \({\mathfrak{F}}\) can be determined by optimisation and model selection respectively. Finally, the parameters \(\hat{\theta}_{\mathfrak{D}}\) of the diffusion tensor \({\mathfrak{D}}\) can be optimised. To find the universal solution \(\widehat{{\mathfrak{U}}}\) five categories of global model \({\mathfrak{S}}\) are constructed
where l is the total number of spin systems used in the analysis and \({\mathfrak{F}}_i\) is one of the modelfree models m0 to m9 for spin system i.
Model I (MI)—local τ_{m}
The value of the local τ_{ m } is dependent on the geometry of the true diffusion tensor and the orientation of the XH bond vector (Barbato et al. 1992; Schurr et al. 1994). The MI diffusion model encompasses all the modelfree models and not simply the single tm2 model which was used in Barbato et al. (1992) to study protein interdomain motions, in Schurr et al. (1994) to avoid artificial nanosecond motions when diffusion anisotropy is not taken into account, and in Bruschweiler et al. (1995) to determine the ellipsoidal diffusion tensor.
Although the introduction of model MI significantly increases the number of universes \({\mathfrak{S}} \in {\mathfrak{K}},\) where originally \({\mathfrak{K}} = \{{\mathfrak{S}}_1, {\mathfrak{S}}_2, \ldots, {\mathfrak{S}}_{n \cdot m^l}\},\ n\) is the number of Brownian rotational diffusion models, m is the number of modelfree models, and l is the number of spin systems, for the subset MI \(\subset {\mathfrak{U}}\) a complete collapse of the complexity of the global problem occurs. As no global parameters exist in these models the space \({\mathfrak{S}}\) can be broken into l independent components or spaces \({\mathfrak{T}}_i = {\mathfrak{D}}_i \cup {\mathfrak{F}}_i\) where i is spin system number. The spaces \({\mathfrak{T}}\) are synonymous with modelfree models tm0 to tm9 defined in Paper I. The complexity reduces to \(\dim {\mathfrak{T}} = 1 + k\, \leqslant \,6,\) where 1 represents the single local τ_{ m } parameter and k is the number of modelfree parameters. Due to this dimensionality collecting six relaxation data sets at a minimum of two field strengths is essential. This drastic dissolution of complexity is key to solving the chickenandegg problem of the dual optimisation of the diffusion tensor and the modelfree models.
To find the solution in MI, defined as the space \({\mathfrak{S}}\) which minimises Δ_{ K–L } in Eq. 1 solely for the subset MI \(\subset {\mathfrak{U}},\) three simple steps are required. Firstly and separately for each spin system the parameters of modelfree models tm0 to tm9 are optimised using Newton minimisation as described in Paper I. Failed models are then eliminated as described in d’Auvergne and Gooley (2006). The last step is to select between models tm0 to tm9 using AIC model selection to minimise the value of Δ_{ K–L } (d’Auvergne and Gooley 2003).
Model II (MII)—the sphere
This subset of models represents the diffusion as a sphere, or isotropic diffusion. The initial stage of optimisation involves setting the modelfree models to those of MI but with the local τ_{ m } parameter removed. The modelfree parameter values, taken from MI, are then held constant while the single global diffusion parameter τ_{ m } is optimised.
The space \({\mathfrak{S}}\) which has now been isolated, although very close to the solution of Eq. 1 for the subset MII, may not actually be the space which minimises Δ_{ K–L } due to the approximate nature of model MI. Therefore a repetitive procedure, similar to the standard iterative methodology of the diffusion seeded modelfree paradigm, is necessary to slide between universes \({\mathfrak{S}}\) to find the solution within the MII subset of \({\mathfrak{U}}.\) By holding the optimised diffusion parameters constant modelfree models m0 to m9 can be optimised. Failed models are then eliminated and the best model is selected using AIC model selection. Finally all diffusion and modelfree parameters of the isolated space \({\mathfrak{S}}\) are optimised simultaneously. These steps are repeated until convergence—defined as identical modelfree models (\({\mathfrak{S}}_i \equiv {\mathfrak{S}}_{i1},\)) equal modelfree and diffusion parameter values (\(\theta_i = \theta_{i1} = \hat\theta,\)) and equal chisquared values between iterations (χ^{2} _{ i } = χ^{2} _{ i−1}).
Model III (MIII)—the prolate spheroid
This subset represents the axially symmetric diffusion of the prolate spheroid. The procedure for optimising this model is the same as for MII except that the diffusion set \({\mathfrak{D}}\) = {\({\mathfrak{D}}_{iso}, {\mathfrak{D}}_a, \theta, \phi\) } is minimised. In addition, the constraint \({\mathfrak{D}}_a\, \geqslant \,0\) is implemented to isolate the prolate spheroid subspace.
Model IV (MIV)—the oblate spheroid
This subset also represents axially symmetric diffusion but of the oblate spheroid. The technique is again the same as for MII except that the diffusion set \({\mathfrak{D}}\) = {\({\mathfrak{D}}_{iso}, {\mathfrak{D}}_a, \theta, \phi\) } is minimised together with the constraint \({\mathfrak{D}}_a \,\leqslant \,0\) to isolate the oblate spheroid subspace.
Model V (MV)—the ellipsoid
This subset represents the rhombic or fully anisotropic diffusion of the ellipsoid. Applying the methodology used in MII, although using the diffusion set \({\mathfrak{D}}\) = {\({\mathfrak{D}}_{iso}, {\mathfrak{D}}_a, {\mathfrak{D}}_r, \alpha, \beta, \gamma\) }, the solution for this subset MV \(\subset {\mathfrak{U}}\) can be found.
The universal solution \(\widehat{{\mathfrak{U}}}\)
Once all the global diffusion models have converged to satisfy Eq. 1 for their respective subsets of \({\mathfrak{U}}\) the universal solution \(\widehat{{\mathfrak{U}}}\) can be found by selecting between these global models using AIC model selection. If any of the models MI to MV have failed with diffusional correlation times shooting towards infinity or diffusion rates of zero these should be removed prior to model selection (d’Auvergne and Gooley 2006). Finally the parameter errors can be calculated by Monte Carlo simulation. The entirety of the new modelfree optimisation protocol has been written into a single self contained relax script which is packaged with the program.
All optimisations of the modelfree parameters, the diffusion parameters, or both sets simultaneously utilised the Newton line search algorithm combined with the backtracking step length selection technique (Nocedal and Wright 1999) and the GMW Hessian modification (Gill et al. 1981). The iterative Augmented Lagrangian algorithm was used to constrain the parameter values (Nocedal and Wright 1999). These techniques were investigated in Paper I.
Replication and extension of Schurr’s data
Due to truncation artefacts of using the R_{1}, R_{2}, and NOE values in Table 4 of Schurr et al. (1994) the relaxation data was regenerated from scratch. A PDB file of 12 NH bond vectors with the direction cosines between the NH bond vectors and the major axis of the prolate spheroid, \(\delta_z = \hat{\mu}(t) \cdot \widehat{{\mathfrak{D}}_\} = \cos\epsilon,\) set to {1.00, 0.95, 0.85, 0.75, 0.65, 0.55, 0.45, 0.35, 0.25, 0.15, 0.05, 0.00} was created. Using the program relax relaxation data was generated for a prolate spheroid diffusion tensor with τ_{ m } = 8.5 ns and \({\mathfrak{D}}_{ratio}\) = 1.3. Only dipolar relaxation was assumed as in Schurr et al. (1994). The bond length was not specified (ibid.) therefore a value of 1.02 Å was assumed. Modelfree model m2 was chosen with S ^{2} = 0.8 and τ_{ e } = 50 ps. To use the new global optimisation protocol both 500 and 600 MHz data was generated. As a nonstandard chisquared statistic was used for minimisation (ibid.) errors needed to be generated so that the standard chisquared formula could be used. To best reflect experimental errors values of 0.04 and 0.05 were used for the 600 and 500 MHz NOE respectively whereas 2% errors were used for all other data (d’Auvergne and Gooley 2003).
Dynamics of the bacteriorhodopsin fragment (136)BR
The R_{1}, R_{2}, and NOE relaxation data at 500, 600, and 750 MHz of the bacteriorhodopsin fragment (136)BR was extracted from the comments inside the PostScript file of the relaxation data figure from Orekhov et al. (1999a). For all optimisations a CSA value of −170 ppm and a bond length of 1.02 Å was used. All residues were included in the optimisation of the diffusion model MI. For the optimisation of the spherical diffusion tensor in model MII only residues 9 to 31 were selected.
The Olfactory Marker Protein
The R_{1}, R_{2}, and NOE values at both 600 and 800 MHz were taken from the supporting information. To mirror the original analysis values of −160 ppm and 1.02 Å were used for the CSA and amide NH bond length respectively. As the high precision NMR structures, refined using residual dipolar couplings (Wright et al. 2005), which were used in (Gitti et al. 2005) were not yet available from the PDB, the reanalysis of the relaxation data was carried out against the first model of the original NMR structure 1JYT of Baldisseri et al. (2002) as well as the 2.3 Å resolution Xray crystallographic structure 1F35 of Smith et al. (2002).
Results and discussion
Three test systems
To test the new modelfree optimisation protocol three test systems were examined. These include the data of Schurr et al. (1994) which explored the effect of NH bond vector orientations within the diffusion tensor frame when a too simplistic diffusion tensor is utilised; the bacteriorhodopsin fragment (136)BR data of Orekhov et al. (1999a) in which all residues experience nanosecond timescale motions; and the Olfactory Marker Protein data of Gitti et al. (2005) as a test case of a typical globular protein.
Artifacts induced by ignoring parsimony when selecting the diffusion model
Underfitting
If the selected diffusion tensor is too simplistic then underfitting occurs causing artefacts to appear in the dynamic description (Schurr et al. 1994; Tjandra et al. 1996). These artefacts are the manifestation of the bias introduced by not observing parsimony. When the Brownian diffusion of a molecule is that of a prolate spheroid and the internal motions are fast (assuming model m2), Schurr et al. (1994) demonstrated that the use of a spherical tensor together with the extended modelfree formalism (using model m5) induces artificial subnanosecond timescale motions. This is best demonstrated in Table 4 (ibid.) which has been recalculated in Table S1 of the supplementary material.
To illustrate the second effect, revealed by Tjandra et al. (1996) whereby artificial R _{ ex } contributions appear across the protein, model m4 was minimised against the same data (Table S1). Again the spherical approximation of the diffusion tensor was utilised to force underfitting. Comparing models m4 and m5 in Table S1 the diametrically opposing effects of the underfitting of the two models are evident. Whereas the artificially slow subnanosecond motions appear perpendicular to the major axis of the prolate spheroid, the fictitious chemical exchange occurs when the bond vector is parallel to the major axis.
Occam’s razor
Using the new modelfree optimisation protocol the tm2 model was chosen for all bond vectors when solving for the first step of the procedure, model MI. The S ^{2} and τ_{ e } values replicate the original internal motions whereas the local τ_{ m } parameter over and under estimates the isotropic correlation time as the bond vector changes from parallel to perpendicular to the unique axis \(\widehat{{\mathfrak{D}}_\}.\) Despite the triple exponential form of the rotational correlation function of the Brownian diffusion of a spheroid the single exponential of the local τ_{ m } parameter adequately compensates. Table S2 of the supplementary material summarises the five global models (MI to MV) showing the total number of parameters, the global chisquared value, and the AIC criteria. The AIC value of the oblate spheroid is very close to that of the prolate spheroid but even if this global model is used, the S ^{2} and τ_{ e } values are replicated to within 0.2% and 1% respectively (data not shown). Nevertheless the true prolate spheroid with model m2 used to create the data of Table 4 of Schurr et al. (1994) is easily isolated at the end with all parameters refound to within machine precision. Thus, when using the new modelfree optimisation protocol, both underfitting and overfitting are avoided and the principle of parsimony is closely adhered to.
Overfitting
When too many parameters are included within the global model overfitting occurs. This situation does not introduce bias and hence artifacts in the dynamics. If overly complex diffusion tensors are selected, and minimised properly, the diffusion parameters will take the values of the simpler, true model with the additional geometric parameters of \({\mathfrak{G}}\) being statistically zero and the additional orientational parameters of \({\mathfrak{O}}\) being undefined. As Schurr’s data was noisefree this occurred for the ellipsoid diffusion tensor. A similar situation occurs if an overly complex modelfree model is selected whereby the additional parameters take values which are insignificant. No statistically significant artefacts will appear if the diffusion tensor is overfit, the worst consequence being the inclusion of additional noise into the model. Avoiding both under and overfitting is purely the balancing of bias against variance (d’Auvergne and Gooley 2003).
Bacteriorhodopsin fragment (136)BR—testing the new optimisation protocol
Violation of the rigidity assumption
One of the major causes of failure of the diffusion seeded modelfree protocol is the violation of the rigidity assumption. When the majority of the bond vectors of a molecule experience motions on the nanosecond timescale, local optimisation together with model selection combine to hide the slow motions and steer the final solution far from \(\widehat{{\mathfrak{U}}}.\) An excellent test case representing a molecule in which the diffusion seeded modelfree paradigm fails as all residues exhibit motions on the nanosecond timescale is the bacteriorhodopsin fragment (136)BR (Orekhov et al. 1999a). Applying the concept of estimating an initial diffusion tensor and using this as a starting point for modelfree analysis causes the global correlation time to be underestimated. Subsequent minimisation of the modelfree models to this global model will then hide the internal nanosecond motions (Korzhnev et al. 2001).
Avoiding the initial diffusion tensor estimate
In Orekhov et al. (1999a) a novel protocol was presented for avoiding the rigidity assumption and the need for an initial estimate diffusion tensor. Using this procedure, the global correlation time τ_{ m } was found to be 5.77 ns and the average modelfree parameter values were \(\overline{S^2_f} = 0.84,\ \overline{S^2_s} = 0.61,\) and \(\overline{\tau_s} = 2.9\) ns. As the minimised chisquared value was 120 and the number of parameters k was 66, the AIC value for this model is 252. To test the robustness of the new protocol in avoiding the hidden motion problem, the relaxation data of (136)BR was reanalysed. The final global models from Orekhov et al. (1999a) and that of the new modelfree optimisation protocol are very similar. In fact, the parameters of the former are a subset of the latter. In addition to all residues having the parameters S ^{2} _{ f }, S ^{2} _{ s }, and τ_{ s } the new protocol adds the parameter τ_{ f } to the termini of the αhelix (residues 9, 10, 11, 12, 15, and 31) as well as the parameter R _{ ex } to residues 10 and 31. The averages of the common parameters have shifted to \(\overline{S^2_f} = 0.82, \overline{S^2_s} = 0.51,\) and \(\overline{\tau_s} = 3.8\) ns. In comparison with the AIC value of 252 for the isotropic model with all residues set to m5 (ibid.), the model of higher complexity determined by the new protocol is in fact more parsimonious (AIC = 238.09).
Reanalysis of the OMP relaxation data
To demonstrate the utility of the program relax and the application and consequences of new modelfree optimisation protocol the NMR relaxation data of the Olfactory Marker Protein (OMP) from the original analysis of Gitti et al. (2005) has been reanalysed. This system was chosen as it was a recent analysis of the modelfree dynamics of a protein system in which a number of the issues associated with the application of the diffusion seeded modelfree paradigm are evident.
Global model MI—local τ_{m}
The local τ_{ m } values of model MI are shown in Fig. 1. The trend of the values is similar to the R_{2}/R_{1} ratio plot in Figure 2 of Gitti et al. (2005). Interestingly, the number of residues experiencing chemical exchange in this model is significantly lower than what was reported (ibid.). The chemical exchange is restricted to residues {26, 38, 44, 45, 46, 140} with values of {2.8±1.7, 6.6±0.7, 4.1±2.1, 1.4±0.9, 3.4±1.9, 3.4±1.4} respectively. The majority of the chemical exchange originally reported for residues 20 to 35 (helix α1) is not present and the entirety of the R _{ ex } values across residues 84 to 99 (Ωloop 3) and residues 145 to 152 (βhairpin loop 4) is also absent. Overlapping with this absence is an elevation of the local τ_{ m } parameter in the three distinct yet spatially proximal regions of residues 19 to 50 (helix α1 and loop 1), 83 to 99, and 145 to 155.
Iterative optimisation of global models MII to MV—finding the universal solution \(\widehat{{\mathfrak{U}}}\)
To slide from the initial position given by model MI to that of the universal solution, multiple iterations of optimising global models MII to MV are necessary (Fig. 2). Surprisingly, when sliding between different universes \({\mathfrak{S}}\) en route to convergence the chisquared value actually increases. For different macromolecules this is not always the case—during the optimisation of the bacteriorhodopsin (136)BR fragment the value decreased. This apparent inconsistency can simply be explained through the formulation of the universal solution in (1). Although each iteration minimises the chisquared value, by contrast the overall iterative procedure minimises Δ_{ K–L }. The AIC plot in Fig. 2 demonstrates the decrease of the discrepancy across iterations. Since AIC = χ^{2} + 2k (d’Auvergne and Gooley 2003) the increase in the chisquared values of OMP is offset by a large decrease in the number of model parameters k. In total all calculations using the OMP relaxation data required less than one week of computation on a dual processor, dual core Intel Xeon 2.8 GHz machine using the program relax.
The OMP diffusion tensor—comparison of the NMR and Xray structures
Two OMP structures were available from the Protein Data Bank (PDB) for the reanalysis of the OMP relaxation data. The optimisation and model statistics postconvergence of the first model of the NMR structure 1JYT and the higher quality Xray crystallographic structure 1F35 are presented in Tables S3 and S4 of the supplementary material respectively. When the two structures are directly compared through the AIC values of their optimal global models, the structural information is included in the mathematical model together with the diffusion tensor and modelfree parameters of all residues. As such the discrepancy Δ_{ K–L } as reflected through the AIC values deems the diffusion tensor of the Xray structure to be a better description of the NMR relaxation data. The significance of this result is that the OMP relaxation data of Gitti et al. (2005) implies that the backbone NH bond orientations of the Xray structure 1F35 are more accurate than those of the first model of the NMR structure 1JYT.
In Gitti et al. (2005), where the precise RDC refined NMR structures were used, the molecule was concluded to diffuse as a prolate spheroid. The shape of this tensor differs significantly from the prolate spheroid selected in the reanalysis reported here as the original geometric parameters are \(\widehat{\theta_{\mathfrak{G}}}\) = {τ_{ m }: 8.93 ns; \({\mathfrak{D}}_a\): 3.5e^{7} s^{−1}} whereas those of the reanalysis are \(\widehat{\theta_{\mathfrak{G}}}\) = {τ_{ m }: 9.09 ns; \({\mathfrak{D}}_a\): 7.13e^{7} s^{−1}}. If the geometric parameter \({\mathfrak{D}}_{ratio}\) is compared, the original and new values are 1.2 and 1.45 respectively. The diffusion tensor of the universal solution, the prolate spheroid using the 1F35 structure, together with the results of 200 Monte Carlo simulations are presented in Fig. S1. The reason for the greater anisotropy in the reanalysis is explained below.
Creation of a hybrid model
In model MI, four regions of the protein were identified from Fig. 1 as having elevated local τ_{ m } values—helix α1, loop 1, Ωloop 3, and βhairpin loop 4. Significantly these regions of model MI do not demonstrate the extensive chemical exchange contributions present in the original results. Therefore to entertain the possibilities that either these regions experience a slower correlation time than the core of the protein or that the orientations of their backbone NH bond vectors are systematically inaccurate, a hybrid model was constructed whereby the core of the protein was treated separately from the four structural elements. Residues 19–50, 83–99, and 145–155 were excluded and the new modelfree optimisation protocol reapplied to the protein core using the Xray structure. The universal solution using this subset of residues was again a prolate spheroid. Interestingly the diffusion tensor geometry, \(\widehat{\theta_{\mathfrak{G}}}\) = {τ_{ m }: 8.95 ns; \({\mathfrak{D}}_a\): 3.4e^{7} s^{−1}}, is very similar to that of the original results.
In the three loops and helix α1 each residue was assumed to tumble independently, each having its own local τ_{ m } parameter, hence global model MI was used. Subsequently two data sets were loaded into and hybridised within relax: one being the universal solution for the core of the protein whereby the loops have been excluded, the other being model MI applied solely to the loops. As the number of residues and relaxation data sets were identical between the hybrid model and the solution found when the protein is treated as a single unit, AIC model selection is able to choose between the two. For the hybrid the optimisation and model statistics were k = 310, χ^{2} = 227.4, and AIC = 847.4. In comparison the prolate spheroid statistics were k = 294, χ^{2} = 252.8, and AIC = 840.8. Hence, despite the chisquared value of the hybrid being significantly lower than that of the prolate spheroid the hybridisation does not improve parsimony. Although this does not enhance the OMP dynamics description, within other systems such as multidomain proteins treating various components of the system separately and then hybridising each individual component can significantly improve the dynamic description (Horne et al. 2007).
OMP dynamics
The internal modelfree motions
The solution to the modelfree problem, as defined in Eq. 1 and when comparing the two structures, is the prolate spheroid for the 1F35 Xray structure. The final and complete modelfree results from this global diffusion model are presented in Table S5 of the supplementary material. For comparison with the original results of Gitti et al. (2005) both sets of parameter values are plotted in Fig. S2 and superimposed onto the OMP Xray structure in both Figs. S2 and S3. Large differences in LipariSzabo order parameters, effective correlation times, and the R _{ ex } parameter are clearly demonstrated in the three figures.
Amplitudes of the internal motions
A number of discrepancies between the original S ^{2} values and the reanalysis exist across the protein (Figs. S2a, 3b, and 3c). The greatest anomaly, which will be discussed below, occurs within residues 20–34 of helix α1. In addition both the Nterminus and residues 39–41 of loop 1 are more mobile in the reanalysis whereas the βhairpin loop 4 is more restricted. Although not statistically significant on a per residue basis, systematic increases or decreases in mobility of distinct secondary structural elements has occurred. For instance all residues of helix α2 are slightly more mobile in the reanalysis. The validity of the new order parameters are strongly supported by the NMR relaxation data—many of the trends present in the R_{1}, R_{2}, and NOE values shown in Figure 2 of Gitti et al. (2005) are combined and reflected in the new amplitudes of motion.
Rigidity of helix α1
The most striking difference between the new and the old analysis, as illustrated by Figs. S2 and 3, is the rigidity of the helix α1. In the original analysis (ibid.) helix α1 was one of the most mobile regions of the protein yet in the new analysis the helix is the most rigid secondary structure element in the protein. This rigidity is strongly supported by the original NOE values. Not only are there significant differences in the internal motions on the picosecond to nanosecond timescales (Figs. S2a, 3b, and c) but large quantities of chemical exchange which were present in the original results are absent from the reanalysis (Figs. S2c, 3d, and e). Although the R_{2} values of α1 are elevated above the protein average and appear to support the presence of chemical exchange the elevation is in fact caused by the geometry of the diffusion tensor. The maximum correlation time of a vector attached to a prolate tensor is when it is parallel to the long axis which, in the case of the reanalysis, is approximately 10.5 ns. The local τ_{ m } values of α1 are very close to this number (Fig. 1). As was demonstrated in Table S1 and in Tjandra et al. (1996) underestimation of the global correlation time experienced by a bond vector induces artificial R _{ ex } values to appear. Notably helix α1 is parallel to the major axis of the prolate diffusion tensor (Fig. S1) hence the halving of the anisotropy will result in the underestimation of the correlation times.
The reason for the underestimation of the anisotropy of rotational diffusion in the original analysis relates to the NH bond vector distribution. A number of empirical rules were used to exclude residues from the initial tensor estimate including the low NOE rule (Kay et al. 1989; Stone et al. 1992; Barbato et al. 1992), deviations from the R_{2}/R_{1} ratio (Clore et al. 1990a; Barbato et al. 1992; Tjandra et al. 1995b), and utilising solely residues within distinct secondary structure elements (Habazettl et al. 1996; Dosset et al. 2000). The consequence of implementing these commonly used exclusion rules for OMP is evident in Fig. 4—almost all residues perpendicular to the unique axis of the diffusion tensor have been removed from the analysis. Hence there is a paucity of information concerning the \({\mathfrak{D}}_\\) eigenvalue within the limited subset of the relaxation data and extracting the true and full anisotropy of the tensor is not possible. The result is the appearance of artificial chemical exchange.
The new modelfree optimisation protocol solves this issue by using all the available relaxation data for determining the diffusion tensor. No rules are used for excluding spin systems. As can be seen in Fig. 4c and d the coverage of space by the OMP amide NH bond distribution is more even and much denser. Importantly a large number of vectors sample the space parallel to the unique axis of the diffusion tensor. Hence information about all components of the diffusion tensor are adequately contained within the full set of relaxation data.
The correlation between structural quality and artificial motions
When the diffusion of the macromolecule under study is anisotropic, the accuracy of the modelfree results is dependent upon the quality of the structure underlying the analysis. For a perfectly spherical probability distribution of vectors centred at the origin, the projection of the vectors onto the major axis of a spheroid will form a sinusoidal probability distribution. This distribution has zero probabilities at the poles and a maximal probability at the equator. If the orientation of an arbitrary vector attached to the molecule is slightly randomised with equal probability in all directions the mean projection of many randomisations will shift towards the equator. The projectional bias, which is purely a geometric phenomenon, has important consequences for the modelfree analysis of nonspherical proteins and can have two opposing effects. If the molecule diffuses as a prolate spheroid the bias will be away from the unique, long axis causing a mean underestimation of the effective global correlation time and hence favour artificial R _{ ex } values over artificial nanosecond motions. If the molecule diffuses as an oblate spheroid the bias will be away from the unique, short axis of the tensor. The result will be a mean overestimation of the effective global correlation time and therefore artificial nanosecond motions are favoured.
The R_{ex} values of OMP
Although loop 1, Ωloop 3, and βhairpin loop 4 all show significant chemical exchange in both Gitti et al. (2005) and the reanalysis, both of which chose the prolate spheroid, the scarce appearance of R _{ ex } contributions in global model MI may be an indication that the R _{ ex } values do not correspond to real chemical exchange. In the reanalysis where the Xray crystallographic structure 1F35 was employed the residues in which R _{ ex } values appear (Fig. 3e) are all located in regions which vary significantly between the different PDB structures, as demonstrated by Figure 5 (ibid.). Because the molecule diffuses as a prolate spheroid inaccuracy in these regions of the protein will bias the modelfree analysis favouring the appearance of artificial R _{ ex } values. As all R _{ ex } values in the OMP reanalysis could in fact be explained by imprecise NH backbone bond orientations either reanalysis using the RDC refined OMP structure (Wright et al. 2005) or relaxation dispersion experiments could be used to prove the presence of true chemical exchange. Alternatively the R _{ ex } contribution to the R_{2} relaxation rate could be eliminated prior to modelfree analysis (Farrow et al. 1995; Phan et al. 1996; Kroenke et al. 1999; Butterwick et al. 2004).
Failure of the diffusion seeded paradigm
The reason for the artificial R _{ ex } values of helix α1 was identified as a failure of the diffusion seeded modelfree paradigm rather than an optimisation, model selection, or model failure issue. By taking the diffusion parameters of the prolate core of the hybrid model (vide supra) as a starting point for modelfree analysis, the diffusion seeded protocol was employed within relax. The prolate spheroid was chosen by AIC model selection. Convergence of this model occurred after six iterations and the final geometric parameters were \(\widehat{\theta_{\mathfrak{G}}}\) = {τ_{ m }: 9.00 ns; \({\mathfrak{D}}_a\): 4.5e^{7} s^{−1}}. Sliding between universes to reach the universal solution \(\widehat{{\mathfrak{U}}}\) did not occur and the artificial motions of the protein were still present. Finding the solution was only possible using either the new modelfree optimisation protocol or that of Orekhov et al. (1999a).
The internal correlation times
Another major difference between the original results and the reanalysis, as demonstrated in Figs. S2b and S3, is the internal modelfree correlation times. Originally only 42 correlation times were extracted whereas in the reanalysis 102 correlation times were selected, the additional correlation times spanning from 10 ps to well into the nanosecond range. The differences are primarily due to the more parsimonious AIC model selection. In the original analysis the ANOVA stepup hypothesis testing model selection which is coded into the FASTmodelfree interface (Cole and Loria 2003) to the Modelfree program (Palmer et al. 1991; Mandel et al. 1995) and based on the stepup methodology of Mandel et al. (1995) was employed. A significant patch of nanosecond motions occurs on the βhairpin loop 4 side of the β clam fold. However as these motions are not present in global model MI (data not shown) and the loop positions are quite variable between the Xray and two NMR structures (Baldisseri et al. 2002; Smith et al. 2002; Wright et al. 2005), these slow nanosecond motions may be artificial (Schurr et al. 1994).
Parameter uncertainties
In Fig. S2 it is evident that the parameter errors in the reanalysis are greater than those of the original results. This is due to two factors: the effects of underfitting and the higher precision optimisation coupled with Monte Carlo simulations. As more parameters are utilised in the reanalysis, greater amounts of noise from the collected relaxation data are transferred into the model (d’Auvergne and Gooley 2003). The deliberate underfitting of the ANOVA stepup model selection (Mandel et al. 1995) of the original analysis not only skews the dynamic picture but also results in an underestimation of the parameter uncertainties. Higher precision optimisation also results in greater, yet real, parameter uncertainties. The modelfree parameter errors are determined via Monte Carlo simulation whereby each simulation is minimised using the same optimisation algorithms as the original data. The initial position for MC simulations is set to the optimised modelfree parameter values hence if optimisation terminates early due to low precision or other issues (Paper I) then the affected simulation does not move as far away from the mean as it should. The result is that the parameter errors are underestimated.
Conclusion
The diffusion seeded modelfree paradigm of using an initial estimate of the diffusion tensor has been used in most modelfree analyses presented in the literature. There are, however, a number of problems associated with the approach (d’Auvergne and Gooley 2007b). To avoid these this paper presents a new modelfree optimisation protocol which completely reverses the logic of the diffusion seeded modelfree paradigm. Rather than starting with the diffusion tensor the protocol begins by optimising the modelfree models free of any global diffusion parameters. This is done by constructing the global model MI in which each bond vector has a local τ_{ m } parameter. Modelfree models tm0 to tm9 are optimised and the best model selected. In the next step of the protocol the local τ_{ m } parameter is removed from the models, the modelfree parameters are held fixed, and the spherical diffusion tensor (global model MII), prolate spheroid (MIII), oblate spheroid (MIV), and ellipsoid (MV) parameters are optimised. Iterative steps of optimisation of models m0 to m9 with the diffusion parameters fixed, model elimination, AIC model selection, and then optimisation of all spin systems are performed until convergence. This protocol is designed for robustly finding the universal solution \(\widehat{{\mathfrak{U}}},\) defined in Eq. 1. By using the synthetic data from Schurr et al. (1994) and the bacteriorhodopsin fragment (136)BR data (Orekhov et al. 1999a) the new protocol is shown to avoid all of the problems associated with modelfree analysis. These include artificial nanosecond motions (Schurr et al. 1994), artificial chemical exchange (Tjandra et al. 1996), two minima of spheroidal parameter space (Paper I), and violation of the rigidity assumption and hiding of nanosecond motions.
In using AIC model selection to choose between the modelfree models as well as the diffusion tensors (d’Auvergne and Gooley 2003); implementing model elimination to remove failed models (d’Auvergne and Gooley 2006); employing Newton optimisation together with the backtracking line search (Nocedal and Wright 1999) and Gill, Murray, Wright Hessian modification (Gill et al. 1981) and constraining the parameters with the Augmented Lagrangian algorithm (Nocedal and Wright 1999; d’Auvergne and Gooley 2007a); minimising the numerically stabilised modelfree equations (d’Auvergne and Gooley 2007a); and utilising the new modelfree optimisation protocol to find the universal solution \(\widehat{{\mathfrak{U}}},\) a significantly improved and refined picture of the dynamics of a macromolecule can be obtained.
Abbreviations
 AIC:

Akaike’s Information Criteria
 ANOVA:

Analysis of variance
 BIC:

Schwarz or Bayesian Information Criteria
 CSA:

Chemical Shift Anisotropy
 Δ_{ K–L } :

Kullback–Leibler discrepancy
 \({\mathfrak{D}}\) :

Set of diffusion tensor parameters
 \({\mathfrak{F}}_i\) :

Set of modelfree parameters for a single spin system
 \({\mathfrak{G}}\) :

Set of geometric diffusion parameters
 GMW:

Gill, Murray, and Wright Hessian modification
 \({\mathfrak{K}}\) :

Set of all global models \({\mathfrak{S}}\)
 MC:

Monte Carlo
 \({\mathfrak{O}}\) :

Set of orientational diffusion parameters
 OMP:

Olfactory Marker Protein
 \({\mathfrak{S}}\) :

The global model, space, or universe
 \({\mathfrak{T}}_i\) :

Set of modelfree parameters and local τ_{ m } for a single spin system
 \({\mathfrak{U}}\) :

Universal set
 \(\widehat{{\mathfrak{U}}}\) :

Universal solution
 XH bond:

Heteronucleusproton bond
References
Akaike H (1973) Information theory and an extension of the maximum likelihood principle. In: Petrov BN, Csaki F (eds) Proceedings of the second international symposium on information theory. Budapest, Akademia Kiado, pp 267–281
Baldisseri DM, Margolis JW, Weber DJ, Koo JH, Margolis FL (2002) Olfactory marker protein (OMP) exhibits a betaclam fold in solution: implications for target peptide interaction and olfactory signal transduction. J Mol Biol 319(3):823–837
Barbato G, Ikura M, Kay LE, Pastor RW, Bax A (1992) Backbone dynamics of calmodulin studied by 15N relaxation using inverse detected twodimensional NMR spectroscopy: the central helix is flexible. Biochemistry 31(23):5269–5278
Brüschweiler R, Liao X, Wright PE (1995) Longrange motional restrictions in a multidomain zincfinger protein from anisotropic tumbling. Science 268(5212):886–889
Burnham KP, Anderson DR (1998) Model selection and inference: a practical informationtheoretic approach. SpringerVerlag, New York
Butterwick JA, Loria PJ, Astrof NS, Kroenke CD, Cole R, Rance M, Palmer AG 3rd (2004) Multiple time scale backbone dynamics of homologous thermophilic and mesophilic ribonuclease HI enzymes. J Mol Biol 339(4):855–871
Chen J, Brooks CL, 3rd, Wright PE (2004) Modelfree analysis of protein dynamics: assessment of accuracy and model selection protocols based on molecular dynamics simulation. J Biomol NMR 29(3):243–257
Clore GM, Driscoll PC, Wingfield PT, Gronenborn AM (1990a) Analysis of the backbone dynamics of interleukin1 beta using twodimensional inverse detected heteronuclear 15N1H NMR spectroscopy. Biochemistry 29(32):7387–7401
Clore GM, Szabo A, Bax A, Kay LE, Driscoll PC, Gronenborn AM (1990b) Deviations from the simple 2parameter modelfree approach to the interpretation of N15 nuclear magneticrelaxation of proteins. J Am Chem Soc 112(12):4989–4991
Cole R, Loria J (2003) FASTModelfree: a program for rapid automated analysis of solution NMR spinrelaxation data. J Biomol NMR 26(3):203–213
d’Auvergne EJ (2006) Protein dynamics: a study of the modelfree analysis of NMR relaxation data. Ph.D. thesis, Biochemistry and Molecular Biology, University of Melbourne. http://eprints.infodiv.unimelb.edu.au/archive/00002799/
d’Auvergne EJ, Gooley PR (2003) The use of model selection in the modelfree analysis of protein dynamics. J Biomol NMR 25(1):25–39
d’Auvergne EJ, Gooley PR (2006) Modelfree model elimination: a new step in the modelfree dynamic analysis of NMR relaxation data. J Biomol NMR 35(2):117–135
d’Auvergne EJ, Gooley PR (2007a) Optimisation of NMR dynamic models I. Minimisation algorithms and their performance within the modelfree and Brownian rotational diffusion spaces. J Biomol NMR. doi:10.1007/s1085800792142
d’Auvergne EJ, Gooley PR (2007b) Set theory formulation of the modelfree problem and the diffusion seeded modelfree paradigm. Mol BioSyst 3(7):483–494
DeLano WL (2002) The PyMOL molecular graphics system. http://www.pymol.org
Dosset P, Hus JC, Blackledge M, Marion D (2000) Efficient analysis of macromolecular rotational diffusion from heteronuclear relaxation data. J Biomol NMR 16(1):23–28
Farrow NA, Zhang OW, Szabo A, Torchia DA, Kay LE (1995) Spectral densityfunction mapping using N15 relaxation data exclusively. J Biomol NMR 6(2):153–162
Gagné SM, Tsuda S, Spyracopoulos L, Kay LE, Sykes BD (1998) Backbone and methyl dynamics of the regulatory domain of troponin C: anisotropic rotational diffusion and contribution of conformational entropy to calcium affinity. J Mol Biol 278(3):667–686
Gill PE, Murray W, Wright MH (1981) Practical optimization. Academic Press
Gitti RK, Wright NT, Margolis JW, Varney KM, Weber DJ, Margolis FL (2005) Backbone dynamics of the olfactory marker protein as studied by 15N NMR relaxation measurements. Biochemistry 44(28):9673–9679
Habazettl J, Myers LC, Yuan F, Verdine GL, Wagner G (1996) Backbone dynamics, amide hydrogen exchange, and resonance assignments of the DNA methylphosphotriester repair domain of Escherichia coli Ada using NMR. Biochemistry 35(29):9335–9348
Horne J, d’Auvergne EJ, Coles M, Velkov T, Chin Y, Charman WN, Prankerd R, Gooley PR, Scanlon MJ (2007) Probing the flexibility of the DsbA oxidoreductase from Vibrio cholerae—a 15N–1H heteronuclear NMR relaxation analysis of oxidized and reduced forms of DsbA. J Mol Biol 371(3):703–716
Kay LE, Torchia DA, Bax A (1989) Backbone dynamics of proteins as studied by 15N inverse detected heteronuclear NMR spectroscopy: application to staphylococcal nuclease. Biochemistry 28(23):8972–8979
Korzhnev DM, Billeter M, Arseniev AS, Orekhov VY (2001) NMR studies of Brownian tumbling and internal motions in proteins. Prog NMR Spectrosc 38(3):197–266
Korzhnev DM, Orekhov VY, Arseniev AS (1997) Modelfree approach beyond the borders of its applicability. J Magn Reson 127(2):184–191
Kroenke CD, Rance M, Palmer AG (1999) Variability of the N15 chemical shift anisotropy in Escherichia coli ribonuclease H in solution. J Am Chem Soc 121(43):10119–10125
Kullback S, Leibler RA (1951) On information and sufficiency. Ann Math Stat 22(1):79–86
Lee LK, Rance M, Chazin WJ, Palmer AG (1997) Rotational diffusion anisotropy of proteins from simultaneous analysis of N15 and C13(alpha) nuclear spin relaxation. J Biomol NMR 9(3):287–298
Linhart H, Zucchini W (1986) Model selection, Wiley Series in Probability and mathematical statistics. John Wiley & Sons, Inc New York
Lipari G, Szabo A (1982a) Modelfree approach to the interpretation of nuclear magneticresonance relaxation in macromolecules I. Theory and range of validity. J Am Chem Soc 104(17):4546–4559
Lipari G, Szabo A (1982b) Modelfree approach to the interpretation of nuclear magneticresonance relaxation in macromolecules II. Analysis of experimental results. J Am Chem Soc 104(17):4559–4570
Luginbühl P, Pervushin KV, Iwai H, Wüthrich K (1997) Anisotropic molecular rotational diffusion in 15N spin relaxation studies of protein mobility. Biochemistry 36(24):7305–7312
Mandel AM, Akke M, Palmer AG, 3rd (1995) Backbone dynamics of Escherichia coli ribonuclease HI: correlations with structure and function in an active enzyme. J Mol Biol 246(1):144–163
Mandel AM, Akke M, Palmer AG, 3rd (1996) Dynamics of ribonuclease H: temperature dependence of motions on multiple time scales. Biochemistry 35(50):16009–16023
Nocedal J, Wright SJ (1999) Numerical optimization, Springer Series in Operations research. SpringerVerlag, New York
Orekhov VY, Korzhnev DM, Diercks T, Kessler H, Arseniev AS (1999a) H1N15 NMR dynamic study of an isolated alphahelical peptide (136) bacteriorhodopsin reveals the equilibrium helixcoil transitions. J Biomol NMR 14(4):345–356
Orekhov VY, Korzhnev DM, Pervushin KV, Hoffmann E, Arseniev AS (1999b) Sampling of protein dynamics in nanosecond time scale by 15N NMR relaxation and selfdiffusion measurements. J Biomol Struct Dyn 17(1):157–174
Orekhov VY, Pervushin KV, Korzhnev DM, Arseniev AS (1995) Backbone dynamics of (171)Bacterioopsin and (136)Bacterioopsin studied by 2dimensional H1N15 NMRspectroscopy. J Biomol NMR 6(2):113–122
Palmer AG, Rance M, Wright PE (1991) Intramolecular motions of a zinc finger DNAbinding domain from Xfin characterized by protondetected natural abundance C12 heteronuclear NMRspectroscopy. J Am Chem Soc 113(12):4371–4380
Phan IQH, Boyd J, Campbell ID (1996) Dynamic studies of a fibronectin type I module pair at three frequencies: anisotropic modelling and direct determination of conformational exchange. J Biomol NMR 8(4):369–378
Schurr JM, Babcock HP, Fujimoto BS (1994) A test of the modelfree formulas. Effects of anisotropic rotational diffusion and dimerization. J Magn Reson B 105(3):211–224
Schwarz G (1978) Estimating dimension of a model. Ann Stat 6(2):461–464
Smith PC, Firestein S, Hunt JF (2002) The crystal structure of the olfactory marker protein at 2.3 A resolution. J Mol Biol 319(3):807–821
Stone MJ, Fairbrother WJ, Palmer AG, Reizer J, Saier MH, Wright PE (1992) Backbone dynamics of the Bacillus subtilis glucose permeaseIIA domain determined from N15 NMR relaxation measurements. Biochemistry 31(18):4394–4406
Tjandra N, Feller SE, Pastor RW, Bax A (1995a) Rotational diffusion anisotropy of human ubiquitin from N15 NMR relaxation. J Am Chem Soc 117(50):12562–12566
Tjandra N, Kuboniwa H, Ren H, Bax A (1995b) Rotational dynamics of calciumfree calmodulin studied by 15NNMR relaxation measurements. Eur J Biochem 230(3):1014–1024
Tjandra N, Wingfield P, Stahl S, Bax A (1996) Anisotropic rotational diffusion of perdeuterated HIV protease from 15 N NMR relaxation measurements at two magnetic fields. J Biomol NMR 8(3):273–284
Woessner DE (1962) Nuclear spin relaxation in ellipsoids undergoing rotational brownian motion. J Chem Phys 37(3):647–654
Wright NT, Margolis JW, Margolis FL, Weber DJ (2005) Refinement of the solution structure of rat olfactory marker protein (OMP). J Biomol NMR 33(1):63–68
Zucchini W (2000) An introduction to model selection. J Math Psychol 44(1):41–61
Acknowledgements
We would like to thank Vladislav Orekhov and Dmitry Korzhnev for kindly donating the relaxation data of the bacteriorhodopsin (136)BR fragment.
Author information
Authors and Affiliations
Corresponding author
Electronic supplementary material
Below is the link to the electronic supplementary material.
Rights and permissions
Open Access This is an open access article distributed under the terms of the Creative Commons Attribution Noncommercial License ( https://creativecommons.org/licenses/bync/2.0 ), which permits any noncommercial use, distribution, and reproduction in any medium, provided the original author(s) and source are credited.
About this article
Cite this article
d’Auvergne, E.J., Gooley, P.R. Optimisation of NMR dynamic models II. A new methodology for the dual optimisation of the modelfree parameters and the Brownian rotational diffusion tensor. J Biomol NMR 40, 121–133 (2008). https://doi.org/10.1007/s1085800792133
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s1085800792133