Correlated Dynamics between Protein HN and HC Bonds Observed by NMR Cross Relaxation

Abstract

Although collective dynamics of atom groups steer many biologically relevant processes in biomacromolecules, most atomic resolution motional studies focus on isolated bonds. In this study, a new method is introduced to assess correlated dynamics between bond vectors by cross relaxation nuclear magnetic resonance (NMR). Dipole-dipole cross correlated relaxation rates between intra and interresidual H^N–N and H^α–C^α in the 56 residue protein GB3 are measured with high accuracy. It is demonstrated that the assumption of anisotropic molecular tumbling is necessary to evaluate rates accurately and predictions from the static structure using effective bond lengths of 1.041 and 1.117 Å for H^N–N and H^α–C^α are within 3 % of both experimental intra and interresidual rates. Deviations are matched to models of different degrees of motional correlation. These models are based on previously determined orientations and motional amplitudes from residual dipolar couplings with high accuracy and precision. Clear evidence of correlated motion in the loops comprising residues 10-14, 20-22 and 47-50, and anticorrelated motion in the α helix comprising 23-38 is presented. Somewhat weaker correlation is observed in the β strands 2-4, which have previously been shown to exhibit slow correlated motional modes.

Introduction

Routine NMR approaches to study molecular dynamics assess motions of isolated H-N or H-C bond vectors. ¹⁵N relaxation experiments aim at time scales of picoseconds to tens of nanoseconds ^1-3. More recent approaches extend the scale up to milliseconds. Relaxation dispersion studies define time scales of chemical and conformational exchange ⁴. Residual dipolar couplings (RDCs) ⁵ report simultaneously on time averages of bond orientations and motional effects ^6-9. However, it is not trivial to convert relaxation rates or RDCs into amplitudes and directionalities of collective dynamics of groups of atoms with which many biologically relevant processes are associated ^10-12. For example, protein backbone plane motions are mostly reflected in fluctuations around the around the φ and ψ backbone angles. Generally, methods to assess relative motions between different bond vectors are required to complete the picture of backbone motion.

In previous studies on protein GB3 ¹³, highly precise intraresidual H^N-H^α J-coupling values have been fitted to Karplus curves to set limits on motions around φ angles ¹⁴. Measurements of H^N-H^α RDCs showed that intraresidual values are virtually the same as predicted for a structure exhibiting minimal motion corresponding to C^α-CO fluctuations, whereas sequential values are 9% smaller ¹⁵. Both approaches suggest fluctuations around the φ angles to be very small. Furthermore, large sets of RDCs and scalar couplings across hydrogen bonds have been used to identify slow correlated motions in the β sheet ¹⁶. Recently, the motional mode distribution has been reproduced by Accelerated Molecular Dynamics (AMD) ¹².

Cross correlated relaxation rates (CCR rates) depend on the relative orientation of two tensorial interactions ¹⁷. In multiple quantum coherences, some of these interactions do not share a common spin and can be located several Ångstroms apart from one another. CCR rates have been proposed for determination of backbone torsion angles a decade ago by the Griesinger laboratory ¹⁸. Subsequently, a wealth of experiments have been designed to determine the ψ angle by means of H^α-C^α(i) dipole/H^N-N(i) dipole ^18-25 or H^α-C^α(i) dipole/CO(i) CSA ^{21, 26-28}, and the φ angle by means of H^α-C^α(i-1) dipole/H^N-N(i) dipole ^{22, 24, 25} or H^α-C^α(i-1) dipole/CO(i) CSA ²⁹ CCR rates. Other approaches defining backbone geometry make use of N CSA/CO CSA ³⁰, H^N-CO dipole/H^N-N dipole ³⁰, CO CSA/CO CSA ³¹ or H^N-CO dipole/CO CSA ³².

It has been shown that CCR rates are also reporters on dynamics ^{33, 34}. Importantly, in contrast to autorelaxation rates, these rates are sensitive to motion on all time scales ^{35, 36}. Multiple quantum experiments have been proposed to assess backbone plane motion ³⁷. Sums of H^N-N dipole/CO-C^α dipole and H^N-C^α dipole/CO-N dipole ³⁸ or H^N-CO dipole/CO CSA ³² CCRs have been interpreted with GAF models ^{39, 40}. H^N-N dipole/ H^N-N dipole revealed correlated motions in Ubiquitin ³⁵. However, interpretation of dynamics is difficult. Similarly to the interpretation of RDCs, vector orientation and motion are intertwined and generally underdetermined ^{7 41}. Sensitivity of CCR rates to anisotropy in rotational diffusion ⁴² complicates the formal description further. A full theoretical description taking into account fast and slow dynamics is still an active field ³⁶.

In the present study, CCR rates in GB3 are used to study relative fluctuations between H^N-N and H^α-C^α bonds, presumably dominated by the φ and ψ backbone angles in GB3. Preselected bond orientations are shown to be approximately correct and deviations from predicted CCR rates are interpreted in terms of motional effects. This approach is opposed to the standard procedure, where dynamics and orientation are fitted simultaneously. The following steps are taken to guarantee highest possible precision and accuracy: i) Data quality: CCR rates are measured exclusively between dipolar interactions (Figure 1). With a very accurate structure at hand, these rates can be more conveniently evaluated than rates involving CSA interactions. Pulse sequences featuring minimal systematic errors, based on evolution of all multiplet components in multiple quantum coherences, are used to measure CCR rates and are compared to sequences producing a high signal-to-noise ratio, based on direct measurements of interconversion between inphase and antiphase multiple quantum coherences. ii) Selection of vector orientations: The interpretation crucially depends on the accuracy of the H^N and H^α proton positions. The orientations of the bond vectors H^N-N and H^α-C^α are taken from a study that employed highly accurate RDCs from multiple alignments ⁹. It is shown that these vectors cross-validate best with highly precise ³J_HNHα scalar couplings and intraresidual and sequential H^N-H^α RDCs. iii) Evaluation of the theoretical framework: Models assuming isotropic, axially symmetric and fully anisotropic tumbling are compared. Examination of the derivation of CCR order parameters is lined out and values are simulated for different models of motional correlation. Finally, the experimental rates are compared to the simulated rates. Observation of correlated motion in the loops comprising residues 10-14, 20-22 and 47-50, and anticorrelated motion in the α helix comprising residues 23-38 is supported by a 13 ns molecular dynamics (MD) simulations. Somewhat weaker correlation in the β strands 2-4 is not substantiated by the MD simulations.

Figure 1.

Cartoon representation of bond vectors used in the present study and motions in two peptide backbone planes. Red atoms form the bond vectors of the dipolar mechanism between which CCRs are measured. Pink circles indicate fast bond fluctuation and the blue arrows represent γ motion about the blue axes in the peptide planes.

Theory

Extraction of cross-correlated relaxation rates

Evolution of multiple quantum coherence, MQ, between spins I¹ and I², where I¹ is weakly scalar coupled to the passive spin S¹ (J_I1S1) and I² to S² (J_I2S2) yields 8 peaks corresponding to the coherence order (zero and double quantum, ZQ and DQ) and the spin states of S¹ and S² (αα, αβ, βα, and ββ). In the secular approximation, magnetization is exchanged between these components by cross-correlated relaxation between 6 mechanisms (Chemical shift anisotropy of I¹ and I², and dipolar interaction between I¹-S¹, I²-S², I¹-S² and I²-S¹) resulting in 15 nonuniform contributions to the relaxation rates of the individual peaks ²⁰. In addition, the nuclear Overhauser effect between S¹ and S² must be considered ^{23, 35}. W₀ contributes to the components in the αβ and βα states, and W₂ to those in the αα and ββ states. The cross-correlated relaxation rate of interest, R_{d(I1S1)/d(I2S2)}, can only be extracted together with other terms from peak intensities I of the components of a ZQ or DQ quadruplet:

$$\frac{1}{4T_{\text{MQ}}}ln\left(\frac{I_{\alpha \beta }^{\text{ZQ}}{I}_{\beta \alpha }^{\text{ZQ}}}{I_{\alpha \alpha }^{\text{ZQ}}{I}_{\beta \beta }^{\text{ZQ}}}\right)=R_{\text{d}(\text{I}1\text{S}1)/\text{d}(\text{I}2\text{S}2)}+R_{\text{d}(\text{I}1\text{S}1)/\text{d}(\text{I}2\text{S}1)}+R_{\text{d}(\text{I}1\text{S}2)/\text{d}(\text{I}2\text{S}2)}+R_{\text{d}(\text{I}1\text{S}2)/\text{d}(\text{I}2\text{S}1)}+\stackrel{W_0}{}/\underset{2}{}-\stackrel{W_2}{}/\underset{2}{}$$ (1.1)

$$\frac{1}{4T_{\text{MQ}}}ln\left(\frac{I_{\alpha \alpha }^{\text{DQ}}{I}_{\beta \beta }^{\text{DQ}}}{I_{\alpha \beta }^{\text{DQ}}{I}_{\beta \alpha }^{\text{DQ}}}\right)=R_{\text{d}(\text{I}1\text{S}1)/\text{d}(\text{I}2\text{S}2)}-R_{\text{d}(\text{I}1\text{S}1)/\text{d}(\text{I}2\text{S}1)}-R_{\text{d}(\text{I}1\text{S}2)/\text{d}(\text{I}2\text{S}2)}+R_{\text{d}(\text{I}1\text{S}2)/\text{d}(\text{I}2\text{S}1)}-\stackrel{W_0}{}/\underset{2}{}+\stackrel{W_2}{}/\underset{2}{}$$ (1.2)

T_MQ is the constant time during which the coherences evolve. Note that in both cases the intensities of the inner components are divided by the outer components (or vice versa, depending on the sign of J_I1S1 and J_I2S2). Unwanted additional contributions are 3 dipole-dipole cross correlated relaxation rates and half of W₀ and W₂. However, all terms except for the one of interest depend inversely on the spatial separation of the I¹-S¹ from the I²-S² vector. Identifying I¹ and I² with N and C^α, and S¹ and S² with the attached protons, R_{d(I1S1)/d(I2S2)} is much larger than the other terms and can be approximately extracted from peak intensities I of the components of a ZQ or DQ quadruplet ²³. In the present application, however, highest possible accuracy and precision are crucial and the adverse impact of the smaller terms is evaluated. Simulations show that for intraresidual coherences R_{d(I1S1)/d(I2S2)}, R_{d(I1S1)/d(I2S1)}, R_{d(I1S2)/d(I2S2)}, and R_{d(I1S2)/d(I2S1)} can be as large as 15, 0.7, 3, and 0.3 s^-1, W₀ is typically ≈0.3 s^-1, whereas W₂ is very small. The contributions to the sequential rates are similar. Clearly, the unwanted terms cannot be neglected and therefore the average of the rates obtained from the ZQ and the DQ spectra are used. This is equivalent to using

$$\frac{1}{8T_{\text{MQ}}}ln\left(\frac{I_{\alpha \beta }^{\text{ZQ}}{I}_{\beta \alpha }^{\text{ZQ}}}{I_{\alpha \alpha }^{\text{ZQ}}{I}_{\beta \beta }^{\text{ZQ}}}\frac{I_{\alpha \alpha }^{\text{DQ}}{I}_{\beta \beta }^{\text{DQ}}}{I_{\alpha \beta }^{\text{DQ}}{I}_{\beta \alpha }^{\text{DQ}}}\right)=R_{\text{d}(\text{I}1\text{S}1)/\text{d}(\text{I}2\text{S}2)}+R_{\text{d}(\text{I}1\text{S}2)/\text{d}(\text{I}2\text{S}1)}$$ (1.3)

R_{d(I1S2)/d(I1S2)} cannot be separated from R_{d(I1S1)/d(I2S2)}. Although it can be neglected for sequential rates with a typical value of 0.03 s^-1, this is not valid for intraresidual rates of typically 0.2 s^-1. In the following, both rates must be considered. Under the assumption that radial and spherical motion of the bond vectors are not correlated, the rates are expressed as:

$$R_{\text{d}(\text{I}1\text{S}1)/\text{d}(\text{I}2\text{S}2)}={\left(\frac{{\mu }_0}{4\pi }\right)}^2\frac{{\gamma }_{\text{I}1}{\gamma }_{\text{S}1}{\gamma }_{\text{I}2}{\gamma }_{\text{S}2}{\text{h}}^2}{10{\pi }^2}\frac{1}{(r_{\text{I}1\text{S}1}^{\text{eff}}{)}^3(r_{\text{I}2\text{S}2}^{\text{eff}}{)}^3}{J}_{\text{d}(\text{I}1\text{S}1)/\text{d}(\text{I}2\text{S}2)}(0)$$ (2.1)

$$R_{\text{d}(\text{I}1\text{S}2)/\text{d}(\text{I}2\text{S}1)}={\left(\frac{{\mu }_0}{4\pi }\right)}^2\frac{{\gamma }_{\text{I}1}{\gamma }_{\text{S}1}{\gamma }_{\text{I}2}{\gamma }_{\text{S}2}{\text{h}}^2}{10{\pi }^2}\frac{1}{(r_{\text{I}1\text{S}2}^{\text{eff}}{)}^3(r_{\text{I}2\text{S}1}^{\text{eff}}{)}^3}{J}_{\text{d}(\text{I}1\text{S}2)/\text{d}(\text{I}2\text{S}1)}(0)$$ (2.2)

γ_i is the gyromagnetic ratio of nucleus i, r_ij^eff is the effective distance between nuclei i and j, μ₀ is the permeability of free space, and h denotes Planck's constant. The spectral density function J_d(A)/d(B)(ω) depends on the orientation and dynamics of bond vectors A and B. Equations 2.1 and 2.2 are exact, that is, there are no terms depending on frequencies ω≠ 0.

In the experimental approach presented in Figure 2, all components are minimally manipulated during evolution (referred to as ACE, see NMR spectroscopy) and the peak intensities are obtained from the same spectrum. No invisible systematic errors can be introduced by experimental imperfections such as suboptimal selective pulses. Imperfections rather rescale all components to the same extent and are canceled out in equations 1.1, 1.2 and 1.3.

Figure 2.

Pulse sequences of the 3D ct-HNCA A) and ct-HN(CO)CA B) experiments for intraresidual and sequential R_{d(HN)/d(HαCα)} + R_{d(HαN)/d(HCα)} measurement. For sequence B) the red boxes in sequence A) are replaced. The radio-frequency pulses on ¹H, ¹⁵N, ¹³C^ali and ¹³C′ are applied at 4.7, 118, 56 and 174 ppm, respectively. Narrow and wide bars indicate rectangular 90° and 180° pulses, of which those on ¹³C are applied with a field of Δ/√15 and Δ/√3, respectively, where Δ is the difference between the ¹³C^ali and ¹³C′ carriers in Hz. The single curved pulses represent ¹³C′-selective 180° sinc pulses of length p_C′^π = 150 μs, and the triple curved pulse a ¹³C^α/β-selective ReBURP pulse ⁴⁹ of length p_Cα^π = 400 μs applied at 43 ppm. Vertical lines connect centered pulses. ¹H-decoupling is achieved with WALTZ16 ⁵⁰ at a field strength γB₁ of 2.1 kHz and ¹⁵N-decoupling is achieved with GARP ⁵¹ at a field strength γB₁ of 1.25 kHz. The delays have the following values: τ₁ = 2.3 ms, τ₂ = 14 ms, τ₃ = 18 ms, τ₄ = 2.6 ms, τ₅ = 60 μs, τ₆ = 16 ms, τ₇ =1/(4J_CαC′) = 4.5 ms, Δ = 1/(2J_HN) = 5.4 ms, and T/2 = 14.42 ms − p_Cα^π/2 − 2(p_Cα^π/2)/π, where p_Cα^π/2 is the length of the rectangular ¹³C^ali 90° pulse. Unless indicated otherwise, all radio-frequency pulses are applied with phase x. The phase cycle for the (ZQ+DQ) subspectrum is: φ₁ = {x,-x}; φ₂ = x; φ₃ = {x, x, x, x,-x,-x,-x,-x}; φ₄ = {x, x,-x,-x}; φ₅ = -y; φ₆ = {x, x, x, x, x, x, x, x,-x,-x,-x,-x,-x,-x,-x,-x}; φ_rec = {x,-x,-x, x,-x, x, x,-x}for A) and {x,-x,-x, x,-x, x, x,-x,-x, x, x,-x, x,-x,-x, x} for B). For the (ZQ-DQ) subspectrum φ₃ and φ₄ are increased by 90°. Pulsed field gradients indicated on the line marked PFG are applied along the z-axis with duration/strength of: G₁, 1200 μs / -9 G/cm; G₂, 2000 μs / 12 G/cm; G₃, 2000 μs / 12 G/cm; G₄, 100 μs / 18 G/cm; G₅, 2000 μs / -15 G/cm; G_N1, 200 μs / 18 G/cm; G_N2, 200 μs / -18 G/cm; G₆, 1200 μs / 10.8 G/cm; G₇, 1200 μs / 18 G/cm; G_H, 40 μs / -18 G/cm; G₈, 2000 μs / 12 G/cm; G₉, 1000 μs / 6 G/cm; G₁₀, 2000 μs / 12 G/cm. Quadrature detection in the ¹⁵N(t₁) is achieved by the ECHO-ANTIECHO method ⁵² applied to φ₅ and gradients G6 and G7, and in the MQ[¹⁵N,¹³C^α](t₂) dimension by the States-TPPI method ⁵³ applied to the phases φ₂, φ₄ and φ_rec.

Spectral density function for cross-correlated relaxation

The correlation function describing dipole/dipole cross-correlated motion between the vectorial tensors A and B can be expressed as ³³:

$$C(t)=\frac{4\pi }{5}\sum _{q=-2}^2\langle \frac{{\text{Y}}_{2q}^{\ast }({\theta }_{\text{B}}^{\text{lab}}(t),{\phi }_{\text{B}}^{\text{lab}}(t)){\text{Y}}_{2q}^{\text{lab}}({\theta }_{\text{A}}^{\text{lab}}(0),{\phi }_{\text{A}}^{\text{lab}}(0))}{r_{\text{B}}^3(t)r_{\text{A}}^3(0)}\rangle$$ (3)

The angular brackets denote time averages, Y_2q are the second rank spherical harmonics, r_X is the length of vector X, and the polar angles θ^lab and φ^lab orient the vectors in the laboratory frame.

A. Anisotropically tumbling rigid molecule

Generalizing the expressions for the correlation function in reference ³ to the case of cross-correlated relaxation equation 3 can be written as:

$$C\left(t\right)=\frac{1}{r_{\text{B}}^3{r}_{\text{A}}^3}\sum _{k=-2}^2{C}_k{e}^{-t/{\tau }_k}$$ (4)

where 1/τ_k are the eigenvalues of the anisotropic diffusion operator D ⁴³ and the coefficients C_k contain the orientational dependency on the vectors A and B. Explicit expressions for 1/τ_k and C_k are provided in the Supporting Information.

B. Anisotropically tumbling dynamic molecule

In the following, it is assumed that the time dependencies of r and Y_2q are not correlated and $\langle \frac{1}{r_{\text{B}}^3(t)r_{\text{A}}^3(0)}\rangle$ can be expressed with effective bond lengths $\frac{1}{(r_{\text{B}}^{\text{eff}}{)}^3(r_{\text{A}}^{\text{eff}}{)}^3}$ ⁴⁴. Including fast internal motion requires time averaging of each summand and for the vectors A and B individually in equations S3.1-5. These modified coefficients may be expressed as <C_k>. Lipari and Szabo proposed a single-exponential approximation for C in autorelaxation using an effective correlation time and a generalized order parameter quantifying motion independently of a specific physical model ¹. Ghose et al. extended this approach to cross-correlated relaxation in axially symmetric molecules ⁴⁵. Analog application to fully anisotropic tumbling approximates the correlation function as

$$C\left(t\right)=\frac{1}{(r_{\text{B}}^{\text{eff}}{)}^3(r_{\text{A}}^{\text{eff}}{)}^3}\sum _{k=-2}^2{e}^{-t/{\tau }_k}\left[S_k^2+\left(C_k-S_k^2\right)e^{-t/{\tau }^e}\right]$$ (5)

Here, C_k are functions of the averaged bond orientations. τ^e is the correlation time for internal motion and the 5 generalized order parameters individually associated with the eigenvalues are defined as:

$$S_k^2\equiv \langle C_k\rangle$$ (6)

Even though these order parameters can be very different from one another, a single cross-correlation order parameter has been defined by the sum over all S²_k ³⁴.

Fourier transformation of the correlation function gives the spectral density function:

$$J(\omega )=\frac{1}{(r_{\text{B}}^{\text{eff}}{)}^3(r_{\text{A}}^{\text{eff}}{)}^3}\sum _{k=-2}^2\left[\frac{S_k^2{\tau }_k}{1+{(\omega {\tau }_k)}^2}+\frac{(C_k-S_k^2){\tau }_k^e}{1+(\omega {\tau }_k^e{)}^2}\right]$$ (7)

with

$$\frac{1}{{\tau }_k^e}=\frac{1}{{\tau }^e}+\frac{1}{{\tau }_k}$$ (8.1)

τ_k is typically much smaller than the inverse of the Eigenvalues of the diffusion operator D, and therefore ⁴⁵

$$\frac{1}{{\tau }_k^e}\approx \frac{1}{{\tau }^e}+2\text{tr}\left(D\right)$$ (8.2)

“tr” denotes trace. The spectral density function may be written in the same form as used for autorelaxation. C_k must be simply set as a prefactor:

$$J(\omega )=\frac{1}{(r_{\text{B}}^{\text{eff}}{)}^3(r_{\text{A}}^{\text{eff}}{)}^3}\sum _{k=-2}^2{C}_k\left[\frac{{S\prime }_k^2{\tau }_k}{1+{(\omega {\tau }_k)}^2}+\frac{(1-{S\prime }_k^2){\tau }_k^e}{1+{(\omega {\tau }_k^e)}^2}\right]$$ (9)

where the order parameter S′²_k is then defined as:

$${S\prime }_k^2\equiv \frac{\langle C_k\rangle }{C_k}$$ (10)

This order parameter does not meet the Lipari-Szabo criterion anymore ¹, but recovers the convenient property that it equals 1 for a perfectly rigid molecule.

C. Symmetrically and isotropically tumbling dynamic molecule

More convenient expressions are obtained for simpler models: If the molecular tumbling is axially symmetric, τ_k = τ_{- -k} and equations 6 and 8 can be rewritten as sums over 3 terms ³⁴. If the molecular tumbling is isotropic, all τ_k equal the isotropic tumbling time, and the summation is replaced by a single expression ³⁴:

$$J^{\text{iso}}(\omega )=\frac{1}{(r_{\text{B}}^{\text{eff}}{)}^3(r_{\text{B}}^{\text{eff}}{)}^3}\left(\frac{S^2\tau }{1+(\omega \tau {)}^2}+\frac{({\text{P}}_2(cos{\theta }^{\text{AB}})-S^2){\tau }^e}{1+(\omega {\tau }^e{)}^2}\right)$$ (11)

The only remaining C coefficient is the Legendre polynomial P₂ of the cosine of the projection angle between the vectors A and B, and the order parameter S² is its time average, which is equivalent to:

$$S^2=\frac{4\pi }{5}\sum _{q=-2}^2\langle {\text{Y}}_{2q}^{\ast }({\theta }_{\text{B}}^{\text{lab}}(t),{\phi }_{\text{B}}^{\text{lab}}(t))\rangle \langle {\text{Y}}_{2q}^{\text{lab}}({\theta }_{\text{A}}^{\text{lab}}(t),{\phi }_{\text{A}}^{\text{lab}}(t))\rangle$$ (12)

Equation 7 reduces to

$$J^{\text{iso}}(\omega )=\frac{{\text{P}}_2(cos{\theta }^{\text{AB}})}{(r_{\text{B}}^{\text{eff}}{)}^3(r_{\text{B}}^{\text{eff}}{)}^3}\left[\frac{{S\prime }^2\tau }{1+{(\omega \tau )}^2}+\frac{(1-{S\prime }^2){\tau }^e}{1+(\omega {\tau }^e{)}^2}\right]$$ (13)

The convenience of the definition in equation 9 is now evident. It has become a standard procedure to relate the experimental spectral density function to that calculated for a rigid molecule by

$$J(\omega )=S^{\text{ah}2}{J}_{\text{rigid}}(\omega )$$ (14)

using an ad hoc order parameter S^{ah2 25, 28, 30-32, 46, 47}, which equals S′² for isotropic tumbling.

Experimental section

Sample expression and purification

The protein GB3 was expressed and purified as described previously ⁴⁸. The ¹³C,¹⁵N-labeled NMR sample contained 350 μl of 4 mM protein solution in 95% H₂O, 5% D₂O, 50 mM potassium phosphate buffer, pH 6.5, and 0.5 mg/mL sodium azide.

NMR spectroscopy

The first approach relies on the evolution of all multiplet components of the multiple quantum coherence in a four spin system (see Theory). This approach is referred to as “all components evolution”, or ACE. The 3D ct-HNCA experiment for measurement of intraresidual R_{d(HN)/d(HαCα)} + R_{d(HαN)/d(HCα)} cross-correlated relaxation rates is shown in Figure 2A. ¹H^N(i) polarization is excited and converted into multiple quantum coherences MQ[¹⁵N(i),¹³C^α(i)] via ¹⁵N(i) in two INEPT steps. The MQ coherences are chemical-shift labeled under scalar coupling to ¹H^N(i) and ¹H^α(i) during T = 28.84 ms yielding four components (doublets of doublets) for both the double-quantum and the zero-quantum coherences. Subsequently, the magnetization is converted into single-quantum ¹⁵N(i) for chemical shift labeling and transferred back to ¹H^N(i) for direct detection. Theoretically, magnetization is also converted into a multiple quantum coherence between ¹⁵N(i) and ¹³C^α(i-1) which evolves under scalar coupling to ¹H^N(i) and ¹H^α(i-1), from which sequential relaxation rates can be extracted. Practically, however, it is not possible to obtain optimal sensitivity simultaneously for intraresidual and sequential coherences and peak overlap may further limit this approach. Therefore, experiment 2A is used for optimal transfer to intraresidual ¹³C^α, whereas signals from sequential coherences are rather weak. Experiment 2B provides exclusively signal from sequential coherences. Magnetization on ¹⁵N(i) is transferred to ¹³C^α(i-1) via ¹³CO(i-1) with an additional INEPT step (and a second step during the back transfer). This experiment is essentially the one proposed in references 18 and 23. Two subspectra, (ZQ+DQ) and (ZQ-DQ), are recorded, and subsequently added (subtracted) to obtain the ZQ (DQ) spectra. Experimental details are provided in figure caption 2.

Each subspectrum of the 3D ct-HNCA and ct-HN(CO)CA experiments was recorded with 55(t₁) × 32(t₂) × 256(t₃) complex points, t_1max = 27.5 ms, t_2max = 21.9 ms, t_3max = 63.28 ms, an interscan delay of 1.0 s and 8 scans per increment resulting in a measurement time of 2 days for a pair of subspectra A and B. The time domain data were multiplied with a square cosine functions in the direct dimension and cosine functions the indirect dimensions and zero-filled to 512 × 128 × 2048 complex points.

The approach put forward by the Bodenhausen laboratory is described in detail in the corresponding publications. Two complementary spectra are recorded with peak intensities depending on the interconversion of doubly inphase and doubly antiphase MQ coherences with respect to H^N and H^α. This approach is referred to as “doubly inphase and antihase interconversion”, or DIAI. Intraresidual R_{d(HN)/d(HαCα)} + R_{d(HαN)/d(HCα)} rates were recorded with a pair of 3D pulse sequences as presented in ²² in 2 days. Sequential R_{d(HN)/d(HαCα)} + R_{d(HαN)/d(HCα)} were obtained from pairs of 2D pulse sequences as described in reference ²⁰ run for 3 days, or in reference ²¹ for one day.

All experiments were performed on a BRUKER DRX600 MHz spectrometer, equipped with a z-axes gradient cryogenic probe, respectively, at 298 K.

All spectra were processed and analyzed using the software package NMRPipe ⁵⁴. Peak heights were determined by parabolic interpolation.

Prediction of CCR rates

A. Protein coordinates selection

A variety of coordinate sets of GB3 has been deposited in the Protein Data Bank (PDB). In this study, the positions of the H^N and H^α protons are of particular importance. In X-ray structures, protons can be added at idealized positions. However, large sets of RDCs have been used to demonstrate out-of-plane H^N positions ⁴⁸. Subsequently, highly accurate RDCs have been used to orient H^N–N and H^α-C^α bond vectors with an iterative DIDC method ⁹. ³J_HNHα scalar couplings and intraresidual and sequential D_HNHα RDCs are very sensitive reporters on the proton positions ^{14 15}. Best cross-validation is obtained with an NMR structure (pdb code: 2OED), where the H^N–N and H^α–C^α bond vectors are replaced by those obtained with the DIDC method. The H^α–C^α vectors obtained from protonated samples as presented in reference ⁹ and a new set of H^N–N vectors from deuterated samples give the best cross-validation. The bond lengths were scaled to 1.02 and 1.09 Å, respectively. Such a coordinate set yields rmsd values between predictions and experimental values of 0.32 Hz for ³J_HNHα, 1.15 Hz for intraresidual D_HNHα and 1.66 Hz sequential D_HNHα with the same alignment strength as in reference ¹⁵. Note that all rmsds are lower than in the original publications.

B. Diffusion model selection

Very accurate diffusion tensors for GB3 are available from reference ⁵⁵. The ratio of the main axis to the averaged perpendicular axis is ≈1.4 and there is a small rhombic component. Due to nonhomogenous distribution of H-N vectors, the effective tumbling time calculated as half the inverse trace of the diffusion tensor is ≈3% larger for the isotropic model than for the nonisotropic ones. R_{d(HN)/d(HαCα)} + R_{d(HαN)/d(HCα)} rates were calculated using equations 2.1, 2.2 and 7 for isotropic, axially symmetric and fully anisotropic tumbling. The protein was assumed to be rigid. Correlation plots showing the two extreme cases (isotropic and fully anisotropic) for intraresidual and sequential rates are shown in Figure 3. The slopes indicate an average difference of 4% and 5% with largest changes for residues 7, 33, 39 and 56, and 8, 27, 43 and 51, respectively, for intraresidual and sequential CCR rates.

Figure 3.

Correlation plot of CCR rates simulated for isotropic and fully anisotropic tumbling of GB3. The protein is assumed to be rigid. R_{d(HN)/d(HαCα)} + R_{d(HαN)/d(HCα)} is abbreviated by R. Diffusion tensors are taken from ⁵⁵. Intraresidual and sequential CCR rates are shown in A and B, respectively. The slopes, obtained by a least-square fit, are 1.04 and 1.05 and Pearson's correlation coefficients are 0.996 and 0.997, respectively. Residues that show the largest change are marked in red. Pairwise rmsd values are 0.64 s^-1 and 0.58 s ^-1.

C. Order parameter modeling

Cross-correlation order parameters are simulated for models with various extents of motional correlation. Amplitudes for H^N-N and H^α-C^α bond motion, σ_H^NN and σ_H^αC^α, are obtained from RDC fits assuming Gaussian symmetric motion ⁹. Fluctuations of these two bonds around the rotation axis orthogonal to the N-C^α axis are assumed to be uncorrelated in all cases. Correlated, uncorrelated or anticorrelated motion around the N-C^α axis is expressed as fluctuation of the H^N-N-C^α-H^α dihedral angle:

$${\sigma }_{\text{dihed}}^{\text{corr}}=\left|{\sigma }_{{\text{H}}^{\text{N}}\text{N}}-{\sigma }_{{\text{H}}^{\alpha }{\text{C}}^{\alpha }}\right|$$ (15.1)

$${\sigma }_{\text{dihed}}^{\text{uncorr}}=\sqrt{{\sigma }_{{\text{H}}^{\text{N}}\text{N}}^2+{\sigma }_{{\text{H}}^{\alpha }{\text{C}}^{\alpha }}^2}$$ (15.2)

$${\sigma }_{\text{dihed}}^{\text{anticorr}}={\sigma }_{{\text{H}}^{\text{N}}\text{N}}+{\sigma }_{{\text{H}}^{\alpha }{\text{C}}^{\alpha }}$$ (15.3)

Order parameters are obtained from equation 10. No distinction between different models of overall tumbling is made and thus only integration of the Legendre polynomial of the projection angle is required. Details on the calculation are provided in the Supporting Information. The simplification is justified because the order parameters will be used to estimate corrections to relaxation rates based on a rigid molecule. Such corrections are generally much smaller than the rates themselves. Note that the choice of the axis along which correlated motion is assumed is somewhat arbitrary, but runs approximately parallel to the axis of γ motion for both the intraresidual and sequential case (Figure 1).

Molecular dynamics simulation

Starting coordinates for the protein atoms were taken from the 2OED structure. The protonation states of the ionizable residues were set to their normal values at pH 7. The protein was solvated by a layer of ∼6,500 TIP3P ⁵⁶ water molecules, which extended 12.5 Å from the outermost protein atoms and resulted in a periodic box of the dimensions 52 × 64 × 61 Å. Two Na⁺ ions were placed by the Leap program ⁵⁷ to neutralize the −2 charge of the model system. The parm03 version of the all-atom AMBER force field ⁵⁸ was used for all the simulations.

MD simulations were carried out using the SANDER module in AMBER 8.0. ⁵⁷ The SHAKE algorithm was used to constrain the bond lengths of all bonds involving hydrogen atoms permitting a 2 fs time step. ⁵⁹ A nonbonded pair list with a cutoff of 8.0 Å was updated every 25 steps. The Particle-Mesh-Ewald method was used to include the contributions of long-range electrostatic interactions. ⁶⁰ The volume and the temperature (300 K) of the system were controlled during the MD simulations (with constant volume) by Berendsen's method. ⁶¹

The simulation time was 14.3 ns with a 1.3 ns equilibration period. Coordinates were saved every 10 ps. All of the MD results were analyzed by using the PTRAJ module of AMBER 8.0 and an in-house program. H^N-N and H^α-C^α order parameters S² are calculated using equation 12 and motional amplitudes are extracted under the assumption of Gaussian symmetric motion.

Results and discussion

Validation of measurements

2D planes cut from spectra as obtained in the ACE approach are shown in Figure 4. Residue Ala29 is chosen to represent a lower limit in terms of signal-to-noise. The CCR rates are cross validated with those obtained with the DIAI approach. Relaxation rates obtained from ACE rely on linear combination of relaxation rates of 8 individual components (see Theory). Formula 1.3 shows that potentially different scaling of peak intensities from the two subspectra is canceled out. Different scaling of the 4 components from one subspectrum due to additional relaxation pathways, pulse imperfections etc. are approximately also eliminated, because the MQ evolution is under minimal manipulation. The DIAI experiments suffer from the fact that the ratio of intensities from peaks in two different spectra is calculated. One spectrum serves as a reference and decouples the protons. A second spectrum evolves the MQ coherence into antiphase with respect to both protons which assumes uniform scalar couplings. Undesired pathways superimpose contributions on the evaluated peaks. On the other hand, this approach is very sensitive. The uncertainty of only two (generally strong) peak intensities (as opposed to 8) is propagated and a smaller random error can be expected. Figure 5 shows correlation plots of experimental R_{d(HN)/d(HαCα)} + R_{d(HαN)/d(HCα)} CCR rates and Table S1 in the Supporting Information lists correlation parameters. The reliability of the ACE approach is demonstrated by the comparison to predicted values (vide infra). In addition, the difference between linear regressions for intraresidual and sequential rates is only 1%. The DIAI approach is not as uniform. Taking the ACE rates as a reference, intraresidual DIAI rates are underestimated by ≈7% with the pulse sequence in reference ²². Interresidual rates are relatively largely overestimated by the initially used pulse sequence ²⁰ (≈20%), but are less than 5% underestimated with an elegantly improved version of the pulse sequence ²¹. This observation seems in rough agreement with measurements on human ubiquitin where experimental CCR rates fall on curves predicted for rates with ad hoc order parameters of ≈0.9 and ≈0.76 for the former and latter pulse sequence ^{20, 21}. Note that underestimations of 7% and 5% could hardly be identified using only one experimental approach and no highly accurate structure due to uncertainties in bond vector orientation, tumbling time and dynamics. Clearly, the slopes of the ACE approach must be used for evaluation of overall order parameters. However, rmsds from predicted values may be comparable or better for DIAI than for ACE. After uniformly scaling all DIAI rates to the slope of ACE, the rmsd between ACE and DIAI rates can be used to estimate the random and a fraction of the systematic error (pairwise rmsd for intraresidual 0.62 s^-1 and for sequential 0.67 s^-1). It should be noted that both approaches might produce some identical systemtic errors since the two approaches are based on similar principles. Averages of rates from ACE and rescaled rates from DIAI yield errors of about 0.3 s^-1 (intraresidual 0.31 s^-1 and sequential 0.34 s^-1).

Figure 4.

2D planes and slices cut from the 3D ct-HNCA and ct-HN(CO)CA experiments showing the multiplets of Ala29. DQ spectra are shown on the left and ZQ spectra on the right. Intraresidual multiplets are on top, and sequential multiplets at the bottom. The horizontal axes represent MQ frequencies in Hz units with arbitrary origins. The peak intensities of residue Ala29 constitute a lower limit for multiplets evaluated.

Figure 5.

Correlation plots of CCR rates obtained with the ACE and DIAI approaches. R_{d(HN)/d(HαCα)} + R_{d(HαN)/d(HCα)} is abbreviated by R. Intraresidual and sequential CCR rates are shown in A and B, respectively. DIAI rates are scaled to match a slope of 1 (see text) resulting in pairwise rmsd values are 0.62 s^-1 and 0.67 s ^-1. Pearson's correlation coefficient is 0.988 and 0.991, respectively. Outliers are marked in red.

Fit to rigid structure

A uniform scaling factor relating predicted to experimental CCR rates is obtained from the slope in a linear regression. This factor is subsequently used to obtain motionally corrected predicted rates and the pairwise rmsd from the experimental rates is calculated. The data sets in the ACE approach have rmsd values between 0.8 and 1.0 s^-1, whereas those from DIAI are between 0.5 and 1.0 s^-1 (Table S1 in Supporting Information). Exceptionally good fits are obtained for the intraresidual DIAI approach, suggesting a low random error. The slopes in the ACE approach differ by 1% for intraresidual and sequential rates, whereas a larger spread is obtained in the DIAI approach. Overall, the ACE approach produces a reliable slope, but DIAI proves very sensitive where systematic errors reflected in deficient slopes are mostly uniformly scaled throughout the molecule. In the following, rates obtained from averages of both approaches, where the rates from DIAI are scaled to produce the same slopes as those from ACE are used (Table 1). The change from the isotropic tumbling model to the axially symmetric tumbling model improves the rmsd by ≈0.05 s^-1. An additional improvement of ≈0.01 s^-1 is obtained with the fully anisotropic model. Importantly, choice of the isotropic model yields ≈5% smaller slopes, which would significantly distort the extraction of order parameters. Clearly, the axially symmetric or fully anisotropic model must be used for a proper analysis. Not surprisingly, the small rhombicity of the diffusion tensor has only a small impact on the rmsd and the slope. Figure 6 shows correlation plots of predicted and experimental CCR rates for the anisotropic tumbling model.

Table 1.

Slopes, order parameters, rmsds and Pearson's correlation coefficient r between experimental and predicted CCR rates of GB3.

CCR	diffusion modela	slopeb	Sad2c	S′2c	rmsd [Hz]d	r
intraresidual	isotropic	0.849±0.011	0.971±0.013	0.971±0.013	0.75	0.983
	axially symmetric	0.894±0.011	1.023±0.012		0.70	0.986
	fully anisotropic	0.888±0.011	1.016±0.012		0.69	0.986
sequential	isotropic	0.851±0.015	0.973±0.017	0.973±0.017	0.84	0.986
	axially symmetric	0.904±0.014	1.034±0.016		0.74	0.989
	fully anisotropic	0.897±0.014	1.026±0.016		0.73	0.989

^aThe diffusion tensors are taken from ⁵⁵.

^br_HN = 1.02 Å and r_HαCα = 1.09 Å are assumed. X axis is the predicted and Y axis the experimental rate.

^cr_HN^eff = 1.041 Å and r_HαCα^eff = 1.117 Å are assumed. X axis is the predicted and Y axis the experimental rate.

^dPairwise rmsd value. The predicted values are multiplied by the slope.

Figure 6.

Correlation plots of predicted and experimental CCR rates. R_{d(HN)/d(HαCα)} + R_{d(HαN)/d(HCα)} is abbreviated by R. Intraresidual and sequential CCR rates are shown in A) and B), respectively. For prediction, the fully anisotropic model is used ⁵⁵. Vertical error bars represent the difference between the values obtained from the ACE and DIAI approach. The slopes, obtained by a least-square fit, are 0.888 and 0.897 and Pearson's correlation coefficient 0.986 and 0.989, respectively. Outliers are marked in red. After scaling the rates obtained from the DIAI approach by the slopes, pairwise rmsd values are 0.69 s^-1 and 0.73 s^-1.

As is evident from equation 6 and 10, S² and S′² can only be experimentally quantified for isotropic tumbling without invoking a specific model of motion. However, S^ah2 is model independent and proportional to the slope between the theoretical and experimental CCR rates. Table 1 lists order parameters obtained for all models of molecular tumbling (For a complete list of all experiments see Table S2 in the Supporting Information). Bond lengths r used in the prediction for a rigid molecule are replaced by r^eff (see Theory). Values of 1.041 and 1.117 Å for H^N–N and H^α–C^{α 44}, respectively, result in a multiplication factor of 1.144. Such effective lengths absorb radial and angular fluctuations of isolated bonds, that is zero-point librations and vibrations as well as angular fluctuations around the C^α-C^α axis. The isotropic tumbling model produces order parameters that are typically 5% smaller than for the other models. Note that the effective tumbling time is about 3% larger (see experimental section) and the additional 2% are effects of nonhomogeneity of the vectors involved in CCR. Deviations of more than 1 s^-1 are observed for residues 15, 24, 27, 29, 31 and 39 for the intraresidual, and 9, 18, 26, 34, 37, 46 and 48 for the sequential rates. Residues 15 (intraresidual) and 46 (sequential) have large errors of 0.62 s^-1 and 0.48 s^-1, respectively, and do fit well using only the DIAI values. These outliers are likely to lack accuracy in the ACE approach. For residues 24, 29, 31 and 39 (intraresidual) and 9, 18, 26, 34, 37 and 48 (sequential) both approaches yield deviation, and for residue 27 (intraresidual) no value could be obtained from the DIAI approach leaving the experimental error unknown. Erroneous vector orientation or dynamic effects may be the cause. However, neither experimental ³J_HNHα scalar couplings nor D_HNHα RDCs exhibit unusually large deviations from values predicted from the vectors used here (data not shown). Many of these residues are located in the α helix comprising residues 23-36. Generally, few good fits are obtained for the loop with residues 37-41, due to lack of experimental data (residues 38 and 41 are GLY), lack of accurate vector orientation (residue 40) or large deviation (residues 37 and 39). Interestingly, the large deviation for residue 39 is not present when using the isotropic tumbling model (see figure 3). The only poor fits for β strands are residues 18 (sequential) and 46 (intraresidual). Interestingly, outliers in the α helix fall all to the same side of the slopes.

Overall, the experimental CCR rates can be predicted very closely by setting the order parameter 1 and using effective bond lengths to absorb motional effects. However, these bond lengths are obtained from analysis of motional effects on isolated bonds. In general, their use for prediction of CCR rates is flawed due to presence of correlated motion.

Fit to dynamic structures

A more elaborate description of CCR rates takes the presence of correlated motion into account. Therefore, the effective bond length is adjusted to absorb radial but no spherical fluctuation, that is, 1.02 and 1.09 Å are assumed for H^N–N and H^α–C^α. It should be noted that in a recent publication the effective H^N–N bond length absorbing zero-point vibrations, but no angular fluctuations, has been determined to be 1.015±0.006 Å. ⁶² Here, 1.02 Å is chosen, which is frequently used in pdb files and is within the experimental uncertainty range. Order parameters are obtained from integration over projection angles obtained from models assuming uncorrelated, correlated and anticorrelated bond motion. Figure 7 shows correlation plots relating the experimental relaxation rates to the simulated relaxation rates for 4 models of correlated motion: Fully correlated motion (equivalent to the rigid molecule), correlated and anticorrelated motion around the H^N-N-C^α-H^α dihedral angle, and uncorrelated motion. Generally, order parameters from anticorrelated motions are the smallest, followed by uncorrelated motion and correlated motion. As expected, the rigid model (complete correlation) overestimates the rates for nearly all residues because shortening effective bonds lengths simply rescales the order parameters obtained in the previous section. The prediction closest to the experimental values is about evenly distributed between the three non-rigid models. Some predictions fall clearly outside the experimental error range. The cumulated uncertainties from experimental errors, errors in bond orientation, motional amplitudes of the isolated bonds and simplification of the models do not allow quantification of correlation degrees. Nevertheless, trends in specific structural elements can be observed. Figure 8 shows the H^NN-H^αC^α network. The motional model that predicts the value closest to the experimental rate is assigned to each connectivity (For a similar map with model assignment in letter code see Supporting Information figure S1). In addition, model assignments based on a 13 ns trajectory of a molecular dynamics (MD) simulation is shown. Here, expression 15.2 is subtracted from the rmsd of the dihedral angle. If this value is smaller than -2° (larger than +2°), correlated (anticorrelated) motion is assigned. Experimental evidence of correlated motion in the loops comprising residues 10-14, 20-22 and 47-50, and anticorrelated motion in the α helix comprising 23-38 is supported by MD simulations. Somewhat weaker correlation in the β strands 2-4 is not substantiated by the MD simulations. Such motion may be present on a time scale not sampled by the simulation, but be picked up by RDCs and CCR rates.

Figure 7.

Correlation plots of experimental and predicted CCR rates based on different models for correlated bond motion. R_{d(HN)/d(HαCα)} + R_{d(HαN)/d(HCα)} is abbreviated by R. Intraresidual and sequential CCR rates are shown in A) and B), respectively. The theoretical CCR rates are calculated for the rigid molecule (dark blue diamonds), and models featuring correlated (yellow triangles), uncorrelated (pink squares), and anticorrelated motion (light blue crosses). Bond specific motional amplitudes obtained from H^N-N and H^α-C^α RDCs have been used in equations 15.1-3 to model the extent of motional correlation. Then, order parameters are obtained from integration over the projection angles. The motional model that predicts the value closest to the experimental rate is assigned to each connectivity.

Figure 8.

Dynamics correlation network of GB3. C^α_(i-1)-N_i and N_i-C^α_i vectors are color coded according to the best matching model for motional correlation between H^α-C^α and H^N-N bonds: Correlated motion, red; uncorrelated motion, blue; and anticorrelated motion, yellow. The network labeled “exp” (“sim”) is obtained from the experimental CCR rates (13 ns MD simulation). Bold lines are used if the model assignment is unambiguous within the experimental error range in the experimental approach, and if the difference between the dihedral fluctuation and expression 15.2 is larger than 3.5°. On top of the networks, amplitudes of the γ motion presented in reference 16 are plotted.

Indeed, correlated motion of the β sheet in GB3 in the nano- to millisecond range has been proposed based on measurements of scalar couplings across hydrogen bonds ¹⁶. In this study, Gaussian motional amplitudes in three dimensions, but not the degree of correlation have been calculated with large sets of RDCs (see Figure 8). Comparison of RDC order parameters to Lipari-Szabo order parameters and Accelerated Molecular Dynamics (AMD) ¹² then indicates μs-ms motion in the loops and the β sheet, but not in the α helix. In the strands β₁, β₃ and β₄ an alternation of large and small motional amplitudes is observed pointing to a mode coupled across the sheet. Interestingly, the pattern matches the alternation of strongly hydrophobic side chains buried in the protein core. Such alternation is not present in the CCR pattern. However, the CCR data does not probe exactly the same type of motion and a strict comparison has to be handled with care. In addition, the authors proved the presence of correlated motion across the β sheet. For this purpose, scalar couplings across hydrogen bonds simulated for an ensemble exhibiting correlated motion are shown to cross-validate best with the experimental data. Note that such an approach cannot be applied to the α helix and the loops. In the present study, correlated motion is detected within the polypeptide chain. Apparently, the correlated motion of the β sheet undergoes a collective rotational fluctuation along the polypeptide chain in a manner synchronized with neighboring β strands.

In this context, it is interesting to note that supra τ_c motion has also been proposed along the entire sequence of ubiquitin which has a fold similar to GB3. ^{63 64} Again, this has been concluded from the observation that RDC orders parameters covering a time window up to ms are smaller than Lipari-Szabo order parameters. A major part of this dynamics is concentrated in a single concerted mode related to molecular recognition. ⁶⁴ In particular, β strand residues with solvent-exposed side chains exhibit reduced order parameters relative to those with core side chains. ⁶³ This may be a hint to collective motional β sheet behavior on slow time scale.

Conclusion

A new method is introduced to assess correlated dynamics between bond vectors. Cross correlated relaxation rates are measured with high precision between vectors with accurately known orientation. Experimental rates are compared to rates predicted for a rigid structure. It is demonstrated that the assumption of anisotropic molecular tumbling is necessary to evaluate precisely cross correlated relaxation rates. Deviations are matched to models of different degrees of motional correlation. These models are based on previously determined orientations and motional amplitudes of isolated bond vectors obtained from residual dipolar couplings. It is shown that for GB3 predictions from a static structure using effective bond lengths absorbing libration and vibration (1.041 and 1.117 Å for H^N–N and H^α–C^α) are within 3 % of both experimental intra and interresidual rates. Analysis involving motional models shows clear evidence of correlated motion in the loops comprising residues 10-14, 20-22 and 47-50, and anticorrelated motion in the α helix comprising 23-38. Somewhat weaker correlation is observed in the β strands 2-4, which have previously been shown to exhibit slow correlated motional modes. More experimental data and further refinement of motional models are expected to lead towards individual quantification of correlated dynamics between bond vectors.