Reconstruction of coupled intra- and interdomain protein motion from nuclear and electron magnetic resonance

Abstract

Proteins composed of multiple domains allow for structural heterogeneity and interdomain dynamics that may be vital for function. Intradomain structures and dynamics can influence interdomain conformations and vice versa. However, no established structure determination method is currently available that can probe the coupling of these motions. The protein Pin1 contains separate regulatory and catalytic domains that sample “extended” and “compact” states, and ligand binding changes this equilibrium. Ligand binding and interdomain distance have been shown to impact the activity of Pin1, suggesting interdomain allostery. In order to characterize the conformational equilibrium of Pin1, we describe a novel method to model the coupling between intra- and interdomain dynamics at atomic resolution using multi-state ensembles. The method uses time-averaged nuclear magnetic resonance (NMR) restraints and double electron-electron resonance (DEER) data that resolves distance distributions. While the intradomain calculation is primarily driven by exact nuclear Overhauser enhancements (eNOEs), J couplings, and residual dipolar couplings (RDCs), the relative domain distribution is driven by paramagnetic relaxation enhancement (PREs), RDCs, interdomain NOEs and DEER. Our data supports a 70:30 population of the compact and extended states in apo Pin1. A multi-state ensemble describes these conformations simultaneously, with distinct conformational differences located in the interdomain interface stabilizing the compact or extended states. We also describe correlated conformations between the catalytic site and interdomain interface that may explain allostery driven by interdomain contact.

Graphical Abstract

Introduction

Proteins with multiple domains are the norm, not the exception. While over 80% and 67% of eukaryotic and prokaryotic proteins, respectively, include more than one domain, less than 35% of structures deposited in the Protein Data Bank contain multiple domains¹. Multi-domain proteins are often more stable and easier to fold than a single domain while allowing for greater structural and functional plasticity². Many domains are linked together by an intrinsically disordered region that may act as a hinge to allow for structural heterogeneity between domains while the individual domains internally maintain their 3D structure^3–5. Multiple orientations between domains may be vital for the function and activity of a biomacromolecule. Domain orientation and interdomain distance at equilibrium can change with environmental parameters, e.g. ligand concentration.

Many properties of a protein, including ligand binding, catalysis, and stability, are influenced by a protein’s conformational dynamics. These dynamics range from small movements near an active site to large collective motions over entire domains. In addition, dynamics are believed to be a major factor in allosteric regulation of a protein. Allostery typically refers to the phenomenon of an effector molecule acting at a distal site that regulates the function of a catalytically active site. For calmodulin and Pin1, both two-domain proteins, it has been shown that ligand binding changes the orientation of the two domains^4,6–10. In order to sustain a long-range allosteric mechanism, local motions must be correlated. Although chemical shifts, order parameters, and catalytic activities have been used to characterize allosteric behavior, these signatures only indirectly report on the underlying dynamics and it has thus been difficult to accurately determine conformational differences.

Studying an isolated domain is useful to learn about its basic structure and function, yet tethering multiple domains together has been shown to significantly alter the structure in 50% and dynamics in 90% of the protein domains studied¹¹. Therefore, investigating individual domains is not sufficient for generating a full understanding of a multi-domain protein. Whereas recent computational approaches can predict multi-domain protein configuration and interdomain interfaces by ab initio folding potential¹² and global structural alignments¹³, experimental studies of multi-domain proteins face a number of challenges in orienting multiple domains. Crystal structures typically capture only one conformation and may feature packing artefacts. Many proteins of interest are too small for cryo-electron microscopy (for now), whereas recent advances in solution NMR have led to structural restraints of macromolecules larger than 100 kDa¹⁴. Even though solution NMR allows interrogation of structure and dynamics of such multi-domain systems, relying solely on interdomain nuclear Overhauser enhancement (NOE) is not sufficient to orient multiple domains. ¹⁵N relaxation data, paramagnetic relaxation enhancement (PRE) and residual dipolar couplings (RDCs) have successfully been utilized to describe interdomain orientation and dynamics^9,10,15–20, but these motions have not been linked to intradomain motion.

Here, we introduce a novel method to solve the structure of a two-domain protein that allows for coupling between intra- and interdomain dynamics at atomic resolution. We aim to identify the structural correlations between intradomain structure and interdomain positions. Our approach builds on previous work by the Clore and Vendruscolo labs that proposed to identify correlated within single domains by a combination of conventional NOEs, scalar couplings, RDCs and relaxation order parameters^21–23. The method presented here is primarily based on an innovative combination of both emerging NMR and EPR techniques for precise short- and long-range distance measurements, supplemented by angular restraints. These measurements yield time-averaged short-range distances (within domains) and probability distributions of long-range distances (between domains). The restraints we use in our calculations include exact NOEs, scalar couplings, RDCs, PREs, and double electron-electron resonance (DEER).

Recent advances in the exact quantitative evaluation of the NOE (eNOE)^24–27 allow us to measure proton-proton distances with less than 0.1 Å error up to 5 Å in favorable cases^27,28, and we can detect distances up to ~8 Å. As these eNOEs are motion- and population-averaged observables, structure calculations based on multiple states often achieve a better agreement with experimental data than an averaged model²⁵. As such, our method is able to characterize and depict correlated motions from experimental data.

Besides the few eNOEs we observe between the two domains at the interface, we rely on PRE, RDC and DEER to access longer distances and thereby generate the following restraints. First, PRE allows distance measurements up to 25 Å between the unpaired electron of a paramagnetic spin label and nuclei. Although it does not offer the accuracy of eNOE, it is a powerful tool to characterize the domain orientations and motions. Second, we use RDCs to determine the relative orientation of the two domains. RDCs report on the bond orientation within molecule-fixed frames and thus carry long-range information. Importantly, PREs have also been shown to reduce the degeneracy of RDCs in a multi-domain protein and to detect transient, minor states in an exchanging system^29,30. Both PRE and RDCs can be used to determine interdomain dynamics, orientation and motion, yet the averaging of these parameters in NMR is convoluting. Therefore, we also utilize DEER to measure interdomain distances and their populations. DEER measurements are performed using flash-frozen samples, so that this electron paramagnetic resonance (EPR) technique provides a distance distribution between spin labels instead of a solution-averaged distance. Importantly, the distances obtained at low temperature can be reliably combined with NMR data at room temperature, due to the slow exchange between the interdomain positions. The distance range depends on the spin environment, and for biological systems typically distances between 15 and 50 Å can be characterized by 4-pulse DEER (4pDEER)³¹. Here, we also use a dynamically decoupled version of the experiment with 5 pulses (5pDEER)³². Together with recent advances in microwave technology that can be exploited to achieve favorable excitation bands, 5pDEER can extend the accessible distance range up to 80 ų³. Often used in conjunction with other biophysical techniques such as crystallography, NMR, and small-angle x-ray scattering (SAXS), DEER has determined the orientation of multi-domain systems, namely the HIV-1 RT p66 homodimer³⁴, fibronectin type III domains of integrin α6β4³⁵, tandem POTRA domain pair of BamA³⁶, calmodulin³⁷, and the E. coli 5’-nucleotidase³⁸. While the integrin and BamA DEER distance distributions were narrow and indicative of only one conformation, an additional broader ensemble was detected in case of the p66 reverse transcriptase and nucleotidase domains. The nucleotidase system even showed that substrate binding could change the equilibrium from completely open to a mixed population of an open and closed populations³⁸.

We applied our method to the two-domain mitotic regulator Pin1. Pin1 (protein interacting with NIMA kinase 1) is a 163 residue peptidyl-prolyl isomerase specific for isomerizing prolines that are immediately preceded by a phosphorylated serine or threonine (pS/TP)^39–41. Residues 1–39 form the WW interaction domain (named for two conserved tryptophans) that features a three-stranded, antiparallel β-sheet which binds the pS/TP motif trans-specifically⁴². Residues 50–163 form the catalytic PPIase domain responsible for isomerizing the proline in the same motif, and is composed of a four-stranded core β-sheet with four exterior α-helices^43,44. A 10-residue flexible linker separates the two domains.

According to the first crystal structure of Pin1, the two domains assume a “compact” conformation with an interdomain interface composed of residues 28–32 in the WW domain interacting with residues 137–142 and 145–149 in the PPIase domain (Figure 1A)⁴³. The PPIase interdomain interface is located on the opposite side of the catalytic site. Such interaction between the two domains was lacking in an initial NMR ensemble as the domains were found to be in an extended conformation (Figure 1B)⁴⁵. Subsequent NMR relaxation experiments proved that Pin1 tumbles somewhere between two independent domains as a single, rigid unit. Interestingly, ligands or point mutations change the equilibrium of compact and extended states dependent on the ligand sequence. These shifts lead to changes in the catalytic activity of Pin1 via an allosteric mechanism^42,46–51 and suggest that different conformations of the individual domains exist that stabilize the compact and extended states. Our previously solved eNOE two-state structure of the isolated WW domain of Pin1 supports this notion⁵². Addition of two different ligands either conserves the correlations between the ligand-binding site and the WW/PPIase interface, or induces partial anti-correlation. Importantly, anti-correlation renders both states incompatible with a compact conformation when the PPIase domain is modeled in its typical compact position. This possibly explains why one of the two ligands shifts the equilibrium towards the open conformation. In order to verify this hypothesis, a multi-state representation of the full-length Pin1 at atomic resolution is required. Although many structures of Pin1 have been solved, none of them depicts Pin1 at equilibrium with compact and extended states.

Figure 1.

Structure of Pin1. A) Crystal structure 1pin⁴³ showing the “compact” state with the WW (orange) and PPIase (blue) domains with all eNOEs plotted. Interdomain NOEs are colored in green. B) Conventional NMR structure 1nmv⁴⁵ showing the extended state. MTSL mutations are colored in purple, and DEER restraints and 98-MTSL PRE distances are overlaid in black and grey, respectively.

Yet, some work has been done to determine the general orientations and distances in the extended states without considering the spatial sampling within the domains^9,53. First of all, while NMR relaxation experiments can report on the degree of compactness in a multi-domain protein, this method is unable to provide information about the actual domain positions^8,15. Secondly, RDCs have been utilized on wild-type and I28A mutant Pin1 to generate long-range, orientational bond-vector restraints based on the incomplete averaging of dipole-dipole interactions⁹. These RDCs were used in conjunction with a Langevin dynamics simulation method optimized for large conformational changes^9,20. It is difficult to determine domain positions solely by RDCs due to ambiguities with respect to orientation and insensitivity to translation. Lastly, PRE using a single paramagnetic label has been measured on a Pin1 construct to evaluate the domain distance upon the addition of PEG400⁵⁴. A recent study also made use of PREs induced by single label (at position H27)⁵⁵. This study revealed further interdomain contacts located in the first two α helices and connecting loop. Importantly, it also presents PRE evidence that interdomain separation is correlated with compaction of the WW domain.

We demonstrate the power of our method that integrates eNOE, J coupling, RDC, PRE, and DEER data to elucidate coupled intra- and interdomain motion by solving multi-state ensembles of full-length apo Pin1. Our two-state ensemble satisfies all NMR data and reproduces the compact and extended states. The obtained domain positions are also in close agreement with the DEER distance distributions. We observe distinct intradomain conformations correlated to interdomain distance and propose how the intradomain conformations may stabilize the compact and extended states. Finally, we show structural changes in the catalytic site that are correlated to interdomain contact, supporting a model of interdomain allostery.

Results and discussion

Data collection

Tumbling time-specific and interdomain eNOEs

We applied our eNOE buildup method on Pin1 to extract precise distances between protons with the ultimate goal of determining an accurate multi-state structural ensemble. This method is well-established for single-domain proteins of various sizes as well as RNA^25,27. Here, we apply this method for the first time to a multi-domain protein. In order to convert a cross-relaxation rate into an effective exact distance, it is imperative to determine the tumbling time (τ_c) of the buildup sample. For a single-domain globular protein, the overall tumbling time suffices to determine an accurate distance. In the case of Pin1, its two domains are known to tumble partially independent of one another^8,56 and we therefore rely on R₁ and R_1ρ measurements to quantify the domain-specific tumbling times as 11.3, 14.1, and 3.6 ns for the WW, PPIase, and linker, respectively. By implementing residue-specific τ_c in the eNORA2 program in CYANA, we were able to extract 537 bi-directional and 1731 uni-directional eNOEs as superimposed on the crystal structure in Figure 1A. In addition, 1911 generic-normalized (gn) eNOEs were also determined from spins, where the diagonal decays could not be fitted⁵⁷. Furthermore, we added 124 ³J_HN,Hα, 129 ³J_Hα,Hβ(2,3) (also used for the stereospecific assignment), and 12 aromatic ³J_N,Cγ scalar couplings to increase the intradomain restraint density.

The original NMR structural work of Pin1 (1nmv)⁴⁵ lacked NOEs between the WW and PPIase domain (interdomain NOEs, ID NOEs). However, we were able to identify 20 eNOEs (3 uni-, 10 bi-, and 7 gn-eNOEs) that according to the crystal structure, 1pin⁴³, lie in the region of the interdomain interface (Figure 1A). Presumably, our higher sample concentration (2 mM) and a 900 MHz cryo-probe spectrometer increased the sensitivity towards ID NOEs compared to room temperature measurements performed on a 800 MHz spectrometer using a sample concentration of 0.6–0.8 mM⁴⁵. The interface is mostly composed of hydrophobic residues, so that half of the eNOE spin pairs were associated with the methyl groups of these residues. As a dynamic system the NOEs are averaged, hence, ID NOEs provide evidence for the compact state, but the simultaneous existence of extended states cannot be discounted because the cross-relaxation rate is dominated by short sampled distances.

Residual dipolar couplings to orient domains

We measured RDCs on the full-length protein to aid orienting the two domains. RDCs are still averages over dynamic ensembles, and Langevin dynamic simulations have previously been used on Pin1 to generate conformational ensembles that freeze intradomain motion while allowing collective interdomain motion⁹. Using C12E5 PEG/hexanol, we were able to determine 407 RDCs (140 ¹D_Ni,HNi, 138 ¹D_C’i,Cαi, 103 D_C’i,Ni+1, and 97 D_C’i,HNi+1). Given the size difference of the WW and PPIase domains (39 vs 113 residues, respectively), and the fact that the alignment in PEG/hexanol is primarily steric in nature, we expect that the PPIase will cause a greater alignment with the magnetic field. We note that a more favorable way of achieving the alignment of a two-domain protein would be the attachment of a lanthanide tag^58,59, such that the alignment would be entirely independent of the relative domain positions. In addition to orient one domain relative to another, we also aimed to determine the proper orientation of bond vectors within the domain. While the fit of the WW domain alignment tensor using the X-ray structure (1pin) was excellent (measured vs. back-predicted RDCs, r = 0.84), our previously solved eNOE-structure of the isolated WW domain (6svc⁵²) achieved an even better agreement (r = 0.94), reinforcing the potential of eNOEs. Our ensemble calculation consists of two steps. First, using the WW alignment tensor, we determined bond vector orientations of the WW domain. In the second step, the RDCs of the entire Pin1 were fitted using a PPIase-specific alignment tensor, obtained from an initial fit to 1pin (r = 0.92, after removal of outliers 0.96). For this step, the angles of the WW domain were frozen, allowing the domain to move as a rigid body relative to the PPIase domain.

PRE and DEER mutants for long-range distance restraints

We relied on PRE and DEER for long-distance restraints that define the relative domain translation and rotation. For these paramagnetic techniques that require adding a nitroxide spin label, we engineered constructs of Pin1 to probe interdomain distances with minimal disruption of the wild-type conformation and enzyme activity. As the MTSL spin label conjugates to free cysteines, we needed to i) mutate endogenous cysteines (C57 and C113) and ii) introduce cysteines to various regions of the protein. The fairly conservative C57S mutation caused severe protein precipitation, while the C57A mutation maintained stability. Originally the endogenous C113 was considered essential for isomerase activity^60–62, although recent work has demonstrated that a mutation to aspartate conserves the protein’s activity^63,64. Therefore, all PRE and DEER mutant constructs also feature C57A and C113D mutations. Using the structure as a guide, we introduced cysteines at M15 (in the WW domain), and N90, S98, and Q131 (in the PPIase domain) as shown in Figure 1B. Chemical shift perturbations were measured on all PRE samples to ensure the absence of long-range perturbations that are indicative of disruption of the overall structure and orientations, as shown in Figure S2. Table S1 reports the isomerase activity of WT Pin1 and of these stable PRE constructs; Figure S3 displays example spectra and fits. All PRE mutants maintained some activity, while mutants M15C and N90C displayed nearly identical activity as WT Pin1. These four PRE constructs were then combined into double-mutants for DEER spectroscopy and Table S1 also reports the associated isomerase activities.

Paramagnetic relaxation enhancement across domains

Due to the high gyromagnetic ratio of an electron, PREs can probe distances up to 25 Å⁶⁵. We extracted an electron-nucleus distance by measuring the difference of transverse relaxation between a MTSL-labeled and DTT-quenched sample (Figure S4A), termed “spin-enhanced” relaxation rate (R₂^sp). While acetyl-MTSL is commonly used for the diamagnetic sample, DTT and ascorbic acid can also be used to remove the MTSL allowing for the use of the exact protein sample for both diamagnetic and paramagnetic measurements^54,65–67. Figure 2A displays the measured R₂^sp of construct S98C as a function of residue number. Peaks within 13 Å completely disappear from the spectra while peaks up to 25 Å show quenching due to R₂^sp. Residues with an R₂^sp greater than 15 s⁻¹ are located in a ~19 Å sphere surrounding the MTSL spin label and are highlighted in the crystal structure (Figure 2A). Figure 2B shows the analogous representation for M15C, N90C, and Q131C, with associated residue-resolved graphs in Figure S4 B–D. In addition to intradomain quenching, constructs M15C, N90C, and S98C feature many residues that are quenched across domains, validating this method to probe interdomain distances in Pin1. Similar to the NOE, the R₂^sp is solution-averaged, rendering the deconvolution of multiple sampled distances challenging. Nevertheless, based on the increased quenching at shorter distances, PRE has been used to infer the presence of a transient state that is much more compact than the extended state of the major conformer. ^68,69. For Pin1, we are unable to determine populations of these states through PRE alone.

Figure 2.

PRE of Pin1. A, left) Residue vs spin-enhanced relaxation rates of 98-MTSL. Right) Major R₂^sp from 98-MTSL plotted on the structure 1pin⁴³. B) Major R₂^sp from 15, 90, 131-MTSL plotted on the structure 1pin⁴³.

Double electron-electron resonance for probing distance distributions

We combined the labelling sites used for the PRE constructs to form six double mutants for EPR spectroscopy. DEER measurements exploit the dipolar coupling between two paramagnetic labels to detect distances in the range of 15–80 Å. Importantly, DEER can be used to determine distance distributions, i.e. all distances present in the sample at the time of freezing (samples are flash-frozen before measurement). We utilized this information to supplement the NMR-based methods employed here, which readout averaged distances.

Figure 3 shows distance distributions between the nitroxide labels for five of the six double mutants. 90–131 and 98–131 are intradomain distances in the PPIase domain, which could be determined by means of 4pDEER. For the three interdomain (ID) distances between the WW and PPIase domains, it was crucial to employ 5pDEER to reliably measure the longer distances associated with extended states. Whereas the DEER data recorded for the double mutants 98–131, 90–131, and 15–131 could only be fit using a model-free distance distribution (Figure S6), for the constructs 15–90 and 15–98 two distinct dipolar oscillations were observed which could be equally well described by two Gaussian components (Figure S7A–B). The regularization parameters and artefact correction were optimized as the bimodality in these distance distributions influence the width of the peaks. For the double mutant 90–98 the observed dipolar oscillation (Figure S5D) could not be translated to a distance distribution because the close proximity of the spin labels (~13 Å expected) violates the point-dipole approximation.

Figure 3.

Distance distributions determined by DEER measurements of double-MTSL Pin1 constructs. The experimental population density P(r) and 95% confidence interval are shown for each construct. For 15–90 and 15–98 mutants, the data could be fitted well using a bi-Gaussian P(r), and average distances (r), standard deviations (σ), and populations (p) are reported for the two components.

Upon comparing the distance distributions in Figure 3, it becomes evident that distance distributions involving position 90 are typically narrower than for constructs featuring position 98. This suggests that the spin label at position 90 is more restricted in conformational freedom compared to the other labeling positions. The broadened lineshape of the room-temperature continuous-wave (CW) EPR spectrum of 15–90 compared to 15–98 (Figure S5A) supports this conclusion as it reports on the restricted motional freedom at the center of the helix (residue 90) relative to the helix end (position 98).

Populations of extended and compact states

For 15–90 and 15–98, the DEER-derived distance distributions feature a relatively narrow contribution centered around 21.5 and 23.3 Å, respectively, which matches the compact state seen in the crystal structure 1pin (Figure S8A). In addition, a longer, more dispersed distance is observed for these two constructs which is more akin to the domain distributions seen in the original, extended NMR structure 1nmv (Figure S8B). The accurate description of these distance distributions by two Gaussian components allows us to determine the populations of the two. Based on the 95% confidence interval of the population value for the shorter, compact state that extends from 0.67 to 0.72 for 15–90 and from 0.72 to 0.77 for 15–98 (Table S3), we propose that the populations of the compact and extended states are ~70% and ~30%, respectively. We note that since each measurement is carried out independently, we cannot determine from this data alone if the major/minor populations for these two mutants are correlated. Whereas involved approaches have been proposed to achieve this⁷⁰, the correlation in our specific case will become obvious once we calculated structural ensembles (vide infra).

Furthermore, the DEER distributions were used for cross-validating the multi-state structure calculations based on NOEs, RDCs, and PREs.

Major conformation: compact or extended?

We should note that the only other study that quantified the populations of apo Pin1 reported near opposite populations with 71% for the extended state and 29% for the compact state⁵⁰. This study was based on a chemical shift correlation analysis and on small-angle x-ray scattering (SAXS) data, though neither method directly measures distances between the two domains. It is possible that some distances are beyond the EPR detection limit >80 Å in the extended state (as we see with the 1nmv back-calculated DEER simulations in Figure S8B), but even when we consider that errors in distance are also expected to increase upon distance increase for a given DEER trace length it is unlikely that this contribution is large enough to completely interchange the populations. As seen in previous work^54,71, many NMR peaks in the interdomain interface have a major and minor species in slow-exchange. Our NOE and relaxation data associated with these slow exchange peaks suggests that the major peak originates from the compact state. For example, Figure S9 shows the major and minor peak of WW residue T29 in the ¹⁵N-HSQC, but with NOE cross peaks to PPIase atom 140 Hβ* (also atom 137 Hα, not shown) only in the major state. This observation implies the major state has interdomain contact, and thus corresponds to the compact conformation, while the minor state lacks on average the contact required to produce a NOE. In addition, the slow exchange peaks in the ¹⁵N-HSQC were sufficiently resolved to extract distinct R₁ and R_1ρ relaxation rates (Table S5). Based on the size and domain behavior of Pin1, more flexible regions are characterized by higher R₁ and lower R_1ρ rates, and the interdomain residues in the extended conformation are expected to be less restricted than in the compact conformation. Specifically, for the six interdomain residues with resolved major and minor peaks, the minor peaks are associated with higher R₁ and lower R_1ρ rates, implying occupation of the extended state. This data provides evidence that the compact conformation corresponds to the major state of apo Pin1, which further supports the DEER distance distributions.

Multi-state structure calculation

PREs as distance restraints

We converted the PRE relaxation rates into distances (see Supporting Information). Compared to the high precision eNOEs (tenths of Å versus 2–4 Å), PRE restraints correspond to longer distances and are associated with larger uncertainties. Because in CYANA both restraints are formally used in the same way, the PRE-derived distance restraints would dominate. Therefore, we firstly excluded the intradomain PREs from the structure calculations (for details on the structure calculations, see the Supporting Information), though the distances of the obtained structure correlate well with these intradomain PRE distances (y = 0.97x, r = 0.98, Figure 4B). Secondly, we optimized the weight of the interdomain PRE relative to the eNOEs restraints for a two-state structure calculation (as preliminary evidence suggested that two states are sufficient for satisfying the data, see Figure 6A) that involved all scalar couplings, angle restraints, eNOEs, and RDCs (as described earlier). Note that for this test our multi-state ensemble method enforces a 50:50 population, though the DEER data reports a major population difference between compact and extended states (this will be addressed later). We calculated five two-state ensembles with various PRE weights (equivalents of NOE weights 1, 0.5, 0.1, 0.01, and 0.001) and evaluated the impact of the different weights on the integrity of the local structure (Figure S10). As expected, lowering the weight reduces the target function (Figure S10B), however, the latter nearly levels off at a weight of 0.01, whereas PRE weights greater than 0.01 caused local structural disruptions near spin-labeled sites. This effect is most apparent in the long PPIase helix (Figure S10C), as a bulge formed near N90 in order to accommodate the PRE restraints to the WW domain. We evaluate the domain orientation by back-calculating DEER distance distributions from these ensembles using the EPR distance simulator within MMM⁷², and overlaying them with our experimental distributions (Figure S10A). While PRE weights of 1 and 0.5 pull the ID distributions into unrealistic conformations, these DEER distance distributions independently validate the use of a 0.01 PRE weight. All this data confirmed that weighting the PRE data by 0.01 (or reducing the weight by a factor of 100) is sufficient to run calculations that eliminate local artefacts caused by PREs overwriting local restraints. We therefore applied this weight in the following calculations.

Figure 4.

Domain positioning of two-state Pin1 structures. A) With the PPIase domain overlaid, positions of the center of the WW domain are shown color-coded by the interdomain restraints used to determine the domain orientations. B) PREs as cross-validation in the structure solved with RDCs and ID NOEs with experimental and ensemble error. C) RDCs as cross-validation in the structure solved with PREs and ID NOEs with ensemble error. D) Back-calculated DEER distributions from various structure calculations overlaid with experimental DEER distance distributions including the associated 95% confidence interval (CI). Distributions from individual two-state conformers are shown in Figure S8C in the Supporting Information.

Figure 6.

Coupling of inter- and intradomain spatial sampling of Pin1. The two-state ensemble is used for analysis. A) Extended and compact Pin1 with PPIase overlaid to show the relative position of the WW domain. B) RMSD between mean compact and extended states of domains versus residue number. Error bars show variance of conformers. C) Two-state overlay of WW domain. D) Two-state overlay of PPIase domain. Major RMSD changes from B are labeled on the structures. Major differences in interdomain interface are shown in the inset. The ten conformers with the lowest CYANA TF values are shown.

Reproduction of EPR-derived distance distributions by conformation-averaged NMR probes

It is not clear a priori whether the solution ensemble calculated from averaged NMR restraints will agree with the experimental DEER distance distributions and populations. Therefore, we tested if PRE, RDC, and ID NOEs can be combined to match the range of conformations present in the DEER distance distributions. The NMR restraints for the ensemble calculations are summarized in Table S6.

First, we optimized the calculation to ensure that our ensemble has converged by increasing the number of torsion angle steps and the number of structures calculated (Figure S10D). Each trial was started from 200 calculations and 50,000 torsion angle steps (as this number is ideal for small proteins) and resulted in similar TF of the ten two-state structures with the lowest values. Increasing the number of steps decreased the RMSD in the PPIase domain, while increasing the total number of calculated structures helped to lower the total RMSD in both domains. For the following calculations, we use 400 calculations each consisting of 100,000 torsion angle steps.

Secondly, we calculated two-state ensemble structures using our protocol that integrates all NMR probes (eNOE distance restraints, scalar couplings, PREs, RDCs, and ID NOEs). We compare the WW domain positions relative to the PPIase, as visualized in Figure 4A, by overlaying the associated MMM-simulated DEER distance distributions for the calculated structures (that did not include DEER data in the calculations themselves) with the distance distributions obtained by DEER experiments (Figure 4D). For an NMR ensemble, MMM will uniformly average the distribution over all conformers, and we show an example of the distributions from individual two-state conformers in Figure S8C. The difference in peak intensity is due to the individual conformers producing a narrow distribution while forcing the same population as the ensemble-averaged distributions. We obtained two distance clusters for 15–90 and 15–98. One narrower cluster is positioned at a short distance of around 18 Å for 15–90 and 21 Å for 15–98, while a broader contribution is centered at a longer distance of 38.5 Å for 15–90 and 46 Å for 15–98. Instead, for 15–131 we only obtain one relatively narrow distribution centered at around 44.5 Å. As mentioned in the previous section, we enforce two states of equal population so that by construction our two-state ensembles yield almost a 50:50 population when two clusters are present. Note that this population could be skewed if one population is much smaller so that both states are more likely to be assigned to the same cluster. Yet, our two-state structures place one state into each cluster (rather than both into the compact cluster), which further supports our population estimate of 70:30. Despite the described limitation of our model, the NMR ensemble reproduces the EPR distance distributions remarkably well. First, EPR also detects a bi-modal distance distribution for 15–90 and 15–98 though the cluster at shorter distance is more populated. For 15–90, the latter feature is narrower and centered at a larger distance (21.5 Å in EPR, 18 Å in NMR ensemble), while the second cluster is very broad and difficult to quantify in terms of center and width. Many interdomain orientations in our ensembles are also supported by a previous study⁵⁵. Secondly, the single cluster obtained for 15–131 is in almost perfect agreement with distance distributions obtained by EPR. We conclude that integration of PRE, RDC, and ID NOEs allows modeling the relative positions of the two domains, as the corresponding interdomain distances agree with DEER experiments, which are able to resolve multi-modal distance distributions.

The impact of different restraints on interdomain orientations

After we optimized our calculations, we cross-validated the different probes used in this study to determine the minimal data needed to restrain the two domains. For the two-state ensemble calculated without ID NOEs, the RDCs and PREs themselves restrain the WW domain along a wide plane that includes the compact position. When the PRE restraints are excluded (but the RDCs and ID NOEs are used), the conformation of the two domains is too extended as the back-calculated interdomain PRE distances are too long compared to the experimental PREs (Figure 4B). This is also mirrored in Figure 4 A and D that show that the compact state is not even present in this ensemble. In addition, all ID NOEs were violated in this calculation, suggesting that the ID NOEs alone are not sufficient to produce the compact state, and the PREs are required to generate this conformation. However, we expect that increasing the weight of the ID NOEs may also generate the compact state. On the other hand, removing the RDC restraints does not cause a relevant change in the ensemble because, as long as PREs and ID NOEs are used in the calculations, the relative positions of the two domains are nearly identical (Figure 4 A and D). Back-calculating RDCs from this structure results in a reasonably good agreement with the experimental RDCs for the individual domains (r_ww = 0.74 and r_PPIase = 0.65, Figure 4C), though the correlation coefficient of r_FL= 0.40 obtained for the full-length Pin1 is significantly smaller. Upon including the RDCs in the calculation, the correlations increase to r_ww = 0.89, r_PPIase = 0.87, and r_FL = 0.77. While the RDCs do not affect the relative global positioning of the two-domains, this result suggests that RDCs aid in subtly orienting the individual bond vectors in the internal domain structures. This is perhaps not surprising because RDCs are sensitive to rotation but not to translation. For the compact state, such rotational restraints may be more relevant, whereas the exact orientation of the extended state is less crucial and is possibly partially restricted by the presence of the linker. Overall, the PREs and ID NOEs are sufficient for determining the large-scale conformation of the two domains, while the RDCs support the eNOE restraints in correctly orienting the bond-vectors within the domains as previously demonstrated⁷³.

Two states are sufficient for describing interdomain positions

Combining all the PRE, ID NOE, and RDC restraints in addition to our eNOEs and angle restraints, we calculated multi-state ensembles of Pin1. We determined the minimal number of states needed to satisfy all the data by checking for convergence of the CYANA target function (TF, proportional to the sum of squared violations) without overfitting (Figure 5A). The TF and NMR restraint violations of the ten structures with the lowest TF values decrease significantly between the one- and two-state ensembles (TF reduced to less than one half), as reported in Table S6. Including a third and fourth state further reduces the TF by ~20% and ~10%, respectively. In addition, we also compared the simulated DEER distance distributions of these multi-state ensembles to the distributions obtained from experimental DEER distance distributions as shown in Figure 5D. Two-, three-, and four-state ensembles are all in reasonable agreement with the experimental DEER distances. At first glance, the single-state structure may appear similar to the crystal structure 1pin (Figure S8). However, the MMM simulations demonstrate that the 15–98 distance is too long compared to the major DEER peak in the single-state ensemble (~30 Å in ensemble vs. 23.3 Å in DEER), while the 15–131 distance is too short (~24 Å in ensemble vs. 44.5 Å in DEER). Further, although the contribution to the DEER distribution for 15–90 and 15–98 at longer distances (30–60 Å) is significant, it is not accounted for by the single-state ensemble. As described above, the two-state structure includes both a compact and extended state and matches the experimental data well. In addition, the two-state ensemble models achieve a superior agreement with experimental NOEs buildup intensities accurately compared to the single-state ensemble (Figure 5C, Figure S11). Although a single-state ensemble describes most NOEs well, some NOEs in regions of interest (i.e. WW binding site, interdomain interface, PPIase catalytic site) fit significantly better to a two-state ensemble, suggesting translational motion in these regions.

Figure 5.

Multi-state structures of various numbers of states. A) Number of states versus CYANA target function. B) CYANA target function of the two-state structure calculations versus various populations. C) Experimental unidirectional eNOE buildup intensities (green dots) versus time against back-predicted buildups of representative NOEs for single-state (orange) and two-states (black) ensembles calculated. D) Relative positions of the two domains in multi-state structures with back-calculated DEER distance distributions overlaid with experimental distributions with confidence interval (CI).

When we allow for more than two states, the ensembles feature a WW domain that tends to occupy conformational space comparable to the compact state, while the position nearby the PPIase catalytic site is not further populated (Figure 5D). This is in agreement with the independently determined compact/extended population ratio of 70:30. To further investigate this proposed population, we performed a pseudo three-state ensemble calculation, allowing only two distinct conformers, to determine if a 66.6:33.3 population provides a better fit to the experimental data than the 50:50 population of a conventional two-state calculation. We obtain a 3% increase in the TF for 66.6:33.3 compared to the 50:50 calculation, which is within 5 Ų (Figure 5B), Surprisingly, the TF increased significantly (19 Ų; 12% increase) when we ran a pseudo four-state calculation to test a 75:25 population. This suggests that the minimum is actually between 50:50 and 66.6:33.3. We note that our multi-state ensemble calculation is not optimized for population determination, in that the first step in the structure calculation only involves the WW domain. The 2:1 weighing should ideally be driven not only by intra-WW domain restraints, but also by inter-domain restraints. However, this is not possible with the current protocol. Given this limitation, together with the fact that the ideal population distribution is probably more equal than 66.6:33.3, we chose to carry out our structural analysis with the 50:50 distribution. However, we anticipate that the direct use of DEER distance distributions in the calculations would shift the minimum towards 66.6:33.3. Such a calculation would require a way to generate ensembles that reproduce this distribution across the states, while reproducing the averaged NMR restraints at the same time.

The interpretation of the sampling of our final two-state ensemble follows Occam’s razor approach, that is, we found the simplest representation of the spatial sampling that explains the data well. This is achieved by the minimal number of states (two in our case) and the minimal difference between these state within the domain (bundling restraints²¹ introduced by Clore and Schwieters and optimized for eNOEs⁷⁴ by us). By construction, the averages of each of the two states over the 10 lowest TF conformers yield the two states sufficient to explain the data, and the standard deviation of a specific state across the 10 conformers is then the uncertainty of the average state.

WW domain two-state ensemble lacks structural correlations

The two-state ensemble produces both compact and extended states (Figure 6A) and we rely on this representation to determine the correlations between intradomain structure and interdomain distance. Figure 6B shows the RMSD between the two states (averaged over the 10 calculations with the lowest TF) on a per-residue basis, with error bars representing the standard deviation within each state. In regions of the standard deviation exceeding the RMSD, structural differences are not likely to be important for changes in orientation. Figures 6B and C demonstrate that for the WW domain there are no clear differences between the two. Comparing the WW domain from this two-state structure calculation of the full-length protein to a similar calculation (solved using eNOEs and scalar couplings) of the isolated WW domain (pdb: 1svc)⁵², we observe many changes throughout the domain (Figure 7A). Most notably is the change at the ligand binding site, which may reflect the allosteric influence of the PPIase domain or the rearrangement may be caused by the S18N/W34F mutation introduced in the isolated system. Additionally, there are structural changes at the ID interface, which likely arise due to the presence of the PPIase. We also note that the very high data density for the isolated WW domain allowed for the resolution of two states, revealing structural correlations between the binding and ID interface sites. Because of the lower data density obtained for the full-length Pin1, we are unable to achieve such a resolution.

Figure 7.

A) Comparison of ensembles of the WW domain from full-length Pin1 and as an isolated system (pdb: 1svc⁵²). Two-state ensembles are shown with 10 conformers with the lowest CYANA TF for each state, but the individual states are not colored distinctly in order to draw attention to differences in the WW domain dependent on PPIase presence. B) ID interface residues in compact state. C) Comparison of catalytic sites of the two-state PPIase mean structures and the PPIase of crystal structure 1pin⁴³ bound to ligand (AlaPro peptide + SO₄⁻²). Regions of particular interest are labeled.

PPIase suggests mode of interdomain allostery

Structural changes of interest in the PPIase domain are labeled in Figures 6B and D, including the interdomain interface that involves residues 137–148. In the compact state, the side chains of residues A140 and L141 point into the interface, whereas in the extended state they are oriented into the PPIase (Figure 6D inset). The interdomain interface of both domains is mostly composed of hydrophobic residues, suggesting that the hydrophobic effect drives the formation of the compact state. In the extended state, the methyl groups of these residues are oriented towards the core of the protein, away from the interface. Steric clashing also appears to impact the interface, as upon overlaying the WW domains from our full-length structure and from the isolated system(1svc), the hydrophobic residues I28 and T29 of the latter appear to clash with the hydrophobic residues in the PPIase domain (Figure 7B). Hydrophobic-driven interdomain contact is supported by mutagenesis studies as mutations I28A and S138E reduce the hydrophobicity and cause Pin1 to adopt a more extended state^49,56,64,75.

By superimposing the secondary structural elements, we also observe major changes in structure at the loop connecting helices 1 and 2 (residues 98–102, Figure 6D), which indicates that this loop may act as a hinge point for these helices. In addition to the interdomain interface, the long helix 1 has previously been identified as an additional pathway for interdomain allostery in addition to the ID interface⁵¹. Finally, we note a striking difference between the two states in the catalytic site residues 152–154 (Figure 6D). We suspect that this loop repositioning could alter the conformation of the active site, which may lead to activity modulation. Comparing the mean PPIase structures from the two-state ensemble to the crystal structure bound to ligand (1pin) reveals that multiple loops in the catalytic site of the crystal structure are more similar to the extended state than the compact state (Figure 7C). The backbone RMSD between crystal structure and our NMR ensemble for residues 68, 69, 127–129, and 152–154 is higher for the compact state (1.14 Å) than for the extended state (0.97 Å). These regions are critical for catalysis, as residues 68–69 and the loop encompassing residues 127–129 are responsible for phosphate binding and fixing the C-terminal region of the isomerizing peptide, respectively⁶⁰. Furthermore, residues 152 and 154 are believed to be important for the hydrogen-bonding network of the catalytic site^76,77. The previously mentioned interface mutations that stabilized the extended state, I28A and S138E, also caused Pin1 activity to increase compared to WT ^49,56,64,75. Conversely, the mutation S138A is predicted to stabilize the compact state and it caused a reduction in activity. While this single-point mutation data supports our findings that the extended state has the higher activity in Pin1, another study that involved larger modifications to Pin1 showed no correlation between interdomain contact and activity⁵⁰. The structural similarity between our extended state and the ligand-bound crystal structure would suggest that the catalytic site is pre-organized in the extended state for isomerization even in absence of ligand. Similarly, it was shown that motions necessary for catalysis are an intrinsic property of Pin1, and these motions exist even without ligand present⁶⁰. Therefore, even in absence of substrate, Pin1 is primed for ligand binding and catalysis, and the correlated structural changes between the ID interface and the active site may be the driver of interdomain allostery.

Conclusion

We have introduced a novel approach that models a multi-domain protein in terms of the spatial sampling within each domain, the relative positioning of these domains, and the coupling of the inter- and intradomain sampling by relying on eNOE, scalar couplings, RDC and PRE data. In addition, we cross-validated the domain positions using distance distributions derived from DEER measurements and we have also utilized these EPR data to improve the accuracy of our model. We note that in general populations obtained from DEER may be altered upon freezing, but this is only the case to a small extent in the current study. In addition, populations determined under in vitro conditions may be shifted in physiological cellular environment. To address this issue, the use of molecular crowders in in vitro studies has been proposed⁷⁸.

Our multi-state ensemble of the two-domain protein Pin1 resolves its compact and extended states, which we independently observed in our experimental DEER distance distributions. In absence of ligands, our data supports that Pin1 slightly favors the compact state with a ~70% population. Furthermore, we were able to determine distinct conformational states within the hydrophobic interdomain interface that stabilize the compact and extended conformers, and we report structural correlations between the interface and the catalytic site. These concerted motions begin to describe how ligand binding in the WW domain alters interdomain contact and thus induces an allosteric change in the PPIase. Having a complete structural ensemble of apo Pin1 in hand, we can further elucidate the allosteric mechanism involved in ligand binding by evaluating the intradomain conformational changes that amend the interdomain dynamics. We expect that our method to generate a multi-state ensemble can be optimized and applied to other multi-domain proteins of interest.