Structural mechanism for signal transduction in RXR nuclear receptor heterodimers

A subset of nuclear receptors (NRs) function as obligate heterodimers with retinoid X receptor (RXR), allowing integration of ligand-dependent signals across the dimer interface via an unknown structural mechanism. Using nuclear magnetic resonance (NMR) spectroscopy, x-ray crystallography and hydrogen/deuterium exchange (HDX) mass spectrometry, here we show an allosteric mechanism through which RXR co-operates with a permissive dimer partner, peroxisome proliferator-activated receptor (PPAR)-γ, while rendered generally unresponsive by a non-permissive dimer partner, thyroid hormone (TR) receptor. Amino acid residues that mediate this allosteric mechanism comprise an evolutionarily conserved network discovered by statistical coupling analysis (SCA). This SCA network acts as a signalling rheostat to integrate signals between dimer partners, ligands and coregulator-binding sites, thereby affecting signal transmission in RXR heterodimers. These findings define rules guiding how NRs integrate two ligand-dependent signalling pathways into RXR heterodimer-specific responses.

T he nuclear receptor (NR) superfamily of transcription factors are broadly implicated in metazoan physiology, and modulate gene expression in response to steroids, lipids, bile acids and other small lipophilic molecules or synthetic ligands 1 . NRs harbour a C-terminal ligand-binding and transactivation domain (LBD), a central DNA-binding domain and a variable N-terminal disordered transactivation domain. These receptors transduce signals from ligand binding in the LBD to regulate gene expression by recruiting co-regulator proteins that modify chromatin and the associated transcriptional complex 2 .
The physical mechanisms governing allosteric signalling between NR ligands and coregulator-binding sites remain poorly understood. Allosteric control of NR function is modulated by a number of factors, including cell type-specific co-regulators 3 , post-translational modifications 4,5 , DNA recognition elements [6][7][8] and NR heterodimer partners [9][10][11] . Understanding the complex allosteric signalling of NRs requires first dissecting the signalling mechanisms within individual domains and binding sites, which will facilitate understanding the more difficult questions related to inter-domain communication 12 . Structural studies have revealed mechanisms that direct communication between ligand and coregulator-binding sites within a single LBD 13,14 . The fully active LBD conformer is wellcharacterized [15][16][17] and its conformation is conserved within the context of the full-length receptor 18 . In its agonist-stabilized conformation, the C-terminal helix, helix 12 forms one side, while helices 3-5 form the other sides of a co-regulator-binding site called the Activation Function-2 (AF-2) surface. Some NR antagonists, such as tamoxifen or RU486, contain a pendent side group that physically relocates helix 12 out of the active conformation thus blocking co-activator recruitment 15,19,20 . More recently, we identified a fine-tuning mechanism for indirectly modulating helix 12 conformation, allowing NRs to direct a graded range of signalling outputs from partial to full agonist [21][22][23][24] . We have also defined a structural mechanism whereby graded agonists and non-agonists do not fully stabilize the conformational dynamics of the AF-2 surface 4, [25][26][27] . However, it is poorly understood how ligand binding to one LBD controls co-regulator recruitment to its dimer partner within a NR heterodimer complex.
A subset of NRs functions as heterodimers with retinoid X receptor (RXR), and thus provides a mechanism to integrate two distinct ligand signalling pathways 28 . In some contexts, RXR heterodimers can act as two independent signalling moieties 29 . However, allosteric phenomena between RXR and partner are not well-understood. First, some heterodimer partners, such as the peroxisome proliferator-activated receptor-g (PPARg), farnesoid X receptor and liver X receptor (LXR), are 'permissive' for RXR activity, where the heterodimer is strongly activated by ligands for either partner in the dimer 30,31 . However, the integration of signals varies with both receptor and ligand combinations, which can produce either additive or synergistic effects 32,33 . Second, RXR heterodimers that contain retinoic acid receptor (RAR), vitamin D receptor (VDR) or thyroid hormone receptor (TR), are 'non-permissive' for RXR as they generally do not respond to RXR ligands 34 , or do so only in certain contexts in the presence of the partner ligand 35,36 . The structural mechanisms that generate this spectrum of signalling outcomes are unknown.
Here we present comprehensive structural analyses of a 'permissive' (PPARg/RXRa) and 'non-permissive' (TR/RXR) heterodimeric complex, which defines how a non-permissive dimer partner allosterically silences RXR. Solution nuclear magnetic resonance (NMR) spectroscopy reveals a mechanism by which the liganded state of TR, but not PPARg, uniquely affects the conformational dynamics of RXR. A crystal structure of the TR/RXR heterodimer defines a structural mechanism for this silencing, which occurs through a sequence of conformational relays between the helix 11 pairs that constitute most of the dimer interface, transferred to a rotation of helix 5 in the core of the RXR LBD, leading to disruption of the adjacent co-regulator-and ligand-binding sites. This allosteric signalling pathway is further confirmed by NMR and hydrogen/deuterium exchange (HDX) mass spectrometry. Notably, analysis of other NR dimers reveals that these structural changes are part of an evolutionarily conserved energetic network, defined by a statistical coupling analysis (SCA) method 10 , where helix 5 functions more generally as a signalling rheostat that integrates signals with the dimer interface, ligand and coregulator-binding sites.

Results
Conformational dynamics control RXR permissiveness. The RXR agonist, 9-cis-retinoic acid (9cRA), stimulates transactivation of PPARg/RXRa (Fig. 1a), verifying PPARg as a permissive RXRa partner. However, 9cRA and another RXR agonist, LG100268 (LG268), have no effect on TRb/RXRa (Fig. 1b) and VDR/RXRa ( Supplementary Fig. 1), establishing TRb and VDR as non-permissive RXR partners. To gain insight into the structural basis for this heterodimer-specific signalling, we performed solution NMR on isotopically labelled RXRa LBD alone as a homodimer, and in complex with unlabelled PPARg LBD or TRb LBD as heterodimers. This analysis enabled us to specifically observe conformational effects in RXRa that result from ligand binding to its heterodimer partner.
NMR resonances corresponding to residues within the apo-RXRa ligand-binding pocket and AF-2 surface are missing or have broad linewidths. These regions exist as a dynamic ensemble of conformations, exchanging between two or more conformations in a molten globule-like state on the ms-ms timescale [37][38][39] . Binding of 9cRA to the RXRa homodimer stabilizes the ligand-binding pocket and AF-2 surface 37 (Fig. 1b,c), resulting in the appearance and sharpening of NMR resonances relative to apo-RXRa. Thus, the NMR-observed structural mechanism by which an agonist activates RXRa occurs by stabilizing an active conformation. That is, agonist binding quenches the ms-ms conformational dynamics of the apo-RXRa ligand-binding pocket and surrounding regions, including the AF-2 surface. This mechanism is also supported by HDX mass spectrometry studies, which demonstrate stabilization of the RXRa ligand-binding pocket and AF-2 surface on ligand binding [40][41][42] . This conformational activation phenotype, whereby the dynamics of the apo-NR LBD is affected (stabilized) by agonist binding, has been observed for PPARg 25,26,39 ,VDR 43,44 , constitutive androstane receptorc 45 and other receptors, indicating this may be a general feature for ligand activation of NRs.
To determine the mechanism through which PPARg acts as a permissive dimer partner, we performed differential NMR analysis by adding unlabelled apo-PPARg to 9cRA-bound isotopically labelled RXRa, with and without addition of the full PPARg agonist, rosiglitazone (Fig. 1b,d). NMR chemical shift changes in RXRa are observed on addition of apo-PPARg to the 9cRA-bound RXRa, consistent with complete formation of a heterodimer complex. Addition of rosiglitazone causes subtle but significant NMR chemical shift changes in RXRa (for example, S355 and G429 in the dimer interface) and only minor changes in NMR resonance linewidths for select residues. Thus, although heterodimerization with and ligand binding to PPARg perturbs the conformation of RXRa, neither of these events dramatically affects the ms-ms dynamics of RXRa. This is in contrast to what occurs with the non-permissive RXR partner, TRb, as detailed below.
To determine the mechanism through which TRb acts as a nonpermissive RXR heterodimer partner, we performed differential NMR analysis by adding unlabelled apo-TRb to 9cRA-bound isotopically labelled RXRa, with and without addition of the TRb agonist, T3 (Fig. 1b,e). In contrast to PPARg, addition of apo-TRb exerts a profound effect on the ms-ms conformational dynamics of 9cRA-bound RXRa, where a large number of agonist-bound RXRa NMR resonances revert to an apo-like NMR profile. NMR resonances that are destabilized-missing or have broad linewidths indicating increased ms-ms motion-correspond to RXRa residues in the ligand-binding pocket (for example, I324, G323, T328, G329, G341, G343 and S355), helix 11 (for example, G413 and G429) and other nearby regions such as helix 8 (for example, G368). Even more striking is that the addition of the TRb agonist, T3, re-stabilizes these agonist-bound RXRa residues by decreasing motion on the ms-ms timescale, resulting in a reappearance of NMR resonances for these regions. Notably, many of the missing NMR resonances in the apo-TRb/agonist-RXRa heterodimer correspond to residues in the apo-RXRa homodimer that are stabilized on binding 9cRA. Our NMR studies indicate that these residues are not affected by PPARg heterodimerization or ligand binding to PPARg, but they are significantly affected by TRb heterodimerization and ligand binding to TRb. In total, these data indicate that the mechanism through which RXRa is allosterically silenced by TRb but not PPARg, involves conformational dynamics on the ms-ms timescale.
Structure of the TRbT3/apo-RXRa complex. To further detail the structural mechanism by which TRb allosterically silences RXRa, we crystallized apo-RXRa in complex with TRb, T3 and a co-activator peptide derived from SRC-2. The anisotropic data set  ARTICLE was scaled to a resolution of 3.2-3.8 Å and refined to an R work / R free of 23.6/28.1% (Table 1). Consistent with other RXR heterodimer structures, TRb and RXRa interact via the conserved dimer interface, largely comprised of helix 11 in each monomer, with additional contacts from helices 8 and 10 (Fig. 2a). TRb adopts the active conformation when bound to T3, with helix 12 forming one side of the co-activator-binding site, allowing the docking of the SRC-2 peptide. RXRa displays an inactive conformation with no bound ligand or co-activator peptide while helix 12 docks at the AF-2 surface. Although the asymmetric unit of the TRbT3/apo-RXRa complex is a dimer, the crystal packing reveals a heterotetramer assembly ( Supplementary Fig. 2). Compared with other NR heterodimer structures, TRbT3/apo-RXRa displays a more extreme deviation from the C2 (180°) symmetry of the dimer. RXRa helix 7 forms an extensive hydrogen bond network with TRb helix 9 (Fig. 2a). However, the symmetryrelated RXRa helix 9 and TRb helix 7 are further apart by B3 Å, preventing this sort of interaction. Superposing TRb with an RXRa subunit in the homodimer structure clearly revealed that the overall LBD structure is highly conserved (Fig. 2b). However, superposing these two structures via the RXRa protomer of the dimer subunits revealed a dramatic shift in the dimer interface (Fig. 2c). The amino-terminal end of TRb helix 11 is oriented similarly towards RXRa, while TRb helices 7, 10 and the carboxyl-terminal part of helix 11 are substantially shifted. As discussed below, this altered dimer interface induces conformational changes in RXRa, accounting for its silencing by TRb.
Structural mechanism for silencing of RXR by TR. The TRbT3/apo-RXRa crystal structure revealed that the structural basis of RXR silencing is via an allosteric signal emanating from the middle of the dimer interface. Compared with other RXR dimers fully occupied by ligands-including the RXRa homodimer 46 , and permissive RXR heterodimer complexes with LXR 47 and PPARg 48 -our structure of the TRbT3/apo-RXRa heterodimer shows a marked shift in TRb helix 11 (Fig. 3a). This shift induces a rotation of RXRa helix 11, which is visualized by comparing our TRbT3/apo-RXRa heterodimer with the RXRa homodimer ( Fig. 3b-d) or with the apo-RXRa tetramer ( Supplementary Fig. 3). In the N terminus of helix 11, TRb T426 is shifted towards RXRa P423 (Fig. 3b), which shows a rotation away from TRb in the heterodimer structure (Fig. 3c, Supplementary Fig. 3a). Towards the C terminus of helix 11, TRb is shifted away from RXRa, leading to a further rotation of RXRa helix 11 to maintain van der Waals contacts between RXRa L430 and TRb helix 11 (Fig. 3d, Supplementary Fig. 3b).
Notably, the TRb-directed rotation of RXRa helix 11 in the TRbT3/apo-RXRa structure induces a corresponding rotation of the adjacent RXR helix 5, which in turn disrupts the active conformation of RXRa. In the active conformation of the RXR homodimer, W305 in helix 5 mediates contacts with the bound ligand, M454 in helix 12 and L276 in helix 3, which is part of the AF-2 co-activator-binding surface (Fig. 3e). In contrast, in the TRbT3/apo-RXRa structure TRb-induced rotation of RXRa helix 5 in the heterodimer provokes a clash with the active conformation of helix 12 that pushes both L276 in helix 3 and M454 in helix 12 away from W305 in helix 5 ( Fig. 3f,g). Importantly, rotation of helix 5 is not observed in the apo-RXRa homotetramer ( Supplementary Fig. 3c), and is thus not a consequence of the substantial shift in helix 3 that is observed in both the TRbT3/apo-RXRa and apo-RXRa tetramers, which is rather determined by the tetramer packing. The electron density map allows the clear visualization of the main chain rotation required for interpretation of this data ( Supplementary  Fig. 3d,e), and the rotation of helix 5 is significant ( Supplementary Fig. 4). Thus in the TRbT3/apo-RXRa structure the TRb-induced rotation of RXRa helix 11 and helix 5 disables the active conformation of RXRa.
Structural role of SCA co-evolved amino acids. Work from the Mangelsdorf and Ranganathan labs 10 identified a network of co-evolved amino acids that are energetically coupled and mediate allosteric signalling in RXR heterodimers. A SCA was used to identify a network of 27 amino acids that comprise an allosteric signalling network for communication between RXR and its heterodimer partner. Importantly, an extensive mutagenesis screen showed that mutation of residues in one molecule allosterically impacted ligand response (that is, permissivity) from the partner 10 , although the structural mechanism that drives this effect at the atomic level remained unknown.
Helix 5 lies at the core of the SCA network, and connects the dimer interface, the ligand-binding pocket and the co-activatorbinding site (Fig. 4a,b). This network includes residues in the core of RXRa that promote rotation of helix 5 and subsequent silencing of RXRa, including residues in helix 11 (for example, L425 and R426), helix 5 (for example, E307 and W305) and helix 3 (for example, L276). RXR R426A and W305A mutants, and the analogous mutants in a permissive RXR heterodimer partner afford a dramatic loss-of-function equivalent to helix 12 deletion 10 . However, while RXR E307A (helix 5) has a modest effect on function, its analogous mutation in a permissive RXR heterodimer partner blunts the permissive response with RXR ligand. The rotation of RXRa W305 observed in the TRbT3/apo-RXRa structure would directly impact ligand binding, thereby accounting for the lower affinity of the TR/RXR heterodimer for RXR ligands 34 . The importance of these residues is underscored by the transmission of helix 5 rotation to the co-evolved amino acid residues in helix 3 and helix 4 at the core of the co-activator-binding site (Fig. 4c). Our structural data suggests a model where this co-evolved network controls the rotation of helix 5, thus impacting the dimer interface, and the ligand-and co-activator-binding sites. In our TRbT3/apo-RXRa heterodimer structure, RXRa ligand binding would require RXRa helix 3 to move back into the agonist conformation, driving RXRa L276 on helix 3 towards RXRa W305 in helix 5, leading to a reversal of the TRb-induced rotation of RXRa helix 5. To determine if the conformation of these RXRa regions are affected by TRb heterodimerzation, we performed differential HDX mass spectrometry comparing RXRa in its homodimeric state versus heterodimerized to TRb. In the presence of T3, 9cRA and co-activator peptide, the secondary structural elements of TRb-bound RXRa that were protected from amide exchange centred around P423 in helix 11 and extended to L276 in helix 3, relative to RXRa in the homodimer ( Fig. 4d and Supplementary Table 1). The changes in HDX support our model where these regions direct allosteric signalling within the heterodimer resulting in the silencing of RXR by TR.
To further test the role of helix 5 rotation in connecting the dimer interface with helix 3, we introduced the RXRa helix 3 mutation L276V, as we hypothesized that the smaller valine residue at this site (Fig. 3f) would facilitate packing of helix 3 against the rotated helix 5, even in the presence of RXR ligand. With the wild-type TRb/RXRa heterodimer, 9cRA impaired T3-mediated induction of a TR-responsive luciferase reporter in CV-1 mammalian cells transfected with TRb/RXRa, shown by a highly significant effect of 9cRA in a two-way analysis of variance (ANOVA, F(1,42) ¼ 27; Po0.001). There was also a trend towards an interaction between the T3 dose and 9cRA terms, suggesting a potential effect of 9cRA on making T3 less potent (ANOVA, F(6,42) ¼ 2; P ¼ 0.071). (Fig. 5a). In contrast, the mutated TRb/RXRa L276V heterodimer showed a gain-of-function in response to T3 (10-versus 5-fold activation), (ANOVA, WT þ T3 versus L276V þ T3: L276V effect F(1,42) ¼ 67; Po0.0001) and this mutation abolished the inhibitory effect of 9cRA on TR activation (Fig. 5b). These gain-of-function results are consistent with our model, where the interaction between RXRa W305 in helix 5 and L276 in helix 3 contributes towards the allosteric silencing of RXR by TR.
Role of helix 11 C terminus in cross-dimer signalling. We also tested the role of helix 11 in modulating the allosteric signal. While our studies here point to roles for the N-terminal region of helix 11 in TRb and RXRa in rotating helix 5, we previously noted that the C terminus of helix 11 could be differentially positioned by distinct ligands, thereby controlling the packing of helix 12 in the agonist conformer. This modulation of helix 11  provides the structural basis for partial agonist activity on NRs, by titrating the dynamics or stability of helix 12 as it forms the active conformer 22,23 . Above we also noted that RXRa E434 can form hydrogen bonds across the dimer interface with some heterodimer partners 13   be positioned to potentially interact with TRb S437 in helix 11. We previously demonstrated a role for this hydrogen bond in signal integration between RAR/RXR heterodimers 21 . Mutation of RXRa E434 to asparagine induced a gain-of-function in response to T3, both in drosophila SL2 cells that lack endogenous RXR and TR (Fig. 5c) and in CV-1 monkey cells (Fig. 5d). Thus, the C terminus of RXRa helix 11 plays a critical role in regulating the response of the TRb/RXRa heterodimer to T3.
To extend these results, we performed similar experiments with VDR, which is another heterodimer partner that silences RXR. Our HDX analysis of the full-length VDR/RXRa heterodimer with various combinations of ligands, DNA and co-activators established that vitamin D3 ligand induces stabilization of the same RXRa regions found here to be affected by TRb, including helix 3 and helix 11 (ref. 7). We therefore tested a series of RXRa E434 mutations in complex with VDR, and a mutation of the corresponding residue in VDR, K395 (Fig. 5e). Although RXR ligand has no activity on its own, within the context of the VDR/ RXRa heterodimer it is conditionally permissive because it enhances vitamin D3-induced transactivation. The helix 11 mutations selectively modulate the conditional activation by the combination of vitamin D3 and RXR agonists, and lead to both gain-and loss-of-function. Taken together, these data suggest that similar helix 11-mediated mechanisms control allosteric signalling across the dimer interfaces of TR/RXR and VDR/RXR heterodimers, and that there are several mechanisms for heterodimer signal integration.
Ligand signalling in PPARc/RXRa employs SCA network. To determine if ligand-selective signalling can occur between LBDs in RXR heterodimers, we took advantage of our extensive structural and chemical biology efforts with PPARg to compare ligands that produce different signalling outcomes or graded receptor activation 4,[25][26][27]49 . We performed NMR analysis using isotopically labelled RXRa and unlabelled PPARg to observe conformational changes in RXRa resulting from ligand binding to PPARg. We added several PPARg ligands to the apo-PPARg/9cRARXRa complex ( Supplementary Fig. 5); including a full PPARg agonist (rosiglitazone), a near full agonist (MRL20), a partial agonist (MRL24) and an antagonist/non-agonist (SR1664). Relative to the other liganded states, PPARg full agonists rosiglitazone and MRL20 caused notable NMR chemical shift changes for RXRa residues at the core of the dimer interface (Fig. 6a,b and Supplementary Fig. 6), including residues in helix 11 (for example, S427, I428, G429 and L430) and helix 7 (for example, T351 and K356). Other more modest NMR resonance shifts are observed in the RXRa dimer interface, including residues at the N terminus of helix 10/11 (for example, Q411 and G413). Importantly, the PPARg full agonist-induced NMR chemical shift changes for these residues at the core of the dimer interface were less prominent for the partial agonist and antagonist/nonagonist, suggesting these RXRa residues are structural sensors for PPARg ligand activity. Thus, the RXRa dimer interface responds to PPARg ligands in a manner that tracks with the ligand pharmacology, ranging from full agonist to non-agonist. NATURE COMMUNICATIONS | DOI: 10.1038/ncomms9013 ARTICLE A possible mechanism for this allosteric communication through the RXRa dimer interface involves the C terminus of PPARg helix 12. We previously demonstrated that PPARg agonists, but not partial agonists or antagonists, stabilize helix 12 of PPARg 4,25-27 . In the PPARg/RXRa heterodimer crystal structure 50 , RXRa K431 in helix 11 forms a hydrogen bond with PPARg Y477, the most C-terminal residue in PPARg. RXRa residues affected differently by the graded PPARg ligands are structurally close to this region. Thus, the effect of PPARg full agonists on RXRa residues in the dimer interface are likely mediated through stabilization of PPARg helix 12 and its interaction with RXRa.
However, structural perturbations also penetrate into other regions of RXRa, including the hydrophobic core, ligand-binding pocket and the AF-2 surface. This includes effects on RXRa core residues in helix 5 and helix 7 (for example, G304, E307, A311 and L353) ( Fig. 6b and Supplementary Fig. 6). One of two tryptophan residues in the RXRa LBD core (W305 or W282) is also affected. Helix 5 sits between helix 11 and helix 3 at the nexus of the ligand-binding pocket and the AF-2 surface. The perturbed residues provide a direct connection to structural changes in helix 3, the AF-2 surface and the ligand-binding pocket ( Fig. 6c and Supplementary Fig. 6). This includes small, but notable changes in RXRa helix 3 residue L276, as well as AF-2 surface residues K284 and S290, which are part of a region that forms electrostatic interactions with the bound co-activator peptide in the crystal structures. Thus, RXRa helix 5 is a key part of the LBD core that transmits PPARg ligand-induced allosteric signals from the dimer interface to the RXRa ligand-binding pocket and the AF-2 surface.
An additional region of the RXRa LBD core, helix 9, also mediates PPARg ligand-induced allosteric signalling across the dimer interface in a similar direction. Helix 9 forms part of the core that contacts the dimer interface, and stabilizes the AF-2 surface via interaction with helix 3 and helix 4 residues. PPARg ligand-induced NMR chemical shift changes in RXRa helix 11 (for example, Q411 and G413) and helix 9 (for example, G368), which lie in this region, suggest that helix 9 may also transmit allosteric information from the dimer interface to the AF-2 surface. Additional NMR resonances showing specific changes in response to PPARg ligands include A457 at the C terminus of RXRa helix 12 and A327 in the RXRa ligandbinding pocket.
All together, our NMR data reveal that binding of different ligands to its heterodimer partner, PPARg, can cause subtle but significant changes in the conformation of RXRa. Using HDX mass spectrometry, we confirmed that ligand binding to PPARg imparts structural changes in RXRa ( Supplementary Fig. 7 and Supplementary Table 2). These effects are not only present at the dimer interface, but also extend through the core of the RXRa LBD, to the AF-2 surface, helix 12 and the ligand-binding pocket. As we discuss below, these structural regions involve a network of co-evolved amino acids in NRs, which are energetically coupled and mediate allosteric signalling in RXR heterodimers.  Figure 6 | NMR reveals ligand binding to PPARc affects the conformation of RXR. NMR data are coloured grey for 9cRA-bound RXRa; black for 9cRA-bound RXRa heterodimerized to apo-PPARg or the same bound to the following PPARg ligands: rosiglitazone (magenta), MRL20 (blue), MRL24 (orange) or SR1664 (green); plotted on PPARg/RXRa (PDB 1FM9). (a) NMR data (left) focusing on residues in RXRa helix 7 and helix 10/11 dimer interface that are perturbed by ligand binding to PPARg, which are plotted onto the PPARg/RXRa crystal structure and coloured according to structural location (yellow for helix 7; blue for helix 10/11); coloured dark if shown in the NMR data to the left or light if not. (b) NMR data (left) focusing on residues in core of RXRa that are perturbed by ligand binding to PPARg, which are plotted onto the PPARg/RXRa crystal structure and coloured red; and coloured dark if shown in the NMR data to the left or light if not. (c) NMR data (left) focusing on residues in RXRa helix 12, the AF-2 surface and the ligand-binding pocket that are perturbed by ligand binding to PPARg, which are plotted onto the PPARg/RXRa crystal structure and coloured according to structural location (green for AF-2/helix 12; orange for the ligand-binding pocket); coloured dark if shown in the NMR data to the left or light if not.
Conservation of allosteric signalling through helix 5. Our data support a model where TR-induced rotation of RXR helix 5 drives TR silencing of RXR, and where rotation in the other direction drives the inhibitory effects of 9cRA on TR via the dimer interface. We propose that the co-evolved amino acid network lies at the core of this allosteric mechanism, which is consistent with our mutagenesis screen showing that these co-evolved amino acids impact allosteric signalling in RXR heterodimers 10 . Indeed, when we calculated NMR chemical shift differences between PPARg as a monomer versus heterodimerized to RXRa 51 , structural perturbations were observed in the regions of the evolutionarily conserved residues in helix 11 and helix 5 induced by heterodimerization with RXRa (Fig. 7a).
If an evolutionarily conserved allosteric network directs helix 5 rotation, then it should manifest for other NRs. A comparison of the structures of the LXR homodimer with the permissive LXR/ RXR heterodimer in the presence of co-activator peptide also shows that RXR induces a shift in LXR helix 11 that is transmitted through helix 5, again via the co-evolved amino acid network (Fig. 7b,c). This shift accommodates a flip of LXR W443 in helix 12 into a position against I295 in helix 5 (Fig. 7c) with additional van der Waals interactions and greater buried surface area, thereby stabilizing LXR helix 12 in the agonist conformation. Thus, RXR-induced rotation of LXR helix 5 is also the mechanism through which RXR drives co-activator binding to apo-LXR, a previously unexplained allosteric phenomena in heterodimer signalling 52 . This mechanism is also operational for heterodimers versus monomers of the RAR and the constitutive androstane receptor (Supplementary Fig. 8), supporting a general role for helix 5 rotation in allosteric control of RXR heterodimers. Thus, helix 5 rotation and the evolutionarily conserved SCA network of amino acid residues provide a structural conduit for signalling from the dimer partner, through helix 11, to the ligand-binding pocket and co-regulator-binding surface.

Discussion
While originally conceived as an on/off switch in transcriptional regulation, it is now clear that NRs contain a number of allosteric fine-tuning mechanisms that allow a full range of graded signalling outcomes. NRs can be viewed more generally as dynamic scaffold proteins, where post-translational modifications and interaction with ligands, co-regulators and DNA modify the nature of the scaffold and the signalling outcomes 5,11 . A large body of work has described functional interactions between NR domains and these interacting molecules, which in sum define a NR signalling code 2,14 . Several studies have investigated various aspects of permissiveness in RXR heterodimers 33,[53][54][55] . However, most of the structural features for allosteric signal integration have remained a mystery, limited in part by our insufficient structural understanding of signalling within the individual domains.
Here we used a variety of structural and functional approaches to show how the dimer partner controls the permissivity, or activity, of RXRa in the integration of two distinct ligandregulated receptors into a single transcriptional response using residues comprising the SCA network (Fig. 8a). Using NMR and crystallography, we show that structural differences in RXRa affected by the different dimer partners, TRb and PPARg, initiate distinct allosteric signals that suppress or permit modulation of heterodimer activity through RXRa. These signals are transmitted through amino acid residues including the co-evolved network previously identified by SCA 10 . Our NMR data reveal that in the presence of RXR agonist, dimerization with apo-TRb triggers a considerable change that reinstates an RXRa conformation that exchanges between two or more conformations on the ms-ms timescale. When compared with activated RXR homodimer 46 and permissive RXR heterodimer structures, LXR/RXR 47 and PPAR/ RXR 48 , our TRbT3/apo-RXRa crystal structure suggests that dimerization with the non-permissive partner TRb rotates RXRa helix 11, twists helix 5, drives open the ligand-binding pocket and induces the apo conformer of RXRa (Fig. 8b). This may account for the observation that TR lowers the affinity of RXRa for its ligand 34 . However, this conformational relay mechanism also operates in reverse, mediating the intrinsic activity of the apo-LXR/RXR heterodimer 52 . Here binding of RXR to LXR leads to a compensatory rotation of LXR helix 5, which directly stabilizes LXR helix 12 in the active conformation. Our studies reveal helix 5 as a central locus for allosteric control between the dimer interface, helix 12, ligand-binding pocket and AF-2 surface. The helices that comprise the SCA network act like a set of interlocking gears to integrate information from the functionally important sites in the NR LBD.
We identified a number of distinct routes for signal transduction through the dimer interface. For example, in our NMR studies of ligand-selective signalling in PPARg/RXRa, only full agonists of PPARg, which stabilize helix 12 in the agonist conformation, induce significant alterations in the RXRa dimer interface adjacent to the C terminus of PPARg helix 12. In this way, RXRa is able to 'feel' the position of the partner helix 12 and the degree of partner agonist activity. These types of cross-dimer interactions may also help stabilize helix 12 of the heterodimer partner in the active conformation, as previously suggested 48 . A second set of RXRa regions affected by all of the various PPARg liganded states include the hydrophobic core, helices 8-10 and helices 3-4 of the AF-2 surface (Fig. 8c). These regions in general employ the network of co-evolved residues predicted by SCA. Of these, PPARg full agonists appear to cause a more prominent effect, but the specific role of this structural conduit is not clear. It could mean that PPARg full agonists may provide additional stabilization to the RXRa AF-2 surface, or alter the shape of the AF-2 to give preferences for certain co-activators. Our NMR data further suggests that this interlocking relay system is also modulated by the ligand, as one of the two tryptophan residues in the RXR LBD, W305 in helix 5 or W282 in helix 3, was differentially sensitive to PPAR ligands. We thus envision that structural elements in helix 5 of RXR and the dimer partner can move in a coordinated way with the C-terminal region of the helix 11 dimer interface to coordinate both receptor-and ligandspecific signals into an integrated transcriptional response with the co-evolved amino acids playing a primary role.
We identified the C terminus of helix 11, adjacent to the bound ligands, as also contributing to heterodimer signalling. Within each monomer, there are two known mechanisms through which different ligands can produce a range of signalling outcomes from full agonist to antagonist. One is by direct modulation of helix 12, where the physical contact between ligand and helix 12 determines the percentage of time the active conformation of helix 12 is stabilized, docked across helix 3 and helix 11 to form the AF-2 surface. A second mechanism-indirect modulationoccurs when the ligand can position helix 11 so as to provide suboptimal van der Waals packing with helix 12, and thus indirectly control its stability in the agonist conformation [21][22][23][24] . Our data suggest an extension of this model where the position of helix 11 is also controlled by the heterodimer partner helix 11. Our mutagenesis data further suggests that the C terminus of helix 11 is also positioned by the type of dimer partner, in addition to the specific ligand, contributing to permissive versus non-permissive heterodimer signal integration.

Methods
Protein expression, purification and ligands. Human RXRa LBD (amino acids 223-462), human TRb LBD (residues 202-461) and human PPARg LBD (residues 203-477; isoform 1 numbering) were cloned into a pET vector with a ligationindependent cloning site as TEV-cleavable hexahistidine-tagged (His-tag) fusion proteins. RXRa LBD was induced in BL21(DE3) cells, and purified with immobilized nickel affinity chromatography. The eluted protein was mixed with a 1:30 ratio (by mg weight) of His-tag TEV protease and dialysed overnight in 20 mM Tris pH 8, 50 mM NaCl, 50 mM b-mercaptoethanol and 10% glycerol. The protein solution was again purified using immobilized nickel affinity chromatography to remove uncut protein, the cut His-tag and the TEV protease. The flow through was diluted 2 Â in H 2 O and subjected to gel filtration in buffer consisting of 20 mM Tris 8.0, 50 mM NaCl, 10% glycerol and 5 mM b-mercaptoethanol. TRb LBD was induced in BL21(DE3)Rosetta cells, and purified with immobilized nickel affinity chromatography (Qiagen) in a manner identical to that of RXRa. For crystallography, purified TRb LBD that had not been subjected to TEV proteolysis was incubated with purified RXR lacking a His-tag (at a ratio of 2:1 RXR to TR). The complex was purified with immobilized nickel affinity chromatography using showing how TR structurally silences RXR. The signal that emanates from TR (i) induces a shift in RXR helix 11 (ii), leading to a rotation of helix 5 (iii) resulting in structural arrangements that cause RXR helix 12 to adopt an inactive conformation (iv). (c) Summary of residues affected in the NMR analysis of ligand-selective signalling in PPARg/RXRa, plotted on PDB 1FM6. Helix numbers are indicated for elements of interest. Arrows indicate the flow of the allosteric signal. buffers and gradients as described above and then the TR LBD His-tag was removed by proteolysis with TEV protease overnight while the complex was being dialysed to 20 mM Tris 8.0, 150 mM NaCl, 5 mM BME and 10% glycerol. PPARg LBD was expressed and purified using similar methods, and final NMR sample conditions contained 20 mM KPO4 (pH 7.4) and 50 mM KCl 56 . Ligands were purchased from commercial sources, or in the case of MRL20, MRL24 and SR1664 were synthesized 27,57 .
NMR spectroscopy and analysis. NMR data were collected at 298 K on a 700 MHz Bruker NMR instrument equipped with a conventional TXI triple resonance probe and on a 800-MHz Varian NMR instrument equipped with a cryogenically cooled triple resonance probe. Ligands that were added to proteins were dissolved in DMSO-d 6 . NMR experiments were performed using pulse sequences and standard experimental parameters provided with Bruker Topspin 3.0. RXRa LBD chemical shift assignments 37 were validated and/or transferred to various complexed states using standard 2D and 3D NMR TROSY-based methods, including HSQC, HNCO, HNCA, HN(CO)CA and HN(CA)CB and 15 N-NOESY-HSQC experiments. Data were processed using Bruker Topspin 3.0 or NMRPipe 58 and analysed with NMRViewJ 59 . NMR chemical shift perturbations (DdCSP) for PPARg LBD in the monomer form and heterodimerized to RXRa LBD were calculated from published values 51 as follows: DdCSP ¼ |DdHN| þ (0.154 Â |DdN|) þ (0.341 Â |DdC 0 |); with DdHN, DdN and DdC 0 as the backbone 1 H N , 15 N and 13 C 0 (carbonyl) NMR chemical shift differences between monomer and heterodimer, respectively, and mapped onto the PPARg LBD crystal structure (PDB 2PRG).
Crystallization, structure determination and refinement. The TRbT3SRC-2/ apo-RXRa complex was formed by mixing a threefold molar excess of T3 (Sigma) and SRC-2 peptide (HKILHRLL). Crystal trials were initially conducted using commercially available sparse matrix screens, from which microcrystals were identified and subsequently streak seeded to produce a high resolution diffracting crystal. The well solution consisted of 20% PEG 3350, 100 mM Tris pH 8.0, 0.1 M ammonium acetate and 1 mM 11-Methoxy-3,7,11-trimethyl-2E,4E-dodecadienoic acid. The crystal was flash cooled with liquid nitrogen after briefly immersing in paratone-n as a cryo-protectant. Data was collected at SSRL beamline 11-1 at 100 K. The structure was solved with molecular replacement using Phaser 60 using search models for TR (PDB 3GWS) and RXR (PDB 1G5Y). Initially scaled at 3.5 Å, the model was refined using PHENIX 61 to an R free of 33%. The data set was anisotropic and was therefore rescaled to 3.2 Å and truncated to 3.2 Â 3.8 Å based on a 1.5 signal-to-noise (s) cutoff using the UCLA Diffraction Anisotropy Server 62 . This allowed the refinement to lower the R work /R free to 23.6/28.1%. The reported Rmerge for our structure is from the pre-truncated data set and was not used to determine the resolution cut-off. Instead, an anisotropic cut-off of 1.5s was used based on recommendations of Brunger et al. 63,64 to avoid discarding reflections when working with low resolution structures and with modern refinement practices that can accommodate lower signal reflections. Thus, the very good R free and geometry statistics for this resolution likely reflect the use of higher resolution structures for molecular replacement. The TLSMD server was used to identify optimal TLS groups 65 . The model was rebuilt using COOT 66 . Structural figures were generated with CCP4MG 67 .
HDX mass spectrometry. HDX was performed using a fully automated in-house system 49,68 with some modifications. Briefly, protein samples are incubated with D 2 O-containing buffer at 4°C for 10, 30, 60, 900 and 3,600 s. Following onexchange, forward or back exchange was minimized and the protein was denatured by dilution with 25 ml of quench solution (0.1% v/v trifluoroacetic acid/TFA in 3 M urea). Samples were then passed through an immobilized pepsin column (prepared in-house) at 50 ml min À 1 (0.1% v/v TFA, 15°C) and the resulting peptides were trapped on a C 8 trap column (Hypersil Gold, Thermo Fisher). The bound peptides were then gradient eluted (5-50% CH 3 CN w/v and 0.3% w/v formic acid) across a 2 Â 50-mm C 18 HPLC column (Hypersil Gold) for 5 min at 4°C. The eluted peptides were then subjected to electrospray ionization directly coupled to a high resolution Orbitrap mass spectrometer; Exactive for TR/RXR or QExactive for PPARg/RXR (Thermo Fisher Scientific). For TR/RXR measurements, 4 ml of 10 mM protein was diluted to 20 ml of D 2 O buffer. Following the prescribed on-exchange interval, the reaction was quenched with a cold 3-M urea solution containing 1% TFA and 50 mM TCEP. For PPARg/RXR measurements, 10 mM of HIS-PPARg LBD protein (20 mM KPO4 pH 7.4, 50 mM KCl) in complex with 10 mM FLAG-RXR LBD (20 mM KPO4 pH 7.4, 50 mM KCl) was preincubated with 1:2 molar excess of compound. About 5 ml of protein solution was mixed with 20 ml of D 2 Ocontaining buffer (20 mM KPO4 pH 7.4, 50 mM KCl) to initiate on-exchange. Following on-exchange, forward or back exchange was minimized and the protein was denatured by dilution with 25 ml of quench solution (0.1% v/v TFA in 3 M urea). HDX values are the average of three individual on-exchange experiments acquired in a random order. HDX data analysis was performed with 'HDX Desktop' for TR/RXR samples and 'HDX Workbench' for PPARg/RXR 69,70 . Each HDX experiment was carried out in triplicate and the intensity-weighted average m/z value (centroid) of each peptide isotopic envelope was calculated. Datadependent tandem mass spectroscopy was performed in the absence of exposure to deuterium for peptide identification in a separate experiment using a 60-min gradient. Peptides with a Mascot score of Z20 were included in the peptide sets used for HDX.
Cell culture and luciferase co-transfection assays. CV-1 cells (ATCC) were maintained in DMEM (Invitrogen) with 10% FBS charcoal/dextran -treated (Hyclone). Cells were transfected using Fugene HD (Roche) with a DR-4 luciferase reporter with expression plasmids for RXRa and TRb. After 6 h, cells were passaged and transferred into 384-well plates. Ligands were added the next day and allowed to incubate overnight before processing for luciferase activity. An equal volume of Britelite (PerkinElmer) was dispensed and the luminescence was measured on an Analyst GT plate reader (PerkinElmer). Drosophila SL2 cells (ATCC) were maintained in Schneider Drosophila Medium (Gibco) containing 5% dextran-charcoal-stripped FBS and transfected at a density of 6,500 cells per well in 96-well plates by calcium phosphate co-precipitation. Co-transfection experiments included 50 ng of reporter plasmid, 20 ng of b-galactosidase expression plasmid, 15 ng of each receptor expression plasmid and PGEM carrier DNA to give a total of 150 ng of DNA per well of a 96-well plate. Cells were transfected for 8 h and were treated for 18 h before harvesting and determination of luciferase and b-galactosidase activity. Luciferase data were normalized to the internal b-galactosidase control and represent the mean of triplicate assays plus s.e. Fold induction values were calculated as ligand-induced relative luciferase units/control relative luciferase units and propagated errors were calculated. Experiments used human full-length RXRa, TRb, PPARg and VDR expression plasmids with the appropriate luciferase reporter plasmid 10 , including TREx2-luc (TR-luc), PPREx3-luc (PPAR-luc) or ADH-mSppx3-luc (VDR-luc). RXR mutant expression plasmids were generated using the Stratagene QuikChange Site-Directed Mutagenesis kit and verified by DNA sequencing. Statistical analyses were performed with Graphpad Prism.