Structural modeling of the AhR:ARNT complex in the bHLH-PASA-PASB region elucidates the key determinants of dimerization

Abstract

Elucidation of the dimerization process of the Aryl hydrocarbon Receptor (AhR) with the AhR Nuclear Translocator (ARNT) is crucial for understanding the mechanisms underlying the AhR functional activity, including mediation of the toxicity of environmental contaminants. In this work, for the first time a structural model of the AhR:ARNT dimer encompassing the entire bHLH-PASA-PASB domain region is proposed. It is developed by using a template based modeling approach, relying on the recently available crystallographic structures of two dimers of homologous systems in the bHLH-PAS family of proteins: the CLOCK:BMAL1 and the HIF-2α:ARNT heterodimers. The structural and energetic characteristics of the modeled AhR:ARNT protein-protein interface are determined by evaluating the variations in solvent accessible surface area, the total binding free energy and the per-residue free energy contributions obtained by the MM-GBSA method and the Energy Decomposition Analysis. The analyses of the intricate network of inter-domain interactions at the dimerization interfaces provide insights into the key determinants of dimerization. These are confirmed by comparison of the computational findings with the available experimental mutagenesis and functional analysis data. The results here presented on the AhR:ARNT dimer structure and interactions provide a framework to start analyzing the mechanism of AhR transformation into its functional DNA binding form.

Introduction

Dimerization of the Aryl hydrocarbon Receptor (AhR) with the AhR Nuclear Translocator (ARNT) protein, both belonging to the basic helix-loop-helix Per-ARNT-Sim (bHLH-PAS) family of transcription factors,^1,2 is a key step in the mechanism of AhR-dependent induction of gene expression.^3–5 In this mechanism, upon activation by ligand-binding in the cytosol, the AhR translocates into the nucleus and its dimerization with ARNT results in the release of the chaperone heat shock protein 90 (hsp90) and associated proteins and conversion of the AhR into its high-affinity DNA binding form. Then, binding of the ligand:AhR:ARNT complex to the specific DNA recognition site, the Dioxin Responsive Element (DRE), stimulates expression of adjacent genes and production of a wide range of biological and toxic effects.^3–5 The AhR has been subject of continuous research efforts for more than forty years due to its role in mediating the biochemical response to xenobiotics and the toxic effects of selected environmental contaminants such as halogenated aromatic hydrocarbons (including halogenated dibenzo-p-dioxins).^6–8 More recently, the AhR has attracted renewed interest since the discovery of its regulatory role in a variety of endogenous developmental and immune response processes.^9,10 Elucidation of the AhR:ARNT dimerization mode is crucial for understanding the molecular mechanisms underlying the functional activity of the AhR.

Dimerization of the AhR and ARNT proteins occurs at the N-terminal half of each protein, inclusive of the bHLH motif and the PAS domain (consisting of the PAS-A and PAS-B repeats).^11–13 While dimerization primarily involves the HLH and PAS-A regions, and these regions appear sufficient to allow transformation of the AhR:ARNT complex into its DNA binding form, deletion mutagenesis and DNA binding analysis not only revealed that the PAS-B domain is important for initiation of AhR:ARNT dimerization, but that it also appeared to be involved in the process in a modulatory role.^14–16 No X-ray or NMR-determined structure has been reported to date either for individual AhR domains or for AhR complexes, except for a recently solved AhR PAS-A homodimer structure.¹⁷ Therefore, the use of computational modeling is essential for the molecular understanding of the AhR structure and interactions.¹⁸

The different computational methods available for modeling protein-protein complex structures can be classified in two main groups.¹⁹ The first includes protein-protein docking, in which the complex is constructed by assembling the structures of the interacting components through an exhaustive search and selection of various binding orientation.^20,21 The experimentally determined structures of the individual protein domains are needed. When these are unavailable, models generated by structure prediction methods can be used; however, the docking accuracy is sensitive to the errors in the monomer models.²² Moreover, a major challenge of these methods is including conformational changes of the protein domains upon binding. The second group of methods is template-based modeling (TBM), which constructs structures of unknown targets by copying and refining the structural framework of other related protein-protein complexes with experimentally solved structure. The major advantage is that the structures of the monomer components are not pre-required; in addition, the models are based on complex templates which are already in the bound form that is expected to be structurally similar to that of the target. Starting from the simple extension of methods for homology modeling of single-chain proteins, a number of different TBM techniques have been proposed to create accurate models of protein-protein complexes, as reviewed in.¹⁹ A limitation in the use of TBM methods arises from the quite low number of experimental complex structures available in the PDB, and in most cases structure prediction protocols are based on fold recognition (threading) techniques.¹⁹

The growing structural knowledge on complexes of homologous proteins belonging to the bHLH-PAS family has currently made homology modeling the most appropriate technique to gain insight into the structure of the AhR:ARNT heterodimer. bHLH-PAS proteins show well-conserved domain structures, as well as similarities in their mechanisms of action: all the class I bHLH-PAS proteins (including the AhR) sense environmental signals and form dimers with class II systems (including ARNT) and the resulting heterodimers bind to specific DNA sites to regulate target genes.¹² The first information about putative dimerization modes of individual PAS repeats was derived from the experimental structure of the hypoxia inducible factor α (HIF-2α) PAS-B in complex with the ARNT PAS-B ^23,24 and by an AhR PAS-A homodimer structure.¹⁷ The recent availability of two X-ray structures, the murine circadian locomotor output cycles kaput (CLOCK) in complex with the brain muscle ARNT-like (BMAL1) protein ²⁵ and the human HIF-2α:ARNT heterodimer,²⁶ also provides insight into the whole architecture of bHLH-PASA-PASB dimers.

In a previous work,²⁷ we developed the first models of the individual AhR:ARNT PAS domain dimers by homology modeling, using the X-ray structures of the PAS-B and PAS-A dimers available at the time as templates.^17,24,25 Due to the differences in the reciprocal orientation of the domains in these depositions, we proposed alternative models for both the PAS-A and PAS-B AhR:ARNT dimers and identified the most reliable ones through analysis of the protein-protein interaction (PPI) interfaces, generation of a set of mutants on both the proteins, and evaluation of ligand-dependent DNA binding of the AhR:ARNT heterodimer mutants.²⁷

The aim of the present work was to build upon our previous modeling work to develop the first structural model of the AhR:ARNT dimer encompassing the entire bHLH-PASA-PASB region. Moreover, relying on the structural and energetic characterization of the complete protein-protein dimerization interfaces, we aimed to reveal the intermolecular interactions critical for dimer stabilization. Alternative homology models were developed and compared, based on the bHLH-PAS dimer structures of the CLOCK:BMAL1 ²⁵ and HIF2α:ARNT ²⁶ templates, that show different quaternary architectures. Validation of the proposed dimerization mode was obtained by comparison of the models with the available experimental mutagenesis data. The structural proposal here developed provides insights into the AhR:ARNT dimerization that will be crucial for future analyses of the molecular mechanisms of transformation of the AhR into its high-affinity DNA binding form.

Materials and Methods

Homology modeling

The sequences of the murine isoforms of the AhR and ARNT proteins (AhR: UniProt Q3U5D9, GI 123784256; ARNT: UniProt P53762, GI 341940591) were aligned with those of the template systems in the X-ray complexes CLOCK:BMAL1 (PDB code 4F3L) and HIF2α:ARNT (PDB code 4ZP4).^25,26 Based on the residues available in the selected template structures, the modeled region included the amino-acids: AhR: 34–384 and ARNT: 89–465, in the 4F3L derived model; AhR: 38–384 and ARNT: 98–464 in the 4ZP4 derived model. The secondary structure (SS) profiles, predicted by the PSIPRED web-server ^28,29 for the target systems and attributed by the DSSPcont algorithm ³⁰ for the templates, were taken into account in developing the alignment. The homology models of the individual PAS-A and PAS-B domain dimers, we previously developed and validated,²⁷ were adopted as additional structural templates.

The homology models of the AhR:ARNT bHLH-PAS dimer were built using MODELLER 9v8.^31–33 This software implements an approach to comparative modeling by satisfying spatial restraints, which are extracted from the known related structures and from their alignment with the target sequence. The models are obtained by optimization of a molecular probability density function by employing methods of conjugate gradients and molecular dynamics with simulated annealing.^31–33 The optimal structure of each AhR:ARNT dimer model was selected among 100 models generated by MODELLER according to the best value of the distance-dependent statistical potential (DOPE) score.³⁴ Inter-domain linkers that were not resolved in the crystallographic templates (corresponding to: AhR:88-96 and ARNT:146-151 residues, for the model based on 4F3L; AhR:88-111 and ARNT:143-158, 347–359 residues, for the model based on 4ZP4) were modeled with the ab-initio method included in MODELLER,³³ by imposing structural restraints in the regions with predicted SS elements. To remove bad contacts and adjust non-optimal lengths and angles, the selected models were subjected to energy minimization followed by a short MD simulation with the AMBER 14 software ³⁵ (see Electronic Supplementary Information, ESI, for details). The overall quality of the final models was assessed with PROCHECK,³⁶ which provides information about the stereo-chemical quality, and ProSA validation method,^37,38 which evaluates model accuracy and statistical significance with a knowledge-based potential.

Loop modeling

Missing residues in the PAS-A loops of the X-ray template structures (AhR FG loop, residues 174-209; ARNT FG, GH and HI loops, residues 228-259, 272-301, 315-334, respectively) were built for the AhR:ARNT model using the Rosetta all-atom de novo loop modelling method with the next generation kinematic closure (NGK) procedure.³⁹ This method is a hybrid loop-modeling strategy that combines an ab initio methods with a fragment based approach (see ESI for details). 1000 sets of loop models were generated; then the ensemble of models of each loop was clustered on the basis of the backbone structural similarity by using the Self Organizing Map (SOM) approach previously described.⁴⁰ The models representing the cluster medoids were selected as representative of the conformational variability of that loop.

Binding Free Energy and Energy Decomposition Analysis

The binding free energy (ΔG_binding) for dimer formation was calculated in implicit solvent by means of the Molecular Mechanics Generalized Born Surface Area (MM-GBSA) method implemented in the AMBER software package.^41,42 In this method the ΔG_binding is obtained as the sum of energy associated with complex formation in the gas-phase and the difference in solvation free energies between the complex and the unbound molecules. MM-GBSA calculations were performed on the basis of a conformational ensemble generated by MD simulations. Each energy component is determined by averaging over the contributions from all the conformers. The single-trajectory approach was selected, i.e. the conformational ensemble was extracted from the single trajectory of the complex, instead of the three-trajectory one (separate trajectories of complex, receptor and ligand).^41,42 In particular, the ensemble was sampled in the 1000 ps MD simulation performed for the homology model optimization. Details on the solvation terms are given in the ESI.

The major contributions to the binding free energy were extracted using the Energy Decomposition Analysis.^43,44 Briefly, the free energy was per-residue decomposed into interaction terms (covalent, electrostatic, van der Waals and solvation) that are used to build a matrix describing the residue-residue pair interactions in the protein complex. This matrix is then eigen-decomposed and the resulting main eigenvectors and eigenvalues are used to generate a simplified matrix (energy matrix) summarizing the most relevant stabilizing interactions within the protein structure.

Nomenclature adopted

The models are herein termed according to the PDB code of the template: the 4F3Lmodel, was based on CLOCK:BMAL1; the 4ZP4model, was based onHIF2α:ARNT. In agreement with the proposal of Teichmann and co-workers,^45,46 the following protein-protein interactions are defined as: homomeric, if they involve the same kind of functional domains (e.g.: PAS-A_AhR vs. PAS-A_ARNT); heteromeric, if they involve different functional domains (e.g.: PAS-A_AhR vs. PAS-B_ARNT).

Results and Discussion

Structural templates and homology models

The X-ray structures of the CLOCK:BMAL1 and HIF2α:ARNT complexes used as templates for modeling the AhR:ARNT dimer show remarkable differences in the quaternary architecture of the bHLH-PASA-PASB region, due to different spatial arrangement of the flexible inter-domain linkers (Fig. 1A). While in HIF2α:ARNT the domains of the bHLH-PAS class I protein, HIF2α, show mutual contacts to form a contiguous surface and the class II protein, ARNT, rotate and twist around the outer surface of the partner, in CLOCK:BMAL1 the domains of the class I protein, CLOCK, wrap around the contiguous domains of BMAL1.^25,26 Nevertheless, the two crystal structures show very similar homomeric interactions between the individual bHLH, PAS-A or PAS-B domains (Fig. 1B). This similarity emerges from the root mean square deviation (RMSD) values calculated on the Cα atoms, excluding the loop regions: 0.76 Å, 5.45 Å, 2.33 Å for the bHLH, PAS-A and PAS-B domains, respectively. The slight difference emerging from the superposition of PAS-A homomeric pairs is due to differences in the reciprocal orientation of the β-sheet and the N-terminal extra-domain A′ α-helix, at the dimerization interface. Such difference was already observed by comparing the X-ray structures of CLOCK:BMAL1 and the AhR PAS-A homodimer (PDB code 4M4X).^17,27

Fig. 1. Crystallographic structures of the bHLH-PAS dimers used as templates.

(A) Overall scaffold of the complexes. The class I bHLH-PAS proteins are colored in green, while the class II bHLH-PAS proteins are colored in magenta. (B) Superposition of the individual domain dimers in the HIF2α:ARNT (lighter colors) complex and the CLOCK:BMAL1 complex (darker colors). RMSD values are calculated upon superposition of Cα atoms.

To analyze the effects of the different quaternary arrangements observed in the structural templates on the dimerization mode and on the protein-protein interface characteristics of the AhR:ARNT complex, two alternative homology models were developed, each based on one template. The overall identity (similarity) with the template sequences are: 24% (45%) between AhR and CLOCK and 52% (74%) between ARNT and BMAL1; 30% (50%) between AhR and HIF2α (100% for ARNT). Information about the target-to-template alignments used for modeling are collected in Fig. S1 (ESI). This includes: the pairwise residue similarity scores; the SS elements with the nomenclature usually adopted for these domains; the regions for which no structural information is available from the template depositions, that were modeled by ab-initio methods (as detailed in the Material and Methods section).

The AhR:ARNT dimer models were generated according to the above alignments by the MODELLER software package (see Material and Methods section) and were further optimized as described in the ESI. Both the final models show a good stereo-chemical quality, as assessed by PROCHECK, with 85.5 – 84.5 % of residues found in the most favored areas of the Ramachandran plot and 1.2 – 0.3 % in the disallowed regions; the overall G-factors range from −0.88 to −0.93 for the 4F3L model and the 4ZP4 model, respectively. The overall Z-scores calculated with ProSA range from −4.77 to −4.99, within the experimentally determined values for protein chains in the current PDB (see Fig. S2, ESI).

In the full-length AhR:ARNT dimer models, shown in Fig. 2A, the overall arrangement of the six domains reproduces that observed in the corresponding template structure (compare with Fig. 1A). To identify the key intermolecular interactions involved in the dimer stabilization, we thoroughly investigated the structural and energetic characteristics of the dimerization interfaces. In this analysis the missing regions in the PAS-A domains of the crystallographic templates were replaced with the shorter topological equivalent PAS-B loops, according to the grafting strategy proposed in ²⁷. The de novo loop modeling was applied to these regions at a later time to analyze their interactions with the core dimerization interfaces.

Fig. 2. AhR:ARNT models and dimerization interfaces.

(A) Global view of the AhR:ARNT dimer models based on the HIF2α:ARNT template (4ZP4 model) and the CLOCK:BMAL1 template (4F3L model). The missing regions in the PAS-A domains of the templates were replaced by the topological equivalent PAS-B loops.27 AhR is colored in green and ARNT in magenta. The dimerization interfaces, as defined by variations of Solvent Accessible Surface Area (ΔSASA), are highlighted by representing residues belonging to the interface with blue spheres. The total interface can be divided in four individual PPI interfaces, depicted in more details in the panels, (B) subregion 1, (C) subregion 2, (D) subregion 3, and (E) subregion 4. In the B–E panels, residues belonging to the interface are shown as sticks using a color ramp from white to blue according to the increasing burial degree.

Dimerization interfaces

The dimerization interfaces were defined by evaluation of the variation in Solvent Accessible Surface Area (ΔSASA), using the POPSCOMP software ⁴⁷ and their characteristics are summarized in Table 1. For both the models, sequence identities with the templates at the interfaces are above the “twilight zone” threshold (30 – 40% identity) that was proposed to infer similarity in the interactions of protein-protein complexes.⁴⁸ Even though the dimerization interfaces provided from the two templates are defined by distinct patterns of residues, several overlaps are found in the structural alignment, highlighting a relevant portion of topological equivalent positions (as illustrated in details in Fig. S1, ESI). The interface root mean square deviation (I_RMSD ⁴⁹) values between each model and the related template structure are very low. According to the higher sequence identity with the HIF2α:ARNT template at the interface, the I_RMSD is particularly small for the 4ZP4 model.

Table 1.

Main features of the dimerization interfaces of the homology models

	4F3L model	4ZP4 model
# residues	279 [270]	248 [205]
interface areaa (Å2)	7046 [6950]	6536 [4756]
I_RMSD (Å)	3.10	1.16
Seq. identity (%)	39.2	54.4
Seq. similarity (%)	54.3	64.6
ΔGbindingb(kcal/mol)	−323.55 ±17.67	−325.70 ±16.30
ΔGbinding [ele]c (kcal/mol)	−362.66 ± 55.80	−428.39 ± 10.79
ΔGbinding [vdw]d (kcal/mol)	−502.58 ± 16.40	−619.24 ± 66.52
ΔGbinding [solv]e (kcal/mol)	541.69 ± 49.72	721 ± 58.98

Reference values for the crystallographic templates are depicted in square brackets.

^avalues of ΔSASA, calculated by the PopsComp method.⁴⁷

^bdimerization ΔG_binding, calculated by the MM-GBSA method:⁴² mean values and standard deviations in the 1 ns MD trajectories are reported. For comparison, the same values obtained in 10 ns MD simulations are −322.37 ±13.84 and −325.06 ±16.96 for the 4F3L model and the 4ZP4 model, respectively.

^cnon-bonded electrostatic energy contribution to ΔG_binding.

^dnon-bonded van der Waals energy contribution to ΔG_binding.

^esolvation energy contribution to ΔG_binding.

Globally, the extension and the shape of the dimerization interfaces characterizing the two models show some differences (Fig. 2A). On the other hand, noteworthy similarities emerge comparing the individual PPIs between the AhR and ARNT chains in four main subregions, that are shown in Fig. 2B–E. The residues that are most buried into the interface, i.e. that mainly contribute to the PPIs, are listed in Table S1 (ESI).

Subregion 1 is characterized by an intimate association of the bHLH α-helices, with the two protein chains intertwined each other (Fig. 2B). Together, AhR and ARNT define a crossed four-helical bundle where the most relevant part of PPIs involves the parallel arrangement of the pairs H1_AhR:H2_ARNT and H1_ARNT:H2_AhR. The two models show an impressive structural similarity, with a RMSD of 1.73 Å upon superposition on the Cα atoms, and share several buried residues at the interface (Table S1, ESI). The major structural difference regards the basic region of the bHLH motif, where the N-terminal portions of the H1 helices diverge forming the forceps able to interact with the major groove of the DNA responsive element, as described in the template structures.^26,50

Subregion 2 (Fig. 2C) does not show clearly defined boundaries, since it encompasses a continuum of interconnected PPIs that span from subregions 1 to 3. It includes the PAS-A homomeric interface, that is similar in the two models. However, the different length of the A′ helices (about 10 residues longer in the 4ZP4 model than in the 4F3L model) introduces noticeable differences in the heteromeric bHLH:PAS-A interfaces. In the 4F3L model, the two domains are oriented in a stretched fashion, with the inter-domain linkers of both AhR and ARNT chains involved in the PPIs. Conversely, in the 4ZP4 model these linkers describe a curl that renders the bHLH motif and PAS-A domains tightly packed. In such conformation the H2_AhR and H1_ARNT helices directly interact with the PAS-A domains. Accordingly, many of the residues contributing to the PPI interface of this subregion are different in the two models (Table S1, ESI).

Subregion 3 describes the heteromeric PPIs between the PAS-A and PAS-B domains (Fig. 2D). In the 4F3L model both inter-domain linkers are buried and participate in the interface between the PAS-A_AhR and the PAS-B_ARNT, while PAS-A_ARNT and PAS-B_AhR are in direct contact. Because of the most pronounced twist of the ARNT chain in the 4ZP4 model, subregion 3 is mainly defined by the PAS-A_ARNT:PAS-B_AhR interaction, while PAS-A_AhR and PAS-B_ARNT do not interact.

Finally, subregion 4 defines the homomeric interface between the PAS-B domains (Fig. 2E). The interfaces are very similar in the two models and resemble those we previously characterized for the individual PAS-B dimer model, where the HI loop in PAS-B_ARNT is accommodated into a hydrophobic groove defined by the E and F helices and the AB loop of PAS-B_AhR.²⁷ In both models, few residues have a high burial degree (Table S1, ESI). Subregion 4 seems to define an independent interface, isolated from the other subregions.

Binding free energy and Energy Decomposition analysis

To evaluate the overall stability of the AhR:ARNT dimer and the interaction determinants, the binding free energy (ΔG_binding) of the two models was calculated with the MM-GBSA method (see Material and Methods section), that has been widely used to analyze protein-protein complexes.⁴¹ Although MM-GBSA calculations can be performed based on single structures, we adopted the approach based on conformational ensembles generated by MD simulations, to consider a certain degree of conformational flexibility. Moreover, we selected the single-trajectory approach because it gives less noisy results than the three-trajectory approach, due to cancellation of intramolecular contribution, therefore allowing MM-GBSA analyses based on shorter simulations.^41,51 The length of the MD simulation required for accurate free energy estimates usually ranges from a few ps to several ns, depending on the specific system.^41,51 To assess the adequacy of conformational sampling in the short MD simulations of the AhR:ARNT complexes performed for model optimization, we analyzed the stability of the MM-GBSA ΔG_binding in these trajectories. The results indicate a normal distribution of the ΔG_binding values with very low standard deviation from the mean value (Table 1). The ΔG_binding distribution shows similar characteristics and very similar mean values when derived from 10 ns simulations (Table 1 and Fig. S3, ESI). Thus a stable trajectory was already obtained for both the modeled complexes in the shorter simulation time.

Interestingly, despite the domain arrangement in the two homology models of the AhR:ARNT complex show some differences, yet the global binding free energy (ΔG_binding) shows nearly identical values (Table 1). Decomposition of ΔG_binding in the electrostatic, van der Waals and solvation components suggests that in the 4ZP4 model the non-bonded interactions (mainly the hydrophobic ones) have higher stabilizing contributions although the interface area and the number of residues involved are fewer than in the 4F3L model.

The Energy Decomposition analysis ^43,44 was performed to identify and compare the most relevant residue-residue pair interactions in the two models. As can be inferred from the energy matrices shown in Fig. 3, most of the individual contributions to the ΔG_binding are similar in the two models, in terms of both topology and magnitude (the areas with high similarity are highlighted by blue circles in the figure), and only few stabilizing interactions (in the areas indicated by red circles) are typical of each one. In both of the models, the strongest interactions are in the subregion 2. In particular, the A′ helices in the PAS-A domains are deeply involved for their contacts with both the PAS-A β-sheet of the dimerization partner and the upper portion of the bHLH four-helical bundle. Subregions 1 and 4 have well defined areas in the energy matrices, that describe how their SS elements are coupled in the dimer. By contrast, the subregion 3 is characterized by several sparse spots with some differences between the two models, underlining the variable topology of the PAS-A/PAS-B linkers described above. While subregion 4 is clearly distinguishable in the energy matrices, its small extension provides limited contribution to the total binding free energy of the dimer.

Fig. 3. Energy Decomposition analysis of the AhR:ARNT models.

The energy matrices derived from the per-residue decomposition of the ΔG_binding value for the two models of the AhR:ARNT complex are shown in the upper (4ZP4 model) and lower triangles (4F3L model). External bars illustrate the positions of each domain (AhR chain in green and ARNT chain in magenta) and the related structural elements (helices in light grey; strands in dark grey). The energetic couplings are indicated by spots, and the areas with the most relevant residue-residue pair interactions are emphasized with colored circles and are numbered according to the subregion (1, 2, 3, 4, defined in the text and represented in Fig. 2B, C, D, E, respectively) in which they belong. The areas with the highest topological similarities of the relevant pair interactions between the two models are highlighted in blue, the ones with some differences between the two models are indicated in red.

The importance of the subregions 1 and 2 for dimer stabilization in both our models is in agreement with experimental observations on different deletion mutants of AhR and ARNT proteins. It was previously demonstrated that deletion of the bHLH motifs (in particular the four-helical bundle region) prevents the formation of the AhR:ARNT complex, while deletion of PAS-A domains strongly reduces the ability of AhR and ARNT to dimerize.^11,13 Poellinger and co-workers proposed that PAS-A domains are required for the AhR:ARNT heterodimerization and that their association could drive the correct spatial orientation of bHLH and PAS-B domains.¹⁴ Furthermore, it was shown that PAS-B_AhR deleted mutants are able to dimerize in a ligand-dependent manner making the AhR:ARNT complex constitutively active.^15,16

Comparison with experimental mutagenesis data

The AhR:ARNT dimerization capability has been extensively characterized through mutagenesis experiments ^{17,27,52–55} and several mutated positions that were shown to be critical for AhR:ARNT dimerization lie at the PPI interfaces of our models, as demonstrated by their ΔSASA values and contributions to ΔG_binding (Fig. 4).

Fig. 4. Mutants of AhR and ARNT proteins known to affect AhR:ARNT dimerization.

Central panel: list of mutants known to affect dimerization and histograms showing the ΔSASA and per-residue contribution to ΔG_binding of every mutated position in the 4ZP4 and 4F3L homology models. Left and right panels: mapping of the mutated residues belonging to the modeled PPI interfaces and contributing to ΔG_binding on the homology models. AhR is colored in green and ARNT in magenta; residues are shown as sticks and labeled.

Most of the mutations in the bHLH motif were inserted in the basic region in an attempt to identify those residues involved in DNA binding, but this region was not included in our model.^56–60 Recent co-IP experiments ⁵⁵ demonstrated that single and double mutations of three hydrophobic residues (L112, L132, V136) within the ARNT HLH region to charged residues compromised the stability of ARNT:AhR complex. In both our models, these residues lie at the dimerization interface and give significant contributions to the ΔG_binding (Fig. 4).

Our group ²⁷ and others ^17,52,55 have demonstrated by mutagenesis and functional analysis that several residues in the A′ helices of both AhR and ARNT (AhR: L116, A119, L120; ARNT: L167, I168, A171) as well as in the faced β-strands in the PAS-A of the protein partner (AhR: F260, I262; ARNT: A339) have a critical role in the AhR:ARNT dimerization. These results confirm the importance of these structural elements (in subregion 2) for PAS-A:PAS-A association and for the overall stability of the dimer observed in both our models. Further confirmation is provided by the ARNT:E163K and ARNT:S190P mutations in the PAS-A which were shown to be critical for the selective heterodimerization of AhR and ARNT.⁵² In our models these residues lie in regions where the ARNT PAS-A domain is in direct contact with the upper portion of the AhR bHLH motif (Fig. 4).

For other residues in the PAS-A domains that are at the boundaries of the subregion 3 interface, calculations indicate a limited stabilizing role in the 4F3L model (AhR: G227, F228) or in the 4ZP4 model (ARNT: D217, C265). Mutations at these positions were shown to affect AhR:ARNT dimerization.⁵² Another mutation in this subregion, ARNT:G341D, was found as critical for the dimerization by other investigators.^52,53 Even if a glycine cannot give relevant contributions to both the calculated ΔSASA and ΔG_binding, in both the models this position is near to a highly interconnected region involving the ARNT PAS-A/PAS-B interdomain linker and the upper portion of the AhR A′ helix, providing an explanation for the detrimental effect observed with a mutation at this position.

Most of the mutated residues in the PAS-B domains (AhR: Y316, I324; ARNT: F446, N448, E455, I458) are buried into the modeled PAS-B homomeric interface (subregion 4), as shown by their contribution to the ΔSASA. Mutations of each of these positions listed in Fig. 4 exhibited disrupting effects on AhR:ARNT dimerization,²⁷ thus validating the PAS-B:PAS-B dimerization mode described by both proposed models. The moderate contribution provided by these residues to the calculated ΔG_binding further reflects the limited role of PAS-B domains in dimer stabilization, as previously discussed (Fig. 3).

Other mutations that were shown to affect the AhR:ARNT dimerization (listed in Fig. 4) ^27,52,54 map far away the PPI interfaces we have modeled and characterized (AhR: I160, L218 and ARNT: L221, M267, V306, C308) and some of them lie in the solvent exposed surface of the dimer (AhR: A131, C216). These mutations could impact the AhR:ARNT dimerization due to long-range allosteric effects or they could affect interactions with other partners (e.g. coactivator proteins, AhR repressor, estrogen receptor, RelB, KLF6 and others ^45,61–63). These hypotheses remain to be tested.

Conformational variability of native loops in the PAS-A domains

The missing regions in the PAS-A domains of the crystallographic templates that align to the AhR FG loop and ARNT FG, GH and HI loops (see Fig. S1, ESI), were modeled with a de novo loop modeling approach,³⁹ as described in the Methods section. The four (4F3L model) or three (4ZP4 model) loop conformations representative of the ensemble of 1000 models developed in this way are shown in Fig. S4A (ESI).

Because of their extension (ranging from 19 to 35 residues) these modeled loops show a high conformational variability, with most of their conformations sampling the solvent exposed region. Few conformations the ARNT FG loop contact the AhR protein, providing a minor additional PPI interface (Fig. S4B, ESI). In the 4F3L model this loop faces the helices of the AhR PAS-A and PAS-B domains, while in the 4ZP4 model it is in contact with the central portion of the PAS-A/B linker. However, in both the models the FG loops does not introduce relevant variations in the ΔSASA profile, it seems to not affect the core dimerization interface, and it is not in contact with any of the mutated residues that were demonstrated to affect the AhR:ARNT dimerization.

Conclusions

In this paper, the structure of the N-terminal region of the AhR:ARNT dimer, inclusive of the bHLH, PAS-A and PAS-B domains, is modeled for the first time. In addition to the previously predicted PAS-A:PAS-A and PAS-B:PAS-B dimerization modes,²⁷ we elucidate the bHLH:bHLH homomeric interactions and the complex network of heteromeric PPI interfaces among the different domains. The proposed dimer structure is validated by the observation that most of the residues that have been shown to be critical for dimerization in extensive mutagenesis and functional analysis studies,^{17,27,52–55} lie at the modeled PPI interfaces.

Similarities in the interactions between the individual bHLH, PAS-A and PAS-B domains in the X-ray structures of the CLOCK:BMAL1 and HIF2α:ARNT complexes reflect on similarities between the homomeric interfaces of the models developed on the basis of the two template structures. Some differences emerge in the bHLH:PAS-A and PAS-A:PAS-B heteromeric interactions. However, the calculated MM-GBSA ΔG_binding values indicate nearly identical stabilities for the two modeled dimers, suggesting that the key contribution to the dimer stability derive from core PPIs that are conserved in the two models. Analysis of the relative contributions of the PPIs in the different subregions of the complex to the total ΔG_binding clearly indicates the regions that are mostly involved in dimer stabilization. In agreement with experimental observations,^11,13,14 the crucial role of the interactions between the bHLH and PAS-A domains are clearly apparent. Conversely, these analysis also confirms the limited role of the PAS-B:PAS-B interactions in the dimerization, as observed in experimental studies on PAS-B_AhR deleted mutants.^15,16

The role of the AhR and ARNT PAS-A loops that map in missing regions of the template structures has been also investigated by de novo loop modeling. These loops show a high conformational variability in the solvent exposed region of the dimer and, in both the models, they seem to not affect the core AhR:ARNT dimerization interface. However further studies on their dynamic behavior are required to investigate possible functional roles of some of them. In fact, it is becoming increasingly evident that either disordered loops or linkers connecting protein domains do not act merely as flexible connectors but may have a critical role in protein functional dynamics and in allosteric communication.⁶⁴

On the basis of the studies here presented, the AhR:ARNT dimer models derived from both the CLOCK:BMAL1 and HIF2α:ARNT templates offer a reliable base for future studies on the mechanism of AhR-dependent induction of gene expression. However, the commonality of the dimerization partner suggests that the HIF2α:ARNT quaternary arrangement may represent a more appropriate reference structure. The higher sequence identity and similarity of the AhR with HIF2α than with CLOCK in the entire N-terminal region as well as the higher identity and similarity of the modeled dimerization interfaces with those of the HIF2α:ARNT complex confirm this orientation. Moreover, the recently published crystallographic structures of two additional bHLH-PAS proteins (NPAS1 and NPAS3) in complex with ARNT ⁵⁵ suggest a similar overall architecture of ARNT heterodimers that involve class I partners. The above studies also indicate that the ARNT PAS-B domain can displays lightly different arrangements to accommodate its different dimerization partners. These new findings further support the models proposed here for the AhR:ARNT dimer in the bHLH-PASA portion and suggest that a peculiar arrangement of the ARNT PAS-B and the related PPI interfaces could characterize this complex. In the absence of crystallographic information, this hypothesis could be addressed in future studies based on this proposed AhR:ARNT model and focused toward elucidating the dynamic behavior of the flexible PASA-PASB linker and PAS-B domain of ARNT.

The structural model of the AhR:ARNT dimer proposed here provides an initial description of the system, and this is a necessary step in order to start analyzing the main events in the mechanism of AhR-dependent induction of gene expression. Such a model represents a pivotal starting point for future molecular investigations of the process of AhR transformation into its functional DNA binding form.

Future research directions will include theoretical predictions of the above mechanism based on extensive molecular dynamics studies, validated by experimental analyses. MD simulations have greatly contributed to the understanding of several biological mechanisms by elucidating the role of protein dynamics in ligand-protein and protein-protein interactions, in the differential effects of mutations, in allosteric signal transmission.^65–70 Our planned studies on the AhR:ARNT dynamics should provide a comprehensive picture of signal propagation across the complex, from ligand-binding in the AhR PAS-B domain to DNA-binding in the bHLH region of the AhR:ARNT complex. They will also further elucidate the dynamic features of both the inter-domain linkers and the loops that could affect the complex network of intermolecular interactions involved in the signal transmission process. These studies can also lead to investigations into the mechanisms responsible for similarities and differences in the mechanistic link between ligand-binding and transcriptional regulation of members of the bHLH-PAS protein family.