Abstract
Type-A γ-aminobutyric acid (GABAA) receptors are pentameric ligand-gated ion channels (pLGICs), typically consisting of α/β/γ subunit combinations. They are the principal mediators of inhibitory neurotransmission throughout the central nervous system and targets of major clinical drugs, such as benzodiazepines (BZDs) used to treat epilepsy, insomnia, anxiety, panic disorder and muscle spasm. However, the structures of heteromeric receptors and the molecular basis of BZD operation remain unknown. Here we report the cryo-EM structure of a human α1β3γ2 GABAAR in complex with GABA and a nanobody that acts as a novel positive allosteric modulator (PAM). The receptor subunits assume a unified quaternary activated conformation around an open pore. We also present crystal structures of engineered α5 and α5γ2 GABAAR constructs, revealing the interfacial site for allosteric modulation by BZDs, including the binding modes and the conformational impact of the potent anxiolytic and partial PAM, bretazenil, and the BZD antagonist, flumazenil. These findings provide the foundation for understanding the mechanistic basis of GABAAR activation.
The common architecture of eukaryotic pLGICs is now established, with at least one crystal or cryo-EM structure determined for the cation-selective nicotinic acetylcholine (nAChR) and serotonin type 3 receptors (5HT3R), and the anion-selective GABAA, glycine (GlyR), and GluCl receptors1–9. Each subunit extracellular domain (ECD) of 200-250 amino acids comprises an N-terminal α-helix and ten β-strands forming a twisted β-sandwich. Each TMD comprises an α-helical M1-M4 bundle, the M3 and M4 helices being connected by a large intracellular loop of 85-255 residues10. The two orthosteric neurotransmitter binding sites are situated between ECD principal (P) and complementary (C) faces, which intercalate around a water-filled vestibule that funnels into the ion channel transmembrane pore lined by five M2 helices. The structures have brought considerable understanding of the molecular basis of receptor function in homomeric receptor formats. However, heteromeric pLGIC structures have only been solved for nAChRs1,2. This is a key limitation given that the vast majority of mammalian pLGICs are heteromers possessing heightened complexity both molecularly and pharmacologically.
The GABAAR family comprises 19 different subunit subtypes: α1-6, β1-3, γ1-3, δ, ε, θ, π and ρ1-311. Selective neuronal expression of particular subtypes, along with preferential assembly rules, ensure that the majority of GABAARs in the human brain comprise 2 a-subunits, 2 β-subunits and 1 γ-subunit, with a1, β2, β3, and γ2 subtypes exhibiting widespread overlapping expression profiles12,13. Alternative subtypes perform more specialised roles, for example, the α5 subtype influences cognition14 and moderation of α5β2/3γ2 receptors ameliorates animal model diseases of autism and Down syndrome, associated with cognitive deficits15,16. Modulation of α5β2/3γ2 receptors also improves recovery after stroke in a rodent model17. The homomeric GABAAR structures so far solved, GABAAR β3 and chimaeras that include α subunit TMDs, possess five-fold symmetry and will possess five identical copies of agonist and anaesthetic binding pockets18,19. In contrast, aβγ GABAARs possess two orthosteric GABA binding sites at β-a interfaces between the ECDs, two non-GABA binding a-β and γ-β interfaces, and one BZD binding site located at the a-γ interface20. Similarly, within the TMDs, aβγ GABAARs have at least three non-equivalent potential anaesthetic binding pockets. In the case of GABAAR β3, solved in a desensitized state, each subunit is bound by agonist and the subunit conformations are indistinguishable5. However, it is not known if different subunits within a heteromer adopt equivalent conformations in the activated state.
PAMs, such as BZDs, the intravenous general anaesthetics propofol and etomidate, barbiturates, endogenous neurosteroids, and alcohol (ethanol), bind GABAARs to promote their activation and channel opening, and thereby enhance GABAergic signalling18–25. This makes BZDs essential treatments for hyper-excitability disorders such as anxiety, insomnia and epilepsy24. However, these agents lack receptor subtype selectivity and cause unwanted sedation, addiction, and motor and cognitive impairment26. Agents with improved selectivity can ameliorate these side effects and allow for more selective targeting to treat other neurological disorders, including autism, Down syndrome, neuropathic pain, schizophrenia and stroke15–17,27,28. Structures of heteromeric GABAARs are essential to understand the molecular basis of inhibitory neurotransmission and drug subtype selectivity, and are anticipated to expedite the delivery of selective therapeutics against these disorders.
RESULTS
Structures of GABAAR constructs
Our initial goal was to obtain high resolution structures of the α and γ human GABAAR subunits. In addition to the β subunit structure we previously reported5, these are the essential constituents of the principal heteromeric receptor subtypes in the CNS29. Moreover, the α/γ interfaces harbour the BZD allosteric site30–37, which is of major pharmacological importance. Iterative rounds of screening and engineering of α5 subunits for pentamer monodispersity and yield identified a construct with 12 residue swaps from the β3 subunit, which readily forms homopentamers5, and 11 residue swaps from the BZD site complementary (C) face of the γ2 subunit. This construct recapitulated 100 % residue identity within the BZD site to the wild type α5-γ2 subunit interface (construct designated α5HOM; site designated BZDHOM; Supplementary Fig. 1a-d)35–37. Of particular note, an Asn114Gly substitution (β3 subunit glycine) to remove an N-linked glycosylation site that faces the extracellular vestibule and restricts homomeric assembly of α-subunits (observed in the cryo-EM construct – discussed below) was crucial to achieving monodispersity and pentamerisation (Supplementary Fig. 1e). We solved its structure to 2.6 Å resolution in complex with the drug, flumazenil38 (Anexate), a BZD site antagonist used to treat overdosage (Fig. 1a, b and Table. 1). Notably, although α5HOM bound the BZD radioligands 3H-flunitrazepam and 3H-flumazenil, the affinity was 60-fold and 150-fold lower than for wild-type α5β3γ2 receptors (Supplementary Fig. 2a-e). To solve the structure of a BZD site with higher affinity we performed further screening and engineering on this background for heteromeric combinations containing γ2 subunits. These were transfected at 9:1 ratios of α5 to γ2 DNAs to favour one γ2 per pentamer. This identified a construct formed from α5HOM subunits containing 4 additional residue swaps from the β3 subunit, and a chimera subunit comprising the γ2 ECD and α1 TMD, designated α5γ2HET. We crystallised this construct in complex with the BZD partial PAM, bretazenil39 and solved its structure to 2.5 Å resolution (Fig. 1c, d; Table 1). In contrast to α5HOM, inclusion of the chimeric γ2 subunit resulted in an additional higher apparent affinity site, designated BZDHET (as well as the lower affinity BZDHOM sites), which had only 2-fold, 5-fold and 15-fold lower affinity for 3H-flunitrazepam, 3H-flumazenil and bretazenil respectively, compared to wild-type receptors (Supplementary Fig. 2a-f).
Attempts to crystallize tri-heteromeric GABAARs were unsuccessful, therefore we generated new constructs for single-particle cryo-electron microscopy. We designed α1 and γ2 subunit constructs in which only the M3-M4 intracellular loops were substituted by a short linker, SQPARAA40, and β3 subunits in which the M3-M4 intracellular loop was substituted by a thermostabilised apocytochrome b562RIL (BRIL)41 domain for unambiguous subunit identification. These constructs were co-expressed and assembled into functional α1β3γ2 GABAARs, designated α1β3γ2EM (Supplementary Fig. 3). Only the γ2 subunit possessed a purification tag, in order to guarantee its presence in the pentamers. To facilitate particle alignment, we raised nanobodies against GABAARs and selected one of these, Nb38, that binds to a1 subunits. We obtained a 5.17 Å (FSC=0.143) cryo-EM map of a1β3γ2EM, in the presence of a saturating 1 mM concentration of GABA (Fig. 1e-f and Supplementary Figs. 4 and 5; Table 2). Two Nb38 molecules were bound between non-adjacent ECD interfaces, consisting of a1(P) faces and neighbouring β3(C) and γ2(C) faces, respectively, validating a1 subunit inclusion and receptor stoichiometry. At this resolution, the TMD a-helices of a1 and β3 subunits showed helical turns and side-chain densities for several large hydrophobic amino acids, although β-sheets were not fully separated into strands (Supplementary Fig. 6). The N-linked glycans served as markers that further allowed unambiguous subunit identification. Mannose branching could be clearly seen, for example, at a1 Asn111 and β3 Asn149. The detergent belt, BRILs and nanobody edges were largely disordered (Supplementary Fig. 5g). The γ2 subunit TMD was also less ordered, but otherwise the local resolution for the rest of the pentamer was consistent throughout (Supplementary Figs. 5g and 6). Viewed from above, the cryo-EM map confirms a clockwise subunit arrangement of a1-A, β3-B, γ2-C, a1-D, β3-E, consistent with previous indirect studies42,43. Despite the low resolution of this map, availability of high resolution crystal structures of individual subunits from α5γ2HET and the previously solved GABAAR-β3cryst5 allowed us to build a model of the heteromeric α1β3γ2EM receptor (Fig. 1g; Table 2). Inclusion of BRIL domains between the M3-M4 helices of the β3 subunit did not distort these helices (RMSD = 0.98 Å between a1β3γ2EM β3 and GABAAR β3cryst across 54 M3-M4 equivalent Cα positions).
The BZD binding site and BZD binding modes
The ECD pockets between α5HOM and α5γ2HET subunits contained large positive peaks in the Fo–Fc electron density maps, which were unambiguously assigned to the co-crystallisation ligands flumazenil and bretazenil, respectively (Supplementary Fig. 2i-l). For α5γ2HET the electron density map revealed inclusion of a single γ2 chimera subunit per pentamer, distinguishable by its unique side chains and glycosylation sites. This confirms the transfection ratio of 9α5:1γ2 ensured a pure stoichiometric population, as required to enable crystallization, and for radioligand binding studies. The lower apparent affinity of the BZDHOM sites did not affect the ligand binding mode which was the same as at the BZDHET site (Supplementary Fig. 2m, n). Furthermore, the binding modes of both flumazenil (observed only in BZDHOM sites in α5HOM) and bretazenil are similar (Figure 2a-h). BZDs such as flunitrazepam (which lack the imidazo C(3)-linked ester moiety and instead possess a diazepine C(6) phenyl ring) exhibit a distinct pharmacology from flumazenil and bretazenil because they require a histidine residue in the BZD site of α1/2/3/5 subtypes for binding32,44. Radioligand binding on a modified α5γ2HET construct, α5γ2HETΔ, with BZDHOM sites removed (by reversing the γ2 substitutions in the α5 subunits), so that only the BZDHET site remained, revealed that a His105Arg substitution ablated 3H-flunitrazepam binding but retained the same apparent affinity for flumazenil (Supplementary Fig. 2g, h). Thus, the differential impact of the His to Arg substitution on the binding of distinct classes of BZDs to wild type receptors is reproduced in this construct.
Intermolecular binding contacts between the ligands and receptor occur predominantly through van der Waals (vdW) interactions, which span the interface between subunits. From the α-subunit (P)-face, five hydrophobic residues make vdW contacts with the ligand benzene A-rings: Phe103 and His105 from the β4 strand (historically named loop A10); Tyr163 from the β7-β8 loop (loop B); Ile206 and Tyr213 from the β9-10 hairpin (loop C) (Fig. 2 a-h and Supplementary Fig. 1). Of particular importance are Tyr163, which forms T-shaped π-stacking interactions with the A and I ligand rings, and Tyr213, forming parallel π-stacking interactions with the A rings. Consistent with such interactions, Tyr163Phe and Tyr213Phe substitutions retain BZD potentiation whilst Ser substitutions reduce sensitivity at least 10-fold33. The structures also explain why moving the flumazenil C(8) fluorine to C(9) or C(10) chlorine or adding a C(1) methyl substituent (Compounds 1-3 in Supplementary Fig. 7a-d) reduce binding by over 100-fold45–47. These substituents clash with loop B Tyr163, as shown by in silico docking studies (Supplementary Fig. 7a-d). The lowest free energy binding mode of flumazenil overlays the one observed experimentally in a5HOM, whereas compounds 1-3 are displaced by ∼2 Å away from Tyr163, impairing the T-shaped π-stacking interactions. In contrast, the C8 azide derivative Ro15-451348, a competitive BZD antagonist developed as an antidote to alcohol and more recently used as a PET ligand49, docked similarly to flumazenil, with the azide positioned under the loop C Tyr213, which it chemically photolabels50 (Supplementary Fig. 7e).
The loop-A His105 undertakes a ligand-dependent reorientation within the site (Fig. 2b-d, f-h), its side chain rotating about the Cα-Cβ bond to accommodate either the bretazenil bromine or the flumazenil chlorine atoms (at adjacent positions on the A rings, Fig. 2a, e). As shown by the radioligand data (Supplementary Fig. 6g, h), a His105Arg substitution does not affect flumazenil binding even though it ablates flunitrazepam binding. The explanation for this is observed in the γP/αC site in α5γ2HET, where the equivalent residue to His105 is γ2 Arg114, which is well accommodated under the bretazenil ligand (Supplementary Fig. 7i, j).
From the (C)-face, the β2 strand (loop D) phenylalanine side chain (γ2 Phe77 in the BZDHET site, γ2* Phe68 in the BZDHOM site) forms π-stacking interactions with the ligand I rings (Fig. 2c, g). These explain why a Phe77Ile substitution reduces flumazenil affinity 1000-fold and why GABAA receptors containing the γ1 subunit, where an Ile residue occupies the equivalent position, are much less responsive to BZDs36. Above Phe77, an alanine residue (γ2 Ala79 in the BZDHET site, γ2* Ala70 in the BZDHOM site) demarcates the top of the BZD site and faces the I ring C(3) substituent, consistent with a previous mutagenesis study predicting close apposition of these two elements51 (Fig. 2d, h). The neighbouring β6 strand (loop E) threonine (γ2 Thr142 in the BZDHET site, γ2* Thr133 in the BZDHOM site) forms putative hydrogen bonds with both the imidazole nitrogen and ester carbonyl of the ligands, consistent with original pharmacophore models proposing a dual H-bond contribution from these moities52 (Fig. 2d, h). The ester carbonyl group is essential for flumazenil binding versus ketone or ether functionalities46, whilst other derivatives that maintain the isosteric constraints retain high affinity binding52,53.
Patients on BZDs can experience a variety of adverse events, including cognitive and psychomotor effects, tolerance and, in some cases, paradoxical behaviours such as disinhibition leading to aggression26. To explore whether there is genetic variation within the BZD site, which may contribute to response variability in individuals, we mapped allelic variants of 138,632 unrelated healthy humans (gnomAD database54) for α1-6 and γ1-3 subunit residues within 5Å of ligand bound to α5γ2HET and α5HOM (Fig. 3a-c). Although no common allelic variants were found, many rare allelic variants (ranging from 1 in 1,000 to 1 in 100,000 people) were identified, spanning the four major physiological BZD site subtypes, α1γ2, α2γ2, α3γ2 and α5γ2. The α5 subunit Gly161Val African variant is predicted to sterically disrupt the folding of loop B reducing ligand-binding affinity, as shown previously for the orthosteric GABA site reducing agonist binding 400-fold55. Several African, East Asian and European Non-Finnish variants cluster at loop A His105, including α1 subunit tyrosine or arginine substitutions and α3 subunit tyrosine or asparagine substitutions, all of which have been shown to reduce sensitivity to classical benzodiazepines 10-20-fold56. Finally, the loop C tip (Ser209-Thr210-Gly211 in the α5 subunit) experiences the highest incidence of variation in all four BDZ binding α1/2/3/5 subunits. Substitutions within this region have previously been shown to reduce sensitivity to potentiation by classical BZDs and the sedative zolpidem, a non-BZD that binds the BZD site, by 5-10-fold34,57. Whilst it might be expected that reduced binding will result in a suboptimal patient response to BZDs, for some of these allelic variants the responses to anxiety or epilepsy treatment might actually be improved. The GABAAR α1 subunit is linked to unwanted sedative and addictive side effects58 so individuals possessing α1 His102 substitutions might, intriguingly, experience less side effects.
Nb38 is a novel positive allosteric modulator
While the cryo-EM map resolution limits our ability to define precise side-chain interactions, it is clear that major receptor-Nb38 contact points exist between: the CDR2/3 loops and β9/β10 strands and β9-β 10 hairpin (loop C), with CDR2 inserting into the pocket under loop C; CDR1-3 loops and the β6/β7 strands and β6-β7 loop (Cys-loop); CDR1/2 loops and the (C)-face β1/β2 strands and β8/β9 loop (loop F) (Fig. 4a, b and Supplementary Fig. 8a, b). Nb38 binding affinity (KD) for detergent-solubilised α1β3γ2EM receptors, as determined by surface plasmon resonance (SPR), was increased 6.5-fold in the presence of 1 mM GABA, from 1.61 nM to 248 pM, suggesting that it favours binding and stabilises an activated receptor conformation (Supplementary Fig. 8c, d). Whole cell patch-clamp recording confirmed this. 10 μM Nb38 strongly potentiated EC10 GABA currents by 480 ± 30 % (n = 7) for α1β3γ2EM and 290 ± 20 % (n = 7) for wild-type (WT) α1β3γ2 (Fig. 4c), greater than achieved by the BZD diazepam, 180 ± 20 % (n = 12) for α1β3γ2EM and 130 ± 20 % (n = 9) for α1β3γ2WT (Supplementary Fig. 3b). Application of Nb38 alone at concentrations up to 10 μM had only a weak direct agonist effect (Fig. 4c). Importantly, Nb38 was α1 subunit selective, eliciting no potentiation of wild-type a2-a6β3γ2 receptors (Fig. 4d). Thus, Nb38 is a novel pharmacological tool with superior efficacy and selectivity for a1 subunit receptors over benzodiazepines. The Nb38 ability to potentiate GABAARs was also observed for spontaneous inhibitory post-synaptic currents (sIPSCs) from dentate gyrus granule cells (DGGCs) in acute hippocampal slices, which primarily stem from a1-subunit containing receptors59. Without significantly affecting sIPSC frequency or rise-time, 2 μM Nb38 increased amplitude 37 ± 9 % (p < 0.05) and prolonged decay times by 96 ± 19 % (p < 0.05) to increase GABAergic drive (in terms of charge transfer calculated as the average sIPSC surface area) by 94 ± 14 %, n = 4 (p < 0.01) (Fig. 4e, f). By comparison, 500 nM diazepam also significantly prolonged sIPSC decay time, although to a lesser extent (39 ± 10 %, p < 0.01 %), and did not significantly increase GABAergic drive.
Impact of N-linked glycosylation on pentamerization
The a1β3γ2EM cryo-EM map revealed multiple N-linked glycans attached to the a1 (Asn111, β5-β5’ loop), β3 (Asn80, β3-strand and Asn149, β7-strand) and γ2 (Asn208, β9-strand) subunits (Fig. 1c). Of particular note, the two a1 Asn111 glycans occupy the ECD vestibule and, unexpectedly, adopt well-ordered conformations (Fig. 5a). The a1-D glycan projects upwards into the extracellular space (Fig. 5b). In contrast, the a1-A glycan projects horizontally to form putative CH-π interactions between the pyranose ring of a mannose moiety and the apposing Trp123 side chain from the γ2 β5-β5’ loop (Fig. 5c, d). This glycosylation site is conserved across all a-subunits (a1-6) and Trp123 is conserved across all γ-subunits (γ1-3), but not β-subunits (Fig. 5e). Thus, this glycan pairing is expected to exist in all aβγ GABAARs in the human brain, a feature absent from all previously solved pLGIC structures10. Enzymatic glycosylation of a1 Asn111 precedes exit from the endoplasmic reticulum (ER)60 and requires access to the inner face of a1, meaning that it must precede assembly and closure of the pentameric ring. The non-random orientation induced by the α1-A glycan interaction with γ2 Trp123 suggests that pentamerization comes from pre-assembled αβγ trimers with a horizontal glycan, followed by the addition of a second αβ dimeric unit (Fig. 5f). Furthermore, the Asn111 glycosylation represents a stoichiometric control mechanism, preventing the inclusion of more than two α-subunits via steric hindrance. Accordingly, we could only obtain the pentameric α5HOM and α5γ2HET constructs described earlier after mutating the Asn111 residue (Supplementary Fig. 1e).
The extracellular region conformations
Superpositions of α5HOM, α5γ2HET and GABAAR-β3cryst5 single ECDs, reveal that all three subunits comprising the major synaptic heteromer format (2αn-2βn-1γn) adopt highly similar β-sandwich arrangements (RMSD in the 0.59-0.76 Å range between different chains over 208 equivalent Cα positions), with a notable divergence in the flexible β8-β8’ and β8’-β9 loops (loop F) (Supplementary Fig. 8e). Global superposition of the α5HOM, α5γ2HET and α1β3γ2EM ECD pentamers reveal that all the subunits have adopted positions relative to the pore axis similar to agonist-bound GABAAR-β3cryst and the homologous agonist-bound glycine receptor6 (GlyR), rather than antagonist-bound GlyR (RMSD in the 0.7-1 Å range versus GABAAR-β3; 1.0-1.4 Å versus agonist-bound GlyR; 2.1-2.3 Å versus antagonist-bound GlyRs, over 208 equivalent Cα positions) (Supplementary Fig. 8f-j). The activated ECD bases swing out from the pore allowing it to widen while the loop C tips close inwards. Of note, the same assumed ECD positions between subunits means the α1β3γ2EM BZD site is highly similar to the BZD sites of α5HOM and α5γ2HET (Fig. 6a-c). Consistent with this, in silico docked flumazenil and bretazenil assumed the same binding modes in α1β3γ2EM (Fig. 6d-e). In the case of α5HOM and α5γ2HET this means the polar Thr208-Ser209-Thr210 residues at the tip of loop C sterically trap BZD in the site (Fig. 2b, f). The tight closure is incompatible with the inverted concave tetracyclic ring series of the (R)-bretazenil enantiomer47, as confirmed by in silico docking in which the lowest energy binding mode of (S)-bretazenil recaptures the binding mode in a5γ2HET whereas (R)-bretazenil is pushed out of the pocket (Supplementary Fig. 7f, g). The activated ECD conformations of a5γ2HET and a1β3γ2EM are consistent with them being bound by PAMs, i.e. bretazenil and Nb38, respectively. That all five subunits of a1β3γ2EM adopt equivalent ECD conformations despite non-equivalent occupancy of the inter-subunit pockets (two β3P/a1C orthosteric sites presumed to be bound by GABA; one a1P/β3C and one a1P/γ2C site bound by Nb38; one a1P/β3C site presumed empty) holds to the Monod-Wyman-Changeux (MWC) theory that postulates transitions between states (e.g. closed-to-open states) preserve symmetry of the oligomer61. In the case of a5HOM, the bound ligand, flumazenil, is a neutral BZD antagonist with no preference between activated or resting states53, which will bind either equally well. In this instance a5HOM has adopted the activated conformation as is the case for a5γ2HET, which might reflect an intrinsic preference of the construct ECDs to favour the activated conformation, as previously observed for β3 ECDs in the absence of bound ligand62. Interestingly, increasing the size of the C(3)-linked ester substituent of flumazenil analogues correlates with PAM activity53. A flumazenil analogue where the ester is substituted by an alkyne of the same length was a neutral modulator (antagonist), whilst addition of a tert- butyl group, as is present in bretazenil, conferred PAM activity46. In silico docking of this PAM analogue (compound 4) into a5γ2HET resulted in the lowest free energy binding mode directly overlapping with bretazenil (Supplementary Fig. 7h). In both cases, the tert-butyl groups are precisely positioned to bridge the gap between loop C Thr208-Ser209 and the adjacent subunit at β1-strand Asp56-Tyr58 (Fig. 2b and Supplementary Fig. 7h). Closure of loop C is primarily a consequence of the quaternary motion undertaken by the pLGIC ECDs during activation6,8,10,20. This suggests that, by bridging the subunits at this critical juncture, bretazenil and other BZDs with appropriately large C(3) substituents stabilise the integrally linked process of loop C closure and ECD quaternary activation via their larger contact surfaces compared to flumazenil.
Ion permeation pathway and pore conformation
α1β3γ2EM possesses two positively charged rings, both previously observed in GABAAR-β3cryst, halfway down the vestibule and at the intracellular end of the pore (Supplementary Fig. 9a-c), and previously proposed as ion selectivity filters in pLGICs63–65. In the crowded environment of the glycan-filled vestibule, the ion permeation pathway is restricted to ∼ 5Å diameter (Fig. 7a), less than a fully hydrated Cl− ion (6.1 Å)66 but larger than a dehydrated Cl− ion (Pauling radius of 1.8 Å). However, the polar, semi-flexible glycan chains are not expected to impede the Cl− flux significantly. The membrane spanning pore, lined by five M2 helices, one from each subunit, narrows from greater than 8 Å diameter on the extracellular side (excluding the uppermost concentric residue ring, designated 20’, which comprises flexible polar side chains that will not impede conductance), to ∼5.6 Å on the intracellular side, at the - 2’ ring (α1 Pro253, β3 Ala248, γ2 Pro263; Fig. 7b). Although slightly narrower than a fully hydrated Cl− ion, this is wider than the closed −2’ desensitization gate of GABAAR-β3cryst5,67 (3.2 Å), and also dehydrated Cl− ions, and is consistent with an open-channel state, within the 5.5-6 Å range previously estimated on the basis of different-size anion permeability68. It is important to state that although the M2 helices (lining the pore) of α1 and β3 subunits are well ordered in the cryo-EM map, the γ2 subunit TMD region is mobile. This appears to be an artefact caused by detergent solubilisation, as we observed a reduction of this motion upon addition of lipids (Supplementary Fig. 4a). Iterative rounds of particle classification led to a map in which the γ2 subunit TMD model could be unambiguously placed, but not accurately refined. Therefore, to ascertain the conductance state, we performed 100 ns molecular dynamics simulations69 on the TMD pore embedded in a POPC membrane with water and 0.15 M NaCl on each side and a + 0.12 V transmembrane potential (where the sign of the voltage refers to that of the cytoplasmic face). Water molecules occupied the length of the pore, indicating there was no significant hydrophobic barrier leading to de-wetting of the channel70. There was an influx of multiple Cl− ions, equivalent to an estimated conductance of the order of 100 pS, sufficient to account for physiological conductances (25-28 pS71) (Fig. 7c and Supplementary Fig. 9d-i). In contrast, Na+ ions failed to traverse the pore despite having a smaller radius (1.2 Å), as expected for an anion selective pore. Consistent with the pore assuming an open state that is narrowest at the intracellular side, single channel recordings of GABAARs in membranes show reduced conductance upon mutation of the only polar side chain in the lowest two residue rings, γ2 2’ Ser, to non-polar alanine or valine72. Furthermore, the a1β3γ2EM pore topology differs from the closed conformation of the related GlyR6,7, where the M2 helices were drawn together at the midpoint 9’ leucine residue ring to create a hydrophobic conductance barrier (2.8 Å diameter) (Fig. 7b). Nevertheless, the GABAAR open pore we observe is narrower than the pore of glycine-activated GlyR (8.8 Å diameter6) which, fittingly, possesses a higher conductance (60-90 pS73). From previous single channel studies wild type a1β2γ2 receptors in HEK cell membranes in the presence of saturating GABA concentrations occupy the open state 25 % of the time relative to 75 % for the desensitised state74. Presumably, the different experimental conditions imposed here, such as extraction from membranes into detergent micelles along with freezing in vitreous ice, have shifted the equilibrium in favour of the open state, resulting in its observation rather than the desensitised state.
pLGICs in which proposed closed and open conformations of the same receptor have been determined, nAChR75 and GlyR6, reveal that pore opening arises from a relative ‘rocking’ motion between the ECD and the TMD regions. For α1β3γ2EM, the activated ECD base loops, in particular the β 1-2 loop and β6-7 loop (Cys-loop), are displaced away from the pore axis similarly to those of agonist-bound GlyR versus closed GlyR (Supplementary Fig. 8n, o). These base loops form contacts along the M2-M3 loop to concomitantly withdraw the TMD. Superposition of single ECDs from antagonist (strychnine), agonist (glycine) and glycine/ivermectin bound GlyRs on the α1β3γ2EM β3 subunit reveal the impact of this ‘rocking’. Relative to the pore axis, closed GlyR M2 is over-straightened by θ = −3.3° and rotated clockwise by φ = 17.6°, which contrasts with the θ = 7.1° recline of the α1β3γ2EM β3 subunit (Fig. 7e). As expected for the agonist-bound GlyR M2s, these assume similar reclines to α1β3γ2EM β3, θ = 10.8° and 9.9°, respectively, with glycine alone bound GlyR also rotated anti-clockwise by φ = 20.7° (Fig. 7f, g).
To evaluate the possible transition motions from the α1β3γ2EM open pore to the desensitized state5,67, we superposed an ECD from the desensitized GABAAR-β3cryst5 onto either α1, β3 or γ2 ECDs from the open pore α1β3γ2EM. This analysis revealed that the desensitized conformation also results from a combined tilting and rotation of subunit TMDs relative to the ECDs: the lower portions of the TMDs tilt 10-15° and rotate 9-11° relative to the α1β3γ2EM subunits, moving towards the pore axis and closing the channel at the −2’ ring (Fig. 7h-j). A similar motion is observed for GlyR, in which the glycine plus ivermectin activated conformation undertakes an inward motion of the lower portions of the TMDs relative to the glycine bound alone activated state6.
DISCUSSION
The α5HOM and α5γ2HET crystal structures reported here reveal the binding modes of the neutral ligand and BZD antagonist, flumazenil, and the partial PAM, bretazenil. These BZD sites retain 100 % residue identity to the wild type receptor site. Thus, these engineered templates serve as potent tools for future studies to unambiguously determine the binding modes of many of the small molecules previously found to bind the BZD site24. Of note, the binding affinities are in the range of one or two orders of magnitude lower for the BZDHET and BZDHOM sites, respectively. Nevertheless, the binding modes for flumazenil and bretazenil are maintained regardless of the site and are consistent with the previously established functional data discussed throughout the results. Importantly, introduction of a His to Arg substitution into the BZDHET site reproduced the same differential impact on flunitrazepam versus flumazenil binding as previously described for wild-type GABAA receptors32,44. Furthermore, superposition of these sites over the BZD site in α1β3γ2EM reveal highly similar positioning of the secondary structure elements and relative appositions of the inter-subunit (P) and (C) faces. Thus, the lower affinities are not due to the sites being distorted or belonging to artefactual structures. Instead, differences likely reflect differing receptor energetics to assume favourable bound conformations for α5HOM and α5γ2HET versus wild type receptors in which two of the α-subunits are replaced by β-subunits.
Here we report activated homomeric and heteromeric GABAAR states bound by a BZD antagonist, and BZD and nanobody PAMs, respectively. These structures reveal fundamental insights into the activation and allosteric modulation processes, and lay the foundation to understand important aspects of GABAAR inhibitory neurotransmission. Furthermore, although BZDs are widely used to treat epilepsy, insomnia, anxiety, panic disorder and muscle spasm, they lack subtype selectivity and cause unwanted sedation, addiction, and motor and cognitive impairment26. Subtype selective drugs against the BZD site will ameliorate side effects and broaden the therapeutic repertoire to include treatments for autism, Down syndrome, neuropathic pain, schizophrenia and stroke15–17,27,28. We anticipate that the mechanistic insights and crystallographic platforms described here will expedite the arrival of structure-guided design of therapeutics to treat these disorders.
AUTHOR CONTRIBUTIONS
Molecular biology, protein expression, purification and crystallization, radioligand binding assays, whole cell electrophysiology: PSM; X-ray data collection: A.R.A., P.S.M., T.M.; X-ray data processing: A.R.A., P.S.M.; cryo-EM data collection: S.M., A.K., Z.S., P.S.M., J.T.H.; cryo-EM data processing: S.M., J.T.H; cryo-EM model building: S.M., L.D.C., S.L., B.F., F.D.M.; small-molecule docking: T.M.; human genome bioinformatics: S.C., M.M.B.; molecular dynamics simulations: S.R., G.K., S.J.T., M.S.P.S; software and hardware: J.M.D, A.S., R.M.E.; nanobody generation: E.P., J.S.; SPR: S.M.; slice electrophysiology: S.H., T.G.S. The manuscript was written by P.S.M, S.M. T.M. and A.R.A, with input from all co-authors.
METHODS
Construct design, α5HOM and α5γ2HET
Details of the α5 subunit constructs design, including protein sequences, are shown in Supplementary Figure 1. The chimeric γ2-ECD:α1-TMD subunit of α5γ2HET comprises mature sequence (Uniprot P18507) γ2 residues 39 to 232 (QKSDD…DLSRR) appended to α1 (Uniprot P62813) from 223 to 455 (IGYFVI…PTPHQ) with a single β3 substitution, (α1 P280A). The α5 intracellular M3-M4 loop amino acids 316-392 (RGWA…NSIS) (Uniprot P31644) and the α1 intracellular M3-M4 loop amino acids 313-390 (RGYA…NSVS) were substituted by the SQPARAA sequence5,40 to enhance the recombinant protein yield and facilitate crystallisation. Constructs were cloned into the pHLsec vector76, between the N-terminal secretion signal sequence and either a double stop codon or a C-terminal 1D4 purification tag derived from bovine rhodopsin (TETSQVAPA) that is recognised by the Rho-1D4 monoclonal antibody (University of British Columbia)77,78.
Construct design, α1β3γ2EM
The protein sequences used were: human GABAAR α1 (mature polypeptide numbering 1-416, QPSL…TPHQ; Uniprot P14867), human GABAAR β3 (mature polypeptide numbering 1-447, QSVN…YYVN; Uniprot P28472), human GABAAR γ2 (mature polypeptide numbering 1-427, QKSD…YLYL; Uniprot P18507). These constructs were cloned into the pHLsec vector76, after the N-terminal secretion signal sequence and before a double stop codon unless stated otherwise below. The a1 intracellular M3-M4 loop amino acids 313-391 (RGYA…NSVS) were substituted by the SQPARAA sequence5,40. Four β3 ECD residues were replaced by β2 subunit residues (Gly171Asp, Lys173Asn, Glu179Thr, Arg180Lys), which block homomer assembly79, and the β3 intracellular M3-M4 loop amino acids 308-423 (GRGP…TDVN) were substituted by a modified SQPARAA sequence containing the E.coli soluble cytochrome B562RIL41 (BRIL, amino acids 23-130, ADLE…QKYL, Uniprot P0ABE7) to give an M3-M4 loop of sequence SQPAGT-BRIL-TGRAA. The γ2 intracellular M3-M4 loop amino acids 323-400 (NRKP…IRIA) were substituted by the SQPARAA sequence, and a C-terminal GTGGT linker followed by a 1D4 purification tag derived from bovine rhodopsin (TETSQVAPA) that is recognised by the Rho-1D4 monoclonal antibody (University of British Columbia)77,78. Nb38 was identified from a previously described nanobody library20.
Large-scale expression and purification of a5HOM, a5γ2HET and a1β3γ2EM
Twenty-litre batches of HEK293S-GnTI− cells (which yield proteins with truncated N-linked glycans, Man5GlcNAc280,81) were grown in suspension to densities of 2 × 106 cells ml−1 in Protein Expression Media (PEM, Invitrogen) supplemented with L-glutamine, non-essential amino-acids (Gibco) and 1% foetal calf serum (Sigma-Aldrich). Typical culture volumes were 200 ml, in 600 ml recycled media bottles, with lids loose, shaking at 130 rpm, 37°C, 8 % CO2. For transient transfection, cells from 1 litre of culture were collected by centrifugation (200 g for 5 min) and resuspended in 150 ml Freestyle medium (Invitrogen) containing 3 mg PEI Max (Polysciences) and 1 mg plasmid DNA, followed by a 4 h shaker-incubation in a 2 litre conical flask at 160 rpm. For α5γ2HET DNA plasmids were transfected at 9:1 ratio (i.e. 0.9:0.1 mg) α5 construct DNA without a 1D4 tag to the chimera γ2-ECD:α1-TMD with a 1D4 purification tag. For α1β3γ2EM DNA plasmid ratios were 1:1:0.5, respectively. Subsequently, culture media were topped up to 1 litre with PEM containing 1 mM valproic acid and returned to empty bottles. Typically, 40-70 % transfection efficiencies were achieved, as assessed by control transfections with a monoVenus-expressing plasmid82,83. 72 h post-transfection cell pellets were collected, snap-frozen in liquid N2 and stored at −80 °C.
Cell pellets (approx. 200g) were solubilised in 600 ml buffer containing 20 mM HEPES pH 7.2, 300 mM NaCl, 1 % (v/v) mammalian protease inhibitor cocktail (Sigma-Aldrich, cat. P8340) and 1.5 % (w/v) dodecyl 1-thio-β-maltoside (DDTM, Anatrace) for α5HOM and α1β3γ2EM or 1.5 % (w/v) decyl β-maltoside (DM, Anatrace) for α5γ2HET, for 2 hours at 4 °C. Insoluble material was removed by centrifugation (10,000 g, 15 min). The supernatant was diluted 2-fold in a buffer containing 20 mM HEPES pH 7.2, 300 mM NaCl and incubated for 2 hr at 4 °C with 10 ml CNBr-activated sepharose beads (GE Healthcare) pre-coated with 50 mg Rho-1D4 antibody (3.3 g dry powdered beads expand during antibody coupling to approximately 10 ml). Affinity-bound samples were washed slowly by gravity flow over 2 hours at 4 °C with 200 ml buffer containing 20 mM HEPES pH 7.2, 300 mM NaCl, and either 0.1 % (w/v) DDTM (approximately 20 x CMC) for α5HOM, or 0.2 % (w/v) DM (approximately 3 x CMC) α5γ2HET, or for α1β3γ2EM 0.1 % (w/v) DDTM and 0.01 % (w/v) porcine brain polar lipid extract (141101C Avanti; chloroform was evaporated under argon then 100 mg lipid film was dissolved in 10 ml 10 % (w/v) DDTM (1000 mg) in water and stored at −80 C until needed). Beads were then washed in a second round of buffer: 20 mM HEPES pH 7.2, 300 mM NaCl, and either 0.01 % (w/v) DDTM (approximately 3 x CMC) for α5HOM or 0.2 % (w/v) DM (approximately 3 x CMC) for α5γ2HET, or for α1β3γ2EM, 1mM GABA, 0.01 % (w/v) DDTM (approximately 4 x CMC), 0.001 % (w/v) porcine brain polar lipid extract. Protein samples were eluted overnight in 15 ml buffer containing 15 mM HEPES pH 7.2, 225 mM NaCl, 500 μM TETSQVAPA peptide (Genscript), and corresponding detergents. The eluate was centrifuged (30,000g, 15 min) and the supernatant was concentrated by ultrafiltration to 1-2 ml at 1-5 mg/ml using 100-kDa cut-off membranes (Millipore). The concentrated sample was centrifuged (30,000 g, 15 min) and the supernatant was aliquoted in 0.5-1.5 mg protein per 0.7 ml aliquots and either snap-frozen for storage at −80 °C or gel filtrated as appropriate. A single aliquot was loaded onto a Superose 6 10/300 Increase gel filtration column (GE Healthcare) equilibrated in 10 mM HEPES pH 7.2, 150 mM NaCl, and either: 0.007 % (w/v) DDTM, 50 μM flumazenil, 50 μM pregnanolone for a5HOM or 0.2 % (w/v) DM, 50 μM bretazenil for a5γ2HET, or for a1β3γ2EM, 1 mM GABA, 0.007 % (w/v) DDTM. The peak fractions were approximately 0.5 mg/ml. The fractionated protein was concentrated by ultrafiltration to 3-5 mg/ml, using 100 kDa cut-off membranes (Millipore), for crystallisation trials. Typical final yields were 0.1-0.2 mg protein per litre of cells grown in suspension (10 g cell pellet). In the case of a1β3γ2EM, Nb38 was added to a1β3γ2 at 4-fold molar excess and the complex was concentrated by ultrafiltration to 2.5 mg/ml, using 100 kDa cut-off membranes (Millipore).
Nb38 purification
Nb38 was produced and purified in milligram quantities from WK6su− E. coli bacteria84. Bacteria were transformed with ∼200 ng of nanobody expression plasmid pMESy4 containing the nanobody of interest and selected on Lysogeny broth (LB)-agar plates containing 2% glucose and 100 μg/ml ampicillin. 2-3 colonies were used for preparing a preculture, which was used to inoculate 0.5 L Terrific broth (TB) cultures supplemented with 0.1 % glucose, 2 mM MgCl2 and 100 μg/mL ampicillin. Cultures were grown at 37 °C until their OD600 reached 0.7, at which point Nb38 expression was induced with 1 mM IPTG. After induction, cells were grown at 28 °C overnight and harvested by centrifugation (20 min, 5000g). Nanobodies were released from the bacterial periplasm by incubating cell pellets with an osmotic shock buffer containing 0.2 M Tris pH 8.0, 0.5 mM EDTA, and 0.5 M sucrose. The C-terminally His6-tagged Nb38 was purified using nickel affinity chromatography (binding buffer: 50 mM HEPES pH 7.6, 1 M NaCl, 10 mM imidazole; elution buffer: 50 mM HEPES pH 7.6, 0.2 M NaCl, 0.5 M imidazole), followed by size-exclusion chromatography on a Superdex 75 16/600 column (GE Healthcare) in 10 mM Hepes pH 7.6, 150 mM NaCl. Nb38 stocks were concentrated to 5-10 mg/mL, snap-frozen in liquid nitrogen and stored at −80 °C.
Crystallization and data collection
α5γ2HET and α5HOM contain 15 N-linked glycosylation sites each, bringing a considerable extra volume, flexibility and potential occupancy heterogeneity. Therefore, prior to crystallization, concentrated protein samples (6 mg/ml α5γ2HET and 4 mg/ml α5HOM) were incubated with 0.01 mg ml−1 endoglycosidase F185 for 2h at room temperature. Sitting drop vapour diffusion crystallization trials were performed in 96-well Swisssci 3 well crystallisation plates (Hampton Research), at three ratios: 200 nl protein plus 100 nl reservoir, 100 nl protein plus 100 nl reservoir, 100 nl protein plus 200 nl reservoir. Drops were dispensed by a Cartesian Technologies robot86, and plates were maintained at 6.5 °C in a Formulatrix storage and imaging system. In the case of α5γ2HET, crystals appeared in a range of conditions87 within 1-28 days, with the best diffracting crystals (to ∼2.5 Å resolution) taking 4 weeks to grow in: 22 % poly-ethylene (PEG) 400, 0.37 M potassium nitrate, 0.1 molar 2-(N-morpholino)ethanesulfonic acid (MES) pH 6.5. For α5HOM, crystals also grew in a range of conditions, typically within 2 weeks, and in the first instance diffracted up to intermediate resolution (>5 Å). Following additive-based optimization (MemAdvantage, Molecular Dimensions), crystals diffracting to ∼2.6 Å resolution were identified, grown in: 19 % PEG 1000, 0.1 M sodium chloride, 0.15 M ammonium sulphate, 0.1 M MES pH 6.5, 2.5 mM sucrose monodecanoate (sucrose monocaprate). Crystals were cryoprotected by soaking in reservoir solution supplemented with 30 % ethylene glycol, and then cryocooled in liquid nitrogen. Diffraction images were collected at the Diamond Light Source beamline I04, λ=0.9795 Å, 0.1° oscillation (bretazenil-bound α5γ2HET) and 0.2° oscillation (flumazenil-bound α5HOM), on a Pilatus 6M-F detector. X-ray data were indexed, integrated and scaled using the HKL2000 package88. Diffraction from both α5γ2HET and α5HOM crystals was severely anisotropic, therefore scaled but unmerged data were processed with STARANISO89, allowing for the anisotropic diffraction cut-offs to be applied before merging with Aimless90,91, within the autoPROC toolbox92. Upon ellipsoidal truncation, resolution limits were 2.33 Å, 3.15 Å and 3.73 Å (in the −0.022 a* + c*, b* and 0.945 a* - 0.327 c* directions, respectively) for α5γ2HET, and 2.49 Å, 3.13 Å and 4.63 Å (in the 0.872 a* - 0.490 c*, b* and 0.842 a* + 0.540 c* directions, respectively) for α5HOM. Data collection and merging statistics are detailed in the Table 1.
Structure determination, refinement and analysis
α5γ2HET and α5HOM structures were solved by molecular replacement using the human GABAAR-β3cryst homopentamer5 (PDB ID: 4COF) as a search model in Phaser93. Polypeptide chains were traced using iterative rounds of manual model building in Coot94 and refinement in BUSTER-TNT95, Refmac96 and Phenix97,98. Automated X-ray and atomic displacement parameter (ADP) weight optimisation, and torsion angle non-crystallographic symmetry (NCS) restraints, were applied. Ligand coordinates and geometry restraints were generated using the grade server99. The α5γ2HET and α5HOM models contain one homopentamer per asymmetric unit. Crystal packing impaired map quality in regions where ECD from certain subunits were near of detergent micelles of neighbouring molecules. Nevertheless, complete polypeptide chains could be built, with the exception of 14 N-terminal α5 residues (QMPTSSVKDETNDN), 22 N-terminal γ2 residues (QKSDDDYEDYTSNKTWVLTPKV) and the C-terminal purification tags, presumably disordered. Strong additional electron density peaks were clearly visible in the BZDHOM and BZD sites, that could be unambiguously assigned to flumazenil in a5HOM and bretazenil in a5γ2HET, respectively, based on shape, coordination and refinement statistics. Furthermore, electron density corresponding to five pregnanolone molecules, one per inter-subunit interface, could be observed at the TMD interfaces of a5HOM, as previously described20, and five well-ordered detergent molecules (decyl β-maltoside) could be built inside the a5γ2HET pore region. The a5 and γ2 extracellular regions have three N-linked glycosylation sites each, and we could observe clear electron density for six NAG moieties in a5HOM and five in a5γ2HET, the others being disordered. Stereochemical properties of the models were assessed in Coot94 and Molprobity100. Refinement statistics are provided in the Table 1. Protein geometry analysis revealed no Ramachandran outliers, 97.96% residues in favoured regions and 2.04% residues in allowed regions for a5HOM, and one Ramachandran outlier (0.06%), 95.97% residues in favoured regions and 4.03% residues in allowed regions for a5γ2HET. Molprobity clash scores after adding hydrogen atoms is 9.16 (99th percentile) for a5HOM and 11.06 (94th percentile) for a5γ2HET. Overall Molprobity scores are 1.50 (100th percentile) for a5HOM and 1.84 (98th percentile) for a5γ2HET. Structural alignments were performed in PyMOL using the align function. Structural figures were prepared with the PyMOL Molecular Graphics System, Version 1.8, Schrödinger, LLC., and the UCSF Chimera package, developed by the Resource for Biocomputing, Visualization, and Informatics at the University of California, San Francisco101.
Electron microscopy and image processing
CryoEM samples were prepared using C-flat™ Holey Carbon grids (R2/1, 200 mesh, 53% of collected images) and UltraAuFoilTM grids (R1.2/1.3, 200 mesh, 47% of collected images). Carbon substrate grids were glow discharged for 10 s, then 3.5 μL of protein sample (4.2 mg/mL) was applied for 15 s. The sample was blotted for 3 s using VitroBot Mark IV (FEI) and flash-frozen in liquid ethane. Gold substrate grids were glow discharged for 15 s and 2.5 μL sample (2.5 mg/mL) was applied for 30s, which was followed by 7.5 s blotting. In both cases, blotting was performed at room temperature and 90-100% humidity. All cryo-EM data were collected using a Tecnai F30 Polara electron microscope (FEI) operating at 300 kV, fitted with a K2 Summit direct electron detector (Gatan) and a GIF-Quantum energy filter (Gatan). SerialEM was used to manually record zero energy-loss (20 eV slit) images at a calibrated magnification of 37,000x (1.35Å/pix) in counting mode. Images were recorded as movies consisting of 47 frames with total dose of 38 e−/ Å2 and exposure of 14.1s. Nominal defocus values ranged from −2.0 to −3.5 μm.
Classification of α1β3γ2EM particles in the decylmaltoside neopentylglycol, (DMNG) detergent revealed that the particles had a ‘collapsed’ TMD region, with the γ2 TMD occupying the central pore (Supplementary Fig. 4a). We discarded this sample, considering it to be a sample preparation artefact, presumably caused by extraction of the receptor from its native cell bilayer into DMNG alone. Preparation of a second sample in a different detergent, dodecyl 1-thio-β-maltoside (DDTM) and in the presence of porcine brain polar lipid extract (see methods section on Large-scale expression and purification) yielded the predominant particle population with TMDs pseudo-symmetrically arranged around the pore, i.e. the γ2 TMD was no longer ‘collapsed’ into central pore. This sample was used for all subsequent processing. A total of 8,548 movies were motion-corrected at micrograph level with UCSF MotionCorr102 and CTFFIND4103 was used to estimate the contrast function parameters. Data processing was performed using RELION software package104. First, ∼5,000 particles were picked using EMAN2105, grouped in RELION using reference-free 2D classification and class averages were used as references for RELION automated particle picking in the same software. False-positives were manually removed leading to a dataset of 525,124 particles. Reference-free 2D classification was performed and particles belonging to classes showing GABAAR features were selected (436,320 particles). An initial 3D model was generated from the GABAAR-β3cryst structure5 (PDB ID 4COF) and low-pass filtered to 40 Å resolution. Particles were first oriented in 3D by imposing C5 symmetry, and further classified in 3D imposing no symmetry (C1) by only allowing rotation around the symmetry axis. The best class showing features of two BRILs and two nanobodies was selected and used as a new initial reference model. Particles selected from 2D classes were 3D-classified into 10 classes using C1 symmetry. Particles assigned to the best 3D classes (186,786 particles) were used for particle ‘polishing’ step in RELION, where particle motion and radiation damage for particles from each movie frame was estimated. ‘Gold standard’ refinement of the resulting ‘shiny’ particles resulted in a 5.65 Å map. Movie motion correction with MotionCor2106 using 25 patches and dose-weighting scheme improved the resolution of the refined map to 5.25 Å. These particles were further 3D-classified into 10 volumes. Particles assigned to the best 9 classes (165,621 particles) were combined to yield a map of 5.17 Å. The resolution was estimated using relion post-process with the FSC criteria of 0.143. The final unsharpened and unfiltered map was globally autosharpened using phenix.auto_sharpen107 (5.17 Å high-resolution cut-off) to maximise the map detail while maintaining the connectivity of the map.
Model building and refinement of α1β3γ2EM
Crystal structures (2.9 Å β3cryst (PDB ID 4COF), 2.5 Å a5γ2HET (PDB ID XXXX) and 3.2 Å a5TMD20 (PDB ID 5O8F)) were used to build the GABAAR α1β3γ2EM heteromer and Nb38 models into the 5.17 Å map. First, the a5 subunit and Nb2520 (5O8F) structures were manually mutated to match the a1 and Nb38 sequences, respectively. TMD of the γ2 subunit was prepared by mutating a5 TMD (5O8F) residues to match the γ2 sequence. The Nb38 model and the corresponding ECDs and TMDs of α1, β3 and γ2 subunits were docked into the cryo-EM map using Chimera. Map was cut around each subunit and nanobody (4 Å radius) using Chimera. Rosetta-CM108 was used to refine the models into the resulting cryo-EM densities to improve model geometry and fitting in the density. N-linked glycan models from α5TMD crystal structure (5O8F) were docked into the cryo-EM map and added to the model. Rotamers were manually adjusted to match the high resolution structures; where no prior information was available, the most common rotamers were chosen. The model was further optimised by rounds of manual correction in Coot94 and iterative refinement in real space with phenix_real_space_refine109 using secondary structure and NCS restraints. The final model contains α1 subunit residues 12-312 (TTVF…YFTK) and 391-414 (KIDR…YWAT); β3 subunit residues 10-307 (SFVK…YIFF) and 422-447 (AIDR…LYYV); γ2 subunit residues 27-322 (VTVI…YFVS) and 401-424 (KMDS…YWVS) and Nb38 residues 1-123 (QVQL…TVSS). For model validation, the final model coordinates were randomly displaced by 0.2 Å and then this model was refined with phenix_real_space_refine against one of the half-maps produced by RELION104. FSC curves were then calculated between the refined model and half-map used for refinement (‘work’) and between the second half-map, not used for refinement, (‘free’). No significant separation between FSCwork and FSCfree curves was observed, indicating the model was not over-refined. The stereochemistry of the final model was evaluated using MolProbity110. Figures were prepared using UCSF Chimera package, developed by the Resource for Biocomputing, Visualization, and Informatics at the University of California, San Francisco101. Structural alignments were performed in SHP111. Pore dimensions were analysed using the Coot implementation of Hole112 and with CAVER113 PyMOL plugin with a probe radius of 2.5 Å.
Small molecule docking
Molecular docking of small molecules to a5γ2HET and a5HOM crystal structures was performed using AutoDock Vina114. Stereochemistry of small molecules was optimised using Grade webserver115. Structures of the receptor were kept rigid during docking. The region selected for docking encompassed whole benzamidine-binding pocket.
Molecular dynamics simulations
The protein portion of the receptor structure, either full length (for uniform electric field simulations) or with only the transmembrane domain included (for water equilibrium simulations), was embedded within a POPC lipid bilayer, in a simulation box containing an aqueous solution of Cl− and Na+ ions at 0.5 M or 0.15 M concentration, respectively, following a previously established protocol116 and employing the TIP4P water model117. Simulations were performed using GROMACS version 5.1118,119, and the OPLS united-atom force field120. The temperature and pressure were maintained at 37 °C and 1 bar, respectively, using the velocity-rescale thermostat121 in combination with a semi-isotropic Parrinello and Rahman barostat122, with coupling constants of τT = 0.1 ps and τP = 1 ps. A Verlet cut-off scheme was applied, and long-range electrostatic interactions were measured using the Particle Mesh Ewald method123. Bonds were constrained using the LINCS algorithm124, and an additional harmonic restraint at a force constant of 1000 kJ mol−1 nm−2 was placed on protein backbone atoms to preserve the conformational state of the original, experimentally determined structure. The integration time-step was 2 fs. For the water distribution and free energy profiles: 50 ns repeats were each initiated from an independently assembled simulation system. Free energy profiles for water permeation along the channel pore were derived (using the CHAP channel annotation package, http://www.channotation.org/) from the equilibrium distribution of water molecules upon simulation through an inverse Boltzmann calculation-based method70. Pore radius profiles (i.e. radii at positions along the pore axis) were estimated using the programme HOLE112. To determine the behaviour of ions a +0.12 V potential difference was imposed across the membrane during 100 ns simulations by setting a linear potential difference of 0.5 V across the z-length of the simulation box. The conductance was estimated based on the number of Cl− ions passing through the channel.
Radioligand binding experiments
GABAAR constructs containing a single BZD site (α1β3γ2EM, α5γ2HETΔ and α5β3γ2WT) at 2 nM and five sites (α5γ2HET and α5HOM) at 0.4 nM, were used, in 10 mM HEPES pH 7.2, 150 mM NaCl, and: 0.05 % (w/v) DDTM, 0.005 % (w/v) porcine brain polar lipid extract; 0.2 % (w/v) DM detergent for α5γ2HET; 0.05 % (w/v) detergent (decylmaltoside neopentylglycol (DMNG) 5:1 (molar ratio) cholesterol hemisuccinate (CHS) for α5β3γ2WT; 0.05 % n-dodecyl-beta-maltoside (DDM) for α5γ2HETΔ and α5HOM. Samples were incubated with WGA YSI beads (bind N-linked glycans, beads at 2 mg/ml, Perkin Elmer) for 30 minutes at 4 °C under slow rotation. 50 μl aliquots of the GABAA receptor-bead mix were added to 50 μl aliquots of 2 x radioligand ([3H]-flunitrazepam or [3H]-flumazenil) concentrations ranging from 0.06-2000 nM (Perkin Elmer) in Serocluster 96-Well ‘U’ Bottom plates (Corning) and incubated for 60 minutes at room temperature (20-22 °C) and [3H] cpm were determined by scintillation proximity assay using a Microbeta TriLUX 1450 LSC. The same ligand binding assay was performed in the presence of 50 μM flumazenil to ascertain the non-specific binding (NSB), which was subtracted from the total radioligand cpm to obtain the specific binding values. [3H]-flunitrazepam binding affinity (Kd) was calculated in OriginPro2015 using the one-site binding curve fit equation (y = Bmax*x/(k1+x)), or two-site binding curve fit equation (y = Bmax1*x/(k1+x) + Bmax2*x/(k2+x)), or using the Hill equation (y = Bmax*xn/(k1n+xn) where Bmax values are maximal binding for each site and k1 and k2 are Kd for each site, n is Hill slope, x is ligand concentration, y is proportion of binding. Displacement curves were performed by adding ligand (bretazenil or diazepam) over the concentration range 1-50000 nM to aliquots of GABAA receptor-bead mix for 30 minutes, then adding this to aliquots of radioligand ([3H]-flumazenil or [3H]-flunitrazepam, respectively) at final concentrations corresponding to approximately 10 x Kd. Diazepam displacement curves were plotted on log concentration axis and fitted using the logistic equation (y = A2+(A2-A1)/1+(x/x0)^p) where A2 and A1 are maximal and minimal binding respectively, x0 is IC50 and p is the Hill coefficient. IC50 values of displacement curves were converted to Ki values according to the Cheng-Prusoff equation125, Ki = IC50/1+([L]/Kd) referring to the [3H]-flumazenil Kd and the bretazenil IC50, and where L is the concentration of [3H]-flumazenil used in the displacement assay.
Thermostability binding experiments
Information regarding the thermostability of a detergent-solubilised protein can be determined by heating protein samples over a range of temperatures for equal time periods and then measuring the reduction in the intensity of the monodisperse SEC profile for each protein sample126. With increasing temperature an increased proportion of protein is denatured, aggregates and is lost from the monodisperse peak when the protein is subsequently run on SEC. A measure of protein stability can then be obtained by plotting the decay in peak UV absorbance against increasing temperature, for example to obtain a 50 % melting temperature (Tm). Purified a1β3γ2EM at 0.02 mg ml−1 (100 nM) in 150 mM NaCl, 10 mM HEPES pH 7.2, 0.007 % DDTM (w/v) was separated into 50 μl aliquots in PCR tubes, and heated at a range of temperatures from 30-80 °C for 1 hour. Samples were then run on a high-performance liquid chromatography system with automated microvolume loader (Shimadzu) through a Superdex 200 Increase 3.2/300 column (GE Healthcare) maintained in 300 mM NaCl, 10 mM HEPES pH 7.2, 0.007 % DDTM (w/v). Monodisperse peak reduction with increasing temperature was measured relative to an unheated control sample maintained at 4 °C.
Importantly, because some drugs when bound thermostabilise detergent-solubilised protein126, the thermostability assay offers an efficient strategy to measure protein sensitivity to drugs in the detergent-solubilised environment. Purified GABAAR Cryo-EM construct was separated into PCR tubes, supplemented with picrotoxin at a range of concentrations and heated at Tm30 % (the temperature at which the monodisperse peak was reduced by 70 %) for 1 hour. Afterwards samples were run on a high-performance liquid chromatography system with automated micro-volume loader (Shimadzu) through a Superdex 200 3.2/300 column (GE Healthcare) maintained in 300 mM NaCl, 10 mM HEPES pH 7.2, 0.007 % DDTM. Drug dose-response curves were generated by plotting UV absorbance against drug concentration.
Surface plasmon resonance analysis (SPR). SPR measurements were performed using a Biacore T200 (GE Healthcare) machine at 25 °C. All reagents and consumables for SPR were purchased from GE Healthcare. The carboxyl groups on CM5 chip flow channels were activated with a 10-minute injection of a 1:1 mixture of 0.1 M N-hydroxysuccinimide (NHS) and 0.4 M 1-ethyl-3-dimethylaminopropyl-carbodiimide (EDC). Streptavidin (Sigma-Aldrich) was covalently bound by a 5-minute injection until an immobilization level of 5000 RU was reached. Free activated carboxyl groups were quenched with a 10-minute injection of 1 M ethanolamine-HCl pH 8.5. The working flow-cell was functionalized by injecting GABAA receptor electron cryo-microscopy construct containing a biotinylated C-terminus containing a biotin ligase recognition sequence, until 350-375 RU were reached, whereas the reference cell containing streptavidin was left unmodified. Running buffer contained 10 mM HEPES pH 7.5, 150 mM NaCl, 0.007% DMNG:CHS (5:1, w/w). For experiments testing GABA effect on Nb38 binding, 1 mM GABA was supplemented to the running buffer. Measurements were performed by injecting nanobodies at concentrations ranging from 2.5 to 40 nM (two-fold serial dilutions) during a single cycle. For reliable single cycle kinetics (SCK) data fitting, the final dissociation phase was set to 15 minutes. Biacore T200 evaluation software was used to analyse all the SCK data. A 1:1 binding model was used to fit the experimental results.
HEK cell preparation and electrophysiology
One day prior to experiments, 8 ml of Dulbecco’s Modified Eagle Medium (DMEM) was pre-incubated for 10 min at room temperature with 96 μl lipofectamine2000 (Thermofisher) and 48 μg plasmid DNA, then added to a single T175 cm2 flask containing HEK293T cells (30-50 % confluency) and 2 ml DMEM (supplemented with 10 % fetal calf serum, L-Gln and non-essential amino acids). After 3 hrs this media was removed and replaced by DMEM supplemented with 10 % fetal calf serum. For GABAAR heteromers, pHLsec plasmids containing human cDNA constructs were mixed in 1:1:2 ratio (a:β:γ), supplemented with 3% plasmid encoding enhanced green fluorescent protein (EGFP) to assess transfection efficiency. The electron cryo-microscopy a1β3γ2 construct used was as described in ‘Construct design’. The wild-type cDNA inserts used for heteromeric receptor expression were as follows: human GABAAR a1 mature protein sequence (a1 Uniprot P14867 entry, Gln28 is Gln1, 1-429 QPSL…PTPHQ) and human β3 mature protein sequence (β3 Uniprot P28472 entry isoform 1, Gln26 is Gln1, 1-448 QVSN…LYYVN) cloned into the pHLsec vector34 between the N-terminal secretion signal sequence and a double stop codon; Human GABAAR γ2 mature protein sequence (γ2 Uniprot P18507 isoform 1 entry Gln40 is Gln1 1-428 QKSDD…SYLYL) cloned into the pHLsec vector34 between the N-terminal secretion signal sequence followed by streptavidin binding protein (MDEKTTGWRGGHVVEGLAGELEQLRARLEHHPQGQREP) and a C-terminal 1D4 purification tag. Transfection efficiencies were typically 50-80 % (cells expressing EGFP, as estimated by fluorescence microscopy). Eighteen to twenty-four hours later cells were washed with phosphate buffered saline, incubated in 4 ml TrypLE (Gibco) for 7 min at 37 °C, suspended in 21 ml DMEM supplemented with 10 % fetal calf serum and L-Gln, centrifuged at 100 g for 1.5 min, then suspended in 50 ml Freestyle 293 Expression Medium (Gibco) and placed in a shaking incubator (130 rpm, 37°C, 8 % CO2) for 30 min. 25 ml cell suspension was then centrifuged at 100 g for 1.5 min, and suspended in 4 ml external recording solution. This solution contained (mM): 137 NaCl, 4 KCl, 1 MgCl2, 1.8 CaCl2, 10 HEPES, and 10 D-Glucose, pH 7.4 (≈ 305 mOsm). The internal recording solution contained (mM): 140 CsCl, 5 NaCl, 1 MgCl2, 10 HEPES, 5 EGTA, 0.2 ATP, pH 7.35 (≈ 295 mOsm). Electrophysiological recordings were performed at room temperature using an Ionflux16 (Molecular Devices) in ensemble mode, with series resistance compensation set at 80 % and cells held at −60 mV. Diazepam and picrotoxin (Sigma-Aldrich) were dissolved in DMSO as 100 mM and 1 M stocks respectively prior to dilution in external recording solution. Diazepam and Nb38, or picrotoxin, were coapplied with EC10-15 GABA or EC70 GABA doses respectively to generate dose response curves. Expression of heteromeric receptors as assemblies of αβγ subunits was confirmed by response to GABA, which requires co-assembly of α1 and β3 subunits, and efficient inclusion of the γ2 subunit into αβγ heteromers was verified by measuring low sensitivity to 100 μM Zn2+ inhibition, defined as less than 50 % inhibition of an EC50 GABA response127.
Brain slice preparation and electrophysiology
All work on animals was carried out in accordance with the UK Animals (Scientific Procedures) Act 1986 under project and personal licenses granted by the UK Home Office. 250 μm thick transverse acute brain slices were prepared from adult (P67-84) male C57BL/6J mice using a Leica VT1200S vibroslicer. After terminal anaesthesia with isoflurane the brain was rapidly removed and sliced in an ice-cold slicing solution composed of (mM): 85 NaCl, 2.5 KCl, 1 CaCl2, 4 MgCl2, 1.25 NaH2PO4, 26 NaHCO3, 75 sucrose, and 25 glucose, 2 kynurenic acid, pH - 7.4; bubbled with 95% O2 and 5% CO2. The slicing solution was exchanged at 37 °C for 60 min under gravity flow with a recording solution or artificial cerebrospinal fluid (ACSF) containing (mM): 125 NaCl, 2.5 KCl, 2 CaCl2, 1 MgCl2, 1.25 NaH2PO4, 26 NaHCO3, 2 kynurenic acid, and 25 glucose, pH - 7.4.
Spontaneous inhibitory postsynaptic currents (sIPSCs) were recorded from hippocampal dentate gyrus granule cells (DGGCs) using patch electrodes of 4-5 MΩ resistance filled with an internal solution containing (mM): 120 CsCl 1 MgCl2, 11 EGTA, 30 KOH, 10 HEPES, 1 CaCl2, and 2 K2ATP; pH - 7.2. sIPSCs were recorded at −60 mV using 5 kHz filtering and optimal series resistance and whole-cell capacitance compensation.
sIPSCs were analysed using WinEDR and WinWCP (John Dempster, University of Strathclyde, UK). Percentage change in IPSC amplitudes, frequencies, rise times, decay times and charge transfers were calculated in the presence of nanobody or diazepam in comparison to controls in normal recording solution (aCSF). The statistical significance of sIPSC parameters in the presence or absence of treatments in each cell was assessed using a paired t-test in Graphpad Instat.
Natural genetic variation in ligand-binding residues
To investigate the extent of natural genetic variation in the BZD binding site, ligand-binding residues were identified using inhouse written Perl scripts, available upon request from authors. Residues making atomic contacts with the ligands within 5Å distance were classified as ligand-binding residues. The number of such contacts were also calculated and are provided in the asteroid plots128. Conservation profiles for ligand binding positions across all human GBRA and GBRG subunits were then generated. We retrieved genetic variation data in different GBRA and GBRG receptors in 138,632 unrelated healthy humans from gnomAD database54. Genetic variation data was retrieved only for non-engineered positions and those that have identical residues as that of GBRA5 or GBRG2.
ACKNOWLEDGMENTS
We thank staff at the Diamond Light Source beamline I04 for synchrotron assistance; K. Harlos and T. Walter for technical support with crystallization; R. Masiulyte for assistance in cryo-EM particle picking; G. Murshudov, I. Tickle and G. Bricogne for advice regarding anisotropic X-ray data processing; S. Scheres and D. Tegunov for advice regarding cryo-EM data processing; Y. Zhao for tissue culture advice; E. Beke for technical assistance during nanobody discovery; M. Duta for help at the Oxford Advanced Research Computing facility; D. Laverty for comments on the manuscript. This work was supported by the UK Biotechnology and Biological Sciences Research Council grants BB/M024709/1 (A.R.A. J.T.H. and P.S.M.) and BB/N000145/1 (S.J.T. and M.S.P.S.); UK Medical Research Council grants MR/L009609/1 (A.R.A.), MC_UP_1201/15 (A.R.A. and S.M.) and MC_U105185859 (M.M.B. and S.C.); Wellcome Trust studentships 084655/Z/08/Z (S.M.) and 105247/Z/14/Z (S.S.); Human Frontier Science Program grant RGP0065/2014 (A.R.A.) and long-term postdoctoral fellowship LT000021/2014-L (T.M.); Cancer Research UK grant C20724/A14414 (T.M.); European Research Council grant 649053 (J.T.H.); EPSRC grant EP/R004722/1 (M.S.P.S.) and Wellcome grant 208361/Z/17/Z (M.S.P.S.). We thank INSTRUCT, part of the European Strategy Forum on Research Infrastructures and the Research Foundation-Flanders (FWO) for funding nanobody discovery. Further support from the Wellcome Trust Core Award 090532/Z/09/Z is acknowledged. The OPIC electron microscopy facility was founded by a Wellcome Trust JIF award (060208/Z/00/Z) and is supported by a WT equipment grant (093305/Z/10/Z).