Summary
Type A GABA receptors (GABAARs) are the principal inhibitory receptors in the brain and the target of a wide range of clinical agents, including anesthetics, sedatives, hypnotics, and antidepressants. However, our understanding of GABAAR pharmacology has been hindered by the vast number of pentameric assemblies that can be derived from a total 19 different subunits and the lack of structural knowledge of clinically relevant receptors. Here, we isolate native murine GABAAR assemblies containing the widely expressed α1 subunit, and elucidate their structures in complex with drugs used to treat insomnia (zolpidem and flurazepam) and postpartum depression (the neurosteroid allopregnanolone). Using cryo-EM analysis and single-molecule photobleaching experiments, we uncover only three structural populations in the brain: the canonical α1β2γ2 receptor containing two α1 subunits and two unanticipated assemblies containing one α1 and either an α2, α3 or α5 subunit. Both of the noncanonical assemblies feature a more compact arrangement between the transmembrane and extracellular domains. Interestingly, allopregnanolone is bound at the transmembrane α/β subunit interface, even when not added to the sample, revealing an important role for endogenous neurosteroids in modulating native GABAARs. Together with structurally engaged lipids, neurosteroids produce global conformational changes throughout the receptor that modify both the pore diameter and binding environments for GABA and insomnia medications. Together, our data reveal that GABAAR assembly is a strictly regulated process that yields a small number of structurally distinct complexes, defining a structural landscape from which subtype-specific drugs can be developed.
Main
Regulation of brain excitability by activation of neuronal GABAARs is essential for normal brain development and function1–3. Deficits in GABAAR activity are associated with health problems ranging from epilepsy to intellectual disability4. A large number of ions and small molecules modulate GABAAR activity, including Zn2+ (ref. 5) and picrotoxin6 as well as therapeutic agents such as benzodiazepines6–8, barbiturates8, and propofol8. Indeed, the GABAAR modulators flurazepam and zolpidem are widely used to treat insomnia. Lipophilic neurosteroids are also potent endogenous modulators of GABAARs. Allopregnanolone (ALP; 3α-OH-5α-pregnan-20-one), synthesized chiefly in the brain9–11, potentiates GABAAR activity in a subunit-dependent manner12–14, and its anxiolytic and sedative effects have proved to be effective for the treatment of postpartum depression15. In addition, ganaxolone, a synthetic mimetic of allopregnanolone, has recently entered the clinic as an anticonvulsive agent.
Because modulation of receptor function is dependent upon subunit composition and arrangement, a knowledge of native GABAAR architecture is crucial to understand how these different molecules elicit distinct physiological responses. However, the potential diversity of pentameric GABAARs is vast due to the existence of 19 different receptor subunits (α1-6, β1-3, γ1-3, ρ1-3, δ, ε, π, and θ). Moreover, studies in vitro suggest that variations in subunit expression levels can modify subunit stoichiometry. Despite progress in resolving the architecture of recombinant di- and tri-heteromeric GABAARs5, 7, 16–18, there is no structural understanding of the various GABAARs that are present in the brain. Indeed, although the presence of specific subunits in native receptor assemblies has been determined by immunoprecipitation19–21, the number and arrangement of subunits remains unknown.
To elucidate the ensemble of GABAARs that define the molecular action of endogenous and therapeutic modulators, we isolated native α1 subunit-containing GABAARs (nα1GABAARs) from mouse brain using an engineered high-affinity, subunit-specific Fab fragment7. Because the α1 subunit is ubiquitously expressed throughout the brain, and is a subunit of both synaptic and extrasynaptic receptors, this approach enabled us to analyze 60%–80% of native GABAARs 22–24. Furthermore, it permitted an investigation of three clinically-relevant molecules, allopregnanolone (ALP), didesethylflurazepam (DID), and zolpidem (ZOL), all of which target α1-containing receptors. With isolated nα1GABAARs in hand, we were able to count the number of α1 subunits in these complexes by single-molecule fluorescence bleaching experiments, investigate protein composition by mass spectrometry, and elucidate high resolution structures of nα1GABAAR assemblies by single-particle cryo-EM. Our results reveal a surprisingly small number of receptor complexes, whose structures provide a framework from which targeted drugs could be developed.
Isolation of functional nα1GABAARs from the mouse brain
We engineered the 8E3 α1 subunit-specific Fab fragment7 to include a GFP fluorophore, affinity tag, and 3C-protease site to enable release it the affinity resin (8E3-GFP; Kd = 0.5 nM; Extended Data Figure 1). The 8E3-GFP Fab was then used to isolate nα1GABAARs from solubilized mouse brain tissue (excluding the cerebellum) while monitoring the purification workflow via GFP fluorescence. Detergent (lauryl maltose meopentyl glycol) treatment routinely solubilized the majority of nα1GABAARs, accompanied by the inhibitory synapse marker neuroligin2 (Extended Data Figure 1). Following nearly complete capture of receptors on affinity resin (Extended Data Figure 1), nα1GABAARs complexes were reconstituted into lipid-filled nanodiscs25 then eluted by 3C-protease treatment. Further purification by size exclusion chromatography yielded an ensemble of nα1GABAAR:Fab complexes. Radioligand binding assays showed that the purified pentameric preparations were functional and retained high-affinity flunitrazepam binding (Kd = 6.0 ± 0.2 nM (mean ± s.e.m.); Extended Data Figure 1)26. Furthermore, analysis of the purified native receptor complexes by mass spectrometry identified all α and β subunits as well as the γ1, γ2, and δ subunits, demonstrating that α1-dependent isolation captured receptors containing most of the 19 GABAAR subunits.
nα1GABAARs comprise three structural populations
To elucidate the composition and arrangement of native receptors, we collected cryo-EM data from nα1GABAAR:Fab complexes in the presence of DID, ZOL plus GABA, and ALP plus GABA (Extended Data Table 1), and carried out single particle analysis. The 2D class averages derived from all three datasets showed prominent Fab features at the periphery of the receptors. In contrast to a previous study on recombinant α1-containing tri-heteromeric GABAAR complexes in which all receptors contained two α1 subunits7, we observed class averages with only one Fab bound (Extended Data Figure 2–4), demonstrating the presence of receptors with a single α1 subunit.
We subsequently used extensive 3D classification to rigorously define subunit composition and arrangement of nα1GABAAR:Fab complexes. An inverse mask of the entire transmembrane domain (TMD) allowed us to exclude structural heterogeneity in the region of the pore and enabled classification to be driven by the α1-specific Fab and N-glycosylation patterns unique to each α, β, and γ subunit. After combining classes with the same Fab and N-glycosylation features, we consistently obtained three different 3D classes: a single class with two Fabs (two-Fab) and two classes with one Fab (one-Fab) (Extended Data Figure 2–4). We defined the two one-Fab classes as meta-one-Fab and ortho-one-Fab according to the relative position of their α1 and γ subunits. In all three classes from all three data sets we observed two α subunits, two β subunits, and one γ subunit arranged in an α*-β-α-β*-γ clockwise order when viewed from the extracellular side of the membrane (asterisks denote subunits adjacent to the γ subunit). This pentameric configuration therefore represents the dominant form of nα1GABAARs. Remarkably, we found no evidence for receptors with a β-β interface despite the high abundance of β subunits in native receptor assemblies, in contrast to recombinant α1-β3-α1-β3-β35 and δ-containing18 assemblies. Thus, heterologous expression of GABA subunits appears to yield receptors in configurations that are not abundant in native brain tissue.
In the ALP/GABA dataset, 3D reconstructions at resolutions of 2.5 Å, 2.6 Å, and 2.6 Å were achieved for the two-Fab, ortho-one-Fab, and meta-one-Fab assemblies, respectively. This resolution was sufficient for subunit identification, small molecule positioning, and model building (Figure 1a; Extended Data Figures 5–6; Extended Data Table 2). The identities of β and γ subunits were determined from a combination of glycosylation patterns and sidechain densities. Receptors in the two-Fab class had a tri-heteromeric α1*-β2-α1-β2*-γ2 arrangement, consistent with genetic27, immunohistochemistry21, and electrophysiology28 data suggesting that this is the most abundant subtype in the brain.
In contrast, each of the meta-one-Fab and ortho-one-Fab classes contained mixed receptor ensembles that we categorized as α2/3/5*-β1/2-α1-β1/2*-γ2 and α1*-β1/2-α2/3/5-β1/2*-γ2, respectively. A single α1 subunit in one position and either α2, α3, or α 5 at the second α position, together with either β1 or β2 at the β position, yielded receptors with at least four and as many as five unique subunits in the pentameric assembly. Despite subunit ambiguity in the density maps, we evaluated the overall agreement between density map and protein sequences and modeled the α2/3/5 subunit from one-Fab structures as α3 and β1/2 as β2 to facilitate structural comparison among datasets. The common presence of a γ2 subunit in α1-containing GABAARs suggests a favorable association between these two highly expressed subunits. Furthermore, the highly ordered N-glycosylation of α subunits observed in the extracellular vestibule of the two-Fab class7 is conserved in both one-Fab classes, and includes a polysaccharide bridge between the γ2 subunit and the non-adjacent α subunit. Intriguingly, α2/3/5 subunits may have a fucose sugar attached to the asparagine-linked N-acetylglucosamine, which is absent in the α1 subunit (Extended Data Figure 6).
The two one-Fab classes comprise 45% of particles in the ALP/GABA dataset and 38% particles in the ZOL/GABA dataset (Figure 1; Extended Data Figure 2–4), demonstrating that receptors containing only one α1 subunit are more abundant than previously thought29–31. To independently measure the α1 subunit stoichiometry within nα1GABAARs, we measured photobleaching of the GFP fluorophore in purified 8E3-GFP complexes using single-molecule total internal reflection fluorescence (TIRF) microscopy32. Roughly 50% of photobleaching events comprised a single step (Figure 1b), indicating that about half the purified receptors have just one α1 subunit, in agreement with our cryo-EM data.
We used the two-Fab and ortho-one-Fab structures from the ALP/GABA dataset as paradigms to compare interdomain arrangements in receptors containing one or two α1 subunits. Despite containing highly homologous subunits (74% sequence similarity between α1 and α3), we observed striking differences between the one-Fab and two-Fab complexes. The extracellular domains (ECD)s and TMDs are almost identical in α1 and α3 subunits in equivalent positions, having backbone RMSDs of 0.45 Å and 0.35 Å, respectively. However, when aligned by TMD, the RMSD of ECD increases to 1.05 Å, suggesting significant inter-domain displacement between these α1 and α3 subunits. Furthermore, both meta-one-Fab and ortho-one-Fab have markedly shorter separations between the ECD and TMD center of masses (50.1/50.6 Å for meta-one-Fab α3*/α1 and 50.9/51.0 Å for ortho-one-Fab α1*/α3) than two-Fab complexes (51.9/52.0 Å for α1*/α1 subunits). This observed shortening is also apparent as a reduction of angles between the primary axes of the ECDs and the TMDs in one-Fab receptors (Extended Data Figure 7).
ALP is a ubiquitous modulator of nα1GABAARs
Neurosteroids, such as ALP and allotetrahydrodeoxycorticosterone (THDOC), are endogenous ligands that confer anxiolytic, sedative, hypnotic, and anesthetic properties by potently and selectively potentiating GABAARs, and by direct activation at higher concentrations (≥ 100 nM)33–36. To investigate the molecular basis of neurosteroid modulation, we compared the structures of two-Fab, meta-one-Fab and ortho-one-Fab assemblies in complex with GABA and ALP (Figure 2a). In the two-Fab structure, two ALP molecules are bound in the TMD region, each approximately 60 Å ‘below’ one of the two GABA binding pockets in the ECD (Figure 2b). The ALP pockets are at the interface between transmembrane helices 1 and 4 (TM1 and TM4) of an α1 subunit and TM3 of the adjacent β2 subunit, which form an almost rectangular box lined by primarily aromatic and hydrophobic residues: α1-W245 on one side, β2-Y304 and β2-L301 at the base, and β2-L297, α1-V242 and α1-I238 on another side. Remarkably, lipid acyl chains are present on the other two long sides and the box is capped by α1-P400 and α1-Q241, the amide oxygen of the latter forming a hydrogen bond with the 3’-OH of ALP (Figure 2c).
Incorporation of ALP remodels the conformation of TMD helices, enlarging the channel’s pore compared to a recombinant α1β3γ2 structure without neurosteroid (PDB 6I53). Local alignment of the β2 and α1* TMDs in the two structures reveals a 2.7° rotation of the line connecting the Cα atoms at the base and top of the ALP box (β2-Y304 and α*1-Q241) while the length of this line remains constant. In addition, the α1* TMD rotates by 2.8° around an axis between the center of mass of the entire TMD and the center of mass of the α*1 TMD (Figure 2d). We observed a similar but smaller effect at the β2*/α1 ALP box, with an α1 TMD rotation of 1.8°, suggesting that the two ALP pockets in the pentamer have a different molecular pharmacology. Global TMD alignment, on the other hand, highlights a greater tilt of the M2 helices with respect to the pore axis, collectively yielding an enlarged and more symmetric ion pore in our ALP-bound structure compared to that without ALP (Figure 2e; Extended Data Figure 8). In particular, the sidechains of the 9’-Leu residues, which are crucial for channel gating, are rotated out of the pore in the presence of ALP (Figure 2e).
Neurosteroids achieve GABAAR potentiation by enhancing the ability of agonists to gate the channel12, 37, 38. Such enhancement must be due to allosteric rearrangements in the GABA-binding ECDs, which we indeed observe in our structure. Specifically, global TMD alignment reveals a concerted ∼2° (between 1.5° and 2.5°) counter-clockwise rotation of individual ECDs compared to the ALP-free structure when viewed from the extracellular side (Extended Data Figure 8). This likely accommodates expansion of the TMDs via interactions between the ECD Cys loops and TMD TM2-TM3 loops. Although GABA binding remains largely unchanged (Extended Data Figure 8), concerted ECD rotations may pose an additional energy barrier to GABA release, thus slowing its unbinding and increasing channel gating. Furthermore, because agonist-induced gating is known to be accompanied by counter-clockwise rotation of the ECDs6, our observed conformational changes are fully compatible with allosteric potentiation of nα1GABAARs by ALP. Thus, despite both molecular models in this structural comparison being in a desensitized state with the 2’ gate closed, ALP-induced remodeling of TMDs and ECDs explains how the receptor opens more readily in the presence of ALP. Furthermore, our data suggest that direct activation by neurosteroids is mediated via the same two binding pockets, as no additional ALP molecules were resolved in samples prepared with ALP concentrations as high as 5 μM.
Neurosteroids have unusually slow on- and off-rates compared to more hydrophilic ligands39, 40. This behavior has been attributed to their lipophilic nature and tendency to be enriched in the membrane41, but consideration of lipids in our ALP-bound structure offers an additional explanation for this phenomenon. An annulus of lipids with distorted acyl tails completely buries ALPs in their binding sites. In total, we resolve nine lipid-like molecules at the β2+/α1- interface (+ denoting the principal face and – denoting the complementary face), three being less than 5 Å from ALP (Figure 2f). As a consequence, ALP molecules must coordinate with the motions of these annular lipids to secure an exit pathway from the pocket, and partially disassociated ALP molecules may effectively re-engage the receptor without leaving the pocket via the housing provided by these lipids.
In addition to their prevalence in the ALP binding pockets, lipids structurally engage the receptor at other sites. The greater TMD tilt in our ALP-bound structure creates five inter-subunit pockets near the center of the membrane’s plane. All five pockets, including two general anesthetic binding sites, are occupied with lipid tails bent like a snorkel (Figure 2g). Collectively, these lipids serve as small wedges that stabilize the expanded conformation of the TMD.
Both meta-one-Fab and ortho-one-Fab have ALP bound at their β2+/α3*- or β2*+/α3- pockets, demonstrating that neurosteroid binding at the β+/α- interface is independent of subunit identity and arrangement within the pentamer (Extended Data Figure 7). Consistent with this notion, residues involved in binding ALP are conserved in all β and α subunits. Thus, we propose that neurosteroid potentiation of all nα1GABAARs with a β+/α- interface involves a mechanism similar to the one we have described for ALP binding to the native tri-heteromeric α1β2γ2 receptor. Nevertheless, our structures suggest there may be differences in potency or efficacy at each of the neurosteroid binding sites. Although the sequences forming the immediate ALP pockets are identical, the W245 residue (α1 numbering) in other α subunits adopts a different sidechain conformer, and can serve as a longer and more effective lever for ALP to reshape the TMD and potentiate receptor activity, consistent with previous electrophysiology experiments42, 43.
Strikingly, we observed similar neurosteroid densities in the ZOL/GABA dataset in the absence of added neurosteroid, which are best modeled as ALP molecules (Extended Data Figure 8). Although we did not locate any distinct neurosteroid densities in the DID dataset, this is likely due to the inferior map resolution. Analysis of our purified ZOL/GABA sample by high performance liquid chromatography and mass spectrometry confirmed the absence of neurosteroid in the buffer and lipids used for protein purification. However, the same analysis uncovered 115 ng/mL (362 nM) of neurosteroid in the ZOL/GABA cryo-EM sample (containing ∼250 nM pentameric receptor), with more than 95% being ALP rather than another of the other three possible stereoisomers (Extended Data Figure 9). The nearly identical TMD configuration of the ALP/GABA and ZOL/GABA structures supports this chemical assignment (Extended Date Figure 7). Thus, endogenous ALP co-purifies with nα1GABAARs and its stoichiometric presence in our native receptor structures highlights its abundance in the brain and high affinity for nα1GABAARs relative to other endogenous neurosteroids.
DID and ZOL augment GABA-induced rearrangements
GABAARs are the target of a range of insomnia medicines, including flurazepam and ZOL. To investigate the molecular effects of insomnia treatments on nα1GABAARs, we examined the interactions with either DID, the major metabolite of flurazepam44 (Ki 16.9 ± 1.7 nM), or ZOL (Ki 22.9 ± 2.7 nM) and nα1GABAARs (Figure 3a and 3b). Both compounds engage the receptor ECD at the α1*+/γ2- interface, which is spatially equivalent to the GABA pockets, each sandwiched at a β+/α- interface. The binding of DID in the two-Fab dataset is reminiscent of recombinant GABAAR structures in complex with diazepam/alprazolam6. DID makes extensive interactions with the receptor, including a hydrogen bond between its carbonyl and the α1-S204 sidechain; two hydrogen bonds between its A ring chloride and the α1-H101 and γ2-Ν60 sidechains; and several π-π/CH interactions with α1-F99, α1-Y159, α1-Y209, and γ2-Y58 (Figure 3c).
ZOL binds to the α1*+/γ2- ECD interface in tri-heteromeric α1β2γ2 receptors at roughly the same position as DID, but engages α1-H101 via π-CH interactions rather than a hydrogen bond. In addition, its amide oxygen forms a hydrogen bond with the α1-S204 sidechain, and the imidazole nitrogen forms a separate hydrogen bond with the α1-T206 sidechain (Figure 3d). We hypothesize that this hydrogen bond duet is preserved in interactions with α2 and α3 subunits but not the α5 subunit in which a threonine residue substitutes for S204. This difference would provide an explanation for the greater than ten-fold weaker affinity of ZOL for α5-containing receptors45. Like DID, ZOL forms π-π interactions with α1-Y159, α1-Y209, and γ2-Y58, as well as γ2-F77. The latter interaction explains why ZOL is more sensitive to the γ2-F77I mutation than diazepam46. During the preparation of this manuscript, a recombinant GABAAR structure in complex with ZOL was published47, revealing a similar binding pose for ZOL in the ECD. This structure also captured ZOL in the general anesthetic binding pockets at the β2+/α1- TMD interface using a similar ZOL concentration to that used in our study. We hypothesize that remodeling of the TMDs by endogenous ALP prevented ZOL from binding to the general anesthetic pockets in nα1GABAARs.
Binding of DID or ZOL causes only moderate conformational changes in their binding pocket. We observed a slight opening of loop C due to a 1.2 Å displacement of the γ2-S205 Cα, as well as sidechain reorganization of α1-H101, γ2-Y58, γ2-Ν60, and γ2-F77, which enlarges the pocket to accommodate the ligand. These subtle changes suggest that ZOL-like medications (Z-drugs) potentiate GABAARs via a benzodiazepine-like mechanism, namely, strengthening of the α1*+/γ2- interface and facilitation of GABA-induced ECD rotation6. Indeed, when the TMDs of the ZOL/GABA structure were aligned to the closed, resting structure6, the ECDs showed concerted counterclockwise rotations ranging from 2 to 5° for individual ECD centers of mass (Figure 3e).
We also observed ZOL binding to the α1*+/γ2- (ortho-one-Fab) and α3*+/γ2 (meta-one-Fab) ECD interfaces. Despite sequence differences, the immediate α3*+/γ2 pocket shares the same chemical environment as the α1*+/γ2- pocket, but the sidechains adopt different conformations. Accordingly, we observed significantly different structural consequences of ZOL binding in the α3*+/γ2 pocket, including a binding pose closer to loop C on the α3 subunit and concerted shifts of the ligand and the protein (Extended Data Figure 7). As mentioned above, our data suggest that α2/3/5 subunits have a greater intrinsic bend between their ECD and TMD than α1 subunits, causing different global rearrangements when incorporated into the pentamer. Although this variation in ECD/TMD coupling causes relatively small structural perturbations to orthosteric and allosteric ligand binding, it has the potential to affect channel gating and ligand modulation, which depend on inter-domain cross-talk.
Intriguingly, the minimum pore radius in the DID structure is 2 Å, large enough to pass dehydrated Cl- with a radius of 1.81 Å48 (Figure 3f). The capture of this potentially conductive state, which has not been observed before, is likely due to prevention of GABA-induced desensitization (GABA was omitted during DID sample preparation) and potentiation by DID (micromolar concentrations of benzodiazepines potentiate GABAARs49, 50). Accordingly, we observed incomplete loop C closure – the structural hallmark of GABA-dependent allostery – in the GABA binding pockets in the DID structure.
Conclusion
Our study reveals that nα1GABAARs comprise three structural populations: tri-heteromeric α1β2γ2 receptors that constitute half the total population and two distinct assemblies containing one α1 subunit and one α2/3/5 subunit. Because only three distinct receptor assemblies were identified from a total of 72 possible arrangements of two α, one β, and one γ subunit18, we propose that neuronal assembly of nα1GABAARs is a highly regulated process. The finding of exclusively α-β-α-β-γ2 assemblies is in contrast to the diversity of subunit combinations that are formed in recombinant expression systems, where β-β interfaces and assemblies with two γ subunits are observed18. Although our cryo-EM study validates the use of tri-heteromeric α1β2γ2 receptors as a model of nα1GABAAR pharmacology, it also reveals the prevalence of mixed α subunit receptors in the brain and challenges the conventional practice of classifying GABAARs according to α subunits.
The molecular structures of ligand-nα1GABAAR complexes have revealed the binding poses of the postpartum depression medication, allopregnanolone, and two insomnia drugs, didesethylflurazepam and zolpidem. Our work also highlights the conformational changes induced by neurosteroid binding to native receptors and thus the structural basis for neurosteroid-dependent positive modulation (Figure 4). Finally, the serendipitous finding that endogenous neurosteroids remain bound to nα1GABAARs after isolation and purification emphasizes the importance of considering background neurosteroid modulation when investigating the pharmacology of nα1GABAARs.
Author Information
Hongtao Zhu
Present address: Institute of Physics, Chinese Academy of Sciences, Beijing, China
Authors and Affiliations
Vollum Institute, Oregon Health and Science University, Portland, OR, USA
Chang Sun, Hongtao Zhu, Sarah Clark, Eric Gouaux
Contributions
C.S. and E.G. designed the project. C.S. prepared cryo-EM samples, carried out biochemical characterizations. C.S. and S.C. performed single-molecule photobleaching experiments. C.S. and H.Z. carried out the cryo-EM data analysis and C.S. built the molecular models. C.S. and E.G. wrote the manuscript.
Data availability
The cryo-EM maps and coordinates for the native GABA receptor in complex with didesethylflurazepam and endogenous GABA (two-Fab-DID) have been deposited in the Electron Microscopy Data Bank (EMDB) under accession number EMD-29728 and in the Protein Data Bank (PDB) under accession code 8G4O. The cryo-EM maps and coordinates for the native GABA receptor in complex with zolpidem, GABA, and endogenous neurosteroids have been deposited and accessed via EMD-39727/8G4N (two-Fab-ZOL), EMD-29743/8G5H (ortho-one-Fab-ZOL), EMD-29742/8G5G (meta-one-Fab-ZOL). The cryo-EM maps and coordinates for the native GABA receptor in complex with GABA, and allopregnanolone have been deposited and accessed via EMD-29350/8FOI (two-Fab-ALP), EMD-29741/8G5F (ortho-one-Fab-ALP), EMD-29733/8G4X (meta-one-Fab-ALP).
Methods
Expression and purification of the α1 specific 8E3-GFP Fab
The α1-specific mouse monoclonal antibody 8E3 was generated and screened as previously described1. The coding sequences of 8E3 Fab light and heavy chains were determined from hybridoma mRNA, and a construct to express the Fab portion of the antibody was designed by including sequences to encode an N-terminal GP64 signal peptide. Codons were optimized for expression in insect cells. To facilitate recombinant antibody detection and purification, a 3C-cleavage sequence, an EGFP gene, and a twin-strep II tag were added to the C-terminus of the heavy chain. Synthetic genes for both chains were then cloned into the pFastBac-Dual vector under the polyhedrin promoter. The recombinant baculovirus was prepared as previously described2. Sf9 cells at a density of 3 million per mL were infected with the recombinant baculovirus, with a multiplicity of infection of 2, and further cultured for 96 hours at 20° C. The antibody-containing supernatant was collected by a 20-minute centrifugation at 5,000 g and then the pH was adjusted to 8 with 30 mM Tris base, incubated in the cold room overnight to allow non-Fab protein precipitation, and clarified by another 20-minute centrifugation at 5,000 g. The supernatant was concentrated and buffer-exchanged three times with TBS (20 mM Tris 150 mM NaCl pH 8) using a tangential-flow concentrator equipped with a 15-kDa filter. The concentrated supernatant was then loaded onto a 15-mL streptactin column, which was washed with at least 20 column volumes of TBS and eluted with 5 mM desthiobiotin in TBS. Selected fractions were pooled, concentrated, and buffer exchanged to TBS using microconcentrators with a 50-kDa cutoff. Concentrated 8E3-GFP Fab (∼100 μM) was aliquoted and stored at –80 ° C until use.
Purification of nα1GABAARs from mouse brains
One-month-old BL/6 mice of mixed sex (∼50 mice per preparation) were used for native receptor isolation. The mice were first euthanized and decapitated. The whole brain was isolated from the skull using a laboratory micro spatula and stored in ice-cold TBS. Cerebella were removed from the whole brain, frozen in liquid nitrogen, and stored at -80 ° C for a separate study. After being washed twice with ice-cold TBS, brain tissue was resuspended with ice-cold TBS (1 mL per brain) supplemented with 0.2 mM phenylmethyl sulfonyl fluoride (PMSF). The suspension was processed with a loose-fit Potter-Elvehjem homogenizer for 20 full up-and-down strokes and further sonicated (1 min per 50 mL) at a setting of 6, typically at a 40 W output. The suspension was centrifuged at 10,000 g for 10 minutes, resulting in a hard pellet of mainly the nuclear fraction and a “runny” soft pellet containing a significant amount of nα1GABAARs. The supernatant was further centrifuged at 200,000 g for 45 minutes to pellet the membranes. About 0.1 g of hard pellet and 0.2 g of soft pellet were obtained from one mouse brain. These membrane pellets were resuspended with an equal volume of TBS buffer containing protease inhibitors (aprotinin/leupeptin/pepstatin A/PMSF). If not used right away, the 50% membrane suspension was supplemented with 10% glycerol and snap frozen in liquid nitrogen.
The following membrane solubilization and affinity chromatography were all carried out at 4 ° C. First, MNG/CHS (10:1 w/w) stock (10% w/v) was diluted in TBS buffer containing protease inhibitors to 2.5%. Then, one volume of the 50% membrane suspension was mixed with two volumes of the diluted detergent stock and incubated for 1 hour on a platform rocker, which routinely resulted in the solubilization of ∼60% of the α1 subunit present in the tissue, estimated based on Western blot (Extended Data Figure 1). Next, Biolock solution was added at 0.1 mL per brain to quench the naturally biotinylated proteins, and the mixture was clarified by centrifugation at 200,000 g for 1 hour. Finally, the 8E3-GFP Fab was added to the solubilized membrane to a concentration between 60 nM and 100 nM. After 1 hour incubation, 3 mL of pre-equilibrated streptactin resin was added to bind the 8E3-GFP Fab and associated nα1GABAARs, for 2 hours in batch mode.
On-column nanodisc reconstitution
MSP2N23 or a recently engineered MSP1E3D1 variant, CSE34, was used for on-column MSP nanodisc reconstitution. The affinity resin, bound with receptor complexes, was washed in batches, first with 20-CV of ice-cold TBS, then with 20-CV of TBS containing 0.05% MNG and 0.01% brain polar lipid (Avanti). During this wash, 40 nmole MSP2N2 and 3.2 μmole POPC:bovine brain extract (Sigma) (85:15) lipids, or 40 nmole CSE3 and 4.8 μmole lipids were mixed to a final volume of 1 mL in TBS and incubated at room temperature for 30 minutes. The beads were transferred to an empty Econo-Pac gravity flow column to drain the buffer. Then the 1 mL pre-incubated MSP:lipids were added and incubated for 1.5 hours. Next, biobeads were added to a 20x weight excess to the MNG detergent. The mixture was incubated with a rotator in the cold room for at least 4 hours. The biobead/resin mixture was washed with 20-CV of ice-cold buffer to remove unbound empty nanodiscs.
Two approaches were used to elute reconstituted nanodiscs: competitive ligand elution and protease cleavage. For ligand elution, 0.5 CV 5 mM desthiobiotin dissolved in TBS was incubated with the streptactin superflow resin for 10 minutes before gravity elution, which was repeated for a total of 6 times. In the case of 3C cleavage, 0.1 mg 3C protease was first diluted to 50 μg/mL with 2 mL TBS and added to the resin. After a 2-hour incubation in the cold room, the elution was collected, and the column was further washed three times with 2 mL TBS to improve the protein yield. 3C protease cleavage offered better protein purity and was used for the zolpidem and the allopregnanolone samples. Pooled elution was concentrated to about 0.5 mL using a 50-kDa cutoff centricon, regardless of the elution methods. The concentrated sample was then injected into a Superose 6 column pre-equilibrated with TBS supplemented with 1 mM GABA and other ligands. Selected fractions corresponding to the nα1GABAAR:Fab complex were combined and concentrated to about 0.1 mg/mL using a centricon with a 50-kDa cutoff.
Mass-spec protein identification
The protein mass-spec analysis was carried out as previously described.5 The native receptor samples were diluted into 100 μl 1% SDS, reduced with Tris (2-carboxyethyl) phosphine hydrochloride (TCEP), and alkylated with iodoacetamide. Proteins were then extracted with methanol-chloroform, mixed with 30 μl of 20 μg/mL trypsin dissolved in 50 mM ammonium bicarbonate, and incubated overnight at 37 °C. The next day, the solvent from the trypsin digestion was evaporated using a speed vac. The resulting pellet was resuspended in 20 μl of 2% acetonitrile (ACN) and 0.1% formic acid (FA) for LC-MS. The sample was run on a Dionex U3000 nanoflow system coupled to a Thermo Fusion mass spectrometer. Each sample was subjected to a 65-min chromatographic method using a gradient from 2–25% acetonitrile in 0.1% formic acid (ACN/FA) for 16 min; from 25% to 35% ACN/FA for an additional 15 min, from 35% to 50% ACN/FA for an additional 4 min, a step to 90% ACN/FA for 4 min and a re-equilibration into 2% ACN/FA. Chromatography was carried out in a ‘trap-and-load’ format using a PicoChip source (New Objective); trap column C18 PepMap 100, 5 µm, 100 A, and the separation column was PicoChip REPROSIL-Pur C18-AQ, 3 µm, 120 A, 105 mm. The entire run was at a flow rate of 0.3 µl/min. Electrospray was achieved at 1.9 kV. The MS1 scans were performed in the Orbitrap with a resolution of 240,000. Data-dependent MS2 scans were performed in the Orbitrap using High Energy Collision Dissociation (HCD) of 30% using a resolution of 30,000. Data analysis was performed using Proteome Discoverer 2.3 using SEQUEST HT scoring. The static modification included dynamic modification of methionine oxidation (+15.9949) and a fixed modification of cysteines alkylation (+57.021). Parent ion tolerance was 10 ppm, fragment mass tolerance was 0.02 Da, and the maximum number of missed cleavages was set to 2. Only high-scoring peptides were considered, using a false discovery rate (FDR) of 1%.
Isotope-dilution quantification of allopregnanolone using LC-MS/MS
Neurosteroids (allopregnanolone, epipregnanolone, isopregnanolone, pregnanolone) and isotope-labeled internal standard allopregnanolone-d5 were purchased from Toronto Research Chemicals (Toronto, ON, Canada). The O-(3-trimethylammonium-propyl) hydroxylamine quaternary amonoxy (QAO) reagent used for derivatization was in the form of Amplifex Keto reagent kit from AB Sciex (Framingham, MA). Solvents for liquid chromatography-tandem mass spectrometry (LC-MS/MS) analysis were from VWR (Tualatin, OR).
Neurosteroid stocks and internal standard (INST) were prepared in methanol. Stocks (5 μL) and the INST allopregnanolone-d5 (5 μL) were mixed with PBS (95 μL) to prepare standard samples with final concentrations ranging from 0.05 to 100 ng/ml. All standards and samples were treated with 1000 µl of acetonitrile, vortexed and mixed using Benchmark Multi-Thermo heat/shaker at 1500 rpm at 22°C for 5 mins, and centrifuged to remove protein at 12,000 g for 5 mins. The supernatant was dried under vacuum and then treated with 75 µl of derivatization reagent. The keto moiety was derivatized with QAO reagent to form a cationic oxime derivative to enable highly sensitive LC-ESI-MS/MS quantification of neurosteroids. The working derivatization reagent was prepared according to vendor instructions. The derivatized samples were diluted 1:4 with 5% acetic acid in methanol before LC-MS/MS analysis. The supernatant was placed in sample vials for analysis by LC-MS/MS using an injection volume of 5 µl. The lower limit of quantification of allopregnanolone was 75 pg/ml, with an accuracy of 101% and a precision of 2.2%.
The samples with INST were analyzed using a Sciex 4000 QTRAP hybrid/triple quadrupole linear ion trap mass spectrometer (Foster City, CA) with electrospray ionization (ESI) in the positive mode. The mass spectrometer was interfaced to a Shimadzu HPLC system (Columbia, MD) with SIL-20AC XR auto-sampler, LC-20AD XR LC pumps, and CTO-20AC column oven. Compounds were quantified with multiple reaction monitoring (MRM). The MS/MS transitions used were optimized by infusion of pure derivatized compounds with method settings, as presented in the table below. The bold transitions were used for quantification, with other transitions used for peak qualification to ensure method specificity. Allopregnanolone was separated from interferents using a Luna 5u C8(2) 50×2 mm column (Phenomenex) kept at 35 °C using a column oven. The gradient mobile phase was delivered at a flow rate of 0.8 ml/min and consisted of two solvents: solvent A (0.1% formic acid in water) and solvent B (0.1% formic acid in acetonitrile). The initial concentration of solvent B was 20%, followed by a linear increase to 60% B in 10 min, then to 95% B in 0.1 min, held for 3 minutes, decreased back to starting 20% B over 0.1 min, and then held for 2 min. The retention time was 3.99 min for allopregnanolone and pregnanolone, 3.64 min for isopregnanolone, and 3.61 min for epipregnanolone. Data were acquired using Analyst 1.6.2 and analyzed with MultiQuant 3.0.3 software.
To further distinguish allopregnanolone and pregnanolone, a different HPLC condition was used. In this case, a Poroshell 120 EC-C18 100×2.1 mm 2.7um column (Agilent) was kept at 35 °C using a column oven. The gradient mobile phase was delivered at a flow rate of 0.4 ml/min (0–5.9min), 0.2ml/min (6.0–8.9min), and 0.4ml/min (9–15min), and consisted of two solvents, A: 0.1% formic acid in water, B: 0.1% formic acid in acetonitrile. The initial concentration of solvent B was 30%, followed by a linear increase to 52% B in 6.5 min, held for 2.5min, then to 95% B in 0.1 min, held for 2.9 minutes, decreased back to starting 30% B over 0.1 min, and then held for 2.9 min. The retention time for allopregnanolone was 6.4 min, pregnenolone was 6.2 min, and 3α-allopregnanolone-d5 was 6.3 min.
Single-molecule photobleaching of nα1GABAAR-Fab complexes
Coverslips and glass slides were extensively cleaned, passivated, and coated with methoxy polyethylene glycol (mPEG) and 2% biotinylated PEG as previously described6. A flow chamber was created by drilling 0.75 mm holes in the quartz slide and placing double-sided tape between the holes. A coverslip was placed on top of the slide, and the edges were sealed with epoxy, creating small flow chambers. A concentration of 0.25 mg/mL streptavidin was then applied to the slide, incubated for 5 minutes, and washed off with buffer consisting of 50 mM Tris, 50 mM NaCl and 0.25 mg/mL bovine serum albumin (BSA), pH 8.0. Biotinylated anti-GFP nanobody at 7.5 µg/mL was applied to the slide, incubated for 10 minutes, and washed off with 30 µL buffer A (20 mM Tris, 150 mM NaCl, pH 8) supplemented with 0.2 mg/mL BSA. nα1GABAAR-Fab complexes in nanodiscs were eluted from the streptactin-XT resin with biotin instead of 3C protease cleavage to preserve the GFP moiety. The sample was further FSEC-purified, and the peak corresponding to the complex was hand collected, which separated the native receptor from free Fab. The sample was diluted 1:30 to about 50 pM based on fluorescence quantitation, applied to the chamber, and incubated for 5 minutes before being washed off with 30 µL of buffer A. The chamber was immediately imaged using a Leica DMi8 TIRF microscope with an oil-immersion 100x objective. Images were captured using a back-illuminated EMCCD camera (Andor iXon Ultra 888) with a 133 x 133 µm imaging area and a 13 µm pixel size. This 13 µm pixel size corresponds to 130 nm on the sample due to the 100x objective.
Photobleaching movies were acquired by exposing the imaging area for 180 seconds. Single-molecule fluorescence time traces of nα1GABAAR-Fab were generated using a custom python script. Each trace was manually scored as having one to three bleaching steps or was discarded if no clean bleaching steps could be identified. A total of ∼450 molecules were evaluated from three separate movies. Scoring was verified by assessing the intensity of the spot; on average, the molecules that bleached in 2 steps were twice as bright as those that bleached in 1 step.
Scintillation proximation assay
YSI Copper SPA beads from PerkinElmer were used to capture the nα1GABAAR in nanodisc via the MSP His-tag. Tritiated flunitrazepam from PerkinElmer was used as the radioligand, and clorazepate was used as the competing ligand to estimate background. During the ligand binding assay setup, nα1GABAAR in nanodisc was first mixed with SPA beads and radioligand (2x bead) while the ligand of different concentrations (2x ligand) and competing ligand (2x background) were prepared using serial dilution. Then, an equal volume of 2x bead was mixed with 2x ligand (in triplicate) or 2x background in a 96-well plate. The final concentrations were 0.5 mg/mL for SPA beads, ∼1 nM for native receptors, 10 nM for 3H-flunitrazepam, and 0.5 mM for clorazepate in the background wells only. The plate was then read with a MicroBeta TriLux after a 2-hour incubation. Specific counts were then imported into GraphPad and analyzed using a one-site competition model.
Negative-stain electron microscopy
Purified nα1GABAAR:Fab complex in nanodiscs was first diluted with TBS to a concentration of ∼0.05 mg/mL. Continuous carbon grids were glow-discharged for 60 seconds at a current of 15 mA. A protein sample (5 μL) was applied to the carbon side of the grid held with a fine-tip tweezer and incubated for 10–30 seconds. The excessive sample was then wicked away from the side with a small piece of filter paper. The grid was quickly washed with 5 μL deionized water, followed by side-wicking, which was repeated for a total of three times. Immediately afterward, the grid was incubated with 5 μL 0.75% uranium formate for 45 seconds, wicked several times from the side, and dried for at least 2 minutes at room temperature.
Cryo-EM sample preparation and data acquisition
We employed a specific setup to prepare grids under different buffer and ligand conditions. First, buffers containing 10x ligand or additive were first prepared and dispensed in 0.5 μL aliquot into PCR tubes. Then, 5 μL purified nα1GABAAR:Fab complex was added and quickly mixed by pipetting. Within 10 seconds, a 2.5 μL sample was applied to a glow-discharged (30 seconds at 15 mA) 200 mesh gold Quantifoil 2/1 grid overlaid with 2-nm continuous carbon and incubated for 30 seconds. The grid was blotted with a Mark IV Vitrobot under 100% humidity at 16 °C and flash-frozen in liquid ethane. For the didesethylflurazepam sample, no GABA was included during the purification, and the didesethylflurazepam (2 μM) was added prior to vitrification. For the zolpidem sample, 1 mM GABA was included throughout the purification, and 5 μM zolpidem was added prior to vitrification using the above-mentioned PCR tube method. For the allopregnanolone sample, 1 mM GABA and 5 μM allopregnanolone were included from the membrane solubilization to the final size-exclusion chromatography.
Cryo-EM data were collected on a 300-keV Titan Krios equipped with a BioQuantum energy filter at either PNCC or the Janelia cryo-EM facility. Data acquisition was automated using serialEM: defoci ranged between 0.9 to 2.5 μm, holes with suitable ice thickness were selected with the hole finder and combined to produce multishot-multihole targets, which allowed the acquisition of six movies per hole in each of the neighboring nine holes. These movies were captured with a K3 direct electron detector. A total dose of 50 electron/Å2 was fractionated into 40 frames, with a dose rate of about 15 electron/(pixel*second) for non-CDS mode or 7 electron/(pixel*second) for CDS mode (Extended Data Table 1).
Cryo-EM data analysis
Super-resolution movies were imported to cryosparc7 v 3.3.1 and motion corrected using cryosparc’s patch motion correction with the output Fourier cropping factor set to ½. Initial contrast transfer function (CTF) parameters were then calculated using cryosparc’s patch CTF estimation. For each dataset, 2D class averages of particles picked by glob-picker from ∼1000 micrographs were used as templates for the template picker. One round of 2D classification and several rounds of heterogeneous refinement seeded with ab initio models generated within cryosparc were used to select GABAAR particles, ranging from 4 to 6 million particles for our datasets. A non-uniform refinement (NU-Refinement) was performed to align these particles to a consensus structure. Two downstream strategies were used for our datasets, as subsequently described.
Data processing strategy #1
Bin 1 GABAAR particles, both images (360×360) and the star file converted using pyem8, were ported into RELION9 3.1. Then a 3D auto refinement job with local search (angular sampling of 1.8 degree) was carried out to fine tune the particle poses in RELION. The refined structure, similar to that generated by cryosparc, had relatively weaker γ subunit transmembrane helices, which was reported previously10. To tackle this issue, we prepared a nanodisc mask in Chimera11 and carried out 3D classification without alignment (15 classes, T=20) using that mask. The 3D classification can robustly give classes with much stronger transmembrane helices of the γ subunit. Those selected particles were imported into cryosparc and further refined using NU-Refinement with both defocus refinement and per-group CTF refinement options turned on. The consensus structure was a two-Fab bound structure, but earlier data processing revealed one-Fab species’ presence. Therefore, a 3D classification job was used with a mask focusing on the two binding sites of 8E3 Fab to isolate the one-Fab species. The one-Fab and two-Fab particles were separately refined with NU-Refinement and further refined with local refinement.
Data processing strategy #2
In this strategy, the heterogeneity in Fab binding was addressed upstream in the data processing pipeline. Like strategy #1, GABAAR particles, at bin3 or 120×120, were imported into RELION for focused 3D classification. A reverse mask was prepared in cryosparc which only excluded the transmembrane domain to allow for Fab binding at all possible positions. 3D classification (10 classes, T=20) gave clear two-Fab and one-Fab classes, and classes with incomplete Fab. Further 3D classification on these incomplete Fab particles produced only incomplete Fab classes, which led us to believe they were damaged particles and should be excluded from downstream processing. The two-Fab particles and the one-Fab particles, on the other hand, were imported into cryosparc, re-extracted at bin1, and separately refined with NU-Refinement. Still, we saw weak transmembrane helices for the γ subunit for the one-Fab and the two-Fab populations. To tackle this issue, instead of the 3D classification in RELION, we used the 3D classification (beta) job in cryosparc with a nanodisc mask, which was less robust but faster. Classes with stronger transmembrane helices were then combined and refined with NU-Refinement and finished with local refinement.
Global sharpening worked sub-optimally for our nα1GABAAR structures because of the local resolution variation and the lower signal-to-noise ratio for the transmembrane domain. The best method to sharpen our maps was achieved with LocScale12, which was used to represent some of our structures in Figure 1. DeepEMhancer13 can yield comparable sharpening for the protein but not for the annulus lipids.
Subunit identification, model building, refinement, and validation
Due to the subunit specificity of 8E3-Fab, the subunit with 8E3-Fab bound is defined as α1. The remaining subunits can be easily classified as α, β, or γ from each subunit’s characteristic N-linked glycosylation patterns. It was clear that all 3D classes obtained are α-β-α-β-γ, clockwise, when viewed from the extracellular side of the membrane. Given the relative subunit abundance from earlier studies, we used α1-β2-α1-β2-γ2 as the starting model of the two-Fab class. We then examined the cryo-EM density maps to test our assignment in the context of sequence information. Specifically, we looked at regions where the sidechain can be unambiguously assigned and positions where a difference of more than 3 carbon atoms or one sulfur atom was found within the subunit group. Regarding the non-α1 α subunit in the one-Fab classes, we further limited our scrutiny to positions showing no significant conformational differences in the corresponding two-Fab structure to ensure the observed density difference was caused by the chemical identity of underlying residues.
For each dataset, the two-Fab bound nα1GABAAR model was built first. The starting structures used were Alpha-fold14 models of mouse GABAAR subunits and the best 8E3 Fab model generated with Rosetta15. These individual chains were first docked into the unsharpened cryo-EM density maps using chimera’s fit-in-map tool to assemble the full receptor-Fab complex. The full complex was then edited to remove unresolved portions and refined extensively to achieve better model-map agreement in Coot16. N-glycosylation was modeled using Coot’s carbohydrate module. Lipid and lipid-like molecules, including POPC, PIP2, dodecane, and octane, were modeled using the CCP4 monomer library. New ligands included in this study, including their optimized geometry and constraint, were generated using phenix.elbow17. After the initial modeling, multiple runs of phenix.real_space_refinement18 and editing in Coot were carried out to improve the model quality.
The optimized two-Fab GABAAR structure was used as the starting model for one-Fab GABAAR structures. Although the one-Fab population likely consists of a mix of α2/3/5 subunits, we decided to use α3 for the modeling because of its best overall agreement with the density maps. The two-Fab structure was first docked into the one-Fab cryo-EM map using the “fit in map” tool of chimerax19. Then the aligned structure was edited in Coot to remove the extra-Fab, replaced and renumbered the α1 sequence with the α3 sequence. This edited structure was further fitted and refined in Coot, first with secondary structure restraints generated with ProSMART20, and then without the restraints. Furthermore, certain residues and lipids were removed due to less clear density, and the glycosylation trees were remodeled. Similarly, this initial model was subjected to multiple runs of phenix.real_space_refinement and editing in Coot.
Animal use statement
Mouse carcass donated from other labs of the Vollum Institute were used to establish and optimize the native GABAA receptor isolation workflow. The quantity of purified native receptor from each mouse was estimated using the fluorescence from the recombinant antibody fragment, which was then extrapolated to give the minimum number required for cryo-EM and biochemical analysis. For each native GABAA receptor preparation, 50 one-month-old (4–6 weeks) C57BL/6 mice (both male and female) were ordered from Charles River Laboratories. No randomization, blinding or experimental manipulations were performed on these animals. All mice were euthanized under Institutional Animal Care and Use Committee (IACUC) protocols, consistent with the recommendations of the Panel on Euthanasia of the American Veterinary Medical Association (AVMA) and carried out only by members of the E.G. laboratory approved on IACUC protocol TR03_IP00000905.
Cell line statement
Sf9 cells for generation of baculovirus and expression of recombinant antibody fragment are from Thermo Fisher (12659017, lot 421973). The cells were not authenticated experimentally for these studies. The cells were tested negative for Mycoplasma contamination using the CELLshipper Mycoplasma Detection Kit M-100 from Bionique.
Extended Data Figures
Acknowledgements
We thank J. Luo and A. DeBarber for mass-spec neurosteroid analysis, J. Guidry for mass-spec protein identification, D. Claxton and D. Cawley for the monoclonal antibody. We thank the use of OHSU Bioanalytical Shared Resource/Pharmacokinetics Core Facility expertise and instrumentation (Research Resource Identifier (RRID): SCR_009963). We thank OHSU Multiscale Microscopy Core (MMC), the Pacific Northwest Cryo-EM Center (PNCC), and the cryo-EM facility at Janelia research campus for microscope use. C.S. thanks J. Myers (PNCC) and C. López (MMC) for cryo-EM training. PNCC is supported by NIH grant U24GM129547 and accessed through EMSL (grid.436923.9), a DOE Office of Science User Facility sponsored by the Office of Biological and Environmental Research. This work was supported by NIH grant 5R01GM10040 to E.G. and E.G. is an investigator of the Howard Hughes Medical Institute.