Resolving native GABA_A receptor structures from the human brain

Abstract

GABA_A receptors mediate most fast inhibitory signaling in the brain and are targets for drugs treating epilepsy, anxiety, depression, insomnia, and anesthesia^1,2. These receptors, composed of a complex array of 19 related subunits, form pentameric ligand-gated ion channels. The composition and structure of native GABA_A receptors found in the human brain have been inferred from subunit localization in tissue^1,3, functional measurements, and structural analysis from recombinant expression^4–7 and from mice⁸. Defining which subunits co-assemble physiologically is needed to understand receptor signaling and for development of targeted therapeutics. However, the subunit arrangements in native human GABA_A receptors remain unknown. Here, we isolated α1 subunit-containing GABA_A receptors from epilepsy patients. Using cryo-electron microscopy, we defined a set of twelve native subunit assemblies and their 3D structures. We address inconsistencies between earlier native and recombinant approaches, and reveal details of previously undefined subunit interfaces. Drug-like densities at a subset of these interfaces led us to uncover unexpected activity on the GABA_A receptor by antiepileptic therapeutics and resulted in localizing one to the benzodiazepine site. Proteomics and further structural analysis suggest interactions with the auxiliary subunits neuroligin 2 and GARLH4, which localize and modulate GABA_A receptors at inhibitory synapses. This work provides a structural foundation for understanding GABA_A receptor signaling and targeted pharmacology in the human brain.

GABA_A receptors assemble in homo- and hetero-pentameric combinations of subunits. This combinatorial puzzle generates a broad spectrum of sensitivities to the neurotransmitter GABA, channel biophysical properties, and drug-binding interfaces^9,10. The 19 human receptor subunits are α1–6, β1–3, γ1–3, ρ1–3, δ, θ, π, and ε¹¹. These subunits arrange in a ring surrounding a central ion-conducting channel through the membrane. Binding of GABA to β-α interfaces favors channel opening, which increases chloride conductance, and generally acts to inhibit neuronal activity¹². The α1 subunit exhibits the overall highest expression, and receptors comprised of α1, β2, and γ2 subunits are thought to be the most abundant^1,13.

While a nearly limitless number of subunit combinations is theoretically possible, only a small subset is understood to form physiologically relevant GABA_A receptors¹. In situ hybridization and immunohistochemical approaches have determined the levels of individual GABA_A subunits in human, non-human primate, and rodent brains^10,13–16. Several distinct assemblies have been structurally determined using recombinant systems^4–7. One recent study determined native mouse GABA_A receptor structures⁸. Together, these approaches inform on differences among species and on the likelihood of receptors to contain certain combinations of subunits. However, the absolute arrangements of subunits in human brains remain ambiguous. Notably, the diversity of GABA_A receptor subunit compositions is greater in primates compared to rodents¹⁵, and even more heterogeneous in humans than in non-human primates¹⁶. Another important element of native receptor complexity is regulation by auxiliary subunits. The localization, clustering, and anchoring of GABA_A receptors are modulated by synaptic proteins including GARLHs, neuroligin 2 (NL2), gephyrin, and neurexins^17–20. Despite their importance, we have almost no structural understanding of how they interact with the intact receptor.

Here, we explored the compositions and arrangements of native GABA_A receptors in the human brain using a high-affinity antibody fragment (Fab) targeting the α1 subunit, mass spectrometry, and cryo-EM. We collected tissue samples from 81 patients with temporal or frontal lobe epilepsy; the samples were pooled into two groups for receptor analysis. Our study revealed a minimum of 12 distinct α1-containing GABA_A receptor assemblies, comprising α1, α2, α3, β1, β2, β3, and γ2 subunits, defining an extensive repertoire of native structures. Several of the structures exhibited drug-like signal at subunit interfaces that led us to uncover modulator activity for two epilepsy drugs not known to act on the GABA_A receptor. Cryo-EM mapped binding of one, lamotrigine, to the α1-γ2 benzodiazepine site. Additionally, we observed non-receptor densities adjacent to an α1-γ2 subunit interface in the transmembrane domain, suggesting an interaction with GARLH4 and NL2, supported by mass spectrometry.

Receptor purification from resected brain tissue

Drug resistant epilepsy can be treated with surgical resection²¹. This procedure commonly results in the removal of otherwise healthy tissue outside of the epileptic focus, verified by histopathology to confirm absence of structural abnormalities in the tissue for research. To investigate the composition of native GABA_A receptors in the human brain, we employed anti-α1 Fab²² (Fab_1F4) affinity purification in conjunction with cryo-EM, leveraging nominally healthy neurosurgical specimens obtained from epilepsy patients. The surgical resections primarily targeted the temporal lobe, with a smaller portion from the frontal lobe; collections spanned from 2016–2023 (Fig. 1a). Tissue collection procedures are detailed in the Methods section. Tissue samples (0.5–6 g each) were pooled into two groups sufficient for purification. The first contained tissue from 45 patients, and the second from 36 patients. The pooled samples are from patients ranging from 19 to 77 years old, of which 39.5% were female and 60.5% were male. Further details regarding patient demographics can be found in Extended Data Tables 1 and 2.

Figure 1|. Patient brain tissue samples, approach, and proteomics data.

a, Locations and demographics of tissue samples. The regions of the human brain collected from patients are distinguished by two colors: green for the frontal lobe and yellow for the temporal lobe. The first dataset includes 45 patients: 22 females and 23 males, in blue. The second dataset includes 36 patients: 10 females and 26 males, in red. Distribution of samples is indicated. b, Overview of tissue processing (cross-section of brain areas utilized indicated in inset with colors relating to panel a) and receptor purification. Inset table shows mass spectrometry results from gel band (red box) from α1-containing GABA_A receptors in the brain.

The resected tissues were homogenized to isolate cell membranes (Fig. 1b). Receptors were then extracted using lauryl maltose neopentyl glycol to preserve interactions with synaptic binding partners²³. Native GABA_A receptors were isolated using Fab_1F4, which targets the α1 subunit, known for its high expression level and its ability to assemble with other subunits including α2–6, β1–3, γ1–3, and δ^8,24–27. Subsequently, native receptors were reconstituted into spMSP1D1-lipid nanodiscs²⁸. This nanodisc scaffold was selected for its ability to produce larger, circularized nanodiscs, which may minimize artifacts in ion channel conformation²⁹. We confirmed the presence of isolated α1-containing GABA_A receptors using mass spectrometry followed by preparation of cryo-EM grids. Thirteen GABA_A receptor subunits were identified: α1–6, β1–3, γ1–3, and δ (Fig. 1b, Extended Data Figs.1a and b). Among them, α1, β2, and γ2 comprised the largest proportion, which is consistent with reports that these three subunits co-assemble into the most abundant GABA_A receptor pentamer population^1,3. As a method with remarkable sensitivity, mass spectrometry suggests numerous potential compositions of α1-containing GABA_A receptors in the human brain. However, it does not provide insights into stoichiometries and assemblies.

Heterogeneity in subunit assemblies

To determine the subunit compositions of major α1-containing GABA_A receptor subtypes, we conducted an analysis on two large single-particle cryo-EM datasets derived from tissue samples from 45 patients and 36 patients, referred to as Datasets 1 and 2, respectively (Extended Data Table 3). The protein samples for both datasets were prepared following the procedures outlined in Fig. 1b, with the Dataset 2 sample also undergoing chemical crosslinking to stabilize interactions with synaptic binding partners or auxiliary subunits; this modification resulted in a very minor improvement in map quality but no noticeable differences in receptor conformations. Two-dimensional (2D) classification revealed two major subsets of GABA_A receptor particles: one with receptors bound by two Fabs and another with particles bound by a single Fab (Extended Data Figs. 2-4), suggesting variable numbers of α1 subunits within the receptors.

We employed several rounds of focused three-dimensional (3D) classification, followed by non-uniform refinement to resolve distinct cryo-EM density maps corresponding to unique subunit arrangements within pentameric receptors. The classification was focused on the extracellular domain (ECD), which includes strong asymmetric features such as bound Fabs and subunit-specific N-glycosylation patterns. Ultimately, we obtained a total of seven distinct cryo-EM density maps from Dataset 1 with overall resolutions ranging from 2.5–3.3 Å (Fig. 2 and Extended Data Fig. 2), and seven maps from Dataset 2 (2.7–3.3 Å overall resolution; Fig. 2 and Extended Data Fig. 3). Three compositions overlapped between the datasets; particles from these subunit assemblies were combined to generate the final maps (Extended Data Fig. 5).

Figure 2|. Architecture of native human α1-containing GABA_A receptors.

a-c, Cryo-EM maps of the three GABA_A receptor assemblies present in both Datasets 1 and 2. d-g, Cryo-EM maps obtained from Dataset 1. h-k, Cryo-EM maps derived from the Dataset 2. Each panel includes a schematic cartoon illustrating the composition, subunit arrangement and Fab_1F4 binding. It also includes the overall cryo-EM map top and side views, resolution, and percentage. The percentage is derived from the fraction of particles that gave rise to the map of that assembly (Extended Data Figs. 2 and 3).

The eleven high-resolution maps enabled distinction of subunits through progressively finer levels of analysis. Subunit identities were first assigned based on selective binding of Fab_1F4 to α1, followed by identification of the distinct N-glycosylation patterns among α, β, γ, and δ subunits. The relatively weak density of the γ-subunit transmembrane domain (TMD)^4,8 was further used in coarse identification of subunits. Next, finer features were examined: the presence of GABA at β-α interfaces, and the density of residues with side chains that can be unambiguously assigned to receptor subunits (Supplementary Figs. 1-11). In all cases, subunit assignment and 3D classification were performed iteratively, with ambiguity at a subunit position leading to further classification. Five maps indicated the presence of a mixture of two subunits at one position that could not be separated (Fig. 2c, e, f, h, i). In these cases, models corresponding to unique subunit assemblies were built and analyzed. In total, these eleven distinct cryo-EM maps enabled building of twelve models of unique subunit assemblies (Extended Data Tables 4 and 5).

The predominant subunit composition in both datasets was β2-α1-β2-α1-γ2 (Fig. 2a, Extended Data Figs. 5a and 6a, Supplementary Fig. 1), aligning with prior functional investigations^1,3 and a recent structural study from mouse brain⁸. There is good agreement in receptor conformations among the closely related native and recombinant protein structures (Extended Data Fig. 7). The second most abundant population assembled as β2-α1-γ2-β2-α2, resulting in a previously undefined α2-β2 interface (Fig. 2b, Extended Data Figs. 5b and 6b, Supplementary Fig. 2). Despite being expressed at lower levels than α1 in the brain, α2 holds promise as a drug target for neuropsychiatric illnesses, substance abuse, and pain³⁰. A similar composition included a mixture of β2 and β3 at one position, resulting in β2/3-α1-β2-α2-γ2 (Fig. 2c, Extended Data Figs. 5c and 6c, Supplementary Fig. 3). The β3-containing model is the first to report β3 assembling into native α1-containing GABA_A receptors. Despite the close similarity to β2 (80% amino acid identity), β3 has unique functional properties that distinguish it in GABA_A receptor-mediated inhibition^31,32.

Beyond the three major assemblies identified in both datasets (Fig. 2a-c), eight maps were reconstructed uniquely from separate analysis of Datasets 1 and 2. Dataset 1 contained an easily distinguishable β2-α1-β2-α2-γ2 assembly that overlapped with one of the mixed compositions common to both datasets (Fig. 2d, Extended Data Fig. 6d, Supplementary Fig. 4). This finding underscores the abundance of receptors comprising both α1 and α2 subunits, and reveals details of receptors with an α2-γ2 benzodiazepine site. Dataset 1 also contained a minor population of β2-α1-β2/3-α2-γ2, indicating rare α1-β3 and β3-α2 subunit interfaces (Fig. 2e, Extended Data Fig. 6e, Supplementary Fig. 5). The compositions containing α3 subunits are also heterogeneous with a clear mixture at one position, β2-α1-β2-α2/3-γ2. (Fig. 2f, Extended Data Fig. 6f, Supplementary Fig. 6). The α3 subunit is implicated not only in regulating sleep³³ but also specifically in anxiety, with distinct roles compared to α2³⁴. The last composition in Dataset 1, β2-α1-γ2-β1-α2, represents a small fraction of particles classified, yet it effectively illustrates the heterogeneity present, with 5 different subunit types assemble into a single receptor molecule (Fig. 2g, Extended Data Fig. 6g, Supplementary Fig. 7). The last three subtypes were not found in Dataset 2, likely due to different limits of detection in the two sets of pooled samples, or differences among the sampled patient sets.

Both datasets revealed assemblies containing mixtures of β subunits. In Dataset 2, the β2-α1-β2/3-α1-γ2 map shows β3 occupying a pentameric position typically held by β2, which may stem from β3’s close similarity to β2, but not to β1 (Fig. 2h, Extended Data Fig. 6h, Supplementary Fig. 8). We also identified a mixed composition of β3-α1-γ2-β2/β3-α2 (Fig. 2i, Extended Data Fig. 6i, Supplementary Fig. 9), a new stoichiometry incorporating two β3 subunits. A third assembly from Dataset 2, β2-α1-β1-α2-γ2, has the same stoichiometry as one found in Dataset 1, but with γ2 and β1 subunits swapping positions (Fig. 2j, Extended Data Fig. 6j, Supplementary Fig. 10). All previously mentioned assemblies adhere to a common arrangement of two α subunits, two β subunits, and one γ subunit. However, the final composition in Dataset 2 deviates from this pattern. The β2-α1-β1-β1-γ2 features a β1-β1 interface (Fig. 2k, Extended Data Fig. 6k, Supplementary Fig. 11). The existence of the β-β interface has been implicated in functional studies³⁵ and has also been observed in recombinant GABA_A receptor structures, for instance β3-α4-β3-β3-γ2⁵.

The twelve distinct receptor subunit assemblies stemming from eleven unique maps represent the diversity of native α1-containing GABA_ARs found in both the temporal and frontal lobes of the human brain, providing valuable insights into pharmacology. The β1-β1 subunit interface appears unique to human structures, contrasting with the study in mice. Furthermore, the presence of β3 in α1-receptor assemblies detected by cryo-EM suggests that this subunit in particular is present in higher amounts than in the mouse study, where it was only detected by mass spectrometry⁸. The structural distribution of GABA_A receptor subunits generally corresponds to the findings of mass spectrometry (Fig. 1b). The presence of α4–6, γ1, γ3, and δ subunits in mass spectrometry, but absence in the cryo-EM reconstructions, suggests they are indeed present in α1-containing receptors, but only in minor assemblies beyond our detection limit.

Structures yield insight into epilepsy drug mechanisms

While assigning subunits based on model-map agreement, we noted strong density features at several subunit interfaces in the ECD, in locations analogous to where benzodiazepines bind. In the most abundant and highest-resolution assembly of β2-α1-β2-α1-γ2, a shrimp-shaped density was found at the α1-γ2 benzodiazepine binding site (Fig. 3a)³⁶. In two β3-containing assemblies (β2-α1-β2/3-α1-γ2 and β3-α1-γ2-β2/3-α1), the density in the α1-γ2 interface was similarly strong but elongated (Fig. 3b and c). We superimposed structures of several experimental GABA_A receptor-drug complexes to ask whether these densities may correspond to locations of chemical groups of known active compounds. We found that diazepam^37,38 and the negative modulator DMCM³⁹ partially fit this α1-γ2 density in the most abundant subunit assembly (Extended Data Fig. 8e, f), while DMCM better fits the elongated density at the α1-γ2 interface in the β2-α1-β2/3-α1-γ2 map (Extended Data Fig. 8g, h). The presence of these varied densities across the three compositions suggests that multiple types of ligands may bind at this site, with the density appearing as an average of several components.

Figure 3|. Densities in native α1-containing GABA_A receptors and related pharmacology.

Cryo-EM densities were observed at the interfaces formed by the ECDs of the α1 and γ2 subunits in the composition of β2-α1-β2-α1-γ2 (a), β2-α1-β2/3-α1-γ2 (b) and β3-α1-γ2-β2/3-α2 (c). Thresholds for maps shown are given in the Methods section. Residues surrounding the density and known determinants of benzodiazepine and DMCM binding are labeled. Experimental density located at the interface is shown as a semi-transparent surface. Each panel includes a schematic illustrating the composition. d, Chemical structures of six drugs taken by patients. e, Bar graph displaying potentiation of GABA + 50 μM lamotrigine (LTG, orange) and GABA + 250 μM levetiracetam (LEV, green) on α1β2γ2 and α1β2. Data are represented as mean ± SEM. n=7 (LTG, α1β2γ2), 11 (LEV, α1β2γ2), 5 (LTG, α1β2), 6 (LEV, α1β2) recordings from independent cells. Error bars, mean ± s.e.m.; Welch’s t-test was used. f, Cryo-EM density map and structural model of the recombinant human GABA_A receptor (β2-α1-β2-α1-γ2) in complex with lamotrigine. The receptor subunits are color-coded as follows: α1 (light pink), γ2 (khaki), and β2 (cornflower blue). Glycans are labeled for clarity. g, Close-up views of the lamotrigine-binding pocket, with nearby residues displayed in stick representation.

Structural biology approaches have yet to reveal drug binding to an α-β interface though modulators are suggested to bind at this site^40,41. Here, we observed density features suggestive of small molecule binding (Extended Data Fig. 8a-c). While the contributions of interfaces from these three compositions vary among the subunits (α2/β2, α1/β2/3, and α2/β3), the residues packing around the densities are conserved and adopt similar conformations (Extended Data Fig. 8i). Together, the densities at α-β interfaces are distinct in size and shape, suggesting that these pockets exhibit pharmacological preferences. In contrast to the structurally undefined pharmacology at α-β interfaces, recombinant approaches revealed histamine exerting agonist activity through the β3-β3 interface⁵. In the native structures, density at the β1-β1 interface in the β2-α1-β1-β1-γ2 composition is coordinated by the same set of residues as for histamine in the β3-containing receptor (Extended Data Fig. 8d and j).

The diverse densities detected in the native α1-containing GABA_A receptors prompted us to consider potential candidates for ligands among the antiepileptic drugs taken by the patients, as well as endogenous neurosteroids and endozepines. Neurosteroids target the TMDs of GABA_A receptors^42,43, which are far from the observed densities. The small molecule endozepines are likely not at high enough concentrations physiologically to exert meaningful binding⁴⁴, and the diazepam binding inhibitor (DBI)^45,46 was not detected by mass spectrometry. Among the drugs taken by the 81 patients, six types were prevalent (Fig. 3d). These ligands act through multiple pathways⁴⁷, and were often taken in combination (Extended Data Tables 1, 2). The two most common drugs, lacosamide (35 patients) and lamotrigine (28 patients), act principally through enhancing sodium channel inactivation. Levetiracetam (28 patients) targets synaptic vesicle 2A (SV2A) proteins. Zonisamide (16 patients) inhibits sodium and calcium channels and also potentiates inhibitory glycine receptors⁴⁸. A total of 18 patients took carbamazepine or oxcarbazepine, which principally inhibit sodium channels but also positively modulate α1β2/3γ2 GABA_A receptors^49,50. The final category comprises benzodiazepines, including clobazam and clonazepam, which were taken by 6 patients, and are well defined GABA_A receptor positive modulators^51,52.

We were curious to determine if any of the prescribed drugs in the patient population might contribute to the density observed at native receptor subunit interfaces. We thus tested the activity of the four most commonly prescribed drugs, which were not thought to affect GABA_A receptors: lacosamide, lamotrigine, levetiracetam and zonisamide. We examined their modulation of GABA_A receptors, at pharmacological concentrations⁵³, via patch-clamp electrophysiology of transfected HEK cells. Unexpectedly, lamotrigine and levetiracetam demonstrated small but significant potentiation of the α1β2γ2 receptor (Fig. 3e, Extended Data Fig. 8k, l), while lacosamide and zonisamide had no effect. To test whether the γ2 subunit was important for the activity of lamotrigine and levetiracetam, we tested modulation of GABA responses on the α1β2 GABA_A receptor. Here, these two drugs caused inhibition (Fig. 3e, Extended Data Fig. 8m, n). These findings raise the intriguing possibility that these commonly used anti-epileptic drugs derive some of their efficacy from potentiating synaptic α1β2γ2 GABA_A receptors. Further, the switch from enhancement to inhibition suggests that the β2-β2 interface in the α1β2 receptor, absent in α1β2γ2, may harbor the inhibitory site for these two drugs. Taken together, hints of potential drug occupancy in the native receptor structures led us to discover unpredicted drug activities from classically prescribed anti-epileptic therapeutics. We suggest that the densities in these interfaces represent a mixture of several compounds that indicate the potential of these sites for modulation. To more confidently map binding sites for levetiracetam and lamotrigine, we prepared cryo-EM samples of a recombinant α1β2γ2 receptor in the presence of the drugs. While the results for levetiracetam were inconclusive, we identified unambiguous density for lamotrigine at the α1-γ2 interface, in a position overlapping with where benzodiazepines bind (Fig. 3f, g, Extended Data Fig. 9). Accordingly, this site may underlie a component of the therapeutic efficacy of lamotrigine in treating epilepsy.

Identification of synaptic binding partners

Our first goal in studying native GABA_A receptors was to define the major arrangements of subunits. To that end, we processed cryo-EM data with a focus on subunit-distinguishing features, which are mainly in the ECD. However, in parallel, we conducted classification based on differences in the TMD (Extended Data Fig. 4a). Unexpectedly, analysis of Dataset 1 revealed a map featuring two Fab_1F4 molecules bound to a receptor with five extra helical densities in the TMD. Subsequent refinement yielded a 4.1 Å resolution map (Fig. 4, Extended Data Fig. 4b, c and d). Interestingly, the approximate 5-fold symmetry of the pore was disrupted, likely because the nanodisc scaffold could not otherwise contain the whole receptor plus the additional interacting partner(s) (Extended Data Fig. 4a). Distinctive N-glycosylation patterns along with Fab_1F4 binding allowed us to assign the receptor composition as βx-α1-βx-α1-γx (Fig. 4a). Notably, the unidentified five helices, situated near the interface between the α1 and γ subunits, primarily interact with the TMD of the γ subunit (Fig. 4b and c).

Figure 4|. Proteins associated with native α1-containing GABA_A receptors.

a-c, Cryo-EM map of the GABA_A receptor-TMD partner complex, in three orientations. The interaction partner density is in dark orange. d, Mass spectrometry results of GABA_A receptor-associated proteins from brain tissue. The left and right panels are from two stages of receptor purification. Left, sample in detergent after affinity chromatography, from a pilot purification; right, sample in lipid nanodisc from purification for structure determination. e, Interaction partner density map from c aligned with models of NL2 (residues 674–704, red) and GARLH4 (residues 41–210, orange). f, Cross-section schematic of suggested interactions between the α1 subunit, γx subunit, NL2, and GARLH4 based on b.

The emergence of the new TMD density prompted us to revisit our mass spectrometry findings to aid in identifying the mystery helices. In addition to the many GABA_A receptor subunits, several proteins involved in GABA_A receptor trafficking, clustering, and anchoring co-purified with the receptor (Fig. 4d, Extended Data Figs. 1a and c). A subset of these directly associate with GABA_A receptors. For instance, NL2 and the GARLH family, specifically GARLH4, form complexes with GABA_A receptors through the TMD of γ subunits^23,54. NL2 is a single pass transmembrane protein while GARLH4 has 4 TM helices. Gephyrin helps cluster the receptors but is a soluble protein that interacts with the receptor’s intracellular loop. Neurexins modulate GABAergic synaptic responses by associating with postsynaptic GABA_A receptors¹⁸, but through reaching across from the presynaptic side. There is no evidence for direct interactions of SV2B and SYT1 with GABA_A receptors; also, these proteins are mainly presynaptic. The mostly likely GABA_A TMD-binding candidates among those that co-purify with the receptor are GARLH4 and NL2.

We rigid-body-fitted the TMDs of NL2 (residues 674–704), and GARLH4 (residues 1–210), both obtained from AlphaFold2⁵⁵, in conjunction with the model of β2-α1-β2-α1-γ2, into the density map (Fig. 4e). The combination of NL2 and GARLH4 fit well into the density without needing local refinement. The fitting is consistent with previous functional studies suggesting that GARLH4 bridges the connection between NL2 and the TMD of the γ subunit (Fig. 4f)²³. The spatial arrangement of helices within GARLH4 matches those in an experimental structure of LHFPL5 (a homolog of GARLH4), and the site of NL2 binding is similar to a bound helical partner of LHFPL5⁵⁶. Our findings thus provide a view into likely synaptic interactions of NL2 and GARLH4 with native α1-containing GABA_ARs. However, further validation and investigation are required to elucidate a well-ordered complex structure; at this resolution, side chain densities are unclear.

Discussion

Previous structural investigations have enabled an extraordinary level of understanding of how GABA_A receptors are built and how drugs act on them. Efforts in using recombinant expression systems and studying native receptors from tissue come with their own strengths and limitations. The recombinant studies on major synaptic and extrasynaptic receptor subtypes revealed details at high resolution about several subunit assemblies^4–7. However, they did not directly address whether these assemblies are present physiologically, or whether they are abundant or minor components in brain tissue. The first study on native GABA_A receptor structures was from mouse brains, which mapped three abundant compositions and uncovered the presence of a native neurosteroid modulator⁸. While a critical step in the field, this approach is inherently limited in making inferences about the human brain. The complexity in physiological assemblies was already thought to be greater^15,16, which has been used to explain failures in therapeutic development when translating results from rodents to humans⁵⁷. Our study on human brain tissue directly addresses the outstanding question of what are the major subunit assemblies in the temporal and frontal lobes. The structures led serendipitously to discovering unexpected drug activities, as well as likely interactions with synaptic partners. However, human brain tissue is an exceptionally precious resource. Only very small amounts are available from a single surgery, which precludes comparisons among individuals that could inform on disease mechanisms. There are also regional limitations in obtaining human brain tissue from surgeries, for example, drug-resistant temporal lobe epilepsy is relatively commonly treated with resection, so a higher proportion of the tissue available comes from the temporal lobe. Further, caveats in interpretation can stem from unknown consequences on subunit levels as a function of drug regimen, or individual genetic variants, though we found no evidence for mutations when analyzing the structural data. With the increasing prevalence of non-resective epilepsy treatments, such as laser interstitial thermal therapy and responsive neurostimulator implants, access to neurosurgical specimens may decrease further in the future. Nonetheless, by pooling the tissue from the diverse population of patients, we are able to survey the most abundant human receptor assemblies, which were common to both datasets.

Multiple sources of variation may explain the difference in the minor structural assemblies found in the two datasets, including patient sex, age, duration of epilepsy, site of resection, and drug regimen. Dataset2 includes a higher percentage of males (72%) than dataset 1 (51%); higher levels of some GABA_A subunits are found in male vs. female temporal cortex; in contrast, age related differences have not been found in the relatively young patient population we surveyed^58,59. Comparisons between tissue from epilepsy surgeries outside the epileptogenic zone, and ‘healthy’ tissue removed during brain tumor surgeries, suggest that when pooled from a given patient, these tissues are not significantly different in GABA_A receptor subunit levels⁶⁰. This finding suggests that the presence of the disease, and variations related to the disease (including duration), likely have a minimal impact on our results. The resection sites are balanced between the two datasets, with ~90% for both coming from the temporal lobe. Chronic epilepsy drug treatment, specifically benzodiazepines, can indeed affect GABA_A subunit levels, based on mouse studies⁶¹. Five patients in dataset 1 were chronically prescribed benzodiazepines, versus one patient in dataset 2 (Extended Data Tables 1-2). The rodent studies suggest these drugs can trigger a decrease in α1, β2, α3, and δ subunit levels; differences in subunit assemblies between the two datasets are not in line with a benzodiazepine effect, likely because only a small fraction of the patients was prescribed this specific drug class. Overall, we cannot confidently attribute the differences in subunit compositions between the two datasets to variations in patient demographics or conditions, suggesting instead that these differences are likely due to the detection limits of cryo-EM for minor species, along with other undefined variations within the patient pools.

Understanding GABA_A signaling in the brain rests on knowing the explicit receptor subunit arrangements. Different subunits and interfaces between them tune pharmacology and channel biophysics. There are many other essential pieces of the signaling puzzle, including spatial and cellular distributions of specific subunit assemblies. However, as a starting point, we require a foundation of knowing the extent of heterogeneity in subunit compositions. Here we tackled this basic question by purifying receptors from human tissue. We studied the population of receptors containing the α1 subunit, as this subunit appears to be the most promiscuous in its subunit partners. Based on this focus, we can unequivocally state that the predominant assembly from the human temporal lobe is β2-α1-β2-α1-γ2. Also abundant are assemblies that illustrate profound heterogeneity in compositions. The α1, α2, α3 and β1, β2, β3, and γ2 positions can swap with each other, and assemble in multiple ratios, giving rise to single receptors that contain 5 different subunits, as well as receptors with a β-β interface. The implications for pharmacology are diverse. As one example, drugs deemed or engineered to be specific for α2-containing receptors will likely also bind to receptors that contain an α1 subunit. While the complexity is daunting, it serves as a basis for developing tools targeting the now explicitly defined native interfaces. Emerging from the native structural data are interactions with epilepsy drugs formerly not known to be active on GABA_A receptors, as well as interactions with additional proteins. We find from the mass spectrometry and features in the experimental map that NL2 and GARLH4 are the likely bound protein partners, identified earlier to interact with the receptor in mouse cerebellum¹⁷. Together, our findings illustrate a great level of complexity both in GABA_A receptor subunit assemblies and in assembly with other proteins in the human brain.

Methods

Human neurosurgical tissue

The study included 81 temporal and frontal lobe epilepsy subjects ages 20–70 who were undergoing surgical resection as part of their treatment plan for medication-resistant epilepsy. Participants came from the University of Texas Southwestern (UTSW) epilepsy surgery program across a time span of 7 years. The protocol was approved by the UTSW Institutional Review Board on Human Subjects Research prior to data collection, and all participants provided informed written consent. After resection, the brain samples were rinsed with cold PBS and any tissue damaged by electrocautery was dissected away and discarded. The tissue was then divided into cryo-tubes and flash frozen in liquid nitrogen. All tissue processing was completed on ice within an hour of resection, typically within 20 minutes. Human subject information is summarized in Extended Data Tables 1 and 2.

Fab_1F4 expression and purification

The coding sequence of Fab_1F4, which recognizes the human GABA_A receptor α1 subunit, was determined previously²². DNA fragments encoding the heavy and light chains were synthesized and subcloned into the pCEP4 mammalian expression vector (Invitrogen). A signal peptide from mouse IgG (MGWSCIILFLVATATGVHS) was added to the 5ʹ ends of both genes in place of the original signal peptides to boost expression. To facilitate affinity purification, 3XFLAG and 8xHis tags were added to the C terminus of the heavy chain. Recombinant Fab expression was carried out using the ExpiCHO expression system (Thermo Fisher) by transient transfection. 400 mL ExpiCHO cells were transfected with 0.4 µg total plasmid DNA at a cell density of 7 million per mL at 37 °C, 8% CO₂. The culture was harvested after 9 days by centrifugation for 20 min at 6,000 g. The supernatant was filtered (0.22 μm) and concentrated using a 30 kDa cutoff Vivaflow 200 (Sartorius) after adjusting pH to 8.0. The Fab_1F4 was purified using a 5 ml HisTrap HP column (Cytiva) with binding buffer (50 mM Tris-HCl, pH 8.0, 200 mM NaCl, 20 mM Imidazole, pH 8.0) and elution buffer (20 mM Tris-HCl, pH 8.0, 200 mM NaCl, 500 mM Imidazole, pH 8.0). The eluted fractions were pooled for concentration and gel filtration assay after analyzing by SDS-PAGE gel. The selected fractions (~30 mg) were stored at −80 °C for future use.

Native α1-containing GABA_A receptor purification

Purifications were performed from two sets of pooled samples; samples from 45 patients led to Dataset 1, and samples from 36 patients led to Dataset 2. For purification, the brain tissue was washed 3 times in ice-cold 1X PBS (pH 7.4) by centrifugation for 10 min at 6000 g followed by resuspension. The final pellet was resuspended in ice-cold buffer A (20 mM Tris-HCl, pH 7.4, 200 mM NaCl, 2 mM GABA) supplemented with 0.2 mM phenylmethyl sulfonyl fluoride (PMSF, Sigma-Aldrich). The suspension was processed with a Dounce homogenizer and further disrupted using an Avestin Emulsiflex. This homogenate was centrifuged at 10,000 g for 20 minutes, and the supernatant was further centrifuged at 186,000 g for 2 hours to pellet the membrane. The membrane pellet was harvested, then mechanically homogenized and solubilized with buffer A containing 1% (w/v) lauryl maltose neopentyl glycol (LMNG; Anatrace) and protease inhibitors (aprotinin, leupeptin, pepstatin A, and PMSF) for 3 h. The membrane solubilization and subsequent affinity chromatography were carried out at 4 °C. Next, the mixture was clarified by centrifugation at 186,000 g for 1 hour. Fab_1F4 was added to the supernatant (6 mg for 45-patient sample; 4 mg for 36-patient sample), and incubated for 2.5 h. Then, pre-equilibrated M2 anti-FLAG resin (Sigma-Aldrich) was added to bind the Fab_1F4 and native α1-containing GABA_A receptors for 3.5 h while rotating. The resin was washed with 5 column volumes buffer A supplemented with protease inhibitors, 0.01% (w/v) LMNG, and 0.01% (w/v) brain polar lipids (Avanti).

Receptor-nanodisc reconstitution

The on-bead nanodisc reconstitution was performed with the spMSP1D1 scaffold protein (a gift from Huan Bao, Addgene #172482)²⁸ and brain polar lipids. GABA_A receptor-bound FLAG resin was resuspended with buffer A containing 0.01% (w/v) LMNG and 0.01% (w/v) brain polar lipids and incubated for 30 mins at 4 °C. Then, spMSP1D1 was added and incubated for another 30 mins at 4 °C. The molar ratio of receptor, lipids and spMSP1D1 was 1:100:10. Detergent was removed by adding biobeads SM2 (Bio-Rad) to a final concentration of 60 mg ml⁻¹ while rotating overnight at 4 °C. The next day, excess lipids and spMSP1D1 were removed by washing with buffer A. Next, the Fab_1F4-bound receptors were eluted with buffer A containing 0.15 mg ml⁻¹ 3XFLAG peptide (Sigma-Aldrich). The eluted fractions were concentrated and run over a Superose 6 Increase 10/300 GL column (Cytiva). For just the 36-patient sample, two separate purifications were performed after isolating membranes, to test two different crosslinking reagents. In these cases, after affinity chromatography, the eluted protein was incubated with either 5 mM glutaraldehyde, or 5 mM Bs3 crosslinking agent (both from Sigma-Aldrich) for 2 h prior to preparative size exclusion chromatography. The peak fractions were analyzed by SDS-PAGE, then collected and concentrated to an absorbance at 280 nm (A₂₈₀) of 6.4 for the 45-patient dataset (Dataset 1) and 1.4 for the 36-patient dataset (Dataset 2).

Cryo-EM sample preparation and data collection

Cryo-EM grids were prepared using a Vitrobot Mark IV (FEI). 300-mesh copper R1.2/1.3 holey carbon grids (Quantifoil) were used for Dataset 1. 300 mesh copper R 2/1 overlaid with 2-nm continuous carbon grids (Quantifoil) were used for the Dataset 2. Before freezing grids, 0.5 mM fluorinated Fos-Choline-8 (Anatrace) was mixed with the protein sample to induce random particle orientations. 3 μL of sample were applied to glow-discharged grids and blotted for 3.5 seconds at 4 °C with 100% humidity, and then plunge-frozen in liquid ethane. Electron microscopy images were collected at the UCSD Cryo-EM Facility on the Titan Krios G4 (Thermo Fisher Scientific) at 300 keV equipped with a Gatan BioContinuum energy filter. The total exposure was 50 e⁻/ Ų and the defocus range was set to −2.2 μm to −1.0 μm. Details for both datasets are in Extended Data Table 3.

Cryo-EM data processing

Both datasets were processed using a CryoSPARC 4.4⁶² workflow. First, the images were motion and gain corrected using Patch Motion Correction. Contrast transfer function (CTF) estimation was performed with Patch CTF. For each dataset, ~1 million particles were subjected to 3 rounds of 2D classification to remove junk particles. The good 2D classes were selected to generate Ab initio models with 4 seeds, followed by 2 rounds of heterogenous refinement with junk volumes to further remove bad particles. The retained particles were re-extracted at full size and aligned with Non-Uniform (NU) refinement. The consensus maps were obtained at 2.6 Å (Dataset 1) and 2.8 Å (Dataset 2). To resolve the heterogeneity of native GABA_A receptors, two kinds of strategies, both using focused (masked) 3D classification without alignment were applied, described below. In processing of Dataset 2, we found no differences between the two crosslinking agents mentioned above, so images were combined from both samples.

Dataset 1 processing

The first classification approach was focused on resolving distinct GABA_A receptor subunit assemblies. Focused 3D classification with an ECD mask was carried out to separate classes with one Fab vs. two Fabs bound. Then, particles from 3D classes with two Fabs, or one Fab, were pooled with each other; these particle sets were aligned in 3D using NU refinement. For the two-Fab particles, further focused 3D classification was used to identify the heterogeneity but no indications of mixtures of subunits at any positions were found in the composition of β2-α1-β2-α1-γ2. For the one-Fab particles, focused 3D classification was carried out and followed by NU refinements. The processing workflow is shown in Extended Data Figs. 2. Chain IDs are always represented in an anticlockwise order as follows: βx (chain A)-α1 (chain B)-X (chain C)-X (chain D)-X (chain E) (Supplementary Figs. 1-11). Five types of masks were tested in the secondary 3D classification: a mask focused on chain B; a mask focused on chain D; a mask focused on chain A and B; a mask focused on chain C, D and E; and a mask focused on chain D and E. All the masks were generated with UCSF ChimeraX (1.7.1)⁶³. The focused 3D classification with the mask on chain A and chain B was the most effective in separating particles. This masked classification was followed by further focused 3D classifications, from which 6 compositions were revealed including: β2-α1-γ2-β2-α2, β2/3-α1-β2-α2-γ2, β2-α1-β2-α2-γ2, β2-α1-β2/3-α2-γ2, β2-α1-β2-α2/3-γ2, and β2-α1-γ2-β1-α2.

The second classification approach was designed to investigate interesting possibilities in the transmembrane domain of GABA_A receptors, including compositional or conformational heterogeneity, and interacting proteins. Two rounds of focused 3D classification with a TMD mask, followed by NU refinements, were performed (Extended Data Fig. 4). After the first focused 3D classification, one class with a poorly ordered TMD was identified. Ultimately, density indicative of five TM helices in addition to the TMD of the GABA_A receptor was discovered. Through docking of likely candidates identified in mass spectrometry, GARLH4 and NL2 are the suggested bound proteins.

Dataset 2 processing

Focused 3D classification with an ECD mask was also applied to resolve distinct GABA_A receptor subunit assemblies. Then, 3D classes with two-Fabs, and one-Fab, were pooled with each other; these particle sets were aligned in 3D using NU refinement (Extended Data Fig. 3). For the two-Fab particles, further focused 3D classification was used to resolve heterogeneity, and a total of 2 compositions including β2-α1-β2-α1-γ2 and β2-α1-β2/3-α1-γ2 were identified. For the one-Fab particles, focused 3D classification with a mask on chains A and B was carried out and followed by NU refinements. Classification and refinement of the one-Fab particles yielded 5 subunit assemblies: β2-α1-γ2-β2-α2, β2/3-α1-β2-α2-γ2, β2-α1-β1-α2-γ2, β3-α1-γ2-β2/3-α2, and β2-α1-β1-β1-γ2.

From both datasets, we observed three common compositions: β2-α1-β2-α1-γ2, β2-α1-γ2-β2-α2, and β2/3-α1-β2-α2-γ2. Particles for these compositions, from the two datasets were combined and NU refinement was performed to generate the final maps for these three compositions. For these 3 abundant assemblies, additional processing was performed to further improve map quality, related to a common feature observed in GABA_A receptors, a disordered TMD of the γ subunit. Focused 3D classification without alignment on the entire transmembrane domain was performed (Extended Data Fig. 5), as there were sufficient particles available after combining these particle sets. Particles with a 5-fold symmetric TMD were selected for NU refinement. The overall resolution for all maps was estimated using the gold-standard Fourier shell correlation 0.143 criterion.

Model building, refinement and validation

A four-step procedure was developed to identify and assign the GABA_A receptor subunits in the maps. First, the subunit bound by Fab_1F4 is assigned as α1. Second, identification is based on the N-linked glycosylation patterns of the subunits, such as the hallmark glycosylation site in the Cys-loop unique to β subunits. Third, subunits involved in GABA binding are identified as α and β subunits based on their binding interfaces. Finally, the side-chain density is checked after docking the target subunit models. A total of 11 distinct maps were obtained from these two datasets. Five of these maps contain a currently unresolvable mixture of two subunits at a single position. The full set of maps includes: β2-α1-β2-α1-γ2 (Supplementary Fig. 1), β2-α1-γ2-β2-α2 (Supplementary Fig. 2), β2/3-α1-β2-α2-γ2 (Supplementary Fig. 3), β2-α1-β2-α2-γ2 (Supplementary Fig. 4), β2-α1-β2/3-α2-γ2 (Supplementary Fig. 5), β2-α1-β2-α2/3-γ2 (Supplementary Fig. 6), β2-α1-γ2-β1-α2 (Supplementary Fig. 7), β2-α1-β2/3-α1-γ2 (Supplementary Fig. 8), β3-α1-γ2-β2/3-α2 (Supplementary Fig. 9), β2-α1-β1-α2-γ2 (Supplementary Fig. 10), β2-α1-β1-β1-γ2 (Supplementary Fig. 11). All the details used in subunit assignment are listed in the Supplementary Figures.

For model building and refinement of β2-α1-β2-α1-γ2, the recombinant β2-α1-β2-α1-γ2 with Fab_1F4 bound (PDB ID:6X3T) and β3-α1-β3-α1-γ2 (PDB ID:6I53) were used as the initial models after docking into the map using UCSF ChimeraX. The refined model of β2-α1-β2-α1-γ2 was used as the initial model for the additional 11 models. The map with the composition of β3-α1-γ2-β2/3-α2 was used to build two models: β3-α1-γ2-β2-α2 and β3-α1-γ2-β3-α2. The models underwent a series of iterative adjustments manually in Coot (v.9.8.92)⁶⁴, followed by Phenix (1.20.1) real space refinement⁶⁵ with secondary structure restraints. The density for the γ-TMD was too poor to model that domain of the subunit in the following assemblies: β2-α1-β2-α2-γ2, β2-α1-β2/3-α2-γ2, β2-α1-β2-α2/3-γ2, β2-α1-γ2-β1-α2, β2-α1-β2/3-α1-γ2, β3-α1-γ2-β2/3-α1, β2-α1-β1-α2-γ2; accordingly, in these cases, the TMD of the γ2 subunit was not modeled. Models were validated using MolProbity v4.5⁶⁶. The validation statistics of the PDB models are provided in Extended Data Tables 4 and 5. Subunit interface densities in Fig. 3 and Extended Data Fig. 8 are shown at the following thresholds, rendered using ChimeraX: Fig. 3a, 0.14; Fig. 3b, 0.14; Fig. 3c, 0.14; Extended Data Fig. 8a, 0.15; Extended Data Fig. 8b, 0.21; Extended Data Fig. 8c, 0.12; Extended Data Fig. 8d and j, 0.11; Extended Data Fig. 8e and f, 0.14; Extended Data Fig. 8g and h, 0.14. The pore diameter was calculated by HOLE v.2⁶⁷ and generated using PyMOL (v.2.5.5, Schrodinger, LLC).

Mass spectrometry

Proteins were identified after SDS-PAGE separation by mass spectrometry using standard methods⁶⁸. Proteins from Dataset 1 were analyzed at the UT Southwestern proteomics facility, and proteins from the Dataset 2 at the UCSD proteomics core. Dataset 1 sample were digested overnight with trypsin (Pierce) following reduction and alkylation with DTT and iodoacetamide (Sigma-Aldrich). The samples then underwent solid-phase extraction cleanup with an Oasis HLB plate (Waters) and the resulting samples were injected onto either a Q Exactive HF or Orbitrap Fusion Lumos mass spectrometer coupled to an Ultimate 3000 RSLC-Nano liquid chromatography system. Samples were injected onto a 75 μm i.d., 15-cm long EasySpray column (Thermo) for Q Exactive HF samples or a 75 μm i.d., 75-cm long EasySpray column (Thermo) for Orbitrap Fusion Lumos samples and eluted with a gradient from 0–28% buffer B over 90 min. Buffer A contained 2% (v/v) ACN and 0.1% formic acid in water, and buffer B contained 80% (v/v) ACN, 10% (v/v) trifluoroethanol, and 0.1% formic acid in water. The mass spectrometer was operated in positive ion mode with a source voltage of 2.5 kV (Q Exactive HF) or 2.2 kV (Orbitrap Fusion Lumos) and an ion transfer tube temperature of 275 °C. MS scans were acquired at 120,000 resolution in the Orbitrap. For Q Exactive HF samples, up to 20 MS/MS spectra were obtained in the ion trap for each full spectrum acquired using higher-energy collisional dissociation (HCD) for ions with charges 2–8. For Orbitrap Fusion Lumos samples, up to 10 MS/MS spectra were obtained in the ion trap for each full spectrum acquired using higher-energy collisional dissociation (HCD) for ions with charges 2–7. Dynamic exclusion was set for either 20 s (Q Exactive HF) or 25 s (Orbitrap Fusion Lumos) after an ion was selected for fragmentation. Raw MS data files were analyzed using Proteome Discoverer v. 3.0 (Thermo), with peptide identification performed using a tryptic search with Sequest HT against the human reviewed protein database from UniProt. Fragment and precursor tolerances of 10 ppm and either 0.02 Da (Q Exactive HF) or 0.6 Da (Orbitrap Fusion Lumos) were specified, and three missed cleavages were allowed. Carbamidomethylation of Cys was set as a fixed modification, with oxidation of Met set as a variable modification. The false-discovery rate (FDR) cutoff was 1% for all peptides.

The mass spectrometry results from Dataset 2 were obtained using the following method. The gel slices were cut to 1 mm by 1 mm cubes and destained 3 times by first washing with 100 μl of 100 mM ammonium bicarbonate for 15 minutes, followed by addition of the same volume of acetonitrile (ACN) for 15 minutes. The samples were dried, then reduced by mixing with 200 µl of 100 mM ammonium bicarbonate and 10 mM DTT and incubated at 56 °C for 30 minutes. The liquid was removed and 200 μl of 100 mM ammonium bicarbonate, 55 mM iodoacetamide was added to the gel pieces and incubated at room temperature in the dark for 20 minutes. Gels were washed with 100 mM ammonium bicarbonate for 15 minutes, then the same volume of ACN was added to dehydrate the gel pieces. The solution was then removed and samples were dried. For digestion, ice-cold trypsin (0.01 μg/μl) in 50 mM ammonium bicarbonate was added to cover the gel pieces and set on ice for 30 min. After complete rehydration, excess trypsin was removed, replaced with fresh 50 mM ammonium bicarbonate, and left overnight at 37 °C. The peptides were extracted twice by the addition of 50 μl of 0.2% formic acid and 5 % ACN and mixed at room temperature for 30 min. The supernatant was removed and saved. A total of 50 µl of 50% ACN-0.2% formic acid was added to the sample, which was vortexed again at room temperature for 30 min. The supernatant was removed and combined with the supernatant from the first extraction. The combined extractions were analyzed directly by liquid chromatography (LC) in combination with tandem mass spectroscopy (MS/MS) using electrospray ionization. Trypsin-digested peptides were analyzed by ultra-high-pressure liquid chromatography (UPLC) coupled with tandem mass spectroscopy (LC-MS/MS) using nano-spray ionization. The nanospray ionization experiments were performed using a TimsTOF 2 pro hybrid mass spectrometer (Bruker) interfaced with nano-scale reversed-phase UPLC (EVOSEP ONE). The evosep method of 30 SPD (samples per day) was utilized using a 10 cm × 150 μm reverse-phase column packed with 1.5 μm C18-beads (PepSep, Bruker) at 58 °C. The analytical columns were connected with a fused silica ID emitter (10 μm ID; Bruker Daltonics) inside a nanoelectrospray ion source (Captive spray source; Bruker). The mobile phases comprised 0.1% FA as solution A and 0.1% FA/99.9% ACN as solution B. The mass spectrometry settings for the TimsTOF Pro 2 used the PASEF method for standard proteomics. The values for mobility-dependent collision energy ramping were set to 95 eV at an inversed reduced mobility (1/k0) of 1.6 V s/cm2 and 23 eV at 0.73 V s/cm2. Collision energies were linearly interpolated between these two 1/k0 values and kept constant above or below. No merging of TIMS scans was performed. Target intensity per individual PASEF precursor was set to 20,000. The scan range was set between 0.6 and 1.6 V s/cm² with a ramp time of 166 ms. 14 PASEF MS/MS scans were triggered per cycle (2.57 s) with a maximum of seven precursors per mobilogram. Precursor ions in an m/z range between 100 and 1700 with charge states ≥3+ and ≤8+ were selected for fragmentation. Active exclusion was enabled for 0.4 min (mass width 0.015 Th, 1/k0 width 0.015 V s/cm²). Protein identification and label free quantification was carried out using Peaks Studio X (Bioinformatics solutions Inc).

Mass spectrometry results after analysis are shown in Fig. 1b and Extended Data Fig. 1. Analysis of Dataset 1 revealed the presence of the short form of γ2, while Dataset 2 revealed both the short and long forms of the γ2 subunit.

Recombinant human GABA_A receptor structure in complex with lamotrigine

The recombinant human α1β2γ2 GABA_A receptor was expressed in a stable cell line using the Sleeping Beauty transposon system, as previously described⁶⁹. Briefly, a tri-cistronic construct encoding the three subunits was cloned into the pSBtet vector (pSBtet-GP, Addgene plasmid #60495). This was co-transfected with SB100X transposase (pCMV(CAT)T7-SB100, Addgene #34879) into HEK293S GnTI⁻ cells (ATCC CRL-3022). These two Addgene plasmids were gifts from Eric Kowarz⁷⁰ and Zsuzsanna Izsvak⁷¹, respectively. Transfected cells were selected with 1 μg/mL puromycin before being transferred to suspension culture. A total of 6.4 L of cells, at a density of 3.5–4 × 10⁶ cells/mL, were induced with 2 μg/mL doxycycline and incubated at 30 °C with shaking (130 rpm) for 48 hours in 8% CO₂. To enhance expression, 3 mM sodium butyrate was added with doxycycline.

Cells were harvested by centrifugation and resuspended in buffer A with 0.2 mM PMSF. After mechanical lysis, the lysate was centrifuged at 10,000 × g for 20 minutes. The resulting supernatant, containing cell membranes, was then centrifuged at 186,000 × g for 2 hours. Membrane pellets were homogenized using a Dounce homogenizer and solubilized at 4 °C for 2 hours nutating in buffer A supplemented with 1% (w/v) lauryl maltose neopentyl glycol (LMNG; Anatrace). The solubilized membranes were clarified by centrifugation at 186,000 × g for 40 minutes, and the supernatant was passed through a Strep-Tactin XT Superflow affinity resin (IBA-GmbH). The resin was washed with buffer A containing 1 mM LMNG and 0.01% (w/v) porcine brain polar lipids (Avanti). The protein was eluted with wash buffer supplemented with 50 mM biotin (Sigma-Aldrich).

Nanodisc reconstitution was performed using SPMSP1D1 under conditions with 0.5 mM lamotrigine. Concentrated GABA_A receptors were mixed with porcine brain polar lipids (Avanti) and incubated at 4 °C for 30 minutes. SPMSP1D1 was then added, and the mixture was further incubated for 30 minutes. The final molar ratio of protein, lipids, and SPMSP1D1 was 1:100:10. To remove detergent, Bio-Beads SM-2 (Bio-Rad) were added at a concentration of 100 mg/mL, and the mixture was rotated overnight at 4 °C. The next day, the Bio-Beads were removed, and the sample was collected for size-exclusion chromatography.

Purified GABA_A receptor at a concentration of 1.8 mg/mL was mixed with 0.5 mM lamotrigine and 2 mM GABA, then incubated on ice for 30 minutes prior to cryo-EM grid preparation. To promote random particle orientations, 0.5 mM fluorinated Fos-Choline-8 (Anatrace) was added to the protein sample just before freezing. Cryo-EM grids were prepared using a Vitrobot Mark IV (FEI). A 3 μL sample was applied to glow-discharged 200 mesh copper R 2/1 grids overlaid with 2-nm continuous carbon (Quantifoil), using 100% humidity in the chamber at 4 °C. The grids were blotted for 3.5 seconds and plunged into liquid ethane cooled with liquid nitrogen, then transferred to liquid nitrogen for storage.

Cryo-EM data collection was performed at the UCSD Cryo-EM Facility using the Titan Krios G4 (Thermo Fisher Scientific) operating at 300 keV and equipped with a Gatan BioContinuum energy filter. Data collection parameters are detailed in Extended Data Table 3, called Dataset3. This dataset3 was processed using a CryoSPARC 4.4 workflow. Briefly, approximately 0.7 million particles were subjected to three rounds of 2D classification to remove junk particles (Extended Data Fig. 9). A consensus map was generated by Nu-refinement at 2.6 Å resolution. Three rounds of 3D classification focused on the transmembrane domain (TMD), followed by non-uniform refinement, yielded the final map for model building at 2.95 Å. Additionally, 3D classification and local refinement focused on the extracellular domain (ECD) improved the density of the lamotrigine-binding pocket to 2.7 Å, used for figure preparation. Both maps were uploaded to the EMDB during PDB deposition.

The human α1β2γ2 GABA_A receptor bound to GABA (PDB: 6X3Z) was used as the initial model for building and refinement. Restraints for lamotrigine were generated using the Grade Web Server. Manual adjustments and further refinements were performed using Coot and Phenix, as previously described. Model refinement statistics are summarized in Extended Data Table 5 (Dataset3: LTG). Structural figures were generated using UCSF ChimeraX.

Electrophysiology

Whole cell voltage-clamp recordings were made from adherent HEK293S GnTI⁻ cells transiently transfected with pEZT-based plasmids, as well as with GFP protein in pCDNA for cell selection. Each 10 mm well of cells in a 12-well dish was transfected with 0.1–0.4 µg plasmid mixture of each subunit in a ratio of 1 α1:1 β2 or 1 α1:1β2:10γ2 (to ensure the incorporation of γ2 subunit), using Lipofectamine 2000 reagent (Invitrogen). The transfected cells were incubated at 30 °C. After 48 h post-transfection, cells were re-plated on 35 mm dishes and allowed to settle for at least 3 hours. Recordings were made 48–96 hours after transfection. The bath solution contained (in mM): 140 NaCl, 2.4 KCl, 4 MgCl₂, 4 CaCl₂, 5 HEPES and 10 glucose pH 7.3. Borosilicate pipettes were pulled and polished to an initial resistance of 2–4 MΩ, filled with the pipette solution containing (in mM): 100 CsCl, 30 CsF, 10 NaCl, 10 EGTA, and 20 HEPES pH 7.3. Whole cell currents were recorded using pClamp 11 with an Axopatch 200B amplifier, sampled at 20 kHz, and low-pass filtered at 2 kHz using a Digidata 1550B (Molecular Devices). Cells were held at −75 mV. Solution exchange was achieved using a gravity driven RSC-200 rapid solution changer (Bio-Logic). Whole cell currents data were analyzed with Clampfit 11 software (Molecular Devices). GABA and ligand (lamotrigine, levetiracetam, lacosamide, and zonisamide, Sigma-Aldrich) solutions were prepared in bath solution from concentrated stocks: 1 M GABA in water, and 300 mM lamotrigine, 500 mM Levetiracetam, 300 mM lacosamide and 300 mM zonisamide in DMSO. For the statistics in Fig. 3 (bar graph), results are presented as normalized peak currents. Rundown was consistently observed in α1β2 recordings; repeated measurements of the response to GABA application were made until stable peak currents were observed, then the drug applications were tested. We observed that only half of the cells expressing the α1β2 receptors responded to LEV and LTG; we report in Fig. 3e the currents from responsive cells. Replicate numbers from independent cells are labeled in each bar. Statistical analysis was performed using GraphPad Prism 10.2.0 software (GraphPad software, Inc, La Jolla, CA). Data are expressed as means ± S.D. Two-tailed Welch’s t-test was used. A p-value of ≤ 0.05 was considered statistically significant (***, p ≤ 0.001; **, p ≤ 0.01; *, 0.01 ≤ p ≤ 0.05).

Extended Data

Extended Data Fig. 1|. Characterization of native human α1-containing GABA_A receptors and associated proteins from brain tissue.

a, Sample analysis focus on the GABA_AR and non-GABA_AR bands. b, Same analysis conducted as in Fig.1b. c, The top three unknown bands in the SDS-PAGE gel were individually excised for mass spectrometry analysis. Detailed results regarding associated proteins and their related homologues are provided in the table. The table groups are color-coded to match the boxes in a. Results shown in this representative gel were similarly obtained from n=4 gels from independent purifications. Molecular weights of standards were given on the left of the gel in kD.

Extended Data Fig. 2|. Cryo-EM data processing of Dataset 1 from tissue from 45 patients.

~28k dose-fractionated micrographs were collected from multiple grids stemming from one purification using anti-α1 Fab (1F4) as the affinity reagent. All data processing was performed in CryoSPARC v4.4. Particles were picked and subjected to multiple rounds of 2D classification; final results are shown, with ~1.3M particles selected from this step. Ab initio reconstruction was performed with 3 classes, which were used as inputs for heterogeneous refinement of the whole particle set. A single particle set emerged from heterogeneous refinement including ~1.0M particles, which was subjected to non-uniform (NU) refinement. This map and particle set served as the starting point for all 3D classification. The primary 3D classification used a focused mask around the whole ECD and was set up with 10 classes (Class0 - Class9). The ECD mask was chosen due to the highest resolution features being mainly found in the ECD, and thus this region should allow more robust separation of compositional heterogeneity. Class0 and Class5 contained two strong Fab densities, so were combined and refined to a single subunit arrangement easily definable during model building as β2-α1-β2-α1-γ2. This map and subunit arrangement are indicated as a final, high confidence endpoint in this data processing by listing the subunit order in bold font. The FSC-based map resolution estimate and particle number are given below all maps. Classes 1, 2, and 4 were pooled due to the presence of one strong Fab, refined, and yielded a ~2.6 Å resolution map. Heterogeneity in TMD signal in the input classes (1, 2, 4) motivated us to conduct a secondary 3D classification on this ~420k particle set (8 classes). Briefly, however, from the primary 3D classification, Class3 and Class6-Class9 were excluded from further analysis due to anisotropic resolution features suggestive of preferred orientation, or low resolution overall (Class3 and 9). Secondary classification of 1-Fab containing particles using focus map on Chains A-B ECD. This region was selected for masking after testing a large panel of masks (single subunits, pairs, adjacent triplets). All 3D classes were subjected to Nu refinement, then atomic model building was done to begin assigning subunits to map positions (Supplementary Figs. 1-11). Class0 and Class5 exhibited features of a common assembly, so were pooled and refined, to a final subunit arrangement of β2-α1-γ2-β2-α2. All processing is summarized in this flow chart for the dataset with the above explanations serving as representative decision-making approaches. Each composition has been also labeled as (a-g), corresponding to the labels in Fig. 2 and Extended Data Fig. 6. In some cases, a single position in a map was clearly a mixture of two subunits, for example α2/α3. In other cases, we designate the heterogeneity as a “mixture,” which instead means that one or more positions are not yet at that step clearly an α vs. β vs. γ. While we cannot rule out the possibility of contamination in a “final” particle set, arising from a minor fraction of incorrect particles, we can conclude with confidence that for each of the “final” maps, the model built represents the clearly dominant species in that particle set.

Extended Data Fig. 3|. Cryo-EM data processing of Dataset 2 from tissue from 36 patients.

Approximately 45k dose-fractionated micrographs were collected, resulting in 1.8 million particles after several rounds of 2D classification. The workflow was conducted according to the procedures outlined in Extended Data Fig. 2. All models were constructed for the reconstructions highlighted in the back boxes. Each composition is also labeled as (a-c) and (h-k), corresponding to the labels in Fig. 2 and Extended Data Fig. 6.

Extended Data Fig. 4|. Cryo-EM analysis of the native α1-containing GABA_A receptors in complex with interacting partner density.

a, Single-particle cryo-EM image processing workflow, focusing on the transmembrane domain (TMD). Refer to the Methods section for further elaboration on the workflow. Fourier shell correlation (FSC) curves with gray line indicating FSC = 0.143 (b), angular distribution plot (c) and local-resolution maps (d) for the final 3D reconstruction.

Extended Data Fig. 5|. Focused classification for γ2-TMD map reconstruction in the three subunit compositions abundant in both datasets.

All cryo-EM data processing steps were performed using CryoSPARC v4.4. Further NU refinement combined particles from both datasets. Focused 3D classifications on the entire TMD region were conducted, and the class with a strong γ2 subunit signal was selected for the final reconstruction. All models were constructed for the reconstructions highlighted in the outlined black boxes. The determination of the γ2 subunit in the compositions β2-α1-β2-α1-γ2, β2-α1-γ2-β2-α2, and β2/3-α1-β2-α1-γ2 is displayed in panels a, b, and c, respectively. Each final composition is also labeled as (a-c), corresponding to the labels in Fig. 2 and Extended Data Fig. 6.

Extended Data Fig. 6|. Characteristics of cryo-EM reconstructions.

The angular distribution plots of particles used for the final cryo-EM reconstructions are shown in the top left. The cryo-EM maps, colored according to local resolution estimation in CryoSPARC v4.4, are displayed in the top right (color scale in Å). FSC curves were calculated between half-maps, with the overall resolution estimated using the FSC = 0.143 criterion (below). The detailed compositions of the eleven cryo-EM maps are labeled in the tables of FSC curves.

Extended Data Fig. 7|. Comparative analysis of the most abundant native GABA_A receptor with previously reported structures.

a, Structural comparison of the β2-α1-β2-α1-γ2 (light coral) with 6X3Z³⁷ (burly wood), 6I53⁷² (light green), and 8G4N⁸ (sky blue) demonstrates high overall similarity. b, Detailed comparison of the transmembrane domain regions, with colors consistent with panel a. Each subunit is labeled accordingly. c, Pore conformation of the β2-α1-β2-α1-γ2 receptor, highlighting the opposing β2 and γ2 M2 α-helices as ribbons, with the pore-lining side chains shown as sticks. Purple and green spheres illustrate the pore shape. Distances shown within the pore represent diameters at the desensitization gate (−2ʹ) and resting gate (9ʹ) positions. d, Comparative pore diameter versus distance plot for the structures depicted in panel a, aligned at y = 0 at the level of the −2ʹ desensitization gate. e, Superposition of structures based on the extracellular domains of adjacent β and α subunits, as shown in panel a.

Extended Data Fig. 8|. Densities in native α1-containing GABA_A receptors, ligand docking and electrophysiological recordings.

Cryo-EM densities were observed at the interfaces formed by the ECDs of the α and β subunits in the composition of β2-α1-γ2-β2-α2 (a), β2-α1-β2/3-α1-γ2 (b) and β3-α1-γ2-β2/3-α2 (c). Additionally, density was observed at the interfaces formed by the ECDs of the β1 and β1 subunits in the composition of β2-α1-β1-β1-γ2 (d). Ligands of diazepam (DZP) and DMCM aligned with the composition of β2-α1-β2-α1-γ2 (e, f) and β2-α1-β2/3-α1-γ2 (g, h) at the interface of α1 and γ2 subunits. DZP pose is from PDB: 6X3X³⁷ and DMCM from PDB: 8DD3³⁹. i, Superposition of surrounding residues from the compositions of β2-α1-γ2-β2-α2, β2-α1-β2/3-α1-γ2 and β3-α1-γ2-β2/3-α2 at interfaces of α and β subunits. j, The alignment of histamine (HSM) with the composition of β2-α1-β1-β1-γ2 at the interface of β1 and β1 subunits; histamine binding pose from Sente et al.⁵ k-l, Representative responses are shown for application of GABA, followed by GABA plus LTG with α1β2γ2 and α1β2, and same for LEV (m-n).

Extended Data Fig. 9 |. Cryo-EM analysis of recombinant GABA_A receptor in complex with lamotrigine.

a, Single-particle cryo-EM image processing workflow, with several rounds of focused 3D classification on the transmembrane domain (TMD), as shown in b. c, Fourier shell correlation (FSC) curves, with the gray line indicating FSC = 0.143. d, Local-resolution map and e, angular distribution plot for the final 3D reconstruction. To enhance lamotrigine density, focused 3D classification and local refinement were applied to the extracellular domain (ECD), as shown in f. g, FSC curves, with the gray line indicating FSC = 0.143, and h, local-resolution maps for the final 3D reconstruction. In f and h, black box boundaries approximate mask used in focused refinement and local resolution estimate; local resolution for regions outside of the mask (e.g., TMD helices) are not meaningful.

Extended Data Table 1 |.

Characteristics of 45 Epilepsy Patients

Sample	Age	Sex	Region	AEDs at time of surgery	Epilepsy duration (y)
25	35	M	Temporal	Phenytoin, Zonisamide	25
26	56	M	Temporal	Lacosamide, Lamotrigine	19
27	33	F	Temporal	Carbamazepine, Lamotrigine, Levetiracetam	26
28	43	F	Temporal	Lamotrigine, Topiramate	17
29	46	M	Temporal	Zonisamide, Lacosamide, Gabapentin, Valproate	9
30	74	M	Temporal	Brivaracetam, Zonisamide	3
35	22	M	Temporal	Brivaracetam, Lacosamide	9
37	64	M	Temporal	Lacosamide, Levetiracetam	8
39	22	M	Temporal	Lacosamide, Topiramate	3
40	45	M	Temporal	Levetiracetam, Oxcarbazepine	41
42	51	F	Temporal	Zonisamide, Lacosamide, Gabapentin, Valproate	50
43	29	M	Temporal	Clobazam, Lacosamide, Levetiracetam	28
48	50	F	Temporal	Phenytoin, Lacosamide, Gabapentin	4
52	24	F	Temporal	Levetiracetam, Oxcarbazepine	1
53	51	M	Temporal	Phenytoin, Gabapentin	43
57	59	F	Temporal	Carbamazepine, Clonazepam	36
58	68	F	Temporal	Lacosamide	23
59	38	F	Temporal	Eslicarbazepine	31
63	26	M	Temporal	Brivaracetam, Eslicarbazepine	1.5
65	35	F	Temporal	Unknown	2
66	54	F	Temporal	Lacosamide, Clobazam, Brivaracetam	22
70	29	M	Temporal	Lacosamide, Zonisamide	10
72	33	M	Temporal	Lamotrigine	4
74	36	M	Temporal	Carbamazepine, Levetiracetam	18
75	36	F	Temporal	Lamotrigine, Levetiracetam, Oxcarbazepine	31
76	24	M	Temporal	Lamotrigine, Zonisamide	16
77	23	M	Temporal	Brivaracetam, Lacosamide	10
78	54	F	Temporal	Eslicarbazepine, Lamotrigine	5
82	37	F	Temporal	Lamotrigine	5
83	23	F	Temporal	Lacosamide, Levetiracetam	5
84	44	F	Temporal	Levetiracetam, Lamotrigine, Lacosamide	41
85	19	M	Temporal	Lacosamide, Lamotrigine	11
87	35	F	Temporal	Brivaracetam, Lamotrigine	25
88	34	F	Temporal	Eslicarbazepine, Cenobamate	4
89	57	F	Temporal	Lacosamide	11
90	46	M	Temporal	Brivaracetam, Lamotrigine	12
91	51	M	Frontal	Lacosamide, Clonazepam	17
92	55	F	Temporal	Lacosamide, Topiramate	15
94	43	M	Frontal	Zonisamide	4
95	21	F	Temporal	Oxcarbazepine, Phenobarbital, Clonazepam, Lacosamide, Levetiracetam, Pregabalin	7
96	67	F	Frontal	Zonisamide	25
97	31	M	Frontal	Lamotrigine, Levetiracetam	15
98	44	M	Temporal	Carbamazepine, Lamotrigine, Levetiracetam	5
99	58	F	Temporal	Zonisamide, Gabapentin	5
100	50	M	Temporal	Lacosamide, Oxcarbazepine	20

Extended Data Table 2 |.

Characteristics of 36 Epilepsy Patients

Sample	Age	Sex	Region	AEDs at time of surgery	Epilepsy duration (y)
101	31	M	Temporal	Valproate, Lamotrigine	7
102	26	F	Temporal	Lacosamide	10
103	44	F	Frontal	Phenytoin	22
106	31	M	Temporal	Lacosamide, Topiramate	2
107	57	M	Temporal	Clobazam, Carbamazepine, Levetiracetam	27
108	40	M	Temporal	Oxcarbazepine, Zonisamide	2
109	40	M	Temporal	Carbamazepine, Lacosamide	40
110	47	F	Temporal	Carbamazepine, Lamotrigine	28
130	48	M	Temporal	Topiramate, Carbamazepine, Lamotrigine	45
131	61	M	Temporal	Levetiracetam	53
132	59	M	Temporal	Lacosamide, Oxcarbazepine	5
133	59	M	Temporal	Eslicarbazepine	20
134	25	M	Temporal	Valproate, Levetiracetam	1
135	55	M	Temporal	Lamotrigine	13
138	55	M	Temporal	Lacosamide, Lamotrigine	51
141	36	M	Temporal	Oxcarbazepine	2
142	64	M	Temporal	Lacosamide, Levetiracetam	62
143	45	M	Temporal	Zonisamide, Oxcarbazepine	32
144	28	M	Temporal	Brivaracetam, Lamotrigine	15
145	37	F	Frontal	Zonisamide, Lacosamide, Lamotrigine, Levetiracetam	3
146	43	M	Frontal	Zonisamide, Lacosamide, Levetiracetam	5
147	41	F	Temporal	Fosphenytoin, Levetiracetam	3
148	26	M	Temporal	Lacosamide, Oxcarbazepine	7
149	44	M	Temporal	Lamotrigine, Levetiracetam	18
150	33	M	Temporal	Zonisamide, Levetiracetam	18
153	38	F	Temporal	Lamotrigine, Levetiracetam	20
154	58	M	Temporal	Zonisamide, Lacosamide	7
155	41	F	Temporal	Zonisamide, Levetiracetam, Lamotrigine	38
156	42	F	Temporal	Lacosamide, Lamotrigine	37
158	49	M	Frontal	Brivaracetam, Lacosamide, Cenobamate	35
159	41	F	Temporal	Levetiracetam	21
162	23	M	Temporal	Cenobamate, Lamotrigine, Levetiracetam	4
163	31	F	Temporal	Fosphenytoin, Lamotrigine	12
165	23	M	Temporal	Valproate, Levetiracetam, Lacosamide, Fosphenytoin	20
166	39	M	Temporal	Lacosamide, Levetiracetam	36
167	38	M	Frontal	Felbamate, Fosphenytoin, Lacosamide, Lamotrigine, Levetiracetam	33

Extended Data Table 3 |.

Cryo-EM data collection parameters

	Dataset1 Native GABAAR (+GABA)	Dataset2 Native GABAAR (+GABA)	Dataset3 GABAAR (+GABA and LTG)
Data collection and processing
Microscope	UCSD Krios G4	UCSD Krios G4	UCSD Krios G4
Electron Gun	FEG	FEG	FEG
Detector	Falcon4	Falcon4	Falcon4
Magnification	130K	130K	130K
Voltage (kV)	300	300	300
Electron exposure (e−/Å2)	50	50	50
Defocus range (μm)	1.0 to 2.2	1.0 to 2.2	1.0–2.2
Pixel size (Å)	0.935	0.935	0.935
Symmetry imposed	C1	C1	C1
Number of collected movies	27,849	45,062	18,308
Initial particle images (no.)	1,290,608	1,808,823	789,923

Extended Data Table 4 |.

Refinement and validation statistics

Dataset	Dataset1+2	Dataset1+2	Dataset1+2	Dataset1	Dataset1	Dataset1	Dataset1
Map	β2-α1-β2-α1-γ2	β2-α1-γ2-β2-α2	β2/3-α1-β2-α2-γ2	β2-α1-β2-α2-γ2	β2-α1-β2/3-α2-γ2	β2-α1-β2-α2/3-γ2	β2-α1-γ2-β1-α2
Model	β2-α1-β2-α1-γ2	β2-α1-γ2-β2-α2	β3-α1-β2-α2-γ2	β2-α1-β2-α2-γ2	β2-α1-β3-α2-γ2	β2-α1-β2-α3-γ2	β2-α1-γ2-β1-α2
EMDB ID	EMD-45878	EMD-45884	EMD-45890	EMD-45894	EMD-45908	EMD-45914	EMD-45920
PDB ID	9CRS	9CRV	9CSB	9CT0	9CTJ	9CTP	9CTV
Initial model used (PDB code)	6X3T/6I53	6X3T/6I53	6X3T/6I53	6X3T/6I53	6X3T/6I53	6X3T/6I53	6X3T/6I53
Final particle images (no.)	101,120	30,526	14,361	39,793	10,975	13,527	14,782
Map resolution (Å) FSC threshold=0.143	2.90	3.18	3.34	3.19	3.74	3.62	3.36
Map sharpening B factor (Å2)	87.1	70.5	57.0	74.3	54.1	53.8	53.8
Model composition
Non-hydrogen atoms	17623	15727	15525	14736	14479	14609	14578
Protein residues	2122	1886	1873	1761	1755	1751	1751
Glycans	22	21	16	19	19	17	19
Ligands	4	4	1	4	0	2	2
Lipids	2	1	2	2	0	2	1
B factors (Å2)
Protein	66.45	79.80	103.26	95.86	129.00	113.03	104.00
Ligands	79.93	102.99	166.91	145.70	103.32	131.68	131.91
R.m.s. deviations
Bond lengths (Å)	0.006	0.006	0.007	0.007	0.007	0.007	0.006
Bond angles (°)	0.972	0.877	1.000	1.080	1.087	1.075	0.873
Validation
MolProbity score	1.56	1.62	1.89	1.97	2.09	1.88	1.54
Clashscore	5.10	5.94	7.42	7.65	8.71	8.51	6.07
Poor rotamers (%)	0.90	0.65	1.37	1.58	1.97	1.14	0.57
Ramachandran plot
Favored (%)	95.85	95.80	94.54	94.02	94.00	94.45	96.76
Allowed (%)	4.15	4.20	5.46	5.98	6.00	5.55	3.24
Disallowed (%)	0	0	0	0	0	0	0

Extended Data Table 5 |.

Refinement and validation statistics

Dataset Map Model	Dataset2 β2-α1-β2/3-α1-γ2 β2-α1-β3-α1-γ2	Dataset2 β3-α1-γ2-β2/3-α2 β3-α1-γ2-β2-α2	Dataset2 β3-α1-γ2-β2/3-α2 β3-α1-γ2-β3-α2	Dataset2 β2-α1-β1-α2-γ2 β2-α1-β1-α2-γ2	Dataset2 β2-α1-β1-β1-γ2 β2-α1-β1-β1-γ2	Dataset3: LTG β2-α1-β2-α1-γ2 β2-α1-β2-α1-γ2
EMDB ID PDB ID	EMD-45983 9CXA	EMD-45985 9CXC	EMD-45980 9CX7	EMD-45984 9CXB	EMD-45986 9CXD	EMD-47132 9DRX
Initial model used (PDB code)	6X3T/6I53	6X3T/6I53	6X3T/6I53	6X3T/6I53	6X3T/6I53	6X3Z
Final particle images (no.)	74,535	29,584	29,584	30,789	28,909	124,678
Map resolution (Å) FSC threshold=0.143	3.04	3.30	3.30	3.33	3.36	2.95
Map sharpening B factor (Å2)	89.4	80.6	80.6	79.7	76.9	83.5
Model composition Non-hydrogen atoms  Protein residues  Glycans  Ligands  Lipids	16521 1989 20 4 2	14577 1754 19 4 0	14584 1755 19 4 0	14479 1749 19 2 1	15403 1874 9 2 0	17398 2121 21 3 0
B factors (Å2)  Protein  Ligands	73.25 94.65	89.62 111.21	92.88 116.73	97.43 96.22	94.68 111.31	143.57 133.19
R.m.s. deviations  Bond lengths (Å)  Bond angles (°)	0.006 0.936	0.005 0.755	0.007 1.112	0.006 0.950	0.004 0.896	0.008 1.232
Validation  MolProbity score  Clashscore  Poor rotamers (%)	1.59 5.50 1.0	1.49 4.10 0.06	2.10 8.10 2.23	1.74 6.38 1.0	1.61 5.18 0.42	1.71 5.03 1.43
Ramachandran plot  Favored (%)  Allowed (%)  Disallowed (%)	95.82 4.18 0	95.79 4.21 0	94.17 5.83 0	94.33 5.67 0	95.14 4.86 0	95.39 4.61 0

Data availability

All atomic models and cryo-EM maps have been deposited in the Protein Data Bank and Electron Microscopy Data Bank: β2-α1-β2-α1-γ2 (PDB 9CRS, EMD 45878), β2-α1-γ2-β2-α2 (PDB 9CRV, EMD 45884), β3-α1-β2-α2-γ2 (PDB 9CSB, EMD 45890), β2-α1-β2-α2-γ2 (PDB 9CT0, EMD 45894), β2-α1-β3-α2-γ2 (PDB 9CTJ, EMD 45908), β2-α1-β2-α3-γ2 (PDB 9CTP, EMD 45914), β2-α1-γ2-β1-α2 (PDB 9CTV, EMD 45920), β2-α1-β3-α1-γ2 (PDB 9CXA, EMD 45983), β3-α1-γ2-β2-α2 (PDB 9CXC, EMD 45985), β3-α1-γ2-β3-α2 (PDB 9CX7, EMD 45980), β2-α1-β1-α2-γ2 (PDB 9CXB, EMD 45984), β2-α1-β1-β1-γ2 (PDB 9CXD, EMD 45986), LTG bound β2-α1-β2-α1-γ2 (PDB 9DRX, EMD 47132). Raw electrophysiology and mass spectrometry data are included as Source data with the paper.