Structural insights into kainate receptor desensitization

Abstract

Kainate receptors (KARs) belong to the ionotropic glutamate receptor (iGluR) family and play critical roles in mediating excitatory neurotransmission and regulating neurotransmitter release. Receptor desensitization is a critical factor for regulating the strength of synaptic transmission. Notwithstanding their overall structural similarity to AMPA receptors, KARs exhibit a desensitized conformation that is distinct from that of most other iGluRs. Despite extensive studies on KARs, a fundamental question remains unresolved: why do KARs require large conformational changes upon desensitization? Here we show cryo-electron microscopy structures of GluK2 containing double cysteine mutations, captured in non-active and various desensitized conformations. In the shallow-desensitized conformation, two cysteine crosslinks stabilize the receptors in a conformation resembling the typical desensitized state of non-KAR iGluRs. Our patch-clamp recordings and fluctuation analysis suggest that KARs in the shallow-desensitized state remain ion-permeable. This finding indicates that the lateral rotational movement of the KAR ligand-binding domains is critical for complete channel closure and stabilization of the fully desensitized receptor. Overall, this study elucidates the mechanism and conformational dynamics of KARs during desensitization.

Kainate receptor (KAR) desensitization is structurally and functionally distinct among iGluRs. Here, authors used cryo-EM and electrophysiology to reveal that ligand-binding domain rotations are essential for full channel closure and stabilization of the desensitized state of KARs.

Introduction

Kainate receptors (KARs) are member of the ionotropic glutamate receptor (iGluR) superfamily, together with α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid (AMPA), N-methyl-D-aspartate (NMDA) and delta receptors. These cation-selective ligand-gated ion channels are activated by the excitatory neurotransmitter glutamate and mediate fast excitatory synaptic transmission in mammalian brains^1,2. In addition to their post-synaptic role, KARs also function pre- and extra-synaptically, regulating excitatory and inhibitory neurotransmitter release and modulating synaptic excitability^3–7. Abnormal expression and dysfunction of KARs are associated with various neurological diseases and disorders, including schizophrenia, mood disorders, epilepsy, and Huntington’s disease^8–17. Accordingly, KARs are considered potential therapeutic drug targets^18–21.

KARs assemble as homo- or heterotetramers comprising five subunits: GluK1–5^22–26. The complexity of their subunit composition²⁵ leads to highly diverse functional and pharmacological properties, enabling the sensitive regulation of synaptic transmission^2,27–36. Extensive studies have revealed that KARs adopt a Y-shaped architecture with four distinct layers: the amino-terminal domain (ATD), which is critical for receptor assembly; the ligand-binding domain (LBD), which contains ligand-binding sites and regulates channel gating; the transmembrane domain (TMD), which forms the ion channel within the plasma membrane; and a short disordered C-terminal domain (CTD), which interacts with intracellular proteins^1,37.

In addition to the diversity of KAR subtypes, the physiological and pharmacological properties of native KARs are regulated by multiple factors in the brain. Most native KARs form heteromers (e.g., GluK2/K5) and co-assemble with the neuropilin and tolloid-like (Neto) auxiliary proteins NETO1 and NETO2. These auxiliary subunits are critical modulators of KAR activity, conferring distinct channel properties such as slower desensitization kinetics^38–45. Furthermore, external monovalent ions are essential for KAR function. By binding at the LBD dimer interface, these ions stabilize dimer formation^46–50 and regulate receptor gating, including desensitization and recovery from desensitization^51–57. Post-translational modifications such as glycosylation and phosphorylation also alter KAR expression, trafficking and function^8,58–65. This functional diversity suggests that native KARs are tuned for sustained signaling under physiological conditions. In contrast, recombinant GluK2 KARs demonstrate rapid desensitization and over 50-fold slower recovery from desensitization⁶⁶ compared to AMPARs^66,67.

While the detailed assembly of heteromeric KARs with auxiliary proteins in the native lipid bilayer remains elusive, recently determined structures of recombinant KARs in various functional states have revealed the structure–function correlations underlying the fundamental gating mechanisms of KARs, including activation/desensitization^{25,59,68–71}, allosteric modulation⁷², inhibition⁷³ and channel block⁷⁴. A key structural feature of KARs is their distinct desensitized conformation. Structures of glutamate-bound GluK2, first reported in 2014 by the Mayer/Subramaniam group, showed that KAR LBD dimers are stabilized in a fully dissociated configuration upon desensitization. This conformation is accompanied by a large ( > 110°) horizontal-plane rigid-body rotation of the LBDs in the BD subunits, leading to a four-fold symmetric conformation^37,69. Due to the rapid, millisecond-scale onset of glutamate-induced desensitization, these glutamate-bound KAR structures are considered to represent fully desensitized states^{25,59,68–71,73,75–77}.

LBD separation also occurs in other major iGluRs (e.g., GluA2 or GluA2 complexed with auxiliary subunits such as TARP γ2 or GSG1L). However, the extent of dimer dissociation is typically significantly smaller^{37,75,78–80}, with some exceptions described below. Desensitization in typical iGluRs is primarily driven by a tilting motion of the LBD dimers. This motion separates the upper (D1) lobes while simultaneously drawing the lower (D2) lobes into closer proximity. AMPARs comprise four subunits (GluA1–4), with GluA2-containing tetramers representing the predominant population in native tissues^81–83. In their native environment, AMPARs assemble with auxiliary subunits such as transmembrane AMPA receptor regulatory proteins (TARPs) and Cornichons^1,84–87. In Caenorhabditis elegans, NETO homologues Sol-1/2 interact with GLR-1 iGluRs⁸⁸. These subunits, especially TARPs stabilize the overall Y-shaped architecture and maintain two-fold symmetry^78,79,89,90. AMPARs lacking the GluA2 subunit or auxiliary proteins appear to be more dynamic, enabling them to adopt splayed LBD conformations upon desensitization, a state similar to that seen in KARs^{75,83,91–93}. Additionally, unconventional NMDA receptors, composed of GluN1 and GluN3 subunits, and activated solely by glycine (rather than glutamate), similarly adopt splayed LBD conformations upon desensitization⁹⁴. In contrast, both homomeric and heteromeric KARs adopt either a four-fold symmetric or asymmetric conformation upon desensitization^69,70. This structure deviates from the classical two-fold desensitized conformation observed in other major iGluRs, regardless of co-assembly with auxiliary proteins⁷¹. Recent structures of KARs bound to full or partial agonists have revealed asymmetrical conformations^{59,60,68,71,73}: one LBD dimer is completely dissociated, resembling the deep-desensitized state of KARs, while the other exhibits a disrupted D1–D1 interface but retains a closely associated D2–D2 interface, similar to the desensitized conformation of AMPAR LBDs. These findings suggest that upon desensitization, KAR LBD dimers can adopt either partially or fully dissociated conformations, and that the dissociation of the two LBD dimers can occur independently. A desensitized KAR structure with a two-fold symmetrical conformation similar to that of AMPARs has not yet been reported, likely due to its energetic instability or transient nature.

Despite their distinct structural and functional features, the mechanism by which KARs undergo such large conformational changes upon desensitization and how these drastic rearrangements regulate receptor gating remains elusive. Studying desensitization in recombinant homomeric GluK2 receptors is valuable for uncovering the intrinsic structural mechanisms that govern KAR gating, which is a crucial baseline for understanding how native regulatory mechanisms then fine-tune or counteract desensitization in more complex physiological environments. Furthermore, these intrinsic gating properties are increasingly relevant in pathological conditions, where excessive or prolonged glutamate release^95–99 or significant changes in pH^100,101 and extracellular ion concentrations^99,102 alter KAR kinetics, including desensitization, which leads to brain diseases such as epilepsy and ischaemia^{17,20,28,36,103–106}.

To address the fundamental question of function-dependent KAR conformations, here, we present cryo-electron microscopy (cryo-EM) structures of GluK2 KAR stabilized in multiple functional states: non-active, shallow-desensitized, intermediate, and deep-desensitized conformations. These states are captured using cysteine cross-linking in the presence of glutamate, either with or without the positive allosteric modulator, BPAM344. Structural comparison reveals that the large in-plane rotation of LBDs in both BD subunits and the resulting twisted motion of the LBD–TM3 linkers are essential for tight ion channel closure and stabilization of receptors in the deep-desensitized state, a feature not observed in major AMPARs and NMDARs. Functionally, the crosslinked GluK2 mutant exhibits large fractional steady-state currents, as revealed by patch-clamp recording. The stationary noise analysis results indicates that the GluK2 mutant remains conductive when desensitized, resembling the behavior of the conductive GluA2(R) AMPAR under similar conditions¹⁰⁷. Overall, our structural and functional data demonstrate that robust in-plane LBD rotations are essential for proper ion channel closure of KARs. This study provides insights into the desensitization mechanism of the KAR subfamily.

Results

Trapping GluK2 KARs in a non-desensitizing state

Previous studies have shown that LBD dimer reorientations switch the gating of KARs^68–70. To gain insight into the detailed conformational transitions from active to desensitized states, we first used SWISS-MODEL^108–112 to generate a homology model of the intact open GluK2 receptor, comprising the ATD, LBD, and TMD layers, based on the structure of the activated GluA2 AMPAR in the presence of glutamate, a potentiator and an auxiliary subunit in its higher conductive open state (Protein Data Bank [PDB] code, 5WEO)^78,113,114 (Supplementary Fig. 1A, B). In the model, the G/E helices of subunit A and the K helix of subunit B come into close proximity (Supplementary Fig. 1C, D). This also resembled the recently determined structure of GluK2 bound to glutamate, an allosteric positive modulator BPAM344, and concanavalin A (ConA) in an open/active state, using time-resolved cryo-EM (PDB code, 9B36)⁶⁸. Specifically, the distances between the Cα atoms of K676 (subunit A) and N802 (subunit B) differed across functional states. These distances measured 6.2 Å in the homology model and 6.7 Å in the active/open GluK2 structure. In contrast, substantially greater distances—9.0 Å, 9.9 Å and 38.5 Å—were observed in the BPAM344-bound no agonist-supplemented state (PDB code, 8FWS)⁷² the antagonized state (PDB code, 7F5B)⁷¹, and the desensitized state (PDB codes, 9B38 and 5KUF)^68,69, respectively (Supplementary Fig. 1D).

Cysteine crosslinking approaches have been used to analyze the dynamics and channel functions of iGluRs^90,115–119. Additionally, a recent report showed that the conformational changes of KARs upon activation resemble those of AMPARs⁶⁸. Based on these reports, we hypothesized that an engineered disulfide bond between residues K676 and N802 would restrict the mobility of the LBD and trap a dimer-of-dimers LBD conformation, altering gating kinetics. To test this hypothesis, we introduced a pair of cysteine residues designed to form spontaneous disulfide bonds between the LBD dimer pairs and assessed the channel activity of the full-length rat GluK2 K676C/N802C mutant using whole-cell patch-clamp electrophysiology. Note that the distance between the Cα atoms of two cysteines required to form a disulfide bond is <7 Ź²⁰.

GluK2 wild-type (WT) demonstrated rapid desensitization, as observed previously⁶⁶. Both the GluK2 K676C and GluK2 N802C single-cysteine mutants showed nearly complete desensitization, similar to GluK2 WT (Supplementary Fig. 2A). In contrast, the GluK2 K676C/N802C double-cysteine mutant exhibited slower desensitization and an increased steady-state to peak current amplitude, with 81% desensitization (Fig. 1A, B). The desensitization time constants (τdes) of GluK2 WT and GluK2 K676C/N802C activated by 1 mM glutamate were 7.5 ± 1.1 ms (n = 12) and 14.4 ± 0.5 ms (n = 15), respectively (Supplementary Fig. 2B). Furthermore, we performed western blot analysis of GluK2 WT and the GluK2 K676C/N802 C mutant, alongside two additional mutations (C595A/C871A), mutating cysteine residues which form inter-protomer crosslinks in the GluK2 WT^74,121 (Supplementary Fig. 2C). The GluK2 K676C/N802C double cysteine mutant formed inter-protomer disulfide bonds, as evidenced by non-reducing Western blots. Conversely, the disulfide bridge requiring both residues was absent in controls using single cysteine mutants (K676C or N802C) or reducing conditions (Supplementary Fig. 2D). These findings unequivocally confirm that disulfide bond formation between K676C and N802C drives the generation of the steady-state current.

Fig. 1. Functional and structural characterization of GluK2 K676C/N802C.

A Representative whole-cell patch-clamp trace showing the response of GluK2 WT or GluK2 K676C/N802C to a 20-s application of 1 mM glutamate in the presence and absence of BPAM344 (BPAM). Arrows indicate the maximum peak current (I_peak) and the steady-state current (I_ss). B Quantification of percentage desensitization for currents evoked by 1 mM glutamate in GluK2 (dark gray; n = 10), and the K676C/N802C mutant (light gray; n = 7), and for currents evoked by 1 mM glutamate plus 500 μM BPAM344 in GluK2 WT (green; n = 6) and the K676C/N802C mutant (cyan; n = 10) (C) Ratios of peak current amplitudes (I_peak) for GluK2 WT (left; n = 6), and K676C/N802C (middle; n = 10), and of steady-state current amplitude (I_ss) for GluK2 K676C/N802C (right; n = 10), measured in the presence and absence of BPAM344. For B and C, black circles denote independent biological replicates. Data are mean ± SD; whiskers indicate the standard deviation. Statistical significance was calculated using a two-sided two-sample t-test, with significance assumed if P < 0.05; the exact P-values are shown in the figure. D Cryo-EM map and model of GluK2 K676C/N802C complexed to glutamate and BPAM344 in a non-active state. N-glycans are highlighted in yellow. E The LBDs shown in views parallel (top) and perpendicular (bottom) to the membrane. BPAM344 and glutamate are depicted as space-filling models, coloured pink and cyan, respectively. Disulfide bonds formed between K676C and N802C, and the BPAM344 binding sites, are indicated by black and cyan dashed boxes, respectively. F Schematic diagram illustrating LBD with cysteine crosslinks, which are indicated by black lines (i) and (ii). In the diagram, glutamate is represented by a pink circle, and BPAM344 is represented by a cyan circle. G Comparison of the apo GluK2 LBD (PDB 9CAZ) and the glutamate-bound GluK2 LBD in the non-active LBD conformation. The D1 lobe is superimposed, and the rotation angles to align the D2 residues were calculated. The glutamate-bound LBD exhibits a 20° closure relative to the apo conformation. H Chemical structure of BPAM344. I Close-up view of the BPAM344 binding site at the AD subunit interface. J Cryo-EM densities for disulfide bonds between K676C and N802C in AB and CD subunits. Source data are provided as a Source Data file.

A previous study showed that BPAM344 significantly slows the desensitization kinetics of GluK2 WT and enhances the glutamate-evoked current by 140-fold and 21-fold, respectively¹²². Therefore, we assessed how BPAM344 modulates the activation and desensitization of the GluK2 K676C/N802C mutant. Indeed, the percent desensitization of GluK2 K676C/N802C elicited by 1 mM glutamate was dramatically decreased in the presence of 0.5 mM BPAM344, with only 25% desensitization observed after 20 s (Fig. 1A, B). While the peak current amplitude (I_peak) was only increased 1.5-fold (1.5 ± 0.3, n = 10), similar to the WT in the presence of BPAM344, the steady-state current (I_ss) amplitudes were significantly increased, by ~4.7-fold (4.7 ± 0.6, n = 10) (Fig. 1C). Thus, we sought to capture the conformation of receptors in this non-desensitizing state using cryo-EM.

GluK2 K676C/N802C adopts multiple conformations

We expressed the intact homotetrameric GluK2 KAR incorporating K676C/N802C double cysteine mutations in HEK293S GnTl⁻ cells and purified the proteins in the detergent, n-dodecyl-β-D-maltoside (DDM). To prepare samples for cryo-EM analysis, 1 mM glutamate was added to the final protein sample (Supplementary Fig. 2E, F) in the presence or absence of 0.5 mM BPAM344. We first performed three-dimensional (3D) refinement without imposing symmetry (Supplementary Table 1). Our initial global reconstruction of the glutamate/BPAM344-supplemented dataset generated six main classes (Supplementary Fig. 3a). One of the classes exhibited a classical Y-shaped conformation and retained intact LBD dimers (Fig. 1D). To improve the density, we further performed signal subtraction on the ATD and LBD-TMD layers separately. The local refinements with focused masks yielded 3.7 Å and 3.9 Å reconstructions of the ATD and LBD-TMD, respectively, enabling reliable model building (Supplementary Fig. 3a, Supplementary Fig. 4A–C). In the presence of crosslinks between two cysteines and BPAM344, the receptors were stabilized in a two-fold asymmetrical arrangement of the LBD layer (Fig. 1E, F). The extracellular domains resembled the open/active state (PDB code 9B36)⁶⁸, while the ion channel pore remained closed. Therefore, we refer to this class as a non-active conformation. In contrast, the remaining five classes showed no BPAM344 molecules at the LBD D1–D1 interface, resulting in D1–D1 rupture. In all of these five classes, the LBD bi-lobes of all four subunits were tightly closed, with closure degrees ranging from 23° to 27°, indicating glutamate binding to receptors. The overall structures of these five conformations were stabilized in two- or four-fold symmetrical, or asymmetrical arrangements.

To analyze conformations in the absence of allosteric potentiators, we also collected cryo-EM data of GluK2 K676C/N802C supplemented with glutamate. These data revealed four main 3D classes (Supplementary Fig. 3b), each of which was superimposable with the glutamate-bound, BPAM344-unbound classes from the BPAM344-supplemented sample (Supplementary Fig. 3a). Thus, we combined these two data sets and further processed and refined them to improve the density quality. Because D1–D1 dissociation is considered a hallmark of KAR desensitization, these five classes represent different functional states with varying degrees of desensitization. We defined these classes as shallow-desensitized, intermediate, and deep-desensitized states, based on the observations described below. Ultimately, the final resolutions of the desensitized classes ranged from 3.7 Å to 4.0 Å (Supplementary Fig. 4).

GluK2 K676C/N802C mutant structure bound to glutamate and BPAM344 represents a non-active conformation

While all six 3D reconstructions showed three domains—ATD, LBD, and TMD—as previously reported^{25,59,68,69,71–74,123}, each class exhibited distinct conformations (Supplementary Fig. 3). In the non-active conformation, all four LBDs are closed in the presence of glutamate, compared to the open LBD bi-lobe conformation in the apo GluK2 structure (PDB code, 9CAZ)⁶⁰ (Fig. 1G). Two molecules of BPAM344 per LBD dimer bind and stabilize the LBD D1–D1 interfaces, with each BPAM344 forming hydrogen bonds with the side chain of P532, as recently reported^68,72,124 (Fig. 1H, I, Supplementary Fig. 5A). We also observed clear densities of glutamate (Supplementary Fig. 5B). In addition, we observed clear densities corresponding to disulfide bonds between cysteines at positions C676 and C802 (Fig. 1J). Together, the positive allosteric modulator and crosslinks trap the LBDs in a dimer-of-dimers arrangement, in the presence of the agonist (Fig. 1D–F).

Focusing on the LBD layer, the clamshell closure of the LBDs in the presence of glutamate causes separation of the D2 lobes, resulting in expansion of the LBD tetramer at the center of the layer, while the D1–D1 interface remains intact, stabilized by BPAM344 (Fig. 2A, B). This conformation closely resembles that of the active/open class (PDB code, 9B36)⁶⁸, with comparable D1–D1 and D2–D2 distances within the LBD dimer, as well as a similarly expanded LBD tetramer (Fig. 2C, D). Specifically, the distance between the Cα atoms of S670, located at the bottom of the D2 lobes in the AD and BC subunit pairs, is equally separated in the non-active state and in the open state (Fig. 2B, D). To further evaluate the patterns of subunit arrangement, we calculated the center-of-mass (COM) of each lobe and measured the distances between the D1–D1 and D2–D2 lobes in non-active and open structures. This analysis revealed a highly similar overall conformation between the two structures.

Fig. 2. Structural mechanism of GluK2 upon activation and desensitization.

A A top-down view of the cryo-EM reconstructions and models for the BPAM344- and glutamate-bound GluK2 LBD in its non-active conformation. The distances at the LBD dimer–dimer interface, measured between S700 on the AC subunits and E788 on the BD subunits, are highlighted in red and blue, respectively. B Cryo-EM reconstructions and models of the LBD dimers in the non-active state, viewed perpendicular to the membrane. The centers of mass (COMs) for the D1 and D2 lobes are represented by blue and green spheres, respectively. The COM distances between the D1 and D2 lobes of the AD subunits (indicated by circles) are displayed below the models. The distances between the Cα atoms of P773 and S670 within the AD subunits are also presented. Arrows denote bi-lobe closure. C, D A top-down and side views of the BPAM344-, glutamate-, and concanavalin A-bound GluK2 LBD in the open state (PDB 9B36). E Comparison of LBD–TM3 linkers and TM3 formation in BPAM344-bound, agonist-unbound GluK2 (PDB 8FWS)⁷² and non-active GluK2. S670 is shown as spheres in corresponding colours for the non-active state and in grey for BPAM344-bound GluK2. Locations of the TM3 gating hinge at A656 (subunit BD) and E662 (subunit AC) are highlighted in red. The dotted line indicates the positions of pore-lining residues: (1) M664, (2) T660, (3) A656, and (4) T652. (2’) E662, which serves as the gating hinge, is also highlighted. F Comparison of LBD–TM3 linkers and TM3 formation in non-active GluK2 and BPAM344-, ConA- and glutamate bound open GluK2 structure (PDB 9B36)⁶⁸. The locations of the gating hinge at L655 in the A–D subunits of the open GluK2 structure are highlighted in grey. Arrows indicate the rearrangement of TM3. G Pore profile in the non-active structure. Pore-delineating dots are coloured according to pore radius: red for regions with a radius <1.1 Å and grey for regions with a radius >1.1 Å, which would allow passage of a dehydrated ion (1.1 Å for calcium). H The pore radius for the non-active (red) and open/active (PDB code, 9B36, grey) conformations, calculated using HOLE. Source data are provided as a Source Data file.

Key features of the open/active-state conformation of iGluRs include the closure of the LBD bi-lobes, which increases tension in the linkers between the bottom of the LBDs and the pore-forming TM3, resulting in expansion of the ion channel pore, particularly at the gate region^1,125. Given the resemblance of the LBD conformation to the open state, we next assessed the LBD–TM3 linker arrangement, TM3 helix orientation and channel opening to confirm the functional state of the non-active conformation. In this state, clear densities for all TM3 helices were observed (Supplementary Fig. 6A).

To analyze specific conformational features of the non-active state, we compared it with the two-fold symmetrical BPAM344-bound (orthosteric ligand-free) (PDB code, 8FWS)⁷² and open⁶⁸ conformations. In the non-active conformation, the TM3 helices exhibit distinct structural features: they are unwound at A656 in the BD subunits and at E662 in the AC subunits (positions 3 and 2’, respectively) (Fig. 2E). As a result, T660 in the BD subunits and M664 in the AC subunits are displaced away from the central axis of the ion channel. In contrast, the BPAM344-bound, orthosteric ligand-free GluK2⁷² structure has an extra half-turn of the helix at the pore entrance. In this conformation, the TM3 helices unwind at higher positions, specifically at T660 in the BD subunits and M664 in the AC subunits (positions 2 and 1, respectively) (Fig. 2E). Consequently, comparison of these two structures revealed that the top of the gating region, which contains the conserved SYTANLAAF motif^126,127, is more dilated in the non-active conformation than in the more tightly closed BPAM344-bound, orthosteric ligand-free GluK2 conformation.

Next, we compared the non-active state to the open/active conformation (PDB code, 9B36)⁶⁸. In the open state, all four TM3 helices exhibit a pronounced kink at a deeper position at L655 (position 4’) (Fig. 2F). However, in the non-active conformation, the TM3 helices are kinked one helical turn higher. The rearrangement and accompanying rotational movements of the TM3 helices in the open state displace T652 away from the central axis, facilitating channel opening. These movements are absent in the non-active conformation, resulting in a more closed ion channel than in the open conformation (Fig. 2F).

While the critical role of the ATD as an allosteric modulator of gating is well established in both NMDARs^128–132 and AMPAR^91,93,133, this role remains underexplored in KARs. To better understand the functional role of KAR ATDs, we examined the functional state-dependent assembly and orientation of the ATD, focusing on its relative positioning with respect to the LBD layer. In the non-active state, the COM distance between the ATD and LBD increases by 6-7 Å relative to the apo or BPAM344-bound agonist-unbound GluK2 (Supplementary Fig. 7A). This upward ATD shift is consistent with observations in the open state GluK2 (stabilized by a ~50 kDa ConA dimer)⁶⁸, and in the GluK2 complexed to NETO2 (PDB code, 7F5A) (Supplementary Fig. 7A)⁷¹. While the ATD layer undergoes rigid-body movement in the non-active conformation, the ATD layers remain superimposable across all structures (Supplementary Fig. 7B). In addition, the ATD in the non-active class undergoes a 45° counterclockwise rigid-body rotation. This tendency is also observed in the open KAR and open AMPARs¹³⁴, whereas no such rotation is observed in the BPAM344-bound or desensitized conformations (Supplementary Fig. 7C). This observation may imply the function-dependent ATD conformational changes upon activation and desensitization.

Disulfide bonds between K676C and N802C stabilize GluK2 KARs in a typical desensitized AMPAR-like conformation

The glutamate-bound, potentiator-unbound GluK2 K676C/N802C generated five classes (Fig. 3A, B). The LBD layer conformation in these classes appears desensitized; however, they represent different levels of desensitization. In one of these desensitized classes, the absence of BPAM344 results in D1–D1 dissociation, a hallmark structural feature of iGluR desensitization, while two disulfide bonds between K676C and N802C (Fig. 3C) retain LBD dimers in a dimer-of-dimers conformation. We refer to this as the shallow-desensitized conformation to distinguish it from the typical fully or deep-desensitized conformations of KARs⁶⁹. Another class, termed the intermediate state, contains a LBD dimer on one side and a disrupted dimer on the other. In this class, one LBD dimer pair maintains its dimer formation with D1–D1 dissociation, resembling the shallow-desensitized state, while the disrupted dimer adopts a conformation characteristic of deep-desensitized classes. Therefore, we propose that this structure represents an intermediate state between the shallow- and deep-desensitized conformations. In contrast, the other three structures lack inter-subunit disulfide bonds between K676C and N802C, resulting in the LBD layers stabilizing an approximately four-fold symmetric conformation, albeit with varying in-plane rotation angles of the LBDs (Fig. 3B). We observed the formation of a desensitization ring, a network of polar interactions on the lower lobes of the LBDs in all of these deep-desensitized conformations, consistent with previous reports^60,69,70 (Fig. 3D, E). This ring formation plays a key role in regulating desensitization and recovery from desensitization^69,70,135.

Fig. 3. Structures of GluK2 in desensitized states.

A Three-dimensional (3D) cryo-EM reconstructions of glutamate-bound GluK2 K676C/N802C in desensitized conformations. B Top views of glutamate-bound GluK2 K676C/N802C, focusing on the LBD layers. The in-plane rotation angles of the LBDs are indicated. Disulfide bonds formed between the two LBD dimers are highlighted in red dotted boxes. C Close-up view of the disulfide bonds between K676C and N802C at the CD subunit interface (left) and the AC subunit interface (right) in the shallow-desensitized conformation. D Top view of the LBD layer in the deep-desensitized conformation, highlighting helices E and G, which form the desensitization ring. The view is parallel to the membrane, with colouring consistent with A. E The desensitization ring formed in three deep-desensitized conformations, viewed perpendicular to the membrane. The in-plane rotation of the LBD shown in B shifts the LBD–TM3 linker, positioning it above the ion channel pore formed by the TM3 helices.

Comparison of the non-active and shallow-desensitized conformations revealed multiple structural rearrangements in the context of the full-length receptor structure. In the shallow-desensitized state, the distance between the ATD and LBD is reduced by 9 Å compared to the non-active state, making it more similar to the deep-desensitized conformations (Supplementary Fig. 7A). Focusing on the LBD layer, the LBDs in the AC subunits are rotated by 5° parallel to the membrane when viewed from the top, compared to the non-active conformation due to the D1–D1 dissociation (Fig. 4A). Crucially, the D1–D1 interface in this structure lacks the Na⁺ and Cl⁻ ions, a direct consequence of the D1 domain separation (Supplementary Fig. 5C, D). This contrasts with the clear electron densities visible in reported active-state structures, where the D1–D1 association is stabilized by two Na⁺ and one Cl⁻ ions^47,72,136 (Supplementary Fig. 5C, D). Viewed from the side, the entire LBDs appear rolled down by 21°, associated with the rearrangement of the dimer-of-dimers formation and the rupture of the D1–D1 interface (Fig. 4B).

Fig. 4. GluK2 stabilized in a shallow-desensitized-like conformations resembles the desensitized conformation of AMPARs.

A, B Comparison between the non-active and shallow-desensitized conformations, viewed parallel (A) and perpendicular (B) to the membrane. Black arrows indicate the degree of LBD domain rotation. C, D Structural comparison of LBD layers in GluK2 K676C/N802C in the shallow-desensitized state and GluA2 AMPAR in the desensitized state (PDB 7RYZ)⁸⁰, viewed parallel (C) and perpendicular (D) to the membrane. Distances between S700 in the AC subunits and E788 in the BD subunits are shown in red and blue, respectively. The locations of the centers of mass (COMs) of D1 and D2 are indicated by blue and green spheres, respectively, with the COM distances between D1 and D2 of the AD subunit of the LBDs displayed below. Arrows indicate D1 lobe rotations compared to the non-active (KAR) or active (AMPAR) states. Distances between the Cα atoms of P773 (GluK2)/P745 (GluA2) and S670 (GluK2)/S635 (GluA2) in the AD subunits are indicated. E Comparison of LBD-TM3 linkers and TM3 helices in non-active and shallow-desensitized conformations. S670 is shown as spheres in corresponding colours for the shallow-desensitized conformation and in gray for the non-active conformation. The cross-dimer distances between the Cα atoms of S670 are indicated. Residues forming the ion channel pore are displayed as sticks. The locations of the TM3 gating hinge at T660 (subunit BD) and M664 (subunit AC) in the shallow-desensitized conformation are highlighted in red. The shallow-desensitized conformation contains an additional half-turn of helices at T660 compared to the non-active conformation. F Comparison of LBD–TM3 linkers and TM3 formation in shallow- and deep-desensitized GluK2. In the deep-desensitized conformation, the ion channel is sealed at M664, completely closing the ion channel pore, in contrast to the shallow-desensitized conformation, which has a kink at T660 in the BD subunits. G Extracellular views of the ion channel in the shallow-desensitized and deep-desensitized states, showing M664 positioned at the top of the gate region as sticks (top). The schematic illustration compares these two desensitized states, highlighting how M664 partially or completely seals the pore (bottom).

These conformational changes are also observed during AMPAR desensitization^{78–80,89,137}. Therefore, we next compared the shallow-desensitized conformation of GluK2 K676C/N802C with one of the desensitized conformations of AMPARs (rat GluA2 with auxiliary subunit GSG1L complexed to agonist quisqualate, PDB code, 7RYZ)⁸⁰. Viewed from the top, the distance between the Cα atoms of S700 (S662 in GluA2 AMPAR) in both structures is reduced compared to their respective open states, although the arrangements in these two structures differ due to the restricted mobility of GluK2 LBD caused by cysteine crosslinks (Fig. 4C). The D1 lobes of GluK2 roll down about 25°, in contrast to the 33° rolling-down motion observed in the D1 lobes of GluA2 (Fig. 4D). The D2 distances in the shallow-desensitized GluK2 and desensitized GluA2 are comparable, measuring 20 Å and 24 Å, respectively. Overall, the shallow-desensitized conformation of the GluK2 mutant appears to resemble the desensitized conformation of AMPARs. However, this KAR conformation is presumably energetically unstable and was not observed in the absence of crosslinks.

To better understand the functional characteristics of the shallow-desensitized state, we next compared its TM3 helices with those in the non-active and shallow-desensitized states. We observed clear densities of TMDs except TM2 helices (Supplementary Fig. 5E, Supplementary Fig. 6A). This structural comparison revealed occlusion of the pore in the shallow-desensitized state at T660 in the BD subunits and M664 in the AC subunits (positions 2 and 1, respectively) (Fig. 4E). Clear densities for these two residues and their associated linkers were observed, confirming their positions (Supplementary Fig. 6B, C). In contrast, the pore in the non-active state is more dilated, with kinks observed at the lower positions at A656 in the BD subunits and E662 in the AC subunits as discussed earlier (positions 3 and 2’, respectively) (Fig. 2E, Fig. 4E). A comparison between the shallow- and deep-desensitized conformations revealed that the pore in the deep-desensitized state is more tightly and completely sealed at M664 (position 1) in both the AC and BD subunit pairs (Fig. 4F). Viewed from the extracellular side, the pore in the deep-desensitized state appears well blocked at M664 and T660 (Fig. 4G, Supplementary Fig. 6B, C). In contrast, the shallow-desensitized conformation exhibits only partial blockage due to the outward displacement of M664 in the BD subunits (Fig. 4G). These observations suggest that the canonical deep-desensitized conformation results in a more tightly sealed ion channel pore at the top of the gate.

A large in-plane rotation of the LBD is required for proper ion channel closure during desensitization in GluK2 KAR

Given the substantial conformational differences in the extracellular domains and TMD, we next compared the ion channel pore across desensitized conformations. The shallow-desensitized state shows restriction points at T652, A656, and T660 but maintains a more dilated pore than other desensitized conformations (Fig. 5A). The transition state, an intermediate functional state between the active and fully desensitized states, resembles the shallow-desensitized state but features a tight constriction (<1 Å) at T652 (Fig. 5A, B). In contrast, all three deep-desensitized conformations exhibit multiple tight constrictions, particularly at the top of the gate region in addition to the T652 site at the bottom of the gating region (Fig. 5A, C). Notably, an additional constriction at M664 at the top of the channel pore is generated by the pronounced in-plane rotation of the LBDs in the BD subunits upon desensitization, which twists the LBD–TM3 linkers (Fig. 4F). This arrangement is not seen in major AMPARs and NMDARs.

Fig. 5. Ion channel pore of desensitized GluK2 KAR.

A Permeation pathway dictating the pore diameter of GluK2 K676C/N802C in five desensitized states. Pore-delineating dots are colour-coded based on radius: red indicates regions with a radius <1.1 Å, while grey denotes regions >1.1 Å. B, C Pore profile comparisons. A comparison of the pore profiles between the shallow-desensitized and intermediate states (B), and a comparison of the pore profiles among three distinct deep-desensitized states (C). D Permeation pathway of the desensitized GluA2 AMPAR (PDB code 7RYZ)⁸⁰. The pore radius is colour-coded according to the scheme in A. E A pore profile comparison between GluK2 K676C/N802C in the shallow-desensitized conformation and the desensitized GluA2 AMPAR (PDB code, 7RYZ). Source data are provided as a Source Data file.

As previously reported^69,70, the desensitization ring at the bottom of the LBD layer in the four-fold symmetrical deep-desensitized conformations (Fig. 3D, E) is stabilized by polar interactions that regulate desensitization and desensitization recovery. The formation of this ring in the deep-desensitized state stabilizes the twisted LBD–TM3 linkers, thereby securing the tightly closed pore, particularly at M664 (Fig. 5A). In contrast, the shallow-desensitized state lacks this pore-locking mechanism at T660 and M664, resulting in a less stable and less tightly sealed channel.

Furthermore, although the extracellular domains of the shallow-desensitized GluK2 KAR and the desensitized GluA2 AMPAR adopt similar conformations, their ion channel pore structures differ markedly (Fig. 5D). Notably, the pore diameter at T660 and T652 in the shallow-desensitized GluK2 KAR is wider than that in the desensitized GluA2 AMPAR which contains tight constriction points at T625 and T617 (PDB code, 7RYZ)⁸⁰ (Fig. 5E). This observation also suggests that the pronounced lateral rotation of LBDs in KARs contributes to secure pore closure and stabilization, distinguishes from major AMPARs.

Overall, based on these functional and structural observations, we hypothesize that the GluK2 KAR K676C/N802C mutant, which introduces cysteine crosslinks, undergoes a shallow desensitization distinct from that of the canonical KAR desensitization pathway. This altered mechanism, characterized by unstable or partial channel closure, may allow the receptors to remain ion-permeable during desensitization, as observed in our functional data (Fig. 1A). Moreover, we propose that this mutant may recover more rapidly from desensitization, similar to the rapid recovery observed in AMPARs.

K676C–N802C cysteine crosslinks partially block receptors from undergoing deep desensitization

To test our hypothesis, we first assessed the kinetics of the GluK2 K676C/N802C mutant using whole-cell patch-clamp electrophysiology. Remarkably, we observed steady-state currents in the cross-linked mutant even at the highest concentration of glutamate (50 mM), in contrast to GluK2 WT, which exhibited complete desensitization under the same conditions. (Fig. 6A). There were also no significant differences in the current density at peak current between GluK2 WT and the mutant (GluK2 WT: 830 ± 50 pA/pF, n = 6; GluK2 K676C/N802C: 700 ± 150 pA/pF, n = 9), or in the EC₅₀ for glutamate (GluK2 WT: 110 ± 30 µM, n = 7; GluK2 K676C/N802C: 150 ± 50 µM, n = 5), indicating that the cysteine crosslinks do not significantly affect agonist binding affinity or receptor activation (Fig. 6A–D). GluK2 K676C/N802C was only partially desensitized even with 0.01 mM glutamate, a concentration below the EC₅₀ (~0.17 mM). This suggests that even partial occupancy is sufficient to initiate desensitization, consistent with earlier findings^138,139. While GluK2 WT was desensitized almost completely with a time constant of 8 ± 1 ms (n = 14), GluK2 K676C/N802C receptors in the presence of 0.01–50 mM glutamate exhibited ~60% desensitization, except when activated by 0.001 mM glutamate where only the steady-state current was observed with a decreased peak amplitude (Fig. 6E). Therefore, the degree of desensitization was not glutamate concentration-dependent.

Fig. 6. Functional characterization of GluK2 K676C/N802C receptor.

A Representative whole-cell patch-clamp recordings of GluK2 WT (left) and GluK2 K676C/N802C (right) in response to 0.001–50 mM glutamate. The black bar represents the glutamate application. B Quantification of current density (pA/pF) at the peak current for GluK2 WT (gray, n = 6) and K676C/N802C (blue, n = 9) activated by 10 mM glutamate. A two-sided two-sample t-test was performed, and no significant differences were observed (P = 0.324). C Glutamate dose-response curves for GluK2 WT (n = 12) and K676C/N802C (n = 8), normalized to the maximal response (50 mM). Glutamate was applied for 1 s, and the peak current was measured at each concentration. All concentrations were tested in the same cell. D EC₅₀ values for glutamate activation of GluK2 WT (gray, n = 7) and K676C/N802C (blue, n = 5). A two-sided two-sample t-test was performed, and no significant differences were observed (P = 0.471). E Ratio of steady-state to peak current (I_steady-state/I_peak) (50 mM, n = 6; 10 mM, n = 10; 1 mM, n = 6; 0.01 mM, n = 6; 0.001 mM, n = 6). The values of the steady-state current were measured at the end of the activation period. F Representative current traces of GluK2 WT (black) and K676C/N802C (blue) during a two-pulse protocol, showing the interval between glutamate exposures ranging from 50 ms to 20 s. G Recovery from desensitization at increasing inter-sweep intervals, fitted with the Hodgkin-Huxley (I(t)=[(I_max^1/m-(I_max^1/m-I₀^1/m)) ×e^-t/τ]^m for GluK2 WT and two-term exponential functions for GluK2 K676C/N802C. The K676C/N802C mutant recovers faster than WT. H Time constant (τ) of recovery from desensitization, showing a significant reduction in τ for the K676C/N802C mutant (P < 0.0001, two-sided two-sample t-test). The data for the WT were obtained from the fit in (G), and for K676C/N802C, a mean τ was calculated using τ_mean = ((A1 τ1 + A2 τ2) / (A1 + A2)). Biologically independent measurements: n = 11 for GluK2 WT and n = 10 for K676C/N802C. Black circles represent biological independent measurements. Data are presented as mean ± SD; whiskers indicates the Standard Deviation (SD). Statistical analysis was performed by applying a two-sided two-sample two-tailed t-test. Exact P-values are shown in the figure; significance is assumed if P < 0.05. Source data are provided as a Source Data file.

To assess the desensitization state of the crosslinked mutant, we next analyzed the rate of recovery from desensitization. As previously reported^36,60,69,140, recovery from desensitization in the GluK2 WT is very slow, requiring 1150 ± 90 ms (n = 9) for full recovery. In contrast, the GluK2 K676C/N802C mutant exhibited faster recovery from desensitization, characterized by two distinct components: τ₁ (100 ms, 60%) and τ₂ (1000 ms, 40%) (n = 9). The mean τ of the mutant was 440 ± 40 ms (n = 9), approximately three times faster than that of GluK2 WT (Fig. 6F–H). This shift indeed brings its kinetics closer to those of AMPARs, which recover in the order of tens to hundreds of milliseconds (e.g. GluA1 and GluA2 WT recovery constants are ~200 ms and 9–50 ms, respectively)^91,141,142. This result highlights the impact of cysteine cross-linking in accelerating desensitization recovery by stabilizing the dimer-of-dimers conformation.

We then assessed the ion permeation pathway of the double-cysteine mutant using polyamines. Polyamines, such as spermine (SPM), are permeant channel blockers that bind within the ion channel pore and inhibit ion flow⁷⁴. The peak current of the GluK2 K676C/N802C mutant was reduced by ~50% in the presence of 20 mM spermine, similar to the WT, while its steady-state current showed a slightly lower degree of inhibition (Supplementary Fig. 8A). Specifically, GluK2 WT exhibited peak current inhibition of 54% ± 6% (n = 5), while GluK2 K676C/N802C showed peak and steady-state current inhibition of 53% ± 7% and 40% ± 16% (n = 6), respectively (Supplementary Fig. 8A, B). These data confirm that ion permeation in the double-cysteine mutant occurs through the ion channel pore and that the mutant retains a functional selectivity filter that is not restricted by the disulfide crosslinks.

GluK2 KAR stabilized in a shallow-desensitized conformation remains ion-permeable

To assess whether the non-active current and faster recovery kinetics result from crosslinking between the K676C and N802C, we performed whole-cell recordings in the presence of 5 mM dithiothreitol (DTT), applied for 7 min prior to the second glutamate activation. The peak intensity of both GluK2 WT and GluK2 K676C/N802C, in the presence and absence of DTT, showed no significant changes, confirming our earlier analysis that the crosslinks do not affect activation kinetics (Fig. 7A). In contrast, a 7-min application of 5 mM DTT reduced the steady-state current of GluK2 K676C/N802C by >90% (Fig. 7A), with the steady-state current ratio (Iss/Ipeak) shifting from 0.4 ± 0.1 to 0.03 ± 0.03 (n = 8) (Fig. 7B). A similar trend was observed in the presence of BPAM344 and DTT (Supplementary Fig. 8C–F). In addition, recovery from the desensitization of the DTT-treated cells expressing GluK2 K676C/N802C was 1,040 ± 50 ms (n = 5), similar to the WT (Fig. 7C, D). This experiment, together with the assessment of the single-cysteine mutants, GluK2 K676C and GluK2 N802C (Supplementary Fig. 2A), confirms that the observed non-desensitizing efficacy is specifically due to disulfide bond formation, which stabilizes the two-fold symmetrical non-active or shallow-desensitized conformations as originally intended.

Fig. 7. Effects of DTT and BPAM344 modulation on the steady-state current of GluK2 K676C/N802C.

A Representative whole-cell patch-clamp recordings of GluK2 WT (left) and GluK2 K676C/N802C (right) activated by 10 mM glutamate before (black) and after (red) treatment with 5 mM DTT for 7 min. B I_ss/I_peak ratio in GluK2 K676C/N802C (n = 7), showing a significant reduction in steady-state current (I_ss) (P < 0.0001, paired t-test). C Recovery from desensitization at increasing inter-sweep intervals after treatment with 5 mM DTT. The recovery kinetics of GluK2 K676C/N802C post-DTT treatment were fitted with the Hodgkin-Huxley (I(t)=[(I_max^1/m-(I_max^1/m-I₀^1/m)) ×e^-t/τ]^m). D Time constant (τ) of recovery from desensitization, showing that DTT treatment restores the recovery kinetics of GluK2 K676C/N802C (n = 5) to a level comparable to WT (n = 9) (P = 0.258, two-sided two-sample t-test). E Representative whole-cell responses to 10 mM glutamate (200 ms, −70 mV; black bar) of GluK2 WT (left, average current in black, n = 60 individual sweeps in gray) or GluK2 K676C/N802C (right, average current in blue, 100 individual sweeps in gray). Inset: Current-variance relationship, with WT data fitting a parabolic function (dotted line), whereas the K676C/N802C mutant does not follow this trend. F Whole-cell recordings of GluK2 K676C/N802C in response to 10 mM glutamate in the absence (blue) and presence (green) of 500 µM BPAM344. Right: Representative noise traces at the end of glutamate application ( ~ 10 s). Background noise was taken from the current recording before glutamate application. G Variance-mean current relationship for individual cells (n = 6), indicating an increase in channel variance-current relationship in the presecen of BPAM344. Data were obtained from currents shown in panel F. H Unitary conductance determined by variance-mean analysis (n = 6), showing no significant difference with and without BPAM344 (P = 0.280, two-sided paired t-test). I Estimated number of open channels, showing a significant increase after BPAM344 application (P < 0.0001, two-sided paired t-test). The black circles represent biologically independent measurements. Data for glutamate and BPAM344 conditions were obtained from the same cell. Data are presented as means, with whiskers representing standard deviation (SD). Statistical significance is indicated as NS (not significant), and p-values are provided in the figure legend or shown in the figure. Source data are provided as a Source Data file.

Next, we aimed to determine whether crosslinked receptors remain ion-conductive while exhibiting steady-state current. A key feature of KARs is their exceptionally small conductance and low open-channel probability^39,143, which makes it technically challenging to directly observe channel activity at the single-channel level, unlike other channels, such as AMPARs. Therefore, we conducted noise analysis to assess KAR unitary conductance. We first performed nonstationary fluctuation analysis (NSFA) on the desensitizing phase of macroscopic currents from GluK2 WT and GluK2 K676C/N802C, activated by glutamate pulses (10 mM, 100 ms). The variance of fluctuations around the expected current was analyzed. The weighted mean single-channel conductance of GluK2 WT was 23 ± 7.8 pS (n = 5), consistent with a previous report³⁹. While the current–variance relationship of the WT was well-fitted by a parabolic function, the analysis of GluK2 K676C/N802C revealed an anomalous current-variance relationship, preventing the calculation of its conductance (Fig. 7E). However, the mutant generated a large fractional steady–state current. Thus, we analyzed stationary fluctuations in the steady-state current of the desensitized receptors in the presence and absence of BPAM344.

To assess unitary conductance and the number of active channels, we analyzed the macroscopic currents of GluK2 K676C/N802C activated by 10 mM glutamate for 30 s under desensitizing conditions (Fig. 7F). When activated, the mutant exhibited a unitary conductance of 10 pS, less than half of the WT conductance calculated by NSFA, and was similar to the lowest conductive level (O1) previously determined for KARs³⁹. Current–variance analysis across individual cells showed that BPAM344 shifted receptor behavior toward higher variance and mean current values (Fig. 7G), suggesting an increase in the number of open receptors and in their open probability. Interestingly, the unitary conductance of the mutant activated by 10 mM glutamate was similar in the presence and absence of BPAM344 (GluK2 K676C/N802C, 10 mM glutamate: 10 ± 3 pS, n = 6; GluK2 K676C/N802C, 10 mM glutamate + 0.5 mM BPAM344: 10 ± 2 pS, n = 6) (Fig. 7H). In contrast, the number of active channels in the presence of BPAM344 was approximately four-fold higher than in its absence (GluK2 K676C/N802C, 10 mM glutamate: 1040 ± 600, n = 6; GluK2 K676C/N802C, 10 mM glutamate + 0.5 mM BPAM344: 4,080 ± 460, n = 6) (Fig. 7I). In summary, these experiments indicate that desensitized GluK2 K676C/N802C KARs remain conductive at the single channel level, although with significantly lower conductance compared to GluK2 WT channels activated by glutamate.

Discussion

Receptor desensitization is a fundamental property of ligand-gated ion channels. Recent advances in iGluR research have revealed significant differences in the kinetics and conformational dynamics of KARs during desensitization compared to other iGluRs^{1,25,59,69,70,73}. The desensitized conformation of KARs is characterized by quasi-four-fold symmetrical LBDs with complete LBD dissociation. This contrasts with the desensitized states of other major iGluRs, particularly the most abundant native AMPARs and NMDARs, which typically retain two-fold symmetrical conformations with some exceptions described above. Functionally, KARs also differ from their AMPAR counterparts; most KARs enter deep-desensitized states, with recovery from desensitization occurring 5- to 130-fold slower than AMPARs.

Here, we employed a classical cysteine crosslinking approach and unveiled the structures of the GluK2 KAR, revealing the arrangements associated with receptor activation and desensitization. In the structure of the non-activated state of GluK2 in the absence of modulatory proteins, the extracellular domain adopts an active-like conformation, characterized by D1–D1 association and LBD bi-lobe closure, which are stabilized by glutamate, the positive allosteric potentiator BPAM344 and cysteine crosslinking. Interestingly, even in the absence of ConA and NETO proteins, which sit between the ATD and LBD layers and raise the ATD^68,71, we also observed the elevated ATD in this state. This represents a marked contrast to apo and desensitized receptor conformations, which adopt more compact extracellular domain arrangements, with the ATD layer positioned closer to the LBDs. Given that ATD-deleted KARs remain functional as ion channels^47,68, further investigation is needed to determine how these conformational changes in the ATD layer contribute to receptor activation. Despite the active-like arrangement of the extracellular domain, the ion channel in this state remains closed, with kinks observed at A656/E662 in the upper gate. In contrast, full channel opening requires rotation of TM3 and kinking at L655 located at the lower gate region across all four subunits, as seen in the GluK2 structure in the maximal conductance O4 active state⁶⁸ (Fig. 8).

Fig. 8. Schematic representation of conformational rearrangements during activation and desensitization.

This schematic model details how conformational changes within the LBD layer regulate ion channel permeation during desensitization. The LBD tetramer conformation for each state is depicted within a circle. The position of the gating kink varies across the non-active, open, and desensitized states. The shallow-desensitized state is specifically observed when the LBD dimer-of-dimers conformation is stabilized by cysteine cross-linking. While the shallow-desensitized conformation exhibits incomplete closure at the top of the channel gate, the pore of the deep-desensitized states is tightly and stably sealed by the T660 and M644 constriction sites. For comparison, the previously determined open state (PDB code: 9B35⁶⁸) is shown. The apo state conformation is represented by two structures (PDB codes: 9CAZ⁶⁰ and 8FWS⁷²).

Structures of the glutamate-bound GluK2 receptor in the shallow, deep, and intermediate desensitized conformations revealed the fundamental reason why KARs require robust rotational movements of their LBDs upon desensitization. Two inter-dimer crosslinks in the shallow-desensitized state stabilize receptors in a two-fold symmetrical conformation. However, in the absence of BPAM344, we observed that the D1–D1 interface is ruptured, likely due in part to its instability without cation and anion stabilization at that interface. The D2 domains, meanwhile, remain closely associated. Thus, the overall conformation of this state resembles the desensitized state of major AMPARs^{78–80,89,113}.

Although asymmetrical KAR conformations have been reported^59,68,71,73, the shallow-desensitized conformation described here has not been observed in prior studies likely due to its energetic instability. This stands in contrast to major AMPARs, for which a four-fold symmetrical deep-desensitized conformation is energetically unfavorable for the predominant subtypes containing GluA2 subunit complexed to auxiliary proteins¹⁴⁴. As a result, most predominant AMPARs remain in a two-fold symmetrical arrangement during desensitization, unlike the four-fold symmetrical desensitized state of KARs. Previous mutagenesis studies demonstrated that various mutations in the D2 lobe of KARs alter desensitization and recovery from desensitization but that no mutation completely blocks desensitization, unlike AMPARs¹³⁵. This suggests that multiple non-conserved residues contribute to the stability of the four-fold desensitized conformation of KARs, despite the overall architectural similarity between AMPARs and KARs.

The shallow-desensitized conformation revealed a dilated ion channel, unlike the narrower pores found in desensitized AMPARs and fully (or deeply) desensitized KARs. Specifically, the deep-desensitized states have a tightly sealed channel pore, capped by residues at the top of the gating region. This cap results from a large rotational movement of the LBDs, and this conformation is locked by the desensitization ring, as observed previously^69,70. Accordingly, our structural observations indicate that the existence of KARs in a two-fold symmetrical conformation mimicking the desensitized major AMPAR does not fully and stably close the pore. Instead, substantial rotational movement of the LBDs is required to achieve complete channel closure and/or to greatly stabilize this channel-closed desensitized state. This conformational rearrangement reinforces receptor desensitization and significantly influences the kinetics of desensitization recovery^69,70.

Our electrophysiological data supports the idea that the crosslinked receptors can form weakly conductive channels. Specifically, whole-cell patch-clamp recordings demonstrated that cysteine crosslinking prevents the receptors from undergoing complete desensitization. This intervention significantly reduced the extent of desensitization and markedly accelerated the recovery kinetics, making the receptors functionally resemble AMPARs rather than GluK2 WT. Crucially, treatment with 5 mM DTT restored the recovery kinetics to those of GluK2 WT, indicating a direct correlation between LBD conformation, ion conductance, and the processes of desensitization and recovery in GluK2 KARs. Furthermore, our fluctuation analysis of the crosslinked GluK2 receptor also demonstrated that the desensitized, crosslinked GluK2 remains conductive at the single-channel level, albeit with a conductance ~30% lower than that of GluK2 WT (∼23 pS versus ∼10 pS), regardless of the presence of the positive allosteric modulator BPAM344. This indicates that GluK2 WT and the crosslinked mutant have distinct distributions of conductance levels³⁹. Despite the spatial proximity and disulfide bond-forming capability of the engineered cysteines, it remains unclear whether this crosslinked conformation mimics a transient intermediate state between active and deep-desensitized states or represents an artificial configuration. Notably, a conducting desensitized state of AMPARs and acetylcholine receptors has been reported^107,145. In addition, it was microscopically shown that the partially desensitized KARs are conductive, as defined by the number of desensitized subunits per KAR tetramer⁵⁴. Overall, our structural and functional observations suggest that robust LBD rotation is essential for complete channel closure, a mechanism that appears to be critical for KARs and not conserved across other iGluRs.

In addition to the function-dependent conformation, we also observed glycan chains in our structures (Fig. 1D). We and others have previously reported function-dependent receptor–glycan interactions at the ATD-LBD domain and at the protomer interfaces, as well as functional modulation of receptors by N-glycans^58–60. Notably, density corresponding to the N275 N-glycan was observed forming inter-protomer interactions exclusively in the shallow-desensitized state, consistent with our previous observations. These interactions were absent in the non-active state and deep-desensitized state, likely due to ATD elevation and LBD arrangement, respectively. This spatially restricted glycan localization of the glycan to the ruptured D1–D1 interface supports its role in modulating receptor function.

The conformational motions described here are inferred transitions between static, high-resolution snapshots obtained via cryo-EM. We acknowledge this limitation and interpret these transitions as approximations of the dynamic gating process. Nevertheless, this study addresses a fundamental question by demonstrating that KARs require substantial conformational rearrangements of the LBD layer and stabilization in a four-fold symmetrical conformation to achieve proper channel closure and full desensitization into a non-conductive state (Fig. 8). Our study utilizes the recombinant homomeric GluK2 receptor, which undergoes strong and rapid desensitization upon prolonged agonist exposure. While native heteromeric KARs, often complexed with auxiliary proteins in brain tissue, typically show reduced desensitization, our dissection of the intrinsic properties of the GluK2 subunit provides a fundamental framework for deciphering the more complex physiological behavior of KARs.

Methods

Plasmid construction

For Cryo-EM studies, the construct used for the structural studies was full-length rat GluK2 KAR (GenBank 54257, Uniprot code P42260) with two mutations of N802C and K656C. The gene was synthesized, subcloned into the pFW vector (gift from Dr. Hiro Furukawa at Cold Spring Harbor Laboratory) and fused to a C-terminal human rhinovirus 3 C protease recognition site and a 1D4 affinity tag (Genscript). The construct was referred to as GluK2 K656C/N802C. For western blot analysis to assess crosslinks between K676C and N802C, a series of five constructs was designed. These were created by introducing the following mutations into the GluK2 backbone: C595A, C595A/C871A, C595A/C871A/K676C, C595A/C871A/N802C, and C595A/C871A/K676C/N802C. The native cysteines, C595 and C871, which are known to form inter-protomer disulfide bonds, were mutated to alanine to prevent the observation of disulfide bonds naturally formed in the wild-type GluK2 protein.

Cell culture

For structural studies, HEK293S GnTI⁻ cells were grown to a density of 3.2 × 10⁶ cells / ml in FreeStyle 293 medium (Supplementary Table 2) at 37 °C and 8% CO₂ supplemented with 2% fetal bovine serum. The rGluK2 N802C/K676C mutant bacmid and baculovirus were generated as previously described. The P1 and P2 viruses were produced in Sf9 cells (Supplementary Table 2). Cells were infected with the baculovirus harboring GluK2 EM and incubated at 37 °C for 12 hrs. Cells were supplemented with 10 mM sodium butyrate and 20 µM DNQX and temperature was shifted to 30 °C at 12 h post-infection. The cell culture was incubated additional 60 hrs and harvested by low-speed centrifugation at 5000 g for 20 min and stored at −80°C until use.

For electrophysiology experiments, human embryonic kidney 293 T (HEK293T) cells were cultured in Dulbecco’s modified Eagle medium (DMEM, Corning) supplemented with 10% FBS at 37 °C in a 95% O₂–5% CO₂ atmosphere (Supplementary Table 2). Wild-type rat GluK2 and rat GluK2 K676C/N802C DNAs were cloned into the pCAG-IRES-EGFP vector (Addgene plasmid #119739). HEK293T cells were transfected with 1 µg/µl cDNA using the TransIT2020 transfection reagent (Mirus) according to the manufacturer’s instructions. Cells expressing GluK2 wild-type were transfected for 24 h, while those expressing K676C/N802C were transfected for 18 h. After transfection, cells were dissociated using Accutase (Innovative Cell Technologies, Inc.), resuspended, and plated onto 35 mm poly-D-lysine-coated dishes (Neuvitro). Electrophysiological recordings were performed 2–4 h later⁴⁸.

Electrophysiological recordings

All whole-cell recordings were performed using HEKA EPC10 amplifiers (HEKA Elektronik, Lambrecht, Germany) with thin-wall borosilicate glass pipettes (2–5 MΩ) coated with dental wax to reduce electrical noise. Currents were recorded at a holding potential of −70 mV, with a sampling frequency of 10 kHz, and filtered at 2.6 kHz. The external solution contained (in mM): 145 NaCl, 2.5 KCl, 1.8 CaCl₂, 1 MgCl₂, 5 glucose, and 5 HEPES. The internal pipette solution contained (in mM): 105 NaCl, 20 NaF, 5 Na₄BAPTA, 0.5 CaCl₂, 10 Na₂ATP, and 5 HEPES. The pH and osmotic pressure of the external and internal solutions were adjusted to 7.4 and 300–290 mOsm/kg, respectively. L-glutamate was applied using theta glass tubing mounted on a piezoelectric stack (MXPZT-300 series, Siskiyou) driven by a HEKA EPC10 amplifier. Typical 10–90% rise times were 250–300 µs, measured from junction potentials at the open tip of the patch pipette after recordings. Data acquisition was performed using PULSE software (HEKA Elektronik, Lambrecht, Germany).

Data analysis

The macroscopic rate of desensitization (τ_desensitization) was measured by fitting an exponential function to the decay of current from ~80% of its peak amplitude (I_peak) to baseline in recordings where glutamate was applied for 1 s. For the double mutant N802C/K676C, only the first exponential component, which exhibited desensitization kinetics, was considered, while the steady-state current was omitted. Desensitization kinetics were fitted using a single-exponential, one-term Levenberg-Marquardt fitting approach.

To evaluate recovery from desensitization, two protocols were employed: (1) a two-pulse application of 10 mM glutamate for 100 ms per pulse, with interpulse intervals increasing in 100 ms steps from 50 ms to 2 s; and (2) a second protocol with interpulse intervals increasing in 2 s steps, ranging from 50 ms to 20 s. In all cases, the first peak was considered the control (100% activation), while the second peak represented recovery at a given time point. Data from both protocols were combined, and a recovery curve was generated, plotting the percentage of recovery against time. The experimental data for the WT were fitted to the Hodgkin-Huxley model:

$$I\left(t\right)={\left[I_{max}^{1/m}-\left(I_{max}^{1/m}-I_0^{1/m}\right)\times e^{-t/\tau }\right]}^m$$ 1

While the double mutant was fitted to a double exponential function:

$$I\left(t\right)=A_1*\left(1-e^{-\frac{t}{{\tau }_1}}\right)+A_2*\left(1-e^{-\frac{t}{{\tau }_2}}\right)$$ 2

To analyze the EC₅₀ for glutamate in wild-type (WT) and mutant receptors, varying concentrations of L-glutamate were applied to a single cell using a theta glass tubing mounted on a piezoelectric stack (MXPZT-300 series, Siskiyou), driven by a HEKA EPC10 amplifier, for 1 s. One input of the theta glass tubing was connected to the control solution, while the other was connected to a six-inlet manifold for different L-glutamate concentrations. A 2-min wash was used between concentration changes to ensure complete exchange before applying the next solution. The L-glutamate concentrations tested were 50, 10, 1, 0.1, 0.01, and 0.001 mM, with 50 mM serving as the normalization reference. The normalized data are presented as mean ± S.E.M. in a dose-response curve, with response on the y-axis and concentration on the x-axis. Experimental data from independent biological measurements (each defined as data from one cell with at least five concentration responses) were fitted to the Hill equation. The resulting EC₅₀ values were then averaged and presented as mean ± S.E.M.

Non-stationary fluctuation analysis (NSFA)

NSFA was performed on the decay phase of currents evoked by 200 ms applications of 10 mM glutamate (60–150 successive applications). Currents were sampled at 50 kHz and filtered at 5 kHz. The variance and mean current were calculated for each set of currents and then binned into 15 bins. The single-channel current and total number of channels (N) were determined by plotting the binned variance against the binned mean current and fitting the data with a parabolic function:

$${\sigma }^2=iĪ-\frac{{Ī}^2}{N}+{\sigma }_B^2$$ 3

Where ${\sigma }_B^2$ represents the background variance, which was calculated from the baseline before glutamate application. The ${\sigma }_B^2$ value was used to correct the variance, while the mean I_B was used to correct the mean current.

Stationary fluctuation analysis (SFA)

HEK293T cells transfected with GluK2 K676C/N802C were activated with 10 mM glutamate for 30 s, followed by a 2-min incubation with 500 µM BPAM344. The cells were then reactivated with 10 mM glutamate + 500 µM BPAM344. Currents were sampled at 10 kHz and filtered at 2 kHz. For each condition (glutamate alone and glutamate + BPAM344), the last 10 s of the steady-state current were used to calculate the variance and mean current. Since the stationary current represented ~35% of the peak current, we assumed that the open probability (P₀) in the steady-state current was P₀ < 1. The unitary current (i) was then estimated using the following equation:

$$i=\frac{{\sigma }^2}{Ī}$$ 4

The unitary current was then used to estimate the unitary conductance (γ) as follows:

$$\gamma =\frac{i}{Vm-Vr}$$ 5

where V_m is the holding potential ( − 70 mV) and the reversal potential (Vᵣ) is 0 mV.

The number of channels (N) was estimated using the equation:

$$N=\frac{{Ī}^2}{{\sigma }^2}$$ 6

Statistical analyzes were performed using OriginLab 2025. Statistical significance was calculated using a two-sided two-sample t-test, with significance assumed if P < 0.05.

Cysteine cross-linking and western blots

For western blots, HEK293T cells were cultured for 48 hrs post-transfection. Harvested cells were sonicated in TBS buffer (20 mM Tris pH 8.0, 150 mM NaCl) and solubilized in 200 uL of lysis buffer (20 mM Tris pH 8.0, 150 mM NaCl, 1% DDM, 1 mM PMSF) for 2 hrs. Samples were centrifuged at 186,000 g. Supernatants were subjected to SDS-polyacrylamide gel electrophoresis (4-15 %) in the presence and absence of 300 mM β-mercaptoethanol (β-ME). The proteins were dry transferred to polyvinylidene fluoride (PVDF) membranes. The membrane was blocked with 5% milk in TBST (20 mM Tris-HCl pH 8.0, 150 mM NaCl and 0.05% Tween-20), then incubated with the anti-GluK2 monoclonal antibody (2000-fold dilution, 04-921, Sigma-Aldrich), followed by horseradis peroxidase (HRP)-conjugated anti-rabbit antibody (10,000-fold dilution, ab97051, abcam). Protein bands were detected by chemiluminescence detection kit (Cytiva).

Protein purification

Cell pellets were resuspended in ice-cold lysis buffer containing 20 mM Tris-Cl (pH 8.0), 150 mM NaCl, 0.2 mM PMSF, 20 µg/mL DNase I, and 20 µM DNQX. Cells were lysed by sonication using a Qsonica sonicator with an amplitude of 40% for 10 minutes total (10 s on, 10 s off). The lysate was centrifuged at 7000 g for 20 min to remove cell debris and unbroken cells, and the supernatant was subjected to ultracentrifugation at 200,000 g for 30 min to pellet the cell membranes. The membrane pellet was homogenized and solubilized by gentle nutation for 2 hours at 4 °C in buffer consisting of 20 mM Tris-Cl (pH 8.0), 150 mM NaCl, 1% Dodecyl-β-D-Maltosid (w/v, DDM), and 20 µM DNQX, followed by centrifugation at 200,000 g for 30 min to remove insoluble material. The supernatant was incubated with Rho 1D4 affinity resin overnight at 4 °C. The protein-bound resin was extensively washed with 100 mL of buffer consisting of 20 mM Tris-Cl (pH 8.0), 500 mM NaCl, 1 mM DNQX, and 0.02% DDM at room temperature, and then slowly washed with 200 mL of buffer consisting of 20 mM Tris-Cl (pH 8.0), 300 mM NaCl, and 0.02% DDM at room temperature at a rate of approximately 0.3 ml/min. The protein was eluted in 50 mL of buffer with 20 mM Tris-Cl (pH 8.0), 300 mM NaCl, 0.02% DDM, and 0.25 mM 1D4 peptide. The eluted protein was concentrated and injected into a Superose 6 10/300 GL size-exclusion column equilibrated with buffer containing 20 mM Tris-Cl (pH 8.0), 150 mM NaCl, and 0.02% DDM. The peak fractions corresponding to tetrameric GluK2 were pooled and concentrated to 3.4 mg/mL for grid freezing.

Grid preparation and cryo-EM data

The cryo-EM grids were prepared as follows. UltrAuFoil R1.2/1.3 300 mesh gold grids were glow-discharged for 50 s before sample application. A final concentration of 1 mM glutamate was added to the sample, and then a droplet of 3 µL of protein sample was rapidly applied onto the grids. For the BPAM344/glutamate sample, proteins were pre-incubated with 500 µM BPAM344 for 30 min on ice, followed by the addition of 1 mM glutamate to the sample. 3.5 µl of the protein sample was applied to the grids, and the grids are frozen within 10 s. An FEI Vitrobot Mark IV (Thermo Fisher Scientific) was used to plunge-freeze the grids into liquid ethane after sample application at 4 °C and 100% humidity, with a blot time of 3 s and a blot force of −5.

All cryo-EM data were collected on a Titan Krios 300 kV microscope (Thermo Fisher Scientific) equipped with a GIF-Quantum energy filter (Gatan) with the slit set to 20 eV and a Gatan K3 Summit direct electron detection camera (Gatan) at Case Western Reserve University. Raw movies were collected using EPU in super-resolution mode, with a physical image pixel size of 1.06 Å and a defocus range of −0.8 to −1.6 µm. The total dose of around 50 e⁻/Ų was used for each movie, distributed across 50 frames for a total exposure time of ~3.2 s.

Cryo-EM image processing

For dataset a (GluK2 K676C/N802C supplemented with 1 mM glutamate and 500 µM BPAM344), the same motion correction procedure was applied using Patch Motion Correction in CryoSPARC4. CTF calculations were also carried out on the motion-corrected images using Patch CTF Estimation in CryoSPARC4. Micrographs unsuitable for further analysis were manually removed. Initial particle selection from 1000 micrographs was done with Blob Picker, followed by 2D classification to generate good templates. Template Picker was used to pick 903,654 particles, which were extracted with a pixel size of 2.12 Å. After several rounds of 2D classification, 612,540 particles were selected for initial 3D model generation via ab initio reconstruction. A total of 352,269 particles were selected and re-extracted with a pixel size of 1.06 Å following multiple rounds of heterogeneous refinement. NU-refinement for each class resulted in maps with resolutions of 6.16 Å, 5.80 Å, 4.14 Å, 4.00 Å, 4.00 Å, and 3.83 Å, respectively.

For dataset b (GluK2 K676C/N802C supplemented with 1 mM glutamate), the raw movie stacks were motion-corrected using Patch Motion Correction in CryoSPARC4. The Contrast Transfer Function (CTF) was calculated from the motion-corrected images using the Patch CTF Estimation in CryoSPARC4. Micrographs deemed unsuitable for further analysis were removed through manual inspection. Initial particle selection was performed using Blob Picker from 2000 micrographs, followed by 2D classification to generate good templates. Subsequently, 1,005,623 particles were picked using Template Picker and extracted with a pixel size of 2.12 Å. After several rounds of 2D classification, 765,893 particles were selected to generate the initial 3D model through ab initio reconstruction. A total of 277,369 particles were selected and re-extracted with a pixel size of 1.06 Å after multiple rounds of heterogeneous refinement. NU-refinement for each class yielded maps with resolutions of 3.92 Å, 4.35 Å, 3.85 Å, and 3.82 Å, respectively.

To improve the structure resolution for classes present in both datasets, particles from each class were combined and subjected to several rounds of heterogeneous refinement to remove junk particles. Separate NU-refinements for each class generated maps with resolutions of 3.67 Å, 4.02 Å, 3.81 Å, and 3.74 Å, respectively. For Class 2, focused masks around the ATD and LBD-TMD regions were created separately and used for local refinements of the corresponding regions. As a result, 3.42 Å reconstruction of the ATD region and 3.86 Å reconstruction of the LBD-TMD region were obtained through local refinement with focused masks. For the non-active state, particle subtraction was performed for both ATD-focused and LBD-TMD-focused heterogeneous refinement. After several rounds of heterogeneous refinement, 80,325 particles were used for subsequent NU-refinement, resulting in a 3.70 Å resolution of the ATD 3D reconstruction. A total of 117,714 particles were subjected to NU-refinement for the LBD-TMD region, generating 3.93 Å maps. The Gold-standard Fourier Shell Correlation (FSC) 0.143 criteria were used to estimate the map resolution for each class. For the shallow-desensitized state, focused masks around the ATD and LBD-TMD regions were created separately and used for local refinements of the corresponding regions. As a result, 3.42 Å reconstruction of the ATD region and 3.86 Å reconstruction of the LBD-TMD region were obtained through local refinement with focused masks. Composite maps were generated as previously described^93,146–148.

Calculation of structure motion and pore radius

The closure and opening of the LBD bi-lobe was quantified by measuring the D2 lobe’s shift relative to its position in the apo structure. This was determined in ChimeraX by aligning the two D1 lobes and utilizing the measure rotation command (https://www.rbvi.ucsf.edu/chimerax/docs/user/commands/measure.html). Pore sizes for each structure were calculated using the program HOLE, and the corresponding figures were generated in PyMOL.

Model building and refinement

The model building for the non-active class was initiated by individually fitting rigid-body components of the LBD extracted from the BPAM344-, and ConA-bound GluK2 structure (PDB code 9B36), and the TMD and ATD from the BPAM344 bound GluK2 structure (PDB code 8FWQ) into the cryo-EM map using Chimera. Subsequent manual adjustments, including fitting the backbone and side chains of the ATD-LBD and LBD-TM3 linkers into the density, as well as deleting unresolved side chains and making subtle local adjustments to minimize clashes, were performed in Coot.

For the shallow-desensitized class, the single LBD, entire ATD, and TMD were separately extracted from the BPAM344 bound GluK2 structure (PDB code 8FWQ) and manually docked into the corresponding cryo-EM map using Chimera. Further manual adjustments of the backbone and side chains were completed in Coot.

For intermediate and deep-desensitized classes, the single LBD, entire ATD, and TMD were extracted from the deep-desensitized GluK2 structure (PDB: 5KUF) and individually docked into the corresponding cryo-EM maps for each class using Chimera. Additional manual adjustments to the backbone and side chains were completed in Coot.

All four initial models were then refined in real space using Phenix against the corresponding cryo-EM maps. The models were validated through comprehensive validation procedures in Phenix. Glutamate ligands and BPAM344 compounds were extracted from the cryo-EM structure (PDB code 9B36) and manually fitted into the corresponding densities in Coot.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

C.Z., G.S.-C., and N.T. conceived the project and designed the experimental procedures. C.Z. designed the mutants/constructs and conduced the protein expression and purification, performed the EM data collection, data analysis, model building, and refinement. G.S.-C. conducted electrophysiological recordings. C.Z., G.S.-C., and N.T. wrote the manuscript. All authors reviewed the final draft.

Peer review

Peer review information

Nature Communications thanks the anonymous reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Data availability

The cryo-EM maps and coordinates for the GluK2 in the non-active state have been deposited in the Electron Microscopy Data Bank (EMDB) under accession numbers EMD-48765 (ATD), EMD-48766 (LBD-TMD), and EMD-48767 (full-length, composite map), and in the PDB under accession codes 9MZQ (ATD), 9MZR (LBD-TMD) and 9MZS (full-length), respectively. The cryo-EM maps and coordinates for the GluK2 in the shallow-desensitized state have been deposited in the EMDB under accession numbers EMD-48762 (ATD), EMD-48763 (LBD-TMD), and EMD-48761 (full-length, consensus map), and in the PDB under accession codes 9MZN (ATD), 9MZO (LBD-TMD), and 9MZM (full-length) respectively. The cryo-EM maps and coordinates for the GluK2 in the intermediate state, deep-desensitized state 1, deep-desensitized state 2, and deep-desensitized state3 have been deposited in the EMDB under accession numbers EMD-48760, EMD-48759, EMD-48758, and EMD-48757, and in the PDB under accession codes, 9MZL, 9MZK, 9MZJ, and 9MZI, respectively. Source data are provided with this paper.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-025-65920-8.