Protein expression and purification

The complementary DNA (cDNA) constructs of a rat GluA2flip(Q) isoform, tagged with a FLAG epitope near its C terminus39,51, and rat TARPγ2 (stargazin) were cloned into the DualTetON plasmid as described previously69,50 to generate a plasmid named DualTetON-A2iQFLAG, which doxycycline-dependently expresses both proteins simultaneously. The two proteins were coexpressed without using any tether. A stable TetON HEK cell line was generated by cotransfecting DualTetON-A2iQFLAG and a plasmid that confers hygromycin resistance, using established methods39,51,69. A clone was isolated in the presence of 30 µM NBQX and 120 µg ml−1 hygromycin. Clone3-#39 was chosen on the basis of its growth rate and the expression level of the complex and adapted to FreeStyle293 medium (Gibco, Thermo Fisher) in suspension.

Next, 1.2 L of a near-saturated suspension culture of clone3-#39 in FreeStyle293 medium supplemented with 30 µM NBQX and 1:500 diluted anticlumping agent (Gibco, Thermo Fisher, cat. no. 0010057DG) was used as a starting material. Cells were induced with 7.5 µg ml−1 doxycycline, 1 mM sodium butylate and 1% fetal calf serum (FCS) for 28 h as described39. The subsequent procedures were conducted on ice or at 4 °C. Cells were centrifuged at 931g for 10 min, washed with Dulbecco’s PBS once and centrifuged again; the pellet was flash-frozen in liquid nitrogen for storage at −80 °C. Approximately 10–12 ml of frozen pellets were resuspended in Resuspend buffer (25 mM Tris-HCl pH 8.0, 150 mM NaCl, 2 mM TCEP, 15 µM NBQX and protease inhibitors: 1 mM PMSF, 10 µg ml−1 aprotinin, 0.5 mM benzamidine, 1 µg ml−1 pepstatin A and 5 µg ml−1 leupeptin), making the final volume 90 ml. Then, 10 ml of 10× digitonin (25 mM Tris-HCl pH 8.0, 150 mM NaCl and 7.5% digitonin) was added and the mixture was nutated at 4 °C for 2.5 h to dissolve the membrane. The large debris was removed by low-speed centrifugation (3,000 r.p.m. for 10 min at 4 °C) and its supernatant was ultracentrifuged at 235,400g in a 45Ti rotor (Beckman) for 1 h. The resulting supernatant was incubated in a batch with 1 ml of FLAG M2 agarose beads (Sigma) for 2 h. The beads were collected by centrifugation at 58g for 5 min and transferred into an empty column. The beads were washed with four column volumes of wash buffer (0.03% glyco-diosgenin (GDN), 20 mM Tris-HCl pH 8.0 and 150 mM NaCl). The proteins were eluted using 6 ml of wash buffer containing 0.5 mg ml−1 FLAG peptide. The eluate was concentrated down to 0.55 ml using Ultrafree 100-kDa molecular weight cutoff (MWCO) ultrafiltration (Millipore). The concentrated sample was ultracentrifuged at 75,325g for 15 min and applied to a Superdex200 Increase column (GE Healthcare) equilibrated with GF buffer (0.03% GDN, 20 mM Tris-HCl pH 8.0 and 150 mM NaCl). The peak fractions were combined and concentrated down to 30 µl using Ultrafree 100-kDa MWCO ultrafiltration. Purity was checked by SDS–PAGE (Extended Data Fig. 6a). The final protein concentration was approximately 10 mg ml−1.

Grid preparation

The purified complex was split into two. The first half was used to obtain structures in acidic condition. Then, 4 µl of protein was mixed with 1 µl of 50 mM citric acid buffer (the 50 mM citric acid buffer was prepared by diluting 0.5 M citric acid–sodium citrate buffer at pH 4.0) immediately before applying the sample to the grid. The pH after mixing was measured using a pH-indicator strip to be pH ~5.0–5.5. The time from mixing to freezing was less than 30 s. Next, 2 µl of protein solution was applied to an UltraAuFoil R1.2/1.3 (300 mesh) and plunged into liquid ethane using Vitrobot Mark4 (Thermo Fisher). The freezing parameters were as follows: blot force, 12; blot time, 4.5 s; temperature, 4 °C; humidity, 100%; wait time, 10 s; and drain time, 0 s. Filter paper was doubled to facilitate blotting. To prevent aggregation in acidic conditions, it was critical to use UltraAuFoil and to reduce the time between acidifying and freezing. Optimal freezing conditions were determined by inspecting the grids using Glacios (Thermo Fisher). The second half was frozen directly to prepare the vitrified grid without any treatment at pH 8.0. We note that a Quantifoil R1.2/1.3 (300 mesh, Cu–C membrane) was used for pH 8.0, as the choice of Quantifoil over UltraAuFoil had no effect on the conformation of the NTDs at pH 8.0 (ref. 70) and the beam alignment was simpler to monitor during the EPU session with Quantifoil carbon membrane grids.

Cryo-EM imaging

All data were collected using a Titan Krios G4i (Thermo Fisher) equipped with a BioQuantumK3 detector at Vanderbilt University’s cryo-EM facility. Images were collected at 50 frames per video. The aberration-free image shift function was used in EPU (Thermo Fisher) semiautomated data collection software. The microscope was equipped with fringe-free optics, which enabled a smaller beam diameter for imaging. An objective aperture was not used. The detector dose rate was at 15.6–15.7 e− per pixel per s (measured over ice). The total dose was at 52.8–55.6 e− per Å2 (measured over vacuum). Each video contained 50 frames. The detailed parameters used for data collection in each sample are summarized in Table 2. Data collection was completed in a single EPU session for the sample at pH 8.0 (21,898 videos) but in two sessions (10,240 + 9,444 = 19,684 videos) for the sample at pH 5.5. Representative motion-corrected images are shown in Extended Data Fig. 6d–f.

Table 2 Cryo-EM data collection, refinement and validation statisticsImage processing of cryo-EM data

All image processing was performed using RELION 4 and 5 (refs. 71,72). Each raw video stack (50 frames) was motion-corrected (at 4 × 4 patches) and dose-weighted using MotionCor2 (ref. 73). CTFFIND4 was used to estimate the contrast transfer function (CTF) from non-dose-weighted images using 1,024 × 1,024 pixel tiles74. No symmetry was imposed throughout. Initial particles were identified using Autopick (pH 8.0, 7,593,346 particles; pH 5.5, 7,028,468 particles). Templates for Autopick containing 2D class averages of particles were centered at the gate region. Thus, using an optimal circular mask, the 2D and 3D classification was guided mostly from the signals in the LBD, TMD and TARPγ2. Before 2D classification, particles were extracted from a box size of 360 × 360 pixels and rescaled to 128 × 128 pixels. Parameters for 2D classification were the VDAM algorithm (variable-metric gradient descent algorithm with adaptive moments estimation) with 200 minibatches and regularization parameter T = 2. Mask circles were chosen at 180-Å diameter to purposefully cut off a portion of the NTDs, such that the alignment would be dominated by the signals in the LBD, TMD and TARPγ2. Particles belonging to 2D class averages with secondary-structure features of AMPAR were selected (pH 8.0, 1,886,582 particles; pH 5.5, 3,051,235 particles). The heterogeneity of the NTD layer was substantially different at the two pH values, even at the initial 2D class averages. The 2D class averages in the main figures with complete NTDs were produced by re-extracting the aligned particles by recentering them to ensure the entire architecture was contained in the circular mask.

Before 3D classification, particles were extracted from a box size of 360 × 360 pixels and rescaled to 180 × 180 pixels to optimize computational load. The 3D classification was performed for 40 iterations at T = 4 without a mask and using EMD-29386 (GluA2flip(Q) in complex with TARPγ2(KKEE)) as the initial model50. We also conducted 3D classification using masks that incorporated conformational heterogeneity of the NTDs but the outcome was not substantially different, which confirms that the alignment at this stage was guided mainly from the signals in the LBD, TMD and TARPγ2. Four and six classes were specified for pH 8.0 and pH 5.5, respectively. Well-defined classes with clear features of transmembrane helices of GluA2 and TARPγ2 were selected (class 1 and 3 for pH 8.0; class 6 for pH 5.5). The particles in two classes were combined in pH 8.0. The numbers of particles selected after 3D classification were 1,108,462 particles (pH 8.0) and 813,615 particles (pH 5.5). Particles were then re-extracted from a box size of 360 × 360 pixels without binning. Further 3D refinement (Refine3D) was performed using a mask that covered the LBD, TMD and TARPγ2 (LBD–TMD mask). The 3D refinement at this stage was performed using RELION 4 with the ‘--external_reconstruct’ flag and SIDESPLITTER75. The 3D refinement was followed by postprocessing, which produced maps of around 2.8-Å overall resolution in both pH conditions. The maps were further improved by CTF refinement, followed by another iteration of Refine3D and postprocessing, which produced a consensus map of the LBD, TMD and TARPγ2 at an overall resolution of 2.8 Å (Table 2 and Extended Data Fig. 6e). Focused refinement of the NTD at pH 8.0 was conducted by first recentering the NTD and then refining using an NTD mask (the consensus NTD map is deposited as an associated map of EMD-44232). In the consensus reconstruction, the LBD adopted greater heterogeneity in the acidic conditions, noticeable as ill-defined LBDs. The local resolution, calculated by ResMap76, of the LBD in the consensus map of the D1 lobe that is closer to the NTD was much lower at pH 5.5 (Extended Data Fig. 6f). The alignment was guided toward improving the resolution of the membrane-embedded region at the cost of degrading the alignment of the LBD because the former contains many bundles of α-helices that generate strong signals.

To resolve the heterogeneity of the LBD and the NTD, the particles from the consensus alignment were re-extracted from a box size of 360 × 360 pixels and rescaled to 128 × 128 pixels, preserving the alignment parameters, and refined using the LBD–TMD mask with a local search. For each pH condition, a mask that covered the NTD and LBD was generated. To generate the mask, one round of 3D classification without alignment was conducted without a mask to sample the conformational heterogeneity of the NTDs at each pH. Representative classes that defined the range of heterogeneity were added in Chimera to guide mask production. Next, the signals outside the NTD and LBD were subtracted using the above mask and the particles were classified into 20 and 40 classes without alignment using regularization T = 4 for 40 iterations (representative classes are shown in Extended Data Fig. 7b,c). The extent of heterogeneity was low at pH 8.0 and, thus, classification into 20 classes was sufficient to sample the entire range of heterogeneity because similar conformations were present among classes. In contrast, classification into 40 classes produced a variety of splayed NTD conformations with unique NTD 3D arrangements. At pH 8.0, classes 1 (containing 47,782 particles) and 12 (containing 48,399 particles) were chosen as representative classes for further refinement of the full complexes. At pH 5.5, class 23 (containing 29,945 particles; Fig. 5c) was the only class that contained solid signals of both NTD dimers and was, thus, subjected to further refinement of the full complex. The reason for the conformational stability in class 23 is because of both NTDs approaching the LBDs, possibly making weak contacts. We note that, because of extensive conformational heterogeneity, one of the two NTD dimers was always weaker in other classes at pH 5.5, which prevented 3D reconstruction of the full complex. To produce the full map of above classes, focused refinement was conducted using the LBD–TMD mask to obtain the maps containing the LBD, TMD and TARPγ2, whereas the particles were recentered to the NTD layer by shifting the center by 64.8 Å in the z direction and refined using the NTD mask to obtain the NTD maps. The NTD layer in class 23 at pH 5.5 still contained conformational heterogeneity that prevented high-resolution 3D reconstruction, which resulted in an overall resolution of 5.9 Å. All other maps of subclassified conformations were refined to a final resolution ranging from 3.4 to 3.7 Å (Table 2).

To further understand the LBD–TMD conformations relative to the NTD conformations, the LBD, TMD and TARPγ2 portions of various classes were investigated by focused refinement using the LBD–TMD mask. In addition to class 23 introduced above, classes 8, 9, 16, 18, 19, 20, 29, 31, 37 and 40 were chosen as representative splayed NTD conformations at pH 5.5. Similarly, in addition to classes 1 and 12 introduced above, classes 4, 6, 8 and 15 were chosen as representative compact NTD conformations at pH 8.0. For each pH condition, a mask that covered the NTD and LBD was generated. Particles in each class were re-extracted from a box size of 360 × 360 pixels and rescaled to 180 × 180 pixels. Each class was subjected to focused refinement using the LBD–TMD mask. Refine3D and postprocessing produced maps at overall resolution ranging from 3.7 to 4.0 Å at pH 5.5 and from 3.4 to 3.9 Å at pH 8.0. The differences between the LBD conformations were characterized as translation and rotation between the two LBD dimers, which were small conformational differences in the organization of the LBDs in the gating ring55. The overall resolutions of the maps were estimated using a Fourier shell correlation (FSC) = 0.143 cutoff in RELION77. The image processing and model statistics are summarized in Table 2. Angular distributions of assigned angles were inspected to ensure the coverage of the Fourier space. Visual inspection of the map showed no signs of artifacts.

Model building and refinement

The model building and refinement for the consensus maps, maps of class 1 and 12 at pH 8.0 and the map of class 23 at pH 5.5 (Fig. 5c) were conducted as follows: the reference models, PDB 8FPG (TMD and TARPγ2) and PDB 8FPK (LBD), were rigid-body fit into the EM density map using Chimera78. The fit was further adjusted using the jiggle fit function in Coot79. Further manual adjustment with the real-space refine zone function in Coot was used to generate an atomic model. The generated model was further refined using the real_space_refine tool in Phenix80. Real-space refinement was conducted by imposing secondary-structure restraints by annotating helices and sheets in the PDB file. To prevent overfitting of the models into the density, refinement was run for five cycles with strict geometric restraints of 0.005–0.01 for bond length and 0.5–1 for bond angle. MolProbity and Mtriage were used for validation. To interpret the full architecture, the map produced by Refine3D, the unmasked and unsharpened map, was used to position the NTD map using rigid-body fit in Chimera. The maps of classes 4, 6, 8 and 15 at pH 8.0 and classes 8, 9, 16, 18, 19, 20, 29, 31, 37 and 40 at pH 5.5 were interpreted as follows: The maps were thresholded at an optimal level to visualize the densities of the NTDs. The NTD atomic model from class 12 at pH 8.0 was rigid-body fit into the map using Chimera. No further model refinement was conducted for the NTDs. The atomic model of the consensus map was rigid-body fit into the LBD, TMD and TARPγ2 densities using Chimera. The model was subjected to jiggle fit and all-atom refinement in Coot with the Geman–McClure self-restraint at 4.2 Å (refs. 27,81). PyMOL (Schrödinger) and Chimera were used to further analyze the structure and generate figures.

Size-exclusion chromatography with MALS

The molecular mass of NTDA2 was determined in solution using size-exclusion chromatography (SEC) with MALS (SEC–MALS). Measurements were performed using a Wyatt Heleos II 18 angle light scattering instrument coupled to a Wyatt Optilab rEX online refractive index detector. Samples of 100 µl were resolved in 10 mM HEPES and 150 mM KCl (pH 7.5) buffer on a Superdex S200 10/300 analytical gel filtration column coupled to an Agilent 1200 series liquid chromatography system running at 0.5 ml min−1 before then passing through the light scattering and refractive index detectors in a standard SEC–MALS format. Protein concentration was determined from the excess differential refractive index based on 0.186 ∆RI for 1 g ml−1. The measured protein concentration and scattering intensity were used to calculate the molecular mass from the intercept of a Debye plot using Zimm’s model as implemented in the Wyatt ASTRA software.

The experimental setup was verified using a BSA standard run of the same sample volume. The monomer peak was used to check mass determination and to evaluate interdetector delay volumes and band-broadening parameters that were subsequently applied during the analysis of NTDA2.

DNA constructs and culture for electrophysiology

Sequences for rat GluA2 and rat GluA1 were flip variants; GluA2 was unedited at the 586Q/R site and edited at the 743R/G site. All cDNA constructs used for transfection were generated using in vivo assembly cloning as previously described82. Constructs were cloned in either pRK5 or IRES vectors. GluA2delNTD was made by deleting the NTD sequence from position 1 to 394 (mature peptide). GluA1NTDA2 was made by replacing the NTD from A1 (mature peptide residues from 1 to 390) with the residues from A2. Similarly, for GluA2NTDA1, we replaced the NTD of GluA2 (mature peptide, residues 1–394) with the corresponding sequence from A1.

HEK293T cells (American Type Culture Collection, cat. no. CRL-11268, RRID: CVCL_1926, lot 58483269; identity authenticated by short tandem repeat analysis, Mycoplasma negative), cultured at 37 °C and 5% CO2 in DMEM (Gibco; high-glucose, GlutaMAX, pyruvate, cat. no. 10569010) supplemented with 10% FBS (Gibco) and penicillin–streptomycin, were transfected using Effectene (Qiagen) according to the manufacturer’s protocol. Transfected cells were identified by cotransfection of a pN1-EGFP plasmid or by EGFP–mCherry coexpressed from the pIRES2 plasmid. Where cotransfected, the DNA ratio of AMPAR to TARPγ2 was 1:2. To avoid AMPAR-mediated toxicity, 30 μM NBQX (Tocris or HelloBio) was added to the medium immediately after transfection.

Electrophysiology

Recording pipettes were pulled with a P1000 horizontal puller (Sutter Instruments) using borosilicate glass electrodes (1.5 mm outside diameter, 0.86 mm inside diameter; Science Products). Electrode tips were heat-polished with an MF-830 microforge (Narishige) to final resistances of 2–4 MΩ (whole cell) and 6–10 MΩ (outside-out patches). Electrodes were filled with an internal solution containing (in mM) CsF (120), CsCl (10), EGTA (10), HEPES (10), Na2-ATP (2) and spermine (0.1), adjusted to pH 7.3 with CsOH. The extracellular solution contained (in mM) NaCl (145), KCl (3), CaCl2 (2), MgCl2 (1), glucose (10) and HEPES (10), adjusted to pH 7.4 using NaOH. We used 1 M HCl to adjust the pH to 5.5 and 6.4 for the recordings in acidic conditions. Recordings were performed at room temperature (~21–23 °C). Currents were recorded with an Axopatch 700B amplifier (Molecular Devices) 24–48 h after transfection. Signals were prefiltered at 10 kHz with a four-pole Bessel filter, sampled at 100 kHz with the Digidata 1550B (Molecular Devices), stored on a computer hard drive and analyzed using pClamp 10 (Molecular Devices), Excel and GraphPrism software.

On the day of recording, cells were plated on poly(l-lysine)-treated glass coverslips. Fast perfusion experiments were performed with a two-barrel theta tube glass with a diameter of approximately 250 µm. The theta tube was mounted on a piezoelectric translator (Physik Instrumente) and the command voltage (9 V) was filtered with a 500-Hz Bessel filter to reduce mechanical oscillations. The theta tube was filled with pressure-driven solutions (ALA Scientific Instruments). The speed of solution exchange at the theta tube interface was measured as 20–80% of the rise time of the current generated with 50% diluted extracellular solution and was on average about 120 µs (outside-out patches) or 400 µs (whole cell). Patches were voltage-clamped at −60 mV (voltage not corrected for junction potential of 8.5 mV). Series resistance was not corrected for outside-out recordings. For whole-cell recordings, series resistance was never higher than 8 MΩ and was compensated by 90%.

Recovery from desensitization was measured with a two-pulse protocol. A conditioning pulse of 10 mM glutamate with a duration of 200 ms or 100 ms (Fig. 1e) was followed by 15-ms glutamate pulses delivered at intervals increasing by 5 or 10 ms (GluA2 constructs) or 20 ms (GluA1 wt and GluA1NTD_A2 constructs). Desensitization (200-ms glutamate pulses) and deactivation (1-ms glutamate pulses) time constants were obtained by fitting the current decay (Chebyshev algorithm, built-in Clampfit 10.2; Molecular Devices) of the glutamate application from 90% of the peak to the steady-state or baseline current with one or two exponentials. Where biexponential fits were used, the weighted τdes is reported, calculated as follows: τw,des = τf(Af/(Af+ As)) + τs(As/(Af + As)), where τf/s and Af/s represent the fast/slow component time constant and coefficient, respectively. The rise time constant was obtained by fitting the current rising phase (from 1-ms glutamate application) with one exponential from 20% to the peak.

Recovery from desensitization was fitted by a Hodgkin–Huxley-type equation:

$$f\left(t\right)=_+(\;_-_)\times (1-\exp (-))^$$

where y0 and ymax are the minimum and maximum, k is the rate constant, t is the interpulse interval and m is the slope. GluA2 receptors have a steeper recovery profile; therefore, we fixed the slope to 2 (ref. 83). Recovery profiles of GluA1 are much slower than GluA2; therefore, the recovery time constant for GluA1 receptors was obtained with m = 1, which gives a single exponential function.

The dose–response relationship of GluA2 + TARPγ2 at pH 7.4 and 5.5 was measured from whole-cell currents at a holding voltage of −40 mV. Six concentrations of glutamate were applied with a theta tube to a lifted whole cell to obtain the dose–response relationship.

The dose–response relationship for each cell was fitted with GraphPrism software using the Hill equation:

$$I=\frac_^}}}}^}}}+}}}_^}}}}$$

where Imax is the maximum response, EC50 is the concentration of glutamate that gave half of the maximum response and nH is the Hill coefficient.

For the presentation, dose–response relationships from each cell were normalized to the response of 10 mM glutamate and pooled together (Extended Data Fig. 3d).

Nonstationary fluctuation analysis (NSFA) was performed on the desensitizing current phase of macroscopic currents evoked with glutamate pulses (10 mM, 200 ms) from outside-out patches containing GluA2 + TARPγ2. The same patch was exposed to pH 7.4 and 5.5 and at least 30–100 successive responses were collected for each condition from the same patch. The mean current and variance from successive responses were calculated in Clampfit and imported to a custom-written Python script, where the variance σ2 was grouped in ten amplitude bins, plotted against the mean current and fitted with a parabolic function38:

where i is the single-channel current, \(\bar\) is the mean current, N is the number of channels and \(_^\) is the background variance. The weighted mean single-channel conductance γ was obtained from the single-channel current and the holding potential (−60 mV, not corrected for the liquid junction potential).

Data visualization and statistical analysis were performed using GraphPad Prism.

AlphaFold and energetic modeling

Predicted structural models of the homomeric and heteromeric GluA1 NTD (UniProt P19490; residues 19–400) and GluA2 NTD (UniProt P19491; residues 25–398) were generated using AlphaFold2-Multimer47 through ColabFold84. The highest-ranked predictions were all validated against predefined established criteria (PAE, pTM, pLDDT, DockQ, MolProbity and QS-score). The Dynamut2 server48 was used to investigate missense substitutions on the GluA2 BD NTD interface using a model from the previously published GluA2/A1 complex52.

Dissociated hippocampal cultures

All procedures were carried out under PPL 70/8135 in accordance with UK Home Office regulations. Experiments were licensed under the UK Animals (Scientific Procedures) Act of 1986 following local ethical approval. All animals were housed with food and water ad libitum on a 12-h light–dark cycle at room temperature (20–22 °C) and 45–65% humidity.

Cultures were prepared according to the protocol described in Beaudoin et al.85. Hippocampi from postnatal P0–P1 C57BL/6JOla wt mice were dissected in ice-cold HBSS (Ca2+ and Mg2+ free; Gibco, cat. no. 14175095) containing 0.11 mg ml−1 sodium pyruvate (Gibco, cat. no. 12539059), 0.1% glucose and 10 mM HEPES (Gibco, cat. no. 15630056) and dissociated for 20 min at 37 °C with trypsin (0.25% w/v; Gibco, cat. no. 15090-046). Neurons were plated onto glass coverslips (24 -mm round coverslips 1.5; Glaswarenfabrik Karl Hecht, cat. no. 1001/24_15 92100105080) coated with poly(l-lysine) (0.1 mg ml−1; P2636, Sigma-Aldrich) following resuspension in equilibrated plating medium containing 86.55% MEM (Gibco, cat. no. 21090022), 10% heat-inactivated FBS (Gibco, cat. no. 11573397), 0.45% glucose, 1 mM sodium pyruvate and 2 mM GlutaMax (Gibco, cat. no. 35050038). Cultures were kept at 37 °C and 5% CO2 in equilibrated maintenance medium containing 96% Neurobasal plus medium (Thermo Fisher, cat. no. A3582901), 1× B-27 plus supplement (Thermo Fisher, cat. no. A3582801) and 2 mM GlutaMax. Half of the medium was replaced every 3–5 days.

FRAP

Dissociated hippocampal neurons were made to express SEP-tagged AMPAR constructs using either Lipofectamine 3000 (Thermo Fisher) at 11 days in vitro and imaged at 14 days in vitro. SEP-tagged GluA2 (SEP–GluA2) was created by inserting the fluorescent protein-coding region between the third and fourth residues of the mature GluA2 protein. In addition, a SEP–GluA2 F231A mutant was generated. In all constructs, the SEP tag was preceded and followed by an A-S dipeptide linker. The SEP sequence was kindly provided by J. Hanley.

At 14–15 days in vitro, neurons were imaged in artificial cerebrospinal fluid containing (in mM) NaCl (150), KCl (2.5), MgCl2 (2), CaCl2 (2), HEPES (20) and glucose (10) at pH adjusted to 7.4 or 5.5 with either NaOH or HCl in a heated chamber at 37 °C. Images were acquired on a Zeiss 780 laser scanning confocal microscope using a ×40 (1.2 numerical aperture) water-immersion objective with a pixel size of 100 nm. Photobleaching was achieved by repeated xy scanning of the region of interest (2 μm2) at high laser intensity, using excitation at 405 nm for Fig. 6c. The imaging protocol consisted of 3 images and 20 images taken before and after bleaching, respectively, at 30-s intervals. Analysis was performed using EasyFRAP-web86. Photobleaching because of image acquisition was corrected by normalization to the fluorescence of the distant nonphotobleached spine (2 μm2) and to the background fluorescence. Normalized data were further postprocessed and fitted to a single exponential curve using GraphPad Prism.

MDGluA2 NTD MD simulations

We used the highest-resolution GluA2 homodimeric NTD crystal structure obtained at pH 8.0 (PDB 3H5V; 2.33 Å)42 as the starting model for NTD-only constant-pH simulations. To prepare the physiological homotetramer (dimer of dimers) from the deposited trimeric asymmetric unit (ASU), a copy of the unit was rotated 180° around the vertical axis (perpendicular to the membrane in the full-length GluA2 receptor), superimposed onto the unrotated ASU with Chimera MatchMaker78 and had repeated domains removed. The created NTD homotetramer was validated against NTD domains of full-length GluA2 crystal structures with a QMEANDisCo score (SWISS-MODEL QMEAN webserver, version 3.1.0)87,88 of 0.87 ± 0.05 (score range 0–1; scores > 0.6 indicate good agreement with experimental structures).

Fixed protonation states were assigned to all titratable residues with the Henderson–Hasselbalch equation on the basis of the pH being simulated and pKa calculations using PROPKA3 (version 3.4.0)89,90; pKa values were calculated on NTD homotetrameric (created as above) crystal structures obtained at pH 8.0 (PDB 3H5V)42 and pH 4.8 (PDB 3HSY)18 for basic and acidic simulations, respectively. By assigning acidic protonation states to the basic pH starting model, we simulated an instantaneous change in pH at the start of the acidic simulations. The NTD tetramer was placed at the center of a 175 × 175 × 175 Å3 box with periodic boundary (PB) conditions (buffer distance between the protein and PB set to 1.0 nm), solvated with simple point-charge water91 and charge-neutralized with sodium and chloride ions. This resulted in a system with 544,812 atoms.

All simulations were run in GROMACS (version 2019.3)92,93,94. The system was first energy-minimized over 10,000 steps with a step size of 0.01, followed by three 1-ns equilibrations with 2-fs steps: temperature (NVT) equilibration to 300 K with the v-rescale thermostat, first pressure (NPT) equilibration to 1 bar with the Berendsen barostat and second NPT equilibration to 1 bar with the isotropic Parrinello–Rahman barostat for greater accuracy; temperature was controlled for the protein and solvent groups separately. The protein movement was fully constrained during the first two equilibrations to not destabilize the system. Finally, the system was simulated with the v-rescale thermostat and Parrinello–Rahman barostat for 100 ns. All simulations were run with the Verlet cutoff scheme, LINCS H-bond constraints, a 1.2-nm van der Waals cutoff and particle mesh Ewald electrostatics.

Reporting summary

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