Model building, refinement and validation

The nominally higher-resolution conformation B model was built first, focusing initially on the well-ordered ECD. A homology model for each subunit was made using the crystal structure of the β3 homopentameric GABA-A receptor (RCSB: 4COF) via Swiss-Model⁶⁰. From side chain and glycosylation features in the ECDs we were able to unambiguously assign the α1, β2 and γ2 subunits; rigid docking of the homology models into these densities supported the assignments. Chain IDs in the models are: A, β2; B, α1; C, β2; D, α1; E, γ2. Density for Fab fragments was observed extending from the ECD of α1 subunits roughly parallel to the membrane plane. Swiss-Model was used to generate a homology model of the Fab light chain using PDB entry 1UYW and of the heavy chain using PDB entry 4WEB and these chains were docked into the EM density at one Fab site using Chimera⁶¹. Manual adjustments of the receptor-Fab structure were then performed in Coot^62,63. The ECD and TMD halves of each subunit and the variable half of each Fab were rigid body fitted into the density map. The variable domain of the Fab was rebuilt into unambiguous density; the density associated with the Fab constant domain was too disordered for both Fab copies to allow building of an atomic model; as this portion of the Fab was not of biological interest, the final model includes just the Fab variable domains. Once this first Fab copy was rebuilt, it was copied into the additional sites on the second α1 subunit and manually adjusted. Well-ordered N-linked glycans were built in the ECD channel vestibule and shorter N-linked chains along the outer surface of the ECD. Strong density for ligands was observed at β-α interfaces (modeled as GABA) and at the α-γ interface (flumazenil). The extended conformation of flumazenil is consistent with its crystal structure⁶⁴. No unaccounted-for density was observed at the α-β or γ-β interfaces in either conformation. The ECD of conformation B was docked into the map for conformation A; we noted no meaningful conformational differences except in the loops contacting the TMD.

In conformation A, the TMD portions of the α1 subunits (chains B, D) are well ordered and there was no ambiguity regarding register. The β2 subunits are less well-ordered but we are still confident about register as the backbone adopts internally consistent conformations. In the γ subunit, which undergoes a conformational rearrangement to fill the pore with its M2 helix, the linker connecting β10 to M1 and the M2-M3 loop are disordered. Nonetheless, the 4-helix bundle from this subunit holds together in a conformation akin to the well-ordered subunits, and to that observed in the β3 homopentameric structure, and thus we are confident about the amino acid register in the γ subunit TMD in conformation A.

In conformation B, the TMD is comparatively less well-ordered than in conformation A. The α1 subunits remain, as in conformation A, well ordered, with clear side chain density. The β2 subunit at the chain A position is well-ordered, however the chain C β2 subunit, which packs opposite the pore from the γ2 subunit, is not well ordered. Nonetheless, its conformation in less sharpened maps was clear enough to dock the 4-helix bundle in a conformation similar to that observed in chain A. The γ2 subunit in conformation B is comparatively disordered and its modeling is problematic as the helix bundle is not held together in a familiar arrangement. We modeled the M1 and M2 helices with amino acids placed tentatively based on side chain density. The M3 and M4 helices were built as poly-alanine chains. Due to a lack of strong interaction of the γ2 subunit in this conformation with its ECD half, the tip of the γ2 subunit Cys-loop did not have clear density and we omitted 3 residues from this region. Otherwise, the ECD modeling is continuous from the first amino acid to the end of β10. In the TMD, we modeled many strong oblong features as CHS; these occupy distinct sites between the two receptor conformations. After manual building in Coot, global real space coordinate and B factor refinement with NCS restraints were performed in Phenix⁶⁵. The refined model quality was assessed using Molprobity (Extended Data Fig. 5). The following segments of the receptor were not modeled due to weak density features in the corresponding regions: in Conformation A, β2 (chain A and C): N-7,341-C; α1 (chain B): N-9, 346-C; α1 (chain D): N-9, 348-C; γ2 (chain E): N-24, 233-236, 287-291, 356-C; in Conformation B, β2 (chain A and C): N-7, 341-C;(chain B): N-12, 346-C; α1 (chain D): N-10, 348-C; γ2 (chain E): N-24,158-160, 288-296, 319-326, 347-C.

The validation to test for overfitting of the model was performed as previously described⁶⁶. Briefly, the atom positions of the final refined models were randomly displaced by a maximum of 0.5 Å using PDBSET in the CCP4 suite⁶⁷. This perturbed model was then refined in Phenix in real space against the first half map of the reconstruction comprising 50% of the particles. A map vs. model FSC comparison was made for this model vs. the map used in its refinement (“work”), as well as the same model vs. the half map not used in refinement (“free”). The FSC curves of work and free half maps vs. model agree well (Extended Data Fig. 5).

Schematic interaction analysis of GABA and flumazenil was performed by Ligplot⁺⁶⁸. Subunit interfaces were analyzed using the PDBePISA server⁶⁹. Structural biology software packages were compiled by SBGrid⁷⁰.

For display settings in figures, density maps for the ECD were sharpened as shown in Extended Data Fig. 5a with a B factor −186 Ų for conformation A and −153 Ų for conformation B. Density maps displayed for TMD were sharpened with a B factor of −100 Ų for both conformations. In Figs. 2 and ​and3,3, density maps for GABA and flumazenil were rendered in Chimera at threshold levels of 0.018 and 0.04, respectively. In Extended Data Fig. 6, density maps were displayed at the following threshold levels: a-h, 0.024; i-k, 0.0158; l, 0.06; m and n, 0.03. In Extended Data Fig. 7, density maps were displayed at the following threshold levels: a-h, 0.024; i-k, 0.015; l, 0.05; m and n, 0.027. In Figs. 1, ​,55 and Extended Data Fig. 9, density maps were rendered in Chimera at a threshold level of 0.024. In Extended Data Fig. 10a-d, density maps were rendered in Chimera at a threshold level of 0.02.