A Mechanically Weak Extracellular Membrane-Adjacent Domain Induces Dimerization of Protocadherin-15

Cloning, expression, and purification of bacterially expressed PCDH15 EC10-MAD12

Pig and mouse (Mus musculus [mm]) PCDH15 including EC10, EC11, and MAD12 (ss and mm PCDH15 EC10-MAD12) (Tables S1 and S2) were subcloned into NdeI and XhoI sites of the pET21a vector. These constructs were used for expression in Rosetta (DE3) competent cells (Novagen, Madison, WI) cultured in terrific broth at 37°C and induced at OD₆₀₀ = 0.6 with 1 mM of isopropyl β-D-1-thiogalactopyranoside at 30°C for ∼16 h. Cells were lysed by sonication in denaturing buffer (20 mM TrisHCl [pH 7.5], 6 M guanidine hydrochloride, 10 mM CaCl₂, and 20 mM imidazole). The cleared lysates were incubated with Ni-Sepharose (GE Healthcare, Chicago, IL) for 1 h and eluted with denaturing buffer supplemented with 500 mM imidazole. Both pig and mouse PCDH15 EC10-MAD12 proteins were refolded at 4°C by fast or drop-wise dilution of 20 mL of eluted denatured protein (1–2 mg/mL) into 480 mL of refolding buffer (20 mM TrisHCl [pH 8.0], 150 mM KCl, 5 mM CaCl₂, 400 mM L-arginine, 2 mM dithiothreitol or 1 mM tris(2-carboxyethyl)phosphine (TCEP), and 10% glycerol). For fast dilution refolding, the denatured protein was dispensed into the refolding buffer over a 1-min period with vigorous stirring. For drop-wise refolding, the denatured protein was dispensed over a 20-min period with gentle stirring. In both cases, the protein solution was kept under constant and gentle stirring after mixing. All protein solutions were concentrated to 10 mL using Amicon Ultra-15 filters at 3000 revolutions per minute (rpm) with mixing every 20 min. Refolded protein was further purified on a Superdex-200 column (GE Healthcare) in holding buffer containing 20 mM TrisHCl [pH 8.0], 150 mM KCl, 5 mM CaCl_2, and 1 mM TCEP. Purity of the recombinant protein was analyzed by sodium dodecyl sulfate-polyacrylamide gel electrophoresis, after which the protein-containing fractions were pooled and used for further experiments. Pure proteins were concentrated by centrifuge ultrafiltration (Vivaspin-6 10 kDa) to 3–5 mg/mL for crystallization trials and biochemical assays. Protein concentrations were determined using the NanoDrop (Thermo Fisher Scientific, Waltham, MA).

Cloning, expression, and purification of mammalian-expressed mm^∗ PCDH15 EC10-MAD12

Use of mouse mammalian expressed protein is indicated by mm^∗. Suspension Expi293F cells (Thermo Fisher Scientific) were cultured in Expi293 medium at 37°C, 8% CO₂, and incubated on an orbital shaker at 125 rpm. The coding sequence for mm PCDH15 EC10-MAD12 was inserted into a pHis-N1 expression vector (modified version of the pEGFP-N1 vector from Clontech [Mountain View, CA] where the EGFP has been substituted for a hexahistidine tag; James Jontes, The Ohio State University) using XhoI and KpnI restriction sites. The native signal sequence was included before EC10. For a typical 30 mL expression, cells were prepared at a density of 2.9 × 10⁶ cells/mL and exchanged into 25.5 mL of fresh media. The DNA complexes were prepared by diluting 30 μg of DNA up to 1.5 mL with Opti-MEM and diluting 81 μL of ExpiFectamine 293 up to 1.5 mL. After 5 min, the two solutions were combined, gently mixed, and incubated for an additional 20 min at room temperature. The complexes were then added to the cells and incubated for 20 h. The next day, 150 μL of ExpiFectamine 293 Transfection Enhancer 1 and 1.5 mL of ExpiFectamine 293 Transfection Enhancer 2 were added. Cells were grown for 4 days, and the conditioned media (CM) was collected. The CM was dialyzed overnight against 20 mM TrisHCl [pH 7.5], 150 mM KCl, 50 mM NaCl, and 10 mM CaCl₂ to remove EDTA. The CM was concentrated using Amicon Ultra-15 10 or 30 kD concentrators and incubated with Ni-Sepharose beads for 1 h. The beads were washed three times with 20 mM TrisHCl [pH 8.0], 300 mM NaCl, 10 mM CaCl₂, and 20 mM imidazole, and the target protein was eluted with the same buffer containing 500 mM imidazole. The protein was further purified on a Superdex-200 column (GE Healthcare) in 20 mM TrisHCl [pH 7.5], 150 mM KCl, 50 mM NaCl, and 2 mM CaCl₂. The protein was concentrated and the final concentration was determined by measuring absorbance using the NanoDrop (Thermo Fisher Scientific).

Crystallization and micro-seeding

Initial needle-like crystals of ss PCDH15 EC10-MAD12 were grown by vapor diffusion at 4°C by mixing equal volumes (0.6 μL) of protein (purified without TCEP) and reservoir solutions (0.1 M HEPES [pH 7.5], 25% w/v PEG 2000 MME) and by equilibrating against 75 μL of reservoir solution in a sitting drop setup. A seed stock was prepared by mixing of crystal-containing solution (2 μL) with reservoir solution (50 μL), followed by vortexing with microseed beads (MD2-14; Molecular Dimensions, Newmarket, UK). The resulting solution was diluted to prepare a seed stock and used for random micro-seeding (57). Crystallization droplets were composed of equal volumes (1 μL) of protein solution (repurified without TCEP; 4.7 mg/mL) and reservoir solution along with 0.5 μL of the seed stock. The droplets were equilibrated against 75 μL of reservoir solution in a sitting-drop setup at 4°C. Final hits were obtained in 0.1 M HEPES [pH 7.5], and 20% v/v Jeffamine M-600.

Data collection and structure determination

Crystals were cryoprotected in reservoir solution plus 25% v/v PEG400. All crystals were cryo-cooled in liquid N₂. X-ray diffraction data were collected as indicated in Table 1 and processed with HKL2000 (58). The final structure was determined by molecular replacement using PHASER (59) and Buccaneer (60). Initial search models were based on the structure of Homo sapiens (hs) PCDH15 EC10 (Protein Data Bank [PDB]: 4XHZ) (45) and a separate homology model for the mm PCDH15 EC11 repeat obtained using SWISS-MODEL (61) and the PDB: 5DZX (32) as a template. MAD12 was built de novo and refined by using COOT (62). REFMAC5 (63) was used for restrained translation-libration-screw-rotation refinement. Data collection and refinement statistics are provided in Table 1. Coordinates for ss PCDH15 EC10-MAD12 have been deposited in the PDB: 6BXZ.

Analytical ultracentrifugation

Sedimentation velocity analytical ultracentrifugation (AUC) experiments were performed in a ProteomeLab XL-I AUC (Beckman Coulter, Brea, CA) following standard procedures (64, 65, 66). Briefly, size-exclusion chromatography (SEC) purified protein samples were loaded into AUC cell assemblies with Epon centerpieces and 12 mm path length. To achieve chemical and thermal equilibrium, the An-50 TI rotor with loaded samples was allowed to equilibrate for ∼2 h at 20°C in the centrifuge. The rotor was spun at 50,000 rpm and data were collected using absorption optics. Data analysis was performed with the software SEDFIT (http://sedfitsedphat.nibib.nih.gov), using a continuous sedimentation coefficient distribution model c(S). Standard values for buffer viscosity (0.01002 poise), density (1 g/mL), and partial specific volume (0.73 mL/g) were used, and confidence level was set to 0.68 as routinely done. The obtained c(S) distribution was loaded in GUSSI (67). AUC experiments were done with at least two biological replicates, which were also accompanied by at least one duplicate to ensure data were obtained accurately.

SEC-multiangle laser light scattering

SEC-multiangle laser light scattering (MALLS) data were collected by using an ÄKTAmicro system connected in series with a Wyatt miniDAWN TREOS system. The protein samples were separated on a Superdex-S200 3.2/30 column (GE Healthcare), and both the absorbance at 280 nm and the light scattering were monitored. The scattering information was subsequently converted into molecular weight using a rodlike model. SEC-MALLS experiments were done with at least two biological replicates and accompanied by at least one duplicate.

SEC–small-angle x-ray scattering

Small-angle x-ray scattering (SAXS) data were collected at SIBYLS beamline 12.3.1 in the Advanced Light Source (ALS) (Berkeley, CA) as described (68, 69) (Table S3). Purified samples of ss PCDH15 EC10-MAD12 (6.3 mg/mL) and mm PCDH15 EC10-MAD12 (5.3 mg/mL) were analyzed. Data were collected using an Agilent 1260 series HPLC with a Shodex KW-803 analytical column at a flow rate of 0.5 mL/min using 20 mM TrisHCl [pH 8.0], 150 mM KCl, 5 mM CaCl_2, and 1 mM TCEP as a mobile phase at 20°C. Each 50 μL sample was run through SEC and 3 s x-ray exposures were collected continuously during a ∼40 min elution. Samples were examined with incident light at λ = 1.03 Å and at a sample to detector distance of 1.5 m resulting in scattering vectors, q, ranging from 0.01 to 0.6 Å⁻¹, where the scattering vector is defined as q = 4π sinθ/λ and 2θ is the measured scattering angle. SAXS frames recorded before the protein elution peak were used to subtract all other frames. Buffer subtraction and data reduction were performed at the beamline with SCÅTTER (70). Buffer matched controls were used for buffer subtraction.

Further data analysis of the merged SAXS data was carried out with PRIMUS (71) and the ATSAS program suite (72). Estimates of the radius of gyration (R_g) from the Guinier region were measured with PRIMUS. Maximum dimension (D_max) of particles was estimated from an indirect Fourier transform of the SAXS profiles using GNOM (73). Values of D_max between 120 and 140 Å provided the best solutions for both samples. The oligomeric state of the samples was assessed by estimating their molecular weight using the method implemented in the SAXSMoW2 server (74). Three-dimensional spatial representations of ss and mm PCDH15 EC10-MAD12, classified as flat using DATCLASS and with an AMBIMETER shape ambiguity score of 2.3 for both samples (possibly ambiguous) (75), were obtained by ab initio modeling with DAMMIF (76) considering q < 0.2, P1 symmetry, and prolate anisometry. For each sample, 30 models were generated and averaged with DAMAVER (77) with a mean normalized spatial discrepancy (NSD) of 1.07 ± 0.16 (one model rejected) and 0.98 ± 0.13 (two models rejected) for ss and mm PCDH15 EC10-MAD12, respectively. A final bead model refined against the SAXS data was produced with DAMMIN/DAMMSTART (78) with an estimated resolution of 44 ± 3 and 37 ± 3 Å (79). Model scattering intensities were computed from ss PCDH15 EC10-MAD12 (6BXZ) and fitted to the experimental SAXS data using FoXS (80). Search of conformational changes of the dimeric structure in solution was performed by normal modes analysis with SREFLEX (81).

Molecular dynamics simulations

The structure of ss PCDH15 EC10-MAD12 (PDB: 6BXZ) was used to build monomeric and dimeric systems. Missing residues in one of the dimer subunits were included in the structure using coordinates obtained from the other superposed subunit. Final models included residues p.G1009–p.I1342. The psfgen, solvate, and autoionize visual molecular dynamics plugins were used to build all systems (Table S4) (82). Hydrogen atoms were automatically added to protein and crystallographic water molecules. Residues D, E, K, and R were charged. Histidine residues were neutral, and their protonation states were chosen to favor the formation of evident hydrogen bonds. Additional water molecules and randomly placed ions were used to solvate the systems (150 mM KCl) in boxes large enough to accommodate stretched states and to prevent interactions between periodic images.

Molecular dynamics simulations were performed using NAMD 2.12 (83), the CHARMM36 force field for proteins with the CMAP correction and the TIP3P model for water (84). A cutoff of 12 Å (with a switching function starting at 10 Å) was used for van der Waals interactions along with periodic boundary conditions. The Particle Mesh Ewald method was used to compute long-range electrostatic forces (grid point density of >1 Å⁻³). A uniform 2 fs integration time step was used together with SHAKE. Langevin dynamics was utilized to enforce constant temperature (T = 300 K; γ = 0.1 ps⁻¹). Constant pressure simulations (NpT) at 1 atm were done using the hybrid Nosé-Hoover Langevin piston method with a 200 fs decay period and a 100 fs damping time constant.

Constant-velocity stretching simulations used the SMD method and the NAMD Tcl forces interface. Constant-velocity SMD simulations (85, 86, 87, 88) were performed by attaching Cα atoms of N- and C-terminal residues to independent virtual springs. For some simulations (Table S4), these springs were connected to virtual slabs connected to a third spring. All springs had stiffness k_s = 1 kcal mol⁻¹ Å⁻². The free ends of the stretching springs were moved away from the protein in opposite directions at a constant velocity. Applied forces were computed using the extension of the virtual springs. Maximum force peaks and their averages were computed from 50 ps running averages used to eliminate local fluctuations.

Sequence analyses, structural modeling, and analysis tools

To compare protein sequences among various species, 18 sequences from 17 species were obtained for their longest PCDH15 isoforms from the National Center for Biotechnology Information protein database (Table S1) and were split into the corresponding EC10, EC11, and MAD12 fragments before alignment. These fragments were aligned as in (47). The ClustalW algorithm (89) on Geneious (90) was used to obtain percent sequence identity for each fragment. Alignment files from Geneious were colored according to % of sequence identity in JalView with 45% conservation threshold (91). Similarly, sequences among MADs from different cadherin subfamilies (Table S2) were split for sequence alignments (colored according to % sequence identity in JalView with 1% conservation threshold), which were used as input into the Sequence Identity and Similarity (SIAS) server (92) to obtain sequence identity matrices. Residue numbering throughout the text and in the structures corresponds to the processed protein, which does not include the signal peptide. Numbering is based on entries hs PCDH15 isoform 1 precursor, National Center for Biotechnology Information Reference Sequence: NP_001136235.1 (signal peptide with 26 residues). Details of interatomic interactions were analyzed using Maestro (93) and visual molecular dynamics (82). Molecular stereo images were prepared with PyMOL (94). Plots and curve fits were prepared with XMGrace or QtiPlot. Protein-protein contacts were analyzed with the Protein, Interfaces, Surfaces, and Assemblies (PISA) server (95). Coordinates for structural models of hs CDH23 MAD28, hs PCDH24 MAD10, hs FAT1 MAD35, hs FAT2 MAD34, hs CELSR1 MAD10, hs CELSR2 MAD10, dm FAT MAD35, as well as dimeric hs CDH23 EC26-MAD28 and hs PCDH24 EC8-MAD10 are available upon request.

Protein-protein docking and MM-GBSA calculations

Homology models of hs CDH23 EC26-27 and hs PCDH24 EC8-9 were obtained using SWISS-MODEL (61) and published structures (PDB: 5TFK and 5SZR) (33, 47) as templates. RaptorX MAD models were connected to models of the preceding EC repeats to build L-shaped structures using COOT (62) and ss PCDH15 EC10-MAD12 as a guide. Models were minimized to reduce steric clashes with Macromodel and the OPLS3e force field using standard parameters (implicit solvent, charges from forcefield, PRCG method with 2500 iterations in gradient conversion mode, convergence threshold of 0.05) without constraints (96). Resulting monomers were built as dimers based on ss PCDH15 EC10-MAD12 and further minimized (97). Monomeric protomers were then used for protein-protein docking with BioLuminate using the “Dimer mode” with 70,000 rotations for the ligand. Final docking poses were selected after automatized clustering and structural superposition to ss PCDH15 EC10-MAD12. Final models were further minimized and used as inputs for molecular mechanics-generalized Born surface area (MM-GBSA) calculations using Prime (96), the VSGB solvation model and the OPLS3e force field with standard parameters, the “side chain only” minimization mode, and no constraints.

Crystal structure of pig PCDH15 EC10-MAD12 reveals a ferredoxin-like fold for MAD12

Diffracting crystals were obtained for the ss PCDH15 EC10-MAD12 protein fragment, for which an x-ray crystal structure was solved and refined to 2.1 Å resolution (Table 1). Two protomers were observed in the asymmetric unit, comprising residues p.G1009–p.I1342 and p.I1011–p.T1340, with missing residues not observed in the electron density map. These protomers adopted an L-shaped architecture (Fig. 1 B) and interacted with each other in a parallel cis arrangement, as also recently reported in (49) (Fig. S4) and as discussed below. The long end of the L-shaped bent structure has two EC repeats (EC10 and EC11) with typical Greek key folds (seven β strands labeled A–G for EC10; A’–G’ for EC11; Fig. 1 C) and a canonical linker region with three bound Ca²⁺ ions at sites 1, 2, and 3 (Figs. 1 B and 2 A; Fig. S1 A). The ss PCDH15 EC10 repeat shares structural features with its orthologs from human and mouse (45), with some differences in its FG connector, which folds into a stable α helix in the presence of the preceding EC9 repeat to be part of a bent, Ca²⁺-free EC9-10 linker region. Instead, in one chain of our structure, this loop is extended forming a β hairpin (F^∗G^∗) involved in crystal contacts (Fig. 1, B and C; Fig. S4, A and B), whereas electron density is absent for the other chain. The ss PCDH15 EC11 repeat has a canonical fold with an atypical and highly conserved p.N1223-XL-p.D1226 linker that forms a 3₁₀ helix leading to MAD12 (Figs. 2 B and S2), a domain that is positioned next to the EC11 face formed by β-strands A’, F’, and G’. MAD12 tucks against EC11 to form the short arm of the L-shaped protomer.

The overall architecture of MAD12 is similar to that observed for the ferredoxin-like fold in the Nup54 αβ domain (PDB: 5C2U, core RMSD 1.64 Å) (98). MAD12 has four β strands and two kinked α helices in a βαββαβ fold with kinks dividing α1 into α1’ and α1”, as well as α2 into α2’ and α2” (Figs. 1 C and 2 B). Both MAD12 and Nup54 have a short β strand preceding α2’ (labeled β3^∗) that forms a β-hairpin with the end of β3. There are also key differences between these folds, the most significant being related to kinks in MAD12 that divide β2 into β2’ and β2”, and β3 into β3’ and β3”. The unique β2”–β3’ hairpin, bridged by a 3₁₀ helix, bends to form a “hook” reaching in the direction of the membrane, thereby establishing contacts that support the α3 linker connecting to PCDH15’s transmembrane helix (Figs. 1 B and 2 B; Fig. S1, B and D).

The L-shaped structure of the PCDH15 EC10-MAD12 protomer harbors various intramolecular contacts between EC11 and MAD12 involving both hydrophobic and hydrophilic surface residues located at the EC11-MAD12 linker, at the beginning of β2’, at the end of β3”, and at the β3”-β3^∗-α2’ connection (Fig. 2, C–E). This interaction surface is ∼640 Å̊² (for one protomer) and the hydrophobic nature of some residues (F1127, I1129, L1218, V1222, V1264, Y1290, I1292, P1294, A1299) that stabilize the interface between EC11 and MAD12 is strictly conserved across species (Figs. 2 C and S2), with some variations in sidechain size, which suggests that this intramolecular contact is somewhat robust to residue substitutions as long as their hydrophobicity is maintained. Interestingly, a missense mutation (p.G1130R) that causes inherited deafness (99) alters a site at this interface (Figs. 2, C–E and S1 C). The sidechain of the mutated residue would not point toward the interface but would likely destabilize the hydrophobic core of EC11. This suggests that disruption of the structural integrity of EC11 and the intramolecular EC11-MAD12 contacts might be deleterious for PCDH15 function.

Pig PCDH15 EC10-MAD12 forms parallel cis dimers in crystallo

The two protomers observed in the asymmetric unit form a parallel cis dimer with three contact points (Fig. 3). The first one involves interactions between the two copies of repeat EC10, which face each other bringing together residues on the surface of their β-strands A, G, and F (Fig. 3, A, C, and E). The surface area of this point of contact is small (∼185 Å²) with interactions mediated by charged and polar residues forming hydrogen bonds and salt-bridges (p.E1017, p.E1018, p.R1020, p.R1085, p.K1108, and p.Y1110). There is strict conservation of charge for all five charged sites in the pig construct, whereas p.Y1110 is conserved only in mammals and replaced by hydrophobic residues in birds, reptiles, and some fish (Fig. S2). Under MAD12-induced dimerization, details of this contact may minimally vary across species, and although not required for dimerization (45, 48, 49), the conserved EC10-EC10 contacts may help stabilize a subset of possible conformations for the PCDH15 EC9-MAD12 subdomain (49). Interestingly, the dynamics of this EC10-EC10 cis interface might be allosterically controlled by Ca²⁺-binding to the EC10-11 linker positioned right below it, thereby providing some Ca²⁺-dependent rigidity at the lower end of the tip link.

The other two contact points, also recently reported in (49), are equivalent and involve the tail end of one of the L-shaped molecules interacting with the elbow of the other, together forming a closed ring arrangement with some additional contacts stemming from residues pointing toward the center of the ring on the membrane-proximal end of the dimer (Figs. 3, A–E and S5). Total surface area for these interfaces is ∼975 Å², with hydrophobic and hydrophilic residues from EC11’s A’B’ loop (p.R1137, p.M1138, and p.F1139), C’D’ loop (p.E1172), β-strand D’ (p.V1175), the end of β-strand E’ (p.K1184 and p.A1186) and the beginning of the E’F’ loop (p.L1188), as well as residues from MAD12’s α1’ helix (p.P1236, p.T1237, and p.E1240) and its β-strands β2’ (p.E1266) and β3’ (p.E1280 and p.Y1282). Residue p.T1237, which forms a hydrogen bond with p.E1172 at this interface, is associated to inherited deafness. The mutation p.T1237P is correlated with a deafness phenotype, although causality has not been established (100). Interestingly, the ring-like arrangement (Fig. 3 B) features an opening exposed to solvent and crossed by p.R1137 residues from the two protomers. These sidechains interact with either the backbone carbonyl group of p.D1135 or with the sidechain of p.E1266 from the neighboring EC11. These interactions bring AB loops from each EC11 protomer together to stabilize the dimeric interface. Charge is again strictly conserved throughout all the species analyzed (Fig. S2), while hydrophobicity is less conserved for two of the sites (p.1138 and p.1188). The total surface area involved in the potential cis dimerization interface, including all three points of contact (∼1160 Å²), and the sequence conservation of residues involved in it suggests that parallel dimerization might be robust and relevant for PCDH15 function.

Pig and mouse PCDH15 EC10-MAD12 form dimers in solution

To determine the oligomeric state of PCDH15 EC10-MAD12 in solution and to test the propensity of this fragment to dimerize, as seen in the x-ray crystal structure, we used pig and mouse protein fragments for AUC experiments, as well as SEC coupled to MALLS and SAXS. Sedimentation velocity AUC data for the refolded ss PCDH15 EC10-MAD12 (produced in bacteria; theoretical monomeric mass of 39.96 kDa) show a major peak with a sedimentation coefficient (S) of ∼4 and a minor peak at S ∼3 (Fig. 4 A). From these AUC measurements, the molar mass of the major peak is ∼60 kDa, yet uncertainty in buffer density, viscosity, specific molar volume, and friction coefficients precluded a more accurate estimate. Nevertheless, we assigned these peaks to dimer and monomer forms, respectively. Refolded mm PCDH15 EC10-MAD12 (produced in bacteria; theoretical monomeric mass of 39.97 kDa) also showed similar peaks (Fig. 4 A). Consistent with AUC, although less conclusive, SEC-MALLS experiments suggest that the ss and mm PCDH15 EC10-MAD12 proteins are dimeric in solution (M_w = 62.13 kDa ± 0.058% and 61.13 kDa ± 0.043%, respectively) (Fig. S6, A and B). Molecular weight estimates from SEC-SAXS experiments for both the pig and the mouse fragments (76.20 and 74.92 kDa; 4.7–6.3% discrepancy with sequence-derived molecular weight) are in better agreement with the expected mass for dimeric PCDH15 EC10-MAD12 in solution.

Similarly, glycosylated mm^∗ PCDH15 EC10-MAD12 produced in a mammalian expression system behaves as a dimer in solution. AUC experiments show that mm^∗ PCDH15 EC10-MAD12 is predominantly a dimer with a calculated mass of 89.4 ± 0.3 kDa. The c(S) distribution (Fig. 4 A) also shows a small monomer peak corresponding to a mass of 54.6 ± 2.1 kDa). The expected mass is ∼55 kDa, as determined by SDS-gel analysis, with the additional mass due to posttranslational modifications on the protein (Fig. S3). Estimates from SEC-MALLS (130.3 kDa ± 0.299%) (Fig. S6 A) are consistent with the AUC data and with the expected mass of a dimeric mm^∗ PCDH15 EC10-MAD12 in solution. Together, this data strongly support a dimeric arrangement at the membrane proximal region of PCDH15, as also recently reported for similar protein fragments (48, 49).

Data from SAXS experiments can also be used to obtain information about the size and shape of a protein in solution. The radius of gyration (R_g) for both ss and mm PCDH15 EC10-MAD12 was estimated using two approaches. First, the Guinier analysis was used to obtain R_g values of 35.37 ± 0.27 and 35.53 ± 0.53 Å, respectively (Fig. 4, B and C; Table S3). In the second approach, the indirect Fourier transform of the SAXS profile yielded R_g values of 36.30 ± 0.09 Å for the pig and 36.34 ± 0.08 Å for the mouse protein, with maximum dimensions (D_max) of 131 and 122 Å, respectively (Fig. 4 D; Table S3). These estimates are in excellent agreement with each other and with the R_g obtained using the ss PCDH15 EC10-MAD12 structure (34 Å), suggesting that the overall shape of the dimer observed in the crystallographic structure is maintained in solution. However, comparison of the SAXS data to x-ray intensities modeled from the crystal structure revealed some discrepancies reflected in large χ² values obtained for the fitting (5.09 and 3.42 for ss and mm PCDH15 EC10-MAD12, respectively) and by a clear dip in the q-region 0.1–0.25 Å⁻¹ observed in the modeled intensities but absent in the experimental data (Fig. 4 B). These discrepancies may indicate loss of symmetry in solution and may arise from the relaxation and reaccommodation of the dimer interface and “wobbling” of the protein. Analysis of the Kratky plot (Fig. S6 C) indicates that while both proteins are folded in solution, they exhibit significant flexibility, likely arising from interrepeat motion or the presence of a C-terminal tail including residues p.I1342–p.E1353 (not seen in the structure) and a His-tag used for protein purification.

The presence of a flexible or conformationally heterogeneous dimer in the samples was further highlighted by ab-initio modeling of the SAXS data. 30 models produced for shape reconstruction considering P1 symmetry and prolate anisometry (see Materials and Methods) were averaged to generate final models that were refined against the SAXS data for each sample using DAMMIF/DAMMIN/DAMSTART (76, 78). Predicted x-ray intensities of the resulting bead models are in good agreement with the experimental data (χ² = 1.45 for pig and χ² = 1.15 for mouse) and with the crystallographic structure (Fig. 4, E and F). However, the bead models superposed to the crystallographic structure highlight the lack of symmetry of the dimer in solution and the likely flexibility of the EC10-EC10 interface.

To determine if the crystallographic structure could be reaccommodated in a way that fits the experimental data, we carried out normal-mode analysis on it followed by fitting to the SAXS data with SREFLEX (81) for data from both pig and mouse samples. X-ray intensities calculated from these models fit the experimental data better than the crystal structure (χ² = 1.64 for pig and χ² = 1.23 for mouse) and lack the dip in the q-region 0.1–0.25 Å⁻¹. The adjusted models have an overall root-mean-square-deviation (RMSD) for Cα atoms of ∼7 Å when compared with the crystallographic structure. These models show changes in the orientation and rotation of the ECs with loss of symmetry and with one EC10 repeat tilted and leaning toward the adjacent, straighter EC10 (Fig. S6 D). Overall, SAXS data suggest a flexible and perhaps asymmetric arrangement of EC10 repeats while strongly supporting a dimeric PCDH15 EC10-MAD12 in solution.

Steered molecular dynamics simulations predict the mechanics of ss PCDH15 EC10-MAD12

The topology and structure of MAD12 has important implications for the transmission of force to the inner-ear transduction channel. Unlike the β-sandwich Greek-key fold of EC repeats, in which N- and C-terminal points are at opposite ends of the folding unit, the N-terminus of MAD12’s βαββαβ fold is spatially nearby its C-terminal end (within ∼5 Å). While stretching and mechanical unfolding of an EC repeat requires large forces to shear-out β-strand A or G (101, 102, 103), unfolding of MAD12, when pulling from the N- and C-termini, would require unzipping of β4 from β1 and subsequent unraveling of α2. Experiments and simulation have suggested that unzipping of β strands is easier than shearing (104, 105, 106, 107, 108), indicating that unfolding of MAD12 may require less force than the force needed to unfold EC repeats.

To explore the elastic response of ss PCDH15 EC10-MAD12, its unfolding pathway, and its mechanical strength, we carried out SMD simulations in which force was applied through independent springs attached to the ends of the protein fragment (86, 87, 88, 109, 110). The free ends of the springs were in turn moved at constant speed in opposite directions, mimicking in vivo force application. Simulation systems, encompassing up to ∼279,000 atoms (Table S4), included explicit solvent and ions along with either a monomer or a dimer of ss PCDH15 EC10-MAD12. Stretching of the monomer at 0.1 nm/ns revealed quick unraveling of α3, followed by opening of the EC11-MAD12 interface (separation of p.R1144 from p.E1244) and unrolling of MAD12 accompanied by a 90° swing of MAD12’s hook (Figs. 5 A and S7, A and B; Simulation S1d; Table S4; Video S1). A final unfolding event involved both unzipping of β4 from β1 with rupture of backbone hydrogen bonds, and splitting of two salt bridges formed during the SMD trajectory (p.D1286–p.R1334 and p.R1271–p.E1332) (Fig. 5, A and B; Fig. S7, A–G). Rupture of this “electrostatic lock,” which linked β4 with β2” (hook) and with β3”, correlated with the largest force peak (477.9 pN) monitored throughout the simulation (Fig. S7 C). Further unfolding of MAD12’s α2 ensued at drastically smaller forces without noticeable unfolding of EC10 or EC11. A similar scenario was observed in the slowest stretching simulation at 0.02 nm/ns (Table S4, S1e), with the magnitude of force peaks decreasing at slower stretching speeds as expected (111). Monomer force peaks at all stretching speeds were associated to MAD12 unfolding and were smaller in magnitude than unfolding forces reported for other CDH23 and PCDH15 EC repeats stretched under similar simulation conditions (Fig. 5 G) (45, 46, 103). In addition, MAD12’s unfolding force peaks were smaller than predicted unbinding force peaks for the PCDH15-CDH23 monomeric complex (43, 112), suggesting that a single tip-link handshake bond formed by EC1-2 repeats of CDH23 and PCDH15 can withstand similar or larger forces than MAD12.

Stretching of the ss PCDH15 EC10-MAD12 dimer was carried out using two constant-velocity SMD protocols and revealed a more complex response. First, we applied forces in opposite directions through independent springs attached to each of the four ends of the protein fragments (Fig. 5 C). At the slowest stretching speed (0.1 nm/ns) (Table S4, S2d), one of the subunits’ β4 detached from β1 as MAD12 was separating out from EC11 and unrolling, which correlated with a broad and rather undefined force peak (data not shown). In the other subunit MAD12 separated from EC11 and unrolled after unzipping β4 from β1 and rupture of the electrostatic lock described above, which correlated with a clear force peak (Fig. 5 D). In a second approach, pairs of N- or C-termini were connected to virtual slabs through independent springs (Fig. 5 E). The slabs were in turn connected through springs to SMD atoms moving at constant speeds in opposite directions. In the simulations that used this approach at the slowest stretching speeds (0.1 and 0.02 nm/ns) (Table S4, S2g and S2h), MAD12 from one subunit separated from EC11 and unrolled before β4 detached from β1, while MAD12 from the other subunit remained partially attached to EC11, with unzipping of β4 from β1 happening first (Figs. 5 E and S8; Video S2). In all cases, additional intermolecular interactions formed throughout trajectories and MAD12’s hook swung as monitored during stretching of the monomer of ss PCDH15 EC10-MAD12. Average unfolding force peaks of MAD12 were smaller than unfolding force peaks of PCDH15 EC repeats and unbinding force peaks of the PCDH15 EC1-2 + CDH23 EC1-2 bond (43, 45, 46, 103, 112), again suggesting that MAD12 is mechanically weak, with the tip-link bond being of similar or larger strength at the stretching speeds tested in this study (Fig. 5, G and H).

Secondary structure analyses and modeling predict MAD folds in other cadherins

The topology and structure of MAD12 has important implications for other cadherins as well. The high structural similarity observed between MAD12 and Nup54 (98), despite their very poor sequence identity (Figs. S9 and S10), led us to search for other βαββαβ domains that might be hidden within the cadherin superfamily (Fig. 6). We looked for extracellular segments that lacked predicted domain architecture, referred to as unknown regions according to SMART (113). In parallel, we predicted the secondary structure of sequences from all cadherins in the human genome using RaptorX (114), which had correctly predicted the secondary structure of MAD12 and identified Nup54 (PDB: 5C2U) (98) as a good template for “homology” modeling. We then focused on sequences predicted to have the βαββαβ fold, and found the pattern in CDH23, PCDH24 (also known as CDHR2), all FAT cadherins, and all CELSRs (Figs. S9 and S11). These predicted βαββαβ-fold containing segments were classified as unknown regions by SMART.

The potential presence of MADs in CDH23 and PCDH24 is intriguing (MAD28 and MAD10, respectively) (Fig. 6). In both cases, these domains would be located close to their membrane insertion points, linking cadherin EC repeats to their associated transmembrane helices. Both proteins had been identified as members of the Cr-2 subfamily based on sequence similarity of their extracellular tips (1, 4, 103, 115), and the presence of a ferredoxin-fold domain in both these proteins further confirms their structural relationship. Perhaps even more intriguing is the existence of potential MAD folds in human FAT and CELSR cadherins. Analysis of the Drosophila melanogaster (dm) FAT also shows the potential existence of a similar βαββαβ fold (MAD35) (Fig. S11). In all these cases, the predicted “unknown region” is between the last, C-terminal EC and subsequent EGF or LAMG domains (MAD35 for FAT1; MAD34 for FAT2; MAD33 for FAT3; MAD35 for FAT4; and MAD10 for CELSR1-3) (Fig. 6 A).

To gain further insights into the structural similarities among cadherin MADs, we used RaptorX to create structural models for each of these domains in the human “MAD bearing” cadherins (PCDH15, CDH23, PCDH24, FAT1-4, and CELSR1-3) and the fly sequences (dm FAT) mentioned above (11 models) (Figs. 6, B–E and S11, A–G). Visual inspection confirmed the βαββαβ fold for eight of these models (hs PCDH15, hs CDH23, hs PCDH24, hs FAT1, hs FAT2, hs CELSR1, hs CELSR2, and dm FAT). In addition, we classified models according to the hydrophobicity of their core. Hydrophobic cores without polar or charged residues pointing into it were observed for PCDH15 MAD12, CDH23 MAD28, FAT1 MAD35, and CELSR1-2 MAD10, which also had their charged residues distributed on their surfaces and loops, as expected. Models for PCDH24 MAD10, FAT2 MAD34, and dm MAD35 were satisfactory, with one or two polar residues lying at the interface between their hydrophobic cores and loops, which could be easily accommodated by rotamer rearrangement. A structural alignment of these models to ss PCDH15 MAD12 using STAMP (116, 117) revealed high structural conservation at the core secondary structure elements, and poor structural conservation of loops for the eight models mentioned above (Fig. 6).

The dimeric structure of ss PCDH15 EC10-MAD12 suggests that other MADs, and their preceding EC repeats, could form cis homodimers as well. However, part of the dimer interface involves the β2”–3₁₀–β3’ hooks that are not well predicted by RaptorX (see “control” prediction in Fig. 6 C). Nevertheless, the β2”–3₁₀–β3’ hook observed in ss PCDH15 MAD12 might be present in CDH23 MAD28. Although the hook is not observed in our model, CDH23 MAD28 has similar length and distribution of charged residues in this region (Fig. S9). Similarly, in the secondary structure prediction by RaptorX (Fig. S9), the well predicted short β3^∗ strand interacting with the preceding EC repeat exist in CDH23 and CELSR1 and 2, thus suggesting an architecture that might be similar to the L-shaped PCDH15 MAD12 favoring dimerization.

To determine whether parallel dimerization facilitated by MADs in other cadherins is possible in silico, we built models of hs CDH23 EC26-MAD28 and hs PCDH24 EC8-MAD10 and computed their interaction energies. Homology models for EC repeats were coupled to RaptorX models for MADs using the ss PCDH15 EC10-MAD12 structure as a guide. The resulting, models with L-shaped protomers were minimized as dimers and used for “blind” molecular docking to predict binding poses for CDH23 and PCDH24, which resulted in cis parallel dimers that are similar to ss PCDH15 EC10-MAD12 (Fig. S12). In both cases, the L-shaped architecture of hs CDH23 EC26-MAD28 and hs PCDH24 EC8-MAD10 favors the formation of a closed ring-like arrangement involving several hydrophobic residues that could stabilize their dimeric interfaces. In addition, free energy binding calculations (via MM-GBSA) predict strong binding for these admittedly biased dimers (Table S5), suggesting that MAD-induced parallel dimerization is possible in CDH23 and PCDH24. Taken together, our modeling analyses suggest that MADs are present in all the aforementioned proteins, with some variations in fold at loops and at the β2”–3₁₀–β3’ and β3”–β3^∗ connections that might determine their oligomerization states and function.