Novel tau filament fold in corticobasal degeneration, a four-repeat tauopathy

Corticobasal degeneration (CBD) is a neurodegenerative tauopathy that is characterised by motor and cognitive disturbances (1–3). A higher frequency of the H1 haplotype of MAPT, the tau gene, is present in cases of CBD than in controls (4, 5) and genome-wide association studies have identified additional risk factors (6). By histology, astrocytic plaques are diagnostic of CBD (7, 8), as are detergent-insoluble tau fragments of 37 kDa by SDS-PAGE (9). Like progressive supranuclear palsy (PSP), globular glial tauopathy (GGT) and argyrophilic grain disease (AGD) (10), CBD is characterised by abundant filamentous tau inclusions that are made of isoforms with four microtubule-binding repeats (4R) (11–15). This distinguishes 4R tauopathies from Pick’s disease, filaments of which are made of three-repeat (3R) tau isoforms, and from Alzheimer’s disease and chronic traumatic encephalopathy (CTE), where both 3R and 4R tau isoforms are found in the filaments (16). Here we report the structures of tau filaments extracted from the brains of three individuals with CBD using electron cryo-microscopy (cryo-EM). They were identical between cases, but distinct from those of Alzheimer’s disease, Pick’s disease and CTE (17–19). The core of CBD filaments comprises residues K274-E380 of tau, spanning the last residue of R1, the whole of R2, R3 and R4, as well as 12 amino acids after R4. It adopts a novel four-layered fold, which encloses a large non-proteinaceous density. The latter is surrounded by the side chains of lysine residues 290 and 294 from R2 and 370 from the sequence after R4. CBD is the first 4R tauopathy with filaments of known structure.

approximately 1,300 Å and a maximal width of 260 Å with similar minimal width as narrow ones. We named these filaments Type I (narrow) and Type II (wide) CBD filaments, respectively. The ratios of Type II to Type I filaments ranged from 3:1 to 1:1, depending on cases. Co-pathologies are often found in CBD (22,23). Small amounts of assembled TDP-43 were present in frontal cortex of CBD cases 1 and 2; CBD case 3 was negative (Extended Data Figure 2).  Table 1). In the three cases, Type I filaments are composed of a single protofilament and adopt a novel four-layered fold. Like CTE filaments, each protofilament of CBD contains an additional density that is surrounded by density of the tau protein chain. Unlike CTE (19), the additional density is found in a positively charged environment. Type II filaments consist of pairs of identical protofilaments of Type I, related by C2 symmetry, with less well-resolved maps at the ends of the cores than in their central parts. For case 1, we obtained maps of Type I and Type II tau filaments at overall resolutions of 3.2 Å and 3.0 Å, respectively (Extended Data Figure 3a Type I and Type II tau filaments contain a common protofilament, whose core structure (CBD fold) is composed of residues K274-E380, i.e. the last residue of R1, all of R2-R4, and 12 amino acids after R4.
The central four layers are formed by β7, β4, β3 and β10. Strands β3 and β4 are connected by a sharp turn, whereas β7 and β10 are connected through β8 and β9, which wrap around the turn. On the other side, β2, β5 and β6 form a three-layered structure. β2 packs against one end of β5 and β6 packs against the other end. The first and the last strands, β1 and β11, pack against each other and close a hydrophilic cavity formed by residues from β2, β3, β10, β11 and the connections between β1 and β2, as well as between β2 and β3 Each tau repeat contains a PGGG motif (16). In the CBD fold, the PGGG motif of R2 (residues 301 to 304) forms a tight turn between β3 and β4, which is essential for the formation of the four-layered cross-β packing. The PGGG motif of R3 (residues 332 to 335) adopts an extended conformation between β6 and β7, compensating for the shorter lengths of these strands compared to the opposing We hypothesise that the extra density is made of non-proteinaceous, polyanionic molecules with a charge of -3 per rung. The buried nature of the negatively charged molecules, their high occupancy and presence in most filaments at end-stage disease indicate that they are continuously incorporated during filament formation. It is therefore possible that these molecules stabilise the CBD fold during initial filament assembly and/or subsequent seeded aggregation. tauopathy of known filament structure. Differences in protofilament structure are observed between diseases, but not between subjects with a given disease, consistent with the existence of distinct conformers of assembled tau in different tauopathies ( Figure 3).
The cores of tau filaments from human brain of known structure all contain R3, R4 and 10-12 amino acids after R4 (17-19).
Filaments comprising 3R+4R tau in Alzheimer's disease and CTE do not have R1 or R2 in their cores, and those comprising 3R tau in Pick's disease have part of R1, whereas filaments comprising 4R tau in CBD contain the whole of R2. This makes the CBD fold the largest known tau fold, with 107 ordered residues. Therefore, filament disassembly may come at a relatively high energetic cost, which may in turn have implications for seeded aggregation and disease progression. Tau assemblies from CBD brains have been shown to seed specific aggregation (29-31). It is likely that filaments from other 4R tauopathies, such as PSP, GGT and AGD, have also at least part of R2 in their cores, but these structures remain to be determined.
It was previously not known why only 4R tau isoforms are present in the filaments of CBD. Our structure reveals that S305, the last residue of R2, which starts β4, is located at a position, where the side chain of K274, the last residue of R1, cannot fit ( Figure 2b). Moreover, if R1 were incorporated instead of R2, K294 would be replaced by T263, which would weaken the interaction with the extra density (Figure 2b,d). In support, the sarkosylinsoluble fraction from CBD cases 1-3, which was used for cryo-EM, seeded aggregation of soluble tau in SH-SY5Y cells expressing fulllength 4R, but not 3R, human tau (Extended Data Fig. 6). Similarly, CBD-tau recruited only soluble 4R tau into insoluble aggregates in primary neurons, whereas Alzheimer's disease-tau recruited both 3R and 4R tau (31). In contrast, tau filaments extracted from the brain of a patient with Pick's disease seeded aggregation of 3R, but not 4R, human tau (18). Templated misfolding of this type may explain why only 4R tau is incorporated into CBD filaments.
Despite differences between folds, with the structures of tau filaments from four human tauopathies now known, common patterns are beginning to emerge ( Figure 3). Packing between β1 and β11 in the CBD fold resembles that between β1 and β8 in the Residues 374 HKLTFRE 380 are missing from the widely used tau constructs K18 and K19, which end at E372 (36). The hairpin-like structure of β4-β7 in the CBD fold resembles that of β3-β6 in the Pick fold, with the exception of C322, which points inwards to form the sharp turn in the CBD fold, and which points outwards in the Pick fold. Interestingly, in all four tau filament folds from human brain, β-strands are formed by approximately the same residues ( Figure 3); this is also true of tau filaments assembled in vitro using heparin (37). It suggests a model for the diversity of tau folds, where β-strands form fixed building blocks and the loops and turns between strands provide diversity. Tau repeats contain many glycine and proline residues, all of which are located in loops and turns.
This work may shed light on why tau folds differ between diseases, which may in turn reveal mechanisms that lead to ordered assembly. Post-translational modifications may be important. Thus, deamidation of N279 in R2 of tau takes place in Alzheimer's disease, but not in CBD or PSP (38). This residue, which is located in β1 of CBD, is outside the structured core of AD and CTE filaments.
Association of non-proteinaceous cofactors with tau filaments from human brain was unexpected, even though it is well established that such factors can induce assembly of soluble, unmodified tau protein in vitro (16). In the same way that hydrophobic molecules inside the β-helix may shape the CTE fold (19), polyanionic molecules inside the positively charged cavity may help to form the CBD fold. We previously speculated that the extra densities near K317 and K321 of tau on the periphery of the Alzheimer fold (17), which are also observed in the CTE fold (19), may be formed by 7 EFE 9 of tau. They are believed to be essential for the formation of the straight tau filaments of Alzheimer's disease. Their similarity to the extra density in the CBD fold raises the possibility that nonproteinaceous molecules may also play a role in providing specificity for tau assembly into Alzheimer and CTE folds.      The APOE genotypes of CBD cases 1-3 were: case 1 (ε3/ε4), case 2 (ε3/ε3), case 3 (ε3/ε3).
Whole-exome sequencing. Target enrichment made use of the SureSelectTX human all-exon library (V6, 58 Mb, Agilent) and high-throughput sequencing was carried out using a HiSeq4,000 (2x75-bp paired-end configuration, Illumina). Bioinformatics analyses were performed as described (43). Following a 10 min. centrifugation at 20,000 g, the supernatants were spun at 100,000 g for 20 min. The pellets were resuspended in 700 µl/g extraction buffer and centrifuged at 9,500 g for 10 min.
For CBD cases 1 and 2, the supernatants were diluted 3-fold in 50 mM Tris-HCl, pH 7.5, containing 0.15 M NaCl, 10% sucrose and 0.2% sarkosyl and spun at 166,000 g for 30 min. For CBD case 3, the supernatant was spun at 100,000 g for 60 min and the pellet resuspended in 700 µl/g extraction buffer and centrifuged at 9,800 g. The supernatant was then spun at 100,000 g for 60 min.
For immuno-EM the samples were diluted 5-10-fold. Sarkosylinsoluble pellets of approximately 2 g frontal cortex were used for cryo-EM.

Immunolabelling, histology and silver staining.
Western blotting and immunogold negative-stain EM were carried out as described (45). For Western blotting, the samples were Electron cryo-microscopy. Extracted tau filaments were centrifuged at 3,000 g for 30 s, before being applied to glowdischarged holey carbon grids (Quantifoil Au R1.2/1.3, 300 mesh) and plunge-frozen in liquid ethane using a Thermo Fischer Vitrobot Mark IV. Images were acquired on a Gatan K2-Summit detector in counting mode using a Thermo Fischer Titan Krios microscope at 300 kV. A GIF-quantum energy filter (Gatan) was used with a slit width of 20 eV to remove inelastically scattered electrons. Further details are given in Extended Data Table 1.
Helical reconstruction. Movie frames were gain-corrected, aligned, dose-weighted and then summed into a single micrograph using MOTIONCOR2 (54). The micrographs were used to estimate the contrast transfer function (CTF) using Gctf (55). All subsequent image-processing steps were performed using helical reconstruction methods in RELION 3.0 (25,56,57). Both types of filaments were selected manually in the micrographs, and the resulting data sets were processed independently. For Type I (narrow) filaments, 135,646 segments were extracted with an inter-box distance of 14.1 Å and a box size of 920 pixels. Initial reference-free 2D classification was performed with images that were downscaled to 230 pixels to speed up calculations. Segments contributing to suboptimal 2D class averages were discarded. Assuming a helical rise of 4.75 Å, a helical twist of -0.9° was estimated from the crossover distance of filaments in the micrographs. Using these parameters, an initial 3D reference was reconstructed from the 2D class averages de novo. We then re-extracted the selected segments without downscaling them, and with a smaller box of 330 pixels. Using these segments and the de novo initial model low-pass filtered to 15 Å, we carried out 3D auto-refinement. We then used the refined reconstruction, low-pass filtered to 15 Å, as reference for a 3D classification without further image alignment. The segments contributing to the best 3D class were used for subsequent 3D auto-refinement of 24,073 selected segments.
Refinement of the helical parameters converged onto a helical twist of -0.845° and a helical rise of 4.786 Å. After Bayesian polishing and CTF refinement, the reconstruction was sharpened with a Bfactor of -26.23 Å 2 (Extended Data Table 1) using the standard post-processing procedure in RELION. The overall resolution of the final map was estimated as 3.2 Å from Fourier shell correlations at 0.143 between the two independently refined half-maps, using phase-randomisation to correct for convolution effects of a generous, soft-edged solvent mask that extended to 30% of the height of the box (58). Local resolution estimates were obtained using the same phase-randomisation procedure, but with a soft spherical mask that was moved over the entire map. For Type II (wide) filaments, 129,812 segments were extracted with an interbox distance of 14.1 Å and a box size of 460 pixels, which were downscaled to 230 pixels for reference-free 2D classification and particle selection. An initial 3D reference was again reconstructed from the 2D class averages de novo by assuming a helical rise of 4.75 Å and a helical twist of -0.65°, based on the estimated crossover distance of filaments. The resulting reconstruction suggested the presence of two-fold symmetry in the structure. To confirm whether this symmetry also had a translational component, we performed two 3D auto-refinements, one with C2 symmetry and one with C1 symmetry, but imposing a pseudo-2 1 screw axis. By comparing the results of the refinements, we found that the map obtained following the application of C2 symmetry was of higher resolution. The C2 map also showed separation of β-strands along the helical axis, which was absent in the map refined with the screw symmetry. We applied C2 symmetry in all subsequent refinements.
Using 3D classification without image alignment, the segments contributing to the best 3D class were selected and used for 3D auto-refinement with optimisation of helical twist and rise. A final 3D auto-refinement of 20,752 selected segments in boxes of 330 pixels without downscaling converged onto a helical twist of -0.61° and the helical rise at 4.786 Å. Following Bayesian polishing (59) and CTF refinement, the overall resolution of the final map was estimated as 3.0 Å. For Type I and Type II tau filaments in CBD case 2 and CBD case 3, the dataset processing steps were similar to Type I and Type II tau filaments in CBD case 1, with the initial 3D references reconstructed de novo independently for each dataset.
Model building and refinement. The models of the cores of Type I and Type II filaments were built de novo in combination of both 3.2 Å and 3.0 Å resolution reconstructions from CBD case 1 using COOT (60). We started model building from the 301 PGGG 304 motif, with the aromatic side chains of H299 and Y310 not far away, and worked our way towards the N-and C-terminal regions by manually adding amino acids and by targeted real-space refinement in the high-resolution core part of Type II filaments. Tracing of the chain was confirmed by the fitting of the 332 PGGG 335 motif, which neighbours the side chains of H329 and H330. Since the density of side chains of N368-E380 was weak in the 3.0 Å reconstruction of Type II filaments, we assigned the side chains in this region following the 3.2 Å reconstruction of Type I filaments. The structure of one protofilament from Type II filaments was rigid-body fitted into the reconstruction of Type I filaments to build the model of Type I filaments. Because of the lack of interaction between two protofilaments, the conformations of the side chains of K343 and K347 in Type I filaments were assigned differently from their counterparts in Type II filaments, according to the reconstruction.
Each model was then translated to give a stack of three consecutive monomers to preserve nearest-neighbour interactions for the middle chain in subsequent refinements. Because most residues adopted a β-strand conformation, hydrogen-bond restraints were imposed to preserve a parallel, in-register hydrogen bonding pattern in earlier stages of Fourier-space refinements. Local symmetry restraints were imposed to keep all β-strand rungs identical. Side chain clashes were detected using MOLPROBITY (61) and corrected by iterative cycles of real-space refinement in COOT and Fourier-space refinement in REFMAC (62) and PHENIX (63). For each refined structure, separate model refinements were performed against a single half-map, and the resulting model was compared to the other half-map to confirm the absence of overfitting (Extended Data Figure 3a,b). The final models were stable in refinements without additional restraints. Statistics for the final models are shown in Extended Data Table 1. and Type II (pink).
Extended Data Figure 5.

Protofilament interface in CBD Type II tau filaments.
Packing between residues 343 KLDFKDR 349 of the two protofilaments.
Inter-protofilament hydrogen bonds are shown in yellow. Intraprotofilament hydrogen bonds are shown in green.
Extended Data Figure 6.