Data reporting

No statistical methods were used to predetermine sample size. Experiments were not randomized. The investigators were not blinded to allocation during experiments and outcome assessment.

Donor and ethical information

Postmortem brain tissue was obtained through the Netherlands Brain Bank (Amsterdam, The Netherlands, In compliance with all ethical standards, brain donors signed informed consent regarding the usage of their brain tissue and clinical records for research purposes. This study was performed at Netherlands Brain Bank, Amsterdam University Medical Centres (location VUmc), and the University of Leeds. This study was approved by both VUmc and the University of Leeds Research Ethics Committee. Brain dissection and neuropathological diagnosis were performed according to international guidelines of Brain Net Europe II (BNE) consortium ( and NIA-AA54.

Unfixed, flash-frozen, rapid autopsy postmortem AD and non-demented donor postmortem brain were stored at −80 °C and provided a source of tissue for these studies. The control case was a 90-year-old man with a history of depression and prostate cancer. At the age of 87, the donor was admitted to a nursing home. In the last phase, he was passive with a concentration disorder, but he was not demented, with normal language skills, speaking skills and communicative ability. Neuropathological examination revealed slight atrophy of the temporal lobe. No β-amyloid plaques or neurofibrillary tangles were observed in the temporal lobe.

Neuropathology of donor tissue

A tissue block containing mid-temporal gyrus that was adjacent to the HPF unfixed tissue, was formalin-fixed and paraffin embedded. Sections of 5 µm thickness were prepared and mounted on Superfrost+ microscope slides (VWR). After overnight incubation at 37 °C, slides were deparaffinized using xylene and alcohol and subsequently washed in phosphate buffered saline (pH 7.4).

The histochemical detection of plaques and neurofibrillary tangles was as previously described55. In short, tissue was pretreated using 5% w/v periodic acid for 30 min. Subsequently, the tissue was silver impregnated using a 0.035% w/v silver nitrate solution for 30 min. After silver impregnation, the bound silver was developed using a reduction reaction induced by the development solution (2.5% w/v sodium carbonate, 0.1% w/v silver nitrate, 0.5% w/v tungstosilicic acid hydrate, 0.1% w/v ammonium nitrate and 0.1% w/v formaldehyde). The development was stopped by washing in 0.5% w/v acetic acid for 5 min and unbound silver was removed by washing in 5% w/v sodium thiosulfate for 5 min. Sections were counterstained using haematoxylin (Diapath). The sections were dehydrated using alcohol and xylene and coverslipped using Depex (BDH Laboratories Supplies).

For immunohistochemistry, deparaffinized sections were pretreated with 0.3% hydrogen peroxide in phosphate buffered saline for 30 min to block endogenous peroxidase activity, followed by autoclave heating (121 °C for 20 min) in 10 mM sodium citrate buffer (pH 6) for antigen retrieval. Primary antibodies were incubated overnight at room temperature and diluted in antibody diluent (Sigma-Aldrich) as follows: anti-pTauSer202/Thr205 clone AT8 (Thermo Fisher) 1:800, anti-amyloid beta clone 4G8 (Biolegend) 1:1,000, pTau-Thr217 (Thermo Fisher) 1:6,400, P62-lck (BD Biosciences) 1:1,000, anti-alpha-synuclein (phospho-S129) (Abcam) 1:500, and anti-pTDP-43 Ser409/410 (Cosmo Bio) dilution 1:6,000 and anti-TMEM106B (C terminal, Sigma-Aldrich) dilution 1:1,000. Envision mouse/rabbit HRP (DAKO) was used in the secondary detection step, and 3,3′-diaminobenzine (DAKO) was used as a chromogen. Immunostained sections were counterstained using haematoxylin, dehydrated using alcohol and xylene, and coverslipped using Depex.

Immunohistochemistry and confocal fluorescence microscopy

Free-floating (200 μm) acute brain slices were incubated for 1 h in carboxygenated NMDG buffer to which 15 μM MX04 was diluted. Next, the slices were transferred to fresh NMDG buffer (93 mM NMDG, 2.5 mM potassium chloride, 1.2 mM sodium hydrogen carbonate, 20 mM HEPES, 25 mM glucose, 5 mM sodium ascorbate, 2 mM thiourea, 3 mM sodium pyruvate, 10 mM magnesium sulfate heptahydrate, 0.5 mM calcium chloride dihydrate, pH 7.4, 300–315 mOsmol)56 and fixed with 4% v/v paraformaldehyde. Slices were permeabilized with 2% v/v Triton X-100 for 30 min and incubated for 1 h in blocking buffer (3% w/v BSA, 0.1% v/v Triton X-100, 50 mM Tris-HCl, 150 mM NaCl, pH 7.4) at room temperature. To detect β-amyloid and tau inclusions, slices were incubated in 1:750 dilution 6E10 (Biolegend, catalogue no. 803001) or 1:750 dilution 4G8 (Biolegend, catalogue no. 803001) and 1:750 dilution AT8 (Thermo Fisher) in blocking buffer at 4 °C for 16 h, respectively. After three washes in TBS (50 mM Tris-HCl, 150 mM NaCl, pH 7.4) for 5 min each, slices were incubated in 1:1,000 diluted antimouse IgG2b AF-633 (Thermo Fisher, catalogue no. A21126) and 1:1,000 diluted antimouse IgG1 AF-568 (Thermo Fisher) in blocking buffer for 2 h at room temperature. Following three washes in TBS for 5 min each, slices were mounted with Vectashield (Vector Laboratories) on a microscope slide (Erpredia, catalogue no. J1810AMNZ). Images were acquired with a confocal laser scanning microscope (ZEISS LSM 700) using a ×10/0.3 and a ×63/1.4 numerical aperture (NA) air objective lens, with frame sizes of 1,024 × 1,024 and 512 × 512 pixels, respectively. MX04, AF-568 and AF-633 were detected with excitation and emission maxima of 405 and 435, 579 and 603 and 639 and 669 nm, respectively.

Sarkosyl-insoluble tau purification and immunoblotting

Sarkosyl-insoluble tau purification followed a previously published protocol57. In brief, freeze–thawed postmortem brain tissue was homogenized in 10 vol (w/v) of homogenization buffer (10 mM Tris-HCI (pH 7.4), 0.8 M NaCI, 1 mM EDTA, 10% w/v sucrose). The homogenate was centrifuged at 20,000g for 20 min, at 4 °C and the supernatant was retained. The pellet was rehomogenized in 5 vol (w/v) of homogenization buffer and recentrifuged. Both supernatants were combined, brought to 1% iN-lauroylsarcosinate (w/v) and were centrifuged at 100,000g for 1 h at 21 °C. The sarkosyl-insoluble pellets were resuspended in 50 mM Tris-HCl, pH 7.4 (0.2 ml per g of starting material) and stored at 4 °C for immunoblots. Samples were analysed using 4–12% Bis-Tris gels (Thermo Fisher) and transferred onto polyvinyl difluoride (PVDF) membranes using iBlot gel transfer stacks (Thermo Fisher). The PVDF membrane was blocked (2.5% w/v casein in 0.1% v/v Tween 20, 50 mM Tris-HCl pH 7.4, 100 mM NaCl) for 1 h at room temperature. The following primary antibodies were diluted in 1.25% w/v casein in TBS-T (0.1% Tween 20, 50 mM Tris-HCl pH 7.4, 100 mM NaCl): 1:2,000 Tau 46 (amino acids (aa) 404–441, T9450, Merck), 1:1,000 AT8 (pS202/pT205 Tau, MN1020, Thermo Fisher), 1:1,000 4-repeat tau (aa 275–291, catalogue no. 05-804, Merck), 1:500 3-repeat tau (aa 267–316, catalogue no. 05-803, Merck) and 1:1,000 C-terminal domain TMEM106B (Merck, catalogue no. SAB2106778). The PVDF membranes were incubated with primary antibodies at 4 °C overnight. The membranes were washed five times with TBS-T for 5 min, followed by incubation with secondary antibody for 40 min at room temperature, then washed five times in TBS-T for 5 min. PVDF membranes with ECL reagent (Lumigen) were imaged on an iBright 1500 (Thermo Fisher).

Single-particle cryo-EM structure of sarkosyl-insoluble tau

The purification of sarkosyl-insoluble tau was as previously described58. In brief, 0.435 g of postmortem brain (cingulate gyrus) tissue was homogenized in 20 vol (v/w) of homogenization buffer (10 mM Tris-HCI (pH 7.4), 0.8 M NaCI, 1 mM EGTA, 10% w/v sucrose). The homogenate was brought to 2% w/v sarkosyl, incubated for 30 min at 37 °C and then centrifuged at 10,000g for 10 min, at 4 °C. The supernatant was retained and centrifuged at 100,000g for 25 min at 4 °C. The sarkosyl-insoluble pellets were resuspended in 700 μl g−1 extraction buffer (per gram of tissue) and centrifuged at 5,000g for 5 min at 4 °C. The supernatant was diluted threefold in 50 mM Tris-HCl, pH 7.4, containing 0.15 M NaCl, 10% w/v sucrose and 0.2% w/v sarkosyl, and spun at 166,000g for 30 min at 4 °C. The pellets were resuspended in 300 μl g−1 EM buffer (20 mM Tris-HCl, pH 7.4, 100 mM NaCl).

The sample (4 µl) was applied to Quantifoil R1.2/1.3 (300 mesh) grids after a 60 s plasma cleaning step (Tergeo, Pie Scientific). Grids were blotted and plunge-frozen in liquid ethane using a Vitrobot Mark IV (FEI) with the chamber maintained at close to 100% humidity and 6 °C. The cryo-EM dataset was collected using EPU v.3.0 (Thermo Fisher) at the Astbury Biostructural Laboratory (University of Leeds) using a Titan Krios electron microscope (Thermo Fisher) operated at 300 kV with a Falcon4i detector in counting mode. A nominal magnification of ×96,000 was set yielding a pixel size of 0.83 Å. A total of 10,860 videos were collected with a nominal defocus range of −1.5 to −2.7 µm and a total dose of roughly 44 e2 over an exposure of 4 s, corresponded to a dose rate of roughly 7.6 e/pixel s−1.

The raw EER videos were initially compressed and converted to TIFF using RELION v.4.059, regrouped to give 38 frames with a dose per frame of 1.2 e2. The TIFF stacks were aligned and summed using motion correction (MotionCorr2 v.1.2.1)60 in RELION (Extended Data Fig. 1c) and contrast transfer function (CTF) parameters were estimated for each micrograph using CTFFIND v.1.1461. Tau fibrils from roughly 100 micrographs were picked manually and used to train a picking model in crYOLO v.1.9.662 for automated picking with an inter-box spacing of 3× layers (roughly 14 Å). Next, 321,041 segments were extracted 2× binned with roughly 560 Å2 box dimensions. Two rounds of two-dimensional (2D) classification were performed to remove picking artefacts, with all classes corresponding to fibrils kept (Extended Data Fig. 1d) yielding 279,590 segments for further processing. An initial 3D template was generated from a PHF-like 2D class average and an estimated helical twist from measured crossover lengths (roughly 80 nm) using the relion_helix_inimodel2d command63. The first 3D classification was run using all of the 2× binned segments, with a sampling of 1.8° and strict high-resolution limit of 6 Å (Extended Data Fig. 1e), from which two classes presented tau PHF folds (72% of segments in total, the more ordered class containing 24% of segments was selected for further processing) and one class presented a tau SF fold (14% of segments). Each subset was extracted unbinned (336 pixel2, 276 Å2) and further classified with 0.9° sampling, without the high-resolution limit and with local searches of the helical twist (Extended Data Fig. 1f). The helical rise was set to 2.4 Å for the PHF subset and 4.8 Å for the SF subsets, on the basis of known structural data. The SF subset map improved to show a backbone fold that was identical to published tau SF structures24, but did not contain enough segments to refine to a high-resolution structure. The PHF subset resolved to give a 3.0 Å (gold-standard, Fourier shell correlation (FSC) of 0.143) refined map (Extended Data Fig. 1g–i) after CTF refinement and Bayesian polishing, with a sharpening value of −57 Å2 applied during postprocessing. The final refined helical rise and twist values were 2.405 Å and 179.44°, respectively (Supplementary Table 8).

Model building of postmortem donor tau PHF cryo-EM structure

A published tau PHF fibril structure (Protein Data Bank (PDB) 5o3l)24 was docked into the refined cryo-EM map and one chain was adjusted to fit into the density using real-space refine in Coot v. The chain was duplicated and docked into the density to create five layers of dimeric tau polypeptide chains. The model was real-space refined using Phenix v.1.17.165 with noncrystallographic symmetry restraints applied to limit inter-chain divergence. The quality of the final model was assessed using MolProbity v.4.5.266. The final model was near identical to the template model, barring slight side chain adjustments and discrepancies in pixel size and/or magnification, with a root mean-squared deviation of 0.67 Å between a chain of each model for 73 Cα positions (Supplementary Table 8). The model was therefore not deposited to the PDB, as several identical models of tau PHF are already present24.

MX04-labelling and high-pressure freezing freeze–thawed postmortem acute brain slices

Flash-frozen postmortem brain samples were thawed at room temperature for 5 min, then placed in ice-cold carboxygenated NMDG buffer (93 mM NMDG, 2.5 mM potassium chloride, 1.2 mM sodium hydrogen carbonate, 20 mM HEPES, 25 mM glucose, 5 mM sodium ascorbate, 2 mM thiourea, 3 mM sodium pyruvate, 10 mM magnesium sulfate heptahydrate, 0.5 mM calcium chloride dihydrate, pH 7.4, 300–315 mOsmol)56. Then 100–200 μm slices of postmortem AD and control donor brain were sliced along the horizontal or coronal plane (speed 0.26 mm s−1) using a vibratome (catalogue no. VT1200S, Leica) in ice-cold carboxygenated NMDG buffer (roughly 30 min). Next, postmortem AD and non-demented control acute brain slices were incubated in carboxygenated NMDG buffer, to which 15 μM MX04 was diluted for 1 h at room temperature before slices were washed three-times in carboxygenated NMDG for 5 min each. Grey matter biopsies (2 mm diameter) were incubated in cryoprotectant (5% w/v sucrose and 20% w/v dextran 40,000 in NMDG buffer) for 30 min at room temperature. Then, 100-μm-deep wells of the specimen carrier type A (Leica, catalogue no. 16770152) were filled with cryoprotectant, and the tissue biopsies were carefully placed inside to avoid tissue damage. They were then covered with the flat side of the lipid-coated specimen carrier type B (Leica, catalogue no. 16770153) and HPF (roughly 2,000 bar at −188 °C) using a Leica EM ICE.


HPF samples were imaged using a cryogenic-fluorescence microscope (Leica EM Thunder) with a HC PL APO ×50/0.9 NA cryo-objective, Orca Flash 4.0 V2 sCMOS camera (Hamamatsu Photonics) and a Solar Light Engine (Lumencor) at −180 °C. A DAPI filter set (excitation and bandwidth 365 and 50, dichroic 400; emission and bandwidth, 460 and 50) was used to detect MX04-labelled amyloid. A rhodamine filter set (excitation and bandwidth 546 and 10, dichroic 560; emission and bandwidth, 525 and 50) was used as a control imaging channel. The images were acquired with a frame size of 2,048 × 2,048 pixels. Tile scans of HPF carriers were acquired with 17% laser intensity for 0.1 s. z stacks of ultrathin cryo-sections were acquired with 30% intensity and an exposure time of 0.2 s. Images were processed using Fiji ImageJ.


HPF sample carriers were transferred to a cryo-ultramicrotome (Leica EM FC7, −160 °C) equipped with trimming (Trim 20, T399) and CEMOVIS (Diatome, cryo immuno, catalogue no. MT12859) diamond knives. A trapezoid stub of tissue measuring 100 × 100 × 60 μm was trimmed, which contained the target amyloid. Cryo-sections (70 nm thick) were then cut at −160 °C with a diamond knife (Diatome, cryo immuno, catalogue no. MT12859) and adhered onto a glow discharged (Cressington glow discharger, 60 s, 1 × 10−4 mbar, 15 mA) 3.5/1, 300 mesh Cu grid (Quantifoil Micro Tools) using a Crion electrostatic gun and gold eyelash micromanipulators.

Cryo-CLEM of cryo-sections

The location of amyloid plaques in tissue cryo-sections was assessed by cryogenic-fluorescence microscopy on the basis of MX04 fluorescence (excitation 370 nm, emission 460–500 nm). Grid squares that showed a signal for MX04 were selected for cryoET. The alignment between cryo-FM images and electron micrographs were carried out using a MATLAB script67,68, in which the centres of ten holes in the carbon foil surrounding the region of interest were used as fiducial markers to align the cryo-FM and cryo-EM images.

Cryo-CLEM and cryo-FIB-SEM of lift-out lamellae

Carriers containing HPF samples for cryo-FIB-SEM lift-out were transferred to a cryo-ultramicrotome (Leica EM FC7, −160 °C) equipped with a Trim 45 T1865 diamond knife. The surface of the carrier and HPF tissue within were trimmed to remove surface ice contamination and the top layer of vibratome-damaged tissue69. To achieve this, three 300-μm-wide steps were trimmed back on four sides of the carrier: the outermost step was trimmed roughly 30 μm deep, the next was trimmed roughly 20 μm back and the innermost step was trimmed roughly 10 μm back. The resulting protruding square of tissue was then trimmed roughly 2–5 μm back to achieve a smoother less contaminated surface for cryo-FIB-SEM and lift-out.

The cryo-CLEM workflow was performed on a ZEISS Crossbeam 550 FIB-Scanning Electron Microscope equipped with the Quorum cryo-system, ZEISS Cryo-accessories tool kit and Omniprobe 350 cryo-micromanipulator (Oxford Instruments) and operated at 30 keV. For cryo-FM a Zeiss LSM (laser scanning microscope) 900 based on an upright Axio Imager stand equipped with AiryScan and Linkam cryo-stage was used.

At the Quorum prep desk, the HPF sample carrier (Leica Microsystems) was mounted on the corresponding Zeiss universal sample holder (USH, Zeiss cryo-accessories tool kit). First, the Zeiss USH was placed on the Zeiss adaptor for the Linkam cryo-stage used for performing cryo-FM. The assembly was transferred into the Linkam cryo-stage in liquid nitrogen using the Zeiss transfer box. A LM ZEN Connect project was acquired using the plugin ZEN Connect of the ZEN Blue v.3.6 software (Zeiss Microscopy). One overview image of the sample and the holder was acquired with a ×5, 0.2 NA C-Epiplan Apochromat air objective with an Axiocam 503. MX04, reflection (for correlation with the EM images) and a control channel were imaged with 385, 511 and 567 nm LEDs, respectively, in combination with a quad-band filter (Excitation BP 385 ± 15 nm, BP 469 ± 19 nm, BP 555 ± 15 nm, BP 631 ± 16.5 nm, Emission QBP 425 ± 15 nm, 514 ± 15 nm,  592 ± 12.5 nm, and 709 ± 50 nm). The image was acquired with a frame size of 2.79 × 2.10 mm and 1,936 × 1,460 pixels; 5, 3 and 60% LED power and 50, 6 and 300 ms exposure time for MX04, reflection and control, respectively.

For sample quality assessment, a large 3D volume (1.27 × 1.27 × 0.108 mm, 2,824 × 2,824 pixels, 49 z slices) was scanned with a 10×, 0.4 NA C-Epiplan Apochromat air objective in confocal mode. MX04 (0.04% laser power, 405 nm laser, detection window 410–546 nm, pixel dwell time 0.74 µs) and the reflection (0.01% laser power, 640 nm laser, detection window 630–700 nm, pixel dwell time 0.74 µs) were imaged.

Regions of interest were scanned with a 100×, 0.75 NA LD EC Epiplan-Neofluar air objective (125.15 × 125.15 × 19.8 µm and 1,140 × 1,140 pixels, 37 z slices, pixel dwell time 1.8 µs) with the following settings: MX04 (0.2% laser power, 405 nm laser, detection window 410–544 nm) and reflection (0.05% laser power, 640 nm, detection window 639–700 nm). A linear deconvolution was run on all confocal images (10× and 100×) using the Zeiss LSM plus plugin.

Airyscan z stacks were acquired from the regions of interest using the Airyscan 2 detector and the ×100 objective detailed above. MX04 was detected using the 405 nm laser line (1.5% laser power, 405 nm laser, pixel dwell time 2.18 µs, detection window 400–650 nm, image size 62.31 × 62.31 × 13.65 µm and 946 × 946 pixels, 36 z slices). Airyscan images were processed using the Airyscan joint deconvolution (jDCV) plugin of the Zeiss ZEN blue software. Cryo-CLEM alignment was performed with maximum projections of confocal and Airyscan image z stacks prepared in Zeiss ZEN blue. After finalizing light microscopy, the USH connected to the Linkam adaptor was transferred into the Quorum prep box. The USH was detached from the Linkam adaptor and mounted on the ZEISS cryo lift-out sample holder with the USH in flat orientation and a mounted upright, standing half-moon Omniprobe grid clipped into an AutoGrid (Thermo Fisher). Using the Quorum cryo-shuttle the sample holder was transferred into the Quorum prep chamber attached to the Crossbeam 550 FIB-SEM. The temperatures of the cryo-stage and anticontaminator in the main and prep chamber were set to −160 °C and −180 °C, respectively. The sample was sputter-coated with platinum for 45 s (5 mA current). After the sputter coating, the sample was transferred on the Quorum cryo-stage in the main chamber.

Cryo-FM-targeted cryo-FIB-SEM lift-out was carried out driving the Zeiss Crossbeam FIB-SEM stage within a ZEN Connect imaging project, in which cryo-FM, SEM and FIB images were correlated using surface features (cryoplaning markings and ice contamination) as fiducial markers70 (Extended Data Fig. 11). At normal view (0° stage tilt), 5 mm working distance and 2.3 kV acceleration voltage an overview of the HPF carrier and a zoomed-in SEM image were acquired and loaded into a new SEM session of the existing ZEN Connect project (imaging parameter SE2, Everhart-Thornley) detector, 98 pA SEM current, 4,096 × 3,072 pixels, 800 ns dwell time, line average with 23 iterations and 800 and 420 nm pixel sizes, respectively. The SEM session was correlated with the already acquired cryo-FM session by using the reflection mode channel images for alignment.

As image navigation was desired in FIB mode, the stage was tilted to 54° allowing normal FIB view. The sample was brought into the coincidence point of SEM and FIB with a 5 mm working distance and an overview FIB image (imaging current 50 pA, SE2 detector, 2,048 × 1,536, 1,6 µs dwell time, pixel average, 300 nm pixel size) was taken for FIB session alignment with the former SEM/cryo-FM session. The coincidence point was fine adjusted to the region targeted for lift-out and the milling box for coarse cross-sectioning was positioned on the basis of the alignment between cryo-FM and FIB images. A roughly 80-μm-wide, 35-µm-high and 30-μm-deep trapezoidal cross-section was milled from the front side using a 30 nA FIB probe. As ice contamination, especially on top of region of interest, was observed, the sample holder was transferred back into the Quorum prep box for cleaning. Under cryogenic conditions the sample surface was cleansed by using a brush and by wiping using a swab. After another sputter coating in the Quorum prep chamber (see above for parameters), a cold deposition of platinum precursor was applied in the Crossbeam main chamber. For cold deposition, the distance between sample and gas injection capillary was about 3 mm and the gas reservoir valve was opened for 45 s. The gas reservoir temperature was about 28 °C (unheated gas reservoir state). Using the saved stage position, the cross-section at the region of interest was targeted, the coincidence point alignment was checked. An FIB image was taken and aligned with the former FIB session using the already milled cross-section as reference.

Next, a 30 nA FIB probe was used to mill a second corresponding cross-section from the back side and a roughly 60-μm-wide left side cut that left a roughly 80-μm-wide, 10-μm-thick, 30-μm-deep tissue chunk attached on to its right side. At the front side, the cross-section was further polished using a 15 nA FIB probe. The stage was tilted to 10° tilt for milling a roughly 80-μm-wide L-shaped undercut, leaving a small connection on the left side (7 nA FIB probe). As a lift-out tool, a roughly 5-µm-thick copper block was attached to the Omniprobe manipulator tip. The stage was at 10° tilt to allow access of the micromanipulator while bypassing the AutoGrid ring. Before lift-out, the roughly 5 μm copper block was attached to the right side of the tissue chunk using redeposition milling of copper material (three 2.5 × 5 μm milling windows with 700 pA FIB probe and 140 mC cm−2 dose). Next, the tissue chunk was cut free from the left side to achieve lift-out. At 10° stage tilt the roughly 80-μm-wide chunk was attached by redeposition milling to a half-moon EM grid (Omniprobe) clipped into an AutoGrid and cut in half, leaving the distal roughly 40 μm chunk attached to the EM grid. The remaining proximal roughly 40 μm chunk attached to the Omniprobe was attached to a second location on the EM grid. The stage was tilted to 56° before two 8–10-μm-wide, roughly 15-μm-deep lamella windows were milled in each chunk half. At 56° stage tilt, the angle between tissue chunk and/or grid plane and FIB beam is 2° allowing to bypass the outer AutoGrid ring for lamellae thinning. The lamellae were sequentially thinned from both sides to roughly 2 µm, then roughly 1 μm, then roughly 500 nm and finally 130 to 200 nm using 700, 300, 100 and 50 pA FIB probes, respectively. Each lamella window was framed with unmilled tissue at the left, right and bottom sides.

Cryo-electron tomography

Cryo-electron tomography was performed using a Thermo Fisher 300 keV Titan Krios G2, X-FEG equipped with a Falcon4i detector and Selectris energy filter in the Astbury Biostructure Laboratory at the University of Leeds. A dose symmetric tilt scheme71 was implemented using Tomo5.15 (Thermo Fisher) to collect tilt series from −60° to +60° in 2° increments with a 100 μm objective aperture and 5–6.5 μm defocus. Each tilt increment received roughly 2.3 s of exposure (fractionated into eight frames) at roughly 2 e2 per tilt for a total dose of roughly 120 e2 per tilt series with a pixel size of 2.38 Å.

In AD samples, locations were chosen for the collection of tilt series in three different ways: (1) the presence of MX04 fluorescence, (Extended Data Fig. 3a), (2) the appearance of dense filamentous structures that resembled amyloid filaments in medium magnification micrographs (Extended Data Fig. 3b) and (3) areas surrounding MX04 fluorescence or other tissue areas in which membrane compartment features could be seen in medium magnification micrographs (Supplementary Tables 1, 2 and 4). Control non-AD samples lacked both MX04 signal (Extended Data Fig. 2a,b) and filamentous structures in medium magnification micrographs. Consequently, tilt series locations were picked in areas with visible membrane compartments in medium magnification micrographs (Supplementary Table 3).

PMI and freeze–thaw step on mammalian brain tissue and amyloid architecture

We previously reported the architecture of Aβ-plaques in the AppNL-G-F mouse model by cryoET30. In contrast to postmortem tissues with a 6 h PMI and freeze–thaw step, AppNL-G-F tissues were cardiac perfused in NMDG buffer56, were not freeze thawed and had a much shorter PMI30 (hereon referred to as AppNL-G-F-HPF samples). To control for the effect of 6 h PMI and a freeze–thaw step on the molecular architecture of tissues, we prepared AppNL-G-F mouse (c57b/l6 background) tissues for cryoET under similar conditions as postmortem AD donor tissues, with a 6 h PMI and freeze–thaw step (hereon referred to as AppNL-G-F-PMI-FT-HPF samples). Animals were treated in accordance with the UK Animal Scientific Procedures Act (1986) and National Institutes of Health guidelines. Oversight and approval was provided by the University of Leeds Animal Welfare and Ethics Review Board and licensed by the UK Government Home Office.

To prepare AppNL-G-F-PMI-FT-HPF, a 10-month-old male AppNL-G-F mouse received intraperitoneal injection of 5 mg kg−1 MX04 24 h before culling30. The carcass was left at room temperature for 6 h to mimic the postmortem delay before forebrain was collected and flash-frozen in liquid nitrogen. Next, forebrain was taken through the workflow for postmortem human tissue (Fig. 1c) including a roughly 5 min thaw step preceding acute slice preparation in ice-cold carboxygenated NMDG buffer, 1 h incubation in carboxygenated NMDG buffer with 15 µM MX04, 3 × 5 min washes in carboxygenated NMDG buffer, and a 30 minute incubation in cryoprotectant (5% w/v sucrose and 20% w/v dextran 40,000 in NMDG buffer) at room temperature before high-pressure freezing. A MX04-labelled β-amyloid plaque was located within HPF tissue by cryo-FM, from which roughly 70-nm-thick cryo-sections were collected and attached to EM grids. Tissue sections were imaged by cryo-FM to locate MX04-labelled β-amyloid plaques before tomographic tilt series were collected from 14 and 46 locations with and without MX04 cryo-CLEM, respectively. Data collection parameters and constituents of these tomograms are detailed in SupplementaryTable 7.

All MX04-labelled β-amyloid plaque tomograms (14 out of 60) from AppNL-G-F-PMI-FT-HPF cryo-sections contained β-amyloid fibrils arranged in parallel bundles or a mesh (Extended Data Fig. 5). This fibril architecture was indistinguishable from that previously reported for AppNL-G-F-HPF plaques that were HPF directly without a 6 h PMI and freeze–thaw step30 (Extended Data Fig. 5).

We next compared non-amyloid constituents of amyloid plaques and subcellular compartments of the tissue that surrounded amyloid plaques in AppNL-G-F-PMI-FT-HPF versus previously published AppNL-G-F-HPF30. (1) In contrast to AppNL-G-F-HPF tomograms30, microtubules were absent in AppNL-G-F-PMI-FT-HPF tomograms. This was expected because microtubules rapidly depolymerize at cold temperatures or when a source of nucleotide triphosphates is compromised72,73. Microtubules were also absent in postmortem AD and non-demented donor tomograms. (2) In contrast to AppNL-G-F-HPF, AppNL-G-F-PMI-FT-HPF tomograms contained a subset of mitochondria with swollen cristae and a depleted mitochondrial matrix (Extended Data Fig. 6c,d and Supplementary Table 7). This architecture is indicative of necrotic and/or apoptotic respiratory collapse74,75. Similarly damaged mitochondria were also observed in postmortem AD and non-demented donor tomograms (Extended Data Fig. 6a,b and Supplementary Table 7). (3) In contrast to AppNL-G-F-HPF, 2 out of 60 AppNL-G-F-PMI-FT-HPF tomograms contained a membrane fragment, suggesting the PMI and/or freeze–thaw step may produce membrane fragments in brain tissue that contains amyloid filaments. Burst plasma membrane were also observed in 10 out of 88 postmortem AD tomograms but were absent in 64 non-demented control postmortem tomograms. (4) In contrast to AppNL-G-F-HPF, 10 out of 60 AppNL-G-F-PMI-FT-PMF tomograms contained a burst plasma membrane, suggesting the PMI and freeze–thaw step can disrupt membrane integrity in these mice (Extended Data Fig. 5). Burst plasma membranes were observed in 11 out of 80 postmortem AD tomograms but absent in 64 non-demented control postmortem tomograms.

Tomographic reconstructions and subtomogram averaging

Subtomogram alignment and averaging of Aβ fibrils and tau filaments was initially performed on a per-tomogram basis to assess the relationship between amyloid structure and its subcellular context. Because of the dense architecture of in-tissue amyloid, only tomograms with axial (oriented in the z axis of the reconstructed tomogram) fibrils and filaments provided sufficient contrast to accurately pick subvolumes for subtomogram averaging.

Dose-fractionated video frames were imported into Warp (v1.1.0b)76 for frame alignment and initial estimation of the CTF. Tilt series stacks were generated in Warp and imported into etomo IMOD (v.4.12.35)77,78 for fine alignment using patch tracking. Lower quality aligned frames were excluded from stack generation before import. Fine aligned tilt series were then imported back into Warp and 3D-CTF corrected tomograms were reconstructed at a pixel size of 9.52 Å (bin4). See Supplementary Tables 5 and 6 for per-tomogram subtomogram averaging details of tau filaments and Aβ fibrils, respectively.

For tau filaments, subtomogram averaging was performed on nine cryoET volumes (seven from cryo-sections and two from lift-out lamellae). Between 55 and 278 tau filaments (24 × 100 × 24 box size) were manually picked from 4× binned (9.52 Å voxel size) tomographic reconstructions in 3dmod (IMOD v.4.12.35). Coordinates of each filament were represented as a two-point contour with respective ‘head’ and ‘tail’ model points positioned at the poles of each filament. The slicer function of 3dmod was used to rotate and translate each tau filament, ensuring different models were approximately centred along the filament axis. Using these model files as input, a script invoking the particle estimation for electron tomography (PEET) ‘stalkInit’ command was run in default mode to generate new, single-point model files (containing coordinates of the head, centroid and tail), initial motive lists and rotation axes files, to supply for alignment and averaging in PEET (v.1.17.0a)79,80. To minimize alignment to the missing wedge and to verify the accuracy of model point coordinates, initial averages were generated by restricting rotational and translation searching from the centroid of the de novo reference filament axis. In subsequent PEET alignment iterations, rotational and translational searching was performed, using the updated centroid coordinates and orientations as input. Where visible improvements in map resolution were observed, createAlignedModel was run to generate new models, motive lists and rotation axes containing the updated locations and orientations of particles from the alignment. Cylindrical masks with blurred edges were applied to the stalkInit reference volume during subvolume alignment. StalkInit subtomogram averages were generated without applying symmetry or helical reconstruction methods. The helical twist of tau filaments in each tomogram was measured from PEET stalkInit averages.

To improve alignment further, centroid models were added every 1 voxel along the stalkInit filament axis (99,664 new model coordinates from the nine cryoET volumes) using the ‘addModPts’ command. AddModPts models were used as new inputs for subvolume alignment and averaging in PEET. CreateAlignedModel was also run after addModPts alignments, outputting new models and initial motive lists of the updated locations and orientations of particles.

After initial alignment iterations in PEET, coordinates and orientations were imported into RELION v.4.059 (toRelionCoords) for further refinement with Warp bin4 tomographic reconstructions59. This was achieved using the ‘imodinfo’ command with the ‘-l’ option to output the model information to a text file. Helical prior information was generated using a custom python script (available from GitHub at to modify the STAR file to contain rlnHelicalTubeID, rlnHelicalTrackLength, rlnAngleTiltPrior (set to 0), rlnAnglePsiPrior (set to 0) and rlnAnglePsiFlipRatio (set to 0.5). This modified STAR file (generated from ‘toRelionCoords’ and custom python scripts) was used in Warp to generate subtomograms (24 × 24 × 24 box size, 9.52 Å) that were then extracted for helical 3D auto-refinement and classification in RELION (using helical twist estimated from PEET stalkInit averages and without C2 symmetry). Refined subtomograms were then re-extracted in Warp at pixel sizes of 4.76 Å (48 × 48 × 48 box size) and followed by helical 3D auto-refinement and classification. Final rounds of helical 3D auto-refinement and 3D classification were performed on subtomograms re-extracted in Warp at pixel sizes of 2.38 Å (96 × 96 × 96). Final averages were sharpened using RELION postprocessing and global map resolutions were estimated at a FSC of 0.143 between two independently refined half-maps. Refined particle coordinates and orientation from RELION 3D auto-refinement that reached subnanometre resolution (Fig. 3h) were imported into M (v.1.0.9)81 for multi-particle refinements.

For Aβ fibrils, subtomogram averaging was performed using two cryo-section volumes collected from MX04-labelled β-amyloid plaques that contained axially oriented fibrils. The same manual picking and subtomogram averaging procedures were followed (using stalkInit) as described for tau filaments. Initially, 100 fibrils were picked from one tomogram, with fibrils represented as two-point ‘head’ and ‘tail’ models, positioned at opposite ends of each fibril. This model was used as input to the PEET ‘stalkInit’ command to generate new model files (head, centroid, tail), initial motive lists and rotation axes files, for alignment and averaging in PEET. The subtomogram average of 100 fibrils generated a featureless, smooth tube, with reducing detail observed with increasing fibril numbers used for averaging, a possible indication of particle heterogeneity (Extended Data Fig. 10). To assess fibril heterogeneity, the widths of individual fibrils were manually measured in IMOD. On the basis of average fibril width, three subpopulations of fibrils were picked (3–5 nm protofilament-like rods, 4–9 nm fibrils and 6–12 nm wide fibrils). Subtomogram averaging of 20 protofilament-like rods, 42 fibrils and 42 wide fibrils from two tomograms was performed with PEET as before (Supplementary Table 6 and Extended Data Fig. 10).

To compare the structural similarity of in situ subtomogram averages with ex vivo purified tau conformers, available AD tau filament atomic models (PDB 5o3l and 5o3t) were fitted into half-maps with EM placement v.1.2.237,38 in ChimeraX v.1.582,83 giving LLG scores of fit (Figs. 3h and 4h,j, Extended Data Fig. 9 and Supplementary Table 5). When running EM placement, the map resolution was set to the value determined using the 0.143 half-map FSC threshold. Theory and experience with single-particle cryo-EM data suggest that LLG scores of 60 or greater indicate a non-random fit, with higher values being observed for more accurate models and higher resolution maps38. The LLG score is related to how much of the information content of the map is explained by the model. Because the total information content of the portion of the map covered by the model depends both on its local (including anisotropic) resolution and its volume, an LLG score above the threshold can be reached by any combination of map quality with model size and quality. The confidence threshold itself (LLG 60) is not expected to vary with map resolution or issues arising from preferential orientations38. For reference, fitting a PHF atomic model24 into a 3 Å resolution single-particle cryo-EM (Extended Data Fig. 1g) half-maps gave LLG 11,889 and correlation coefficient 0.66. Atomic model (PHF or SF)24 versus subtomogram averaged map Fourier shell correlation was also calculated with Phenix v.1.17.0a65, avoiding artefacts at the edges of the averaged map by restricting the comparison to a central sphere with a radius of 80 Å (Extended Data Fig. 13).

Initial subtomogram averaging (without helical symmetry in PEET) and after-helical averaging in RELION were assessed by visualizing Euler angles in ChimeraX using the ArtiaX v.0.4.784 indicating good overall agreement. Filaments were assigned a polarity orientation in which those with more than or equal to an 80% Z Euler angle were oriented in the same direction (Extended Data Fig. 12c–f).

Preparation of cryoET figures

All tomographic slice figures were one voxel thick prepared from 4× binned (9.52 Å voxel size) tomographic reconstructions generated with MotionCorr2 v.1.2.160 aligned frames that were reconstructed with AreTomo v.1.3.085 and deconvolved with Isonet v.0.286. Segmenting membranes within tomograms was performed in Dynamo v.1.1.53287 using the manual surface modelling tool. Dynamo tables containing coordinate information of the tomogram membrane models were converted into CMM files and visualized in ChimeraX. The neural network-based tomogram segmentation pipeline in EMAN2 v.2.9988 was used to segment tau filaments and Aβ fibrils. In Fig. 3c–h, regions of the raw tomographic volumes containing β-amyloid fibrils and tau filaments were trimmed using the IMOD rubber band tool and trimvol command. The raw tomographic density was then visualized in ChimeraX using the segger tool to colour each Aβ fibril/tau filament by connectivity.

Annotation of constituents in tomographic volumes

The constituents of tomographic volumes of tissue cryo-sections from β-amyloid plaques, tau pathology, non-demented control and cryo-FIB-SEM lift-out of tau threads were detailed in Supplementary Tables 1, 2, 3 and 4, respectively. Constituents were initially identified by two curators independently. Next, two curators consolidated and verified each annotation. The boundary of intracellular and extracellular regions of tomographic volumes were determined by the presence of myelinated axons or by lipid bilayers enclosing intracellular organelles within each tomographic volume and the corresponding electron micrograph used for cryo-CLEM. The following constituents were identified: (1) amyloid fibrils or filaments were assigned on the basis of MX04 cryo-CLEM labelling and rod shape: fibrils and filaments were cross-checked for their absence in tomographic volumes reconstructed from non-demented control brain donors; (2) subcellular compartments, defined by plasma membrane containing membrane-bound organelles or a higher tomographic density than the extracellular space (excluding amyloid fibrils), that is, consistent with the higher concentration of proteins in the cell cytoplasm compared to the extracellular space; (3) mitochondria, defined by the double membrane including outer membrane and inner mitochondrial cristae; (4) putative F1FO-ATPases were identifiable within the inner mitochondrial membrane (for example, see Fig. 2c); (5) damaged mitochondria, double membrane-bound compartments with swollen cristae and a less densely packed mitochondrial matrix than other mitochondria to the extent that parts of the mitochondria seem empty (Extended Data Fig. 6); (6) myelinated axon, defined as five or more layers of 6–8 nm membrane lipid bilayer enclosing a subcellular compartment89; (7) open membrane sheets, defined by roughly 5-nm-thick membrane lipid bilayer sheets in interstitial space that did not form closed compartments; (8) burst compartment, membrane compartments that are open in the xy plane of the tomogram; (9) extracellular vesicles were defined as membranes that were closed and situated in extracellular spaces; (10) C-shaped vesicles were defined as cup-shaped membrane within the lumen of a vesicle; (11) multilamellar bodies were defined as 40–250 nm vesicle or subcellular compartment wrapped in three or more concentric rings of 4.5–6 nm membrane lipid bilayer; (12) extracellular droplets, 30–250 nm amorphous and smooth spheroidal droplets35,36 that were situated in extracellular locations; (13) extracellular cuboidal particles, 27–200-nm-diameter particles that were situated in extracellular locations (these particles contained striated layers of higher tomographic density: 2.5–2.8 nm apart, Extended Data Fig. 7); (14) vesicles were defined as closed spheroidal membranes; (15) F-actin, defined as roughly 7-nm-diameter filaments composed of a helical arrangement of globular subunits90 (F-actin was only observed intracellularly in a minor subset of both AD and non-demented control donor tomographic volumes); (16) ribosomes: spherical 25–30-nm-diameter dense particles; (17) microtubules, 25-nm-diameter tubes with 13 tubulin subunits and (18) knife damage, tissue cryo-sections contained regions in which the sample seems compressed, sometimes leaving a crevasse in the tissue that were readily identified as holes within the tissue39.

Statistics and reproducibility

All cryo-CLEM (Figs. 1g and 4b and Extended Data Figs. 3a,b and 11), tomographic slice images (Figs. 1h, 2a,b, 3 and 4e,f and Extended Data Figs. 3c–e, 4, 7 and 8a) and electron micrographs (Fig. 4d and Extended Data Figs. 1c and 3a,b) were representative of datasets from a single AD (42 β-amyloid plaque cryo-section, 25 tau pathology cryo-section, 13 tau pathology lift-out lamella tomograms, see Supplementary Tables 1, 2 and 4, respectively) and non-demented control (64 cryo-section tomograms, Supplementary Table 3) postmortem brain donor. All immunohistochemical imaging (Fig. 1a and Extended Data Fig. 1a), immunoblot (Fig. 1b and Extended Data Fig. 1b), immunofluorescence (Fig. 1c and Extended Data Fig. 2a) and cryo-FM imaging (Figs. 1e,f and 4c and Extended Data Fig. 2b–d) experiments were representative of 3–4 technical replicates of these donor tissues. tomographic slices in Extended Data Figs. 6 and 7 were representative of cryoET datasets of HPF (23 cryo-section tomograms, n = 2 biological replicates)30 and freeze thawed (60 cryo-section tomograms, n = 1 biological replicates, see Supplementary Table 7) AppNL-G-F knock-in mice.

Reporting summary

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

Source link