Human tissues. Fresh frozen human tissue samples of the inferior temporal gyrus were provided by the Massachusetts Alzheimer’s Disease Research Center (ADRC) with approval from the Mass General Brigham IRB (1999P009556) and with informed consent of patients or their relatives. In total, 7 human participants with AD and 9 controls were selected from the Massachusetts Alzheimer's Disease Research Center. For quantitative imaging experiments a sex-matched group (3F/3M) of donor tissues with and without AD pathology were selected for comparison. Additional samples were utilized for biochemistry and protocol optimization experiments. All samples were rigorously evaluated using standard histological methods by trained neuropathologists in the ADRC and detailed information for all donor tissues including sex, age at death, Thal stage, Braak stage, CERAD scores, post-mortem interval, APOE genotype, and comorbidities are listed in Table 1 [19].

Isolation of blood vessels. Blood vessels were isolated from 200 to 300 mg of frozen mice and human tissue. Brains were minced in 2 mm sections using a razor blade in ice-cold B1 buffer (Hanks Balanced Salt Solution with 10 mM HEPES, pH 7; Thermo Fisher Scientific). Then samples were manually homogenized using a Dounce homogenizer with 12 strokes. Homogenate was then transferred into a conical tube filled with 20 mL of B1 buffer and centrifuged at 2000g for 10 min at 4 ºC. Supernatant was discarded and the pellet was vigorously resuspended for 1 min in 20 mL of B2 buffer (B1 buffer with 18% dextran, Sigma-Aldrich) to remove myelin. Samples were centrifuged at 4400g for 15 min at 4 ºC. The myelin layer was carefully detached, and the pellet was resuspended in 1 mL of B3 buffer (B1 buffer with 1% Bovine Serum Albumin, BSA, Sigma-Aldrich). Afterward, homogenate was filtered through a 20 µm mesh (Millipore) previously equilibrated with 5 mL of ice-cold B3 solution. Brain blood vessels were rinsed with 30 mL of ice-cold B3 solution and then the blood vessels were detached from the filters by immersing them in 30 mL of B3 ice-cold solution. Vessels were centrifuged at 2000g for 5 min at 4 ºC. Finally, the pellet was resuspended in 1 mL of ice-cold B1 solution and again centrifuged at 2000g for 5 min at 4 ºC and the supernatant was discarded. Vessel-containing pellets were stored at − 80 ºC.

Protein assays. Protein was extracted from human and mice brain blood vessel homogenates, which were sonicated at 20% amplitude in 10 pulses in PBS supplemented with protease and phosphatase inhibitors (cOmplete Mini and PhosSTOP EASYpack; Roche). Then, samples were centrifuged at 3000g for 5 min at 4 ºC and the supernatant was discarded. Proteins were analyzed using a capillary-based electrophoresis instrument (SimpleWes, Biotechne). Three mg of protein were used per sample. Protein separation and detection were performed by capillary electrophoresis, and the binding of antibodies and HRP conjugated secondaries was done in the SimpleWes machine. Antibodies used were phospho-T181 (mouse 1:50, MN10050, Invitrogen), phospho-S202 (rabbit 1:25, 39357S, Cell Signaling), phospho-T217 (rabbit 1:25, 44–744, Invitrogen), phospho-T231 (rabbit 1:50, #44–746, Invitrogen), Tau13 (mouse 1:50, 835,201, Biolegend), Tau46 (mouse 1:50, 4019S, Cell Signaling) and total tau (rabbit 1:50, A0024, DAKO). Specific SimpleWes secondary antibodies HRP conjugated were acquired from the manufacturer (Biotechne). Protein quantification was analyzed in Fiji (https://doi.org/10.1038/nmeth.2019). The total intensity of signal in each lane was measured and normalized to the average of the three control samples.

Tissue slicing. Brain samples were placed in 4% paraformaldehyde (Thermo Fisher Scientific, cat No. 50980487) for 24 h at 4 °C. Tissue was then rinsed three times with 50 ml phosphate-buffered saline (PBS) for 10 min each, then placed in fresh PBS overnight at 4 °C and rinsed with fresh PBS. Fixed tissue underwent three rinsing cycles in 10-min increments using 50 ml of PBS, and then were placed in fresh PBS overnight at 4 °C. In preparation for tissue slicing, tissue was transferred to individual 35 mm Petri dishes and embedded in a gel block by pouring warm 4% agarose gel solution in PBS (4 g/100 ml) (Promega, cat No. V3121) over the tissue. The gel was then cooled to solidify and cut into a block to provide rigidity for cutting even slices. The tissue was secured on a vibratome (Leica Biosystems, VT1000 S Vibrating Blade Microtome) by super gluing the bottom of the agarose block. The vibratome was then used to slice 0.5–1 mm thick tissue sections. Each slice was then removed from the agarose through gentle manipulation with blunt forceps or paintbrushes and placed in a crosslinking solution, described below.

Delipidation. Tissue was then placed into sodium dodecyl sulfate (Sigma-Aldrich, cat No. L3771) 28.83 g/500 ml PBS-clearing solution supplemented with sodium borate (Sigma-Aldrich, cat No. S9640) on shaker at 100 rpm and 37 °C for ∼3 days. After delipidation, the brain slices were rinsed with 50 ml PBS five times over 24 h.

Immunohistochemistry. Each brain slice was placed in a 2 ml Eppendorf tube that could hold the slice so its large, flat sides could be exposed to solution. PBST (PBS with 0.2% Triton X-100, Thermo Fisher Scientific) was added to just cover the top of the samples (∼500 μl). Tissue was heated to 50 °C for 1 h in PBST and then cooled to room temperature prior to incubation with antibodies. Conjugated antibodies against the following epitopes were then added to the solution containing each tissue slice: phospho-tau Ser202, Thr205 (AT8, 1.6:500, Thermo Fisher, cat No. MN1020) conjugated to Alexa Fluor 647 (Thermo Fisher, cat No. A37573), HuD Antibody E-1 (1.6:500, Santa Cruz Biotechnology, cat No. sc-28299) conjugated to Alexa Fluor 555 (Thermo Fisher, cat No. A37571), Glut1 antibody conjugated to Alexa Fluor 488 (EMD Millipore, 07-1401-AF488) and 4′,6-diamidino-2-phenylindole dihydrochloride (DAPI, 1.6:500, Sigma-Aldrich, cat No. 10236276001). An additional slice from each donor was separately incubated in an antibody mixture containing amyloid-beta conjugated to Alexa 488 (Cell Signaling Technology, cat No. 51374S), phospho-tau Ser 202, Thr205 conjugated to Alexa 555, and Glut1 conjugated to Alexa 647 antibodies. Tissue was incubated with primary antibodies for one week at 4 °C with gentle shaking. Following incubation, tissue was washed in fresh PBST 3 × 10 min and set on shaker for one week at 4 °C with gentle shaking.

Refractive index matching. After immunohistochemical staining, the samples were incubated with 80% glycerol and 20% deionized water for 24 h at room temperature with gentle shaking. Samples were then placed on a glass microscope slide with a 3D-printed ring that allowed the tissue to remain in a pool of glycerol during imaging. The ring was 3D-printed to match the thickness of the tissue (Formlabs) so a glass coverslip could be placed on top and seal the tissue in the glycerol.

Imaging. The tissue was imaged using Olympus Inverted Confocal FV3000 with a 10 × air objective, and multi-region images were stitched together using the microscope software (Fluoview FV31S-SW, Version 2.5.1.228). Additional higher resolution images were collected by placing the tissue in a bath of 80% glycerol in a Petri dish and imaged using a 20 × immersion objective (Zeiss Clr Plan-Neofluar 20x/1.0 Corr) with an inverted Zeiss 980 confocal microscope. Image Z-stacks were then reconstructed and visualized using Imaris microscopy image analysis software. Alternatively, a LifeCanvas Technologies megaSPIM light sheet microscope equipped with a 3.6 × lens was used, and refractive index matching was performed using EasyIndex (LifeCanvas Technologies).

Analysis with Ilastik (1.4.0). Imaging data for tau and HuD were converted to HDF5 format using Ilastik’s ImageJ plugin [7]. The staining was then individually segmented for each image using Ilastik’s pixel classifier workflow. In short, a paintbrush was used to draw over the signal and background to help train the classifier on how to segment each image. All images were then processed through the trained pixel classifier, and probability maps were exported as HDF5 formatted images. Pixel probability maps and raw data were loaded into Ilastik’s object classification workflow and used to train object classifiers for each image. Tau object classifiers were trained by manually classifying objects as noise or tangles, and HuD object classifiers were trained by manually classifying objects as noise or neurons. Data were exported as object identities and spreadsheets with information about the objects’ classification and characteristics, which were then loaded into MATLAB (r2023b) code to match objects from each channel with colocalized objects.

Separating objects into cortical layers. Imaris surface generation was used to draw regions around each cortical layer on individual imaging planes within the Z-stack, which was then merged into distinct volumes that contained each cortical layer. These volumes were then used to generate a new channel by masking the pixels contained within each volume and setting them equal to the cortical layer (i.e., pixels in layer 1 = 1, pixels in layer 2 = 2, etc.) and pixels not within a clearly defined layer equal to zero. This channel was then exported as a single multipage tiff stack, which could be loaded into our MATLAB code to identify the cortical layer for each object output by Ilastik.

Blood vessel segmentation. Individual blood vessels were manually segmented from clear brain images using a virtual reality image analysis software (Syglass). Segmentation was performed blinded to disease status, using the GLUT1 channel alone. Based on initial qualitative analysis, where tau accumulation along capillaries was not apparent, blood vessels with diameters of approximately 20 µm that spanned multiple cortical layers were selected for segmentation. Using these criteria resulted in the selection of 16–26 blood vessels, or roughly all traceable non-capillary vessels, were manually masked and segmented in each sample. Diameters were measured by drawing a line across the center cross section of each blood vessel and averaging three measures taken from separate locations. Individually masked blood vessels were then realigned with their original image in Imaris, and distance transforms were calculated and exported for each blood vessel.

Intensity and density calculation and binning (MATLAB r2023b). MATLAB scripts were developed to calculate the intensity of tau staining, neuron density, and tau tangle density, along and away from the segmented blood vessels’ surfaces. First, tau data, segmented blood vessel distance transform images, and segmented cortical layer images were loaded simultaneously in one MATLAB script to align all images and export coordinates for data within 100 microns of the blood vessels’ surfaces. These coordinates contained data for each pixel in this range, with their X, Y, and Z positions; tau intensity; and cortical layer. To account for staining differences in each sample, tau intensity was normalized between samples using a piecewise linear normalization. The pixel intensity for background, autofluorescence, and AT8 positivity were recorded in each sample at 3 depths in the images’ z-stacks, with 10 measurements in each category per depth. Then, measurements were averaged for each category and linear functions between them were calculated to normalize the data in a singular dataspace.

To analyze tau staining intensity along the blood vessel, the exported coordinates were input into a script that calculates the pixel distance, in microns, along the surface of the blood vessel. This script calculated a centerline through the blood vessel, found the nearest position on the centerline to each tau pixel, and calculated that position’s distance from the start of the centerline [11]. The data were then exported as the original coordinates with their distance along the vessel appended. A similar script was used for the neuron and tangle density analysis, where the object coordinates obtained from Ilastik were input along with the intensity coordinates. This script calculated the distance along the vessel for each object within 30 microns of the vessel surface.

Finally, these data were binned into groups based on their distance along and away from the blood vessel. Immunolabeling intensity, neuron, and NFT data were grouped in 10-micron intervals along the vessel surfaces. For tau intensity, the mean intensity was calculated for each bin, and data within 3 microns from each vessel surface was used to determine the surface tau percentile [25]. For neurons and tangles, their density was calculated by measuring the number of objects within 30 microns from the vessel surface and comparing their quantity to the spatial volume of each bin. Last, mean tau intensity was normalized to control samples, and R Studio software was used to create heatmaps of tau intensity as a function of vessel surface distance (µm).

All statistical analysis was performed in GraphPad Prism (version 10.2.2). For comparisons with two groups, Shapiro–Wilks normality tests were applied followed by two-tailed Student’s t tests or Mann–Whitney U tests. A priori, an outlier test (ROUT; Graphpad Prism) was used to assess the biochemical data and no outliers were detected—all data have subsequently been included. A repeated-measures ANOVA was used to examine the association between tau intensity along blood vessels and nearby tangle density. In all cases, statistical significance was defined as p < 0.05.