Spontaneous pathology in PS19 tauopathy mice progresses via brain networks

Abstract

Tauopathies are progressive neurodegenerative diseases characterized by cellular accumulation of the microtubule-associated protein tau. Evidence suggests tau is a prion, propagating pathology across brain networks via unique transmissible assemblies which mediate distinct neuropathologies in model systems. Neuroimaging has identified network alterations reflecting distinct patterns of brain atrophy in tauopathy patients. Preclinical studies confirmed transmission of pathological tau between connected brain areas, but relied on inoculation of pathogenic tau protein, leaving a gap in experimental evidence that spontaneous tau aggregates act as prions. We used anti-phospho-tau nanobodies in combination with serial two-photon tomography to immunostain and image whole brains from male and female PS19 mice, which have pan-neuronal expression of full-length human tau containing the P301S mutation. We analyzed patterns of phospho-tau deposition across established brain networks at multiple ages, testing the relationship between structural connectivity and patterns of progressive pathology. We identified core regions with early phospho-tau deposition, and used network propagation modeling to determine the link between tau pathology and connectivity strength. We found that tauopathy progression correlated with structural connectivity, consistent with the prion model. Spontaneous tau propagation was biased in the retrograde direction. These data suggest that despite widespread pathological human tau expression in PS19 mice, spontaneous phospho-tau pathology initiates and propagates along specific brain networks. This work establishes new preclinical methods for studying tau accumulation and propagation, and fills a major gap in our understanding of spontaneous tauopathy. Our novel approach establishes a fundamental role for brain networks in tau propagation, with implications for human disease.

1. Introduction

Neurodegenerative tauopathies are characterized by progressive neuronal and glial accumulation in amyloid assemblies of tau protein (Lee et al., 2001). Several lines of evidence indicate tau propagates pathology among neurons. Tau assemblies enter cells to trigger intracellular aggregation (Frost et al., 2009; Holmes et al., 2014), and inoculation of mice with tauopathy homogenates produces local pathology that spreads beyond the injection site (Clavaguera et al., 2009). Prion protein (PrP) inoculation also propagates pathology in neuroanatomical networks (Koshy et al., 2022). Human studies based on functional connectivity and tractography have suggested brain networks could play a role in tauopathy progression, consistent with the prion hypothesis (Braak and Braak, 1991; Seeley et al., 2009; Drzezga, 2018; Frontzkowski et al., 2022). Tau also forms strains, which are defined assembly conformations that propagate faithfully in vivo, creating unique transmissible pathologies (Sanders et al., 2014; Kaufman et al., 2016). Finally, immunotherapies targeting extracellular tau reduce pathology in mice (Yanamandra et al., 2013). However, despite being a core feature of the prion model, this hypothesis remains uncertain.

In humans, magnetic resonance imaging- (MRI) based studies using functional MRI (fMRI) and tractography cannot precisely map pathology at the cellular level. Instead, they rely on atrophy patterns or relatively low-resolution positron emission tomography (PET) imaging of tau deposition, and histopathological analysis of end-stage disease (Seeley et al., 2009). Second, while inoculation in rodents has been widely used by our lab and others to induce focal pathology that spreads from a defined brain region (Lee et al., 2001; Clavaguera et al., 2009; Sanders et al., 2014; Kaufman et al., 2016; Detrez et al., 2019; Cornblath et al., 2021; Danis et al., 2022), these methods induce brain trauma, cannot exclude trafficking of inoculated seeds within a network, and do not reflect the spontaneous development of pathology that occurs in patients. Finally, while traditional methods of assessing tau pathology via immunostaining of serial (2-dimensional; 2D) brain sections have supported trans-neuronal propagation of pathological tau in PS19 mice (Wu et al., 2016), these methods pose a barrier to complete, unbiased assessment of pathological progression. Advances in image registration and analysis of serial sections improve this capacity (Furth et al., 2018; Cornblath et al., 2021), but remain laborious and limited to a subset of anatomical levels. Consequently, the hypothesis of trans-neuronal propagation of endogenous tau pathology, while intriguing, has remained fundamentally untested in vivo.

We now report novel methods for labeling, volumetric imaging, and analysis of the progression of spontaneous tau pathology in PS19 mice, which overexpress P301S mutant human tau and develop progressive accumulation of hyperphosphorylated and insoluble tau protein (Yoshiyama et al., 2007). We have tested the network model by combining three-dimensional (3D) maps of progressive pathology and the mouse structural connectome (Oh et al., 2014), without confounds introduced by inoculation or expression within restricted neuronal populations. Using a fluorescently tagged camelid nanobody (VHH) directed against pathological human tau phosphorylated at residue S422 (VHH-A2-488) (Li et al., 2016), we determined phosphorylated-tau (p-tau) deposition in intact, uncleared brains of PS19 mice at 3 to 12 months. Despite widespread transgene expression, high-resolution whole-brain images revealed p-tau neuropathology that began spontaneously in a small number of distinct brain regions. Whole brain analysis of tau pathology defined characteristic, non-random patterns of p-tau deposition that progressed over time, with a bias towards retrograde propagation.

2. Materials and methods

2.1. Animals

PS19 mice expressing 4R1N P301S human tau under the murine prion promoter (Yoshiyama et al., 2007) and mice lacking endogenous tau (TauKO) (Dawson et al., 2001) at ages from 3 to 12 months were used. The mice used in this study included both males and females, with the exception of the 6mo group of PS19 mice which were all female (Table 1). All mice involved in this study were housed under a 12-h light/dark cycle and were provided food and water ad libitum. All experiments involving animals were approved by the University of Texas Southwestern Medical Center Institutional Animal Care and Use Committee (IACUC).

Table 1.

Sex and phospho-tau cluster assignment information for PS19 mice analyzed in this study.

Sample ID	Age (mo)	Sex	Cluster
297	12	F	5
298	12	F	4
296	12	M	4
371	12	M	5
299	11	F	2
300	11	F	2
301	11	M	4
620	11	M	4
625	11	M	2
711	10	F	5
716	10	F	3
717	10	F	2
715	10	M	3
719	10	M	4
721	10	M	3
308	9	F	1
309	9	F	3
710	9	M	2
748	6	F	1
749	6	F	1
750	6	F	1
751	6	F	1
752	6	F	1

Within the limitations of our study that was not powered to detect sex differences as a primary outcome, sex did not appear to affect either the pathological tau burden or its accumulation pattern at any age.

2.2. Experimental design and statistical analyses

Information on statistical comparisons is found in the relevant methods subsections below, entitled “Regional distribution analysis” which relates to Fig. 5 and “Computational models of pathology spread” which relates to Fig. 6.

2.3. Code accessibility

Registered 3D image stacks of p-tau probability maps, the matrix of p-tau intensity per CCFv3 annotated brain region for each brain, and associated custom code used in visualization and analyses are available in the Zenodo repository, https://doi.org/10.5281/zenodo.7942188.

2.4. Tissue harvest and immunostaining

PS19 animals were sacrificed at 3 (n = 5), 6 (n = 5), 9 (n = 3), 10 (n = 6), 11 (n = 5), or 12 (n = 4) months of age via transcardial perfusion with PBS followed by 4 % paraformaldehyde (PFA; w/v) in PBS. Three TauKO mice at 12 or 14 months of age were additionally sacrificed and perfused in the same manner. Brains were extracted and post-fixed overnight in 4 % PFA at 4 °C. Brains were rinsed in PBS before staining and permeabilized in 10 mL of 0.2 % Triton-X-100 (Surfact-Amps X-100; ThermoFisher #28314) + PBS for 48 h at room temperature (RT) while gently shaking, with three total washes. Brains were incubated in 162.5 μg/mL VHH-A2 previously conjugated to AlexaFluor 488 (VHH-A2-488; a nanobody which specifically recognizes pathological tau phosphorylated at S422) (Li et al., 2016), diluted in PBS + 0.2 % (v/v) Triton X-100 to a total volume of 1 mL for 6 days at 4 °C with constant rocking. Nanobody solution was removed and brains were washed in 0.2 % Triton-X-100 + PBS for 2 days at 4 °C with gentle shaking and solution was changed twice a day. Brains were then rinsed well in PBS in preparation for agarose embedding as described below.

2.5. Agarose embedding

A modified agarose embedding method was employed for the majority of the brains used in this study, which incorporated infiltration of the immunostained brains with acrylamide prior to agarose embedding. The brains were soaked overnight at 4 °C in a solution of Surecast (ThermoFisher #HC2040), with a total concentration of 4 % Surecast and 0.5 % VA-044 activator, in excess volume (5 mL per brain), diluted in Phosphate Buffer (PB; 0.42 g/L Monobasic sodium phosphate, 0.92 g/L Dibasic sodium phosphate). The brains were then allowed to equilibrate to room temperature for 1 h while gently shaking. Each brain was transferred to a cryomold (VWR, #15160-215), which was filled with the prepared acrylamide solution. The mold was covered securely with foil and incubated in an oven at 40 °C for two hours. After polymerization was complete, the excess acrylamide was carefully removed from around the brain. The acrylamide-infused tissue was then embedded in oxidized agarose for vibratome sectioning and TissueCyte imaging as previously described (Poinsatte et al., 2019; Ramirez et al., 2019; Ortega et al., 2020). A subset of brains was embedded in agarose but did not receive the pre-treatment with acrylamide using an earlier version of the embedding protocol.

2.6. Serial two-photon tomography (STPT)

STPT was performed on the TissueCyte 1000 (TissueVision, Inc., Newton, MA) instrument essentially as described (Poinsatte et al., 2019, Ramirez et al., 2019, Ortega et al., 2020). The agarose blocks containing the brain samples were attached to a custom magnetic slide with superglue and placed on a magnetized stage within an imaging chamber filled with PB. STPT imaging, a block-face technique, was used to acquire a series of 2-dimensional (2-D) mosaic images in the coronal plane (Ragan et al., 2012). For this study, three optical planes were imaged at 25, 50, and 75 μm below the cut surface, followed by a vibrating microtome cut at 75 μm (blade vibration frequency of 70 Hz, advancement velocity 0.5 mm/s). The excitation laser (MaiTai DeepSee, SpectraPhysics/Newport, Santa Clara, CA) wavelength was tuned to 800 nm to efficiently excite AlexaFluor 488. Fluorescent signals from three emission channels were collected with a predetermined photomultiplier tube voltage of 720 V using appropriate emission filters encompassing red (>560 nm), green (500–560 nm), and blue (<500 nm) wavelengths. This process produced 190 physical sections and 570 2-D stitched coronal section images (9 by 13 mosaic collected for each coronal section) with a lateral resolution of 0.875 μm/pixel and axial resolution of 25 μm, generating ~280 gigabytes of raw data per brain.

2.7. Image processing

Raw image tiles were trimmed and subjected to flat field correction, then stitched into 2-D mosaic coronal section images using Autostitcher software (TissueVision, Inc.). Using custom written macros (MATLAB), 3-channel merged, contrast adjusted, maximum intensity projection (MIP) images were generated for each physical section (using the 3 optical planes collected for each physical section). The merged MIP images were downsampled to 1.5 μm/px, and contrast was linearly adjusted to match intensity histograms as much as possible to support the subsequent analysis pipeline. A manual quality control process was employed for each immunostained brain to eliminate samples with obvious structural flaws or imaging artifacts. Brains showing excessive vascular autofluorescence signal, no identifiable neuronal phospho-tau staining, or missing portions of the images due to sectioning or dissection artifacts were excluded from analysis (5 total excluded brains).

2.8. Automated image quantification

Our pipeline for quantification of STPT whole brain image datasets was previously described in detail elsewhere (Poinsatte et al., 2019; Ramirez et al., 2019; Ortega et al., 2020). Briefly, the pipeline consists of supervised machine learning based pixel classification, registration of the brain volumes into the reference atlas (Allen Institute CCFv3.0; (Wang et al., 2020)) and automated quantification of fluorescent signals in all annotated brain regions. Supervised machine learning-based pixel classification was performed using a pixel-wise random forest (RF) classifier in ilastik (“Pixel Classification” applet; (Berg et al., 2019)), which allowed isolation of fluorescent signals of interest (e.g., VHH-stained tau deposits) across brain regions that had variable fluorescent backgrounds. A subset of images from each brain was included in the RF model training set (2–3 images per brain). A single RF classifier was trained and used for all brains in the dataset. Probability maps for were exported as scaled 8-bit images (a pixel intensity value of 255 equated to a 100 % probability that a pixel belonged to a signal label). Probability maps were subjected to median filtering prior to analysis to exclude high contrast noise artifacts. Registration to the CCFv3.0 was performed using the raw red (>560 nm) channel autofluorescence signal using a combination of translation, affine, and bspline transformations with SimpleElastix (Marstal et al., 2016). The registration transformation parameters for each brain were applied to the probability map datasets, resulting in registered probability maps of phospho-tau deposition. Finally, the normalized intensity of all voxels lying within CCFv3.0 annotated regions was quantified using custom MATLAB scripts, producing a matrix of phospho-tau signal intensity for each sample, brain region, and hemisphere.

2.9. Background subtraction

Our automated classifier was not always able to differentiate between true positive signal (Supp. Fig. 4A) and autofluorescence from lipofuscin or other background fluorescence in the tissue (Supp. Fig. 4B), but true positives and false positives could be differentiated by eye. Background subtraction was performed in order to guard against potential interpretation errors in the face of regional variability in the probability value distribution (i.e., to account for false positive probabilities caused by autofluorescence from lipofuscin or other sources). Background subtraction was based on probability maps (not raw fluorescence images) from samples with a confirmed lack of p-tau staining including both aged TauKO mice (which should have a similar lipofuscin burden as aged PS19 mice), as well as 3-month old PS19 mice. To remove the automatically detected false positive values in our per-structure quantification, we took advantage of two TauKO brains that were processed alongside our experimental samples as negative controls. Brains from five additional 3-month old PS19 mice – an age with minimal tau pathology (Yoshiyama et al., 2007) – were immunostained and imaged using the same methods to assist in the determination of background staining in PS19 brains (data not shown). We also manually scored a subset of regions in twelve brains. We visually checked 17–112 structures per brain in all 3- and 6-month-old brains plus three 12-month-old brains (1019 structures total). Since fewer than 10 % of the structures in the 3-month-old brains were true positives (36/336) and most of this signal was extremely sparse, we included the 3-month-old brains with the tau KO controls and defined a threshold using the maximum value per structure in the 3-month-old and tau KO brains, excluding the values for the 36 true positive structures. We then subtracted these threshold values from the quantified density per structure in all brains. The distribution of density values for all true-positive and true-negative structures is plotted in Supp. Fig. 4C. This per-structure background subtraction eliminated 488 false positives and resulted in 70 true positive structures falling below the threshold and being assigned a zero value (Supp. Fig. 4D). The whole brain p-tau density for each brain is shown in Supp. Fig. 4E before and after baseline subtraction.

2.10. Regional distribution analysis

To test whether the distribution of tau pathology in each brain was related to structural connectivity, we took advantage of a brain-wide model that converts the density of axonal projections measured by stereotaxic injections in the Allen Mouse Brain Connectivity atlas (Oh et al., 2014) into a structure-wise connectivity weight for all annotated structures in the common coordinate framework ((Knox et al., 2019, Szelenyi et al., 2024) regionalized voxel model, normalized connection density). For each brain, we first calculated the sum of the connectivity weights for all structures containing tau pathology, then compared it to a distribution of randomly sampled networks of the same size with similar inter-regional distances. The procedure for this selection is as follows:

Given a set of N structures containing tau pathology:

1. Randomly draw a set of N structures.

2. Compute the pairwise inter-regional distances for the set of sampled regions.

3. Compute the Kolmogorov-Smirnov statistic to measure the difference in distributions of distances for the sampled and tau correlated networks.

   1. If the KS statistic shows a significant difference in distributions (having a p-value <0.01), reject the sample and return to (1).

   2. Else, return the sample.

The above procedure was repeated 1000 times, after which the normalized connection density of the tau network was compared to the normalized connection density of the random sample networks. Since these connectivity measures are log-normally distributed (Oh et al., 2014; Knox et al., 2019), the t statistic is computed in log-transformed space to test the significance of the difference.

2.11. Computational models of pathology spread

Network diffusion models have been used to characterize and predict pathological protein spread patterns implicated in Alzheimer’s and Parkinson’s disease in mice and humans (Henderson et al., 2019; Pandya et al., 2019; Vogel et al., 2020; Cornblath et al., 2021; Raj et al., 2021). These models hypothesize linear diffusion of pathology over the brain connectome composed of the white-matter tracts. Thus, the pathology spread from a given brain region to its neighbors is a function of their 1) strength of connections and 2) pathology differentials. In this work, we applied the network diffusion model initialized with multiple seed structures on a directed brain graph to predict regional spread patterns of p-tau with respect to age.

To identify seed structures for the network diffusion model, we included regions that were manually validated as true positives in all the 6-month-old brains and were informatically identified as true positives at all subsequent ages. We then included only structures in the top 5 % of all quantified values at 6 months for either total tau or tau density per structure. Only seven structures met all of these criteria: AHN, PA, PGRNl, B, LC, ENTm, and MV’ (see Fig. 6A).

We constructed the directed brain graph $G=\{V,E\}$ with $V$ nodes (gray matter regions) and $E$ edges (white matter projections). The nodes (${\mathrm{N}}_{\mathrm{v}}=625$) comprise 310 bilateral and 5 medial cortical and subcortical gray matter regions defined in the Allen Reference Atlas ontology (i.e. CCFv3 “summary structures” - see Fig. 3). The edge strengths were quantified using the green fluorescent protein measures from the Allen Mouse Brain Connectivity Atlas (Oh et al., 2014).

We represent the adjacency matrix of $G$ as $\mathrm{A}=\left[{\mathrm{A}}_{i\mathrm{j}}\right]$, where i′th row contains connection strengths from the region $\mathrm{i}$ to all other ipsi-, medial, and contra-lateral brain regions.

We denote the measured pathology burden of all nodes at a given time epoch (i.e. mouse age relative to the seed time point) as a vector $\mathrm{x}\left(\mathrm{t}\right)$. We initialize the model with seed structures at the 6 month time point and then compute the predicted tau burdens, $\hat{x}\left(t\right)$ at subsequent $t=\mathrm{3,4},\mathrm{5,6}$ month-intervals with the following equation:

$$\widehat{x}\left(t\right)=e^{-\beta Lt}x\left(0\right)$$

and,

$$L=D-A$$

where, $L$ is a graph Laplacian matrix calculated from the graph out-degree matrix $D$ (i.e. sum of outgoing connection strengths for each node) and the adjacency matrix $A$. Hyperparameter $\beta$ is a diffusion rate constant which scales the global speed of spread for a given network comprising nodes and edges quantified in arbitrary and relative units of pathology and projection strengths, respectively.

We evaluated the model performance with Pearson correlation coefficient between $\mathrm{l}\mathrm{o}\mathrm{g}\left(\widehat{x}\right(t\left)\right)$ (predicted) and $\mathrm{l}\mathrm{o}\mathrm{g}\left(x\right(t\left)\right)$ (measured) p-tau levels at each age. The optimal model was identified by fitting model instances for a range of $\beta$ values $\left(\beta \in \left[1{\mathrm{e}}^{-5},100\right]\right)$ and selecting the one that maximized the average Pearson correlation coefficient over all time points. We checked that the selected $\beta$ remained well within the predefined bounds for each fitted model, and the model performance remained stable over small perturbations of $\beta$. With this setup, we carried out two experiments with the seven identified seed structures.

In the first experiment, we compared the model performance across three anatomical adjacency matrices and a single Euclidean adjacency matrix (see Fig. 6B, C). The anatomical adjacency was derived using 1) anterograde 2) retrograde, and 3) bidirectional projection strengths. The connectivity matrix used for these analyses is an anterograde matrix, constructed by applying a spatial model to anterograde axonal tracing experiments (Knox et al., 2019). The connection (i.e. directional graph edge) strength is modeled as “normalized connection density” from source to target regions. We also created a retrograde adjacency matrix, which is simply the transpose of the anterograde matrix, so that a region’s weight in the retrograde matrix represents the strength of axonal projections into that region, rather than the strength of axonal projections out of that region as in the anterograde matrix. We also defined a bidirectional adjacency matrix which is the sum of the anterograde and retrograde connectivity weights. The weights of the bidirectional matrix are the sum of incoming and outgoing axonal projections in that region, or the overall strength of its connections to the rest of the network. This formulation of directional networks using Allen atlas has been previously used by Cornblath et al. (Cornblath et al., 2021). Euclidean adjacency was calculated as the inverse Euclidean distance between the center-of-mass vertices of anatomical regions. Four separate models were fitted for each of these adjacency measures.

In the second experiment, we compared the performance of the four models (Euclidean, anterograde, retrograde and bidirectional) with and without three mice with substantial levels of cortical p-tau (see Fig. 6D and Supp. Fig. 6) in order to assess the potential basis for relatively lower correlations for model predictions in the 12mo age group. This involved refitting the models with exclusion of one brain at 10mo and two brains at 12mo.

Finally, we fitted three sets of null models to evaluate retrograde model performance against the specificity of seed structures as well as the network configuration using non-parametric permutation tests (see Fig. 6E). In the first set, we fitted separate models with each region from the entire set of bilateral structures set as a single seed. In the second set, we evaluated models with 500 alternate combinations of seven seed regions. Both of these sets of models estimate the null performance distributions with single and multiple random seed initializations. Lastly, in the third set, we evaluated performance of 500 rewired network configurations that preserve higher-order graph properties (Sporns and Betzel, 2016; Vogel et al., 2020; Vasa and Misic, 2022). The goal is to generate a randomized network that has the same number of nodes and edges as the observed network, thereby preserving density and degree. This allowed us to test the hypothesis that p-tau spread is dependent on the actual biological topology defined by the anatomical regions and projections, rather than a function of the total number of connections at a given region in any network with the same number of total nodes and edges. In other words, the “rewired” network null models test whether the performance is driven by the topology of the anatomical connectome represented by A or if it could be obtained with alternative graphs with similar higher-order graph statistics. The p-values were calculated as the percentage of times null models yielded better performance than the retrograde model at each age.

2.12. HJ8.5 Immunostaining and quantification of total tau levels

PS19 animals at 3, 6, 9 and 12 months of age were subjected to transcardial perfusion with PBS and 4 % PFA. Brains were extracted and postfixed in 4 % PFA overnight at 4 °C and transferred to 30 % sucrose (w/v) in PBS until sunk. Coronal sections were prepared on a freezing microtome at 50 μm thickness. Selected sections spanning the rostrocaudal extent of the hippocampus were immunostained with HJ8.5 monoclonal antibodies (Yanamandra et al., 2013) to detect expression of total human tau. Briefly, free-floating sections were permeabilized for 15 min at RT in PBS + 0.2 % (v/v) Triton-X100 and blocked in NGS blocking solution (10 % normal goat serum, 5 % (w/v) BSA, 0.2 % Triton X-100 in PBS) for at least 1 h at RT. The sections were incubated with HJ8.5 antibodies (1:400) sourced from the Diamond laboratory overnight in well plates at 4 °C diluted in NGS blocking solution. Sections were washed 3 times 10 min each in 0.025 % Triton X-100 in PBS at RT, followed by incubation with goat-anti-mouse AlexaFluor 594 secondary antibodies (1:500; ThermoFisher #A-11032) in NGS blocking solution for 1 h at RT in the dark. Sections were washed 3 times 10 min each in 0.025 % Triton X-100 in PBS at RT, stained with DAPI and coverslipped. Whole slide images of the red (HJ8.5) and blue (DAPI) channels were acquired at 20× magnification on the Zeiss Axioscan.Z1. The measurement tool in Zeiss Zen Lite Blue was used to obtain intensity measurements for HJ8.5 staining in the red channel for each section from each animal. Intensity results from sections stained with primary HJ8.5 antibodies were normalized to section area and the corresponding no primary control section from each animal was used for background subtraction. An additional cohort of brains from three PS19 animals at 6, 9, and 12mo of age was processed for sagittal sectioning on a freezing microtome as described above. HJ8.5 staining and slide imaging was carried out as described above, but with VHH-A2-488 included during overnight primary incubation (20 μg per well).

2.13. AT8 immunostaining

Selected 75 μm sections from brains that had been subjected to whole brain VHH-A2-488 staining and STPT were recovered after imaging and stained with AT8 primary antibodies to assess the overlap of p-tau staining between the two antibodies. Briefly, free-floating sections were permeabilized for 15 min at RT in PBS + 0.2 % (v/v) Triton-X100 and blocked in NGS blocking solution (10 % normal goat serum, 5 % (w/v) BSA, 0.2 % Triton X-100 in PBS) for at least 1 h at RT. The sections were incubated with biotinylated AT8 antibodies (AT8b, Fisher Scientific #ENMN1020B) overnight in a well plate at 4 °C diluted 1:500 in NGS blocking solution. Sections were washed 3 times 10 min each in 0.025 % Triton X-100 in PBS at RT, followed by incubation with AlexaFluor 594-streptavidin (1:500; ThermoFisher #S11227) in NGS blocking solution for 1 h at RT in the dark. Sections were washed 3 times 10 min each in 0.025 % Triton X-100 in PBS at RT, stained with DAPI and coverslipped. Whole slide images of the blue (DAPI), green (VHH-A2-488) and red (AT8) were acquired at 20× magnification on the Zeiss Axioscan.Z1.

3. Results

3.1. Development of whole brain p-tau staining and quantification pipeline

Building on previous studies (Ragan et al., 2012; Li et al., 2016; Poinsatte et al., 2019; Ramirez et al., 2019; Whitesell et al., 2019; Ortega et al., 2020; Wang et al., 2020), we created a high-resolution imaging and informatics pipeline to map the distribution of pathological tau across the whole brain in 3–12mo PS19 mice (Fig. 1A). These animals express 1N4R human tau containing a disease-associated mutation (P301S) driven by the pan-neuronal prion promoter (Yoshiyama et al., 2007). We used a previously characterized, fluorescently conjugated nanobody directed against human tau (pS422) (VHH-A2-488), which stains Alzheimer’s and P301L tauopathy mouse brain (Ramsden et al., 2005) comparably to AT8 (Li et al., 2016), an anti-p-tau monoclonal antibody widely used to monitor pathology (Biernat et al., 1992; Goedert et al., 1995; Xia et al., 2020; Stopschinski et al., 2021). Because of their smaller size and other inherent features (Erreni et al., 2020; Zheng et al., 2022), we expected VHH-A2-488 to better penetrate and diffuse through large tissue volumes—essential for serial two-photon tomography (STPT). We optimized a protocol for permeabilization, immunolabeling and washing compatible with whole brain staining of p-tau (Fig. 1A).

Fig. 1.

Whole brain imaging and analysis of nanobody-stained p-tau pathology. (A) Schematic of experimental workflow. Twenty-three PS19 mice were sacrificed at ages 3, 6, 9, 10, 11, or 12mo and immunostained with VHH-A2-488. The brains were subjected to STPT, generating a library of 3-dimensional whole-brain images of spontaneous p-tau deposition. A custom informatics workflow incorporating supervised machine learning-based pixel classification and registration into the CCFv3.0 was used to quantify p-tau accumulation across the cohort. Network diffusion modeling was used to compare the patterns of brain-wide spontaneous tau pathology to structural brain networks. (B) Upper left, p-tau staining in a section from entorhinal cortex (ENT) of a 12mo PS19 mouse brain and enlarged to show p-tau neuronal morphology. Lower left, staining of 12mo TauKO brains with VHH-A2-488 did not produce neuronal signal. Green = VHH-A2-488 positive p-tau deposits, Red = tissue autofluorescence. Upper and lower right, co-staining of AT8 (red), a canonical marker for p-tau, and VHH-A2-488 (green) showed consistent examples of double labeled neurons. Scale bars in lower panels also apply to the corresponding upper panels. (C) Image taken from piriform cortex (PIR) of PS19 animals stained with VHH-A2-488 at 12mos (upper panels) or 6mos (lower panels). Left side shows raw fluorescence images (VHH-A2-488 in green) and right side shows the corresponding probability map outputs of the same region (segmented p-tau signal in green and atlas template in gray). Scale bar in lower left panel applies to all panels. Additional examples of VHH-A2-488 immunostaining in PS19 mice at various ages and corresponding probability maps are provided as Supplementary Fig. 1. (D) Average brain-wide p-tau distribution in PS19 mice in each age group (6, 9, 10, 11, and 12mos). P-tau probabilities are shown in green and the CCF average template in gray. Top, 3D renderings of the whole average brain from a dorsal oblique view. Bottom, Coronal planes of the 12mo average brain spanning the rostro-caudal extent of the brain. Scale bar in far right panel applies to all single plane image panels. Accumulation of p-tau is visible across many brain regions, notably brainstem, hippocampus and cortex, and these regions exhibited a high p-tau burden in 12mo animals. Number of brains averaged for each age group: 6mo, n = 5; 9mo, n = 3; 10mo, n = 6; 11mo, n = 5; 12mo, n = 4.

STPT images of VHH-A2-488 immunostained brains demonstrated neuronal p-tau staining across brain regions including entorhinal cortex (ENT; Fig. 1B), piriform cortex (PIR; Fig. 1C) and anterior hypothalamic nucleus (AHN; Supp. Fig. 1). As observed previously (Li et al., 2016), neurons stained with VHH-A2-488 were sometimes co-labeled with AT8 antibodies (Fig. 1B). We also observed independent populations of VHH-A2-488- and AT8-positive neurons (data not shown) which might be explained by the different tau phospho-epitopes targeted by each antibody. In tau knockout brains (Dawson et al., 2001) we observed no neuronal p-tau staining by VHH-A2-488, confirming a lack of nonspecific staining of non-tau protein by this nanobody (Fig. 1B and Supp. Fig. 1). We then developed a machine learning model compatible with variations in background intensity or texture to classify pixels as VHH-A2-positive or -negative (Fig. 1C, Supp. Fig. 1). We produced “probability maps” of p-tau that were transformed into the 3D Allen Common Coordinate Framework version 3 (CCFv3) reference atlas (Poinsatte et al., 2019; Ramirez et al., 2019; Ortega et al., 2020; Wang et al., 2020).

We visually inspected the CCFv3-registered p-tau probability maps and observed consistent patterns of tau pathology at all ages (Fig. 1D, Fig. 2, Supp. Fig. 2, Supp. Fig. 3). 6mo animals exhibited low p-tau staining above background, most consistently in brainstem and other caudal regions (Fig. 1D, Fig. 2A–D, Supp. Fig. 2A–D, Supp. Fig. 3A–C). At 9–10mos, p-tau signal increased in brainstem regions, and low-to-moderate signal appeared in more rostral regions, including hippocampus and isocortex (Fig. 1D, Fig. 2E–L, Supp. Fig. 2E–L, Supp. Fig. 3D–I). Despite inter-animal variability at each age, a common set of brain structures with p-tau pathology emerged when comparing 6, 9, and 10mo mice, including the pons (P), medulla (MY), hypothalamus (HY), entorhinal cortex (ENT) and piriform cortex (PIR). This suggested p-tau accumulated non-randomly. The 11–12mo animals often had intense p-tau pathology in many brain structures (Fig. 1D, Fig. 2 M–T, Supp. Fig. 2M–T, Supp. Fig. 3J–O). P-tau positive structures in 11–12mo brains largely overlapped those at 9–10mo. The p-tau burden in specific regions was not necessarily higher than in the younger animals, possibly owing to degenerative loss of p-tau bearing neurons in the older brains. (The 11–12mo brains also exhibited pathology in anterior regions such as somatosensory (SS) and motor (MO) cortices, suggesting the network of tau-susceptible structures expanded with age. The general pattern of p-tau accumulation matched recent reports of brain-wide p-tau accumulation in the P301L tauopathy model (Ramsden et al., 2005; Detrez et al., 2019). A subset of brains showed especially strong p-tau pathology in the isocortex (Movie 1). Staining for human tau (with HJ8.5 antibody (Yanamandra et al., 2013, Holmes et al., 2014)) showed consistent expression of human tau across brain sections spanning the rostrocaudal extent of the hippocampus at 3, 6, 9 and 12 months of age (Fig. 3). Further co-staining experiments with HJ8.5 and VHH-A2-488 in PS19 brain sections at various ages again indicated a consistent brain-wide pattern of human tau expression, but restricted presentation of p-tau recognized by the VHH-A2-488 (Fig. 4). The results of these experiments indicated that p-tau deposition did not correlate with total tau expression (Figs. 3 and 4). Consistent levels of human tau expression across the lifespan of PS19 mice have also been reported in plasma, whole brain (Sun et al., 2020), and brain subregions including brainstem and hippocampus (Ono et al., 2022). Together, these data strongly suggested that p-tau accumulation in the PS19 mouse model was ordered and progressive in nature, and could not be attributed simply to transgene expression.

Fig. 2.

P-tau pathology initiates caudally and progresses rostrally. We show maximum intensity projections (MIP) of p-tau probability maps in each age group at selected anatomical levels. Four coronal sections spanning the rostro-caudal extent of the brain at the levels of the anterior cortex, dorsal hippocampus, midbrain and cerebellum are shown for brains of the indicated ages (6mo = Panels A-D; 9mo = Panels E-H; 10mo = Panels I-L; 11mo = Panels M-P; 12mo = Panels Q-T). P-tau probabilities are shown in green and the CCF average template in gray. The total numbers of brains in the dataset in each age group (6mo, n = 5; 9mo, n = 3, 10mo, n = 6; 11mo, n = 5; 12mo, n = 4) were used to generate MIP images of each physical section. Progressive accumulation of p-tau was visible across many brain regions, notably brainstem, hippocampus and ventral cortical areas including entorhinal (ENT), cortical amygdala (COA) and piriform (PIR) regions. Asterisks on each section indicate the location of the corresponding enlarged images shown in Supplementary Fig. 2. Average brain-wide p-tau distribution in PS19 mice at each timepoint is depicted in Supplementary Fig. 3. Scale bar in Panel T is 1.5 mm and applies to all panels.

Fig. 4.

Distribution of p-tau in PS19 mice does not correlate with expression of total human tau. Whole sagittal sections from 12mo PS19 brains were subjected to immunostaining with HJ8.5 to detect total human tau (red) and VHH-A2-488 to detect pathological p-tau (green). Panels A, C, E show staining with both HJ8.5 and VHH-A2-488, and Panels B, D and F show staining with VHH-A2-488 alone (no HJ8.5). Robust staining of total human tau is seen across the entire section, whereas the distribution of p-tau is restricted to sparse populations of cells predominantly localized to the brainstem (A, C, E). No positive staining of HJ8.5 is observed in the negative control section (B, D, F). Panels G-I show brainstem regions of PS19 mice at ages 6, 9 and 12 mo (Panels G, H and I respectively) stained with HJ8.5 (red) and VHH-A2-488 (green) as in above panels. Consistent expression of human tau (red) is observed across ages, as well as prominent accumulation of p-tau (green) in the 12mo animal. Panels J-L show enlarged field of view from an additional 12mo P301S mouse showing the detailed patterns of p-tau accumulation (green) and total human tau expression (red) are not the same. Scale bar in Panel F is 1 mm and applies to Panels A-F. Scale bar in Panel I is 500 μm and applies to Panels G-I. Scale bar in Panel L is 20 μm and applies to Panels J-L.

Fig. 3. Total levels of human tau do not vary with age in PS19 mice.

Brains were harvested at 3, 6, 9, and 12 months and serial frozen sections were subjected to immunostaining with HJ8.5 antibodies to label total human tau. A) Representative whole section images of the series of three sections at different anatomical levels from a 9 month old brain stained with HJ8.5 (Top right and bottom 2 panels) or a no primary control section from the same brain (top left panel). HJ8.5 staining is shown in red and DAPI nuclear staining is shown in blue. Scale bar in top left panel of 3 A is 2 mm and applies to all panels in 3 A. B) Full resolution (20×) images of hippocampal area CA1 from brains at each of the indicated ages showing robust HJ8.5 staining in red. Scale bar in 3 month image is 50 μm and applies to all panels in 3B. C) Quantification of HJ8.5 staining intensity across six serial sections for each age group (one series of three sections at anatomical levels as shown in Fig. 3A from two brains per age group). Error bars = standard deviation. Scatter plots of individual replicates (n = 6) are shown as open circles. No significant differences were observed in total tau levels between ages.

3.2. Nonrandom patterns of tau pathology emerge across animals and ages

In humans, neuropathological heterogeneity has been proposed to reflect variable onset times, anatomy, and rates of disease progression (Braak and Braak, 1991). We investigated whether p-tau staining patterns represented mild-to-severe variants of pathology in a consistent set of brain regions, or whether they were more random. To approach this question, we used cluster analysis on region-normalized p-tau signals across 23 brains and 315 brain regions annotated in the Allen CCFv3. These “summary structures” covered the entire brain (Fig. 5) (Walther et al., 2009; Wang et al., 2020). This revealed pathological patterns related to, but spanning, age groups (Fig. 5A). We observed no obvious asymmetries between hemispheres in p-tau staining at any age (Pearson’s correlation coefficient 0.9226, p < 1e-12). Cluster analysis confirmed initial visual observations of p-tau pathology in specific brain structures. P-tau signal was generally highest in the pons (P), medulla (MY), and select structures in the cerebellum (CB) and hypothalamus (HY). The cluster with lowest overall p-tau intensity (Fig. 5A, “C1”) contained five 6mo brains, when consistent pathology emerges in the PS19 model (Yoshiyama et al., 2007; Holmes et al., 2014), as well as a 9mo brain. A second cluster with relatively mild caudal staining contained four older (10–11mo) brains (Fig. 5A, “C2”) as well as a 9mo brain. The C1 and C2 cluster composition, where C1 contains a 9mo brain with generally similar p-tau levels to 6mo brains, and C2 contains predominantly older brains with a lower pathological burden indicates that some animals were resistant to pathological progression. This pattern is also consistent with a report of widely variable ages of induction of tau pathology in the PS19 model (Woerman et al., 2017). A third cluster of four 9–10mo brains had pronounced caudal staining, including P, MY, and CB (Fig. 5A, “C3”). The fourth cluster included five 10–12mo brains with intense p-tau staining in caudal regions and increased pathology in midbrain (MB), olfactory areas (OLF), and the hippocampal formation (HPF)(Fig. 5A, “C4”). Finally, three 10–12mo brains clustered together that had a distinct pattern of pathology with pronounced staining in isocortex, OLF, HPF, and only moderate staining in P, MY, and CB (Fig. 5A, “C5”). Overall, we found highly nonrandom patterns of p-tau pathology, with C1-C4 appearing to be instances of pathological progression across one distinct subcortical set of regions (“Subcortical”), and C5 appearing to be a late stage of progression across a separate, distinct set of cortical regions (“Cortical”). Remarkably, excepting a few caudal structures, brain-wide patterns of neuropathology in the Cortical and Subcortical groups were almost entirely nonoverlapping (Fig. 5A, B, Movie 1). With the caveat that our study was not powered to assess sex differences as a primary outcome, we found no obvious effects of sex on either pathological burden or presentation of the subcortical vs. cortical pattern of pathology (Table 1).

Fig. 5.

Patterns of spontaneous p-tau pathology match structural connectivity. (A) Raster plots showing automated quantification of region-normalized p-tau probability levels in each of the 630 structures annotated in the Allen Institute CCFv3. Displayed values were truncated at 3 × 10⁶ to visualize lower-intensity regions. Clustering based on brain-wide patterns of pathology identified progressive accumulation of p-tau in a predominantly subcortical pattern (C1-C4), but a subset of three brains exhibited an alternate pattern with heavier pathology in cortical regions (C5). Both “Subcortical” (C4) and “Cortical” (C5) clusters of brains with high p-tau burdens were composed of aged mice (10–12mo). Labels across the top indicate major brain divisions. Abbreviations: OLF: Olfactory areas; HPF: Hippocampal formation; sp.: Cortical subplate; STR: Striatum; PAL: Pallidum; TH: Thalamus; HY: Hypothalamus; MB: Midbrain; P: Pons; MY: Medulla; CB: Cerebellum. A description of the background subtraction method used in this analysis is presented as Supplementary Fig. 4. (B) Maps showing the spatial pattern of p-tau pathology by annotated brain region. Median values are shown for n = 3 brains with pronounced Cortical pathology patterns and n = 5 brains with Subcortical pathology (indicated by boxes in A). Coronal sections at the levels of the hippocampus (left), midbrain (center), and brainstem (right) are shown for each pattern. Heatmap is the same as A. (C) Schematic of method used to evaluate whether p-tau levels were related to connectivity. For each brain, the set of regions (nodes) that contained p-tau signal (red points) were identified and the sum of the anatomical connection strength between regions was calculated by summing all the edges in this graph (red lines). Inter-regional connection strength values came from the regionalized voxel model in Knox et al. (Knox et al., 2019). The strengths of all the connections in this network were summed and compared to results of 1000 alternative networks drawn from the same brain-wide connectivity matrix and containing a similar distribution of inter-regional distances (blue-gray lines). (D) Histogram of the total connection strength of 1000 alternative networks (blue bars) compared to the connection strength of observed a representative p-tau positive network from Sample 301 (red line). The structures positive for p-tau in Sample 301 had a total connection density of ~3.5 × 10⁻⁶, which is higher than ~98 % of the possible networks that could be formed with a similar distance distribution. (E) In most brains, p-tau positive structures were more highly connected than expected by chance, as indicated by their network z scores. There was no relationship between the cluster assignments in panel A and the connectivity strength of the corresponding p-tau positve networks (compare along y-axis), or between cortical-dominant or subcortical-dominant p-tau patterns and network connection strength (compare red points and magenta points). There was no obvious similarity between the two brains that showed lower connection strength in their p-tau positive networks (308 and 371). (F) There was no relationship between the number of p-tau positive structures in a brain and network connection strength, indicating that these networks represent true preferential connectivity and are not simply a result of including more structures in the network in older brains.

3.3. Pathology co-occurs in strongly connected brain regions

Next, we tested for a relationship between regional patterns of p-tau accumulation in any given PS19 brain and underlying network connectivity, using the Allen Mouse Connectivity Atlas as the reference for normal connectivity strengths among all brain regions (Oh et al., 2014; Knox et al., 2019; Szelenyi et al., 2024). We compared the observed sum of connectivity strengths between all regions with detectable p-tau signal bilaterally in each individual brain to the sum of connectivity strengths derived from sets of alternate random networks of similar anatomical (Euclidean) distances (Fig. 5C, D). An individual brain’s p-tau-containing regions were more strongly connected than expected by chance (z-scores >2; Fig. 5E, F). Notably, two brains (308 and 371) exhibited a lower connection strength among their p-tau-positive structures, which was not explained by their pathological stage or p-tau distribution pattern (Fig. 5E). We also found no correlation between the total number of p-tau positive regions with the network connectivity z-score, indicating that this relationship was not enriched in brains with higher p-tau burden (Fig. 5F). We next used the strength of all projections originating from each region (out-strength), the strength of all incoming projections to each region (in-strength), and the sum of all incoming and outgoing projections in each region (total strength) to test for a relationship between the strength of a region’s connectivity and its p-tau burden. We plotted the p-tau burden as a function of in-strength, out-strength, and total strength for all mice at the 6 month and 12 month time points and found no significant correlation for any connectivity metric with p-tau burden at either age (Supp. Fig. 5A–C). Finally, we visually inspected the connectivity matrix for evidence of communities formed by highly connected structures (Supp. Fig. 5D–H). Nodes were sorted by their p-tau levels. P-tau positive nodes are located on the top-left quadrant of the matrices, defined by the white lines. There was no obvious pattern that suggested high p-tau nodes were consistently hyper-connected. Overall, these analyses indicated that brain-wide distribution patterns of p-tau significantly correlated with their network connection strength.

3.4. Brain networks predict patterns of pathology progression

Our previous analysis did not consider whether progression of pathology correlated with network connectivity. Thus we first identified reliable “seed regions,” defined as those with detectable p-tau in 6mo mice that were also positive for p-tau at 9–12mo. This curation produced seven consistent seed regions: anterior hypothalamic nucleus (AHN), posterior amygdalar nucleus (PA), paragigantocellular reticular nucleus, lateral part (PGRNl), Barrington’s nucleus (B), locus ceruleus (LC), entorhinal area, medial part, dorsal zone (ENTm), and medial vestibular nucleus (MV) (Fig. 6A). We used these regions as starting nodes for subsequent analyses to test the performance and validity of four models for predicting progression of tau pathology across ages: (1) Euclidean distance between regions, (2) anterograde connectivity, (3) bidirectional connectivity (retro/anterograde), and (4) retrograde connectivity (Fig. 6B).

Fig. 6.

Brain networks defined with the retrograde connectome best predict spread of pathology. (A) Regions with p-tau signal in 6mo brains were identified as seed structures. Abbreviations: AHN = Anterior hypothalamic nucleus; PA = Posterior amygdala nucleus; ENTm = medial entorhinal cortex; LC = Locus ceruleus; B = Barrington’s nucleus; MV = Medial vestibular nucleus; PGRNl = Paragigantocellular reticular nucleus, lateral part. Heatmap shows the mean density of p-tau signal in each region in n = 5, 6mo brains. (B) Potential mechanisms for tau spread. Neurons in structures send anterograde projections to neurons in other areas and receive projections from neurons in other areas. For simplicity, only two neurons are shown in the diagram representing cells in a region with high p-tau burden (purple) and its one neighboring structure with relatively low p-tau burden (white). If tau spreads in an anterograde direction, the amount of spread should be related to the strength of the anterograde projection (w_i) from the high p-tau region to its low p-tau neighbor(s). If tau spreads in a retrograde direction, the amount of spread should be related to the strength (w_j) of the input from the low p-tau neighbor(s) to the high p-tau region. If tau spreads in both directions, its propagation should be related to the sum of the anterograde and retrograde connection weights. If tau propagation is driven by Euclidean distance, the amount of spread should be related to the distance between regions, not the weight of their anterograde or retrograde connections. (C) Measured levels of p-tau (log scale) plotted as a function of the tau pathology predicted from propagation models based on Euclidean distance between regions (gray), anterograde connection strength (green), bidirectional connection strength (blue), and retrograde connection strength (pink). Each point represents one region from the Allen Institute CCFv3. Pearson’s r values are reported for each model-age group combination. The outliers on the right side of each scatter plot denote the seed structures. (D) Comparison of retrograde model with and without the three brains with the cortical pattern of pathology (Fig. 5A). The model predictions improved at 10 and 12mo when three brains with the cortical pathology pattern were excluded from the analysis. (E) Model performance compared to three sets of null models to evaluate specificity of seed regions and the retrograde network connections for cohorts with (top) and without (bottom) brains with the cortical pathology pattern. The null models included predictions from 315 single seeds (i.e., individual regions, dark-gray), 500 alternate combinations of 7 regions (light-gray), and 500 rewired network configurations that preserved higher-order (i.e., node degree sequence) graph statistics (pink). Each dot represents the performance of one iteration of the null model, while the horizontal dashed line specifies the performance of the proposed model that was statistically compared to the performance of the null models. Significance notation: * P < 0.05, **P < 0.01, ***P < 0.001).

We used diffusion spread models to predict the regional distribution of p-tau pathology at later ages based on the levels of p-tau in the seed regions at 6mo. The models were applied to the brains based on age, since our previous clustering results showed that, as expected, aging was a prominent contributor to development of high levels of p-tau pathology (Fig. 5A, C4 and C5). We compared the predicted p-tau levels with the average measured p-tau per region at 9, 10, 11, and 12mo (Fig. 6C). We observed positive correlations between predicted and measured p-tau for all four models, but some outperformed others. The Euclidean distance-based model performed the worst at all ages, with the lowest Pearson’s r (Fig. 6C, gray, top row). The obvious outlier regions, predominantly seed structures, in the right side of Fig. 6C indicate poor prediction of pathology propagation from the seeds to the rest of the brain. The retrograde connectivity model best predicted p-tau levels in most age groups (Fig. 6C, pink, bottom row). The anterograde connectivity model performed worse than the retrograde model in all but the 9mo age group (Fig. 6C, green). The bidirectional spread model (Fig. 6C, blue) predicted p-tau levels reasonably well and outperformed the anterograde model in every age group.

We also noted relatively lower correlations for all model predictions in the 9 and 12mo age groups. At 9mo, this may be explained by the larger time gap from the 6mo seeding epoch (3mo separation) vs. subsequent age groups (1mo separation), potentially biasing the models to better capture progression at the later ages. For 12mo, we suspected our earlier observation of two distinct deposition patterns, “Cortical” and “Subcortical”, might be a contributing factor (Fig. 5A, B). Therefore, we re-ran the same analysis without the three “Cortical” samples (2 at 12mo, 1 at 10mo). The predictive performance of the retrograde spread model indeed improved, from r = 0.61 to 0.66 (10mo) and r = 0.45 to 0.66 (12mo) (Fig. 6D). The predictive performance of the other three models also improved at 12mo when the three “Cortical” samples were excluded; however, the retrograde model still outperforms the other three models (Supp. Fig. 5). These results suggested the extensive cortical deposition likely reflected output from a different network pathway than that followed in most brains.

Finally, we tested whether the number and/or the specific set of seed regions impacted performance of the spread models. We compared the Pearson’s r values derived from the retrograde spread model against the performance of three different “null” model predictions based on: a single seed region randomly selected from the entire set of summary structures, randomized sets of n = 7 seed regions from the entire set of summary structures, and “rewired,” or plausible alternative sets of connectivity strengths (Sporns and Betzel, 2016; Vogel et al., 2020; Vasa and Misic, 2022) between the n = 7 p-tau seed regions (Fig. 6E). Using non-parametric permutation testing, we found the retrograde spread model with the curated set of seven seed regions significantly (p < 0.01) outperformed both single (dark-gray) and multiple seed (light gray) null models in all age groups regardless of cortical mice inclusion. The retrograde model also significantly (p < 0.05) outperformed the “rewired” network null models at 10 and 11mo with inclusion of cortical brains, and in 10, 11, and 12mo age groups with the three cortical brains excluded. These results further indicated that p-tau spread in PS19 mice was driven specifically from these seven seed regions and was strongly related to the underlying connectome architecture mapped in the Allen Mouse Brain Connectivity Atlas. Overall, these data indicated that neuronal p-tau accumulation patterns across age were primarily propagated via the anatomical connectome as opposed to absolute Euclidean proximity between the regions, and that spread between regions was biased towards the retrograde direction.

4. Discussion

It has remained unknown, but of critical importance, whether the axonal projections that comprise brain networks mediate propagation of spontaneous tau pathology, both for an understanding of disease mechanisms and for preclinical translational studies in mouse models.

4.1. Mapping whole brain tau pathology in 3D

We used a novel staining protocol with a fluorescently conjugated camelid anti-p-tau nanobody and STPT to generate high resolution whole brain images of p-tau accumulation in PS19 mice aged 6–12mos. The use of nanobodies, which can in theory be engineered to bind any substrate, to map proteins at subcellular resolution is a powerful approach with potentially broad applications. Because VHH-A2 was directly conjugated to AlexaFluor 488, whole-brain staining of fixed, permeabilized mouse brains was performed in a single, highly efficient step. This method, combined with STPT, reduces some barriers to whole brain mapping using alternative methods such as traditional histological sectioning or lightsheet microscopy. It was also enabled by our development of an informatics pipeline incorporating supervised machine learning to detect and quantify cellular p-tau pathology, with registration of the acquired image stacks to a standard 3D anatomical reference space—the Allen CCFv3 atlas. We anticipate that this approach could be used to study many aspects of brain biology and neuropathology.

4.2. Progressive tau pathology follows brain networks

Propagation of tau pathology based on prion mechanisms requires that tau protein be expressed in a native, “seedable” state in vulnerable cells, with the progressive aggregation mediated by trans-cellular movement of seed-competent tau. Although proposed by many, it has remained unclear whether spontaneously occurring tau pathology in a human, or in a mouse model, spreads via brain networks. In humans, studies of the role of network involvement have combined functional connectivity or tractography, with imaging of tau pathology based on positron emission tomography, or, as a surrogate, brain atrophy (Seeley et al., 2009). Each of these methods, while state-of-the-art in humans, has poor cellular resolution, orders of magnitude below what was achieved here via STPT. In mouse models, network studies have been based on inoculation of seeds (Sanders et al., 2014; Detrez et al., 2019), which does not reflect a normal disease process, and is prone to experimental artifact. Instead, we have taken advantage of the widely used PS19 mouse (Yoshiyama et al., 2007; Yanamandra et al., 2013; Sanders et al., 2014; Kaufman et al., 2016) to study the development of spontaneous tau pathology. While we and others (Sun et al., 2020; Ono et al., 2022) have not observed any regional differences in transgene expression across the brains of these animals, it remains a possibility that variation in human tau transgene levels across the brain in the PS19 model could affect the regional deposition of pathological tau. The variability which we observed in p-tau burden across the lifespan is consistent with earlier reports showing significant variability in spontaneous hippocampal p-tau accumulation in aged male and female PS19 mice (Sun et al., 2020; Woerman et al., 2017).

The Allen Mouse Connectivity Atlas and CCFv3 resources facilitated our analyses of brain-wide tau propagation patterns across discrete anatomical regions (Oh et al., 2014). By registering whole brain images of p-tau deposition in PS19 mice into the annotated Allen CCFv3, similar to recent studies (Whitesell et al., 2019; Cornblath et al., 2021), we identified seven regions which consistently exhibited pathology at 6mo and all later ages. We determined that the emergence of additional sites of pathology was related to age, but not exclusively. The areas of the brain with tau pathology in any given animal exhibited stronger network inter-connectivity than would be expected from a random set of regions. We also found that the connectivity strengths between regions in the retrograde direction best predicted progression of p-tau levels with increasing age. In contrast, anatomic proximity, with a few exceptions, had relatively weak predictive power. Importantly, the improved fit of the Euclidean model after removal of the 3 “cortical” brains is likely due to the excluded p-tau+ regions being physically further away from the seed regions. Thus, the simplest interpretation of our data is that pathology begins in discrete areas in PS19 mice, before propagating via network connections with a retrograde bias. While we have used the existing connectivity map generated in wild-type mice (Oh et al., 2014) for this study, we note the potential caveat that expression of tau in individual neurons may affect the underlying network connectivity in the PS19 model. However, to our knowledge, no altered network connections in this mouse have been reported.

We made a surprising observation of two relatively discrete patterns of late stage pathology: cortical vs. subcortical. In the three cases with dominant cortical pathology, selected seed regions did not effectively predict their involvement in later ages, reducing the apparent power of the network model. However, when the three mice with cortical involvement were not included in the analysis, the predictive power improved for all four models. The improved fit of the Euclidean model after removal of the 3 “cortical” brains is likely due to the excluded p-tau+ regions being physically further away from the seed regions. Nevertheless, the retrograde model consistently outperformed the other three models with and without the inclusion of the cortical pattern brains. The origins of these two distinct late stage pathological patterns in isogenic PS19 mice are not clear. They could arise from variation in earliest regions of initiation, differences in neural development, strain composition, or other unknown factors that impact the emergence of tau pathology. In this study, we did not observe enough individual mice at intermediate stages of tau progression to distinguish those that might have been on the “cortical” path from the more prevalent “subcortical” path. Future studies can be designed to oversample mice in the earlier and intermediate stages of disease to promote sufficient representation and analysis of animals with both types of pathological progression.

5. Conclusions

It is not yet technically possible to track in real time the physical propagation of protein pathology across distributed brain networks in vivo, i.e., transcellular movement of tau seeds. However, our observation that emergent cellular pathology in a mouse model with pan-neuronal mutant tau expression is accurately predicted by a retrograde propagation model suggests that in mice, as has been long hypothesized in humans (Duyckaerts et al., 1997), spontaneous tau prion phenomena underlie neuropathological progression. It also indicates that the PS19, and perhaps other endogenous tauopathy models, remain useful for analysis of mechanisms of propagation, and development of diagnostic and therapeutic strategies that can be translated to humans.

Supplementary Material

Supplementary data to this article can be found online at https://doi.org/10.1016/j.nbd.2025.107072.

Significance statement.

Our novel methodology for whole brain imaging of p-tau deposition reveals retrograde-dominant network propagation in a tauopathy mouse model. This work establishes new preclinical methods for studying tau accumulation and propagation, and fills a major gap in our understanding of spontaneous tauopathy. Our results establish a fundamental role for brain networks in tau propagation, with implications for human disease.

Abbreviations:

2D

two-dimensional

3D

three-dimensional

AHN

anterior hypothalamic nucleus

B

Barrington’s nucleus

CB

cerebellum

CCFv3

3D Allen Common Coordinate Framework version 3 reference atlas

ENT

entorhinal cortex

ENTm

entorhinal area, medial part, dorsal zone

fMRI

functional magnetic resonance imaging

HPF

hippocampal formation

HY

hypothalamus

LC

locus ceruleus

MB

midbrain

MIP

maximum intensity projection

MO

motor cortex

MRI

magnetic resonance imaging

MV

medial vestibular nucleus

MY

medulla

OLF

olfactory areas

P

pons

PA

posterior amygdalar nucleus

PB

phosphate buffer

PET

positron emission tomography

PGRNl

paragigantocellular reticular nucleus, lateral part

PIR

piriform cortex

PrP

prion protein

p-tau

phosphorylated tau

SS

somatosensory cortex

STPT

serial two-photon tomography

VHH

heavy chain variable domain

Data availability

Data will be made available on request.