Abstract Background Tauopathies are a heterogeneous group of neurodegenerative disorders characterized by the brain-regional aggregation of three-repeat (3R) or four-repeat (4R) tau isoforms. Current fluid and imaging biomarkers rarely discriminate these isoforms, hampering early, pathology‑specific diagnosis. Objective To determine whether proteomic fingerprints of brain‑derived extracellular vesicles (BD‑EVs) isolated from the prefrontal cortex can (i) distinguish 3R from 4R tauopathies and (ii) mirror the histopathological burden of phosphorylated tau. Methods BD‑EVs were purified from post‑mortem prefrontal cortex interstitial fluid of Pick’s disease (PiD; 3R), progressive supranuclear palsy (PSP; 4R), and control cases (CTRL). Nanoparticle tracking analysis quantified the concentration and size of vesicles. Label‑free LC–MS/MS profiled BD‑EV proteomes, followed by differential expression, gene set enrichment (GSEA), weighted gene co‑expression network analysis (WGCNA), and machine‑learning classification. AT8 immunohistochemistry quantified cortical tau pathology, enabling protein–pathology correlations. Results Tau pathology did not alter overall BD‑EV yield but shifted vesicle size distribution in PiD (higher small/large EV ratio). Proteomic analysis identified two discriminant modules: an astrocyte-derived mitochondrial cluster enriched in PiD and a neuron-derived microtubule cluster depleted in PiD relative to PSP and control groups. Combined glial protein abundance (e.g., GFAP, AQP4, S100β, GLAST, ANXA1) classified PiD, PSP, and controls with perfect accuracy (F1 = 1.0). Several BD‑EV proteins—including CAMKV, TMEM30A, NMT1, AK1 (PiD‑specific), and CALB2 (PSP‑specific)—correlated strongly with regional AT8 burden (|ρ| ≥ 0.70, FDR < 0.05). Conclusions BD‑EV proteomic fingerprints robustly differentiate 3R and 4R tauopathies and track disease severity, unveiling astrocytic mitochondrial proteins as candidate biomarkers. Overall, our results indicate that BD-EV profiling may complement existing approaches for distinguishing tau isoforms and, pending further validation, could ultimately be adapted for use in more accessible biofluids. Supplementary Information The online version contains supplementary material available at 10.1186/s13195-025-01865-w. Keywords: Tauopathy; BD-EV; Brain secretome; Tau isoform; Glia, prefrontal cortex Introduction Tau protein is a pleiotropic protein, performing multiple essential roles in the brain. It stabilizes neuronal microtubules, ensuring efficient nutrient transport and cellular communication along axons. Additionally, tau plays a crucial role in synaptic plasticity by modulating neuronal signaling pathways and regulating interactions between various proteins, which are vital for maintaining neural network stability and adaptability [[46]1–[47]4]. However, when tau becomes dysfunctional, it gives rise to a class of neurodegenerative diseases known as tauopathies [[48]5]. The heterogeneity of tauopathies is striking; they manifest in a spectrum of clinical presentations and neuropathological profiles. Each tauopathy has distinct clinical and pathological features but shares the commonality of tau dysfunction due to hyperphosphorylation. The mechanisms leading to tau aggregation are complex and not yet fully understood. They involve the abnormal phosphorylation of tau, which reduces its affinity for microtubules and increases its propensity to form intracellular aggregates. The specific pattern of tau deposition, the isoforms of tau that aggregate, and the regional brain involvement vary across different tauopathies, which can influence the clinical presentation and progression of the disease [[49]6]. For instance, Pick’s disease (PiD) is typically characterized by the presence of “Pick bodies”, inclusions of 3R tau in granular neurons of the dentate gyrus of the hippocampus and layers two and six of the prefrontal cortex, whereas progressive supranuclear palsy (PSP) and corticobasal degeneration (CBD) are characterized by neuronal and glial (“tufted” astrocytes and astrocytic plaques) lesions made of 4R tau in the brainstem, cerebellar, subcortical and cortical brain regions [[50]7, [51]8]. This diversity in tauopathies suggests that while tau is a common pathological thread, disease-specific factors determine the more subtle phenotypes of this family of diseases [[52]9]. Clinicians today still struggle to differentiate and diagnose tauopathies early in the disease progression. This is particularly true of primary tauopathies, for which the main protein driver is tau. Circulating biomarkers like phosphorylated tau (P-tau) in blood and cerebrospinal fluid (CSF) are increasingly accurate at reflecting tau pathology in the brain at earlier time points in disease progression, but they are often only validated for pathologies like Alzheimer’s disease (AD), which have amyloid beta as a co-driver of pathology [[53]10]. Some tau positron emission tomography (PET) tracers can differentiate 4R tau inclusions and thus can identify diseases like PSP and CBD, however, these tracers are not yet validated for clinical use [[54]11, [55]12]. Interestingly, proteomic analysis of tau inclusions in human 3R and 4R pathology is capable of distinguishing specific tauopathies based on post-translational modifications present in the insoluble tau aggregates obtained from brain tissue. While insoluble tau exhibited distinct phenotypes specific to each pathology, soluble tau extracted from the same aggregates was unable to differentiate between these pathologies [[56]13]. Various forms of soluble hyperphosphorylated tau are currently measured in blood and CSF for diagnostic purposes, but to date they do not allow the diagnosis of primary 3R and 4R tauopathies [[57]14]. Extracellular vesicles (EVs) are now a promising diagnostic tool for many brain diseases. Brain-derived EVs (BD-EVs), which are shed by cells into the interstitial spaces of the brain, encapsulate a diverse cargo that reflects the physiological (or pathological) state of their parent cells [[58]15]. This encapsulated cargo includes proteins, lipids, RNA, and other molecules, providing a comprehensive snapshot of the cellular environment at any given time [[59]16]. As such, BD-EVs provide a unique window into brain health, revealing information that is otherwise difficult to obtain. Thanks to their proteomic cargo, BD-EVs could be the key to a better understanding of pathophysiology, and if their content is also available in accessible biofluids such as plasma and saliva, they could be a diagnostic tool for various tauopathies. This potential has not gone unnoticed, as numerous researchers are exploring using BD-EVs as biomarkers for a wide range of neurodegenerative disorders [[60]17]. Recently, and colleagues have demonstrated that the 3R/4R ratio of plasmatic EVs isolated by L1CAM could serve as a suitable diagnostic tool to discriminate between different forms of frontotemporal dementia, ALS, and Parkinson’s disease [[61]18]. Nevertheless, little is known about the protein signature of BD-EVs as a function of the different types of tau inclusions and isoforms. In this study, we explored the potential of BD-EVs to enhance the understanding and diagnosis of 3R and 4R tauopathies. We analyzed the proteomic content of BD-EVs isolated from the prefrontal cortical biofluids of patients with different tauopathies, comparing their cellular and molecular signatures. Using a combination of advanced approaches, we aimed to identify specific proteomic patterns that could reflect the severity and underlying mechanisms of tauopathies. Combined with current diagnostic protocols, these patterns could pave the way for more accurate diagnosis and dementia classification, particularly for non-AD tau pathology. Materials and methods Patient cohort Brain extracts were obtained from Lille Neurobank, fulfilling requirements concerning biological resources and declared to the competent authority. Samples were managed by the CRB/CIC1403 Biobank (BB-0033-00030 for Lille Neurobank). The demographic data are listed in Leroux et al. (2023). AD brain samples were part of the original cohort described in Leroux et al. (2021); however, we deliberately chose not to include AD cases in the present proteomic analysis of BD-EVs. The data is summarized in Table [62]1. Our aim in this study is to isolate and investigate the effects of tau pathology and isoform differences (3R vs. 4R), specifically in the absence of co-pathology with amyloid. Including Alzheimer’s disease cases—which are defined by the concomitant presence of amyloid-beta—would have introduced a major confounding factor, making it more difficult to discern whether the observed BD-EV signatures were linked to tau isoforms or to the presence of amyloid pathology. We believe this focused approach allows for a clearer mechanistic understanding of tau-driven alterations in EV content, which is a particularly unmet challenge in the field of primary tauopathies. Table 1. Demographic, clinical, and pathological characteristics of human brain donors used for brain-derived extracellular vesicle (BD-EV) isolation. All samples were collected from the prefrontal cortex. AT8 immunohistochemistry (IHC) was performed on adjacent sections from the same fresh frozen tissue used for BD-EV extraction. PMD: post-mortem delay; NFT: neurofibrillary tangles; GFT: globose fibrillary tangles; pid: pick’s disease; PSP: progressive supranuclear palsy; NA: not available Diagnostic Gender Age at death Post-mortem time (h) Tau lesions Braak Stage Thal Stage Cause of death AT8 IHC Control M 78 19 no 0 0 invasive aspergillosis yes Control F 82 NA no 0 0 myocarditis yes Control F 23 24 no 0 0 DIC in septic shock yes Control M 59 13 no 0 0 suffocation yes PSP M 74 26 NFT + GFT NA 1 — yes PSP M 90 36 NFT + GFT NA 2 — no PSP F 82 6 NFT + GFT NA 0 — yes PSP M 69 17 NFT + GFT NA 0 — yes PSP M 65 18 NFT + GFT NA 0 — yes PSP M 82 44 NFT + GFT NA 0 — yes PSP M 84 18 NFT + GFT NA 0 — yes PSP F 77 9 NFT + GFT NA 0 — yes PSP M 57 20 NFT + GFT NA 1 — no PSP F 76 18 NFT + GFT NA 0 — yes PiD M 57 22 Pick bodies + NFT NA 0 — yes PiD M 71 21 Pick bodies NA 3 — no PiD F 78 11 Pick bodies NA 0 — yes PiD M 68 15 Pick bodies NA 0 — yes PiD M 68 8 Pick bodies NA 0 — yes [63]Open in a new tab Interstitial fluid isolation from the patient prefrontal cortex tissue Brain-derived fluids (BD-fluids) were isolated as previously described. Briefly, approximately 1.5 g of fresh-frozen patient prefrontal cortex was gently cut into smaller pieces. These were then incubated in a Papain/Hibernate-E solution to gently digest the tissue and release interstitial fluid, theoretically avoiding cellular lysis. After tissue digestion, differential centrifugations at 300 g for 10 min, 2,000 g for 10 min, and 10,000 g for 30 min were performed at 4 °C to remove cells, membranes, and debris, respectively, and the supernatant of the 10,000g centrifugation was then stored at -80 °C until EV isolation was performed. Normalization according to the weight of the brain extracts was systematically done to avoid bias in the results. Patient brain-derived extracellular vesicle isolation from interstitial fluid The procedures to isolate the brain derived EVs (BD-EVs) from the human BD-fluid were carried out in accordance with the minimal information for the studies of extracellular vesicles (MISEV) guidelines that were established and updated in 2024 by the International Society for Extracellular Vesicles [[64]19]. Various controls were applied to validate the enrichment and the content of the BD-EVs, as recommended in these guidelines. However, the procedure described above to recover BD-fluids may still lead to some cell lysis and intracellular contamination. Intraluminal vesicles (ILVs) in the preparations cannot be fully excluded. 500 µL of BD-fluid were loaded on the top of a size exclusion chromatography (SEC) column (10mL column, CL2B Sepharose, pore size 75 nm, Millipore) [[65]20], . A mean of 7.94 × 10^10 (± 3.36 × 10^9 vesicles/g of tissue in F1–4 (n = 36 samples) was recovered for each sample. Isolation was carried out in phosphate-buffered saline (PBS) with a flow of 36–48 s/mL. The first 3mL were eliminated, and the following 20 fractions were recovered (500 µL per fraction). Measurement of brain-derived extracellular vesicle size and concentration Nanoparticle tracking analysis (NTA) was performed on individual fractions diluted in PBS with a NanoSight NS300 (Malvern Panalytical). To generate statistical data, five videos of 90 s were recorded and analyzed using NTA software (camera level: 15; detection threshold: 4). Label-free liquid chromatography tandem mass spectrometry Protein digestion F1-4 fractions were digested according to a modified version of the iST method (named miST method) [[66]21]. Briefly, 50mL solution in PBS were supplemented with 50mL miST lysis buffer (1% Sodium deoxycholate, 100 mM Tris pH 8.6, 10 mM DTT) and heated at 95 °C for 5 min. Samples were then diluted 1:1 (v: v) with water and reduced disulfides were alkylated by adding 1 /4 vol of 160 mM chloroacetamide (final 32 mM) and incubating at 25 °C for 45 min in the dark. Samples were adjusted to 3 mM EDTA and digested with 0.5 mg Trypsin/LysC mix (Promega #V5073) for 1 h at 37 °C, followed by a second 1 h digestion with a second and identical aliquot of proteases. To remove sodium deoxycholate and desalt peptides, two sample volumes of isopropanol containing 1% TFA were added to the digests, and the samples were desalted on a strong cation exchange (SCX) plate (Oasis MCX; Waters Corp., Milford, MA) by centrifugation. After washing with isopropanol/1% TFA, peptides were eluted in 250 mL of 80% MeCN, 19% water, 1% (v/v) ammonia. Liquid chromatography-tandem mass spectrometry Tryptic peptide mixtures were injected on an Ultimate RSLC 3000 nanoHPLC system interfaced via a nanospray Flex source to a high resolution Orbitrap Exploris 480 mass spectrometer (Thermo Fisher, Bremen, Germany). Peptides were loaded onto a trapping microcolumn Acclaim PepMap100 C18 (20 mm x 100 μm ID, 5 μm, Dionex) before separation on a C18 custom packed column (75 μm ID × 45 cm, 1.8 μm particles, Reprosil Pur, Dr. Maisch), using a gradient from 4 to 90% acetonitrile in 0.1% formic acid for peptide separation (total time: 140 min). Full MS survey scans were performed at 120,000 resolution. A data-dependent acquisition method controlled by Xcalibur software (Thermo Fisher Scientific) was used that optimized the number of precursors selected (“top speed”) of charge 2 + to 5 + while maintaining a fixed scan cycle of 2s. Peptides were fragmented by higher energy collision dissociation (HCD) with a normalized energy of 30% at 15’000 resolution. The window for precursor isolation was of 1.6 m/z units around the precursor and selected fragments were excluded for 60s from further analysis. MS and MS data analysis Data files were analyzed with MaxQuant 1.6.14.0 incorporating the Andromeda search engine [[67]22, [68]23]. Cysteine carbamidomethylation was selected as fixed modification while methionine oxidation and protein N-terminal acetylation were specified as variable modifications. The sequence databases used for searching were the human (Homo sapien) reference proteome based on the UniProt database and a “contaminant” database containing the most usual environmental contaminants and enzymes used for digestion (keratins, trypsin, etc.). Mass tolerance was 4.5 ppm on precursors (after recalibration) and 20 ppm on MS/MS fragments. Both peptide and protein identifications were filtered at 1% FDR relative to hits against a decoy database built by reversing protein sequences. Categorical annotation of proteins Proteins were annotated according to several databases. These databases included the Gene Ontology (GO) Human database - further split into the categories of cellular component (CC), biological process (BP), and molecular functions (MF) - as well as the MISEV guidelines for EV-associated and potential contaminant proteins, a database generated by McKenzie and colleagues in their publication for brain cell type specificity, and the Human Protein Atlas databases for cell type specificity and predicted subcellular localization [[69]19, [70]24–[71]27]. We also compared excitatory and inhibitory neuron-specific proteins within the neuronal category according to the Human Protein Atlas [[72]27]. After annotation, according to the McKenzie et al. database, glial abundance was determined by summing the abundance of astrocytic, oligodendrocytic, and microglial material [[73]26]. Differential expression of proteins Proteins with their abundance values stored as label-free quantitations (LFQs) were further processed using the bioinformatic tool Perseus (v2.1.5). First, we calculated relative abundance values (riBAQ) that is a normalized measure of molar abundance [[74]28]. The scores were transformed using the log2(x) function and the columns parameter for normalization. These transformed values were used to prepare volcano plots. We used the log2 fold change value for PSP and PiD in comparison to CTRL, with their respective significance being –log10(p-value). Proteins with fold change of > 1 and –log10(p value) > 1.3 were identified as the enriched proteins in PSP and PiD. And proteins, with a fold change of <-1 and a -log10(p-value) > 1.3, were identified as depleted proteins in PSP and PiD. Gene set enrichment analysis (GSEA) We used the evaluated fold change values for the respective proteins of PSP and PiD to perform GSEA in R [[75]29]. The fold change in decreasing order was used as the ranking gene list for GSEAGSEA. With the specific functions such as gseGO, gseKEGG, and gseWP from the clusterProfiler R package, we obtained the best enriched functional annotations for the differentially expressed proteins in PSP and PiD. We aggregated similarly associated terms into respective modules, such as synaptic regulation, mitochondria and metabolism, microtubule organization, endoplasmic reticulum and secretory pathway, extracellular matrix and adhesion, along with their associated average normalized enrichment score and FDR score for each gene set. We used bubble plots to represent this GSEA analysis using ggplot2 R package. To display the best terms from respective modules, we used a dot plot with the ggplot2 package. In the plot, we present the gene set size and its enrichment score with its significance value. Weighted gene co-expression network analysis Protein co-expression network analysis was performed with the R package WGCNA on the preprocessed proteomic data of all identified proteins with their respective riBAQ scores for each disease group [[76]30]. At first, a correlation matrix for all pair-wise correlations of proteins across all samples was generated and then transformed into a matrix of connection strengths, i.e., a weighted adjacency matrix with soft threshold power β = 16. Then, the topological overlap (TO) was calculated using the connection strengths. Proteins were further hierarchically clustered using 1-TO measure as the distance measure to generate a cluster dendrogram, and modules of proteins with similar co-expression relationships were identified by using a dynamic tree-cutting algorithm with parameters set to minimal module size = 30, deepSplit = 2, and merge cut height = 0.15. Within each module, the module eigenprotein was defined as the first principal component, which serves as a weighted summary of protein expression within the module and accounts for the greatest variance among all module proteins. Further, module membership (kME) was assigned by calculating Pearson’s correlation between each protein and each module eigenprotein and the corresponding p-values. Proteins were reassigned to the module for which they had the highest module membership with a reassignment threshold of p < 0.05. Hub proteins were found from each module using the signedkME function, which explains the protein membership with its module and its strong association within the module. Gene ontology A detailed genetic annotation of each protein within WGCNA modules was performed. These differentially expressed proteins and co-expressed proteins were characterized based on their gene ontologies, using GO Elite (version 1.2.5) Python package [[77]31]. The entire set of proteins identified and included in the network analysis served as the background dataset for the analysis. The presence of significantly overrepresented ontologies within a module was gauged using a Z-score, while the significance of Z-scores was evaluated using a one-tailed Fisher’s exact test, with adjustments for multiple comparisons via the Benjamini-Hochberg FDR method. Threshold analysis included a Z-score cut off 2, a p-value threshold of 0.05, and a minimum requirement of five genes per ontology before ontology pruning was performed. The best Gene Ontology term that explains the molecular and cellular function of each module was used to name it. Brain cell type enrichment GeneListFET in R was used to evaluate the cell type specificity of WGCNA module proteins [[78]32, [79]33]. The cell type specific gene list from RNA-seq in human brain was used [[80]26]. The total list of identified protein groups was used in the background, and the cell-type-specific gene list was filtered for presence in the total protein list prior to cross-referencing. A one-tailed Fisher’s exact test was performed, and correction for multiple comparisons by the FDR (Benjamini-Hochberg) method was applied to evaluate the significance of cell type enrichment in each module. We used the -log10(p-value) to represent the significance of cell type association. Protein-protein interaction network Proteins in WGCNA modules are associated with specific biological entities and functions, and these proteins must also interact between modules. To understand the interactions between proteins from different modules, we used the SIGNOR 3.038 database as our knowledge base to retrieve the interactions. We utilized Cytoscape 3.10.3 for the network visualization and transformation. The obtained networks were merged into one, which was subsequently curated to retain only potential biomarker proteins, along with their direct or one-step neighborhood interactions, and additional protein complexes and phenotypes proposed by SIGNOR. We utilized various styling tools to represent the distinct characteristics of proteins and their interactions. The interaction network is a reduced version of the entire interactome, highlighting important module proteins and their interactions. Machine learning Machine learning was used to support the traditional statistical results. The dataset was preprocessed by applying min–max normalization to the feature matrix and label encoding to the target variable. A stratified train-test split was performed with 40% of the data used for testing and a fixed random seed of 42 to ensure reproducibility. Then, Borderline SMOTE (Synthetic Minority Over-sampling Technique) was performed on the training set to balance the classes before training the model [[81]34]. Classification was performed using a one-vs-rest approach with XGBoost as the base estimator. The classifier was configured with a regularized tree-based model (max depth = 3, subsample = 0.8, colsample_bytree = 0.8) and trained using multiclass log-loss as the evaluation metric., as described in 9,40. The model was trained with the training set for the classification of data into three classes for each feature. The model was evaluated on the testing set and its performance was assessed using F1-score as the metrics for evaluation. The F1-score for each individual feature for three classes was represented in a performance matrix as a heatmap. Histology Automated immunohistochemistry (IHC) was performed on 4-µm-thick formalin-fixed, paraffin-embedded (FFPE) prefrontal cortex tissue sections using the BenchMark Ultra system (Roche Tissue Diagnostics), in combination with the UltraView diaminobenzidine (DAB) detection kit (Ventana). Sections were incubated with the primary antibody AT8 (MN1020; Thermo Scientific, Illkirch, France) at a dilution of 1:500. This monoclonal antibody recognizes tau phosphorylated at serine 202, threonine 205, and serine 208. Following IHC staining, whole-slide imaging was performed using a Zeiss Axioscan Z1 slide scanner equipped with a 20× objective, covering approximately 5 mm² per section. For each patient, two to three sections were acquired. Image files were first converted and pre-processed using FIJI software (v1.54f), then analyzed using NIS-Elements AR (v6.10) with the General Analysis 3 (GA3) module (Nikon Instruments, Japan). A tissue mask was automatically generated to delineate the brain section, and the central region was excluded to focus the analysis on the cortical periphery. AT8 immunoreactivity was segmented using a manually optimized intensity threshold, established from representative PiD and PSP cases and then uniformly applied across all samples. For each section, the total analyzed area, AT8-positive area, number of tau-positive inclusions, and mean inter-inclusion distance were automatically quantified. The AT8 lesions was calculated as the proportion of immunoreactive area relative to the analyzed tissue area. All values were averaged across the available sections per patient (n = 2–3), and all image acquisition and analysis parameters were kept constant to ensure inter-sample comparability. Data were exported for subsequent statistical analysis. Construction of a combined protein score (C-score) correlated with tau pathology severity To optimize the association between protein abundance in brain-derived extracellular vesicles (BD-EVs) and the severity of phosphorylated tau pathology (AT8), we constructed a weighted composite score (“C-score”) based on the most highly correlated proteins. For each group of interest (all patients, CTRL + PSP, CTRL + PiD, PSP + PiD), the five proteins with the strongest positive and negative Spearman correlations with AT8 were identified from the correlation matrix. A weighted score was then calculated for each subject according to the following formula: C-score = (weight₁ × prot₁) + (weight₂ × prot₂) + … + (weightₙ × protₙ), where each weight corresponds to the correlation coefficient (R) for the given protein, normalized to ensure comparability between markers. Coefficients were initially determined by multiple linear regression on the discovery cohort to maximize the overall correlation of the C-score with AT8 inclusions. Final combinations included two positively and two negatively correlated markers, to minimize overfitting and enhance robustness. Performance of the C-score was assessed by Pearson correlation (R) and corresponding p-value with AT8 quantification across the cohort. This approach enables the generation of an optimally informative composite score for tau pathology severity using a minimal set of protein biomarkers, capturing both positive and negative contributors to disease progression. Statistical analyses All data processing and additional statistical analyses were performed in Python (pandas, scipy.stats), R (base, ggplot2), and GraphPad Prism v10 as needed. Statistical significance for group comparisons was evaluated using ordinary one-way ANOVA with Tukey’s multiple comparisons post hoc test (for particle concentration, MAPT abundance, endoplasmic reticulum protein eigengene plot, and cation transport protein eigengene plot). Ordinary two-way ANOVA was used to assess brain-specific versus non-specific protein abundance. Ordinary two-way ANOVA with Dunnett’s multiple comparisons post hoc test was applied to evaluate the proportion of brain-specific protein by cell type, while two-way ANOVA with Sidak’s post hoc test was used for EV-associated and neuronal vs. glial protein abundances. For non-normally distributed data, including LEV vs. SEV quantity, SNCA and APP abundance, mitochondrial protein eigengene plot, and microtubule protein eigengene plot, the Kruskal-Wallis test with Dunn’s post hoc test was employed. For correlation analyses between BD-EV protein abundance (quantified by ribaQ from mass spectrometry) and histopathological measures (e.g., AT8 burden), Spearman’s rank correlation coefficient was systematically used to assess the strength and direction of monotonic associations. R values were visualized in heatmaps for correlation matrices comparing various BD-EV features with histopathological findings. For other correlations, a two-tailed Pearson’s test was used with a 95% confidence interval as appropriate. P-values were adjusted for multiple comparisons using the Benjamini-Hochberg method where applicable. Significance was denoted as follows: *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001. Results Tau isoform-specific pathology does not impact the distribution profile of BD-EVs Prefrontal cortex samples were prepared to isolate brain-derived (BD) fluids according to the protocol described by Leroux and colleagues. 23 Then, BD fluid-derived EVs (BD-EV) were isolated from human BD fluid using differential centrifugations and size exclusion chromatography (SEC). BD-EVs were then subjected to nanoparticle tracking analysis (NTA) to evaluate their size and concentration, as well as liquid chromatography tandem mass spectrometry (LC-MS/MS) to evaluate their protein content (Fig. [82]1A). Firstly, measurements with NTA indicated that PiD and PSP pathology do not affect the concentration of BD-EVs circulating in the patient’s prefrontal cortex fluids (Fig. [83]1B-C). PiD pathology does, however, shift the size distribution of the EV sample towards having a higher ratio of small EVs (< 150 nm) to large EVs (> 150 nm) than CTRL (Fig. [84]1C-D). We used proteomic analysis to evaluate the quality of BD-EV isolations. We revealed that most proteins detected in patient BD-EVs are shared among control and pathological samples, while a few proteins are detected only in PiD or PSP (Fig. [85]1E). Using the minimum information for the study of EVs (MISEV) guidelines, we annotated all proteins according to their known association with EVs as potential EV markers or potential contaminants of EV isolation [[86]19]. This analysis revealed a significantly higher abundance of EV-associated proteins than components of non-EV associated co-isolated structures (Fig. [87]1F) in all three patient groups, indicating that EV samples are highly enriched in vesicular material. Altogether, these results indicate that tau pathology does not influence the quantity or quality of BD-EVs secreted into patients’ brain derived fluid, but may affect other aspects of BD-EVs’ proteomic makeup and size distribution. Despite careful enrichment and annotation following MISEV guidelines, the possibility of co-isolated non-vesicular contaminants or differences in EV subpopulations between groups cannot be excluded, potentially affecting protein quantification. Fig. 1. [88]Fig. 1 [89]Open in a new tab Consequences of tau pathology on concentration, size and quality of BD-EVs. (A) Illustration of the procedure to investigate how BD-EVs may reflect brain tau pathology. Highlighted is the MISEV annotation of proteins in BD-EV samples as EV-associated or non-EV-associated. (B) Violin plot depicting the concentration of BD-EVs detected by NTA by patient group. (C) Area graph depicting the size and number of EVs in each patient group. Particle concentration has been normalized by tissue weight prior to EV extraction. (D) Stacked bar graph showing the ratio of small (< 150 nm) to large (> 150 nm) EVs by patient group. (E) Venn diagram showing the number of proteins present after threshold filtration in the BD-EVs of CTRL, PiD and PSP patients. (F) Bar graph quantifying the abundance of protein associated with EVs vs. potentially contaminating proteins of EV isolation according to MISEV2023 guideline categories. (G) Performance matrix depicting the predictive capability of a classification model based on BD-EV quality control features We then investigated whether size distribution, EV protein abundance, and particle count could be used in machine learning to predict patient disease status. A OneVsRest classifier along with XGBoost model was used to evaluate the abilities of the features described in Fig. [90]1B-F to differentiate between disease classes. Since we have an imbalanced dataset for our classification model, we used the F1 score because it takes into account the different types of error- false positive and false negative, and not just number of incorrect predictions. The F1 score computes the average of precision and recall, where the relative contribution of both of these metrics results in the F1 score. In the first round of machine learning, the model could only reliably predict subjects based on the percentage of large EVs in the sample or the number of proteins in the sample that pass a presence threshold in the case of PiD (Fig. [91]1G). This further corroborates the conclusion drawn from the findings in Fig. [92]1A-F, which indicate that none of these features differ significantly between the patient groups, except for the ratio of small and large EVs in PiD. This feature was not specific enough to differentiate between PiD and PSP, only CTRL from pathology (Fig. [93]1G). Our findings show that tau pathology slightly affects BD-EV size and protein content, particularly in PiD, but does not significantly impact their concentration or quality. Mitochondrial, endoplasmic reticulum, and microtubule protein signatures in BD-EVs as key indicators of tau isoform pathology Given the large number of peptides present in BD-EVs, we first sought to determine whether protein profiles specific to 3R and 4R tauopathies were reflected in their content. To address this, we first compared the BD-EV proteome between controls and each tauopathy subgroup (PiD and PSP) using label-free quantitative mass spectrometry and differential expression analysis (Fig. [94]2A). Venn diagrams revealed that, while the majority of BD-EV proteins were shared between groups, each tauopathy displayed a distinct set of unique proteins (Fig. [95]2B-C). Volcano plots highlighted numerous proteins that were significantly up- (red points) or down-regulated (blue points) in BD-EVs from PiD and PSP compared to CTRL (Fig. [96]2D–E). To gain insight into the functional consequences of these alterations, we performed a gene set enrichment analysis (GSEA) on the differentially abundant proteins in each disease group (Fig. [97]2F–G). In both PiD and PSP, BD-EVs showed significant enrichment for pathways related to mitochondrial metabolism, endoplasmic reticulum (ER) function, microtubule organization, synaptic regulation, and immune response. Notably, mitochondrial and ER-related proteins were especially enriched in PiD, whereas pathways related to extracellular matrix, cell adhesion, and neuroinflammation were more prominent in PSP. These results indicate that, despite a large overlap, each tauopathy subtype is associated with a distinct molecular signature in BD-EVs, characterized by specific patterns of organelle and cytoskeletal protein enrichment. Fig. 2. [98]Fig. 2 [99]Open in a new tab Differential proteomic signatures of BD-EVs in PiD and PSP compared to controls. (A) Schematic overview of the experimental workflow: BD-EVs were isolated from post-mortem prefrontal cortex samples, analyzed by mass spectrometry, and subjected to differential expression and pathway enrichment analysis. (B) Venn diagrams showing the number of shared and unique BD-EV proteins detected in control (CTRL; n = 4) versus Pick’s disease (PiD; n = 5; left) and PSP (n = 10; right) cases. (C–D) Volcano plots depicting differentially abundant BD-EV proteins in PiD (C) and PSP (D) relative to controls. Each dot represents a protein; red indicates significantly upregulated, blue downregulated (FDR < 0.05, fold change > 1 or < 1). Selected proteins of interest are annotated. (E–F) Gene set enrichment analysis (GSEA) of BD-EV proteins differentially abundant in PiD (E) and PSP (F). Dot plots summarize significantly enriched biological processes (Reactome, GO, KEGG, Wikipathways), with dot size representing gene set size, color representing Average normalized enrichment score (Average NES), and fill (transparency) indicating significance level Additionally, we performed a weighted gene co-expression network analysis (WGCNA) to identify protein modules that best explained the variance between patient groups (Fig. [100]3A and B). This approach revealed four significant modules: M1-Endoplasmic Reticulum, M2-Mitochondria, M3-Microtubules, and M4-Trivalent Inorganic Cation Transport (Fig. [101]3B). Our results show that proteins associated with the endoplasmic reticulum and mitochondria are significantly more abundant in BD-EVs from patients with Pick’s disease (PiD) compared to controls (CTRL) and patients with PSP (Fig. [102]3C). In contrast, microtubule-related proteins are significantly less abundant in PiD patients than in controls and PSP patients (Fig. [103]3C). We created a predictive classification model which utilises One-Vs-Rest behavior along with XGBoost classifier to differentiate the pathology based on given features. With the module data as the given features, our model suggests that only the M2-Mitochondria and M3-Microtubules modules enable the good distinction of PiD tauopathies, although they do not clearly distinguish between controls and PSP patients (Fig. [104]3D). Fig. 3. [105]Fig. 3 [106]Open in a new tab Distinct protein signatures of BD-EVs during 3R and 4R tauopathies. (A) Illustration of the procedure to investigate how BD-EVs may reflect brain tau pathology. Highlighted is the Weighted Gene Co-expression Network Analysis (WGCNA) of the BD-EV proteomic dataset. (B) Weighted gene co-expression network analysis emphasizing four protein modules which account for the majority of variance between patient groups: Endoplasmic Reticulum, Mitochondrion, Microtubule, and Trivalent Inorganic Cation Transport. (C) Violin plots of the cumulative abundance of the proteins contained within each of the four highlighted modules, in the form of eigenegene values (y axis) calculated in WGCNA. (D) Performance matrix depicting the predictive capability of a classification model based on abundance of WGCNA module proteins. (E) Barplots representing brain cell type enrichment for corresponding WGCNA modules. We used -log10(p value) for FET values. Each barplot has a threshold with a single line for significant p value. (F) Protein-Protein Interaction Network of prospective biomarker proteins. Node colors are represents brain cell specificity: Blue: Neuron, Red: Astrocyte, Orange: Oligodendrocyte, Yellow: Microglia, Lightgrey: Unspecified; Node border color are representing the associated WGCNA module: Black: Microtubule module, Turquoise: Mitochondria Module, Pink: Cation Transport Module, Green: Endoplasmic Reticulum. Hexagonal node: Protein Complex, Rectangular Node: Phenotype. Edge Color represents type of regulation: Blue: Positive regulation, Red: Negative Regulation. Edge type represents mechanism: Open Circle: Phosphorylation and Dephosphorylation, Closed Square: Binding, Dashed line: Undirect interaction To further understand the origin and cell-type specificity of the proteins constituting BD-EVs, each protein was annotated using cell subtype-specific databases (see Methods). A Fisher’s exact test with Benjamini-Hochberg correction revealed that proteins in the Mitochondrion module are significantly associated with an astrocytic origin, while proteins in the Microtubule module are primarily of neuronal origin (Fig. [107]3E). These results suggest that tau pathology may induce different perturbations depending on the cellular subtype involved. To deepen understanding of BD-EV content, we constructed a protein interaction network with SIGNOR. This interactome revealed that most proteins in the M2-Microtubule module (outlined in black) are neuron-derived (blue) and strongly interact with proteins from other modules (Fig. [108]3F). Proteins in the M3-Microtubule module are known to play a key role in processes such as synaptic vesicle signalling, excitatory synaptic transmission, and synaptic vesicle recycling (e.g., SNAP25, SNAP91, STX1A, STX1B, SYT1, SV2A, DLG3, DLG4, VAMP2, DNM1, SNCA), as well as in neurofilament assembly, neurogenesis, actin cytoskeleton reorganization, axonal growth cone formation, and neurite outgrowth (GAP43, SYN1, SYN2, SYN3, PDXP). Conversely, most of the proteins in the M2-Mitochondrion module (outlined in turquoise) are astrocyte-derived (red) and interact closely with other proteins (IMMT, ITPKB, FLNC, GRM5, GFAP, CLU). These proteins are primarily involved in neuronal tissue regeneration and microtubule polymerization (Fig. [109]3F). In conclusion, these results demonstrate that the protein content of BD-EVs differs depending on whether the tauopathy is primarily 3R or 4R. In Pick’s disease (3R), there is a higher abundance of astrocyte-derived mitochondrial proteins and a lower abundance of neuronal microtubule proteins compared to CTRL and PSP (4R) patients. These differences point to distinct cellular and functional alterations, for example in synaptic function, neuronal cytoskeletal structure, and astrocytic mitochondrial dynamics. Tau isoforms influence the neuronal and glial composition of BD-EVs We investigated whether the type of tauopathy (3R or 4R) was reflected in the specific protein content of the BD-EV cell subtype (Fig. [110]4A). We observed a significant decrease in the proportion of neuron-specific proteins in the PiD and PSP groups compared with the CTRL group (Fig. [111]4B). Fig. 4. [112]Fig. 4 [113]Open in a new tab Neuronal and glial composition of BD-EV during 3R and 4R tauopathies. (A) Illustration of the procedure to investigate how BD-EVs may reflect brain tau pathology. Highlighted is the Human Protein Atlas annotation of the BD-EV proteomic dataset according to brain cell type specificity. (B) Bar graph showing the percentage of brain-specific proteins which can be categorized as neuronal, astrocytic, oligodendrocytic, or microglial per patient group. (C) Bar graph depicting the overall relative abundance of neuron-specific and glia-specific material by patient group. (D) Performance matrix depicting the performance of a classification model based on abundance of glial and neuronal proteins. (E-N) Violin plots depicting the relative abundance of five neuronal (E-I) and astrocytic (J-N) proteins across patient groups. (O) Performance matrix depicting the predictive capability of a classification model based on abundance of specific neuronal and astrocytic proteins depicted in Fig. [114]3E-N At the glial level, the proportion of oligodendrocyte-specific proteins was significantly higher in the PiD and PSP groups, whereas the proportion of astrocyte-specific proteins was significantly increased only in 3R tauopathy (Fig. [115]4B). Measurement of the relative amounts of neuronal and glial proteins (comprising astrocytes, microglia and oligodendrocytes) revealed a marked imbalance in 3R and 4R tauopathies (Fig. [116]4C). Whereas the neuron/glial ratio is balanced at 1:1 in CTRL, this ratio has been radically altered in favour of glial enrichment in 3R and 4R patients, with a greater increase in glial proteins in 3R tauopathy (PiD) compared with 4R (PSP), (Fig. [117]4C). We used One-Vs-Rest-Classifier along with our XGBoost classification model to get one-class decision boundaries for different features and found that the sum of the relative amounts of glial proteins predicted CTRL, PiD, and PSP with 100% accuracy (F1 score = 1), reflecting a perfect balance between precision and recall (Fig. [118]4D). Among the many proteins associated with neuronal and glial categories, we selected five candidates that are abundant in BD-EV and relevant to the biomarker literature. For the neuronal proteins (Fig. [119]4E–I), we observed that tau (Fig. [120]4E), α-synuclein (Fig. [121]4F), GPM6A (Fig. [122]4G), GAP43 (Fig. [123]4H), and SNAP25 (Fig. [124]4I) were all significantly decreased in PiD. Although all of these proteins were significantly decreased in PiD compared to PSP, only GAP43 and SNAP25 showed a significant reduction when compared to the control group (Fig. [125]4E–I). In contrast, astrocytic proteins (Fig. [126]4J-N), such as GFAP (Fig. [127]4J), GLAST, (Fig. [128]4K), S100β (Fig. [129]4L), AQP4 (Fig. [130]4M), and ANXA1 (Fig. [131]4N) were all significantly increased in PiD compared to controls and PSP patients. Using One-Vs-Rest behavior along with XGBoost classifier, we observed that among the selected neuronal proteins, certain proteins such as GAP43 and SNAP25 presented F1 score > 0.89 to distinguish the 3R and 4R tauopathy groups. Astrocytic proteins such as AQP4 and S100b showed perfect prediction scores (F1 score = 1) to distinguish control patients from other pathologies (Fig. [132]4O). Finally, our results show that combining the relative amounts of the five selected astrocytic proteins allows for a perfect differentiation (F1 score = 1) of 3R and 4R tauopathies from controls (CTRL), (Fig. [133]4O). These results demonstrate that individual protein markers in BD-EV may be limiting for the purpose of distinguishing tauopathies, but that a combined analysis of glial proteins in BD-EV could significantly improve the diagnostic accuracy of 3R and 4R tauopathies. BD-EV protein signatures: correlation with tau pathology severity To assess whether BD-EV protein signatures reflect the severity of tau pathology in the prefrontal cortex, we systematically analyzed the correlation between individual BD-EV protein abundances and AT8-positive tau inclusions measured in matched brain sections (Fig. [134]5A–B). Quantitative assessment of AT8 immunoreactivity confirmed a significantly higher tau pathology burden in PiD groups compared to controls (Fig. [135]5C). Data-driven proteomic analysis revealed that several BD-EV proteins are strongly correlated with AT8 pathology severity, both across all groups and within tauopathy subtypes (Fig. [136]5D). In the global cohort, CAMKV, TMEM30A, and NMT1 showed the highest positive correlations, while subgroup-specific markers emerged, such as AK1 for PiD and CALB2 for PSP (Fig. [137]5D). Conversely, proteins including TROVE2 and SEPTIN6 exhibited robust negative correlations with tau pathology (Fig. [138]5D). Collectively, these results demonstrate that BD-EV protein composition encodes a molecular fingerprint tightly linked to tau pathology progression (Fig. [139]5D). While proteins such as CAMKV were consistently correlated with AT8 burden across all groups, others appeared to be more subtype specific. This proteomic profiling of BD-EVs provides a promising basis for identifying candidate protein markers of disease progression and for stratifying tauopathy subtypes, with potential applicability to circulating biofluids for non-invasive diagnosis. Fig. 5. [140]Fig. 5 [141]Open in a new tab BD-EV protein signatures reflect tau pathology severity in the prefrontal cortex. (A) Experimental workflow: BD-EVs were isolated from post-mortem prefrontal cortex, analyzed by mass spectrometry, and correlated with AT8 tau pathology quantified by automated image analysis. (B) Representative AT8 immunostaining images from control (CTRL), Pick’s disease (PiD, 3R), and PSP (4R) cases. (C) Quantification of AT8 inclusions (per mm²) by diagnostic group. Each dot = one patient; bars: mean ± SEM. (D–G) For each comparison (D: all diagnostic groups; E: CTRL + PSP; F: CTRL + PiD; G: PiD + PSP), the left panel shows the top 10 BD-EV proteins most significantly correlated with AT8 histological burden, based on Spearman correlation. The tables indicate the correlation coefficient (R), p-value, number of subjects (N), and direction (positive or negative). The right panel in each case displays a scatter plot showing the relationship between a composite protein score (C-score) and AT8 pathology burden per subject. These composite scores reveal both common and disease-specific BD-EV proteomic signatures associated with tau accumulation in the brain To further refine biomarker selection, we constructed a weighted composite score, referred to as the “C-score”, within each diagnostic context by combining the two BD-EV proteins most positively correlated and the two most negatively correlated with AT8 immunoreactivity (see Fig. [142]5D–G; formulas shown below each table). Proteins were selected based on the strength and significance of their Spearman correlation with histological AT8 burden. For each group, a weight was assigned to each selected protein proportionally to the magnitude of its (positive or negative) correlation coefficient, and these weights were used in a weighted linear combination to compute a single score per subject — an approach that can be described as correlation-weighted composite biomarker modeling. In the full cohort (Fig. [143]5D), the C-score based on NMT1, TMEM30A, TROVE2, and HDHD2 was significantly associated with AT8-positive inclusion density (R = 0.79, p = 0.00046). In the CTRL + PSP subgroup (Fig. [144]5E), the C-score including CALB2, CAMKV, PLS3, and ANXA6 reached R = 0.88 (p = 0.00056). In the CTRL + PiD subgroup (Fig. [145]5F), the combination of AK1, CTNNB1, SEPTIN6, and SEPTIN5 produced a near-perfect association with AT8 pathology (R = 0.98, p = 0.000041). Lastly, in the PSP + PiD group (Fig. [146]5G), the model based on CAMKV, ATP6V0A1, NUDT5, and SLC3A2 also showed a robust correlation (R = 0.94, p = 0.000018). These findings demonstrate that a minimal panel of four BD-EV proteins is sufficient to capture AT8 pathology severity across distinct tauopathy subtypes, supporting their potential as practical and reliable biomarkers for disease monitoring and progression. Discussion This study investigated the proteomic content of brain-derived extracellular vesicles (BD-EVs) isolated from the frontal cortex of patients with tauopathies and correlated these findings with the number and size of P-tau inclusions observed in this region through immunohistochemistry. Our results indicate that, although tau pathology does not significantly impact the concentration or quality of BD-EVs, it alters their size distribution and proteomic composition, particularly showing an enrichment in astrocytic and mitochondrial proteins in Pick’s disease. We also demonstrated that these distinct BD-EV signatures reflect the severity of tau pathology and could eventually serve as potential biomarkers for monitoring tauopathy progression, highlighting the need for integrated proteomic approaches. However, the degree of interindividual variability observed within tauopathy groups highlights the need to validate the robustness and reproducibility of these signatures in larger, independent cohorts. Although the field of biomarker discovery for neurodegenerative diseases, and particularly Alzheimer’s disease, has made considerable progress over the last decade, notably thanks to sophisticated tau proteomics methods and ultrasensitive P-tau 181 and P-tau 217 assays, it is still impossible to accurately differentiate between 3R and 4R tauopathies in the biological diagnostic setting [[147]10, [148]11, [149]35]. Indeed, PET tracers currently struggle to distinguish the two isoforms, and the use of circulating biomarkers runs up against a conceptual problem recently highlighted by Kyalu Ngoie Zola and colleagues [[150]13]. Indeed, although it is easy to classify 3R, 4R and mixed tauopathies in the insoluble fraction, this is not the case in the soluble fraction, suggesting that it will be difficult, despite the progress and precision of the tools, to imagine assaying 3R and 4R tau in biofluids [[151]13]. Our previous studies demonstrated that extracellular vesicles had very different consequences on the propagation of tauopathies and cell toxicity, particularly that of astrocytes [[152]36]. Here, we continue this work by analyzing the complete proteome of extracellular vesicles in patient brains, providing new insights and perspectives regarding extracellular vesicle diagnostics. We observed that pathology could influence the size of vesicles in the patient’s prefrontal cortex. However, our results show that the presence of tau aggregations, whether 3R or 4R, does not drastically change the concentration and quality of extracellular vesicles. We found a significant difference in the ratio of small (< 150 nm) to large (> 150 nm) vesicles in patients with PiD, which could indicate a change in the mechanism of vesicle secretion in non-apoptotic cells, or vesicles with a more extensive corona [[153]37]. Indeed, many free proteins could be secreted and adorn the corona of vesicles in the brain [[154]38]. In addition, large vesicles, derived from plasma membrane budding, may reflect alterations in cell biogenesis processes and cellular metabolism [[155]39]. Several studies have shown that pathological conditions, including certain brain diseases such as Alzheimer’s disease or cancer, can influence the size of the extracellular vesicles secreted [[156]40]. To better understand the global protein signature of BD-EVs, we have examined the cellular signatures present in BD-EVs thanks to databases generated by McKenzie and colleagues [[157]26]. We found a significant difference in the neuronal to glial material ratio in pathological BD-EVs compared with control BD-EVs. Importantly, we also found a significant difference in the amount of glial material between PiD and PSP, indicating that the level of glial material in BD-EV may indicate the accumulation of specific tau isoforms in the brain. These results suggest that a central element between 3R and 4R tauopathies is probably how glial cells are differentially affected by the accumulation of the isoforms [[158]36, [159]41, [160]42]. Our previous work strongly suggested these results when we observed very different functional modifications by exposing astrocytes to EVs derived from neurons accumulating 3R or 4R tau [[161]36]. Many of our significant BD-EV protein candidates closely correlate with the histopathological features observed in the same patients, further supporting the translational relevance of these vesicular protein markers. Among these, the astrocyte-enriched (“glial”) signature appears to be the most prominent, consistent with accumulating evidence that astrocytic activation and gliosis are key drivers of tau pathology and neurodegeneration in both Alzheimer’s disease (AD) and other tauopathies [[162]43]. Notably, proteins such as GFAP have emerged as promising fluid biomarkers of astrocyte response in AD and primary tauopathies [[163]44]. To further dissect the molecular landscape encoded by BD-EVs, we systematically analyzed the correlation between individual EV protein abundances and the histopathological burden of tau (AT8) in matched samples (Fig. [164]5). This unbiased proteomic analysis revealed several strong candidate biomarkers — including CAMKV, TMEM30A (also known as CDC50A), and NMT1—showing robust positive correlations with tau pathology severity, both in the global cohort and within specific tauopathy subtypes. While these proteins are less characterized in the context of clinical tauopathies, TMEM30A has been implicated in synaptic function and endolysosomal trafficking [[165]45] and its dysregulation has been linked to neurodegeneration. CAMKV (calmodulin kinase-like vesicle-associated protein) has also been associated with dendritic spine maintenance and may play a role in synaptic dysfunction in AD [[166]46]. NMT1 is involved in protein myristoylation, a process increasingly recognized in neurodegenerative disease pathways [[167]47]. We also identified markers with subgroup-specific correlations, such as AK1 for Pick’s disease and CALB2 for PSP, suggesting that the BD-EV proteome may encode not only disease severity but also tauopathy subtype [[168]48]. In addition, proteins with robust negative correlations, including TROVE2 and SEPTIN6, further highlight the complexity of the EV fingerprint. The emergence of less well-characterized proteins as strong correlates of tau pathology opens new avenues for research into their possible mechanistic roles and biomarker value in tauopathies. Whether these BD-EV protein alterations precede, accompany, or follow tau aggregation in disease progression remains unknown and will require longitudinal studies and animal models to clarify their temporal dynamics. Interestingly, our other recent live-imaging study of the response of glial-neuronal 3D co-cultures exposed to extracellular paired helical filaments of tau (4R) revealed a strong astrocytic participation in the internalization of the aggregates [[169]49] which may explain the glial signature in the results obtained here. Astrocytes (but not neurons) displayed increases in mitochondrial dynamics, reflecting the strong overlap between mitochondrial and astrocyte-specific proteins observed in the WGCNA analysis presented in this study. We also observed via a protein interaction network the regulatory role of astrocytic proteins from the mitochondrial module in different neuronal mechanisms such as synaptic transmission, neuronal cell differentiation and apoptosis, which highlights the importance of astro-neuro cross communication during tau pathology. Aside from numerous reported roles of astrocytes in the accumulation, processing, and response to tau [[170]36, [171]43, [172]50, [173]51] these observations reinforce the importance of this cell type, and most importantly its mitochondrial system, in the pathology and as target in research for highly specific biomarkers. These findings are further supported by recent transcriptomic data from Pick’s disease brains. A study by Zechuan et al. using single-nucleus RNA sequencing of the prefrontal cortex in PiD reported a prominent astrocytic transcriptional signature, including upregulation of genes involved in mitochondrial function, neuroinflammation, and energy metabolism [[174]52]. These transcriptomic alterations strongly overlap with the astrocyte-derived mitochondrial module we identified in the BD-EV proteome, providing independent validation of a disease-specific glial response to 3R tau pathology. In addition to their biomarker potential, it will be important to assess whether BD-EVs actively contribute to the propagation of tau pathology or to the modulation of neuroinflammatory and neurodegenerative processes. Our study shows that the in-depth analysis of the BD-EV proteome offers new insights into the biological mechanisms underlying 3R and 4R tau aggregation. More importantly, we have shown that the amount of tau protein, alpha-synuclein, and glial signatures in BD-EVs reflect with excellent accuracy the histopathological state of the brain. This study also demonstrated that the amount of tau protein peptides in BD-EVs is a good indicator of the type of tauopathy, 3R or 4R. Indeed, we observed a sharp decrease in tau in the BD-EVs of PiD relative to CTRL patients, making it possible to distinguish significantly between PiD and PSP patients, though not between PSP and CTRL patients. These results suggest that a simple assay of tau in BD-EVs could be indicative of 3R and 4R pathologies in the case where previous tests indicate the presence of a non-AD tau pathology. A more global analysis of molecular signatures by WGCNA shows that microtubule-associated proteins can differentiate PiD from both controls and PSP. This suggests that the difference between the effects of 3R and 4R accumulation on the cytoskeleton of cortical cells can be directly observed in BD-EVs. Several studies corroborate our hypothesis at the level of the protein in CSF. Indeed, p-tau181 and total tau in CSF are decreased in several tauopathies, particularly in comparison with AD. For example, FTLD-tau generally has lower levels of total tau and p-tau than AD [[175]10] but patients with sporadic FTLD-TDP have even lower levels of p-tau181 than PSP, for which they could be used as diagnostic markers [[176]53, [177]54]. Our results also show that the amount of alpha-synuclein protein follows the same trend as tau protein, i.e., it is inversely proportional to the levels of tau pathology in the brain. These results corroborate several studies that have demonstrated that neurons containing Pick’s bodies in the frontal cortex can also acquire Lewy bodies (LB) during disease progression [[178]55, [179]56]. It has also been shown that in Alzheimer’s disease, neurons with tau aggregates are more likely to form LBs [[180]57]. Furthermore, alpha-synuclein in neuron-derived EVs has recently been shown to be a potential biomarker of Parkinson’s disease (PD) [[181]58, [182]59]. Compared with tau protein, alpha-synuclein is 10-fold more abundant in cortex BD-EVs. Our results therefore strongly suggest that measuring the number of tau peptides and alpha-synuclein in extracellular vesicles could be a good indicator for differentiating the progression and type of 3R and 4R tauopathies. Electronic supplementary material Below is the link to the electronic supplementary material. [183]Supplementary Material 1^ (314.6KB, xlsx) Acknowledgements