Proteomic Changes in the Human Cerebrovasculature in Alzheimer’s Disease and Related Tauopathies Linked to Peripheral Biomarkers in Plasma and Cerebrospinal Fluid

Comparative analysis of vascular and bulk proteomes in brain

Proteomic characterization of cerebrovasculature has been challenging owing to low abundance of vascular cells in the brain and inefficient capturing of vascular-specific proteins by bulk quantitative methods. Thus, to investigate the molecular definition of human brain vasculature in comparison to traditional bulk analysis, we isolated cerebral blood vessels from frozen frontal cortical tissues using density-mediated separation. To verify the enrichment of the cerebrovasculature in our preparations, we analyzed proteomic profiles of different brain fractions, including i) whole brain, ii) parenchyma/myelin, and iii) vascular fractions by label-free MS-based proteomics. We identified core clusters of highly enriched fraction-specific proteomic signatures. As expected, the cerebrovascular fraction was enriched in proteins specific to endothelial cells, SMCs, pericytes, and fibroblasts (Supplemental Figure 1A). Further validation of the enrichment of the human brain vasculature by Western blotting showed a significant increase in the levels of vascular specific proteins in the vascular fraction, including markers of endothelial cells Claudin-5 (CLDN5), CD31 (PECAM1) and vimentin (VIM) (Supplemental Figure 1B). Finally, we assessed the purity and morphology of the isolated cerebrovasculature through light microscopy. This examination unveiled sizable blood vessels that align with the characteristics of brain vasculature (Supplemental Figure 1C).

Using TMT-MS based quantitative proteomics we analyzed a total of 67 individual vessel preparations from postmortem frontal cortical tissues from the University of Pennsylvania School of Medicine Brain Bank (UPenn cohort) (Figure 1A, Supplemental Table S1). Importantly, in this study we also leveraged bulk proteomes from cortical samples from the same individuals profiled using an identical TMT-MS approach to better understand the relationship across bulk and vascular enriched proteomes (Figure 1B). A unified semi-quantitative method to assess both amyloid (CERAD) and tau (Braak stage) deposition was used to classify all cases, defined by these postmortem pathological and also clinical criteria, as neurologically healthy controls, AD, and PSP (UPenn cohort: N = 26 controls, 22 AD, and 19 PSP) (Supplemental Table S1). To characterize the vascular enriched fractions, we first evaluated contributions of core brain cell types to each proteome using a marker list extracted in silico from the data of a previously generated single-cell and single-nuclei RNA-seq study (Supplemental Table S2)²³. This analysis showed that the vascular proteome was enriched in endothelial cells in contrast to other brain cell types, with the notable exception of five samples spanning various diagnostic groups. These outliers displayed minimal enrichment of endothelial cells but demonstrated a significant increase in neuron-specific proteins, and thus were excluded from subsequent analyses (Supplemental Figure 2A). Following outlier removal we adjusted for the confounding effects of age, sex, and postmortem interval (PMI)²⁹. Moreover, in order to account for biases in sample fractionation, we normalized for neuronal-specific cell type signatures across all samples. This process resulted in the ultimate quantification of 9,854 vascular-enriched proteins across individual brain samples from controls, AD, and PSP groups (Supplemental Figure 2B and C). Comparably, the subsequent bulk proteome analysis led to the identification of 9,418 proteins. Among the identified unique protein gene products (rather than protein isoforms represented in the above counts) in both vascular and bulk proteomes, 1,183 were unique to the vascular proteome, 782 were only identified in the bulk proteome and the remaining 8,636 proteins showed overlap between both fractions (Supplemental Figure 3A). To investigate a relationship between both fractions in AD and PSP, we compared the effect size (log₂ fold change) of quantified proteins in bulk and vascular proteomes across diagnosis groups, overlaying log₂ fold enrichment in vasculature over bulk tissue protein (Figure 1C and D, Supplemental Table S3 and S4). This revealed a partial overlap between proteomic profiles in AD relative to control, with an enrichment in both proteomes of amyloid-associated proteins, that are related to the matrisome such as COL25A1, MDK, NTN1, SFRP1, SMOC1, HGF, SPOCK3 and reduction of neuronal-specific proteins, including PCSK1, VGF, and NPTX2. Interestingly, proteins associated with SMCs and pericytes (DES, PDLIM3, VIM, PTN, and OLFML3) showed an opposite directionality of change with their higher abundance in the AD bulk proteome and concomitant decrease in AD vasculature (Figure 1C). Vascular and bulk proteomic signatures discovered in PSP showed higher consistency in directionality of proteins compared to AD. More specifically, we found an increase in SAA1, FGA, C4BPA, CA1, and HP proteins and a decrease in many fibroblast-specific proteins (COL1A1, COL1A2, LUM, OGN) (Figure 1D).

• Download figure
• Open in new tab

Figure 1. Deep quantitative proteomics of human cerebrovasculature and whole brain tissue.

(A-B) Schematic representation of the experimental workflow for matched human postmortem tissues from non-demented controls (N = 28), AD (N = 21), and PSP (N = 18) cases from the University of Pennsylvania Brain Bank (UPenn) cohort. Isolated cerebrovasculature was obtained by dextran gradient centrifugation that allowed the separation of parenchyma/myelin from blood vessels. Both vascular and bulk fractions underwent enzymatic digestion followed by labeling with isobaric tandem mass tags and liquid chromatography mass spectrometry (TMT-MS). (C-D) Scatter plot of the log₂ effect size in the bulk and vascular proteome in AD vs control (C) and PSP vs control (D); color scale indicates the mean log₂ (cerebrovasculature/bulk) enrichment. (E) TMT-MS quantified levels of APP and MAPT in vascular and bulk preparations. P values were calculated using one-way ANOVA for overall groupwise difference. Box plots represent median, 25^th and 75^th percentiles while box whiskers extend to non-outlier measurements up to 1.5 times each nearest interquartile range. (F) APP and MAPT abundance levels were positively correlated to CERAD and Braak scores, respectively using biweight midcorrelation (Bicor). Student’s test for correlation significance p values are provided for each correlation with p < 0.05.

Utilizing a pairwise differential expression approach, we sought to identify individual proteins that distinguish the vascular and bulk proteomic datasets and also differentially expressed in AD or PSP compared to control cases. In the vascular fraction of AD, we observed significant increases in i) several transport proteins (ATP7B, SLC4A11, APOF), ii) proteins associated with the ECM (ADMATSL4, PAPPA, COL16A1), and iii) transcription factors with pivotal roles in angiogenesis and vascular homeostasis (FLI1, POU3F, NFTAC2), were significantly increased. Conversely, proteins that showed significant increases or decreases in AD and uniquely identified in bulk fractions in AD were associated with neuronal signaling and axon biology (ORAI2, SYT6, SYT10, GLDN, KLK6, TPH2, SHISA6) (Supplemental Figure 3B and C). Pairwise analysis of protein changes in PSP in the bulk or vascular fractions revealed several overlapping proteins with AD, including APOF, MLKL, SLC4A11, B3GAT2 providing evidence for existence of shared mechanism possibly related to the pathological accumulation of tau (Supplemental Figure 3B and D).

Given that proteomic assessment of amyloid precursor protein (APP) and microtubule associated protein tau (MAPT) provide surrogate measurements of Aβ and tau, respectively³⁴, we next evaluated their TMT-MS quantified levels in vascular and bulk preparations across all disease groups. As expected, the APP levels quantified by TMT-MS were significantly elevated in AD compared to controls and positively correlated with CERAD score, but there was no change observed in PSP (Figure 1E and F). Interestingly, MAPT levels showed significant increase in AD cases compared to controls and high correlation with Braak stages in vascular and bulk preparations potentially suggesting association of tau with the vasculature. Although we also observed an elevation of MAPT in the both bulk and vascular fractions, the effect was less dramatic in PSP than in AD compared to controls, highlighting the expected lower pathological burden of tau in frontal cortex of PSP cases³⁵ (Figure 1E and F). Given that Aβ40 is recognized for its enrichment in brain vasculature³⁶ and Aβ42 is considered a predominant form found in brain parenchyma³⁷, we further examined the MS-quantified levels of the unique C-terminal tryptic peptides (Aβ42: GAIIGLMVGGVVIA and Aβ40: GAIIGLMVGGVV) from both these amyloid species in both the bulk and vascular fractions. Our findings indicate that the isolated cerebrovasculature exhibited an >4-fold enrichment of Aβ40, while >4-fold levels of Aβ42 were observed in the bulk fraction of the brain compared to the vascular fraction. Furthermore, the elevated levels (>2-fold) of the Aβ40/ Aβ42 ratio in AD cases compared to control and PSP within the vascular fractions align with previous evidence linking higher levels of Aβ40 to vascular CAA³⁸ (Supplemental Figure 4A). In summary, AD demonstrated both shared and distinctive patterns of protein alterations in the vascular and bulk fractions with amyloid-associated proteins linked to the matrisome displaying the most significant effects. The overall magnitude of protein changes was notably lower in PSP compared to AD, aligning with the spatially distributed pathology observed in the frontal cortex in these cases.

Cerebrovascular proteomic network analysis identifies modules associated with brain vascular cell types and functions

Our previous work has identified a reproducible network of highly preserved groups of correlated proteins termed modules from bulk tissue that established a biological context for the human AD brain proteome and allowed discovery of new pathophysiological processes related to disease^29,30. To identify such biologically meaningful co-expression patterns in the cerebrovascular proteome, we applied an analogous network-based approach to the vascular-enriched proteome (Figure 2A). The constructed network highlighted expression of 93 protein modules across control and disease brain tissues, ranging in size from the largest module, M1 with n=812 protein members to the smallest module M93 with n=33 members (Supplemental Table S5). Gene ontology (GO) analysis was used to assign principal biology and specific molecular functions to each protein module using top GO terms (Figure 2A, Supplemental Table S6). Given that cellular composition can influence protein co-expression²⁸, we assessed the cell type nature of each protein module using cell type specific protein markers derived from scRNA-seq studies for enrichment of endothelial cells, pericytes, SMCs, fibroblasts, astrocytes, microglia/macrophage, neurons, oligodendrocytes, and oligodendrocyte progenitor cells (OPCs) (Supplemental Table S2)²³. In the present network we observed a significant relationship between module biology, disease status, and cell type specificity. Specifically, we found nine modules with prominent contribution of vascular cells, including three modules enriched with endothelial and pericyte markers: M16 ‘Cell Signaling’, M12 ‘Blood vessel development’, M5 ‘Maintenance of BBB’; four modules had abundant expression of SMC proteins, including M77 ‘Laminin/Endothelial cell (EC) proliferation’, M88 ‘MHC complex’, M28 and M90 Actin cytoskeleton/SMC contraction, and there were two modules that were associated with fibroblasts: M11 ‘ECM/Collagen fibril organization’, and M2 ‘Cell adhesion’ (Figure 2A). Using the ratio of vascular to bulk proteome to annotate a mean value for each vascular-enriched proteomic module, we next found an additional 25 modules comprised of proteins substantially enriched in the vasculature despite no obvious contribution of vascular cell types (Figure 2A). The analysis of paired vascular and bulk samples also allowed us to identify modules with novel proteins showing the strongest correlation to each module eigenprotein (i.e., ‘hub’ proteins). Importantly, we observed that modules enriched in cerebrovasculature had hub proteins that were previously unidentified or relatively depleted in the bulk proteome. This included M11 ‘ECM/Collagen fibril organization’, M35 ‘ECM organization’, M14 ‘Keratinization’, M88 ‘MHC complex’, M28, M90 ‘Actin cytoskeleton/SMC contraction’, M5 ‘Maintenance of BBB’, M77 ‘Laminin/Endothelial cell (EC) proliferation’, and M2 ‘Cell adhesion’ (Figure 2B). Thus, characterization of the protein systems biology of isolated cerebrovasculature significantly diverges from the bulk analysis, leading to identification of novel vascular-specific protein modules highlighting biological heterogeneity.

• Download figure
• Open in new tab

Figure 2. Protein co-expression network reveals novel modules enriched in cerebrovasculature.

(A) Protein co-expression network of cerebrovasculature was built using WGCNA as described in Methods, and consisted of 93 protein co-expression modules (N = 9,854 proteins, each module represented by a different color and number for decreasing rank size). Gene ontology (GO) analysis was used to identify representative the modular biology (outer track labels). Mean log₂ ratio of vascular to bulk proteomic abundance was performed to identify modules enriched in vascular proteome. The cell type association of each module was assessed by hypergeometric overlap of each module’s gene products with the cell type specific marker list for each brain cell type de novo extracted from single-nuclei RNA-seq data²³, curated into 9 primary cell types; one-sided Fisher’s exact test with Benjamini-Hochberg correction was employed, with colors other than white representing significant FDR < 0.05 (inner panel). Scale bars indicate log₂ ratio of vascular to bulk proteome (red indicates enrichment and blue indicates depletion) and cell type enrichment (from modestly significant to darker brown indicating stronger enrichment). Module relatedness is shown in the central dendrogram. (B) Top percent novel hubs identified in the TMT cerebrovascular network novel or ≥3-fold enriched relative to paired bulk protein measurements on average. Color scale represents gene product enrichment calculation using Fisher exact test with Benjamini-Hochberg FDR correction.

To further assess whether the protein modules identified in the vascular proteome were also present in the bulk protein network, we performed network preservation statistics, where the Z_Summary composite statistic indicates the degree of preservation of a module across networks. We found that 74 protein modules in the vascular network were preserved in bulk proteome with 17 modules showing high preservation (Z_Summary >10 or p<1×10^-23) and 57 modules showing moderate preservation (Z_Summary >1.96 or p<0.05) indicating the robustness of the majority of the modules across both proteomes derived from fractionation or the same bulk tissue (Figure 3). Consistent with previous findings²⁹, among the highly preserved modules we found protein communities commonly associated with AD pathophysiology, including M6 and M7 ‘Synapse/Neuron’, M39 ‘Oligodendrocyte/Myelination’, M10, M32 ‘Oxidative phosphorylation’, and finally M55 ‘Mitochondria’. Importantly, we also observed that 19 modules were only present in the vascular and not in the bulk network. Of these, 10 modules were enriched in cerebrovasculature, including M86 ‘Vesicle-mediated transport’, M14 ‘Keratinization’, M2 ‘Cell adhesion’, M77 ‘Laminin binding/EC proliferation’, M88 MHC complex, M90 ‘Actin cytoskeleton/SMC contraction’, M12 ‘Blood vessel development’, M18 ‘Glycosylation’, M31 ‘Glycosylation/ER’, M30 ‘Vesicle-mediated transport’, and M40 ‘Wnt signaling’ (Figure 3). This suggests that bulk proteomic analyses do not entirely capture differences on a network level in brain vasculature, emphasizing the significance of employing enrichment approaches to capture vascular changes in the brain proteome.

• Download figure
• Open in new tab

Figure 3. Cerebrovascular network preservation.

Module preservation of the TMT protein vascular network into the proteomes of paired bulk tissue case samples from frontal cortex of matched individuals. Modules with a preservation Z_summary score less than 1.96 (p value<0.05) were not considered preserved. Protein modules with Z_summary score above 1.96 were considered moderately preserved (modules above black line), whereas protein modules with Z_summary score above 10 (p value< 1×10^-23) were considered highly preserved (above grey line). Each bar is shaded according to the scale of mean pair cerebrovascular enrichment. Modules that showed no preservation in the bulk proteome but were enriched in vasculature are labeled in red.

Correlation of cerebrovascular proteomic modules with distinct neuropathological phenotypes in AD and PSP

Given the robustness of our cerebrovascular network, we set out to uncover pathophysiological alterations associated with protein modules in AD and PSP. To this end, we first correlated modules with disease state and neuropathological hallmarks, including amyloid and tau burden presented as CERAD and Braak scores, respectively, as well as CAA severity score, gliosis, and APOE4 genotype. We found that 66% of protein groups were strongly associated with AD and/or PSP and correlated with at least one neuropathological trait, whereas 34% of modules showed no substantial change in the disease conditions compared to controls. Specifically, among significantly changed modules 19 were altered only in AD, and 31 modules showed the same direction of change in AD and PSP, yet only 4 modules were different in PSP compared to controls, but not in AD (Figure 4A, Supplemental Table S7).

• Download figure
• Open in new tab

Figure 4. Co-expression network of cerebrovascular proteome resolves different proteomic signatures in AD and PSP.

(A) The WGCNA co-expression network constructed from 9,854 proteins yielded 93 co-expression protein modules. The heatmap shows the correlations of module eigenproteins with binary disease/control status, neuropathological hallmarks (CERAD, Braak, CAA, gliosis), and APOE Ɛ4 dose. Two-color heatmap represents strength of positive (red) or negative (blue) correlation, with P values provided for all correlations with P < 0.05. (B) Module eigenprotein (ME) levels grouped by case status were plotted for protein modules of interest. MEs were analyzed using one-way ANOVA test and unadjusted p values are provided for each module. Box plots represent median, 25^th and 75^th percentile while box whiskers encompass actual data points up to 1.5 times the nearest interquartile range.

Next, we assessed differences between disease states across the various vascular-relevant biological systems. The evaluation of eigenproteins, which reflect the first principal component expression profile for all proteins of a module, yielded two main clusters based on their association with disease class and vascular contribution. The first group exhibited significant changes in AD frontal cortex, but not in PSP when compared to controls (Figure 4A and B). It included modules strongly associated with vascular cells. Among these, six modules, including M11 ECM/Collagen fibril organization, M2 ‘Cell adhesion’ M77 ‘Laminin/Endothelial cell proliferation’, M88 ‘MHC complex’, M28 and M90 ‘Actin cytoskeleton/SMC contraction’ demonstrated a significant enrichment with fibroblasts and SMCs and showed negative correlation with neuropathological traits, particularly CERAD and CAA score, reflecting the specificity of vascular changes in response to amyloid pathology in AD. Modules M5 ‘Maintenance of BBB’ and M16 ‘Cell signaling’ were largely enriched with endothelial and pericyte functions, but displayed only a modest increase in the abundance levels in AD, possibly due to capturing less prominent alterations affecting capillaries compared to larger vessels in our cohort. This group also featured a module linked to ‘Matrisome/Heparin binding’ (M89) that showed a significant elevation in AD (p = 0.00001) and positive correlation to amyloid and tau deposition in the brain.

The second group comprised of biologically diverse modules with a significant increase in both AD and PSP and was further divided into modules depleted in blood vessels such as M3 ‘Neddylation’, M29 ‘Nuclear metabolism’, M33 ‘Fatty-acid metabolism’, and M56 ‘Glutamine/glutathione metabolism’ or enriched in the cerebrovasculature, including M47 ‘Blood coagulation/Acute-phase response’ and M87 ‘Erythrocyte metabolism’. Of note, M47 and M87 were the only modules in this group that displayed strong association with Braak scores but no correlation to amyloid burden suggesting the interplay between tau pathology and these vascular-related functional changes (Figure 4A and B). Collectively, cerebrovascular proteome network analysis revealed novel, heterogeneous biology influenced by hallmark pathological features of AD and PSP, and in many cases linked to brain vascular cells.

To examine molecular signatures of isolated cerebrovasculature distinguishing AD and PSP, we set out to identify significant protein alterations in each disease group compared to non-demented controls (Figure 5A and B). AD exhibited a larger number of differentially altered proteins than PSP in relation to controls, possibly due to more prominent structural and functional changes that occur due to amyloid deposition in cerebrovasculature in AD frontal cortex³⁹. Pairwise comparison of AD to controls outlined a large fold increase in M89 ‘Matrisome/ Heparin-binding’ proteins, such as SMOC1, MDK, C4B, SFRP1, and NTN1 with high specificity even as these proteins were not differentially expressed in the PSP versus control comparison. This protein signature, previously mapped to a bulk Brodmann area 9 “Matrisome” module²⁹, has been shown to consistently distinguish AD from other proteinopathies. These proteins either directly bind to amyloid or are associated with amyloid plaques, highlighting their crucial connection to amyloid deposition in the brain^29,33. As expected, MAPT was not uniquely altered in AD but showed higher elevation in AD than PSP, reflecting higher burden of tau pathology in frontal cortex in AD cases. The remaining signature of increased proteins in AD was associated with nuclear metabolism (SDC4, SPOCK3, CEP63) and fatty-acid metabolism (FABP7, TP22, and FKBP1A) (Figure 5A). Although, our previous study identified SDC4 and SPOCK3 as members of the ‘Matrisome’ module in bulk fractions²⁹, here these proteins were mapped to a module featuring MAPT, potentially suggesting their closer relationship to tau than amyloid in vasculature. In contrast, among proteins showing a steep reduction in abundance in AD, a number were involved in SMC function (ACTA2, TAGLN, TPM1, TPM2, PGM5) and extracellular matrix (DES, OGN, CPXM1, PAPPA, FMOD) which may suggest cell loss or subtype shifting in response to amyloid deposition in brain parenchyma and cerebral blood vessels (Figure 5A). In addition, consistent with previous reports²⁹, neuronal markers with known links to synaptic plasticity, such as VGF and NPTX2 showed declining trends in AD. Markers significantly altered in PSP across pairwise comparisons prominently featured members of M47 ‘Blood coagulation/ Acute-phase response’, including SPTA1, SPTB, RHCE, APOB and M87 ‘Erythrocyte metabolism’, including APOF, C4BPA, C4BPB, and PROS1. In summary, the frontal cortex vascular-enriched network outlines key differences in protein signatures of AD- and PSP-associated pathologies enabling differentiation between different disease states (Figure 5, Supplemental Figure 5).

• Download figure
• Open in new tab

Figure 5. Disease status demonstrates different cerebrovascular protein signature.

(A-B) Volcano plots showing differential abundance of proteins between control and AD groups (N = 3,622) (A) or proteins between control and PSP cases (N = 1,825) (B). The x axis illustrates the log₂ fold change (AD vs Control, A) or (PSP vs Control, B), while the y axis represents the -log₁₀ statistical p value calculated for all proteins in each pairwise group. P values were obtained from one-way ANOVA with Tukey’s post hoc test and for Tukey p values less than 10^-8.5, Bonferroni correction of two-tailed unequal variance t-test p values replaced highly significant but imprecisely estimated Tukey p values. Proteins significantly increased in AD (N = 1,541) (A) or PSP (N = 857) (B) are depicted in red (p < 0.05) whereas those significantly decreased in AD (N = 2,081) (A) or PSP (N = 968) (B) are highlighted in blue. Proteins of interest are shown as enlarged dots and shaded according to color of the module membership. Grey shaded dots represent proteins with unchanged levels. (C) Supervised hierarchical clustering of the 49 most significant proteins that were unique to vasculature or 3-fold enriched in vasculature compared to bulk and with volcano significance p < 0.05 across each of the three possible comparisons among control, AD, and PSP groups. Red shaded boxes highlight proteins upregulated and blue shaded boxes indicate downregulated proteins relative to the central tendency. Minimum and maximum log₂ (abundance/central tendency) are clipped at – and +4.

To determine if these differentially expressed proteins could function as classifiers for disease status and pathological abnormalities, we conducted a supervised clustering approach. This analysis was confined to the 49 most significant proteins that were either unique to vasculature or at least 3-fold enriched in vasculature compared to bulk and also differentially expressed (ANOVA+Tukey p<0.05) across all three comparisons between control, AD, and PSP groups. Notably, AD was easily distinguished from controls by significant elevation of proteins associated with M89 ‘Matrisome/Heparin binding’, M47 ‘Blood coagulation/ Acute-phase response’, M87 ‘Erythrocyte metabolism’, M8 and M36 ‘RNA splicing’, M29 ‘Nuclear metabolism’, M3 ‘Neddylation’, and M33 ‘Fatty-acid metabolism’. AD also displayed significant decrease of proteins in modules largely linked to SMC biology, mitochondria, and metabolic processes compared to controls. (Figure 5C). PSP vascular-enriched samples showed a similar pattern to AD for increased proteins but a more heterogenous proteomic profile of decreased proteins compared to AD or controls (Figure 5C). Collectively, the selected differentially altered proteins provide an AD-specific signature that diverges from controls and PSP.

AD GWAS candidate genes are overrepresented in modules enriched in cerebrovasculature

Integrated analysis of genetic association signals derived from large-scale genome wide association studies (GWAS) with the co-expression modules has a great potential to nominate pathways implicated in AD pathophysiology²⁸. To assess whether proteins encoded by GWAS AD risk genes were enriched in any particular modules in our network, we analyzed the single nucleotide polymorphism (SNP) summary statistics for three AD GWAS (Supplemental Table S8)^40-42 to calculate gene level association using three runs of the algorithm Multi-marker Analysis of GenoMic Annotation (MAGMA) as well as a PSP GWAS (Supplemental Table S9)⁴³. The gene-level MAGMA summary p-values for disease risk were considered as an ensemble for all three AD GWAS. Our vascular network featured five modules with overrepresentation of proteins from AD GWAS candidate loci, including M11 ‘ECM/Collagen fibril organization’, M67 ‘Mitochondria’, M77 ‘Laminin/Endothelial cell proliferation’, M89 ‘Matrisome/Heparin binding’, and M5 ‘Maintenance of BBB’ (Figure 6A). Strikingly, M11, M77, M89, and M5 modules showed significant enrichment in human cerebrovasculature, further underscoring the importance of vascular dysfunction in pathogenesis of AD. While a total of 34 candidate risk genes mapped to M11 ‘ECM/Collagen fibril organization’, including APOC1, APOC2, PVR, and GEMIN1, contribution of 14 gene products led to enrichment of M77 ‘Laminin/Endothelial cell proliferation’, including BCAM, CAV1, and CAV2. The strong association of M89 with AD was largely driven by the presence of APOE, the remaining 14 risk-associated genes showing less contribution to the module enrichment, including KNOP1, APP, CHADL, SMOC1, and SMOC2. Finally, M5 ‘Maintenance of the BBB’ constituted of 83 candidate risk-associated genes with mostly small effect size, with the exception of NECTIN1 that showed an overweighted significance of contribution to the enriched risk in the module.

• Download figure
• Open in new tab

Figure 6. AD GWAS risk genes are overrepresented in protein modules enriched for brain vascular cells.

(A) AD GWAS candidate genes were significantly enriched in five modules, including M11, M67, M77, M89, and M5. These modules were significantly up- or downregulated in AD, as shown in Figure 2A. The horizontal dotted lines indicate the thresholds for permutation test false discovery rate (FDRs of 10 or 5 percent), above which genetic risk enrichment was considered significant. Bar colors indicate mean all-protein module member mean log₂(cerebrovasculature/bulk) proteome fractionation enrichment. (B) GWAS risk genes for PSP were mapped to 10 modules with 5% FDR, among which two were significantly upregulated in PSP, including M29, and M47 (Figure 2B).

Gene-level risk enrichment of PSP causally associated gene products by module also identified candidate genes with MAGMA p-values< 0.05, and whose net risk was also significantly overrepresented in 10 modules. As expected, modules enriched for PSP GWAS risk genes differed from AD and were comprised of 4 cerebrovascular-specific modules, including M16 ‘Cell signaling’, M29 ‘Nuclear metabolism, M47 ‘Blood coagulation/ Acute-phase response’, and M43 ‘Immune pathway’ (Figure 6B). We did not find single candidate risk genes with overweighted significance driving the PSP GWAS risk association with protein modules. Instead, we observed modest contributions from multiple risk-associated genes, each contributing equally to the overall module enrichment of risk. Specifically, M16 featured 33 candidate genes, including CYTH1, AFAP1L2, and MICALL1, and M43 consisted of 30 gene products such as ACOX1, PLOD3, and IKBIP). Furthermore, M29 exhibited overrepresentation of genetic risk with 30 candidate risk-associated gene products, like B3GAT2, ARID4B, SDC4, and COL15A1, and finally M47 harbored 37 nominally significant PSP risk genes, including PF4, FBXW7, SERAINEF2, and A2M. Notably, we previously observed a significant increase in the abundance of M29 and M47 eigenproteins in PSP, further indicating the importance of these modules changing in disease development or progression. The other 6 modules enriched with PSP risk gene product proteins were depleted in cerebrovasculature but displayed significant enrichment for PSP GWAS were associated with neuronal and synaptic function, myelination, and metabolic processes (Figure 6B). In summary, these findings offer network-based evidence and organization of potential genetically linked factors contributing to vascular-associated changes occurring in AD or PSP, underscoring the critical role of cerebrovasculature in the etiology and pathophysiology of these diseases.

Matrisome proteins are present in cerebral blood vessels and associate with CAA

As we consistently identified matrisome/heparin-binding proteins as highly specific and sensitive discriminators of AD in both bulk and vascular proteomes, our focus turned to understanding their spatial relationship with neuropathological hallmarks of AD. Module M89, associated with the matrisome, was comprised of 39 members (Figure 7A), of which the vast majority of proteins were mapped to this module for the first time. Only 10 proteins of M89 overlapped with the previously identified Matrisome module (M42) in a consensus bulk proteome of AD brain²⁹, including SMOC1, NTN1, MDK, SFRP1, SLIT2, APP, APOE, NXPH1, TMEFF2, and SLIT1. In contrast, in depth quantification of cerebrovasculature enabled us to identify additional proteins linked to angiogenesis, ECM and SMC functions that overlap with ECM organization and collagen fibrillization. These M89 proteins included HGF, SLIT3, CHADL, ANOS1, SMOC2, HHIPL1, B4GALNT1, and CCN1, among others. As expected, the products of the APP processing, including Abeta40 and Abeta42 were also found in M89. Furthermore, we found that a few members of the previously described M42 ‘Matrisome’ module in bulk tissue, including SDC4, COL25A1, and SPOCK3, mapped to M29 (‘Nuclear metabolism’), which showed a significant correlation with both AD and PSP in our network. This suggests an association of M29 with tau-related changes. Other matrisome module proteins in bulk tissue, such as CTHRC1, OLFML3, FRZB, and COL11A1, were assigned to M11 in our proteome. Despite its association with ECM and collagen fibrilization, M11 exhibited a decrease in AD isolated vasculature, in contrast to the increase observed in M89.

• Download figure
• Open in new tab

Figure 7. Matrisome proteins are found in cerebrovasculature in AD and associated with CAA.

(A) An iGraph represents the top 50 proteins identified in the M89 Matrisome/Heparin binding module. Lines between proteins represent topological overlap matrix weight corresponding to the similarity of correlated patterns of node pairs over the 62 case samples in the cerebrovascular network. (B) Protein abundance of module M89 and selected module members across different CAA severity score. A 0 indicates no CAA, 1 indicates mild CAA, and 2/3 indicates moderate to severe CAA score. One-way ANOVA p values were assessed for each 3-group box plot. Box plots represent the median and 25th and 75th percentiles, with data points up to 1.5 times the nearest interquartile range beyond the box defining the extent of error bar whiskers. (C) Immunohistological evaluation of expression patterns of SMOC1, MDK, SMOC2, SLIT2, HHIPL1, and ITM2C in human postmortem cortical tissues from non-demented control, AD, AD with severe CAA, and PSP. Arrows indicate the presence of these proteins in the cerebrovasculature, the arrowheads indicate their expression in brain parenchyma. Scale bar, 100 μm. (D) Co-localization of SMOC1, MDK, and SMOC2 with amyloid in brain parenchyma and cerebrovasculature in human AD brain tissue. Thioflavin-S (thioS) stain was used to label fibrillar amyloid deposition. Arrows indicate the co-localization of these proteins with CAA, the arrowheads indicate their co-localization with amyloid plaques. Scale bar, 100 μm.

We next assessed whether previously and newly discovered module-member proteins were associated with CAA by plotting the selected M89 markers to each case sample’s CAA burden score, including no CAA (0), mild CAA (1), and moderate to severe CAA reflected by CAA scores of 2 and 3 (Figure 7B). We found that SMOC1, MDK, CHADL, and HGF displayed a stepwise increase in their abundance in relation to the severity of CAA. Meanwhile, SMOC2, SLIT2, and ITM2C showed a significant elevation in cases only presenting with moderate to severe CAA (Figure 7B). To further corroborate these findings, we employed immunohistochemistry to reveal the spatial distribution patterns in the frontal cortex across different disease states. Notably, our analysis revealed abundant accumulation of SMOC1, MDK, SLIT2, SMOC2, HHIPL1, and ITM2C localized to cerebral blood vessels in AD and AD presenting with severe CAA. This suggests that these proteins co-deposit with amyloid in the walls of cerebral vessels. Intriguingly, SLIT2 was also detected in cerebrovasculature in controls and PSP, indicating a specific basal vascular distribution pattern for this protein even in cases lacking vascular pathologies (Figure 7C). This was further supported by strong association of SMOC1, MDK, and SMOC2 expression with amyloid deposition in vasculature as CAA, indicated by thioflavine-S stain (Figure 7D). Although, SMOC1 and MDK also co-localized with fibrillar amyloid plaques in brain parenchyma, SMOC2 was nearly exclusively present in the Aβ-laden vessels, indicating its preferential vascular distribution pattern (Figure 7D). In summary, these findings highlight the discovery of new amyloid-associated proteins in the cerebrovasculature co-expression module, termed here ‘Matrisome/Heparin binding’ M89 module, which were not previously mapped to a ‘Matrisome’ module in bulk fractions²⁹. Importantly, the heightened signature of module members aligned with the prominent co-occurrence of CAA, demonstrates a high level of specificity of this novel proteomic signature in amyloid-related neuropathological changes.

Integrative analyses of brain and biofluid proteomes reveal biomarkers reflective of cerebrovascular changes in AD

Given that cerebrovascular dysfunction contributes to AD pathogenesis, and more than 90% of individuals with AD present with concurrent cerebrovascular disease³⁹, there is a strong rationale to identify peripheral biomarkers that reflect this underlying pathophysiological process. In this study, we employed an integrated analysis to identify plasma and CSF biomarkers that potentially reflect vascular changes in the brain. To this end, we leveraged previous datasets^44,45 (Supplemental Table S10 and S11) to map differentially abundant proteins in both AD plasma and CSF to the cerebrovascular network to prioritize peripheral protein signatures that were enriched in modules associated with vascular biology in brain (Figure 8A and B). Importantly, AD patients in these biofluid studies were confirmed to be biomarker positive by CSF ratios of Aβ42 and tau. Collectively, this includes two separate sets of plasma samples: set 1 (N = 36) and set 2 (N = 85) encompassing a total of 47 control and 62 AD samples in the plasma proteomics study (Supplemental Table S10)⁴⁴ and a single CSF set comprising of 141 control and 140 AD samples in the CSF TMT-MS proteomic study (Supplemental Table S11). Using each of these datasets independently, we identified brain vascular modules that were enriched with proteins significantly increased or decreased in AD CSF and plasma (Figure 8A and B, Supplemental Table S12-S15). Notably, this revealed a highly concordant and significant increase in many members of M89, the ‘Matrisome/Heparin binding’ module, including SMOC1 and SMOC2, in the brain, plasma, and CSF of individuals with AD (Figure 8C-F). Furthermore, we observed a high correlation between these proteins and well-established core biomarkers, including Aβ42, total tau, and phosphorylated tau at position 181 (pT181)^46-48 (Figure 9), making them promising biomarkers of vascular dysfunction in the AD brain. Interestingly, we also found that other M89 members such as MDK demonstrated increased trends in cerebrovasculature and plasma, but displayed a significant reduction in CSF suggesting that other regulatory mechanisms may play a role in their levels across biofluids. We also identified protein modules displaying divergent expression trends across brain, CSF, and plasma, including M11 ‘ECM/Collagen fibril organization’, M35 ‘ECM organization’, M47 ‘Blood coagulation/Acute-phase response’, M3 ‘Neddylation’, and M33 ‘Fatty-acid metabolism’, M74 ‘Glycolysis/Metabolic process, M56 ‘Glutamine/glutathione metabolism’, M58 ‘Purine metabolism’ (Figure 8). More specifically, the proteins associated with brain fibroblast biology and ECM within M11 and M2 were notably decreased in the cerebrovasculature and CSF in AD, yet there was a striking elevation in these proteins in AD plasma. Indeed, proteins such as BGN, CHST15, COL12A1, DCN, and COL14A1, among others exhibited the highest levels in AD plasma and demonstrated a significant correlation with conventional pathological traits (Figures 8 and 9). In summary, integrative cross-fluid proteomic characterization proves to be a promising tool for prioritizing peripheral biomarkers linked to CAA and cerebrovascular dysfunction in AD.

• Download figure
• Open in new tab

Figure 8. Integrated analysis of the brain vasculature and CSF and plasma proteomes in AD.

(A-B) Overlap of proteins identified in the discovery cerebrovascular proteome (N = 26 control and N = 19 AD) with separate plasma cohort (N = 47 control and N = 62 AD) (A) or CSF (N = 141 control and N = 140 AD) (B) discovery datasets. In the plasma proteome N = 2,865 proteins were identified, whereas in the CSF proteome N = 2,067 proteins were quantified. Separate hypergeometric overlap Fisher exact analyses were used to assess the significance of protein overlap. The top row indicates overlap between brain vascular and plasma or brain vascular and CSF proteome (grey shaded scale). The middle row demonstrates the overlap between brain vascular and plasma or brain vascular and CSF proteins significantly increased in AD (p<0.05) (red shaded scale). The bottom row depicts the overlap between brain vascular and plasma or CSF proteins significantly decreased in AD (p<0.05) (blue shaded scale). The intensity of color shading indicates the degree of the overlap. Statistical significance is shown in the heatmap using stars (* FDR< 0.05, ** FDR < 0.01, and *** FDR < 0.001). FDR corrected values were determined using the Benjamini-Hochberg method. (C-D) Volcano plots showing differential abundance of proteins measured in plasma (N = 1,528) (C) or proteins measured in CSF (N = 647) (D) between Control and AD groups. The x axis illustrates the log₂ fold change (AD vs Control), while the y axis represents the -log₁₀ statistical p value calculated for all proteins in each pairwise group, obtained as Tukey post-hoc test p values following one-way ANOVA, except for imprecisely calculated Tukey values below 10^-8.5 which underwent more precise and stringent Bonferroni post-hoc correction of a two-tailed unequal variance t-test. Proteins significantly increased in plasma in AD (N = 675) (C) or CSF in AD (N = 307) (D) are depicted in red (p < 0.05) whereas those significantly decreased in plasma in AD (N = 853) (C) or CSF in AD (N = 340) (D) are highlighted in blue. Proteins of interest are shown as enlarged dots and shaded according to the color of their module membership. (E-F) Scatter plots illustrate a Pearson correlation between log₂ effect size (AD vs Control) of significantly altered proteins in cerebrovasculature and plasma (E) or cerebrovasculature and CSF (F). The significance of Pearson correlation was determined by Student’s t-test for significance of correlation implemented in the WGCNA R package.

• Download figure
• Open in new tab

Figure 9. Specific cerebrovascular enriched proteins are detected in CSF and plasma, and strongly correlate with core AD biomarkers.

Scatterplots represent the correlation measurements of CSF A(₄₂, total tau and p-tau181 and plasma p-tau181 with six plasma (left panel) (total sample measurements: SMOC1 N = 36, SMOC2 N = 80, BGN, COL12A1, APOF, CPN1 N = 109) and CSF (right panel) proteins (total sample measurements for all six proteins N = 281). The Pearson correlation coefficient and Student’s p values are provided for each plot.