Loss of perivascular aquaporin-4 localization impairs glymphatic exchange and promotes amyloid β plaque formation in mice

We first sought to characterize the changes in AQP4 localization associated with clinical AD diagnosis, Alzheimer’s pathology, and measures of global cognitive and functional status in a post mortem case series from the Oregon Brain Bank. In contrast with our preliminary study that examined the effect of age and cognitive status on AQP4 expression and localization in human frontal cortical tissue [20], in the present study, we sought a detailed comparison between AQP4 localization and Alzheimer’s pathology across the AD clinico-pathological spectrum in old age. Thus, we selected only older subjects (>65 years old at death) that exhibited primarily Alzheimer’s neuropathology, excluding those with histopathological evidence of other processes such as vascular disease, Lewy body disease, and hippocampal sclerosis. The case series included frontal cortical tissues from 53 subjects, and hippocampal tissue from 36 subjects. This included subjects with a premortem clinical diagnosis of cognitively normal (CN), mild cognitive impairment (MCI), and AD. Demographic information (including sex, APOE4 allele status, education), cognitive and functional status (Mini-Mental State Examination, MMSE; Clinical Dementia Rating, CDR), and pathological status (Consortium to Establish Registry for Alzheimer’s Disease (CERAD) neuritic plaque score, Braak stage) are provided for subjects with frontal cortical and hippocampal slices in Table 1. Compared to CN subjects (median age 93.5 years, intra-quartile range (IR) [89.3, 98.2]) and MCI subjects (median age 94.9 years, IR [93.0, 98.8]), subjects with AD were significantly younger (median age 88.7 years, IR [81.3, 92.3]; P = 0.031, P = 0.003, respectively; 1-way ANOVA with Tukey’s post hoc correction). These groups did not differ in years of education or gender distribution. They exhibited expected patterns of APOE4 allele distribution, with more APOE4 carriers among the AD group. Expected differences were observed in MMSE, CDR, CERAD score, and Braak stage between CN and MCI groups but were not significant. AD groups exhibited expected reductions in cognitive and functional status, and higher CERAD scores and Braak stages.

Immunofluorescent double labeling for AQP4 and Aβ (4G8 clone) was conducted in both frontal cortical and hippocampal slices. Representative wide-field fluorescence micrographs show that AQP4 IF in CN subjects was largely uniform through the cortical layers between the pial surface and subcortical white matter boundary, with increased IF at perivascular endfeet surrounding the cerebral vasculature (Fig. 1A). In MCI and AD subjects, cortical AQP4 IF appeared more “patchy” and varied between cortical layers (Fig. 1B,C). Quantification of mean AQP4 IF revealed no consistent differences in relative AQP4 expression in the frontal cortical gray matter or white matter between CN, MCI, or AD subjects (Fig. 1D), and no differences in the CA1, CA2, or CA3 hippocampal subfields between CN, MCI, and AD subjects (Supplemental Figure 3A). As described in Supplemental Figure 1, the observed “patchiness” was evaluated by thresholding analysis to calculate the area coverage of cellular AQP4 IF. Using this approach, although cellular AQP4 IF coverage tended to be lower in MCI and AD subjects, measured differences between CN, MCI, and AD subjects were not statistically significant. No significant differences were observed between CN, MCI, and AD subjects in either frontal cortex (Fig. 1E), or in the CA2 or CA3 hippocampal subfields (Supplemental Figure 4B). AQP4 cell coverage was reduced in the CA1 subfield of MCI compared to CN subjects (Supplemental Figure 4B, P = 0.0086, one-way ANOVA with Dunnett’s post hoc correction).

Perivascular AQP4 polarization was quantified as the ratio between AQP4 IF along perivascular endfeet and the intensity in the cellular compartment (as detailed in Supplemental Figure 1). Perivascular AQP4 polarization in the frontal cortical gray matter was significantly reduced in AD compared to CN subjects (Fig. 1F; P = 0.0228, one-way ANOVA with Dunnett’s post hoc test). A similar reduction was observed in MCI compared to CN subjects, but this difference was not significant with correction for multiple comparisons (P = 0.0576, one-way ANOVA with Dunnett’s post hoc test). No significant differences in perivascular AQP4 polarization were observed in the frontal cortical white matter (Fig. 1F) or in the CA1, CA2, or CA3 hippocampal subfields (Supplemental Figure 4C).

We evaluated the association between changes in perivascular AQP4 polarization in the frontal cortical gray matter and measures of global cognitive function. When all subjects were evaluated without regard to clinical diagnosis, no significant association was observed between AQP4 localization and MMSE (P = 0.2852) or the CDR Sum of Boxes (CDR SoB) score (P = 0.455). Because Aqp4 deletion promotes Aβ deposition [18], proposed to be an early event in the Alzheimer’s pathological cascade [31], we evaluated whether changes in perivascular AQP4 localization were associated with cognitive decline prior to the development of frank dementia. Among non-demented subjects (CN + MCI), lower perivascular AQP4 localization was associated with lower MMSE scores (P = 0.035, R² = 0.279) and higher CDR SoB scores (P < 0.001, R² = 0.7832, Fig. 1G–H). Among subjects with AD, no association with MMSE (P = 0.720) or CDR SoB (P = 0.704) was observed.

We next evaluated the association between perivascular AQP4 polarization in the frontal cortical gray matter and either global or local measures of Alzheimer’s-related Aβ and p-tau pathology. Among non-demented (CN + MCI) subjects, reduced perivascular AQP4 localization was associated with higher CERAD neuritic plaque score, an ordinal measure of cortical Aβ plaque density (Fig. 1I, P = 0.017, R²=0.224). No association was observed among subjects with AD diagnosis. No association was observed between changes in perivascular AQP4 localization and Braak stage in either non-demented or subjects with AD (Fig. 1J, P = 0.328, P = 0.388, respectively). Because Braak staging assesses topographic spread rather than local intensity or density of p-tau pathology, we further evaluated pathological phosphorylated tau (p-tau) levels by IF using the AT8 (Ser202/Thr205, human/mouse cross-reactive) antibody in the frontal cortex and hippocampus. As expected, p-tau density was significantly increased in the frontal cortical gray matter (Supplemental Figure 5D, P = 0.011, 2-way ANOVA with Tukey’s post hoc test) and CA1 and CA2 subfields (P < 0.001) of AD subjects (Supplemental Figure 5A-D). Similarly, frontal cortical p-tau density was positively correlated with Braak stage (Supplemental Figure 5E, P = 0.011, R² = 0.139). Among non-demented (CN+MCI) subjects, reduced frontal cortical perivascular AQP4 polarization was associated with increased frontal cortical p-tau IF (Fig. 1K, P = 0.018, R² = 0.220), but no association was observed among subjects with AD (P = 0.172). These data demonstrate that subjects with AD exhibit reduced perivascular AQP4 polarization in the frontal cortical gray matter and that prior to the onset of dementia, loss of perivascular AQP4 polarization is associated with cognitive decline and local frontal cortical Aβ and tau pathology.

We next performed linear regression analysis to control for cognitive status in CN and MCI individuals (cognitive status as a covariate). Using this approach, reduced frontal cortical perivascular AQP4 localization was not significantly associated with reduced MMSE score (P = 0.200, adjusted (Adj) R² = 0.201), but was associated with higher CDR SoB score (P < 0.0001, Adj R² = 0.805) (Supplemental Figure 6A-B). Reduced perivascular AQP4 localization was associated with CERAD neuritic plaque score (P = 0.024, Adj R² = 0.156), was not significantly associated with Braak stage (P = 0.613, Adj R² = 0.002), and was significantly associated with local p-tau density (P = 0.030, Adj R² = 0.149) (Supplemental Figure 6C-E).

We next sought to replicate and extend findings from our prior study [13] to define whether similar changes in AQP4 localization are observed in the aging rodent brain as are observed in aging and AD human subjects ([20], Fig. 1). IF and confocal imaging revealed that as previously reported, cortical AQP4 in 3-month-old animals is predominantly perivascular (Fig. 2A) while in 15-month-old animals, there is an increase in diffuse AQP4 IF in astrocytes surrounding the penetrating vasculature (Fig. 2B). Using fluorescent lectin staining to label the cerebrovascular endothelium, we measured AQP4 IF in capillary-associated perivascular endfeet (“PV Endfoot”) and AQP4 IF in the surrounding non-perivascular neuropil (Supplemental Figure 2). Evaluating AQP4 IF at the level of capillaries, we found that microvascular perivascular AQP4 IF was significantly reduced in 15 compared to 3-month-old mice, while the non-perivascular AQP4 IF was not impacted by age (Fig. 2C, P = 0.0097, mixed effect model with Sidak post hoc test). We evaluated the AQP4 IF profile surrounding penetrating and large intracortical vessels (both arteries and veins). Segmenting the signal into three compartments: the immediate PV Endfoot region (0–1.5 μm from the vessel wall, “PV Endfoot”), the layer of astrocytes immediately bounding these penetrating vessels (“PV Astro,” 1.5–20 μm from the vessel wall), and the neuropil remote from these vessels (20–68 μm from the vessel wall), we observed that while PV Endfeet surrounding these large vessels did not differ in AQP4 IF between 3- and 15-month-old animals, the layer of astrocytes immediately surrounding these penetrating vessels exhibited increased diffuse AQP4 IF throughout their fine non-perivascular processes in 15 compared to 3-month-old animals (Fig. 2E, P = 0.004, mixed effect model with Sidak post hoc correction). When we evaluated the relationship between vessel diameter and AQP4 IF in these three segments, we observed that diffuse AQP4 IF in the PV Astro segment was significantly associated with increasing vessel diameter and that the correlation is more positive in 15-month-old mice than in 3-month-old mice (P = 0.0422; R²_15mo = 0.2410, R²_3mo = 0.03444) (Fig. 2F).

We used the same approach to evaluate the effect of age on AQP4 localization in the hippocampal microcirculation and large vessels of the stratum lacunosum moleculare (Supplemental Figure 7A-B). Unlike the cortex, there was no effect of age on capillary-associated perivascular AQP4 IF nor on non-perivascular AQP4 IF (Supplemental Figure 7C). Cross-sectional analysis of large vessels showed no change in AQP4 IF in PV Endfeet with age, but as in the cortex, increased AQP4 IF in astrocytes surrounding large vessels (PV Astro) was detected in 15-month-old compared to 3-month-old animals (Supplemental Figure 7D-E, P = 0.0003, mixed effect model with Sidak post hoc test). No significant association was observed between AQP4 IF in any segment and vessel diameter at either age (Supplemental Figure 7F). Together, these findings demonstrate that as in the aging and AD human cortex, two distinct changes in AQP4 localization occur in the aging rodent cortex: first, a loss of perivascular localization surrounding the cortical microcirculation, and second, an enhancement of AQP4 distribution over the fine, non-perivascular processes of astrocytes surrounding large penetrating vessels in both the cortex and hippocampus.

We next sought to model the distribution of AQP4 to the fine processes in astrocytes surrounding large penetrating vessels in the cortex and hippocampus, one of the phenotypic changes observed in the aging mouse (Fig. 2) as well as the aged and AD human cortex ([20], Fig. 1), using AAV-based gene delivery. All AAV constructs used in the present study are detailed in Supplemental Table 1. The two most abundant AQP4 isoforms, AQP4-M1 and AQP4-M23, exhibit divergent membrane localization profiles when overexpressed on an Aqp4-null background in primary astrocytes, with AQP4-M1 dispersing over the cell surface and AQP4-M23 forming large clusters within the cell membrane [32]. We first sought to confirm whether this was also observed following overexpression on a wild type Aqp4 background both in vitro and in vivo. We utilized an AAV with an astrocyte-specific promoter (AAV8-GfaABC1D) to drive expression of c-terminus enhanced green fluorescent protein (eGFP)-tagged versions of AQP4-M1 (AQP4-M1-eGFP; AAV8-M1-eGFP) and AQP4-M23 (AQP4-M23-eGFP; AAV8-M23-eGFP) in cultured primary mouse cortical astrocytes (Supplemental Figure 8A-B). Consistent with prior studies, overexpression of AQP4-M1-eGFP resulted in punctate green fluorescence distributed evenly over the cell surface, while overexpression of AQP4-M23-eGFP resulted in larger fluorescence puncta than those for AQP4-M1-eGFP (Supplemental Figure 8C-D). Puncta sizes were quantified, revealing that AQP4-M23-eGFP overexpression increased the ratio of large puncta (≥2 μm²) to small puncta (≤1 μm²) compared to that observed in AQP4-M1-eGFP overexpressing cells (Supplemental Figure 8E, P = 0.0540 unpaired t-test).

We next evaluated whether these differences in AQP4 membrane cluster size related to perivascular AQP4 localization in vivo. Using local intracortical injection of AAV8-M1-eGFP or AAV8-M23-eGFP, with in vivo 2-photon microscopy, perivascular localization of eGFP-tagged AQP4 was evaluated in reference to the cerebral vasculature labeled with an intravascular fluorescent tracer (70 kD Texas Red-conjugated dextran). AAV8-M1-eGFP or AAV8-M23-eGFP injection resulted in widespread cortical astroglial AQP4-eGFP expression (Supplemental Figure 8F). Visualization of individual astrocytes showed that AQP4-M1-eGFP overexpression resulted in small fluorescence puncta dispersed throughout astroglial fine processes, while AQP4-M23-eGFP puncta were larger in size (Supplemental Figure 8G). When perivascular compared to fine process localization of GFP fluorescence was evaluated, we observed that AQP4-M23-eGFP was significantly more perivascular in its localization than AQP4-M1-eGFP (Supplemental Figure 8H, P = 0.0457, unpaired t-test).

To assess the localization of AQP4 isoforms in vivo without an eGFP tag, we utilized a second AAV variant that efficiently crosses the blood-brain barrier and exhibits a high trophism for astrocytes, coupled with an astrocyte-specific promoter, AAV-PHP.B-GfaABC1D [33]. We generated one AAV^PHP eGFP reporter and four AAV^PHP AQP4 viral constructs. The reporter construct drove expression of eGFP under the astrocyte-specific GfaABC1D promoter (AAV^PHP-eGFP). The next two constructs drove expression of unlabeled AQP4-M1 (AAV^PHP-M1) or AQP4-M23 (AAV^PHP-M23) under control of this promoter. To distinguish between background and overexpressed AQP4 expression, the final two constructs used P2A sites to co-express AQP4 isoforms tagged with a hemagglutinin (HA) tag on the second external loop of the AQP4 protein, as well as an eGFP reporter (AAV^PHP-M1-HA, AAV^PHP-M23-HA) (Fig. 3A,B).

Thirty days after intravenous injection of the eGFP reporter construct (1 × 10¹² vg), eGFP expression was observed throughout the brain, including in all layers of the cerebral cortex and in subcortical structures (Supplemental Figure 9A). Immunofluorescent triple-labeling with the astroglial marker GFAP and the neuronal marker NeuN showed that eGFP expression was restricted to stellate GFAP-positive astrocytes and was not expressed in NeuN-positive neurons (Supplemental Figure 9B). Immunofluorescent labeling also demonstrated that GFP reporter expression was not present in CD31-positive endothelial cells (Supplemental Figure 9C-D).

Thirty days following retro-orbital virus injection of AAV^PHP-M1 or AAV^PHP-M23, robust expression of human AQP4 message was observed by qPCR (Fig. 3C). We next used AAV^PHP-M1-HA and AAV^PHP-M23-HA constructs to evaluate the localization of AQP4 isoforms after systemic virus administration. We observed widespread cortical and hippocampal expression of HA-tagged AQP4 with both constructs 30 days after injection. Super-resolution microscopy targeting individual eGFP-expressing cortical astrocytes following staining with a pan-AQP4 antibody (recognizing both wild type mouse and overexpressed human AQP4) revealed that overall, AQP4 localized preferentially to perivascular astrocytic endfeet. However, detection of the HA-human isoform using an anti-HA antibody showed no specific perivascular localization, but rather AQP4 distributed over all fine processes (Fig. 3D, top). In contrast, HA-tagged AQP4-M23 did localize to both perivascular endfeet and fine astroglial processes (Fig. 3D, bottom). Enhanced GFP expression was evaluated for enrichment proximal to large vessels or within capillary beds. Transduction of both large vessel-adjacent and capillary-adjacent astrocytes was observed for both M1 and M23 viruses (Fig. 3E, left). Quantification of the ratio of capillary-adjacent astrocytes to large vessel-adjacent astrocytes revealed a 2.24-fold higher transduction of capillary astrocytes compared to large vessel that was similar between both viruses (Fig. 3E, right). Evaluation of eGFP also permitted assessment of the extent and specificity of construct expression. As observed with the AAV^PHP-eGFP reporter (Supplemental Figure 9), eGFP distributed widely in volume-filling cells with morphology consistent with astrocyte-specific expression (Fig. 3D,E).

We next measured the effect of untagged AQP4-M1 and AQP4-M23 overexpression on overall perivascular localization of cortical and hippocampal AQP4. Immunofluorescence with a pan-AQP4 antibody followed by confocal imaging revealed that AQP4-M1 overexpression resulted in a diffuse pattern of AQP4 labeling in the fine processes of astrocytes throughout the cortex (Fig. 3F) and hippocampus (Supplemental Figure 10A). Quantification of AQP4 localization around cortical capillaries revealed that compared to injection of null vector, both AQP4-M1 and AQP4-M23 overexpression increased AQP4 expression at PV Endfeet (Fig. 3G, left; P = 0.0009, P = 0.0226, respectively, 2-way ANOVA with Tukey’s post hoc test). The non-perivascular neuropil AQP4 IF intensity and the cellular AQP4 coverage were significantly increased in the cortex of AQP4-M1-overexpressng animals (Fig. 3G, left; P = 0.0307, 2-way ANOVA with Tukey’s post hoc test, right; P = 0.0263, 1-way ANOVA with Tukey’s post hoc test) and similar results were observed in the hippocampus (Supplemental Figure 10B). When AQP4 localization surrounding large penetrating vessels was evaluated, we observed that while AQP4-M1 and AQP4-M23 overexpression did not alter AQP4 IF at PV Endfeet or in the surrounding neuropil, significantly increased AQP4 expression in the astrocytes surrounding large vessels (PV Astro) was observed following AQP4-M1 overexpression (Fig. 3H,I, P = 0.0394, 2-way ANOVA with Tukey’s post hoc test). AQP4-M1 overexpression did not alter AQP4 localization in astrocytes surrounding the large vessels of the hippocampal stratum lacunosum moleculare (Supplemental Figure 10C-D). In total, these data demonstrate that AAV-based overexpression of AQP4-M1 is a suitable approach for modeling the increased fine process AQP4 expression observed surrounding large vessels in the aging rodent brain (Fig. 2) [13].

Next, we sought to measure whether increasing fine process AQP4 localization by AQP4-M1 overexpression is sufficient to recapitulate the impairment in CSF tracer influx observed previously in aging mice [13]. Twenty-eight days after AAV^PHP-M1, AAV^PHP-M23, or null vector injection, 10 kD (Cascade Blue conjugate) and 70 kD (Texas Red conjugate) dextrans were co-infused intracisternally into the CSF of wild type animals. Ninety minutes later, the tissue was perfusion-fixed and parenchymal tracer influx was evaluated by imaging five coronal sections across the cerebrum (1, 0, − 1.5, − 2.5, and − 3.5 mm relative to bregma). Overexpression of either AQP4-M1 or AQP4-M23 had no effect on tracer influx compared to mice treated with null virus for either the 10 kD or 70 kD dextrans (Fig. 4B,C). This was observed both globally, and upon assessment of tracer influx into individual subregions (cortex, hippocampus, striatum, and the diencephalon).

To determine whether increased AQP4 localization to diffuse fine processes alters Aβ levels, we determined whether expression of AQP4 using an AAV-based approach would alter Aβ burden in Tg2576 mice [23]. Three-month-old Tg2576 mice were injected with AAV^PHP-M1, AAV^PHP-M23, or null vector, and 3 months later, Aβ levels were evaluated in the parenchyma (Fig. 4D). Assessment of AQP4 localization at this time point in AAVPHP-M1-treated Tg2576 mice showed that AQP4-M1 overexpression and the distribution of AQP4 to astroglial fine processes persisted at this time point (Supplemental Figure 11A-B). No differences were observed in either Aβ_1-40 or Aβ_1-42 levels with AQP4-M1 or AQP4-M23 overexpression compared to null vector-injected controls (Fig. 4E). These data suggest that increasing fine process AQP4 expression with viral AQP4-M1 overexpression does not appreciably alter either perivascular exchange or Aβ levels in the Tg2576 model.

We next sought to model the loss of microvascular perivascular AQP4 localization observed in the cortex of human aged and AD subjects ([20], Fig. 1), and aging mice (Fig. 2). The localization of AQP4 to perivascular astroglial endfeet results from its binding to the dystrophin-associated complex, a multi-protein complex including α-syntrophin (SNTA1), dystrobrevin, dystrophin, and dystroglycan linking AQP4 to the extracellular matrix of the basal lamina (Fig. 5A). Prior studies demonstrate that deletion of the α-syntrophin gene (Snta1) eliminates the perivascular localization of AQP4 [25, 34]. In our hands, Western blot demonstrated that Snta1 deletion abolishes SNTA1 expression (Fig. 5B, P = 0.0036, t-test). As reported in prior studies [35, 36], total AQP4 expression was not significantly different between Snta1^+/+ and Snta1^−/− mice (Fig. 5B). Immunofluorescence and confocal microscopy revealed a virtually complete loss of perivascular AQP4 localization surrounding the cerebral microcirculation (Fig. 5C), which was reflected in a dramatic reduction in capillary-associated PV Endfoot AQP4 IF (Fig. 5D, P < 0.0001, mixed effect model with Sidak’s post hoc test) with no change in AQP4 IF in the surrounding non-perivascular neuropil. Evaluation of AQP4 localization associated with large vessels showed that Snta1 gene deletion eliminated AQP4 enhancement at the PV endfeet (Fig. 5E,F; Supplementary Figure 10B-C) but did not alter the increased expression of AQP4 in astrocytes surrounding large penetrating vessels. Within the hippocampus, Snta1 gene deletion abolished perivascular AQP4 localization associated with the microvasculature (Supplemental Fig. 12A,B) but did not alter AQP4 localization at the endfeet surrounding the large vessels of the stratum lacunosum moleculare (Supplemental Figure 12C-D). This latter finding suggest that localization of AQP4 to the perivascular endfeet surrounding this population of vessels may be dependent upon a different set of adaptor proteins than around other vessels. More generally, these findings demonstrate that the Snta1-null mouse models well the loss of perivascular endfoot AQP4 localization, but not the increase in non-perivascular fine process AQP4 expression, observed in the aging brain.

CSF tracer influx into brain tissue was next evaluated using whole-slice fluorescence microscopy following intracisternal tracer injection in both wild type and Snta1^−/− mice. At 90 min post-injection, penetration of the 10 kD Cascade Blue-conjugated dextran was generally reduced into deep brain structures of the Snta1^−/− compared to wild type mice. Regional analysis showed that this effect was significant in the hippocampus (Fig. 6A,B, top; P = 0.0064, 2-way ANOVA with Sidak’s post hoc test). A trend towards reduced tracer distribution was also observed in the striatum and diencephalon though the difference did not persist following correction for multiple comparisons. Penetration of 70 kD Texas Red-conjugated dextran was impaired in Snta1^−/− compared to wild type mice, with significant reductions in the hippocampus and diencephalon (Fig. 6B, bottom; P = 0.0004, 0.0066, 2-way ANOVA with Sidak’s post hoc test). Consistent with our previously reported dynamic contrast-enhanced MRI studies [17], Snta1 gene deletion did not significantly reduce CSF tracer distribution when evaluated at 30 min post-injection (data not shown), suggesting that the phenotype of the Snta1^−/− is relatively muted compared to that observed with global Aqp4 gene deletion [11, 14, 17, 37].

To define the effect that loss of perivascular AQP4 localization has on interstitial solute efflux, we evaluated the distribution of 10 kD Cascade Blue-conjugated dextran following intraparenchymal injection in Snta1^−/− and wild type mice. Tracer was injected directly into motor cortex and allowed to distribute for 120 min, and interstitial tracer distribution across brain regions was evaluated in fixed coronal slices by whole-slice fluorescence microscopy. Distribution was defined by area coverage of tracer fluorescence within the ipsilateral and contralateral cortex, hippocampus, striatum, and diencephalon integrated through five standard coronal slices per animal (2.5, 1.5, 0, −1, and −2 mm relative to bregma, Supplemental Figure 13A). Following intracortical injection, the 10-kD tracer distributed through the cortex and into the subcortical regions (Supplemental Figure 13B). When compared between wild type and Snta1^−/− mice at 120 min after injection, tracer fluorescence integrated globally across all slices was generally higher in Snta1^−/− than in wild type animals (Supplemental Figure 13C-D, P = 0.0468, t-test), with the most pronounced differences within deeper brain structures such as the striatum (P = 0.0340, 2-way ANOVA with Sidak’s post hoc test). These data demonstrate both CSF tracer influx and interstitial tracer efflux are impaired in Snta1^−/− mice.

To evaluate whether loss of perivascular AQP4 localization increases Aβ levels, we evaluated the effect of Snta1 gene deletion within the Tg2576 amyloidosis line, comparing soluble and insoluble Aβ levels by ELISA in Tg2576:Snta1^−/− mice and Tg2576:Snta1^+/+ littermate controls at 6 months of age (Fig. 6C). Both soluble and insoluble Aβ_1-40 levels were significantly increased in Snta1-null mice (Fig. 6D,E, P = 0.0022, P = 0.0277, respectively; t-test). Soluble Aβ_1-42 levels were significantly elevated with Snta1 gene deletion (Fig. 6F, P = 0.0111, t-test), while the change in insoluble Aβ_1-42 levels did not reach statistical significance (Fig. 6G, P = 0.1395, t-test). To investigate which changes in AQP4 localization are most closely related to these increases in Aβ deposition, we evaluated the relationship between Aβ levels and cortical and hippocampal capillary- and large vessel-associated measures of perivascular and astroglial AQP4 IF measured in from the same animals. The complete results of these association studies are shown in Supplemental Figure 14. The most consistent association observed was between declining cortical capillary-associated PV Endfoot AQP4 IF and increasing soluble Aβ_1-40, Aβ_1-42, and insoluble Aβ_1-40 (Fig. 6D–G). Reductions in cortical large vessel-associated perivascular endfoot AQP4 were similarly associated with soluble Aβ_1-40 and Aβ_1-42. These associations were restricted to changes in cortical AQP4, as AQP4 IF measures in the hippocampus were not significantly associated with changes in any Aβ measure (Supplemental Figure 14). When a similar association study was undertaken evaluating the relationship between changes in capillary- and large vessel-associated fine process AQP4 expression in the AAV-treated tissue described in Fig. 5 above, no association was observed between soluble or insoluble Aβ_1-40 or Aβ_1-42 levels and fine process AQP4 IF either in the cortex or hippocampus (results of the association study are shown in Supplemental Figure 15). These findings suggest that loss of AQP4 from cortical perivascular astroglial endfeet specifically promotes the deposition of Aβ while the distribution of AQP4 to non-perivascular fine processes does not impact Aβ deposition.

In a prior human post mortem histopathological study, we reported that in the frontal cortex perivascular AQP4 localization declined with age and that this decline associated with dementia diagnosis [20]. The present study extends these prior findings in a substantially larger case series including frontal cortical (n = 49) and hippocampal (n = 37) samples from subjects over 65 years of age spanning the clinical spectrum from cognitively normal, to MCI, to AD. This case series was developed to directly examine the relationship between changes in AQP4 localization and the development of Alzheimer’s pathology, selecting only cases with a clinico-pathological diagnosis of either “Alzheimer’s disease” or “Alzheimer’s-related pathological changes,” and excluding subjects with significant non-AD pathology. The resulting case series was largely independent from that of our prior study (only 6 overlapping cases). The present study extends prior findings by demonstrating that while subjects with a clinical AD diagnosis exhibited reduced perivascular AQP4 localization compared to cognitively normal individuals, subjects with a clinical diagnosis of MCI exhibited an intermediate AQP4 localization phenotype. In addition, the present study evaluated AQP4 localization in frontal cortical gray matter, frontal cortical white matter, and in the CA1, CA2, and CA3 hippocampal subfields, observing that the association between reduced perivascular AQP4 localization with cognitive status prior to death was specific to changes within the frontal cortical gray matter.

The inclusion of subjects across the spectrum of clinical and pathological AD permitted us to evaluate the association between changes in frontal cortical perivascular AQP4 localization, local measures of Alzheimer’s pathology, and clinical and functional decline in detail. Our findings from the present rodent studies demonstrate that changes in perivascular AQP4 localization are sufficient to impair CSF-ISF exchange and promote Aβ deposition. Aβ deposition begins early in the AD pathological cascade, while Aβ burden plateaus in the course of clinical progression [38–41]. The present observation that among cognitively intact and MCI subjects reduced perivascular AQP4 localization was associated with higher CERAD neuritic plaque scores is consistent with a role for changes in AQP4 localization in human amyloid β deposition.

Similarly, AQP4-dependent glymphatic exchange has been implicated in the clearance of tau, and the impairment of this perivascular exchange has been proposed to account in part to the development of age-related and post-traumatic tauopathy [14, 42]. When we measured local pathological tau burden within the frontal cortex, we observed that as with amyloid β pathology, declining perivascular AQP4 localization was associated with increased p-tau burden in non-demented subjects. These findings were paralleled by associations between reduced perivascular AQP4 localization and lower scores on measures of cognitive and functional status, the CDR SoB. The fact that these associations were maintained only in non-demented subjects, and not in subjects with clinical AD diagnosis, argues that changes in AQP4 localization may be exerting their effect early in the AD pathological process, rather than in later stages of the disease process. Our observation that the association with CDR SoB score and with local measures of AD-related neuropathology remained significant even when controlling for cognitive status using partial correlation analysis argues that these effects do not simply reflect differences in neuropathology across non-demented groups.

One key limitation of any post mortem histopathological study is the inability to ascertain causality. The association of changes in perivascular AQP4 localization with measures of pathological and clinical AD progression may result from the activation of cortical astrocytes associated with the development of amyloid β plaques. The present experimental studies in rodents demonstrate that such changes in AQP4 localization are sufficient to promote Aβ pathology. However, they do not confirm that this actually occurs in the setting of human disease. Genetic association studies may provide important insight into the role that perivascular astroglial biology may play in the development of AD in human populations. For example, we reported in a limited candidate gene-association study that five naturally occurring single-nucleotide polymorphisms (SNPs) in the human AQP4 gene are associated with cognitive impairment in the setting of AD [43]. In two subsequent studies, SNPs in the AQP4 gene were observed to moderate the relationship between sleep disruption and amyloid β pathology [44] and to be associated with variation in amyloid β burden and clinical AD progression [45]. These studies, while preliminary, support a causal link between AQP4 and the development of AD pathology and AD clinical progression. No study to date has evaluated whether SNPs in the genes whose products determine perivascular AQP4 localization, such as a-syntrophin (SNTA1), dystrobrevin (DTNA), or dystroglycan (DAG1) are associated with variation in AD pathology or progression. This question remains an important subject for future research.

In the initial description of the glymphatic system, perivascular CSF-ISF exchange and interstitial Aβ clearance were observed to be impaired by Aqp4 gene deletion [11], suggesting a key role for glial water transport in perivascular fluid and solute movement. Although one study failed to replicate this observation [46], a subsequent study including findings from five different labs and using five independent transgenic mouse lines [17] confirmed the initially observed role of AQP4 in perivascular CSF-ISF exchange. More recently, studies utilizing an AQP4-specific inhibitor TGN-02 0[47, 48] report that acute inhibition of AQP4 slows perivascular exchange and Aβ clearance [42, 49]. These findings are corroborated by studies demonstrating that Aqp4 gene deletion in mouse models of amyloidosis accelerates Aβ deposition [18, 50].

Our prior findings show an association between loss of perivascular AQP4 localization and impairment in glymphatic function in the aging, post-traumatic, and ischemic brain [13–15], yet these studies did not directly test whether this loss of perivascular localization reduces CSF-ISF exchange or alters the development of AD pathology. The present observation that Snta1 gene deletion, which abolishes cortical perivascular AQP4 localization, disrupts CSF tracer influx and interstitial solute efflux while increasing Aβ deposition supports the overall role of AQP4 in perivascular CSF-ISF exchange. It also provides the first causal evidence that loss of perivascular AQP4 localization impairs CSF-ISF exchange and can promote neurodegenerative pathology. It is important to note that the loss of perivascular AQP4 localization in the Snta1^−/− mouse model is more pronounced than that observed in the post mortem cortex in the setting of human aging and AD. Unlike the development of Aβ and tau pathology in transgenic mouse lines that occurs over the course of months, the pathological accumulation of Aβ and tau aggregates in the human brain occurs over the course of years prior to the onset of AD-related clinical symptoms [40]. In addition, the anatomical distances for perivascular Aβ clearance in the human brain are much greater than those in the rodent brain. These differences in anatomical scale may permit relatively subtle changes in AQP4 localization to contribute to the development of neuropathology in the human brain over the extended timescale observed in the setting of human AD.

Whether changes in SNTA1 expression, or changes in the expression of other elements of the astroglial dystrophin-associated complex underlie age- and injury-associated changes in AQP4 localization has not yet been defined. Indeed, recent transcriptomic profiling studies from mice and human brain tissue have identified several potential perivascular endfoot elements whose expression profiles in development and the setting of AD parallel that of AQP4. Expression levels of these elements predict dementia status and p-tau levels in the aging human brain [51–53] and dystrophin-deficient astrocytes have altered astrocyte function and altered AQP4 levels [54]. Whether these putative endfoot elements contribute to changes in AQP4 localization, perivascular CSF-ISF exchange, or the development of AD pathology remains unknown. Similarly, whether SNTA1 expression regulates their localization or function has not been evaluated.

It is noteworthy that the reduction in perivascular CSF-ISF exchange described in the present study is more modest than observed in Aqp4^−/− mice [11, 14, 17, 37]. For example, differences in CSF tracer distribution between wild type and Snta1^−/− mice were evident 90 min, but not 30 min following intracisternal injection, whereas differences between Aqp4^−/− and wild type mice are readily observed at the 30 min time point. We also observed regional, rather than global changes in CSF tracer influx in the Snta1 knockout line. If AQP4 supports perivascular CSF-ISF exchange, then it is not surprising that a genetic model lacking perivascular localization yet preserving overall AQP4 expression exhibits a regional phenotype relative to a global knockout. The regional changes in CSF tracer influx observed in the hippocampus and diencephalon following Snta1 gene deletion in Tg2576 mice may have important implications for AD. In AD, both Aβ and p-tau pathology arise earliest in specific neuroanatomical locations, with pathology propagating or spreading to other involved regions over time. The hippocampus is an early site of AD pathology [55] and plays a crucial role in memory formation [56]. Thus, the observed vulnerability of hippocampal glymphatic clearance to changes in AQP4 localization in rodents may reflect the role that local differences in baseline rates of glymphatic clearance or in age- and injury-associated impairment in glymphatic function may play in the neuroanatomical distribution of AD-related pathology.

In mouse models of aging [13], traumatic brain injury [14, 21], and ischemic injury [15, 22], the loss of AQP4 from perivascular endfoot domains is accompanied by distribution of AQP4 to fine non-perivascular processes. In the setting of aging and AD, present and prior [20] studies in post mortem human tissue demonstrate the emergence of a dispersed, cortical astrocyte population with intense AQP4 immunoreactivity spread throughout its fine processes which underlies the development of a “patchy” pattern of AQP4-IF throughout the cortical depth. The biological rationale and trigger for this upregulation remain unclear. One possibility is that blood-brain barrier disruption, a process that occurs in aging and neurodegeneration [9, 57], promotes reactive astrogliosis and upregulation of AQP4. It is possible that impaired perivascular clearance resulting from astrogliosis and the loss of perivascular localization occur in parallel with blood-brain barrier disruption, with both processes contributing to slowed Aβ clearance. It is also possible that waste accumulation at the perivascular compartment due to loss of perivascular AQP4 localization is a precursor to widespread cerebrovascular neuroinflammation and downstream blood-brain barrier disruption. Clearly, further investigation into the interrelationship between these processes is warranted.

In one arm of the present study, we tested whether the distribution of AQP4 to astroglial fine processes, rather than its removal from perivascular endfoot processes, could impair perivascular CSF-ISF exchange and alter Aβ deposition. In agreement with prior in vitro studies [32], we observed that the AQP4-M1 and AQP4-M23 isoforms exhibit distinct subcellular localization profiles when expressed both in vitro and in vivo. AQP4-M1 localized to small puncta dispersed over the fine processes of astrocytes while AQP4-M23 clustered into larger puncta that concentrated around perivascular astroglial endfeet. This is consistent with previous reports that suggest that palmitoylation sites specific to the NH₂ terminus of the AQP4-M1 isoform inhibit its assembly into orthogonal arrays and localization to perivascular endfoot domains [58]. These results suggest that the differential expression of AQP4-M1 and AQP4-M23 isoforms in vivo may contribute both to the localization of AQP4 to perivascular astroglial endfeet as well as to the localization of AQP4 to fine non-perivascular processes under pathological conditions.

The viral overexpression of AQP4-M1 resulted in AQP4 localization similar to the distribution to fine non-perivascular astroglial processes observed both in the aging rodent and human AD brain. We observed that neither the overexpression of AQP4-M1 nor of AQP4-M23 altered either perivascular exchange or Aβ deposition. Indeed, while variation in cortical perivascular endfoot localization of AQP4 was significantly associated with Aβ burden among Tg2576:Snta1^−/− and Tg2576:Snta1^+/+ mice, no significant associations between non-perivascular AQP4 localization and Aβ burden were observed with AQP4-M1 or AQP4-M23 overexpression. These results, along with those from the Snta1^−/− mice, suggest that it is the loss of AQP4 from perivascular astroglial endfoot domains, and not the distribution of AQP4 to non-perivascular fine astroglial processes that impairs CSF-ISF exchange and the clearance of Aβ in the aging brain.

A recent study demonstrated that AQP4 localization is regulated dynamically, undergoing changes in subcellular localization an plasma membrane insertion over a much shorter timescale than previously appreciated [59–61]. Whether such dynamic regulation of subcellular AQP4 localization is altered in the setting of aging or AD, and whether such changes impair glymphatic function or contribute to the development of AD pathology remains an important subject for future inquiry. In addition, two independent groups have reported that AQP4 exhibits an exceptionally high rate of translational readthrough, resulting in the common occurrence of a novel isoform AQP4-X (also termed AQP4-ex) [62, 63]. This variant exhibits robust perivascular localization and is less dramatically upregulated in response to reactive astrogliosis than other AQP4 isoforms [63]. While the present study suggests that upregulation of AQP4-M1 and AQP4-M23 likely do not account for impairment of perivascular exchange and Aβ clearance in the aging brain, whether changes in AQP4-X expression or localization regulate perivascular clearance mechanisms, remain an important, unanswered question. Although it is reasonable to suppose that perivascular AQP4-X localization is dependent on SNTA1 expression, this has never been directly tested.

Several open questions remain regarding the fundamental mechanisms of glymphatic function, perivascular CSF-ISF exchange, and their role in the pathogenesis of AD. It is currently unclear precisely how AQP4 supports macroscopic perivascular CSF-ISF exchange throughout the brain. While computational studies of CNS fluid dynamics on the microscopic (cellular) scale have failed to fully explain the role of AQP4 in perivascular exchange [18, 63], findings from an increasing number of studies carried out in Aqp4^−/− [11, 14, 17, 37], in Snta1^−/− mice ([17], present findings) and with AQP4 inhibitors [42] clearly demonstrate a critical role for perivascular AQP4 in CSF-ISF exchange. A second, related limitation is the fact that our analysis of perivascular AQP4 localization in both the human and rodent brain did not distinguish between localization surrounding arterial versus venous segments. The initial characterization of the glymphatic system described the fact that perivascular AQP4 localization was greater surrounding cortical ascending veins than around penetrating arteries (Supplementary Figure 7 in [11]), a pattern of localization that may contribute to the organization of fluid and solute movement through brain tissue. In the present study, neither AQP4-M1 overexpression nor Snta1 gene deletion preferentially targeted artery versus vein-associated perivascular astrocytes. Thus, it cannot shed light on the role that vascular segmental differences in AQP4 localization play in the supporting perivascular exchange and Aβ clearance. Furthermore, our study examined CSF-ISF exchange and Aβ levels in 6-month-old Tg2576 following manipulation of AQP4 localization via either AAV treatment or when crossed with Snta1 KO mice. This 6-month timepoint was chosen in order to be able to detect accelerated progression of Aβ pathology in these mice and to avoid a possible ceiling effect in Aβ levels in older Tg2576 mice. However, similar experiments conducted in further aged mice may show a similar phenotype. Similar results were seen in the APP/PS1 line when crossed with AQP4 KO at 12 months of ages, as published by Xu et al. [18]. There are several limitations to our AAV approach; given that we administered our AAVs intravenously via the retro-orbital sinus, we cannot rule out the possibility that the transduction efficiency around capillaries may have been higher than for large vessels and the wider neuropil. In addition, intravenously administered AAV-PHP viruses have been reported to be more efficient at transducing astrocytes in female mice [64]. Both sexes were represented equally in all groups of mice and obvious stratification was not seen within virus injected groups for any of the analyses; however, the AAV transduction efficiency could have been more efficient in our female mice.

The research protocols for studies in human tissues were reviewed and granted ethical approval by the OHSU Institutional Review Board. All animal experiments were performed in accordance with state and federal guidelines and all experimental protocols were approved by the OHSU institutional animal care and use committee.

Not applicable.

The neuropathological case series included in this study was funded through a Sponsored Collaborative Agreement with GlaxoSmithKline to JJI. The other authors declare that they have no competing interests.

The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.