Loss of aquaporin-4 results in glymphatic system dysfunction via brain-wide interstitial fluid stagnation

All MRI scanning was performed at 9.4T MRI device (BioSpec 94/30USR, Bruker BioSpin, Ettlingen, Germany) in the head-first prone position with animal’s body temperature maintained at 37 ± 1°C with a thermostatically-controlled waterbed and monitored, along with the respiratory rate, by an MR compatible remote monitoring system (SA Instruments, NY, USA).

To achieve high-spatial resolution of MRI CSF space volumetry and cisternography a 3D constructive interference in steady-state (CISS) sequence along with a cryogenically cooled Tx/Rx quadrature-resonator (CryoProbe, Bruker BioSpin) and 240 mT/m gradient coil (BGA-12S) were used. During MRI, the animals (5 KO and 6 WT) were anesthetized under Ketamine/Xylazine (i.p. K/X: 100/10 mg/kg) and underwent acquisition of two 3D-TrueFISP volumes of opposite phase encoding direction (i.e.: 0° and 180°) (Table 1B), for further 3D-CISS image calculation. The complete MRI protocol lasted over an hour so every animal was implanted with a permanent intraperitoneal PE-10 catheter in the abdominal area, connected to a 1 mL syringe filled with K/X solution. The syringe was kept outside MR during whole imaging, and animals received a single supplementary dose of K/X after the first TrueFISP volume acquired. No difference in age (p=0.697 for KO vs. WT using Mann-Whitney U-test), body weight (p=0.7662), respiration rate during MRI (p>0.99) as well as signal-to-noise ratio (SNR) of computed 3D-CISS images (3.5 ± 0.4 vs 3.7 ± 0.2; p=0.1385) was found between KO and WT animals (Table 1A) so no animals were excluded from further analysis.

To assess differences in the brain water mobility between the genotypes, echo-planar-imaging (EPI)-based diffusion-weighted imaging (DWI) was performed using a room-temperature volumetric Tx/Rx resonator (in. ø40 mm) and 1500mT/m gradient coil (BFG6S, Bruker). The animals (6 KO and 6 WT) were anesthetized with K/X (i.p. 100/10 mg/kg) and underwent DWI with 17 b-values measured in 3 orthogonal directions of diffusion encoding gradients (Table 1B). To reduce the effect of respiratory motion (Federau et al., 2013), all DWI images were acquired with respiratory-gating in exhale, assisted by the remote monitoring system (see above). To minimize the influence of deep anesthesia and long scanning time on the measurements, the imaging protocol encompassed solely DWI lasting <40 min so no supplementary K/X was required. No difference in age (p>0.99 for KO vs. WT; Mann-Whitney U-test), body weight (p=0.935), respiration rate during MRI (p=0.632) was found between KO and WT so no animals were excluded from further analysis (Table 1A).

It is worth mentioning, that there is an ongoing debate on efficacy of IVIM modelling in reflecting phenomena in microvascular network or related to tissue microarchitecture (Fournet et al., 2017; Meeus et al., 2017; Paschoal et al., 2018; Schneider et al., 2019; Niendorf et al., 1996). We have aimed to provide an optimal setup for DWI (Liao et al., 2021; Lemke et al., 2011) by measuring MR diffusion signal using 30ms echo time (TE) and 17 b-values with increased averaging for b-values≥1000 s²/mm (see Table 1B). Based on or preliminary assessment (unpublished) application of higher than minimal available (here minimal ~22ms) TE would reduce the influence of ghosting and perfusion-related artefact, and higher averging would reduce the influence of possible Rician noise at high b-values. Although IVIM estimates were reported to depend on TE (Führes et al., 2022; Bisdas and Klose, 2015), mostly for perfusion fraction, our evaluation focused predominantly on the slow diffusion component. Furthermore, by sampling signal up to 3000 s²/mm b-value, which may lead to presented slight lower ADC values, we aimed for depicting dominant signal from extracellular space at high b-values (Le Bihan, 2019; Cihangiroglu et al., 2009; Clark and Le Bihan, 2000; Niendorf et al., 1996). However indicatively useful, higher order models and the models focused on separating hindered MR diffusion signal according to assumption on microarchitecture (Latour et al., 1994; Palombo et al., 2020; Burcaw et al., 2015; Kaden et al., 2016; Wu and Zhang, 2019; Olesen et al., 2022; Pfeuffer et al., 1998) or assessment of diffusion signal distribution (Roth et al., 2008; Benjamini and Basser, 2019; Slator et al., 2021; Ronen et al., 2006) are beyond presented general evaluation.

For DCE-MRI, cisterna magna (CM) cannulation based on the previous studies (Xavier et al., 2018) was performed in mice (5 KO and 6 WT; Table 1A) anesthetized with Ketamine/Dexmedetomidine (i.p. K/Dex: 75/1 mg/kg). After exposing CM, a 30 G copper cannula (out. ø0.32 mm; Nippon Tokushukan, Mfg, Tokyo, Japan), attached to a PE-10 tube filled with aCSF, was inserted into CM and fixed in position with a drop of cyanoacrylate glue followed by a drop of glue accelerator. The incision site and the skull were then covered by a mixture of cyanoacrylate glue and dental cement. Subsequently, the exposed PE-10 tubing was attached to a cannula filled with contrast agent (gadobutrol, 20 mM; Gadovist, Bayer Pharma AG, Leverkusen, Germany) and the animals were moved to MR scanner. The filled cannula was then connected to a syringe in a microinfusion pump. Head movements during the scanning were minimized by fixing the animals head in an MR-compatible stereotactic holder with ear bars, and the animals were put into MR with the head centered under the CryoProbe. The scanning protocol based on the previously described (Stanton et al., 2021) and whole mouse brain pre- and post-contrast T₁-weighted DCE-MRI was acquired with 1 min temporal and 100 µm isotropic spatial resolutions using 3D-FISP sequence (Table 1B). DCE-MRI continued over 90 measurements (90 min), and T₁-enhancing contrast agent was infused into the CM (gadobutrol, 1 µL/min for 10 min) after the first three baseline scans (i.e., after 3 min).

To verify whether ADC and D may reflect the volume of interstitial fluid (ISF) space in the brain parenchyma, we performed an additional DWI measurements by mimicking increasing ISF space volume in 3 water phantoms filled with Sephadex G-25 microbeads (Sephadex G-25; Sigma-Aldrich, St. Louis, MO, USA) of coarse (100–300 μm), medium (50–150 μm), fine particle size (20–80 μm of wet particle size). All microbeads possess the same porosity <5 kD so a similar exchange rate between stored and free water pools is expected. Each phantom was of the same in-house design (in. volume ~0.5 ml.), formed from of thick plexiglas tube (~25 mm in. long, out./in. ø10/5 mm) with a thread plug at both sides. The thread plug was made from the same tube, with a standard bonded polyester microfilter placed inside (Appendix 1—figure 1E) to prevent evacuation of the microbeads. For DWI, microbeads for each phantom were placed initially in a distilled water for 24 hr, to achieve their maximal size. Afterwards, a 2 ml syringe filled with distilled water solution of 0.001 mM/ml gadobutrol was attached to the thread plug on one side of the phantom and the phantom was filled with the microbeads. Gadobutrol was used to obtain optimal MR signal shortening for the echo time used in the same EPI sequence as employed in vivo (Table 1B). Any possible air bubbles remaining between the microbeads were carefully removed using a 1 µl inoculating loop. Subsequently, the other side of the phantom was closed with the thread plug, the phantom was flushed with ~1.8 ml gadobutrol solution in the attached syringe, and the solitary plug was sealed with a rubber syringe cap and parafilm. The syringe with a residual of gadobutrol solution at the other end was left to support pressure equalization inside the phantom.

DWI was performed 6 times in each phantom, with a central slice imaged in 1/3 phantom’s distal portion from the syringe. To verify the diffusion values in a free water environment, DWI was performed 4 times in the same phantom filled solely with gadobutrol solution.

To confirm that the free fluid volume surrounding the microbeads increases with increasing particles size, additional measurements were performed employing MR T₁ mapping and contrast-enhanced micro computed-tomography (µCT).

For T₁ mapping, all phantoms were prepared de novo and scanned jointly using spin-echo sequence (2D-RARE; Table 1B) with variable repetition times. For each phantom, dry microbeads were put for 24 hr directly into the phantom lumen filled with 0.001 mM/ml gadobutrol solution. Before MRI, any possible air bubbles were removed, and the phantom was sealed (as above). Microbeads of different size were expected to differently infiltrate surrounding water, and partially gadobutrol. Thus, the relative concentration of gadobutrol was expected to be altered compared to original solution, and shorter T₁ relaxations were expected to be observed for phantoms with decreasing microbeads size. For reference, T₁ mapping was performed in the phantom filled solely with gadobutrol solution. To correct T₁ maps for B₁-filed inhomogeneities using double-angle method (Insko and Bolinger, 1993; Cunningham et al., 2006), all phantoms were scanned with the same spin-echo sequence with a maximal repetition time used for T₁ mapping, and using two different flip angles (Table 1B).

For free fluid space estimation using µCT, all phantoms previously scanned for DWI were unsealed and flushed with a distilled water solution via 5 ml syringe to remove gadobutrol. Subsequently, distilled water was replaced with a 1:1 dilute of non-ionic iodine Omnipaque 350 contrast agent (Iohexol, 350 mg iodine/ml; GE Healthcare AS) in normal saline via 3 ml syringe. Each phantom was scanned using Vector4uCT system (MILabs, Utrecht, Netherlands) using the scan parameters of 15 μm isomteric resolution, 50 kVp, 0.24 mA (75ms exposure), 360 degrees rotation, 0.2 degree rotation step, 2 frames for averaging, 0.5 mm thick beam aluminum filter, Hann filter with cone-beam reconstruction.

To verify the relation between MR diffusion values and the T₁ relaxation times as well as free fluid estimates using µCT, DWI was performed 4 times in each phantom (as above).

C57Bl/6 brain sections collected previously (Hablitz et al., 2020, ) were newly reimaged and an entirely new analysis done for current experiment. Originally, mice (4 WT; Table 1A) were cardiac perfused with AlexaFluor 488 conjugated wheat germ agglutinin (Thermofisher Scientific) at 15 µg/mL in 20 mL of phosphate buffered saline (PBS), and then perfused with 20 mL PBS with 4% paraformaldehyde (PFA). Subsequently, the brains were extracted and were immersed in 4% PFA overnight. This step labels the vasculature and fixes the brain tissue. Brains were sectioned coronally at 100 µm using a vibratome (Leica VT1200s) and equivalent sections from +1.2 to –1.8 bregma were stained for AQP4. Floating sections were permeabilized with 0.1% Triton-X-100 in PBS, blocked with 7% normal donkey serum (Jackson Immunoresearch) in PBS with 0.03% Triton-X-100, and then incubated with Anti-AQP4 primary antibody (AB3594, Millipore; 1:1000 dilution) overnight. The sections were afterwards washed three times with PBS and incubated with Alexa 594 fluorophore-linked donkey anti-rabbit secondary antibody (A21207, Invitrogen; 1:500 dilution) and DAPI (D1306, Invitrogen; 1:2000 dilution). Stained sections were mounted with Fluoromount G (Thermofisher Scientific).

Vascular density was measured in mice brains (6 AQP4 KO and 6 C57Bl6; Table 1A) prepared in few steps. All mice were cardiac perfused with AlexaFluor 488 conjugated wheat germ agglutinin at 15 ug/mL in 20 mL of PBS, and then perfused with 20 mL 4% PFA. Afterwards, the brains were extracted and immersed in 4% PFA overnight. 100 µm coronal sections were taken using a vibratome (Leica 1200 S). Anterior brain sections at +1.2 mm bregma, and posterior brain sections –1.8 mm bregma were mounted in Prolong Gold media with DAPI (Invitrogen). Equivalent sections were used for all biological replicates.

The measurements were performed in mice (5 KO and 6 WT; Table 1A) as described previously (Liu et al., 2020). Mice were anesthetized with K/X (100 mg/kg / 20 mg/mL i.p.) and 2% isoflurane, and placed in a stereotactic frame. Their scalp was shaved, and the skin was cleaned with a chlorhexidine swab followed by an alcohol wipe to remove the chlorhexidine. An iodine solution was applied and left to dry, and the scalp was opened and the skin retracted. The exposed skull was irrigated with sterile saline and cleaned by applying sterile cotton swabs, and a sterilized stainless-steel light-weight head plate (0.9×19×12 mm dimension), equipped with a round hole of at the center (in. ø9.0 mm), was attached to the mouse skull using a mixture of dental cement with cyano-acrylate glue (Sweeney et al., 2019). Pre-operatively as well as for three days post-surgery the mice received Banamine (1.1 mg/kg) subcutaneously as an analgesic. The mice were trained to tolerate positioning in the head plate stand (Cat# MAG-1, Narishige International USA Inc), as well as a restraint tube in three daily training sessions, each lasting 30 min for 3 days post-surgery.

For measuring CSF production during wakefulness, the mice were anesthetized with 2% isoflurane during cannula implantation. A cannula (30 G needle) attached to artificial CSF filled PE-10 tubing was implanted into the right lateral ventricle through a small burr hole (AP = −0.1 mm, ML = 0.85 mm, DV = –2.00 mm from the postion of bregma). The cannula was fixed to the skull with dental cement and the opposite end of the PE tubing was sealed by high-temperature cautery. Once the cannula was in place, the neck was flexed 90 degrees and the headplate was attached to the head stand. Then a separate cannula was inserted into the CM and advanced into the 4th ventricle. One microliter of mineral oil was infused into the 4th ventricle to block the exit of CSF into the subarachnoid space. All incisions were infiltrated with 0.25% bupivacaine topical anesthetic to prevent the animal from experiencing post-surgical pain. The measurements were collected while the animal rested in an open cylinder restraint tube (9 cm length, in. ø3.5 cm), to which the animals were accustomed during training. Anesthesia was discontinued and CSF production was measured in head-fixed mice in 10 min intervals for 65 min, after a 30-min recovery period.

Real-time iontophoresis with tetramethylammonium (TMA) was performed in mice (8 KO and 20 WT, body weight not recorded) as adapted from the previous studies (Nicholson, 1993; Xie et al., 2013). The single barrel iontophoresis microelectrode (tip out. ø2–3 µm) contained 150 mM tetramethylammonium (TMA)-chloride and 10 µM Alexa 488. A series of currents of 20 nA, 40 nA and 80 nA were applied by a dual channel microelectrode pre-amplifier. For measurements of TMA, microelectrodes (out. ø2–3 µm) were fabricated from double-barreled theta-glass using a tetraphenylborate-based ion exchanger. The TMA barrel was backfilled with 150 mM TMA chloride, while the reference barrel contained 150 mM NaCl and 10 µM Alexa 568. All recordings were obtained by inserting the two electrodes to a depth of 150 µm below the cortical surface. Recording electrodes were inserted 2.5 mm lateral and 2 mm posterior to bregma. The electrode tips were imaged after insertion using 2-photon excitation to determine the exact distance between the electrodes (typically ~150 µm). The TMA signal was calculated by subtracting the voltage measured by the reference barrel from the voltage measured by the ion-detecting barrel using a dual-channel microelectrode pre-amplifier. The Nikolsky equation was used for calibration of the TMA electrodes based on measurements obtained in electrodes containing 0.5, 1, 2, 4, and 8 mM TMA-chloride in 150 mM NaCl. The TMA measurements were acquired relative to similar recordings obtained in 0.3% agarose prepared from a solution containing 0.5 mM TMA and 150 mM NaCl. A custom-made software in Matlab (v. R2019a, The Mathworks, Inc, Natick, MA.), ‘Walter’, developed by C. Nicholson was used to calculate α and λ values (Nicholson, 1993; Xie et al., 2013).

Anesthetized animals (3 KO and 5 WT, gender and body weight not recorded; Table 1A) were decapitated, the whole brains were taken out and weighed immediately (W_wet). Brain tissue was dried at 65 °C for 48 h until it reached a constant weight, and brain were re-weighed (W_dry). The ratio between the difference of W_wet - W_dry and W_dry was considered reflecting the brain water content (ml H₂O/g dry weight).

All acquired images were visually checked and no presence of significant artefacts influencing morphological and functional assessment was found. Further processing pipelines were applied, and included motion-correction, spatial co-registration, and automatic or semi-automatic pre- and final postprocessing.

All described statistical analyses were performed in GraphPad Prism 8 (GraphPad Software) and Matlab. The results coming from statistical comparisons were considered significant for p<0.05 after post-hoc correction, when applicable.

All 3D-CISS volumes were calculated in few steps using in-house pre-processing pipeline (Appendix 1—figure 3A). For each animal, every 3D-TrueFISP volume acquired with two repetitions was motion-corrected (10 times or until no further improvement) and averaged. Subsequently, the second averaged 3D-TrueFISP volume (180° encoding direction) was subjected to rigid-body registration (6 df.) to the first volume (0° encoding direction). Both motion-correction and registration were performed in AFNI (Oakes et al., 2005), and aimed to reduce the influence of random motion on the computed 3D-CISS image. Finally, 3D-CISS image was computed as a maximum intensity projection from 2 co-registered 3D-TrueFISP volumes, resulting in an image of almost completely removed banding artifacts. Afterwards, every computed 3D-CISS volume underwent semi-automatic brain parenchyma image extraction using the ‘Segment 3D’ tool in ITK-SNAP (Yushkevich et al., 2006), to remove the regions outside the brain parenchyma image from further analysis. Brain parenchyma was considered as the brain tissue volume surrounded with dark regions of skull image and including intracerebral vessels. To correct for intensity inhomogeneities coming from the B₀ field and the surface profile of the CryoProbe (B₁), the extracted brain parenchyma image underwent bias field correction using FSL (Zhang et al., 2001) (0.5 sigma, 20 mm FWHM, 4 iterations). At each step of pre-processing, the results were visually checked and confirmed for correctness.

For a single, bias-corrected 3D-CISS volume the ventricular and perivascular CSF spaces were separated from the brain parenchyma image in 3 dimensions using an in-house fully automatic adaptive algorithm in Matlab (Gomolka et al., 2021), in four steps (Appendix 1—figure 3B). First, high-intensity regions, as branches of the optic nerve’s residual after the semi-automatic brain extraction and not adjacent to the parenchyma in all image slices, were excluded using a bounding box enclosing the brain (Appendix 1—figure 3B ‘Bounding box and adaptive thresholding’). The bounding box was automatically computed based on a maximization of the voxels intensity variance slice-wise, separately in three orthogonal planes. The parenchyma volume surrounded by the bounding box was enclosed and the solitary regions were removed based on their geometrical properties calculated slice-wise in the sagittal plane: eccentricity≥0.5, roudness≥0.5, perimeter<0.005% of the brain parenchyma voxels count. The resulting brain image mask was geometrically dilated with a disk kernel of 11 pixels diameter, to enclose potentially removed or non-continuous parenchymal regions. The non-continuity appeared in case of residuals from banding artifacts at the borders of the skull and the ethmoidal bone. Subsequently, brain parenchyma volume was updated according to the resulting mask, for further automatic segmentation of the CSF space.

Second, initial CSF space separation was performed by means of an adaptive intensity threshold and calculation of cumulative distribution of voxels intensities>0 from the separated and bias-corrected brain volume. As the overall distribution of the brain intensities differed slightly between the 3D-CISS images due to their SNR, a threshold-correcting factor (denoted f_c) was calculated by means of the formula:

where µ_d is a mean and σ_d is a standard deviation (SD) of aggregated brain parenchyma voxels intensity distribution. Separation of the CSF space was performed assuming that CSF intensities reflect those ≥95^th percentile of the aggregated intensities distribution. Hence, the intensity of each voxel was rescaled according to the formula:

where I and I_r are the original and the rescaled voxels intensities. The rescaled voxels intensities were encoded using a floating point precision in range between −1.33 × σ_d and a distribution peak close to the image SNR calculated as µ_d/σ_d (i.e. maximum rescaled intensity ~10 with the mean value varying between 3 and 5 among all analyzed images). Third, all the rescaled voxels possessing negative intensity (i.e. brain parenchyma) were assigned to 0, and a new aggregated distribution of the rescaled voxels of >0 intensity was computed. Subsequently, all voxels intensities ≤95.5^th percentile of the new distribution were assigned to 0 to keep only the high intensity CSF seed regions. For the images of SNR >4 (i.e. lower contribution of Rician noise), the new distribution threshold was set of ≤95.5 +f_c. The correction factor f_c accounted for subtle intensity changes and did not result in the threshold exceeding the 97^th percentile. A mask image of the initially separated CSF space was computed by assigning all remaining nonzero voxels to 1, for further application of a region-growing algorithm.

The final segmentation was performed using a 2D region-growing algorithm applied slice-wise, consecutively in sagittal, axial and coronal planes (Appendix 1—figure 3B- ‘in Sagittal plane’, ‘in Axial plane’, ‘in Coronal plane’). In each slice, the algorithm reconsidered the voxels at the boundary of the CSF space in horizontal, vertical and diagonal directions separately. The algorithm based on extension of the method for contrast calculation applied to study properties of hemorrhagic and ischemic regions in clinical CT images (Nowinski et al., 2014; Gomolka et al., 2017). The calculations were performed considering the initially separated CSF space mask and the original 3D-CISS image, in two steps: (1) CSF mask dilation; (2) CSF mask erosion.

For every animal, all DWI volumes acquired were subjected to motion-correction in AFNI to avoid influence of random and subtle frame-to-frame image displacements. The motion-correction was performed (4 times or until no further improvement) in reference to the first image acquired using the first b-value (b0), and the results were visually confirmed for correctness by RSG. Images acquired with different number of repetitions were averaged according to b-value and diffusion encoding direction, for subsequent calculation of diffusion measures in all 3 directions. Separately, an averaging took place considering only b-values, to calculate an average diffusion images. To normalize the image intensities and reduce influence of nonstationary noise, volume-wise normalization of voxels intensities was performed in all images from different b-values using a mean value of a background signal defined in a half-ball VOI including 704 voxels outside the visible tissue image. The VOI was manually set in an averaged b0 image and applied to every averaged volume from each b-value for the background signal depiction. The intensity normalization was performed according to formula:

where DWI_bval and DWI_{bval norm} are the original and normalized voxels intensities, and µ_{bval noise} and σ_{bval noise} are respectively the mean and SD of the background signal at b-value. To reduce the influence of distortion artefacts, all DWI images were subjected to slice-wise spatial smoothing using a 0.5 sigma [3×3] Gaussian filter. To automatically exclude the voxels outside the brain regions in subsequent diffusion curve fitting, a brain mask image was calculated from b0 image and applied to all images from different b-values. The mask image excluded all the voxels below the mean + 0.5 × SD of the b0 image intensity. As initial threshold excluded voxels ventrally from the brain image, the mask was subjected to morphological dilation with a diamond shape kernel of 2 pixels diameter, and subsequent image filling to include sole obsolete voxels inside the mask.

All calculations were performed in Matlab using in-house computational pipelines and curve-fitting implementations based on the least squares method. For a standard DWI model, monoexponential curve-fitting was performed voxel-wise using all 17 b-values. The ADC and estimated S₀ images were calculated voxel-wise based on the formula:

For IVIM model, a two-step algorithm was used as considered providing more robust and reliable results compared to standard biexponential curve-fitting (Lee et al., 2016). In the first step, a perfusion fraction (F_p) was estimated as a voxel-wise ratio of the difference between b0 and approximated S₀ to b0 image. Voxels intensities in the S₀ image were estimated from the linear regression approximation considering the logarithm intensities of the volumes from the IVIM threshold to the highest b-value. The threshold b-value used was closest to 250 s²/mm (here ~238 s²/mm). The slow diffusion D was estimated as the slope coefficient of the linear regression function. Subsequently, for the voxels where a positive and nonzero F_p value was calculated (presence of fast diffusion) a biexponential curve-fitting was performed using the calculated F_p and D:

where D is the slow (‘pure’) molecular water diffusion, D* is the fast diffusion (pseudodiffusion).

For further statistical comparison, the mean and SD of ADC, F_p, D and D* signal intensities were calculated within 21 regions of interest (ROI), for every animal brain image separately. The regions were chosen based on reported high Aqp4 expression (Hsu et al., 2011; Hubbard et al., 2015) and were drawn manually using ITK-SNAP in estimated S₀ image, and overlaid on the calculated diffusion parameters maps. The defined ROIs included from 48 to 250 voxels depending on the anatomical structure and excluding neighboring distortion artefact, and covered the position of cortical (olfactory area - OLF; cingulate/retrosplenial cortex - CA/RSP; visual - VIS; somatosensory - SS; auditory - AUD; hippocampus - HIP; perirhinal - PERI), brain stem (thalamus - TH; habenula - HAB; hypothalamus - HY; midbrain - MB; periaqueductal gray - PAG; hindbrain - HB), cerebral nuclei and tracts (caudate putamen - CP; white matter including striatal regions - WM); CSF space (lateral ventricles - LV; third ventricle - 3 V; fourth ventricle - 4 V; pericisternal space - PCS; perivascular space within the Circle of Willis - CoW) and the cerebellar ROI (cerebellum - CB). Application of automatic segmentation template was avoided due to variable influence of distortion and ghosting artefacts at large span of 17 b-values measured, resulting in erroneous template registration and delineation of signal from specific anatomical structures. The regions were set manually by RSG in reference to the Allen Brain Atlas, and verified by YM and MN.

Due to presence of distortion and ghosting artefacts resulting in possible inhomogeneous noise properties of EPI images acquired, voxel intensities within delineated ROIs were considered coming from independent and non-uniform signal distributions (i.e., comparing ventral vs. dorsal ROI). Therefore, mean ADC, D, F_p, D* and F_p × D* were calculated excluding <0 and>99 percentiles of their intensity distribution within each ROI (see calculated values in the supplementary Excel file), and were compared ROI-wise between KO and WT animals using a nonparametric Mann-Whitney U-test. To confirm reliability of the slow diffusion measures, Pearson’s linear correlation was calculated between the mean ADC and D from all animals.

Time-series of FISP volumes acquired were motion-corrected and spatially normalized with ANTs (Avants et al., 2014) and co-registered to the respective baseline image, subject-wise. For every animal, the co-registration process was repeated twice using rigid-body (6 df.) and twice using an affine transformation (12 df.) to assure accuracy. Subsequently, the intensities from the first volume acquired after the dummy scan (i.e., of unstable steady-state MR signal) were subtracted voxel-wise from succeeding volumes to reveal only the regions enhanced with the contrast agent. To normalize the CSF signal in each time-series their voxel intensities were subjected to Gaussian normalization using the first 3D-FISP volume. The resulting images were smoothed with a 3×3×3 voxels kernel of [0.2, 1, 0.2] weights along each axis, to reduce the influence of possible artifacts after the automatic registration and subtracting the baseline volume.

For further statistical comparison, DCE 3D-FISP signal was derived from 21 ROI reflecting those from DWI analysis (see above), in each animal individually. All ROI were defined in a 3D template volume, based on Reference Space Model from Allen Brain Atlas, of the same spatial resolution as 3D-FISP volumes acquired. The reference volume was automatically registered in AFNI using affine registration (6 df.) to the baseline DCE volume. Mean DCE signal was depicted from every 3D-FISP volume in every animal ROI-wise (90 signal intensity values / ROI), for further statistical comparison. The first volume following the dummy scan was rejected from further consideration as it had previously been used for the signal normalization. To compare ROI-wise DCE signals between KO and WT animals, a repeated measurements Two-way ANOVA along with post-hoc Bonferroni’s correction was used in search for differences between the signal changes at specific time points and compared to the baseline signal before the gadolinium infusion. Further calculation of DCE-derived scores of CSF tracer arrival time, relative time-to-peak, peak intensity, and duration of significant from baseline interstitial tracer accumulation was performed based on the rank-sum scores provided by Dunn’s multiple comparison from nonparametric Friedman’s one-way ANOVA, performed separately in KO and WT groups. Arrival time was considered as a difference in time between the baseline time point, and time point where the negative rank-sum score was at least three times that from the baseline. Time-to-peak was calculated as a time difference between the start of the gadobutrol infusion and the subsequent time point at which the maximum negative rank-sum score was calculated (i.e. corresponding to the peak CSF signal intensity within the group). The peak intensity was considered as a relative signal increase occurring at the time-to-peak. Duration of significant from baseline interstitial tracer accumulation was obtained as a time difference between the last and the first time points, where DCE signal increase was significantly different from baseline. Area under the DCE curve was calculated for each subject as a sum of DCE signal amplitudes over all the measured time points. DCE scores derived for KO and WT groups were compared ROI-wise using nonparametric two-tailed Wilcoxon signed-rank test.

Calculation of T₁ maps was performed using an in-house algorithm in Matlab. The voxel intensities from spin-echo images of the phantoms were approximated over the increasing repetition times, by inversely solving the equation with the least squares method:

where TR is a repetition time, S₀ is the estimate of the proton density, S_TR is the intensity of the considered voxel from the corresponding slice, and ε is an estimator of a random error. To reduce influence of transmit inhomogeneity on T₁ estimations, B₁-correction was applied using dual-angle method based on relationship between an effective and nominal flip angles (Cunningham et al., 2006) and by applying voxel-wise a correction factor dependent on the intensities from two spin-echo volumes acquired with 45° and 90° flip angles.

Evaluation of the free water volume fraction inside the phantoms with microbeads was performed based on percentage of the contrast-filled volume, quantified in 8 bits gray scale axial images. First, the central portion of the phantom volume, reflecting that from DWI analysis, was manually separated in ImageJ (v. 1.53 j, NIH, USA). The free fluid volume regions (filled with Omnipaque 350 solution) were of clearly higher Hounsfield Units (HU) intensities compared to those from the microbeads. Thus, the voxels belonging to the free fluid were empirically verified occupying >75^th percentile of the cumulative HU distribution in all the separated volume images, from all phantoms. Therefore, the voxels belonging to the free fluid were counted three times, considering only voxels above 75^th, 80^th, and 85^th percentile of HU distribution. For each phantom, the free fluid volume fraction was estimated as a mean voxel count from all thresholds, divided by the voxels count in the separated volume image. As objective separation of the fluid space was impossible in the phantom with the fine microbeads and lower HU thresholds would consider voxels affected by partial volume from the microbeads, proposed free fluid volume fraction estimation was found sufficient to correct for the changes in shape of voxels HU distribution from different phantoms.

For each brain hemisphere in each section, all ROIs were manually drawn according to the Allen Brain Atlas and using auto fluorescence from the green channel following visible anatomical landmarks and confirming AQP4 expression visibility. Each subregion measured from 0.03 to 1.69 mm². Therefore, a universal threshold was applied manually to all images as a preferred method for reduction of the background signal influence. Subsequently, a mean area fraction covered by AQP4 in both hemispheres/sections was measured for each ROI. In total, mean AQP4 expression was calculated for 11 ROIs (retrosplenial cortex - RSP, visual - VIS; somatosensory - SS; auditory - AUD; hippocampus - HIP; perirhinal - PERI; thalamus - TH; habenula - HAB; hypothalamus - HY; pericisternal - PCS; white matter - WM) in 4 WT mice. A nonparametric Kruskal-Wallis one-way ANOVA with Dunn’s post-hoc was employed to compare the mean AQP4 channel expressions between ROIs.

ROIs were manually outlined around each anatomical subregion according to the Allen Brain Atlas, and using visible anatomical landmarks. Due to differences in fluorescent labeling intensity, each region was thresholded individually to isolate labeled blood vessels from the background, so each region measured from 0.023 to 4.67 mm² within each hemisphere/section. Area fraction of blood vessels above the threshold was measured for each ROI. In total, mean vascular density was calculated from multiple subregions for 17 ROI (olfactory area - OLF; cingulate cortex - CA; retrosplenial cortex - RSP; primary visual - V1; primary somatosensory - S1; primary motor area - M1; auditory - AUD; hippocampus - HIP; perirhinal - PERI; insular - INS; thalamus - TH; habenula - HAB; hypothalamus - HY; caudate putamen - CP; white matter - WM; pericisternal - PCS; ependymal around lateral ventricles - EPD) in 6 KO and 6 WT animals. Further statistical comparison was performed assuming inhomogeneous signal distribution properties between different ROI (similarly as for DWI). Hence, considering independent measurements of vascular densities among ROI analyzed and due to small group size, nonparametric Mann-Whitney U-test was employed to compare the vessel densities from KO and WT animals ROI-wise.

For ROI-wise correlation analysis, a mean value of AQP4 expressions as well as vascular densities at ROI was calculated from all respective animals strain-wise.