This study is in compliance with the Declaration of Helsinki. Consent for autopsy was obtained from the patient’s next of kin through standardized consenting procedures. No IRB approval was required given that the autopsy material used herein is considered to be non-human subjects. Pre-COVID autopsy material was obtained from donors who died between January 2018–2019, or during 2020–2021 and were negative for COVID-19. COVID-19 cases are previously thoroughly described (al-Dalahmah et al., 2020a; Thakur et al., 2021). Only the medulla oblongata tissue was analyzed in this study. The demographic information and relevant information regarding hospital course, including whether patients had histologic evidence of Diffuse Alveolar Damage (DAD) in the case when a full autopsy had been conducted, are provided in Supplementary Table S1. For some of our control cases, the clinical history is not available. This is because these were brain-only autopsies of patients who died elsewhere (not in our hospital). For these cases, we ensured that no hypoxic changes were neuropathologically detected (i.e., red neurons, nuclear pyknosis, and neuronal shrinkage) so as to use them as non-hypoxic controls. Autopsy brains, fixed in 10% formalin for 10–14 days after removal, were sectioned coronally and samples from representative areas of the CNS were removed and embedded in paraffin blocks, cut at 7 μm thickness, and mounted on charged glass slides. All studies reported herein are from the medulla at the level of the inferior olivary nuclei and hypoglossal nuclei.

All immunostains were conducted on a Leica^© Bond RXm automated stainer. For chromogenic 3,3′-Diaminobenzidine (DAB) stains, a generic immunohistochemistry protocol was employed as per manufacturer protocols. For multiplexing immunostains using antibodies raised in non-overlapping hosts, we used a generic immunofluorescence protocol. Briefly, slides were baked in a 65°C oven for a minimum of 2 h. The following protocol was then used: After a dewaxing step, incubation in BOND Epitope Retrieval Solution 2 (cat# AR9640) for 20 min was used for heat-induced epitope retrieval. Next, the slides were washed in 1X PBS before washing twice in Bond Wash Solution (Ref#AR9590)—10 min/wash. Next, they were incubated in a 10% donkey serum blocking buffer for 60 min followed by the primary antibody diluted in blocking buffer for 60 min. After three washes, the slides were incubated in the secondary antibody containing buffer for 60 min. After three washes, A DAPI containing mounting solution (Everbright TrueBlack Hardset Mounting Medium with DAPI Cat#23018) was used to label nuclei and quench autofluorescence prior to coverslipping. One hundred fifty microliters/slide was the volume we used for all steps. All steps were conducted at ambient temperature—excluding the antigen retrieval step.

For multiplexing immunostains using primary antibodies raised in overlapping hosts (ALDH1L1, MT3 and AQP4 and ALDH1L1, YKL-40 and C3), the Opal 4-color Automation IHC kit Ref#220126024 from Akoya^© Biosciences was used in accordance with the manufacturer protocol. Briefly, two wash steps were followed by incubation in PKI Blocking buffer for 5 min before incubation in the first primary antibody for 30 min. After 3 wash steps, the slides were incubated in Opal Polymer HRP for 10 min followed by 6 wash steps prior to incubation in Opal 520 reagent for 10 min. This was followed by 4 additional wash steps. Next, the slides were incubated in Bond ER 1 solution for 20 min at 95° to elute the antibody complexes before 3 more wash steps. This procedure was repeated twice, once with the second primary antibody and Opal 570 reagent and once with the third primary antibody and Opal 690 reagent. Following the 3 wash steps at the end of the third round, the slides were incubated with Spectral DAPI for 5 min before the final 3 wash steps.

The following primary antibodies and dilutions were used: Rabbit ALDH1L1 (1:100, EnCor, Cat#RPCA-ALDH1L1), Rabbit YKL-40 (1:250, Abcam, Cat#ab255297), Rabbit C3 (1:200, Abcam, Cat#ab200999), Chicken GFAP (1:1000, Abcam, Cat#4674), Goat Clusterin (1:200, Thermo fisher, PA5-46931), Rabbit CD44 (1:100, Abcam, Cat#ab101531), Rabbit MT3 (1:100, millipore, Cat#HPA004011), Rabbit, AQP4 (1:2000, Millipore, Cat# ABN910), Goat IBA1 (1,500, Abcam, Cat#ab5076), Rabbit Trem2 (1,100, Cell Signaling, Cat#91068). Secondary antibodies conjugated to fluorophores: anti-mouse Alexa Fluor 488, 568, and 633, anti-rabbit Alexa Fluor 488, 594, anti-chicken Alexa Fluor 488 and 647, and anti-goat Alexa Fluor 488, 568, 633; all from goat or donkey (1:500, ThermoFisher Scientific, Eugene, OR).

All brightfield images were taken using a Leica Aperio LSM^™ slide scanner under 20X objective. All immunofluorescent images were taken on the Leica Thunder imager DMi8. Images were acquired at 20x using a Leica K5 camera. Leica biosystems LAS X software was used for image capture. Tiles covering the entire ION were taken and stitched. Leica Thunder instant computational clearing was used to remove out of focus light. The images were exported as tiff files for downstream analysis.

All image analysis was done in QuPath 0.30 (Bankhead et al., 2017). Annotations delineating the ventral, lateral and dorsal ION Parenchyma as well as the hilum were manually drawn. To detect cells, we used the “cell detection” function under the analysis menu. The DAPI Channel was selected for the Detection Channel. We modified the background threshold for each image to eliminate non-specific detections. Next, we trained an object classifier to classify the detections for the different channels. Training data were created from each image to delineate cells that are positive for the specific antigens in question. One classifier per channel was trained by calling the “train object classifier” function under classify with the following parameters: type = Random Trees, measurements = Cell: measurements = Cell: Channel X standard deviation, mean, max, and min measurements for the channel in question. To increase the accuracy of the classifier, additional training annotations were created on the image in question until the classification results matched the impression of the observer. Once a classifier was trained for each channel, “create composite classifier” was called to create a classifier consisting of multiple individual classifiers, one for each channel on the image. Classifiers were trained for each image separately. For CD44 and AQP4 analysis, we created a pixel classifier to classify positive and negative pixels. Training annotations were created for each image for positive and negative pixels. “Train pixel classifier” function was then called with the classifier type set to random trees, with a resolution of 2.60 μm/pixel, and selected all the features from only the channel in question.

To measure the minimal distance between microglial and olivary neurons, we first detected microglia using positive cell detection to identify IBA1+ cells. Next, QuPath pixel classifier was used to classify IBA1− ION neurons, which have characteristic large cytoplasm and eccentric nuclei. Using more than 20 manually annotated neurons as the training set, the pixel classifier accurately detected all neurons. We next converted the pixel classifications into annotations which we used to measure the distance against by calling analyze > spatial analysis > distance to annotations 2D measurements function between microglia (as positive cells) and neurons (as annotations). The measurements were exported as .csv files for downstream analysis in R. After −1*log10(1 + value) normalization, and binning into 100 bins, the kernel density distribution of the counts of cells that fall within each bin was used to calculate the modes for each condition using the multimode package in R by calling the locmodes function with the following options (lowsup = 0.00001, uppsup = 6, mod0 = 2, display = T). The supports were chosen to fit the data empirically—the upper support was ≤ to the maximum value in the data. We assumed two modes for the distribution (mod0 = 2). The Gaussian kernel density estimator is employed in the package. The distributions were compared using the ks test (two-sided) in R.

To classify microglia by activation state, an object classifier was used on objects detected by setting the detection channel to the IBA1 channel. This allowed the full tracing of microglial processes. The training images for microglial reactivity were compiled from examples taken from all images included in the analysis. Training objects were assigned by setting the class of microglia as quiescent vs. activated cells. The key characteristics used to identify a quiescent microglial cell were lightly-stained processes and small somata, while the activated microglia were marked by darker stains, larger soma, and thickened and retracted processes.

PCA analysis was done in FactoMineR R package (Lê et al., 2008). A total of 28 donor brains (10 ARDS and 18 COVID) were analyzed in four brain regions (dorsal, lateral, and ventral ION parenchyma (OP), and hilum), for a total of 107 data points representing the results of the image quantification after outlier removal. Metadata was included in the analysis as supplementary variables. Numerical values (age and length of hospitalization) were categorized into three bins. Other qualitative data included presence or absence of diffuse alveolar damage (DAD), intubation, and sepsis, as well as sex, condition, and brain region. −1*Log10 (value+1) normalization was performed on all immunohistochemical data measured as number of positive cells per area; no normalization was performed for data measured by percentage of area covered (AQP4 and CD44). Outliers, denoted in Supplementary Table S1, were identified in both using the Grubb’s method (see Section Statistical Analysis section below) and in cluster analysis. Outliers formed small 1–2 sample clusters. The few missing values, such as those resulting from low quality images, were imputed using the imputePCA function of the MissMDA package in R. Principal component analysis (PCA) was performed on the IHC data alongside supplementary qualitative variables, comprised of the metadata variables (Supplementary Figure S5A). The dimdesc function, part of the FactoMineR package, provided further details of factor analysis of samples (Supplementary Table S2). These results were then used for hierarchical clustering analysis, using the FactoMineR package’s HCPC function with the distance metric set to ‘Manhattan’, to provide four hierarchical clusters of the data (Figure 5B). Proportions of qualitative variables comprising each hierarchical cluster were then calculated using Dplyr functions (Supplementary Figure S5C). Heatmaps were generated using the pheatmap R package.

All statistical analyses were conducted in GraphPad^® Prism 9 or R v4.03. For all data sets, outliers were identified using the Grubbs’ method with an Alpha = 0.2. All statistical tests and graphs were done using the outlier-free data. For analyzing two groups we used two-tailed and unpaired t-tests. For analyzing more than two groups, we used one way Brown-Forsythe and Welch ANOVA correcting for multiple comparisons using Original FDR method of Benjiamini and Hochberg. All data sets that were analyzed using one-way ANOVA were tested for normality using Shapiro–Wilk test and transformed using Y = −1*Log(Y + 1) if they did not pass the normality test. p values reported are those of the transformed data where transformation was done. To further validate our ANOVA test results, a beta regression model was also used as implemented in the betareg package in R. The independent variables used were condition (either ARDS, COVID with microglial nodules (MN), and COVID with no MN, or Control, ARDS, and COVID) and region (dorsal, lateral, and ventral OP), with the counts (microglia per area, proportion of activated microglia, GFAP per area, or percent MT3) as the dependent variable. The results of this analysis are provided in Supplementary Table S4.

Microglial reactivity is a common feature of COVID-19 pathology. We first set out to replicate this finding in our cohort, focusing on the ION. We included control patients who died without COVID-19 or ARDS, patients who died with ARDS but not COVID-19, and patients who died of COVID-19. This allows us to answer the following question: is microglial reactivity in COVID-19 due to hypoxia? Thus, we quantified the number of IBA1+ cells per unit area in different regions of the ION: the lateral, dorsal, and ventral sectors (Figure 1A). ANOVA analysis of IBA1+ cells/area was significant in all three regions of the ION, and there was a significant increase in the number of microglia per unit area in COVID-19 cases compared to the non-hypoxic controls in the lateral ION, but there was no difference between the non-hypoxic controls and the ARDS cohort (Figure 1B). This indicates that factors in COVID-19, in addition to hypoxia, were necessary to drive the increase in microglia in the ION. Beta regression analysis for microglia per area returned, for the COVID condition, a coefficient of 0.472 and p value 2.89E-05 (Supplementary Table S4), suggesting that COVID-19 condition can explain the increased microglial numbers in the COVID-19 cases in our cohort. Because ARDS and controls were not significantly different in the density of ION microglia, and ARDS cases can be considered matching controls for hypoxia, we compared microglial reactivity between the ARDS and COVID-19. We used morphologic attributes of microglia to train a machine learning algorithm to classify microglia into quiescent versus activated (see Section Materials and methods). We wanted a simple way to classify microglia based on morphology, knowing that microglial reactivity falls on a spectrum of states, and that activated microglia generally have retracted thick processes compared to quiescent cells (Davis et al., 2017; Leyh et al., 2021). We opted for a simple binary classification of quiescent vs. activated microglia; we show examples of these classifications in Figure 1C. We also chose to split our COVID-19 group into two groups: a group with high abundance of microglial nodules (MN), and another group with relatively few/no microglial nodules (No-MN). These designations were based on previously reported neuropathologic assessments (Thakur et al., 2021). Comparing the proportion of activated microglia across the ventral, dorsal, and lateral ION in these three groups (ARDS, COVID-19 no-MN, COVID-19 MN) revealed that in the ventral ION, the COVID-19 MN group had a significantly larger proportion of activated microglia (Figure 1D). Subsequent beta regression testing of proportion of activated microglia returned a coefficient of 0.509 with p value 0.000553 for the COVID-MN condition, but the coefficients were not significant for ION regions, suggesting that condition rather than ION region drives microglial activation (Supplementary Table S4). Next, we asked if the minimum distance between any microglial cell and the closest neuron to it is different between the groups. This in effect is a way to quantify the proximity of microglia to neurons. We reasoned that we would see more microglia close to neurons if there is more neuronophagia or microglial nodules. We measured the distance between microglia and neurons in the ION (Figure 1E) and found that the distribution of minimal distance between microglia and neurons is quite different between the groups (Figure 1F—Asymptotic two-sample Kolmogorov–Smirnov test—a non-parametric test to compare distributions). The results are as follows: D values = 0.28, 0.21, 0.12, for MN vs. ARDS, No-MN vs. ARDS, and MN vs. No-MN, respectively, for all comparisons the p-value is less than 2.2e-16. Comparing the distribution across different anatomic sectors of the ION revealed similar results (data not shown). The distribution was truly bimodal in the COVID-19 MN group, and the two modes were 1.36 (higher probability mode) and 3.45 (lower probability mode). Conversely, the modes for the ARDS group were 3.47 (highest probability mode) and 0.11 (lower probability mode) and for the No-MN group were 1.48 (highest probability mode) and 2.45 (lower probability mode). The fact that the mode with the highest probability (density) in the MN was lower than that in the ARDS group can be seen as an indirect measure of the presence of microglial nodules in the MN group—which is previously established. Altogether, we found that microglia are more activated and closer to neurons in the ION in COVID-19, especially the MN group.

We also asked if microglia in COVID-19 brains expressed more TREM2 compared to microglia in ARDS in the ION. TREM2 labels phagocytic microglia (Takahashi et al., 2005). Although we could detect TREM2 in microglia in the white matter surrounding the ION (for example—in the pyramids Supplementary Figure S1A), there was no significant specific labeling of microglia in the ION in COVID-19 or in ARDS (Supplementary Figure S1B).

Given that astrogliosis is a prominent feature of COVID-19 neuropathology, we asked if we could recapitulate this finding in the ION. To address this question, we conducted a series of immunohistochemical and multiplex immunofluorescence studies to quantify the expression of proteins related to reactive astrogliosis or alterations in astrocyte function. First, we quantified the number of Glial fibrillary acidic protein (GFAP) positive astrocytes in the ION in controls, ARDS, and COVID-19 (Figure 2A). Interestingly, we found that compared to control, both ARDS and COVID-19 exhibited increased numbers of GFAP+ astrocytes, defined as GFAP+ somata, per unit area in all ION regions (Figure 2B). Notably, while control samples showed many GFAP+ astrocytic processes, few astrocytic cell bodies were labeled. Additionally, beta regression testing of GFAP per area data returned coefficient −0.0828 and p value 2.1E-08 for the control condition, consistent with our ANOVA results (Supplementary Table S4).

A caveat is worth mentioning here: detecting increased GFAP+ cells does not necessarily suggest that there were more astrocytes in one group vs. the other. Some astrocytes may exhibit lower levels of GFAP below the sensitivity of the assay, and can upregulate GFAP in pathologic contexts allowing its detection. Either way, the downstream interpretation of this phenomenon supports that in hypoxia (ARDS and COVID-19), there are elevated levels of astrogliosis.

Reactive astrocytes upregulate the expression of Aquaporin-4 (AQP4) (Tomas-Camardiel et al., 2004; Tourdias et al., 2011), and redistribute its expression to the cell soma from the astrocytic end-feet, where it is normally localized (Eid et al., 2005; Steiner et al., 2012). Moreover, AQP4 has important implications in hypoxic–ischemic conditions (Shi et al., 2012), and studies have shown that loss of AQP4 protects against early cytotoxic edema associated with stroke (Papadopoulos and Verkman, 2008). Also, AQP4 expression in astrocytes has important implications in neuroinflammation secondary to ischemia, and AQP4 knockout mice exhibit exaggerated post-stroke microglial reactivity (Shi et al., 2012). Thus, we asked if AQP4 levels were altered in COVID-19 vs. ARDS. We measured the area covered by AQP4 in ION (Figure 3A). In patients who had died of COVID-19, there was decreased expression of AQP4 in the ventral ION compared to the ARDS patients (Figure 3B). Together, these findings link lower AQP4 levels to increase neuroinflammation in COVID-19.

Encoded by the CHI3L1 gene, Chitinase-3-like protein (YKL-40) is a secreted glycoprotein primarily expressed in astrocytes in the brain that is a common marker of neurodegeneration (Querol-Vilaseca et al., 2017; Lananna et al., 2020; Zhao et al., 2020; Ferrari-Souza et al., 2022). Astrocytes increase the expression of YKL-40 in several neurodegenerative diseases including AD, tauopathies, and prion disease (Bonneh-Barkay et al., 2010; Llorens et al., 2017; Querol-Vilaseca et al., 2017). In vitro studies showed that YKL-40 could be induced in astrocytes by macrophages (Bonneh-Barkay et al., 2012). A recent study showed that YKL-40 knockout mice exhibit reduced amyloid plaques and increased expression of CD68 in microglia in an AD model, suggesting that YKL-40 suppresses microglial reactivity (Lananna et al., 2020). Thus, we examined the expression of YKL-40 in astrocytes in our cohort (Figure 4A). We were interested in knowing whether hypoxia in general can increase YKL-40, so for this analysis, we included the non-ARDS controls. We quantified the proportion of ALDH1L1 positive astrocytes that were also positive for YKL-40 and found that there were significantly more YKL-40 positive astrocytes in the hilum of ARDS and COVID-19 brains compared to non-ARDS controls (Figure 4B). This was not the case in the ION parenchyma (data not shown). However, there were fewer YKL-40 positive astrocytes in the ION hilum of the COVID-19 cohort compared to ARDS (Figure 4B). We examined astrocytic protein expression in all the comparisons we conducted, and YKL-40 is the only protein that we found dysregulated in the hilum. This is interesting given that it is a secreted protein (Zhao et al., 2020). Together, these results indicate that ION astrocytes behave differently under hypoxia in the setting of COVID-19 systemic infection; they fail to upregulate YKL-40 to the same extent as in ARDS. A caveat is that YKL-40 is a secreted protein, and that changes in YKL-40 levels between COVID-19 and ARDS may reflect changes in secretion patterns.

To further characterize astrogliosis in COVID-19 ION astrocytes, we performed multiplex immunofluorescence for other protein markers associated with reactive astrogliosis. We first quantified the expression of metallothionein-3 (MT3), a zinc-binding protein that has been shown to be upregulated in reactive astrocytes in Huntington disease (al-Dalahmah et al., 2020b). Metallothioneins are thought to be neuroprotective (Stankovic et al., 2007). Quantification of MT3 in different sectors of the ION of ARDS and COVID-19 showed no significant difference in the proportion of astrocytes that label with MT-3 (unpaired t-test, ARDS and COVID-19 mean ± SEM = 8.901 ± 2.731 and 14.27 ± 4.462, respectively, p value = 0.2262), however, when we stratified COVID-19 by the presence or absence of microglial nodules, we detected significantly lower proportions of lateral ION astrocytes in the COVID-19 with microglial nodules compared with ARDS patients (Supplementary Figures S2A,B). In accordance with this result, beta regression testing showed that COVID-MN condition had coefficient of—0.6191 and p value 0.0016, and that ION regions also had significant coefficients (Supplementary Table S4). We next asked if ION astrocytes in COVID-19 increase the expression of complement factor 3 (C3), which is a gene that is upregulated in and therefore a marker of putative neurotoxic “A1” astrocytes (Liddelow et al., 2017). We found no significant increase in the proportion of C3+ astrocytes in COVID-19 vs. ARDS (Supplementary Figures S3A,B). Finally, we examined the expression of CD44, an astrocyte protein expressed in white matter astrocytes [see our preprint (Al Dalahmah et al., 2023)], astrocytes around large vessels, interlaminar astrocytes, and a subset of cortical astrocytes (Sosunov et al., 2014), as well as Clusterin (CLU), which is increased in neurodegenerative astrocytes in AD (Wojtas et al., 2020; Chen et al., 2021). We quantified the area covered by CD44 and again found no significant increase in CD44 labeling in the COVID-19 ION (Supplementary Figures S4A,B). Likewise, we found no significant differences in the proportion of ION astrocytes that were CLU-positive between COVID-19 and ARDS (Supplementary Figures S4A,C).

In our design, we tried to control for relevant demographic and clinical variables (metadata), however this is not always possible. To determine the correlation between metadata variables and biological results, we performed principal component analysis (PCA) on all the immunohistochemical data from images from different regions in ION, using the metadata as supplementary variables, allowing us to predict their PCA coordinates from the IHC data. The input to the PCA analysis is provided in Supplementary Table S2. First, we plotted the brain donors in PCA space and found that ARDS and COVID-19 donors were relatively well separated (Figure 5A). This highlights the biological differences between the two groups. A closer look at the PCA results showed that a number of quantitative IHC variables were responsible for the greatest amount of variation in dimensions 1 and 2 (Supplementary Figure S5; Supplementary Table S3). In PC1, YKL40, C3, and a combination of the two (YKL40.C3) had correlation values of 0.898, 0.839, and 0.910, and p values of 3.27E-39, 1.77E-29, and 4.68E-42, respectively. Interestingly, the proportion of activated microglia was only weakly correlated with PC1. Metadata variables length of hospital stay, condition, sex, and DAD had low correlation (R2 values of 0.206, 0.069, 0.039, and 0.038, respectively—Supplementary Table S3) with PC1, suggesting that our case-control matching is not perfect, but sufficient. CD44 proportion and CLU per area were the most significant IHC variables associated with PC2, with correlation values of 0.620 and 0.508, and p values of 1.09E-12 and 2.35E-08, respectively. Astrocytes per unit area and proportion of MT3 positive astrocytes were significantly and strongly negatively correlated with PC2. The most relevant qualitative variable for PC2 was age, with a relatively low R2 value of 0.080 (Supplementary Table S3). This again shows high significance but low correlation of qualitative variables with the variance shown in the dimension, suggesting that our ARDS- COVID-19 matching was relatively effective. Supplementary Figure S5A shows the correlation circle depicting the IHC variables and their correlation to PC1 and PC2. Supplementary Figure S5B shows the correlation between the metadata variables and the first 5 PCs.

We next asked if clustering the samples (IHC images) based on the PCA dimensions would give us clusters that reflect condition, and/or other relevant variables like anatomic region for example. To achieve that, we clustered the data on the first 5 PC’s using hierarchical clustering on the Euclidian distance matrix derived from the PC1-5 coordinates for each sample. We identified four clusters as shown in Figure 5B. Examination of the hierarchical clustering results show that clusters 1 and 4 were relatively deplete of samples derived from the ION hilum compared to clusters 2 and 3. This is expected because the hilum is composed of white matter harboring axons and glia, unlike the ION parenchyma, which harbors neurons, too. To highlight any relationships between clusters and condition (ARDS vs. COVID-19), we plotted the proportion of images that fell under each cluster against condition in a heatmap (Figure 5C). The results show that ARDS samples were mainly enriched in clusters 1 and 3, while COVID-19 samples were distributed between clusters 2, 3, and 4. Together, these findings demonstrate that our samples cluster based on the major factors that our analysis set out to investigate, biological condition and anatomic locale.

A closer look at the distribution of metadata variables shows cluster 1 appears to be most enriched with old age (13), and short hospital stay (16) samples, and cluster 2 with COVID (20), short hospital stay (22), male (19), DAD (20), and non-septic (23) samples. Cluster 3 is most enriched with median-age (20), ARDS (15), short hospital stay (22), DAD (22), and non-septic (26) samples, and cluster 4 with COVID (18) and female (16) samples (Supplementary Figure S2C). The cos2 value of each variable, a good metric of variable correlation with the circumference of the correlation circle, also shows the same patterns we described by looking at the R2 above. Briefly, correlations in PC1 with length of hospital stay, sex, DAD, disease condition including the presence of microglial nodules in COVID-19, in PC2 with region and age, and in PC3 with region and length of hospital stay, respectively (Supplementary Figure S2B). All together, these data suggest that in addition to condition and anatomic locale, sex, concomitant DAD, and length of hospital stay were variables that correlated with IHC features and contributed to clustering. However, their overall correlation with the PC’s that explain the variance was low, suggesting they had a modest influence on the reactivity of astrocytes and microglia in the ION.