Fragmentation of rest periods, astrocyte activation, and cognitive decline in older adults with and without Alzheimer’s disease

Abstract

Introduction:

Sleep disruption is associated with astrocyte activation and impaired cognition in model organisms. However, the relationship among sleep, astrocyte activation, and cognition in humans is uncertain.

Methods:

We used RNA-seq to quantify the prefrontal cortex expression of a panel of human activated astrocyte marker genes in 1076 older adults in the Religious Orders Study and Rush Memory and Aging Project, 411 of whom had multi-day actigraphy prior to death. We related this to rest fragmentation, a proxy for sleep fragmentation, and to longitudinal cognitive function.

Results:

Fragmentation of rest periods was associated with higher expression of activated astrocyte marker genes, which was associated with a lower level and faster decline of cognitive function.

Discussion:

Astrocyte activation and fragmented rest are associated with each other and with accelerated cognitive decline. If experimental studies confirm a causal relationship, targeting sleep fragmentation and astrocyte activation may benefit cognition in older adults.

1 |. INTRODUCTION

Alzheimer’s disease (AD) and related dementias (ADRD) are a growing public health concern. In 2013, AD was the sixth leading cause of death in the United States.¹ Interventions to prevent cognitive impairment and dementia are urgently needed in our aging population. In older adults, sleep disruption commonly occurs and accumulating evidence suggests that its relationship to cognitive decline and dementia is bidirectional.^2–4 However, the mechanisms underlying these bidirectional effects in older adults are not fully clarified, impeding the development of effective treatments.

Astrocytes are among the most abundant glial cells in the central nervous system and in model organisms, sleep and astrocyte biology appear to be linked. In mice, sleep disruption can lead to astrocytic activation^5–7 that can be associated with cognitive impairment.^8–10

A body of evidence also suggests that astrocyte biology may play an important role in ADRD. AD pathology has been shown to be associated with astrocyte reactivity and a prominent feature of AD pathology is an abundance of activated glia (microglia and astrocytes) in close proximity to amyloid plaques, a pathological hallmark of AD.^11,12 Moreover, astrocyte abnormalities including astrogliosis, impaired synaptic homeostasis, and blood–brain barrier breakdown¹³ may precede the onset of classic AD pathology.¹⁴ Reactive astrocytes may contribute to amyloid beta (Aβ) accumulation, which may in turn contribute to a local inflammatory response.^15–18 Furthermore, persistent activation of microglia and astrocytes may be associated with AD progression.¹⁹

Based on the above, we hypothesized that astrocyte activation may represent a potential mechanism linking sleep disruption to cognitive decline.^5–21 We had previously shown that microglial aging and activation are linked to fragmentation of rest periods, a proxy for sleep fragmentation in older adults.²⁰ However, whether astrocyte activation is linked to sleep fragmentation in humans is uncertain given the challenges in obtaining indices of astrocyte activation, repeated measures of sleep and cognition, and post mortem measurements of ADRD neuropathologies in older adults—an important gap given the differences in astrocytes between model organisms and humans.²¹

To fill this knowledge gap, we used novel clinical and post mortem data obtained from 1076 older adults participating in the Religious Orders Study and Rush Memory and Aging Project (ROSMAP). We quantified the fragmentation of rest periods, a proxy for sleep fragmentation in 411 participants using continuous multi-day wrist activity monitoring. We leveraged available RNA-seq data from the dorsolateral prefrontal cortex (DLPFC) to assess the expression of genes characteristic of activated astrocytes. Although the DLPFC is not a classic sleep-regulatory node nor is it an early site of amyloid or tau accumulation, many of the cognitive domains most strongly associated with fragmentation of rest periods in our previous work²⁷ (e.g., working memory, perceptual speed) are subserved by the DLPFC. We related rest fragmentation to cognitive decline and to the DLPFC expression of these genes, while controlling for various clinical factors and ADRD neuropathologies. We tested the hypothesis that greater fragmentation of rest periods is associated with higher expression of genes characteristic of activated astrocytes. We also tested the hypothesis that higher expression of activated astrocyte marker genes would be associated with faster cognitive decline and lower level of cognitive function, after regressing out the negative effects of amyloid and neurofibrillary tangle pathology on cognition.

2 |. MATERIALS AND METHODS

Additional description of the materials and methods can be found in the eMethods in supporting information.

2.1 |. Participants

As described more fully in the supporting information, we examined participants from two community-based cohort studies of older persons, ROSMAP. Participants enrolled from 1994 to 2020 and were followed up longitudinally until death. Their characteristics are summarized in Table 1. We related fragmentation of rest periods to astrocyte gene expression in 411 MAP participants with at least one actigraphic recording and DLPFC RNA-seq data. Then we related the expression of astrocyte genes to cognition in 1076 participants with at least one cognitive assessment and DLPFC RNA-seq, excluding participants who overlapped with one of the single nucleus RNA sequencing (snRNA-seq) datasets²² used to identify activated astrocyte marker genes^22,23 (see section 2.3).

TABLE 1.

Characteristics of study population

Characteristics	Participants with RNA sequencing and actigraphy*	Participants with RNA sequencing and without actigraphy*	Difference between participants with (n = 411) and without (n = 665) actigraphy: P-value*	Participants with RNA sequencing and cognitive assessment*
Number	411	665		1076
Age at last visit (years)	88.6 [86.8, 94.3]	87.3 [83.5,92.3]	4.3E-9	89.6 [84.7, 93.1]
Age at death (years)	91.1 [87.6, 95.4]	88.6 [84.4, 93.1]	4.6E-10	89.6 [85.5, 93.9]
Female sex	297 (72.3%)	423 (63.6%)	0.0039	720 (66.9%)
European descent	406 (98.8%)	654 (98.3%)	0.60	1060 (98.5%)
Years of education	15.0 [12.0, 16.0]	18.0 [15.0, 20.0]	< 2.2E-16	16.0 [14.0, 18.0]
Mini-Mental State Examination score	25.0 [17.0, 28.0]	25.0 [16.0,28.0]	0.52	25.0 [16.0, 28.0]
Residual composite global cognition	0.18 [–0.28, 0.22]	0.19 [–0.29, 0.22]	0.45	0.18 [–0.28, 0.22]
Clinical diagnosis AD	164 (39.5%)	287 (43.2%)	0.26	451 (41.8%)
Time from last cognitive assessment to death (years)	0.6 [0.4, 0.9]	0.6 [0.3, 0.9]	0.050	0.6 [0.4, 0.9]
Baseline sleep fragmentation (KRA)	0.029 [0.016, 0.082]	NA	NA	NA
Number of actigraphy recordings	3.9 [1.0, 10.0] Median = 3.0	NA	NA	NA
Time from last actigraphy to death (years)	1.9 [0.8,3.6]	N/A	N/A	NA
Post mortem interval (hours)	7.0 [5.8, 8.8]	6.0 [4.5, 9.8]	1.4E-06	6.5 [5.0, 9.0]
Amyloid level (% area)	1.84 [0.00, 4.79]	1.56 [0.00, 4.46]	8.043e-05	1.67 [0.00,4.79]
Tangle density (number per mmˆ2 squared)	2.41 [0.00, 8.86]	2.17 [0.00,7.81]	0.0075	2.26 [0.00, 8.86]
NIA-Reagan pathological diagnosis AD (intermediate/high)	141 (34.3%)	257 (38.1%)	0.19	398 (37.0%)
Presence of Lewy body pathology	100 (24.8%) n = 404	142 (22.0%) n = 643	0.33	242 (22.4%) n = 1047
Macroscopic infarcts present	66 (16.1%)	85 (12.6%)	0.029	141 (13.1%)
Microscopic infarcts present	87 (21.2%)	110 (16.3%)	0.014	187 (17.3%)
Extralimbic TDP-43 pathology	143 (34.6%) n = 413	171 (29.0%) n = 590	0.068	314 (31.3%) n = 1003
Presence of hippocampal sclerosis	36 (8.7%)	53 (8.0%)	0.80	89 (8.2%)

^*Values represent mean, square brackets represent range, and parentheses represent proportion of total population. Abbreviations: AD, Alzheimer’s disease; NIA, National Institute on Aging; TDP-43, TAR DNA-binding protein 43.

2.2 |. Ethics approval

The ROSMAP studies were approved by the institutional review board of Rush University Medical Center and conducted in accordance with the Declaration of Helsinki. All participants provided written informed consent and signed an anatomic gift act for brain donation and a repository consent for data sharing.

2.3 |. Identification of human activated astrocyte marker genes

To identify marker genes for human activated astrocytes, we reanalyzed post mortem human brain snRNA-seq data from two published datasets.^22,23

In snRNA-seq, transcriptomes are assessed at the individual nucleus level rather than at the tissue level, allowing comparison of the gene expression of identifiable cell types. In this study, the assessment of transcriptomes at single-cell resolution allowed us to compare gene expression in transcriptionally identified nuclei from activated astrocytes versus non-activated astrocytes to identify marker genes manifesting significantly higher expression in the former compared to the latter and compared to identified non-astrocyte nuclei.

As described in the eMethods, we re-normalized, re-scaled, and re-clustered these data (Figure 1A). We identified clusters corresponding to broad brain cell types based on expression of cell-type–specific genes: excitatory neurons (SLC17A7 and CAMK2A), inhibitory neurons (GAD1 and GAD2), astrocytes (AQP4), oligodendrocytes (MBP), microglia (CSF1R), oligodendrocyte precursor cells (VCAN), and endothelial cells (FLT1). Then, we subclustered those cells identified as astrocytes (Figure 1B), and identified which subclusters corresponded to reactive astrocytes based on the expression of the known reactive astrocyte genes²⁴ GFAP, CD44, OSMR, and CHI3L1 (Figure 1C). Next, we identified genes significantly overexpressed in these activated astrocytes versus all other astrocytes, and versus all other cells. This constituted our first activated astrocyte marker gene set. We repeated these analyses on a second human brain snRNA-seq dataset.²³ We took the union of both gene sets to create a combined reactive astrocyte marker gene set we used for further analysis. Using a similar approach, we also identified marker genes characteristic of astrocytes in general.

FIGURE 1.

snRNA-seq distinguishes major brain-cell types and astrocyte subtypes.²² Re-analysis of snRNA-seq data from our primary snRNA-seq dataset.²² A, t-SNE plot showing seven major brain cell type clusters: Oli = oligodendrocytes, Mg = microglia, Ex = excitatory neurons, OPC = oligodendrocyte progenitor cell, Ast = astrocyte, Inh = inhibitory neurons, Endo = endothelial cells. Number of individual nuclei = 162,562. B, Subclustering of astrocyte nuclei. C, Dot plot showing expression levels (as color scale) and the percent of cells expressing (as dot size) known astrocyte marker gene, AQP4, and activated astrocyte marker genes:²⁴ GFAP, CD44, OSMR, and CHI3L1, across all clusters. D, Composite expression of activated astrocyte marker genes by cluster

2.4 |. Assessment of rest fragmentation

As described in the supporting information, fragmentation of rest periods was assessed biennially in MAP participants from up to 10 days of continuous (24 hours/day) wrist activity monitoring using the metric kRA.²⁶ Briefly, kRA represents the probability per 15-second epoch of having an arousal (i.e., non-zero activity count), as indicated by movement, after a long (≈5 minutes) period of rest (i.e., sleep). The higher the kRA, the more quickly bouts of sleep/rest end in arousal and thus the greater the degree of sleep fragmentation. Because older adults do not only sleep during the night, kRA was computed on the basis of the entire day to include periods of sleep whenever they occur. We have previously shown that the value of kRA, calculated from the entire day, is strongly correlated with kRA computed from the six consecutive hours of greatest rest²⁶ and so kRA computed from the entire day is largely reflective of fragmentation during the principal sleep period. Moreover, we have shown that kRA correlates well with polysomnographic metrics of sleep fragmentation.²⁵

Participants had a mean (standard deviation) of 3.9 (2.33) actigraphy recordings prior to death. For all participants, we used the first available actigraphy recording for each participant as our rest fragmentation measure, which is more likely to reflect the degree of rest fragmentation to which the participant was exposed during the earlier stages of AD pathogenesis, and is less likely to be affected by terminal changes in health status.

2.5 |. Assessment of cognition

As described in detail in the supporting information, ROS and MAP participants underwent annual structured cognitive assessments from which a previously published composite measure of global cognitive function was computed. The composite measure of global cognitive function was obtained from 19 cognitive tests from five cognitive domains: processing speed, semantic memory, working memory, visuospatial ability, and episodic memory.

Linear mixed effect models, adjusted for amyloid level at autopsy, tangle density at autopsy, age, sex, education, and post mortem interval, were fit to the longitudinal cognitive data in the decade prior to death (Figure S1 in supporting information) to obtain the residual level and rate of cognitive decline after accounting for amyloid and neurofibrillary tangle pathology, which were used as our co-primary cognitive outcomes in all models.

2.6 |. Assessment of clinical and pathologic covariates

Age, sex, education level, history of smoking, alcohol consumption, sleeping pill usage, number of vascular diseases in life, number of vascular risk factors in life, and body mass index were assessed as described in the supporting information. Amyloid level, tangle density, Lewy body pathology, macroscopic cerebral infarcts, microinfarcts, TAR DNA-binding protein 43 (TDP-43) pathology, hippocampal sclerosis, cerebral amyloid angiopathy, cerebral atherosclerosis, and arteriolosclerosis were assessed as described in the supporting information.

2.7 |. Assessment of DLPFC astrocyte and microglial gene expression

As described in the supporting information, RNA sequencing was carried out on frozen DLPFC tissue, as part of the parent ROSMAP studies. Although the choice of the DLPFC for RNA sequencing was driven by the parent studies, we do note that some of the cognitive domains most strongly associated with sleep fragmentation in our previous work²⁷ (e.g., working memory, perceptual speed) are subserved by the DLPFC.

For each gene in our combined activated astrocyte marker gene set, we computed the normalized expression by subtracting the mean for that gene across all samples (in counts per million) and dividing by the standard deviation. For the combined activated astrocyte gene set, and secondarily for each separate gene set, we then computed a summary z-score by taking the mean normalized expression across all genes in the set as previously described.²⁸

2.8 |. Statistical analysis

Our primary objectives were to test the hypotheses that (1) greater fragmentation of rest periods is associated with higher expression of activated astrocyte marker genes, and that higher expression of activated astrocyte marker genes is associated with (2) lower level and (3) faster decline of composite global cognitive performance in the decade before death. For these primary analyses, we used Bonferroni correction for three comparisons and set the threshold for significance at P = 0.05/3 = 0.017. For the remainder of the secondary analyses noted below, we set the nominal threshold for significance at P = 0.05.

To test the primary hypothesis, we used a multiple linear regression model to examine the association between the first available measurement of rest fragmentation (predictor), and the composite expression of activated astrocyte marker genes in our combined gene set (outcome). Then, we used a series of multiple linear regression models (one per gene) to examine the association between fragmentation of rest periods, measured using kRA, and the expression of each individual marker gene in our combined activated astrocyte marker gene set. Second, we repeated these analyses using each separate reactive astrocyte marker gene set, and our general astrocyte marker gene sets. Then, we considered models augmented with terms for ADRD neuropathologies and clinical covariates.

To test our second and third primary hypotheses, we used two separate multiple linear regressions to examine the association between the composite expression of our combined activated astrocyte gene set (predictor) and, in separate models, the residual level and rate of decline of the composite global cognitive summary score (outcomes). Second, we repeated these analyses replacing global cognition with each of the five constituent cognitive abilities. Then, we considered models augmented with terms for ADRD neuropathologies and clinical covariates.

Next, we used mediation analyses (mediate function in the mediation package version 4.5.0²⁹) to consider whether astrocyte activation and fragmentation of rest periods, measured by kRA, mediated each other’s associations with cognition. Last, we used mediation analysis to examine whether the expression of aged microglial marker genes linked fragmentation of rest periods to astrocyte activation, or astrocyte activation to cognitive impairment.

3 |. RESULTS

3.1 |. Study participants

We studied 1076 older community-dwelling adults with DLPFC RNA-seq and cognitive assessment, 411 of whom also had assessment of rest fragmentation by actigraphy. Their characteristics are in Table 1 and supplemental Figure S2A in supporting information. Participants with available actigraphy (drawn exclusively from the MAP cohort) were on average 1 to 2 years older than those without (drawn from both the ROS and MAP cohorts), slightly more likely to be female, and had a median of three fewer years of education.

3.2 |. Identification of human activated astrocyte marker genes

To identify marker genes for human activated astrocytes, we reanalyzed post mortem human brain snRNA-seq data from a study of 162,562 nuclei from older adults.²² After re-normalization and re-clustering, we identified the cluster corresponding to astrocytes based on AQP4 expression (Figure 1A). We then re-clustered within the astrocyte population revealing a total of 15 astrocyte subclusters (Figure 1B). We identified subclusters 2, 6, and 10 as corresponding to activated astrocytes based on expression of known activated astrocyte marker genes GFAP, CD44, OSMR, and CHI3L1 (Figure 1C).

To identify a panel of marker genes for this population of human activated astrocytes, we performed two differential expression analyses: one comparing astrocyte subclusters 2, 6, and 10 against all other astrocyte subclusters, and another against all other clusters (including non-astrocyte clusters). We identified a set of marker genes (n = 25) with significantly higher expression (log2FC > 1, false discovery rate [FDR] q < 0.05) in astrocyte subclusters 2, 6, and 10 in both comparisons (Table S1 in supporting information).

For each single-nucleus transcriptome, we computed a composite measure of activated astrocyte marker gene expression by taking the average normalized expression across all 25 genes. This clearly differentiated activated astrocytes (clusters 2, 6, and 10) from all other cell types (Figure 1D). Using similar methods, we also identified a set of 228 general astrocyte marker genes.

We repeated the above using another human aged brain snRNA-seq data dataset²³ (Figure S3 in supporting information) to generate a secondary set of 32 activated astrocyte marker genes and 313 general astrocyte marker genes.

We then combined genes from both the reactive astrocyte gene sets to create a combined reactive astrocyte gene set used for downstream analysis (Table S2 in supporting information). We repeated the same for the general astrocyte marker genes.

3.3 |. Fragmentation of rest periods is associated with higher expression of activated astrocyte marker genes

Four hundred eleven MAP participants had measurements of both rest fragmentation, quantified from ante mortem actigraphy using the metric kRA, and DLPFC gene expression quantified by bulk RNA-seq. For our first primary analysis, in a linear regression model adjusted for age, sex, education, post mortem interval, the burden of amyloid pathology, and neurofibrillary tangle density, greater fragmentation of rest periods, quantified by kRA, was associated with higher composite expression of activated astrocyte marker genes (estimate = +9.63; standard error [SE] = 0.030; P = 0.0014; Figure 2B). In contrast, kRA was not associated with the expression of general astrocyte marker genes (Figure S4A–B in supporting information). We repeated these analyses considering the composite expression of each of the two substituent activated astrocyte gene sets. Despite an incomplete overlap between the two gene sets (Figure S2B), and differences in the brain regions from the snRNA-seq datasets, results were similar (Figure S5 and S6 in supporting information) supporting the robustness and generalizability of these findings.

FIGURE 2.

Ante mortem rest fragmentation, expression of our activated astrocyte marker genes derived from our combined gene set, and cognition. A, Volcano plots of –log10 (P value) versus effect size for normalized expression of genes characteristic of reactive astrocytes. Y axis is the –log10 (P value), and the x axis is the difference in the normalized expression of the genes characteristic of reactive astrocyte per 1 standard deviation difference in rest fragmentation. Each dot represents a gene and the dotted line indicates unadjusted P < 0.05. False discovery rate–significant genes are labelled. B, Partial residual plot of composite expression of genes characteristic of reactive astrocytes as a function of ante mortem rest fragmentation adjusted for age, sex, education, post mortem interval, burden of amyloid pathology, and neurofibrillary tangle density. Y axis is the composite expression for the activated astrocyte gene set. X axis is ante mortem rest fragmentation. Each dot represents a single participant. Solid line indicates the predicted composite gene expression for an average participant. Dotted lines indicate 95% confidence intervals (CIs) on the prediction. C, Partial residual plot of residual global cognitive score as a function of composite expression of genes characteristic of human activated astrocytes. Each dot represents a single participant. Solid line represents the predicted cognition for an otherwise average participant. Dotted lines indicate 95% CIs on the prediction. D, Partial residual plot of residual cognitive decline, as a function of composite expression of genes characteristic of human activated astrocytes. Each dot represents a single participant. Solid line represents the predicted cognition for an otherwise average participant. Dotted lines indicate 95% CIs on the prediction. SE, standard error

We next looked at individual genes. In linear models adjusted for age at death, sex, education, post mortem interval, the burden of amyloid pathology, and the density of neurofibrillary tangles, the expression levels of 13 activated astrocyte marker genes (IFI16, TPST1, CCL2, ZFP36, OSMR, SLC44A3, ANO6, FAM189A2, C4orf19, RANBB9, GGPS1, UST, and NRP1) were associated with rest fragmentation at FDR q < 0.05 (Figure 2A and Table S3 in supporting information). For each of these, more fragmented rest was associated with higher expression (Figure S7 in supporting information). When we looked back at the snRNA-seq data, these genes were expressed in all our activated astrocyte subsclusters (clusters 2, 6, and 10 in our primary snRNA-seq dataset and clusters 3 and 6 in our secondary snRNA-seq dataset, Figure S8 in supporting information) and did not clearly differentiate one cluster from the other.

Numerous ADRD neuropathologies and lifestyle factors are associated with sleep fragmentation.^25,30,31 However, the association between rest fragmentation and higher composite expression of activated astrocyte marker genes remained significant in models controlling for clinical factors (smoking, alcohol consumption, use of sleeping pills, body mass index, vascular disease, and vascular risk factors) and a broad spectrum of ADRD neuropathologies (amyloid levels, tangle density, Lewy body pathology, TDP-43 pathology, hippocampal sclerosis, large vessel atherosclerosis, arteriolosclerosis, cerebral amyloid angiopathy, microscopic cortical infarcts, macroscopic cortical infarcts) alone or in combination (P < 0.05; Table 2). Additionally, we found that these clinical factors and ADRD neuropathologies were not themselves associated with composite expression of activated astrocyte marker genes (Table S4 in supporting information).

TABLE 2.

Ante mortem fragmentation of rest periods and composite expression of genes characteristic of reactive astrocytes: adjustment for neuropathology and clinical factors

Model	Outcome	Covariates	Predictor	Estimate (per 1 SD greater kRA)	SE	P-value
A	Composite expression of activated astrocyte genes	Age + sex + education + post mortem interval + amyloid + neurofibrillary tangle density	Ante mortem level of rest fragmentation kRA	0.08	0.025	1.45E-03
B	Composite expression of activated astrocyte genes	A + all pathologiesa	Ante mortem level of rest fragmentation kRA	0.09	0.03	7.38E-04
C	Composite expression of activated astrocyte genes	A + all clinical covariatesb	Ante mortem level of rest fragmentation kRA	0.08	0.03	1.65E-03

^aAmyloid level, tangle density, Lewy body pathology, macroscopic cerebral infarcts, microinfarcts, TDP-43 pathology, hippocampal sclerosis, cerebral amyloid angiopathy, cerebral atherosclerosis, and arteriolosclerosis.

^bSmoking, alcohol consumption, use of sleeping pills, body mass index, vascular disease, and vascular risk factors.

Abbreviations: SD, standard deviation; SE, standard error; TDP-43, TAR DNA-binding protein 43.

3.4 |. Higher expression of activated astrocyte marker genes is associated with cognitive impairment

We next examined, in separate models, the association between the expression of genes characteristic of activated astrocytes and residual level and rate of change of cognition in the decade prior to death, after accounting for sex, age, education, post mortem interval, amyloid burden, and tangle density post mortem. We analyzed data from 1076 ROSMAP participants who had DLPFC RNA-seq data and at least one cognitive assessment. There was no overlap between these 1076 participants and participants in the two snRNA-seq datasets used to identify marker genes. Greater composite expression of activated astrocyte marker genes was associated with lower residual level and more rapid residual rate of decline of composite global cognition score (level estimate = −0.142, SE = 0.053, P = 0.0075; slope estimate = −0.012; SE = 0.004, P = 0.0025; Figure 2C–D). In contrast, there was no significant association between composite expression of generic astrocyte marker genes and residual level or rate of change of the composite global cognitive score (Figure S4C–D) indicating that this relationship is specific to activated astrocytes rather than astrocytes in general. In secondary analyses, we also examined each of the five cognitive domains separately. Greater composite expression of activated astrocyte marker genes was associated with lower residual level of visuospatial function, processing speed, and semantic memory, and with more rapid decline in all five cognitive abilities (P < 0.05; Table S5).

In models controlling for the same clinical and ADRD neuropathology factors as with the rest fragmentation analyses, these association remained significant (P < 0.05; Table 3). Thus, the association between the expression of genes characteristic of human activated astrocytes and cognition is not accounted for by these ADRD neuropathologies or clinical factors.

TABLE 3.

Activated astrocyte marker gene expression from primary gene set, and residual and rate of change of composite global cognition: adjustment for clinical factors and neuropathology

Model	Outcome	Covariates	Predictor	Effect of 1 unit greater composite expressionestimate (SE) P-value
A	Residual cognition level	Age + sex + education + post mortem interval + amyloid + neurofibrillary tangle density	Composite expression of activated astrocyte genes	−0.14 (0.05) 0.0074
B	Residual cognition slope	Same as A	Composite expression of activated astrocyte genes	−0.012 (0.0039) 0.0024
C	Residual cognition level	A + all neuropathologiesa	Composite expression of activated astrocyte genes	−0.11 (0.03) 0.039
D	Residual cognition slope	A + all neuropathologiesa	Composite expression of activated astrocyte genes	−0.0097 (0.0039) 0.012
E	Residual cognition level	A +all clinical covariatesb	Composite expression of activated astrocyte genes	−0.071 (0.03) 0.0090
F	Residual cognition slope	A + all clinical covariatesb	Composite expression of activated astrocyte genes	−0.011 (0.0039) 0.00362

^aAmyloid level, tangle density, Lewy body pathology, macroscopic cerebral infarcts, microinfarcts, TDP-43 pathology, hippocampal sclerosis, cerebral amyloid angiopathy, cerebral atherosclerosis, and arteriolosclerosis.

^bSmoking, alcohol consumption, use of sleeping pills, body mass index, vascular disease, and vascular risk factors.

Abbreviations: SE, standard error; TDP-43, TAR DNA-binding protein 43.

3.5 |. Mediation analyses

Next, in the subset of individuals with measurement of rest fragmentation, cognitive function, and activated astrocyte marker gene expression, we used mediation models to examine the extent to which rest fragmentation and activated astrocyte marker gene expression mediated each other’s associations with cognitive trajectories. We considered models in which astrocyte activation was hypothesized to mediate the association between rest fragmentation and residual cognitive level and decline and models in which rest fragmentation was hypothesized to mediate the association between astrocyte activation and residual cognitive level and decline. Examining residual cognitive level, the association between rest fragmentation and cognition remained significant when controlling for astrocyte activation; however, the association between astrocyte activation and cognition was no longer significant when controlling for rest fragmentation (Table S6, model C in supporting information). The average causal mediation effect was not significant for rest fragmentation→astrocyte→cognitive level (Table S7, model 1A in supporting information) but was weakly significant for astrocyte→rest fragmentation→cognitive level (Table S7, model 1B); however, the estimate for the proportion mediated effect did not reach statistical significance (P = 0.066) suggesting that this was a weak effect at best. Examining residual cognitive decline, the association between rest fragmentation and cognition was no longer significant when controlling for astrocyte activation and the association between astrocyte activation and cognition was no longer significant when controlling for rest fragmentation (Table S6, model F). The average causal mediation effect was not significant for the astrocyte→rest fragmentation→cognitive decline relationship (Table S7, model 2B) and the rest fragmentation→astrocyte→cognitive decline relationship (Table S7, model 2A). Overall, these results are inconclusive.

As reported previously, we computed a composite measure of the expression marker genes characteristic of aged microglia.³² There was a moderate correlation between the composite expression of aged microglia marker genes, and activated astrocyte marker genes (Pearson’s r = 0.43 with P < 2 × 10−16, n = 411; Figure S9 in supporting information). In mediation analyses, when we considered the rest fragmentation → microglial aging → astrocyte activation pathway, the association between rest fragmentation and the composite expression of activated astrocyte marker genes was partially mediated by the composite expression of aged microglia marker genes (Table S7, model 3A), with the composite microglia measure accounting for 37% of the association. However, when we considered the rest fragmentation → astrocyte activation → microglial aging pathway, the association between rest fragmentation and the composite expression of aged microglia marker genes was partially mediated by the composite expression of activated astrocyte marker genes (Table S7, model 3B) with the activated astrocyte measure accounting for 43% of the association. When we examined the association among aged microglia, activated astrocytes, and residual level of cognition, the composite expression of aged microglia marker genes did not significantly mediate the association between the expression of activated astrocyte marker genes and cognitive impairment (Table S7, model 4A). When we examined the aged microglia → activated astrocyte → residual cognition pathway, the average causal mediated effect was nominally significant (Table S7, model 4B) but not the proportion mediated (P = 0.16) suggesting a weak effect at best. Taken together, this is in keeping with a complex bidirectional relationship between rest fragmentation, astrocyte activation, and microglial aging, along with the possibility that astrocyte activation may link microglial aging to impaired cognition.

4 |. DISCUSSION

In this cross-sectional study of older community-dwelling adults, the fragmentation of rest periods, a proxy for sleep fragmentation, was associated with higher expression of genes characteristic of human activated astrocytes, and higher expression of activated astrocyte genes was associated with worse residual cognitive function, and faster residual cognitive decline, independent of age, a range of clinical and demographic factors, and ADRD neuropathologies.

4.1 |. The fragmentation of rest periods, a proxy for sleep fragmentation, is associated with greater expression of genes characteristic of human reactive astrocytes

In this study greater fragmentation of rest periods, a proxy for sleep fragmentation, was associated with higher expression of genes characteristic of human reactive astrocytes. Similar results were not seen for general astrocyte marker genes, indicating that these results are specific to activated astrocytes.

Whether fragmentation of rest periods is a contributor to or result of astrocyte activation is not possible to conclude from these cross-sectional data. Data from model organisms suggests that both may be plausible. In mice, sleep disruption can lead to astrocytic activation.^5–7 Meanwhile, astrocyte processing of adenosine plays an important role in sleep homeostasis,^33–35 and in human subjects, polymorphisms in aquaporin 4, a key astrocyte water channel, are associated with differences in sleep depth as measured by electroencephalography (EEG).³⁶

Reactive astrocyte genes associated with sleep fragmentation after FDR correction have functions in inflammation (zinc finger protein 36 [ZFP36]); cellular proliferation (interferon gamma inducible protein 16 [IFI16]); nuclear hormone regulation (geranylgeranyl pyrophosphate synthase 1 [GGPS1]); cytokine signaling (oncostatin M receptor [OSMR]); molecular transport (solute carrier family 44 member 3 [SLC44A4], tyrosylprotein sulfotransferase 1 [TPST1], anoctamin 6 [ANO6]); and cell–cell interactions, cell adhesion, and migration (family with sequence similarity 189 member A2 [FAM189A2], monocyte chemoattractant protein-1 [MCP-1/CCL2], neuropilin 1 [NRP1], RAN binding protein 9 [RANBP9]). Further studies on these genes are needed to clarify their functions and their associations with sleep fragmentation and astrocyte function.

4.2 |. Greater expression of genes characteristic of human activated astrocytes is associated with worse cognitive function

Higher expression of genes characteristic of activated astrocytes was associated with lower residual cognitive function and faster residual cognitive decline, accounting for the burden of amyloid and neurofibrillary tangles, and independent of age, lifestyle and clinical factors, and ADRD neuropathologies. This is concordant with experiments in rats, in which sleep deprivation leads to astrocyte activation and secretion of proinflammatory cytokines that impair spatial memory.³⁷

Reactive astrocytes may affect cognition through several mechanisms. In model organisms, activated astrocytes can increase inflammation and enhance Aβ formation and tau hyperphosphorylation, which are pathological hallmarks of AD.^38,39 They also release cell complement proteins, which enhance inflammatory processes and cause cell lysis,³⁸ which may in turn potentiate cognitive impairment.⁴⁰ Last, activation of alpha-1-antichymotrypsin (ACT), an astrocyte serine protease, inhibits amyloid plaque breakdown and leads to tau hyperphosphorylation.⁴¹

Our mediation analyses relating sleep fragmentation, expression of activated astrocyte marker genes, and residual cognitive function and decline were largely inconclusive. Future longitudinal and experimental studies are needed to better delineate the true causal relationship among sleep, astrocyte activation, and cognition.

4.3 |. Role of microglia

In model organisms, microglial and astrocyte activation are closely linked. Microglia play a role in activating and secreting signals that trigger astrocyte activation and astrocytes have also been shown to influence microglia activation through secreted proteins.⁴² Furthermore, the interaction between activated microglia and astrocytes plays a vital role in the process of neuroinflammation.⁴³

We had previously shown that greater fragmentation of rest periods is associated with microglial aging and activation.⁴³ In the present study, samples with higher expression of aged microglia genes also had higher expression of activated astrocyte genes, supporting a relationship between the two. Controlling for aged microglial gene expression attenuated the association between rest fragmentation and expression of activated astrocyte marker genes, and vice versa, compatible with a bi-directional scenario in which microglial aging and astrocyte activation may partially mediate the others’ associations with rest fragmentation. However, the association between activated astrocyte gene expression and cognition was independent of aged microglial gene expression. Further work in model organisms will be needed to disentangle the causal links among sleep, microglial changes, and astrocyte activation.

4.4 |. Methodological considerations

Several limitations merit discussion. First, this study was cross-sectional, precluding definitive causal inference. Moreover, the post mortem nature of the study precluded contemporaneous measurement of AD pathology, cognition, sleep, and the expression of activated astrocyte genes. Studies relating longitudinal changes in rest fragmentation and human biomarkers of astrocyte activation (e.g., cerebrospinal fluid CH3IL1/YKL-40) may be helpful. Second, although we have shown that kRA is a good proxy for polysomnography-measured sleep fragmentation,²⁷ actigraphy indirectly infers sleep from immobility and cannot distinguish the different causes of sleep of fragmentation (e.g., sleep disorders such as sleep apnea and restless legs versus environmental factors such as light and noise versus other medical comorbidities). Ambulatory polysomnography, including EEG, could provide a more definitive differentiation of wake/sleep, provide additional information regarding sleep staging and depth, and identify some causes of sleep fragmentation (e.g., sleep apnea). While several studies have demonstrated the feasibility of ambulatory EEG in adults with mild dementia,^44,45 it remains challenging to obtain ambulatory EEG in adults with more severe dementia, which would be important in a post mortem study such as ours. Third, bulk RNA-seq cannot distinguish between higher expression of reactive astrocyte genes due to more reactive astrocytes or increased gene expression without change in cell number. Moreover, it cannot exclude ectopic expression astrocyte genes in other cell types. We think that ectopic expression alone is unlikely to account for our results given the specificity of our gene sets in the snRNA-seq datasets examined. However, new studies combining snRNA-seq with sleep and cognitive assessments are needed. Fourth, RNA-sequencing was performed only in the DLPFC. While the DLPFC is involved in several of the cognitive domains most strongly associated with sleep fragmentation, including working memory and processing speed, it is not known as a site of early Aβ or tau aggregation, nor is it a core sleep regulatory node. Further studies examining other regions involved in sleep regulation and/or early Aβ or tau aggregation, such as the lateral orbitofrontal cortex, are needed. Fifth, participants from the study are mainly from European ancestry and the findings here may not be generalizable to populations from other ancestries. Additional studies replicating these results in populations of different ancestries are needed. Sixth, individuals in this study are relatively older (≈90 years old at death) and further studies are needed in younger populations to determine whether the results are generalizable to younger individuals. Seventh, we acknowledge that the proportion of variance in global cognitive outcomes accounted for by astrocyte activation is modest (e.g., partial R-squared of 0.7% for residual global cognitive level). This is somewhat higher for cognitive domains partially subserved by the DLPFC (e.g., partial R-squared of 1% for residual level in processing speed), which is greater than or similar to the proportion of variance accounted for by other ADRD-related brain pathologies (e.g., partial R-squared of 0.4% for microscopic infarcts, 0.2% for macroscopic infarcts, and 2% for Lewy bodies) and when multiplied across the population level, even these relatively small effects can have important clinical consequences. We hypothesize that if astrocyte activation were assessed in other areas relevant to AD-related cognitive deficits (e.g., hippocampus) the correlation may be stronger.

This study also had several strengths. First, we used an objective sleep measure, actigraphy, to assess a validated proxy of sleep fragmentation. Whereas self-report sleep measures can be confounded by poor recall and do not always correlate well with objective sleep measures⁴⁶ we have shown that the actigraphic metric used here correlates well with polysomnographic measures of sleep fragmentation.²⁵ Moreover, compared to in-laboratory polysomnography, actigraphy minimally perturbs natural sleep behavior and can be performed continuously for several days, enabling data collection from individuals who may not tolerate sleeping in a laboratory environment. Second, we derived our set of activated astrocyte marker genes from human snRNA-seq data rather than model organisms and confirmed the specificity of these genes for activated human astrocytes. This is particularly important given the transcriptional differences between rodent and human astrocytes.²¹ Third, we quantified various dementia pathologies using gold-standard histopathology, rather than relying on clinical diagnoses. Fourth, this study examined a cohort of particularly aged adults (≈90 years old), a demographic group whose proportion of the overall population is increasing over time, and yet which remains relatively understudied compared to younger seniors.

4.5 |. Future directions and clinical implications

Our results suggest that astrocyte activation may be an important correlate of sleep fragmentation and that both sleep fragmentation and astrocyte activation are associated with cognitive decline. Our results are compatible with studies in model organisms suggesting a causal relationship between these factors. However, the cross-sectional nature of our study makes it impossible to determine the causal nature of the associations we observed in humans. Looking forward, this study may lay the foundation for future longitudinal and experimental studies in older adults to allow for more definitive causal inference in humans and highlights the potential clinical importance of further work in model organisms to dissect the mechanisms linking sleep disruption, astrocyte activation, and impaired cognition. Several lines of investigation need to be pursued. First our results should be confirmed using cell-specific sequencing approaches such as snRNA-seq. Second, longitudinal studies with in vivo markers of sleep, astrocyte activation, and cognition to disentangle the causal relationships among sleep, astrocyte activation, and cognition are needed. Third, astrocyte activation should be investigated in relation to other clinical correlates of sleep fragmentation in aging. Fourth, experiments are needed to investigate whether modifying sleep fragmentation (i.e., by treating sleep disorders) can affect in vivo markers of astrocyte activation and cognition. Finally, studies are needed to examine the impact of targeting astrocyte biology (e.g., with fluorocitrate⁴⁷ or other therapeutics) on cognition, sleep, and other aging phenotypes in older adults.

Highlights.

• Greater fragmentation of rest periods, a proxy for sleep fragmentation, is associated with higher composite expression of a panel of genes characteristic of activated astrocytes.

• Increased expression of genes characteristic of activated astrocytes was associated with a lower level and more rapid decline of cognitive function, beyond that accounted for by the burden of amyloid and neurofibrillary tangle pathology.

• Longitudinal and experimental studies are needed to delineate the causal relationships among sleep, astrocyte activation, and cognition.

RESEARCH IN CONTEXT.

1. Systematic Review: The authors performed a literature review encompassing published articles investigating the relationships among sleep fragmentation, astrocyte activation, and cognitive impairment in model organisms. However, studies examining the association between sleep fragmentation and astrocyte activation in older humans are lacking.

2. Interpretation: In a cohort study of 1076 participants from the Religious Orders Study and Rush Memory and Aging Project (ROSMAP), 411 of whom had both ante mortem measurement of rest fragmentation, a proxy for sleep fragmentation, by actigraphy, and post mortem evaluation of astrocyte marker genes, we found that increased fragmentation of rest periods was associated with higher expression of reactive astrocyte marker genes, which was associated with a lower level and more rapid decline of cognitive function, independent of amyloid and neurofibrillary tangle pathology.

3. Future Directions: If longitudinal and experimental studies confirm a causal relationship, targeting sleep fragmentation and astrocyte activation may benefit cognition in older adults.