Sleep fragmentation, microglial aging, and cognitive impairment in adults with and without Alzheimer’s dementia

Study participants

We studied 685 adults (>65 years old; 265 with Alzheimer’s dementia at the last available assessment and 420 without) participating in two cohort studies of older persons—the MAP and the ROS. Clinical and postmortem characteristics of the participants are summarized in Table 1, and the overlap between subsets of these participants is shown in fig. S1A.

Sleep fragmentation is associated with aging and activation of microglia

We quantified antemortem sleep fragmentation by actigraphy and postmortem dorsolateral prefrontal cortex gene expression by RNA sequencing in 152 MAP participants. We considered the association between antemortem sleep fragmentation and the expression of sets of microglial marker genes from three published sources: the HuMi_Aged gene set (10) and the Galatro gene set (9) derived from postmortem human brain tissue, and the NeuroExpresso cortical microglial gene set derived from rodents (11). We considered gene expression at the individual gene level. We also computed a summary z score by taking the average normalized gene expression across all genes in that set. Despite the incomplete overlap between gene sets (fig. S1B), their composite expression levels were highly correlated (Pearson R = 0.95 to 0.98).

We first examined the HuMi_Aged gene set. In linear regression models adjusted for age, sex, education, time between last actigraphy and death, postmortem interval, RNA quality (RIN) score, and proportion of ribosomal bases, the expression levels of 352 HuMi_Aged genes were associated with sleep fragmentation at an unadjusted threshold of P = 0.05 (Fig. 1A and table S1). Of these, 279 showed positive associations such that greater sleep fragmentation was associated with higher expression, while 73 showed negative associations such that greater sleep fragmentation was associated with lower expression. Eight were associated with sleep fragmentation at a Bonferroni adjusted threshold of P = 0.00005 (ARID5A, CISH, TNFRSF18, SELL, IRF7, PLAUR, SLC11A1, and RPS19); all were expressed at higher levels with greater sleep fragmentation. Next, to capture the collective change in expression of the HuMi_Aged gene set at the bulk tissue level in each participant, we generated a composite measure of microglial marker gene expression by taking the mean normalized expression of all genes in the HuMi_Aged set as described previously (10). In a linear regression model adjusted for age, sex, education, and the same technical covariates, greater antemortem sleep fragmentation was associated with higher composite expression of HuMi_Aged microglial genes (estimate = +0.066 standard units of expression per 0.01 unit difference in k_RA; SE = 0.027; P = 0.014; Fig. 1B). Each 0.01 unit increase in k_RA, representing roughly 1 SD, was associated with an effect on the composite expression of microglial marker genes (HuMi_Aged gene set) equivalent to 4.5 years of aging [95% confidence interval (CI), 0.9 to 10.5),

To ensure that these results were not specific to the HuMi_Aged gene set, we repeated these analyses using the Galatro and NeuroExpresso gene sets. Despite incomplete overlap between these gene sets (fig. S1B), results were similar (fig. S2). Similar results were seen when we repeated these analyses considering only those genes that were shared between all three gene sets (fig. S3, A and B) and those genes that were unique to each gene set (fig. S3, C to H), supporting the robustness of these findings. Sleep-wake fragmentation can accompany circadian rhythm dysfunction. Therefore, we repeated the above analyses using an actigraphically derived nonparametric metric of circadian regularity, interdaily stability, that has previously been shown to be abnormal in AD (12). Unlike parametric cosinor-based methods, this does not assume that the activity rhythm conforms to any particular shape. We found no association between interdaily stability and the expression of microglial marker genes (fig. S4), suggesting that the association between sleep fragmentation and microglial marker gene expression is not primarily driven by differences in circadian rhythmicity. To examine the specificity of these findings, we repeated these analyses using astrocyte marker genes from the NeuroExpresso gene set (fig. S5A and table S1). No astrocyte marker genes were associated with sleep fragmentation after Bonferroni correction, and there was no statistically significant association between sleep fragmentation and composite expression of astrocytic marker genes (estimate = +0.021 standard units of gene expression per 0.01 unit difference in k_RA; SE = 0.040; P = 0.61; fig. S5B).

The transcriptional phenotype of human microglia changes with age (10). To examine for a differential effect of sleep on age-related microglial transcriptional programs, we repeated the above analyses considering separately those HuMi_Aged microglial marker genes previously identified as being enriched in aged or in young microglia (10, 13). The composite expression of genes characteristic of aged microglia (HuMi_Aged_Old gene set) was positively correlated with age (estimate = +0.009 standard units of gene expression per 1 year of age; SE = 0.003; P = 0.002). Of the 117 genes identified as being enriched in aged microglia, 115 were expressed at higher levels with greater sleep fragmentation at an uncorrected P < 0.05, and 7 (ARID5A, IRF7, SLC11A1, RPS19, CYTH4, HAMP, and RHBDF2) were associated with sleep fragmentation after Bonferroni correction (Fig. 1C; table S1), independent of chronological age at death. Of these, three (CYTH4, HAMP, and RHBDF2) were not significant after Bonferroni correction in our primary analyses of the HuMi_Aged gene set. The reason for this is that there are fewer genes in the HuMi_Aged_Old set, resulting in a less stringent Bonferroni cutoff. In addition to these gene-level findings, composite expression of HuMi_Aged_Old genes was strongly associated with sleep fragmentation (estimate = +0.161 standard units of gene expression per 0.01 unit difference in k_RA; SE = 0.041; P = 0.00014; Fig. 1D). To contextualize this, each 0.01 unit greater k_RA, corresponding to roughly 1 SD, had an effect on the expression of aged microglial marker genes (HuMi_Aged_Old gene set) equivalent to 7.7 additional years (95% CI, 4.0 to 16.1) of chronological age, suggesting that sleep fragmentation is a better predictor of microglial transcriptional aging than chronological age. Conversely, of the 26 genes previously identified as being associated with young microglia (HuMi_Aged_Young gene set), and meeting quality control criteria in our study, more were expressed at lower levels with greater sleep fragmentation (16 of 26) than were expressed at higher levels with greater sleep fragmentation (10 of 26) (Fig. 2E; table S1), and there was no significant relationship between the composite expression of genes characteristic of young microglia and sleep fragmentation (estimate = +0.015 standard units of expression per 0.01 unit difference in k_RA; SE = 0.038; P = 0.69; Fig. 2F). Health behaviors such as smoking and alcohol consumption may plausibly influence sleep and microglial biology. However, the association between sleep fragmentation and the composite expression of genes characteristic of aged microglia remained significant in models controlling for history of smoking and consumption of alcohol.

A number of brain pathologies, including AD pathology, Lewy body pathology (characteristics of Parkinson’s disease), and infarcts, are associated with sleep fragmentation (14–16). However, the association between greater sleep fragmentation and higher composite expression of genes characteristic of aged microglia remained significant in models controlling for the burden of AD pathology, Lewy body pathology, macroinfarcts, microinfarcts, TDP-43 pathology (characteristics of frontotemporal dementia), or hippocampal sclerosis, alone or in combination (estimate = +0.17 standard units of expression per 0.01 unit difference in k_RA; SE = 0.04; P = 0.0001; table S2, model H). Moreover, the association between sleep fragmentation and the expression of genes characteristic of aged microglia (HuMi_Aged_Old gene set) was not significantly different between individuals with and without AD pathology as defined by an National Institute on Aging (NIA)–Reagan classification of intermediate or high (interaction P = 0.82; Fig. 2). Thus, greater sleep fragmentation is associated with higher expression of genes characteristic of aged microglia independent of the presence of AD and other dementia-associated brain pathologies.

Sleep disruption is associated with microglial activation in model organisms (7, 8). We related antemortem sleep fragmentation to neocortical microglial density assessed by immunohistochemistry in 156 MAP participants with quantification of both. In linear regression models adjusted for age at death, sex, education, time between last actigraphy and death, and postmortem interval, we found no significant association between greater sleep fragmentation and the total density of microglia (P = 0.66; Fig. 3A). Next, we examined separately microglia in different stages of morphologic activation: stage I (resting), stage II (activated), and stage III (phagocytic) (fig. S6). We observed a greater proportion of morphologically activated (stages II and III) microglia in association with greater sleep fragmentation (estimate = +0.055, SE = 0.026, P = 0.034; Fig. 3B). This suggests that sleep fragmentation is associated with morphological microglial activation.

We next explored the relationship between composite expression of genes characteristic of aged microglia and morphologic microglial activation quantified by immunohistochemistry. To do so, we analyzed data from the 96 MAP participants who had antemortem quantification of sleep fragmentation and assessment of microglia by both immunohistochemistry and RNA sequencing. In a linear regression model adjusted for age, sex, education, time from last actigraphy and death, postmortem interval of autopsy, RNA RIN score, and proportion of ribosomal bases, the association between sleep fragmentation and the composite expression of aging-associated microglial genes (HuMi_Aged_Old gene set) remained significant (estimate = +0.15 standard units of expression per 1 SD difference in k_RA; SE = 0.05; P = 0.0023; table S3). The overall density of microglia was not associated with composite expression of genes characteristic of aged microglia. In contrast, the proportion of activated microglia was associated with the expression of genes characteristic of aged microglia (Spearman R = 0.27, P = 0.009), although this association was somewhat attenuated when controlling for age, sex, and other covariates as above (estimate = +0.08 standard units of expression per 1 SD difference in the proportion of morphologically activated microglia; SE = 0.05, P = 0.12; table S3). The association between sleep fragmentation and the composite expression of genes characteristic of aged microglia was minimally attenuated when controlling for the overall density of microglia or for the proportion of morphologically activated microglia (estimate = +0.14, SE = 0.05, P = 0.004; table S3, model D), suggesting that the association between sleep fragmentation and an aged microglial transcriptional phenotype is not solely due to morphologically defined microglial activation. By contrast, in a model with the proportion of morphologically activated microglia as the outcome, when we controlled for the composite expression of genes characteristic of aged microglia, the effect of sleep fragmentation was attenuated (table S3, model G), which is consistent with a scenario where the association between sleep fragmentation and microglial morphologic activation is mediated in part by a shift toward an aged microglial transcriptional phenotype (fig. S7).

Microglial transcriptional aging and activation are associated with cognitive impairment

We next examined whether higher expression of genes characteristic of aged microglia was associated with cognition proximate to death. To do so, we considered data from the 621 ROS and MAP participants with available dorsolateral prefrontal cortex (DLPFC) RNA sequencing and at least one cognitive assessment. We found that higher composite expression of genes characteristics of aged microglia from the HuMi_Aged_Old gene set was associated with poorer composite global cognition proximate to death, adjusted for age, sex, education, time from last clinical assessment to death, postmortem interval, RNA RIN score, and proportion of ribosomal bases (estimate = −0.254 standard units of composite global cognition per 1 standard unit of gene expression; SE = 0.097; P = 0.0095; Fig. 4A). To contextualize this, each SD greater expression of genes characteristic of aged microglia was equivalent to an additional 3.4 years of chronological age (95% CI, 0.8 to 7.3) in its association with cognition. This association remained significant when controlling for dementia-associated brain pathologies, including the burden of AD pathology, Lewy body pathology, macroinfarcts, microinfarcts, TDP-43 pathology, or hippocampal sclerosis, alone or in combination (P < 0.05; table S4), indicating that the association between the expression of genes characteristic of aged microglia and cognition is not solely accounted for by the presence of these brain pathologies. Moreover, this association remained significant when controlling for source cohort (MAP versus ROS) and did not differ between source cohorts (interaction, P = 0.65).

We then examined whether a greater proportion of morphologically defined activated microglia was associated with differences in cognition. To do so, we considered data from the 224 ROS and MAP participants with quantification of microglial activation by immunohistochemistry. In a linear model adjusted for age, sex, education, and technical covariates, a greater proportion of activated microglia was associated with poorer composite global cognition proximate to death (estimate = −0.72, SE = 0.21, P = 0.00066; Fig. 4B).

Next, we examined the relationship between sleep fragmentation, cognitive performance, dementia-related brain pathologies, and microglial aging and activation. In concordance with previous results from our laboratory and others (14, 17), in 480 deceased participants with actigraphy, greater sleep fragmentation was associated with poorer composite global cognition proximate to death (estimate = −0.23 standard units of cognition per 0.01 unit difference in k_RA; SE = 0.06; P = 0.0004; Fig. 4C).

Similar results were seen when the analysis was limited to those participants with RNA sequencing data (n = 152; table S4, model I) or with immunohistochemical quantification of microglial density (n = 156; table S4, model M). These associations were partially attenuated in models controlling for the composite expression of genes characteristic of aged microglia (table S4, model J) or for the proportion of morphologically activated microglia (table S4, model N). They were also partially attenuated in models controlling for dementia-associated brain pathologies (AD pathology, Lewy body disease, macroinfarcts, microinfarcts, TDP-43 pathology, and hippocampal sclerosis; table S4, models K and O). In models controlling for both the dementia-associated brain pathologies and for the microglial measures (table S4, models L and P), the association between sleep fragmentation and composite global cognition proximate to death was no longer significant. This is compatible with a scenario in which microglial transcriptional aging or morphological activation partially mediates the association between sleep fragmentation and poor cognition, independent of the presence of dementia-related brain pathologies.

Last, we examined to what extent these relationships differed between those with and without AD pathology. Greater sleep fragmentation was associated with worse cognitive performance in individuals with or without Alzheimer’ disease pathology (interaction P = 0.90; Fig. 5, A and B). However, a pathological diagnosis of AD was a significant modifier of the association between the expression of genes characteristic of aged microglia and cognition (interaction, P = 0.037). Whereas greater microglial gene expression was strongly associated with cognition in those with a pathological diagnosis of AD, this effect was attenuated in the absence of AD pathology (Fig. 5, C and D). Considered another way, high levels of expression of genes characteristic of aged microglia amplified the effect of AD pathology on cognition, with a 1.5-fold greater effect of AD pathology in those in the highest versus lowest quartile of microglial gene expression (Fig. 5, E and F).