Introduction

Microglia are one of few immune cells with long term residence in the brain parenchyma¹. Microglia support neural development by synaptic pruning and neurogenic function in the developing brain, followed by homeostatic maintenance in the adult brain². Alternatively, in the aged brain microglia exhibit dystrophic phenotypes, aggravating perceived insults rather than responding productively to environmental signals³. While microglia have been identified as principal mediators of multiple age-related neurodegenerative pathologies⁴, mechanisms leading to microglia dysfunction during aging remain obscure.

Numerous studies have indicated that aged microglia are inflamed⁵, have reduced phagocytic capacity^6,7, and have decreased motility⁸. Microglia exhibit several hallmarks of aging that potentially contribute to their age-related dysfunction, such as shortened telomeres⁹, altered intercellular communication¹⁰, molecular alterations¹¹, and a loss of proteostasis¹². Furthermore, many recent studies have started to reveal the molecular changes that define microglial aging. Single cell RNA-Seq (scRNA-Seq) analyses indicate that microglia isolated from the entire brain lose homeostasis and activate inflammatory transcriptional profiles with age^13,14. Data rich studies have also revealed partial overlaps between aging microglia and those from disease models, including Alzheimer’s disease^15–19. Interestingly, the microglial response to aging shows regional variation, as white matter-rich regions induce an interferon signature in aged microglia²⁰. Studies using aged plasma administration and heterochronic blood exchange demonstrate that microglia aging is in part driven by the aged systemic environment ^21,22. However, the genesis of age-related dystrophic microglial phenotypes has not been extensively investigated. So, we set out to characterize the progression of age-related hippocampal microglial dysfunction, aiming to uncover intermediate states that could be intrinsic to the aging process. To do so, we undertook complementary cellular and molecular analyses of microglia across the adult lifespan and in heterochronic parabiosis - an experimental model of aging in which the circulatory systems of young adult and aged animals are joined²³.

In this study, we report that microglia in the adult mouse hippocampus, a brain region responsible for learning and memory and susceptible to age-related cognitive decline, advance through intermediate states that drive inflammatory activation during aging. We utilize scRNA-Seq across the adult lifespan to identify intermediate transcriptional states of microglial aging that emerge following exposure to an aged systemic environment. Functionally, we tested the role of these intermediate states using in vitro microglia approaches and an in vivo temporally controlled adult microglia-specific Tgfb1 conditional genetic knockout mouse model to demonstrate that intermediates represent checkpoints in the progression of microglia from homeostasis to inflammatory activation, with functional implications for hippocampal-dependent cognitive decline.

Results

Single-cell transcriptional analysis reveals the dynamics of hippocampal microglia aging

Previously, single-cell transcriptional profiling of microglia uncovered a diversity of responses during development and in response to pathology aggregated over multiple brain regions^{13,14,16,19,24,25}. Microglia have region-specific transcriptional states, so we decided to focus our analyses on the hippocampus¹¹, a region with well-reported functional differences during aging²⁶. Correspondingly, we investigated age-related transcriptional heterogeneity by performing scRNA-Seq on hippocampal CD11b-positive cells isolated from mice at 6, 12, 18, and 24 months of age using a commonly employed pooling strategy^18,27,28 (n = 1 pool of 5 biological replicates for each age) to reveal the dynamics of microglia aging. 82% of cells isolated with CD11b were identified as microglia using canonical markers with a small subset being proliferative (Figure S1A-C). Dimensionality reduction revealed that non-proliferative microglia were grouped into interconnected clusters that contained homeostatic, transition, and activated states, along with a semi-distinct cluster with interferon activation (Figure 1a, Figure S1D-G). Overlaying microglia ages onto the clusters revealed progression from a homeostatic state in younger microglia to an activated state in old microglia, suggesting that gradual transcriptional changes occur during microglial aging (Figure 1B,C, Figure S1H). Furthermore, the heterogeneity of microglial transcriptional programs increased with age (Figure S1I). Expression of Cx3cr1 and Itgam exemplified the homeostatic state, while B2m, Apoe, Cd48, and Lyz2 depict the activated state (Figure 1D). Interestingly, several immediate early genes (Jun, Klf2, Fos), as well as the purinergic receptor gene P2ry12, exhibited transient increases in expression at middle-age (Figure 1D). Alternatively, the interferon cluster represents a small proportion of the overall microglia population with little age-related changes in the number of these cells, but augmented expression of interferon genes during aging (Figure 1A,C,D, Figure S1G). Thus, gene expression changes exhibited unique patterns during aging, with certain genes showing age-related decreases (e.g. Tgfbr1) or increases (e.g. Apoe), while a subset of genes (e.g. Tgfb1) have peaks of expression at middle age (Figure 1D,E,G; Table S1). Several complement genes implicated in age-related synaptic loss^10,29(e.g. C1q and C3) demonstrated progressive age-related transcriptional increases, and the increased complement activity in microglia was confirmed by immunohistochemistry (Figure 1E,F).

Fig. 1

Fig. S1

The effects of aging on microglia could be progressive or sudden, so we investigated if we could find evidence of early aging of hippocampal microglia. We observe relatively few differentially genes between 6- and 12-month (Figure 1G,H, Table S1). The number of differentially expressed genes gradually increased through the 24-month time point (Figure 1G-J, Table S1), suggesting that hippocampal microglia progressively accumulate aged-related expression changes. Furthermore, when investigating the outcome for genes with significant changes for each age, we find that genes that are differentially expressed at 12 months lose their expression change, while genes that change at later ages show progressive expression differences (Figure 1H). GO analysis of the microglia from 6- and 12-month-old mice revealed that activation and translational programs are modestly enriched in genes with increased expression during early aging (Figure 1I). The latter stages of aging display more prominent enrichment of immune activation and translation processes in genes with increased expression, while revealing that genes that regulate transcriptional processes exhibit decreased expression (Figure 1J,K). We corroborated the progressive increase in microglial activation by immunohistochemical analysis for CD68 and p65 (Figure S1J,K). Thus, microglia progressively age with subsets of cells at each age group being forerunners towards inflammatory activation or refractory to aging by remaining homeostatic. This analysis further suggests that microglia could pass through transitory states during aging.

Trajectory analysis indicates that aging microglia pass through intermediate states

To determine the transcriptional progression of microglia aging towards an inflammatory state, we performed pseudotime analysis of the scRNA-Seq data and observed several trajectories (Figure 2A). We find that the pseudotime trajectories progressed along the biological ages of the mice, indicating the suitability of this analysis for modeling aging progression in microglia (Figure 2B). We focus our analysis on one trajectory with two semi-distinct subbranches that terminated in inflammatory activation, since the alternate branch had minimal transcriptional differences between the beginning and end of the trajectory (Figure 2B,C). To gain insight into the transcriptional states of microglial aging along this inflammatory trajectory, we used spatial autocorrelation analysis and identified five expression modules of coregulated genes (Figure 2B,C, Table S2). We named these modules according to the most enriched GO pathways and known roles of prominent genes for microglia. This trajectory proceeds from high expression of homeostatic genes and mitochondrial processes (module 1) to transient expression of stress response genes and Tgfb1 (module 2), of which TGFβ signaling is critical for microglia development³⁰. A concerted increase in expression of ribosomal genes (module 3) precedes one subbranch of inflammatory activation (module 4), while the other myeloid activation subbranch (module 5), characterized by B2m and C1qc expression, proceeds independently of increased ribosomal gene expression. GO analysis of genes specifically upregulated at 12-months showed an enrichment of stress response pathways (Figure S2A,B), corroborating the findings of the pseudotime analysis that intermediate stages of microglial aging pass through stress response pathways. Thus, pseudotime trajectory analysis identifies a progression of intermediate states of microglial aging with potential involvement in age-related maladaptive inflammation.

Fig. 2

Fig. S2

The gene level dynamics of microglia aging are further revealed when interrogating the modules at each timepoint. Expression of several genes in the homeostatic cluster, including Cx3cr1 and Tgfbr1, progressively decline throughout the aging timeline (Figure 2D). Stress response gene expression peaks at early timepoints during aging, while those genes in the translation module have coherent upregulation at 24-months of age (Figure 2D). The top ranked genes in the inflammation module gradually increase their expression throughout aging in a consistent manner (Figure 2D). Alternatively, several genes in the myeloid activation module have expression peaks at 12-(Cst3, Selplg) or 18-months (Ctss, Cd9) of age (Figure 2D). When the modules are collapsed into metagenes, we observe that homeostatic module is reduced early on in aging, while the stress response module follows later (Figure 2E-G). The ribosomal and inflammatory modules display augmented expression at latter ages (Figure 2E-G). Alternatively, the myeloid activation consistently has slightly increased expression throughout aging. These results suggest that with advancing age microglia lose homeostatic and stress response functions, while increasing their translational and inflammatory capacity.

Next, we utilized a publicly available scRNA-Seq datasets of pro-aging systemic interventions and age-related neurodegenerative disease models to determine their impact on hippocampal microglia aging trajectories. We used the heterochronic parabiosis model³¹ to investigate whether exposure to an aged systemic environment could promote progression of young adult microglia along the identified aging trajectory (Figure 2H). First, using immunohistochemistry for IBA1 and CD68, we reveal that an aged systemic environment activates microglia (Figure S2C). Next, we find that hippocampal microglia from young adult heterochronic parabionts have decreased expression of homeostatic genes (module 1), while having increased expression of ribosomal genes (modules 3), inflammatory activation genes (module 4) and myeloid activation genes (module 5) (Figure 2I, Figure S2D). We used the App^NL-G-F transgenic mouse model at 12 months of age to determine the effects of Alzheimer’s disease pathology on microglial aging¹⁷ (Figure 2J). We observe exaggerated advancement along the aging trajectory in App^NL-G-F mice compared to controls, as expression is shifted towards the age-associated modules (Figure 2K, Figure S2E,F). These shift in gene expression posit that the aged systemic and diseased environments drive adult microglia advancement along an aging-associated inflammatory trajectory.

Intermediate states of microglia aging act as checkpoints on inflammatory progression

Next, we sought to corroborate microglial age-related molecular changes associated with individual intermediate states in the aging hippocampus by RNAscope and immunohistochemical analysis. Specifically, we assessed changes in the intermediate stress response (module 2) and translation (module 3) modules by examining expression of Tgfb1 (Figure 3A,B) and the ribosomal protein S6 (Figure 3G,H), respectively. The vast majority (>95%) of Tgfb1 signal localized to microglia at every age (Figure S3A), indicating microglia are the predominant source for hippocampal TGFB1. Consistent with scRNA-Seq data of aging microglia (Figure 1E), we observed highest Tgfb1 expression in microglia by middle-age (Figure 3B). TGFBR1 activation was reduced during aging in hippocampal microglia (Figure S3B), and genes involved in the TGFβ signaling pathway exhibited age-related expression changes with several having peak expression at 12 months of age (Figure S3C). Additionally, we observed an increase in translational components found during the scRNA-Seq analysis by immunohistochemistry for the ribosomal protein S6 (Figure 3H)³².

Fig. 3

Fig. S3

To complement hippocampal aging analysis, we assessed the impact of the aging systemic milieu on both the TGFβ pathway and translational components in the young adult hippocampus following heterochronic parabiosis. We observed a decrease in microglia Tgfb1 and TGFβ pathway expression in young adult heterochronic parabionts compared to age-matched young adult isochronic controls³¹ (Figure 3C, Figure S3D). Additionally, expression of the translation module in microglia increases following exposure to an aged systemic environment in the heterochronic parabiosis model (Figure 3I).

To investigate the role of the intermediate stress response (module 2) and translation (module 3) modules in mediating advancement through microglial activation states, we used an in vitro approach. Primary microglia were treated with LPS, which induced gene expression changes with a significant overlap to aging, as well as expression changes in aging modules genes (Figure S3E,F; Table S3). We administered TGFB1 to activate TGFβ signaling (module 2) (Figure 3A) and CX-5461 to inhibit RNA Pol I synthesis and interfere with translation (module 3) (Figure 3G) in the context of microglial activation. TGFB1 treatment modified the transcriptional states of LPS-treated microglia (Figure S3G) and caused decreased expression of inflammatory immune processes after LPS stimulation (Figure 3D). TGFB1 treatment reduced expression of later aging modules (modules 2, 3 and 4) and restored expression of the homeostatic gene Cx3cr1 (Figure 3E,F), suggesting that TGFBβ signaling acts as an aging checkpoint during microglial stress response. Targeting translation (module 3) with CX-5461 attenuated expression of immune processes after LPS stimulation (Figure 3J; Table S3), as well as decreased expression of genes in modules 2, 3, and 4, while restoring expression of homeostatic genes in module 1 (Figure 3K,L). Interestingly, CX-5461 treatment had no effect on genes in module 5, which is located on an independent myeloid activation trajectory from altered translational gene expression (Figure 3K,L). These in vitro perturbation data identify active roles for intermediate states in mediating advancement along aging-associated inflammatory trajectories.

Mimicking age-related changes in microglia-derived TGFB1 promotes microglial advancement along an aging-associated inflammatory trajectory in vivo

Having uncovered roles for intermediate states in advancing microglia along aging-associated inflammatory trajectories in vitro, we next investigated the role of individual aging modules in adult microglia in vivo. We elected to disrupt the transition from homeostasis to inflammatory activation by targeting microglia-derived TGFB1, a key marker of the stress response intermediate state (module 2). While TGFB1 is critical for microglia development³³, the role of microglia-derived TGFB1 in regulating microglia homeostasis during aging remains to be defined.

We generated Tgfb1^flox/flox, Tgfb1^flox/wt, and Tgfb1^wt/wt mice carrying an inducible Cx3cr1-Cre-ER gene, in which Tgfb1 is excised specifically in adult microglia upon tamoxifen administration (Tgfb1 cKO, Tgfb1 Het, and WT, respectively) (Figure 4A). To disrupt the age-related increase in microglia Tgfb1 observed between mature and middle-age (Figure 1E and Figure 3B), mature adult mice were administered tamoxifen and molecular changes were assessed two months later (Figure 4A). We find that microglia activation, as measured by CD68/IBA1 immunohistochemistry, exhibited genotype-dependent effects (Figure S4A). We performed scRNA-Seq on microglia isolated from WT, Tgfb1 Het and Tgfb1 cKO mature mice (n = 2 pools of 3 animals per genotype), and detected genotype-dependent changes in Tgfb1expression levels, as well as effectors of TGFB1 signaling (Figure S4B). Dimensionality reduction and clustering of cells from scRNA-Seq revealed readily identifiable clusters based on marker expression (Figure 4B). Composition of these clusters were dependent on the microglia genotype, with WT microglia representing the largest fraction of the homeostatic cluster and Tgfb1 cKO microglia representing the largest fraction of the activated cluster (Figure 4C, Figure S4C). Next, we assessed whether deletion of Tgfb1 in mature adult microglia impacted advancement along aging-associated inflammatory trajectories. Interestingly, when we analyze the effects of Tgfb1 dosage on the aging modules, we find that the loss of a single Tgfb1 allele reduces expression of homeostatic genes (module 1), while increasing expression of translation-related genes (module 3) (Figure 4D,E). In addition to a decrease in expression of homeostatic genes (module 1), loss of both Tgfb1 alleles caused a greater increase in expression of translation-related genes (module 3), and led to a profound increase in inflammatory activation genes (module 4) (Figure 4D,F). Results were corroborated in an independent cohort of mature Tgfb1 cKO mice, in which we detected altered levels of immune- and age-related cell surface markers (Figure S4D,E), transcriptional alterations that significantly overlapped with aging (Figure S4F-H; Table S4), and progression along the aging-associated inflammatory trajectories (Figure S4I,J). These data indicate that microglia-derived TGFB1 is both necessary for expression of youth-associated homeostatic genes and prevents aberrant microglia inflammatory activation, further suggesting that TGFBβ signaling acts as a checkpoint in mediating advancement along aging-associated inflammatory trajectories.

Fig. 4

Fig. S4

Mimicking age-related changes in microglia-derived TGFB1 impairs hippocampal-dependent cognitive function

It is becoming increasingly apparent that microglia interact with other cell types in the brain to influence cognitive function^2,34. Given loss of microglia-derived TGFB1 promoted microglia progression along aging-associated inflammatory trajectories, we next investigated its functional consequence on cognition. We assessed hippocampal-dependent learning and memory in WT, Tgfb1 Het and Tgfb1 cKO mature mice using novel object recognition and contextual fear conditioning - behavioral paradigms that are sensitive to age-related impairments^35,36. During novel object recognition testing, WT mice were biased toward a novel object relative to a familiar object while neither Tgfb1 Het or Tgfb1 cKO mice showed any preference (Figure 4G). During contextual fear conditioning testing, Tgfb1 Het and Tgfb1 cKO mice exhibited less freezing compared to WT controls (Figure 4H). Alternatively, we assessed amygdala-dependent fear memory by cued fear conditioning (Figure S4K)³⁷ and short-term spatial reference memory by Y maze (Figure S4L)³⁸, and observed no differences across genotypes. As a control, we also profiled general health using an open field paradigm and observed no differences in total distance traveled or time spent in the center of the open field, indicative of normal motor and anxiety functions (Figure S4M). These behavioral data indicate that loss of microglia-derived TGFB1 not only impacts aging-associated inflammatory trajectories but also promote hippocampal-dependent cognitive impairments characteristic of age-related functional decline.

Discussion

Cumulatively, we demonstrate that microglia aging in the hippocampus advances through intermediate states that precede inflammatory activation, with functional implications for age-related cognitive decline. These transient states, such as activation of the stress response and TGFβ signaling, act as checkpoints on further age-related activation. Furthermore, microglia asynchronously progress through this inflammatory aging trajectory that is in turn influenced by the systemic environment and neurodegenerative disease states.

Analysis of our scRNA-Seq data elucidate cryptic, but necessary, intermediate states of microglia progression to age-related inflammatory activation. These states are present in middle-aged mice and constitute stress response pathways likely responsive to internal and external cellular insults^39,40. Pseudotime analysis suggests a highly plausible sequence of molecular alterations that lead to age-related inflammatory activation of hippocampal microglia. First, mitochondrial dysfunction triggers a stress response, and products of this stress response (e.g. TGFB1) attempt to return microglia to homeostasis⁴¹. Here, chronic activation of the stress response leads to subsequent increased translational capacity that promotes inflammatory activation⁴². Alternatively, another branch of microglial activation advances independently of increased translational capacity, evidenced by translational inhibition having little effect on its activation following LPS stimulation. Thus, we find that increases in translational capacity predicate age-related inflammatory phenotypes, identifying potential targets to counter microglial aging. Analysis of scRNA-Seq data further identify intermediate states of microglial aging in young adult heterochronic parabionts, suggesting a pivotal role for the aged systemic environment in driving advancement towards age-related inflammatory activation.

While we find evidence of the beginnings of inflammatory activation in the hippocampus by middle age (12 months), consistent with the timing recently reported in the whole brain¹⁴, it should be noted that scRNA-Seq analysis of the whole brain did not observe intermediate stages in microglial aging¹⁴. These disparate observations could likely be the result of analyzing multiple regions with highly differentiated transcriptional programs¹¹ that could obfuscate regional aging trajectories. This further underscores the need for both temporal and regional specificity in aging studies, given the level of cellular heterogeneity observed across ages and brain regions. Moreover, studying microglia - or other cell types - at early stages of aging in a region-specific manner could also represent useful models to investigate the ontogeny of aging without the confounding effects of accumulated functional deficits.

Of note, we find that an inflammatory trajectory of microglia aging can be mitigated in vitro by environmental TGFβ signaling, while loss of microglia-derived TGFB1 promoted progression along aging-associated inflammatory trajectories in vivo. Together, these data indicate that TGFβ signaling is necessary for maintaining youth-associated homeostasis and represents a checkpoint in mediating advancement along aging-associated inflammatory trajectories. Functionally, loss of microglia-derived TGFB1 resulted in hippocampal-dependent cognitive impairments, suggesting that advancement through aging-associated inflammatory trajectories may facilitate age-related cognitive decline. TGFβ likely acts in an autocrine fashion to return microglia to a homeostatic state after encountering stressful stimuli. Given that a transient increase in Tgfb1 expression is observed in middle-aged microglia, it is possible that maintaining stress response-mediated signaling into older ages may also provide a means to delay age-related microglial dysfunction and maintain cognitive function. Alternatively, TGFβ signaling impairs microglial responses to Alzheimer’s disease pathology⁴³, suggesting that the dynamic regulation of TGFβ signaling in response to aging-associated pathologies is necessary for tuning microglial responses.

While microglia have been demonstrated to mediate the establishment of cognitive function during development² and play prominent roles in neurodegeneration⁴⁴, little is known about how age-related changes in microglia affect cognitive function. The specificity of the cognitive deficits that we observe in Tgfb1 Het and Tgfb1 cKO mice suggests that age-related changes in microglia-derived TGFB1 predominantly impacts hippocampal-dependent cognitive decline. Furthermore, the deficits observed in Tgfb1 heterozygous mice point to the importance of maintaining youthful levels of microglia-derived TGFB1 to counteract stressors that accrue during aging.

Altogether, our results highlight the need to investigate aging across the lifespan in a region-specific manner to reveal necessary cellular intermediates in the aging process. While the present study focuses on age-related changes in microglia, our findings have broad implications for other studies in determining the extent to which aging sequelae are conserved across different cell types in the hippocampus. Furthermore, therapeutic interventions targeting intermediate states in the aging process hold far reaching potential to counteract the development of age-related pathologies, including those promoting cognitive dysfunction.

Acknowledgements

We thank Dr. Erik Ullian for critically reading manuscript. Funding: This work was funded by National Institute on Aging (AG055292 (J.M.S.), AG055797 (S.A.V.), AG077816 (S.A.V.)), Simons Foundation (S.A.V). The authors declare no competing financial interests.

Author contributions

J.M.S and S.A.V. developed concept and designed experiments. J.M.S conducted experiments, collected data, and performed analysis. J.M.S and S.A.V wrote manuscript. S.A.V supervised all aspects of the project. All authors had the opportunity to discuss results and comment on manuscript.

Competing interest statement

The authors declare no competing interests.

Methods

Mice

All animal procedures were performed in accordance with protocols approved by the UCSF IACUC. Animals were housed in SPF barrier facilities and provided continuous food and water along with environmental enrichment. Aging characterizations were performed on an inhouse C57BL/6J mouse colony where 2-month-old mice were purchased from Jackson Laboratories (Stock 000664) and aged in the UCSF Parnassus barrier facility. Timed pregnant females were obtained for isolation of primary microglia from newborn pups. Male mice were used in all experiments. B6.129P2(Cg)-Cx3cr1^{tm2.1(cre/ERT2)Litt}/WganJ (Cx3cr1Cre^ER) (Stock 021160) and C57BL/6J-Tgfb1^em2Lutzy/Mmjax (Tgfb1-flox) (Stock 65809-JAX) mice were obtained from Jackson Laboratories. Mice were bred to obtain Cx3cr1Cre^ER +/wt with either Tgfb1 wt/wt, Tgfb1 fl/fl. As Cx3cr1Cre^ER is a knock-in/knock-out allele for the microglia homeostatic gene Cx3cr1, all mice tested were Cx3cr1Cre^ER +/wt. To induce recombination and deletion of Tgfb1 by Cre^ER, we injected tamoxifen for 5 days at 90mg/kg daily, and mice were analyzed after 60 days.

Parabiosis

Parabiosis surgery followed previously described procedures⁴⁵. Mirror-image skin incisions at the left and right flanks were made through the skin, and shorter incisions were made through the peritoneum. The peritoneal openings of the adjacent parabionts were sutured together with chromic gut suture (MYCO Medical, GC635-BRC). Apposing elbow and knee joints from each parabiont were sutured together (Coated VICRYL Suture, Ethicon, J386) and the skin of each mouse was stapled (9-mm Autoclip, Clay Adams, 427631) to the skin of the adjacent parabiont. Each mouse was injected subcutaneously with Carpofen, Enrofloxacin, and Buprenex as directed for pain and monitored during recovery. For overall health and maintenance behavior, several recovery characteristics were analyzed at various times after surgery, including pair weights and grooming behavior.

Bulk Microglia RNA-Seq

Microglia were isolated from C57BL6/J, Tgfb1 cHet, and Tgfb1 cKO using FACS. Briefly, mice were sedated with ketamine followed by perfusion with 30mL of ice-cold PBS. The entire brain was removed, then the hippocampus was sub-dissected. Single-cell suspensions were generated by enzyme-mediated (papain) and mechanical dissociation using Miltenyi Neural Dissociation Kits (P) (Miltenyi, 130-092-628) according to the manufacturer’s instructions. Myelin was depleted from the suspensions using Myelin Removal Beads (Miltenyi, 130-096-731). Cells were labeled with CD11b-APC (eBioscience, 17-0112-82, RRID:AB_469343) and CD45-PE (eBioscience, 12-0451-82, RRID:AB_465668), then sorted at 4°C into Tri Reagent with a FACS AriaII (BD). The gating strategy for microglia isolation is presented in Figure S1D, showing the collection of Cd11b⁺Cd45^Intermediate cells. During the collection of control and Tgfb1 microglia, CD48 was detected in Cd11b⁺Cd45^Intermediate cells using CD48-Pacific Blue (BioLegend, 103417, RRID:AB_756139). FlowJo was used to analyze flow cytometry data. RNA was isolated from microglia using Tri Reagent (Sigma-Aldrich, T9424). RNA pellets were resuspended in 5uL TE buffer.

RNA was transformed into RNA-Seq libraries using an adapted version of the Smart-Seq2 protocol⁴⁶. Briefly, 5ng of total RNA was reverse transcribed with SuperScript II (Thermo Fisher, 18064014) supplemented with betaine and MgCl₂ using an oligo-dT RT primer (AAGCAGTGGTATCAACGCAGAGTACT(30)VN) with a PCR binding site and a template switching oligonucleotide (5′ AAGCAGTGGTATCAACGCAGAGTACATrGrG+G) with a homotypic PCR binding site. Betaine and MgCl2 were added to enhance the reaction. After reverse transcription, whole transcriptome amplification by PCR was performed for 12 cycles using KAPA HiFi HotStart ReadyMix (Roche, 7958935001) with a PCR primer (AAGCAGTGGTATCAACGCAGAGT). The PCR reaction was cleaned up with Ampure XP beads (Beckman Coulter, A63881). In addition, after the whole transcriptome PCR amplification step, qPCR was performed to determine the presence of microglia-specific transcripts (Cx3cr1) and housekeeping genes (Gapdh). The amplified DNA was diluted to a concentration of 0.5ng/ul, and subjected to tagmentation with the Illumina Nextera XT kit (Illumina, FC-131-1096). Each sample was PCR amplified with a unique set of Nextera indices. Bulk microglia RNA-Seq libraries were sequenced on an Illumina HiSeq, while in vitro microglia RNA-Seq libraries were sequenced on an Illumina Nova-Seq.

RNA-Seq Analysis

FASTQ reads were pseudo aligned to the mouse transcriptome (GRCm39 cDNA from Ensembl) using kallisto⁴⁷ with default parameters. Next, transcript abundance estimates were imported into DeSeq2⁴⁸. Differential expression analysis was performed using DeSeq2 using the Wald significance test. Gene ontology analysis was performed with Panther^49–51. Principal component analysis plots were generated with the plotPCA function included with DeSeq2 and graphed with ggplot2. Heatmaps were generated with the pheatmap package in R.

The overlap between gene expression differences in aging and LPS treatment were determined using genes significantly changed between the 6- and 24-month-old timepoints in the scRNA-Seq data and the control and LPS treated primary microglia. ξ² test was used to determine the significance of the overlap between these two datasets. The Venn diagram was generated using the VennDiagram package in R.

10x Genomics Single Cell RNA-Sequencing

10x Single-Cell RNA-Seq libraries were generated from Cd11b+ cells isolated from the hippocampi of an aging cohort of mice consisting of 6-, 12-, 18-, and 24-month-old C57BL/6J mice that were all collected and processed on the same day in an interspersed order. Mice were sedated with ketamine followed by perfusion with ice-cold PBS. Subsequently, the entire brain was removed from the mouse followed by sub-dissection of the hippocampus. For each age, hippocampi from five mice were pooled during the dissection step in HBSS at 4°C. Single-cell suspensions were generated by enzyme-mediated (papain) and mechanical dissociation using Miltenyi Neural Dissociation Kits (P). The papain dissociation was done at 37°C for 10 minutes with 3 trituration steps. All other processing steps were performed at 4°C. Myelin was depleted from the suspensions using Myelin Removal Beads (Miltenyi). Microglia were enriched using Cd11b magnetic beads (Miltenyi, 130-126-725). Cell viability was determined to be over 95%. Single-cell RNA-Seq 3’ libraries were generated from the cell suspension using 10x Genomics Chromium Single-Cell 3’ Solution. Libraries were sequenced on the Illumina Nova-Seq. Cell Ranger demultiplexed and mapped (using bcl2fastq) reads, followed by alignment (with STAR), and generation of single-cell expression matrices⁵².

The isolation protocol for the Tgfb1 genetic mouse model was modified. We added transcriptional and translational inhibitors (Actinomycin D (Sigma-Aldrich, A1410), Anisomycin (Sigma-Aldrich, A9789), and Triptolide (Sigma-Aldrich, T3652) to prevent transcriptional changes due to ex vivo activation ⁵³. We utilized the concentrations presented in Marsh et al. 2022⁵³.

Single-cell RNA-Seq Analysis

Seurat: The single-cell gene expression matrices generated by Cell Ranger were loaded into Seurat⁵⁴.

Aging microglia dataset. Cells were filtered according to the following parameters: Genes > 1000 & RNA count > 10000 & mitochondrial percentage < 5. The count data was log normalized. The 2,500 most variable genes were identified using the “vst” method. The expression data was scaled and centered. Principal component analysis was performed, and the first 25 PCAs were used to identify nearest neighbors, and interconnected clusters were identified with a resolution of 0.4. UMAP was used to visualize the results. Preliminary dimensionality reduction and visualization with UMAP using 25 PCAs and 25 nearest neighbors in Seurat identified a large microglia cluster, plus sparse clusters consisting of astrocytes, vascular cells, neutrophils, and macrophages. Hexb, P2ry12, Tmem119, and Slc2a5 are used as markers to identify microglia in the preliminary clustering, while Nrxn1 (neurons), Tm4sf1 (vasculature), Dclk1 (astrocytes), Sdpr (vasculature smooth muscle), Abcb1a (pericyte/vasculature), and Foxq1 (macrophages) identified other cell types. The data was then subsetted on the non-proliferating microglial clusters (to avoid confounds associated with the cell cycle). The top 1,500 most variable genes in the subsetted microglia were identified using the “vst” method. The data was scaled and centered. Principal component analysis was performed, and the first 8 PCAs were used to identify nearest neighbors, and interconnected clusters were identified with a resolution of 0.25. UMAP was used to visualize the results using 8 PCAs and 100 nearest neighbors. Gene enrichment and differential expression analysis was performed on the filtered dataset to determine differences between ages.

Tgfb1 cKO microglia dataset. Cells were prefiltered according to the following parameters: Hexb > 4 & Tmem119 > 1 & Cx3cr1 > 4 & Top2a < 1 & Mki67 < 1 nFeature_RNA > 1000 & nFeature_RNA < 6000 & nCount_RNA > 2000 & nCount_RNA < 40000 & percent.mt <10. Data was integrated using IntegrateData and scaled. Principal component analysis was performed, and the first 50 PCAs were used to identify nearest neighbors, and interconnected clusters were identified with a resolution of 0.35. UMAP was used to visualize the results. Preliminary dimensionality reduction and visualization with UMAP using 50 PCAs identified a large microglia cluster, plus clusters of other cell types. The non-proliferative microglia clusters were subsetted for further analysis (the macrophage cluster was differentiated from the microglia cluster based on increased expression of Mcp1, Pf4, and Ms4a7). The data was scaled and centered. Principal component analysis was performed, and the first 10 PCAs were used for UMAP dimensionality reduction and to identify nearest neighbors, and interconnected clusters were identified with a resolution of 0.20.

Monocle 3: Monocle 3 was used to generate the pseudotime analysis^55–58. Data was imported from Seurat analysis and a Monocle dataset was generated. The dataset was preprocessed with normalization of the data and PCA generation. Next, the preprocessed data underwent further non-linear dimensionality reduction using UMAP. Cells were clustered, and trajectories were determined for cells within a cluster. For determining “pseudotime” aging of microglia, the root node of the pseudotime trajectory was manually placed in the cluster of 6-month cells. The graph segments of the pseudotime trajectories resulting in inflammatory activation were chosen for further analysis. Graph autocorrelation analysis was performed on the inflammatory trajectory using Moran’s I, which determines if genes are expressed in focal regions of graph space. A cutoff q-value of 0.005 was set as significant for focal expression. Subsequently, coregulated modules were determined using Louvain community analysis at an optimized resolution of 0.0078. The modules were then overlaid onto the UMAP plot to determine the region of the trajectory with focal expression of the module. The ten most significant genes of each module, as determined by q-value were used to construct the heatmaps in Figures 3F and 3I. The significant genes of each module were used to determine the effects of manipulations in the presence of LPS in Figure 3J.

Immunohistochemistry

Mice were sedated with ketamine followed by perfusion with 30mL of ice-cold PBS. Brains were collected from PBS perfused mice and fixed overnight at 4°C in 4% paraformaldehyde in PBS. The brains were washed with PBS and immersed in 30% sucrose (in PBS), and then stored at 4°C until the brains sank. 40μM coronal sections were sliced using a microtome at -20°C. Sections were stored in cryopreservation buffer (40% PB Buffer (0.02 M Sodium Phosphate Monobasic, 0.08 M Sodium Phosphate Dibasic, pH 7.4), 30% Glycerol, 30% Ethylene Glycol) at -20°C until staining. Hippocampal sections were washed three times in TBST. Sections were permeabilized in pretreatment buffer (0.1% Triton-X in TBST) with rocking for 1 hour at room temperature. After three washes in TBST, sections were blocked in 5% goat serum for 2 hours (for IBA1/CD68 double stain) or 5% donkey serum (NFKB/Iba1 and C1q/C3/Iba1 stains). Blocking buffer was replaced with primary antibody mixture and incubated overnight at 4°C with rocking. The following primary antibodies were used: rabbit anti-IBA1 (1:1000, Wako, Cat# 019-19741, RRID:AB_839504), rat anti-CD68 (1:250, Bio-Rad, Cat# MCA1957, RRID:AB_322219), rabbit anti-NFKB p65 (1:500, Santa Cruz, Cat# sc-372, RRID:AB_632037), guinea pig anti-Iba1 (1:1000, Synaptic Systems, Cat# 234-004, RRID:AB_2493179), rabbit anti-C1q (1:1000, Abcam, Cat# ab182451, RRID:AB_2732849), rat anti-C3 (1:1000, Abcam, Cat# ab11862, RRID:AB_2066623), and rabbit anti-S6 (1:500, Cell Signaling, Cat# 2217, RRID:AB_331355). The sections were washed three times with TBST. Subsequently, secondary antibody solution was added to the sections, and incubated at room temperature with rocking for two hours. The following secondary antibodies were used: goat anti-rabbit AlexaFluor 488 (Life Technologies, Cat# A-11008, RRID:AB_143165), goat anti-rat AlexaFluor 555 (Life Technologies, Cat# A-21434, RRID:AB_2535855), donkey anti-guinea pig AlexaFluor 488 (Jackson Immuno, Cat# 706-545-148, RRID:AB_2340472), donkey anti-rabbit AlexaFluor 555 (Life Technologies, Cat# A-31572, RRID:AB_162543), and donkey anti-rat AlexaFluor 647 (Jackson Immuno, Cat# 712-606-150, RRID:AB_2340695). The sections were washed three times with TBST, with the first wash containing Hoescht 33342 (Invitrogen, H3570) at a concentration of 1μg/mL. The sections were transferred to phosphate buffer and mounted on frosted slides. Coverslips were mounted with Prolong Gold Antifade (Thermo Fisher, P10144) reagent after the sections were dried. Slides were imaged at 20x magnification on a Zeiss LSM 800 or LSM900 for a final resolution of 0.312μM/pixel. 8-9 z-planes were imaged at 3μM intervals. The hippocampus was imaged, and the tiled images were stitched together with Zeiss ZenPro software. 2-3 hippocampal sections were imaged from each sample.

Images were analyzed with FIJI. Maximum intensity projections of Z-stacked images were constructed. Images were converted to 8-bit images and consistent adjustments were made for each channel within an experiment. The Iba1 channel was used to quantify microglia using the particle counter function in FIJI and the counts were validated by manual visual counting. The presence of CD68, NFKB(p65), C3, or C1q in microglia was determined by creating ROIs within the microglia in a field using the create selection function on the Iba1 images. The ROIs of the hippocampal subregions and the microglia were combined, and CD68 puncta meeting a minimum particle size threshold (100 pixels, equating to 31.2μM) within microglia determined if the microglia were activated (CD68+IBA1+), while the measure function determined the intensity of NFKB(p65), C3, or C1q. To determine percent activation of microglia with IBA1/CD68 staining, we divided the number of microglia that were activated (CD68+IBA1+) by the total number of microglia (IBA1+) in the analyzed regions and multiplied by 100.

In vitro treatment of microglia

Mixed glial cultures were generated from p1 male pups. Cortices and hippocampi were isolated from the pups, and meninges were removed. The tissue was broken up by repeatedly pipetting up and down, followed by dissociation with Trypsin. Single cell suspensions were plated on poly-L-lysine coated flasks in DMEM (Thermo Fisher, 11965-118) supplemented with 10% FBS (Gemini Bio-Products, 900-208) and antibiotic mix (Penicillin/Streptomycin)(Genesee Scientific, 25-512) and incubated at 37°C in 5% CO₂. Media was changed after 4 days to remove debris. Microglia were isolated after 14 days by agitating the plates to dislodge microglia, then transferring the supernatant of each sample evenly amongst 4 wells of a 24 well plate. Microglia were incubated for 16 hours followed by replacement with serum free macrophage SFM media (Thermo Fisher, 12065074) supplemented with M-CSF (10ng/m) (Peprotech, 315-02) and antibiotic mix. After 24 hours microglia were treated with RNA polymerase I Inhibitor 2, CX-5461 (100nM)(EMD Millipore, 509265) or human TGFβ1 recombinant protein (10ng/mL)(ThermoFisher Scientific, PHG9204) for 24 hours followed by addition of lipopolysaccharides (LPS, 200ng/uL) from E. coli (Sigma-Aldrich, L2018) for 8 hours and ATP (10nM)(InvivoGen, tlrl-atpl) for 30 minutes. Tri Reagent was added to the culture plate, and the plates were shaken for 5 minutes. RNA was isolated according to manufacturer’s instructions. RNA-Seq libraries were generated as above for the bulk RNA-Seq.

RNAscope

RNAscope was performed with the RNAscope® Multiplex Fluorescent Reagent Kit v2 (ACD, 323110). 40μM coronal sections were washed with TBST, then incubated with hydrogen peroxide for 45 minutes at 25°C. After washing with TBST, the sections were incubated in Target Retrieval Buffer at 95°C for 10 minutes. The sections were mounted on slides and dried overnight. The sections were treated with RNAscope Protease III for 10 minutes at 40°C. Sections were washed in water four times. Sections were incubated with Tgfb1 probe (ACD, 443571-C2) for 2 hours at 40°C. The sections were washed twice with RNAscope wash buffer. The sections were incubated with AMP1 for 30 minutes, AMP2 for 30 minutes, AMP3 for 15 minutes, HRP-C2 for 15 minutes, TSA Plus Cy3 (1:1000, Perkin Elmer, SKU NEL744001KT) for 30 minutes, and HRP-Blocker for 15 minutes; all at 40°C with washes with RNAscope wash buffer between every incubation. RNA-Protein Co-Detection Ancillary Kit (ACD, 323180) buffer was used to block the sections before incubating with rabbit anti-IBA1 (1:500, Wako, Cat# 019-19741, RRID:AB_839504) overnight at 4°C. Sections were washed three times in TBST. Sections were incubated with secondary antibody solution with goat anti-rabbit AlexaFluor 488 (1:1000, Life Technologies, Cat# A-11008, RRID:AB_143165) for two hours at room temperature. The sections were washed three times with TBST, with the first wash containing Hoescht 33342 (Invitrogen, H3570) at a concentration of 1μg/mL Coverslips were mounted with Prolong Gold Antifade (Thermo Fisher, P10144) reagent after the sections were dried. Slides were imaged at 40x magnification on a Zeiss LSM 900. 10 z-planes were imaged at 1μM intervals.

Images were analyzed with FIJI. Maximum intensity projections of Z-stacked images were constructed. Images were converted to 8-bit images and the Tgfb1 was thresholded. For each section, the intensity of 5-7 individual Tgfb1 puncta were averaged to get the average puncta intensity of that section. The intensity of all Tgfb1 signal in clearly identifiable individual IBA1 cells was measured for 20-25 cells in the molecular layer of the hippocampus. The mean Tgfb1 count per microglia for each sample was calculated by dividing the average Tgfb1 intensity in IBA1 cells by the average puncta intensity.

Contextual fear conditioning

In this task, mice learned to associate the environmental context (fear conditioning chamber) with an aversive stimulus (mild foot shock; unconditioned stimulus, US) during the training phase enabling testing for hippocampal-dependent contextual fear conditioning. To also assess amygdala-dependent cued fear conditioning, the mild foot shock was paired with a light and tone cue (conditioned stimulus, CS) during the training phase. Conditioned fear was displayed as freezing behavior. Specific training parameters are as follows: tone duration is 30 seconds; level is 70 dB, 2 kHz; shock duration is 2 seconds; intensity is 0.6 mA. This intensity is not painful and can easily be tolerated but will generate an unpleasant feeling. More specifically, on day 1 each mouse was placed in a fear-conditioning chamber and allowed to explore for 2 min before delivery of a 30-second tone (70 dB) and light ending with a 2-second foot shock (0.6 mA). Two minutes later, a second CS-US pair was delivered. On day 2, each mouse was first placed in the fear-conditioning chamber containing the same exact context, but with no CS or foot shock. Freezing was analyzed for 2 minutes and represented as the percentage of time that the mouse froze over the 2 minutes. One hour later, the mice were placed in a new context containing a different odor, floor texture, chamber walls and shape. Animals were allowed to explore for 2 minutes before being re-exposed to the CS. Freezing was analyzed for 30 seconds following the CS and represented as the percentage of time that the mouse froze over those 30 seconds. Determination of freezing behavior was performed using FreezeScan video tracking system and software (Cleversys, Inc). Single outliers were removed using the extreme studentized deviate method (Grubbs’ test) with an alpha of 0.05.

Novel object recognition

The novel object recognition task was adapted from a previously described protocol⁵⁹. Specifically, during the habituation phase (day 1), mice could freely explore an empty open field arena for 10 minutes (motor activity and anxiety (time in center) were measured during this phase). During the training phase (day 2), two identical objects were placed in the habituated arena, and mice could explore the objects for 5 minutes. For the testing phase (day 3), one object was replaced with a novel object, and mice could explore the objects for 5 minutes. Time spent exploring each object was quantified using the Smart Video Tracking Software (Panlab; Harvard Apparatus). Two different sets of objects are used. To control for any inherent object preference, half of the mice are exposed to object A as their novel object and half to object B. To control for any potential object-independent location preference, the location of the novel object relative to the trained object is also varied. The objects were chosen based on their ability to capture the animal’s interest, independent of genetic background or age. To determine percent time with novel object, we calculate (Time with novel object)/(Time with Trained Object + Time with Novel Object) * 100. In this preference index, 100% indicates full preference for the novel object, and 0% indicates full preference for the trained object. A mouse with a value of 50% would have spent equal time exploring both objects. Mice that did not explore both objects for 5 seconds during the training phase or testing phase were excluded from analysis. Single outliers were removed using the extreme studentized deviate method (Grubbs’ test) with an alpha of 0.05.

Y maze

The Y Maze task was conducted using an established forced alternation protocol⁶⁰. During the training phase, mice were placed in the start arm facing the wall and allowed to explore the start and trained arm for 5 minutes, while entry to the 3rd arm (novel arm) was blocked. The maze was cleaned between each mouse to remove odor cues, and the trained arm was alternated between mice. The mouse was then removed to its home cage. After 30 minutes, the block was removed, and the mouse was returned to the start arm and allowed to explore all 3 arms for 5 minutes. Time spent in each arm was quantified using the Smart Video Tracking Software (Panlab; Harvard Apparatus). The percent time in the novel arm was defined as time in the novel arm divided by time spent in the novel and trained arms during the task.

Datasets

All RNA-Seq and scRNA-Seq data have been deposited in the Gene Expression Omnibus and are publicly available as of the date of publication. The following datasets were generated for this manuscript and deposited in Gene Expression Omnibus: aging single-cell RNA-Seq (GSE179358), in vitro treated primary microglia (GSE179611), and Tgfb1 cKO microglia RNA-Seq (GSE190007).

Statistics

Researchers were blinded throughout histological assessments with groups being un-blinded at the end of each experiment upon statistical analysis. Data are expressed as mean ± s.e.m with individual sample values being shown. Statistical analysis was performed with Prism 8.0 (GraphPad), R, DESeq2, Seurat, or Monocle 3. Comparisons of means in histology experiments were analyzed with multiple T-tests followed by Holm-Sidak correction or mixed effects analysis followed by Dunnett’s multiple comparisons (Prism). Changes in expression for aggregated gene sets were determined with one sample T-tests with the expected value of 0 in log2 space (null hypothesis of no change). Differential expression analysis in DESeq2 was based on Wald’s significance test of the negative binomial distribution. Differential expression analysis of single-cell RNA-Seq data in Seurat was performed using non-parametric Wilcoxon rank sum tests. Autocorrelation analysis to determine focal expression Monocle 3 was done using Moran’s I. The significance of gene set overlap was determined using the ξ² test with R. All data generated or analyzed in this study are included in this article.