Multi‐Omic Analysis Reveals Astrocytic Annexin‐A2 as Critical for Network‐Level Circadian Timekeeping in the Suprachiasmatic Nucleus

All mouse work was carried out in accordance with the UK Animals (Scientific Procedures) Act of 1986, with local ethical approval from the Animal Welfare and Ethical Review Body (AWERB) of the Medical Research Council (MRC) Laboratory of Molecular Biology (LMB). Mice were maintained and bred at the LMB Ares facility and were maintained on a 12:12 light:dark schedule with food and water available ad libitum. Period2::Luciferase (Per2::Luc) mice (Yoo et al. 2004) were a gift from Dr. J.S. Takahashi (University of Texas Southwestern Medical Center, Dallas, TX) back‐crossed in‐house onto C57BL/6J mice (Jackson Laboratories), which were also used as wild‐types.

Mice of postnatal day 10 (P10) to P12 and of either sex were killed, and SCN explants were prepared as previously described (Hastings et al. 2005; Patton et al. 2020). The brain was removed and transferred to ice‐cold Gey's Buffered Saline Solution (GBSS) supplemented with 5 mg/mL glucose, 3 mM MgCl₂, 0.05 mM DL‐AP5, and 100 nM (+)‐MK‐801. It was then trimmed into a block containing the hypothalamus before 300 μm thick coronal slices were made on a tissue chopper (McIlwain, UK). The suprachiasmatic nucleus (SCN) was dissected from the coronal hypothalamic slices before being cultured as an explant via the interface method on tissue culture inserts (Millipore). The slices were rested in culture medium (CM) consisting of 50% Eagle's Basal Medium, 25% Earle's Balanced Salt Solution, and 25% horse serum supplemented with 5 mg/mL glucose, 2 mM Glutamax (Gibco), penicillin/streptomycin (Gibco) and, additionally, 3 mM MgCl₂, 0.05 mM DL‐AP5, and 100 nM (+)‐MK‐801 for a few hours before being transferred to HEPES‐buffered Dulbecco's Modified Essential Medium‐based recording medium (AM) supplemented with Glutamax, penicillin/streptomycin, fetal calf serum (FCS) and B‐27 for at least 1 week before use.

To heavy‐label SCN slices, a specially formulated SILAC medium based on our CM was used, with the substitution of custom DMEM lacking arginine and lysine (AthenaES), 10% heat‐inactivated horse serum dialyzed (10 kDa, Spectrum Lab) against 50 mM NH₃HCO₃ (Invitrogen), and 0.398 μM Arg + 10 and 0.798 μM Lys + 6 (Cambridge Isotope Laboratories). Wild‐type slices were maintained in this SILAC CM for 8 weeks to create a time‐free SILAC SCN “spike‐in” proteome. The labeling efficiency was measured to ≈95%, based on the number of heavy peptides divided by the number of light peptides (Chen et al. 2015) (LFQ vs. peptide count) (Figure 1A).

During recording, slices were maintained in AM supplemented with 100 μM luciferin in dishes sealed with glass coverslips (Patton et al. 2020). Aggregate bioluminescence was monitored in real‐time in customized light‐tight incubators with individual photon‐multiplier tubes (PMTs; catalogue #H9319‐11 photon counting head, Hamamatsu). Data collected were binned into 6‐min epochs for export and analysis.

At least 1 week after preparation, Per2::Luc slices were transferred to PMTs to record their free‐running bioluminescence. After several circadian cycles, the circadian phase was determined via the Per2::Luc bioluminescence signal, and slices were harvested serially at CT0, 6, 12, 18, and 24 (i.e., CT0 repeat). Thus, although the CT time stamp was absolute, it would have been meaningless to refer to “cycle 1” and “cycle 2” as all slice samples were free‐running independently without common registration to external time. This structure necessitated that our analysis was limited to a one‐cycle design, which therefore informed how we should apply circadian algorithms to analyze our data.

Slices were washed with 0.01 M Phosphate Buffered Saline (PBS) and then mechanically aspirated before lysis in 2% SDS in Tris‐buffered saline (TBS) (pH 7.4, 5 μL/slice). Seven SCN slices were pooled per triplicated sample and boiled for 5 min before sonication in a water bath for 30 s. After centrifugation, the protein concentration of the supernatant was determined by BCA assay (Pierce). The unlabeled circadian time‐point samples were mixed 1:1 with the SILAC‐labeled SCN lysate; subsequently, 50 μg of the protein mix was digested by the FASP method (Wisniewski 2017) with 1 μg trypsin (Sigma).

The resulting peptide mixtures were acidified by the addition of formic acid (final concentration 2% v/v). The digests were analyzed in triplicate (3 μg initial protein/injection) by nano‐scale capillary LC–MS/MS using an Ultimate U3000 HPLC (ThermoScientific Dionex) to deliver a flow of approximately 300 nL/min. A C18 Acclaim PepMap100 5 μm, 100 μm × 20 mm nanoViper (ThermoScientific Dionex) trapped the peptides prior to separation on a C18 PepMap100 3 μm, 75 μm × 250 mm nanoViper (ThermoScientific Dionex). Peptides were eluted with a 120 min gradient of acetonitrile (2%–60%). The analytical column outlet was directly interfaced via a nano‐flow electrospray ionization source, with a hybrid quadrupole orbitrap mass spectrometer (Q‐Exactive Plus Orbitrap, ThermoScientific).

Data‐dependent analysis was carried out using a resolution of 30,000 for the full MS spectrum, followed by 10 MS/MS spectra. MS spectra were collected over an m/z range of 300–2000. MS/MS scans were collected using a threshold energy of 27 for higher energy collisional dissociation (HCD). MS raw files were processed by MaxQuant (Tyanova et al. 2016) version 1.3.0.5 and searched with Andromeda (Cox et al. 2011) against a reviewed mouse proteome retrieved from Uniprot (Oct 2014, 16,650 entries). Searches were done with an MS/MS tolerance of 20 ppm and 1% FDR at both the peptide and protein levels. Carabamidomethylation of cysteine was set as a fixed modification, and oxidation of methionine and protein N‐terminal acetylation were chosen as variable modifications. Two missed cleavages were allowed. One quantified peptide per protein group was allowed since further statistical analysis on the biological replicates was applied. Common contaminants were removed and further filtered and processed with R statistical software (R). As a criterion for quantification, at least two‐SILAC ratios per time‐point were required and SILAC ratios were log2 transformed.

Quantified proteins were analyzed for circadian oscillation using two different and widely used circadian algorithms: RAIN (version 1.34.0) (Thaben and Westermark 2014) and JTK‐Cycle (version 3.1) (Hughes et al. 2010). Both RAIN and JTK‐Cycle were implemented in R running in the RStudio environment. Due to the nature of our data (low‐frequency sampling over 1 cycle with an independently repeated CT24/0), we used a stringent, conservative criterion such that a protein in our dataset could only be considered circadian if it was significant by both RAIN (p < 0.05) and JTK‐Cycle (p < 0.05) (Figure S1), and the overlap was determined using the VennDiagram package (version 1.7.3) (Chen and Boutros 2011).

The expression data heatmap was produced in R using the pheatmap package (version 1.0.12) (Kolde 2019), implementing k‐means clustering to cluster the data into the four expected phases that could reasonably be expected from the experimental structure (peaking at CT0/24, 6, 12 or 18). Data represented in the heatmap are rescaled by protein to reveal the dynamics without consideration of different oscillation amplitudes.

Publicly available data from a previously published scRNA‐seq dataset produced from free‐running SCN explants (Morris et al. 2021) were accessed from the NCBI Gene Expression Omnibus with the accession number: GSE167927↗. These data were re‐analyzed in R using the Seurat package (version 5.1.0) (Hao et al. 2021, 2024) to generate a cell‐type‐specific transcriptome. Data were first filtered to remove cells containing high levels of mitochondrial transcripts before data were log‐normalized and scaled using the NormalizeData and ScaleData functions. The data were then clustered using FindNeighbors(dim = 40) and FindClusters(resolution = 1.2, method = “igraph,” algorithm = 3) before clusters were assigned identities based on their expression of cell‐specific marker genes to allow further analysis of cell‐type specific transcriptomes.

Cell‐type enrichment of Anxa2 and potential interaction partners were examined by first identifying Anxa2 interaction partners using the BioGRID database of curated protein–protein interactions (www.thebiogrid.org↗, BioGRID version 4.4.235; Oughtred et al. 2021; Stark et al. 2006). This analysis was filtered for interactions identified in Mus musculus , and a comprehensive list of 40 potential interactors (including Anxa2) was queried against the cell‐type clustered single‐cell transcriptomic dataset using the FetchData function in the Seurat package and pulling the expression data from the counts data layer along with the cell identities. These were then summed across cell‐types and min‐max normalized across genes using the caret package (version 6.0‐94) (Kuhn 2008). Finally, a heatmap was generated using the pheatmap function. A differential expression test for day‐night expression of astrocyte transcripts was performed using the built‐in Seurat FindMarkers function, using the Wilcox test, subsetting the Seurat object to include only astrocytes and grouping the data by time of day of harvesting, which was added to the metadata. Volcano plots were generated using the EnhancedVolcano package (version 1.18.0) (Blighe et al. 2024).

GO‐term analysis was carried out in R using the clusterProfiler package (version 4.8.3) (Wu et al. 2021). To determine the cluster‐specific GO‐terms from the scRNA‐seq cell clusters, first, a differential expression test was performed for specific clusters using the Seurat function FindAllMarkers with the only.pos = TRUE option to find transcripts whose expression (presence) defines those cells. These results were then filtered to include only significant (adjusted p < 0.05) transcripts and ordered by significance. From this list, the top 250 genes in each cluster were passed to the GO‐term analysis function. These cell‐specific lists were analyzed using the compareCluster function from clusterProfiler, using the enrichGO option to specify the molecular function (MF) ontology and Mus musculus organism database (org,Mm.eg.db, version 3.17.0; Carlson 2023). Even though 11 cell‐type clusters were identified in our transcriptomic dataset, due to overlap in the top two GO‐terms reported for closely related cell‐clusters, 19 final GO terms were identified, not the expected 22.

To determine the MF GO‐terms for the circadian proteome data, astrocyte transcriptome, and intersection between the two, gene lists were passed to the enrichGO function, specifying the Mus musculus organism database and MF ontology as above. The top 10 most significant GO‐terms were then passed to the cnetplot function from the enrichplot package in R to plot category network plots enabling the visualization of contributing genes/proteins.

Probe sets against S100a10 (B1 adaptor, designed from NM_009112.2↗, 12 probes), Anxa2 (B2 adaptor, designed from NM_007585.3↗, 23 probes) and Sox9 (B3 adaptor, designed from NM_011448.4↗, 20 probes) were designed and manufactured by Molecular Instruments. SCN slices were excised from the culture insert and fixed in 4% PFA for 1 h before being washed three times for 15 min each in PBS‐T (0.01 M PBS supplemented with 0.1% Tween‐20). Slices were then dehydrated in a series of graded methanol washes: 100%:0%, 75%:25%, 50%:50%, 25%:75%, and 0%:100% PBS‐T:methanol, before rehydration through the same series of washes in reverse: 0%:100%, 25%:75%, 50%:50%, 75%:25%, and 100%:0% PBS‐T:methanol. These graded washes were incubated for 15 min on ice per step. Slices were then incubated in pre‐warmed Probe Hybridization buffer (Molecular Instruments) supplemented with 1:20 RNAse Inhibitor, Murine (New England Biolabs) for 1 h at 37°C before being incubated with 40 nM S100a10‐B1, 40 nM Anxa2‐B2, and 8 nM Sox9‐B3 probe sets in Probe Hybridization Buffer supplemented with 1:20 RNAse Inhibitor, Murine at 37°C for 24 h.

Primary probes were removed from slices by three washes in pre‐warmed Probe Wash Buffer (PWB, Molecular Instruments) at 37°C for 15 min each before additional graded 5× SSC with 0.1% Tween‐20 (SSC‐T) and PWB, wash steps at 37°C for 15 min each: 0%:100%, 25%:75%, 50%:50%, 75%:25%, 100%:0% SSC‐T:PWB. Finally, slices were run through a further three SSC‐T washes for 15 min at each step at room temperature before incubation in Amplification Buffer (Molecular Instruments) supplemented with 1:20 RNAse Inhibitor, Murine for 1 h at room temperature. Half an hour before the end of the incubation with Amplification Buffer, fluorescently labeled hairpins were snap‐cooled by heating individual hairpins to 95°C for 90 s in a thermocycler, before immediately transferring tubes to a dark drawer to cool over 30 min. Once cooled, 60 nM B1‐Alexa‐647 (h1/h2), B2‐Alexa‐594 (h1/h2), and B3‐Alexa‐488 (h1/h2) amplifiers (Molecular Instruments) were added to Amplification Buffer supplemented with 1:20 RNase Inhibitor, Murine, and slices were incubated for an additional 24 h at room temperature.

Secondary probes were removed by eight sequential SSC‐T washes for 15 min each. Slices were incubated in SSC‐T supplemented with 1 μg/mL DAPI (Tocris, UK) for 1 h before three more SSC‐T washes for 5 min each. Slices were then briefly rinsed in ddH₂O before mounting onto glass slides with Vectashield Vibrance Antifade Mounting Medium (H‐1700, Vector Laboratories).

GFAP.jGCaMP7b was a gift from Thomas Oertner (Lohr et al. 2021) (Addgene plasmid # 171118; http://n2t.net/addgene:171118↗; RRID: Addgene_171118) packaged into AAV serotype 5 particles by VectorBuilder (> 1 × 10¹³ GC/mL). Syn.NES‐jRCaMP1a was a gift from Douglas Kim & the GENIE project (Addgene viral prep # 100848‐AAV1; http://n2t.net/addgene:100848↗; RRID: Addgene_100848) (Dana et al. 2016).

At least 1 week post‐culturing, organotypic SCN slices were transduced with AAV particles by dropping 1 μL AAV directly on top of the SCN slice. Slices were placed in an incubator and super‐transduced with further AAVs in the same manner at an interval of 2–3 days to express different reporters. AAVs were washed out of the slice by complete medium change.

AAV‐expressed fluorescent reporters were imaged in slices without staining. Slices were fixed in 4% PFA diluted in 0.01 M phosphate buffer for 1 h before three washes in PBS‐T for 15 min each wash. Slices were incubated in PBS‐T supplemented with 1 μg/mL DAPI for an hour before being washed three times in PBS‐T for 5 min each wash and finally rinsed in ddH₂O. Slices were mounted on glass slides with Vectashield Vibrance Antifade Mounting Medium and stored overnight in the fridge before imaging.

Slides were stored overnight in a fridge at 4°C before imaging to allow the mounting medium to reach its optimal refractive index. Slices were imaged on a Zeiss 900 inverted confocal microscope using 40×/1.3 NA or 63×/1.4 NA oil immersion objectives. Slices were imaged as tile scans with 10% overlap and 16× averaging.

For combined circadian fluorescence imaging, SCN slices were sealed in glass‐bottom imaging dishes (P35G‐0‐10‐C, Mattek) containing the same HEPES‐buffered DMEM‐based recording medium as in the PMT recordings. Slices were imaged on an LV200 Imaging System (Olympus). Fluorescence channels were acquired sequentially, with acquisition set at 100 ms. The acquisition interval for circadian imaging was 30 min.

A2ti‐1 (2‐(4‐phenyl‐5‐o‐tolyloxymethyl‐4H‐[1,2,4]triazol‐3‐ylsulfanyl) acetamide) and A2ti‐2 (2‐[4‐(2‐ethylphenyl)‐5‐o‐tolyloxymethyl‐4H‐[1,2,4]triazol‐3‐ylsulfanyl] acetamide) (MedChemExpress) were solubilized in DMSO to 100 mM. DMSO was used as the control vehicle treatment at a final concentration of 0.1%–0.2% dependent on experimental conditions. Dilutions for dose response curves were made in DMSO so that the concentration of DMSO across treatments was consistent. Pharmacological treatments were washed out by three serial transfers of the membrane and SCN into fresh medium for 10 min each step, before the final transfer to fresh medium for recording.

PMT recordings were analyzed by identification of peaks and troughs in the raw bioluminescence signal, as described previously (Patton, Morris, et al. 2023). The peak‐to‐peak period was calculated as the time difference between consecutive peaks and the period of an experimental interval was the mean of the peak‐to‐peak periods for that interval. The change in period (ΔPeriod) was the difference in hours between the mean period for the baseline and treatment intervals. Cycle amplitude was the absolute difference between trough and peak, normalized to the amplitude of the cycle before treatment (i.e., the last baseline peak and trough). Phase shifts were calculated as the time difference between the second peak post‐treatment and the predicted second peak post‐treatment, as extrapolated from the last pre‐treatment peak and baseline period. The second cycle post‐treatment accounts for any potential changes in waveform on the first cycle, therefore representing the sustained shift. Statistical analyses for cycle amplitude and phase‐shift meta‐analyses across experiments were carried using a linear mixed model approach where the main factors were treatment phase and treatment, and slice identity was a random effect to account for repeated treatments at different phases.

Relative amplitude error (RAE) values (an inverse measure of oscillation coherence) were determined using FFT‐NLLS measures in BioDARE2 (Zielinski et al. 2014) (https://biodare2.ed.ac.uk/↗). RAE values for treatments were normalized to the baseline. Washout dynamics were calculated by taking the ratio of the amplitude of the rising phase of the cycle (the amplitude between the preceding trough and the peak, Before [B]) and the amplitude of the falling phase (the amplitude between the peak and the following trough, After [A]). This ratio was termed the B/A ratio and was defined at steady state as the second peak post‐washout.

Real‐time fluorescence data were processed in FIJI (Schindelin et al. 2012) as described previously (Patton, Morris, et al. 2023). Fluorescence stacks were background‐subtracted using the built‐in rolling ball algorithm at 2 pixels before taking aggregate raw mean gray values. Data were detrended by subtracting a line fit to the last 72 h of the baseline interval of the recording projected forward into the treatment interval. Finally, smoothing was applied by a 5‐point (2.5 h) moving average. To see the general trend across experiments on pharmacological treatment, slices were temporally pseudo‐aligned by the final jRCaMP1a peak in the baseline interval to create a mean oscillation profile. Individual data points around the treatment interval were excluded from this aggregate if they did not align with other slices due to treatment artifact, thereby creating a gap between the final point in the baseline and the first point during treatment. Complete traces with the missing datapoints that underlie these are provided in Figure S3.

Where possible, slices received paired treatments (vehicle and drug) and were exposed to all concentrations along a dose–response curve (in an interleaved design). Where this was not possible (or slices died), slices were randomly assigned to groups. In all real‐time experiments, slice recordings covered baseline, treatment and (where possible) washout intervals. Baselines acted as additional internal paired controls alongside vehicle. All statistical tests used are listed in the text and were performed in Graphpad Prism with the exception of Rayleigh tests and Linear Mixed Model analyses which were carried out in R using the r.test function from the CircStats package (version 0.2‐6) (Lund and Agostinelli 2018) and the lme function from the nlme package (version 3.1‐162) (Pinheiro and Bates 2000; Pinheiro et al. 2024) alongside the emmeans package (version 1.10.2) (Lenth 2024), respectively.

All data were analyzed in Excel (Microsoft), R (R Foundation for Statistical Computing, Vienna, Austria. R version 4.3.0 [2023‐04‐21]), RStudio (Posit PBC, software version 2023.03.1+446), FIJI (Schindelin et al. 2012), and GraphPad Prism 10 (GraphPad). Statistical tests used are listed in the text and figure legends. All numbers reported in text are mean ± SEM unless otherwise stated. Scripts to reproduce R analyses are reproduced in Supporting Information.

To define the circadian proteome of free‐running SCN tissue, we used SILAC‐MS. Through comparison of peptidergic MS spectra from a labeled “spike” reference control standard and unlabeled circadian samples (Figure 1A), and subsequent bioinformatic analysis, we ultimately identified 3182 SCN proteins. This identified proteome was refined further by applying the criterion that two of the three replicates from each circadian time point should provide reliable quantification of the protein. Consequently, 2082 proteins constituted the quantified proteome of SCN slices sampled over a full circadian cycle (Figure S1, Supporting Information S1). Quantified proteins were tested for significant (p < 0.05) circadian regulation by two parallel approaches. First, the RAIN‐algorithm (Thaben and Westermark 2014) was applied, producing a proteome of 304 rhythmic proteins (~15% rhythmic). Second, the JTK‐algorithm (Hughes et al. 2010) was applied, producing a proteome of 166 rhythmic proteins (~8% rhythmic). We generated a circadian proteome where the methods overlapped, representing 153 proteins from RAIN (~50%) and all but 13 JTK‐identified proteins (~92%). This circadian proteome constituted ~7% of the quantified proteome (Figure S1, Supporting Information S2). Clustering and visualization of the circadian proteome by raster plot revealed approximately equal phase distribution of peak protein expression throughout the circadian cycle (CT0/24 24%, CT6 21%, CT12 25%, and CT18 31%) (Figure 1B). This indicates that identification of the final circadian proteome did not result in any waveform bias.

Having assembled an SCN circadian proteome, we then performed bioinformatic analysis to reveal enriched processes, thereby pinpointing potential contributors to SCN circadian function. We focused on Gene‐Ontology terms (GO‐terms), specifically those that describe Molecular Function (MF). The top MF GO‐terms in our circadian proteome were related to actin, trafficking and transport, and cellular signaling pathways (Figure 1C, Supporting Information S3). These GO‐terms indicated that SCN cellular homeostasis is under circadian control. Further examination of enrichment of proteins related to the “calcium dependent‐protein binding,” “GTPase binding,” “ATP‐dependent protein binding,” and “S100 protein binding” indicated discrete cellular signaling pathways that may form inputs and outputs within the SCN network are also under circadian control.

To characterize this further, we focused on expression profiles of representative proteins. First, the synaptic protein Nsf, identified in a previous, smaller SCN proteome (Deery et al. 2009) was a circadian protein peaking around CT12 alongside the functionally related proteins Snap25 and Snap47 (Kadkova et al. 2019), which shared a common temporal profile (Figure 1D, Synaptic proteins). We then investigated profiles of representative proteins within other significant GO‐terms. Proteins involved in actin binding, encompassed by the two most‐significant GO‐terms “actin binding” and “actin filament binding” had staggered profiles: Gsn peaked sharply during the day at CT6 while, in contrast, Syne2 which binds actin filaments to the nuclear membrane and Fscn1 which bundles actin filaments (Adams 2004) (two potentially related functions) peaked at CT24/0 (Figure 1D, Actin binding). Together, these proteins suggest circadian dynamics to actin remodeling in the SCN as has been shown in fibroblasts (Hoyle et al. 2017). Next, we looked at proteins contributing to the “transmembrane transporter binding,” which may regulate SCN network activity (Patton, Morris, et al. 2023). The three representative proteins in this group: Arrb1, Fhl1, and Fkbp1a displayed differently phased circadian profiles, suggesting differential roles within the SCN clock. Arrb1, peaking at CT6, has been implicated in met‐Enk signaling where, in response to delta‐opioid receptor activation, it translocates to the nucleus to activate the Per2 promoter (Bigliardi et al. 2024) (Figure 1D, Transmembrane Transporter Binding). Accordingly, met‐Enk signaling in the SCN modulates light induced phase‐shifts (Tierno et al. 2002). Finally, we assessed the nucleocytoplasmic transporters (Figure 1D, Nucleocytoplasmic carrier activity). This included Tnpo, Xpo1, and Xpo5, whose knockdown alters circadian period in cell models (Korge et al. 2018). Strikingly, loss of Tnpo shortened circadian period by reducing the rate of Per1 accumulation in the nucleus (Korge et al. 2018). Consistent with this role, Tnpo peaks at CT6 coincident with when Per proteins are accumulating. Nevertheless, the phase staggering of Xpo1 and Xpo5 relative to each other and to Tnpo suggests that they play different roles in clock function (Figure 1D, Nucleocytoplasmic carrier activity). Collectively, the differential temporal dynamics of these proteins indicate that even when their biological function is similar, they can occupy different temporal niches.

The cellular SCN network is heterogeneous, including neurons and astrocytes (Morris et al. 2021; Wen et al. 2020; Xu et al. 2021). Our SCN circadian proteome gave us the circadian profile of SCN proteins but lacked the cellular resolution to identify where in the network these proteins were expressed. To map our proteomic dataset to particular cell‐types, we utilized a published single‐cell transcriptomic dataset produced from similar free‐running SCN slices (Morris et al. 2021). We first clustered this dataset before manually annotating the cell‐types of clusters based on their top identifying genes. This allowed us to identify cell‐types (Figure 2A) with the transcriptomic profiles (defined previously; Morris et al. 2021; Wen et al. 2020) of: SCN neurons (Slc32a1, Nms, Avp, and Vip), extra‐SCN neurons (Slc17a6, Th, Sst, and Agrp), astrocytes (Gfap, Aldh1l1, Aqp4, and Sox9), ependymocytes (Tmem212 and Tctex1d4), oligodendrocytes (Mog and Plp1), NG2‐cells (Pdgfra), radial glia (Ccnb1 and Ube2c), tanycytes (Col23a1), microglia (Ly86, C1qa, and Hexb), endothelial cells (Lum and Dcn), and remnant unclassified cells (exclusion of other markers).

Having clustered SCN cell‐types via transcript expression, we applied an unbiased bioinformatic approach across the cellular clusters to determine whether any cell‐types aligned with the MF GO‐terms from the proteomic analysis. Analyzing the defining 250 genes in each cluster (Supporting Information S4), we plotted the top 2 GO‐terms for each cell‐type, resulting in 19 (partially redundant) examined GO‐terms (Figure 2B). Of interest, only one of the top cell‐type GO‐terms was also found in the circadian proteome: “S100 protein binding,” which to declining degrees of significance was specific to astrocytes, tanycytes, microglia and endothelial cells. Moreover, the astrocytes and tanycytes (two related glial‐cell types; Clasadonte and Prevot 2018) showed more significant enrichment. Given that astrocytes are known to actively participate in SCN network‐level timekeeping (Brancaccio et al. 2017, 2019; Patton, Morris, et al. 2023; Patton et al. 2022), we focused on them. When GO‐term analysis was applied to astrocytes independently of all other cell types, “S100 protein binding” was the fifth most significantly enriched MF GO‐term (Figure 2C, Supporting Information S5, adjusted p‐value < 0.0001), thereby describing one of the defining transcriptomic features of SCN astrocytes.

Having identified astrocytes as the cell‐type enriched in “S100 protein binding” related transcripts, we looked to combine our two‐omics data sets by comparing proteins identified as circadian (153) with astrocyte enriched transcripts (1292). This revealed 33 circadian proteins with transcripts enriched in the astrocytic transcriptome (Figure 2D). We then carried out our analysis on these 33 proteins/transcripts with “S100 protein binding” being the third most significant MF GO‐term (Figure 2D, Supporting Information S6). Between the two datasets, this was driven by the presence of S100 protein family members (S100a1, S100a4, and S100a6) and their interaction partners (Anxa2 and Ezr) (Figure 2D). Interestingly, annexin‐A2 (Anxa2) appeared as a nexus between seven of the top 10 MF GO‐term nodes, contributing to enrichment of: “S100 protein binding,” “metal‐ion transmembrane transporter activity,” “calcium‐dependent protein binding,” “calcium‐dependent phospholipid binding,” “phosphoinositol‐(4,5)‐bisphosphate (PI (4,5)P₂) binding,” “phosphatidylserine binding,” and “modified amino acid binding” (Figure 2D).

To explore the possibility that Anxa2 and its interactors may be important to SCN astrocytic circadian function, regardless of their presence in the circadian proteome, we next assayed the cellular transcript expression of Anxa2 and all of its interaction partners, as defined by BioGRID (Oughtred et al. 2021; Stark et al. 2006) (version 4.4). This highlighted 34 potential genes, including three members of the S100 protein family (S100a6, S100a10, and S100a11) that could encode known interaction partners of Anxa2. When plotted as a heatmap, although some of these proteins were restricted to other SCN cell‐types, only about a third of the potential interaction partners were enriched in astrocytes alongside Anxa2, positioning them as the cellular site of Anxa2 action within the SCN and narrowing the pool of potential interactors (Figure 2E).

As the Anxa and S100 families were identified as important proteins/genes at both proteomic and astrocytic transcriptomic levels, we assessed the temporal protein expression profiles of these family members. Anxa2, 6, and 7 were present in our proteome, along with S100a1, 4, and 6 (Figure 3A). Interestingly, these all showed the same circadian profile, peaking in the middle of the circadian day at CT6 despite only S100a6 being a potential interaction partner (Figures 2E and 3A). To assess a functional role for Anxa2 in the SCN astrocytes, we then re‐examined our pool of potential interaction partners (Figure 2E). Among Anxa2 interaction partners enriched within the SCN astrocytes, co‐expression of Anxa2 and S100a10 stood out due to previous work implicating them as partners controlling a diverse range of cellular processes including exocytosis, ion channel translocation and modulation, and histone modification (Bharadwaj et al. 2021). Indeed, this interaction is so strong that S100a10 protein stability is dependent on interaction with Anxa2 (Puisieux et al. 1996). Although S100a10 was in our identified proteome, it was not quantified due to our initial stringent quality control.

To validate SCN expression of S100a10, and to show that Anxa2 and S100a10 are expressed in the same cells (and thereby positioned to interact), we queried the transcriptomic data for cellular co‐expression of these transcripts. Out of 29,822 cells across the transcriptomic data, we found that 18,969 cells expressed neither S100a10 nor Anxa2 with the remaining cells expressing either Anxa2 or S100a10 alone (1459 and 6745, respectively). Only 2649 cells co‐expressed Anxa2 and S100a10 together, and the majority of these were astrocytes (Figure 3B), with a smaller representation from ependymocytes. Given that half (50.7%) of the astrocytes co‐expressed Anxa2 and S100a10 in the scRNA‐seq dataset, we performed in situ hybridization on SCN slices to confirm this astrocytic co‐expression, using Sox9 as a specific astrocytic marker (Endo et al. 2022; Sun et al. 2017). Importantly, these three transcripts co‐localized, demonstrating Anxa2 and S100a10 co‐expression in individual SCN astrocytes (Figure 3C, Figure S2).

Cellular translocation of Anxa2 to cell membranes is Ca²⁺‐ and S100a10‐dependent, being mediated by the formation of a hetero‐tetramer (Svenningsson and Greengard 2007): A2–A10. A2–A10 is Ca²⁺‐sensitive, and among other roles recruits and modulates various ion channels at the cell membrane, possibly to regulate ion flux (Gerke et al. 2005). This therefore positions the A2–A10 complex both as a potential sensor and as a regulator of cellular calcium levels ([Ca²⁺]_i). Formation of the complex relies on accommodation of the N‐terminus of Anxa2 into a hydrophobic pocket on S100a10, an interaction that can be pharmacologically disrupted by the compound A2ti‐1, which binds to this pocket, preventing hetero‐tetramerization (Reddy et al. 2012). It is therefore a highly potent and specific inhibitor of A2–A10 association and so we used it to disrupt A2–A10 in SCN slices and thereby explore the contribution of A2–A10 to SCN circadian function.

The first target of interest was [Ca²⁺]_i in SCN astrocytes, which reflects cellular activity and is under circadian control, peaking in circadian night when neuronal [Ca²⁺]_i is low (Brancaccio et al. 2017). To monitor [Ca²⁺]_i in astrocytes, we expressed the calcium reporter jGCaMP7b under control of the astrocyte‐specific promoter GFAP via AAV vector (Lohr et al. 2021). Confocal microscopy confirmed jGCaMP7b signal localized to SCN astrocytes, as evidenced by the characteristic stellate cellular morphology (Figure 4A). Slices were also transduced with AAV Syn.jRCaMP1a to monitor neuronal [Ca²⁺]_i. [Ca²⁺]_i levels in astrocytes and neurons were both highly circadian (Figure 4B), with common periods of 25.0 ± 0.2 and 24.8 ± 0.1 h, respectively (paired two‐tailed t‐test t (9) = 1.04, p = 0.32, Figure 4C) but different phases: astrocytic and neuronal peaks were ~12 h apart, as reported (Brancaccio et al. 2017). We have shown previously that neuronal [Ca²⁺]_i peaks at ~CT6 (Patton, Morris, et al. 2023), which provided a phase reference that assigned the peak of the astrocytic [Ca²⁺]_i rhythm to CT17.1 ± 0.5 h (Rayleigh test, R = 0.92, p < 0.0001, Figure 4D), again confirming the astrocytic origin of the signal.

We then treated reporter‐expressing slices with vehicle or 80 μM A2ti‐1 (Figure 4E,F, Figure S3). As expected, vehicle treatment had no effect on the oscillations, whereas 80 μM A2ti‐1 markedly disrupted the rhythms in [Ca²⁺]_i in both cell types (Figure 4E,F). Treatment with A2ti‐1 caused an immediate and sustained elevation of astrocytic [Ca²⁺]_i, (repeated‐measures two‐way ANOVA: treatment effect F (1, 5) = 16.34, p = 0.01; time effect F (8, 40) = 8.41, p < 0.0001; interaction F (8, 40) = 6.10, p < 0.0001) (Figure 4G). In contrast, neuronal [Ca²⁺]_i was suppressed by A2ti‐1 (repeated‐measures two‐way ANOVA: treatment effect F (1, 5) = 13.97, p = 0.018; time effect F (8, 40) = 10.5, p < 0.0001; interaction F (8, 40) = 9.13, p < 0.0001) (Figure 4H). Compared to the rapid effect seen in astrocytes, however, the suppression of neuronal [Ca²⁺]_i, was progressive, moving out of the operating range of vehicle treatment only after 4 cycles (Figure 4H). In both cell‐types, the circadian regulation of [Ca²⁺]_i was rapidly (1–2 cycles) damped, notwithstanding the overall elevation or decline. Thus, blockade of formation of the astrocytic A2–A10 complex disrupted SCN timekeeping across the cellular network, suspending circadian oscillations and functionally pushing astrocytic and neuronal calcium‐reported metabolic states away from one another and to their respective “night‐time” states.

Given the immediate disruption of astrocytic [Ca²⁺]_i levels and delayed neuronal disruption, we then tested the effect of A2ti‐1 on overall SCN network time‐keeping using the Per2::Luciferase (Per2::Luc) translational reporter of the TTFL. To confirm specificity of A2ti‐1 effects, we used a second compound, A2ti‐2, which shares the chemical structure of A2ti‐1 but lacks an ethyl group. This leads to a 10‐fold reduction in potency (A2ti‐1 IC₅₀ ≈ 24 μM; A2ti‐2 IC₅₀ ≈ 230 μM) and so provides a valuable control treatment for any off‐target effects of A2ti‐1 in our tested concentration range (Reddy et al. 2012). Bioluminescence from SCN explants expressing Per2::Luc was recorded before and during chronic treatment with either vehicle, A2ti‐1 or A2ti‐2 at concentration ranges between 20 and 80 μM and also after wash‐out to monitor the reversibility of any effects (Figure 5A). Treatment with vehicle or A2ti‐2 had no obvious effect on the Per2::Luc circadian oscillation. In contrast, A2ti‐1 severely and dose‐dependently reduced the amplitude of the ongoing oscillation within 1 circadian cycle, down to 15% ± 4% of the pre‐treatment amplitude (Figure 5A,B) (mixed‐effects two‐way ANOVA: dose effect F (3, 26) = 24.67, p < 0.0001; treatment effect F (1, 9) = 49.65, p < 0.0001; dose × treatment interaction F (3, 26) = 9.67, p = 0.0002). To confirm the specificity of the compounds, we performed combined treatments with A2ti‐1 and A2ti‐2 and saw no blockade or enhancement of the suppression seen with 40 μM A2ti‐1 when 40 μM A2ti‐2 was also present (Figure S4). Accompanying the suppressed amplitude, A2ti‐1 also dose‐dependently reduced the quality of the ongoing oscillation (assessed by relative amplitude error [RAE]) (Figure 5C) (mixed‐effects two‐way ANOVA: dose effect F (3, 26) = 11.55, p < 0.0001; treatment effect F (1, 9) = 10.15; p = 0.01, dose × treatment interaction F (3, 26) = 7.49, p = 0.0009). Again, A2ti‐2 was without effect on the oscillation. It should be noted, however, that whereas either A2ti‐1 or vehicle treatment did not alter period, there was a dose‐dependent increase in period up to 0.77 ± 0.06 h with A2ti‐2 (Figure 5D) (mixed‐effects two‐way ANOVA: dose effect F (3, 26) = 2.58, p = 0.08; treatment effect F (1, 9) = 30.33, p = 0.0004; dose × treatment interaction F (3, 26) = 5.25, p = 0.006). We interpreted this period‐lengthening by A2ti‐2 as a potential off‐target effect of the drug.

These results show that pharmacological disruption of the astrocyte‐restricted A2–A10 complex had profound impact on SCN timekeeping, but importantly, was reversible, as evidenced by the return of high‐amplitude Per2::Luc rhythms following washout of the compound under all doses tested (Figure 5A,E) (mixed‐effects two‐way ANOVA: dose effect F (3, 34) = 0.69, p = 0.56; treatment effect F (1, 34) = 0.001; p = 0.97, dose × treatment interaction F (3, 34) = 0.19, p = 0.90). Interestingly, although the oscillation returned to expected levels, this was not immediate: there was a progressive increase in amplitude following withdrawal of A2ti‐1 (Figure 5A,F). We therefore explored this formally, using a ratio of the amplitude of the rising slope of the oscillation to that of the falling slope as a metric to assess progressive re‐establishment of the oscillation. This revealed significant differences in the re‐establishment of stable oscillation across the doses of A2ti‐1 compared to vehicle measurements (Figure 5F) (mixed‐effects two‐way ANOVA: cycle effect F (4, 20) = 15.55, p < 0.0001; dose effect F (3, 15) = 14.12; p = 0.0001; cycle × dose interaction F (12, 40) = 5.33, p < 0.0001). Following 40 μM A2ti‐1, the oscillation recovered completely after 2 cycles, whereas after 80 μM A2ti‐1, it took 4 cycles to recover. To examine this more clearly, we compared the ratios of the second peak post‐washout (to account for washouts where the preceding trough was missing) for slices after treatment with vehicle, 80 μM A2ti‐1 or 80 μM A2ti‐2 (Figure 5G) (mixed‐effects one‐way ANOVA, F (2, 17) = 12.63, p = 0.0004). As there was no significant difference between this measure for vehicle and A2ti‐2 (Tukey's multiple comparisons, vehicle vs. A2ti‐2, 0.99 ± 0.002 au vs. 0.96 ± 0.02 au, p = 0.61), this effect was linked to the amplitude recovery following washout of A2ti‐1 and not off‐target effects of the compound. We previously interpreted this progressive “winding up” of the oscillation following drug treatments targetted at astrocytes as a reflection of an iterative reciprocal reinforcement between the TTFLs of the astrocytes and neurons, each providing a progressively stronger signal as network recovery evolved. The current observations are therefore consistent with disruption of an astrocyte‐specific complex compromising network function, with its recovery in astrocytes on wash‐out followed by progressive re‐integration across the network (Patton, Morris, et al. 2023).

To put this recovery into context and probe the time‐course of network sensitivity to A2ti‐1, we performed further experiments in which slices were treated with A2ti‐1 for different durations ranging from 2 to 7 days before washout and then monitored the time‐course of recovery (Figure 5H). This revealed a significant effect of treatment length on the oscillation dynamics following washout (repeated‐measures two‐way ANOVA: treatment effect F (1, 30) = 52.00, p < 0.0001; treatment length effect F (5, 30) = 4.95, p = 0.002; interaction F (5, 30) = 4.61, p = 0.003) (Figure 5H). Furthermore, post hoc multiple comparisons testing revealed that the washout effect was only manifested once SCN were exposed to A2ti‐1 for more than 4 days (Figure 5H). This is consistent with the [Ca²⁺]_i data, whereby changes in neuronal [Ca²⁺]_i were delayed and moved significantly out of their operating range only after 4 cycles of treatment (Figure 4H). These washout dynamics indicate that progressive amplification of the oscillation is likely caused by a cycle‐to‐cycle reversal of the dissociation of neurons and astrocytes within the network.

This effect, driven specifically by disruption of astrocyte function, was confirmed by the phase‐dependent effect of A2–A10 hetero‐tetramer disruption on the acute amplitude of the PER2::Luc report (Figure 5I). Throughout our experiments, 80 μM A2ti‐1 treatment started at different circadian phases and so we performed a meta‐analysis of slices treated in dose–response and treatment length experiments. Analysis of this pooled data revealed bidirectional phase‐specific changes to the amplitude of the first peak post‐treatment (Linear Mixed Model: Treatment effect F (1, 51) = 3.95, p = 0.05; Phase effect F (3, 51) = 12.97, p < 0.0001; Interaction F (3, 51) = 14.37, p < 0.0001) (Figure 5I, Figure S5). Importantly, post hoc analysis revealed that this phase‐specific effect manifests as acute amplitude reductions when drug was applied at phases during which astrocytic [Ca²⁺]_i is rising or peaking (Figure 5I: CT6‐12 Veh 0.93 ± 0.01 au vs. A2ti‐1 0.37 ± 0.18 au, p = 0.0001; CT12‐18 Veh 0.89 ± 0.004 au vs. A2ti‐1 0.46 ± 0.08 au, p = 0.0002) while treatment during phases when astrocytic [Ca²⁺]_i is falling or at low levels either enhanced the following peak or was without effect (Figure 5I: CT0‐6 Veh 0.88 ± 0.02 au vs. A2ti‐1 1.20 ± 0.10 au, p = 0.001; CT18‐24 Veh 0.87 ± 0.01 au vs. A2ti‐1 0.88 ± 0.08 au, p = 0.96). This pattern of acute suppression of the amplitude of the oscillation is comparable with astrocyte‐specific disruption via the glial metabolic toxin fluorocitrate, where CT6‐18 is acutely sensitive to treatment (Patton et al. 2022).

Due to the observed phase‐dependence of changes in the TTFL amplitude, we examined whether A2ti‐1 treatment altered the phase of the ongoing oscillation (Figure 5J). Across these experiments, where we could accurately quantify the time of the second peak (88/89), we saw changes in phase that were specific to the phase of A2ti‐1 application (Linear Mixed Model: Treatment effect F (1, 50) = 9.96, p = 0.003; Phase effect F (3, 50) = 12.88, p < 0.0001; Interaction F (3, 50) = 9.96, p < 0.0001). These phase‐shifts were mainly phase delays occurring in circadian night, with a few advances in mid‐to‐late circadian day, as revealed by post hoc analysis (Figure 5J: CT0‐6 Veh 0.05 ± 0.08 h vs. A2ti‐1 0.54 ± 0.63 h, p = 0.54; CT6‐12 Veh 0.07 ± 0.08 h vs. A2ti‐1 2.27 ± 1.44 h, p = 0.02; CT12‐18 Veh −0.06 ± 0.21 h vs. A2ti‐1 −3.23 ± 0.75 h, p = 0.0003; CT18‐24 Veh −0.11 ± 0.10 h vs. A2ti‐1 −2.79 ± 0.61 h, p = 0.0001). This is the first evidence that direct manipulation of an endogenous astrocytic signaling pathway can alter the ongoing phase of the SCN TTFL oscillation (Patton et al. 2022), elevating the participation of astrocytes into a facet of SCN network level timekeeping beyond the modulation of period and/or amplitude.

The mass spectrometry proteomics data have been deposited to the ProteomeXchange Consortium via the PRIDE (Perez‐Riverol et al. 2025) partner repository with the dataset identifier PXD061617. Single‐cell RNA sequencing data reanalyzed in this manuscript (Morris et al. 2021) are publicly available from the NCBI Gene Expression Omnibus with the accession number GSE167927↗.