Cortical state, defined by population-level neuronal activity patterns, determines sensory perception. While arousal-associated neuromodulators—including norepinephrine (NE)—reduce cortical synchrony, how the cortex resynchronizes remains unknown. Furthermore, general mechanisms regulating cortical synchrony in the wake state are poorly understood. Using in vivo imaging and electrophysiology in mouse visual cortex, we describe a critical role for cortical astrocytes in circuit resynchronization. We characterize astrocytes’ calcium responses to changes in behavioral arousal and NE, and show that astrocytes signal when arousal-driven neuronal activity is reduced and bi-hemispheric cortical synchrony is increased. Using in vivo pharmacology, we uncover a paradoxical, synchronizing response to Adra1a receptor stimulation. We reconcile these results by demonstrating that astrocyte-specific deletion of Adra1a enhances arousal-driven neuronal activity, while impairing arousal-related cortical synchrony. Our findings demonstrate that astrocytic NE signaling acts as a distinct neuromodulatory pathway, regulating cortical state and linking arousal-associated desynchrony to cortical circuit resynchronization.

Recording setup

Animals were given at least 1 week after surgery for recovery and viral expression. They were then habituated on a custom-made circular running wheel over at least two days, and for a cumulative time of at least 2.5 hours, before experimental recordings began. After habituation, mice were head-fixed on the wheel and movements were recorded by monitoring deflections of colored tabs on the edge of the wheel using an optoswitch (Newark, HOA1877-003). 

Pupillometry

Pupil recordings were made using a Genie near-infrared (NIR) camera (1stVision, M640), a telescopic lens (Thorlabs, MVL50TM23), and acquired at 30Hz using the MATLAB Image Acquisition toolbox. A small monitor (Amazon, B06XKLNMW3) showing a consistent teal background color (RGB: 0,1,1) was placed by the mouse to allow for observation of the full range of pupil dynamics in an otherwise dark room. For experiments without imaging, a NIR light (Amazon, B00NFNJ7FS) was used to visualize the pupil.

Two-photon imaging

Two-photon imaging was performed on a microscope (Bruker) with two tunable Ti:sapphire lasers (MaiTai, SpectraPhysics) and a Nikon 16x, 0.8 N.A. water-dipping objective with a 2x-optical zoom (frame rate: 1.7Hz, FOV: 412µm2, resolution: 512x512 pixels). 950nm excitation light with a 515/30 emission filter was used to image green-emitting fluorophores, and 1040nm light with a 605/15 emission filter was used to image red-emitting fluorophores. Recordings lasted from ten minutes to one hour.

Electrophysiological recordings

Visual cortex LFP was recorded at 1KHz and subtracted from the ipsilateral cerebellar LFP before 1KHz amplification (Warner Instruments, DP-304A) and acquired using PrairieView (Bruker) or PackIO.

In vivo pharmacology

Recordings were taken before and after saline, Prazosin-HCl (5mg/kg Sigma-Aldrich, P7791-50MG), A61603(1µg/kg or 10µg/kg, Tocris, 1052), or clozapine N-oxide (CNO, 1mg/kg or 5mg/kg, Tocris, 4936) were injected intraperitonially while animals remained head-fixed on the wheel, to ensure post-treatment recordings were comparable with baseline measures.

Data analysis

All data analysis was done in MATLAB unless otherwise indicated. No statistical methods were used to pre-determine sample sizes, but they are similar to previous reports. Boxplots are shown with the central mark indicating the median and the bottom and top edges of the box  indicating the 25th and 75th  percentiles, respectively. Whiskers extend to the most extreme data point or within 1.5 times the interquartile range from the bottom or top of the box, and all other data are plotted as individual points, as listed in the figure legends. For statistical comparisons, non-parametric tests were used, or where indicated, normality was assumed but not formally tested, and t-tests were used. Hierarchical bootstrapping was performed based on a MATLAB implementation (https://github.com/jenwallace/Hierarchical_bootstrap_Matlab) of the methodology, and used in order to reduce the statistical error rate of comparisons while retaining statistical power. All multiway comparisons were adjusted for using Tukey-Kramer correction. No data was excluded from analyses except for the following (not pre-determined): In hSYN-hM4Di experiments, outliers were excluded across all conditions from small stationary responses to avoid confounding effects from other influences on astrocyte Ca2+, as described in methods. For in vivo pharmacology experiments, electrical artifacts in band power were excluded before analysis, as described in methods. Samples were allocated into experimental groups by cell-type expression of each individual fluorescent sensor. Only adult animals (1–6 months of age) were used in experiments, and both male and female were used and randomly selected. For imaging and electrical recordings of spontaneous activity, blinding was not relevant because cell-type viral expression is evident from expression pattern. For in vivo pharmacology, blinding was not possible because control recordings were taken prior to treatment recordings to avoid confounding the treatment effects. For Adra1afl/fl mice, the experimenter was blinded to genotype before data collection and analysis.

Speed calculations

To compute wheel speed, a detected break in the optoswitch circuit was determined when the absolute value of the derivative of the raw voltage trace was at least 2 standard deviations above the mean. For recordings with very little movement (std < 0.1), this threshold generated false positives so a set threshold of 0.1 was used. The number of breaks in the optoswitch circuit per second was then calculated, and using the circumference and number of evenly spaced colored tabs at the edge of the wheel, the wheel speed was determined. Movement periods were defined by wheel speed ≥ 10 cm/s and movement bouts that were separated by ≤ 2 s were considered one event. To ensure that movement related dynamics were not included in stationary analysis, data was excluded from at least ten seconds around identified movement periods.

Pupillometry

Following acquisition, pupil data was processed through a Python function which used contrast detection to identify the edges of the backlit pupil from the sclera and fit an ellipse whose major radius was taken as the pupil diameter. The diameter was then low-pass filtered to 0.5 Hz and normalized to a range between 0 and 1 to give pupil diameter as percent maximum. The pupil derivate was normalized to the acquisition rate (30Hz) to compute pupil phase and to determine the phase of astrocyte Ca2+ events and the cross-correlations with GRABNE. Stationary arousal dilations and constrictions were identified by the sign of the calculated pupil derivative and only changes in pupil diameter > 10% were used for subsequent analyses.

LFP

All spectral analysis was done using Chronux. Raw LFP data was visually inspected to confirm useable signal was present, and then 60 Hz noise was filtered out and drifting baselines were compensated for using linear fitting. LFP power for frequency bands was computed using built-in Chronux functions with a time-bandwidth of 2.5 and two tapers, no frequency padding, and 5 s moving windows. For changes in LFP band power around arousal or astrocyte Ca2+ events, the median band-limited power was obtained and then normalized to the median band-limited power in the event-triggered time window to get relative band power. The median power before an event onset vs after was combined for each recording, and for Adra1afl/fl mice, the ratio between the two was computed and compared to Adra1a wild-type mice. For changes in LFP power after A61603 administration, the band-limited power for saline and drug data were calculated, outliers were removed to avoid contamination by recording artifacts, and then this power was normalized to the band-limited power for the respective baseline recording.

For total power in Adra1afl/fl and control mice, no baseline correction was done to avoid skewing the analysis. Spectrum power was calculated by concatenating recordings from all mice of each genotype and computing average or individual spectra with a time-bandwidth of 6 and 8 tapers to increase accuracy. The total power was then computed by summing across all frequency bands. Relative power was computing by dividing the spectrum from each mouse by its total power, and relative band power was computed by summing the power from each frequency band and dividing by the total power.

Ca2+ imaging

Astrocyte Ca2+ events and fluorescence was extracted using the AQuA software analysis package. For dual-color imaging with neuronal Ca2+ indicators, particular care was taken to avoid AQuA detection of neuronal activity; the standard deviation of the neuronal channel was taken in FIJI and a mask was created in AQuA to exclude areas of high neuronal activity and soma from analysis. Astrocyte events were only included if they had an area greater than 10 µm, lasted for at least 2 frames, and had an AQuA p-value < 0.05. To obtain the average astrocyte Ca2+ fluorescence, the compensated fluorescence traces which account for spatially overlapping events was taken, normalized to their maximum value, and then averaged together.

Neuronal Ca2+ events and fluorescence were extracted from neuropil background semi-automatically using Suite2P95. We identified Ca2+ events by taking identified spikes in the Ca2+ fluorescence data and thresholding them for only the largest (> 3std over the mean) events. To calculate the average neuronal Ca2+ fluorescence, the trace from each neuron was normalized to its maximum value and averaged together.

Machine-learning based analysis of input contribution to astrocyte Ca2+

As an alternative to assess the contribution of biological inputs on astrocyte Ca2+, data from dual-color Ca2+ imaging was used to train a machine-learning model to predict average Ca2+ activity. To include LFP data, the spectrogram data from each LFP recording was decomposed using Principal Component Analysis based on the eigenvectors from the ipsilateral recording that accounted for the largest proportion of the variance. The PC1 in this data corresponded to cortical synchrony, with positive weights for high frequency and negative weights for low frequencies, matching a previous report but with inverse sign. LFP PC1 for ipsilateral and contralateral recordings, as well as speed, pupil diameter, and average Ca2+ fluorescence for neurons and astrocytes, was then z-scored and resampled to 10Hz before being concatenated. This dataset was then imported into Python for machine learning analysis using the SciKit-learn toolbox. For analysis, randomly generated data was added for comparison, and rows without both ipsilateral and contralateral LFP recordings were excluded from subsequent analysis. Average astrocyte Ca2+ data was used as the target dataset and data was split into training (80%) and testing (20%) before classification using a Random Forest Regression model. The model was validated using the R2 between the predicted average astrocyte Ca2+ fluorescence from the model and the actual average astrocyte Ca2+ data of the test set. Permutation testing of the predictors was then used to determine their relative contributions to model prediction.

GRABNE Analysis

In GRABNE imaging data, we observed background fluorescence fluctuations which we thought might arise from hemodynamic artifacts. To ensure the data reflected the NE signal, hemodynamic artifacts in the data were removed by a custom-designed, data processing pipeline. The predominant hemodynamic artifact in the data was assumed to reflect fluctuating hemoglobin levels altering brain absorptivity causing an attenuation of light. As such, the signal from each pixel could be modeled as where k is index of pixel,  and  are, respectively, the observed curve and the real fluorescence of kth pixel,  is the absorption coefficient for kth pixel,  is the path distance, the term  represents the intensity attenuation, and N(t) is the noise. In our data, the identified hemodynamic signal across pixels was approximately synchronous, but varied in magnitude, thus the attenuation of one pixel can be represented by another, that is  Based on the findings above, we designed the following pipeline:

1. We selected one connected vascular region with minimal fluorescence in the average projection of the data and calculated an initial vascular reference curve. This region was assumed to have the lowest possibility to contain any true GRABNE signal.

2. To avoid compensating for slow changes in the true GRABNE signal, we subtracted the curve of each pixel by a 100-frame moving average.

3. We applied linear regression and fit the logarithm of the processed curve to the initial vascular reference curve. The exponential of the fit data was then taken to represent the hemodynamic effect for each pixel. To account for cases where the initial reference curve was contaminated, we iteratively refined the reference curve before fitting, that is, we calculated the weighted average of the original curve for all pixels (fitting parameter  in  for each pixel is considered as the weight) and subtracted its moving average.

4. We removed the hemodynamic artifact (if any) by dividing the raw pixel curve by the exponential of the fitting data.

Next, only the least (1–25%) hemodynamically affected pixels with the lowest a were taken, excluding the bottom 1% which often contained artifacts, and these were averaged together and used as the final GRABNE signal. For spectral analysis, each recording was concatenated together in ten-minute segments, padding with its median value if necessary, and then run in Chronux with no frequency padding, a time-bandwidth of three, five tapers, and passing frequencies above half the window size (3e-3Hz) and less than the Nyquist frequency (0.9Hz). To identify phasic increases in the GRABNE signal built-in MATLAB functions were used to determine local peaks in the signal and a prominence threshold was used to determine the different magnitudes of GRABNE increases.

Event triggered-averages

All averages were computed by identifying events (e.g., movement offset, pupil dilation, etc.) and taking data in a symmetric time window around the events. The data is subsequently plotted as the mean and standard error across all events, except for spectrograms where the median was used.

Correlations

For comparisons of maximum and change in astrocyte Ca2+/pupil/speed after arousal, values were computed for each trace separately and then linearly correlated with p-values describing the probability of a true R2 relationship between each two metrics. For correlations between astrocyte and neuronal Ca2+ activity, the average population fluorescence was taken for each and z-scored before cross-correlation. For behavioral state-separated cross-correlations, the same procedure applied, but only z-scored data from either moving periods, or stationary periods, was used. For correlations within neuronal and astrocyte populations, the pairwise correlation between each cell (neurons) or event (astrocytes) was computed, the symmetric and autocorrelations were excluded, and the overall mean was taken to obtain a single value indicating the synchrony of Ca2+ dynamics within each cellular population. For cross-correlations between pupil and imaging data (GRABNE and astrocyte Ca2+) all data was z-scored, resampled to either 10 or 30Hz, padded with nan values if unequal in length, and then cross-correlated. For cross-correlations between LFP band power and Ca2+ imaging data, all data was z-scored, averaged, and then resampled to 2Hz to match the LFP resolution before cross-correlation.

hSYN-hM4Di

To estimate the effect of CNO on astrocyte Ca2+, the average Ca2+ event properties and overall event rate for each recording were randomly sampled 104 times and CNO data was subtracted from corresponding baseline data. This procedure generated the range of treatment effects possible from the sampled data, and a p-value was calculated as the proportion of CNO difference from baseline that were less than the maximum, or greater than the minimum, difference found in saline conditions. For calculating the modulation of astrocyte Ca2+ responses to arousal, the absolute change in average astrocyte Ca2+ fluorescence after either movement onset or pupil dilation was determined. Outliers during small stationary responses which might reflect the influence of other variables on astrocyte Ca2+, were excluded. The magnitude of astrocyte Ca2+ responses to arousal after treatment was then compared to baseline responses.

Neuronal arousal PC analysis

PCA was done using the built-in MATLAB function on z-scored neuronal Ca2+ data. The pupil diameter and PC data was then resampled to an effective rate of 10Hz, and the Pearson’s correlation between the pupil diameter and each PC was used to identify the arousal PC for each recording This PC was then normalized to the maximum value before subsequent analysis. 

For comparisons between wild-type and Adra1afl/fl mice, the response to each movement onset or stationary pupil dilation was normalized to the median value in the window, and then the average arousal PC value during the event period was taken with a 2-frame offset to account for a slight lag in the arousal-associated neuronal response.

All data are two-photon imaging, electrophysiology, and associated physiology datasets. All data is formatted to allow for reconstructing the figures in Reitman et al. 2022 with limited user input needed. Data are stored as MAT-Files and the linked code was generated in MATLAB 2020a. The data for each figure is separated into individual MAT-Files, and should be placed in the same folder as the linked code.  Running the ‘MakeFigures.m’ file will result in all Paper figures being created, and to create individual figures one can run the ‘PlotFigX.m’ function within ‘MakeFigures.m’. Abbreviations used in file names: xcorr = cross-correlation; ETA = event-triggered average; floxed = Adra1a^fl/fl