Group ICA of the imaging dataset from three subjects in different brain states yielded independent components (ICs) confined to various regions of the mouse cortex. The ICs were registered on the Allen Brain Atlas (CCFv3) and tagged within the visual, somatosensory, retrosplenial, posterior parietal, and motor groups, based on the specific area, in the atlas, where the maximum value of each component was located (Figure 2A).

Using this approach, we identified 10 somatosensory ICs, 3 somatomotor ICs, 3 retrosplenial ICs, 2 posterior parietal ICs, and 3 visual ICs (Supplementary Figure S2). The reproducibility of the components across sessions of individual animals has been verified by measuring the correlation coefficients between the back reconstructed ICs extracted with group ICA and the ICs extracted by applying ICA separately to each session of each animal. The results of the comparison are presented in Supplementary Figure S3, showing a remarkable high agreement between the two sets of components on a session by session basis (Supplementary Figure S3A).

In addition to providing spatial maps, group ICA also generates time courses for each component (Figure 2B). In Figure 2, the time course of IC₄ in a representative subject is displayed, aligned with the global cortical activity during deep anesthesia and the awake state (Figure 2C).

To determine the contribution of each component to the cortical activity, we applied a threshold to the ICs’ time courses in the deeper anesthesia condition. We used this state of global inactivation as a reference point to identify significant activation of each component, represented by crossing the threshold (see methods). We observed that the number of points crossing the threshold was modulated by the level of anesthesia. Specifically, as the level of anesthesia decreased, there was an increase in overall neuronal activation, resulting in a drastic increase in the number of crossings for all ICs (Figure 2B).

Overall these results suggest that group ICA can be successfully applied to wide-field calcium imaging data yielding components that are stable across sessions of different animals and can be used to decompose cortical activity into several ICs whose time course is coupled to the global cortical activity.

Since each component showed an increased number of threshold crossings with decreasing levels of anesthesia, we aimed to investigate how the relative occurrence of each IC is modulated by the brain state. To achieve this, in each state, we examined the relative proportion of occurrences of each IC and compared it with the expected number of occurrences when considering either the deepest state of anesthesia or the awake state.

When arranging the components in descending order based on their relative frequency of occurrence (Figure 3A), we observed that each state was characterized by a distinct set of components whose relative occurrence surpassed the expected frequency (Figure 3B). In addition, the sorted probability distribution of all brain states exhibited a negative slope, indicating an uneven distribution of IC probabilities, implying that certain groups of ICs were more prevalent than others (Figure 3A).

The results revealed that lighter levels of anesthesia (iso 1–0.5%) exhibited a reduced number of prominent components compared to deeper anesthesia (iso 2–1.5%) and wakefulness. This finding suggests a more balanced contribution of the ICs during the oscillatory state. Moreover, deeper anesthesia states were denoted by the prevalence of retrosplenial and somatosensory components, while lighter anesthesia and wakefulness displayed a broader repertoire of component groups.

The uneven distribution of ICs probabilities arises from anesthesia-induced modulation of their relative occurrence. To assess which component class was modulated by the level of anesthesia we analyzed the proportional occurrence of each class from deep anesthesia to wakefulness. The findings revealed that the retrosplenial group exhibited a statistically significant brain state dependency [RM ANOVA F (5, 10) = 19.46, p < 0.001], whereas other classes showed no significant trends or a uniform distribution (Figure 4). A post hoc analysis revealed that the probability of occurrence of the retrosplenial group was significantly higher in the deepest level of anesthesia compared to the awake state [pairwise repeated t-test with holm correction, t (2) = 21.04, p = 0.03]. These results were also tested with respect to the total number of ICs selected for the group ICA analysis. Specifically, we repeated the same analysis selecting 16, 20, 24, 28, and 32 ICs. Supplementary Figure S5A shows that the result obtained using 24 ICs is preserved over the chosen range of ICs numbers supporting the stability of the result with respect to moderate changes in the chosen number of ICs.

These findings suggest that the relative contribution of the retrosplenial cortices to the overall activation can serve as a predictor of the brain state.

Our results showed that the ICs in the retrosplenial group exhibited a pronounced dependency on brain states, being prominent in anesthetized states and proportionally infrequent during wakefulness. To further investigate the role of the retrosplenial cortex, we measured the correlation among the time courses of ICs across states (Figure 5). Our findings indicated a decrease in overall correlation with diminishing levels of anesthesia. In addition, we observed the highest correlation values between ICs of the retrosplenial group during the deepest state of anesthesia (r = 0.80). These correlations significantly decreased during the awake states (r = 0.38) [RM ANOVA F (5, 10) = 39.4, p < 0.001] (Figure 5B). Finally, to ensure the robustness of our findings, we assessed the stability of the result with respect to a moderate variation in the number of ICs. Supplementary Figure S5B shows that the result obtained using 24 ICs is preserved over the chosen range of ICs numbers, supporting the stability of the result with respect to variation in the selected number of ICs.

These results show that brain states directly modulate the correlations among the time course of the ICs especially in the retrosplenial cortices.

All experiments conducted in this study adhered to the guidelines set forth by the Italian Minister of Health (authorization number 857/2021). Three C57BL/6 J adult mice (8–10 months old) of both sexes were utilized. The mice were housed in a controlled environment with regulated temperature and humidity and provided with unrestricted access to food and water.

Mice were anesthetized with isoflurane (3% for induction, 1–2% for maintenance) and placed in a stereotaxic apparatus (KOPF, model 1900) as previously described (Resta et al., 2022). Ophthalmic gel (Lacrilube) was applied to prevent eye drying, and body temperature was maintained at 36°C using a heating pad. Lidocaine 2% was used as a local anesthetic. The skin and periosteum were cleaned and removed, and bregma was marked with a black fine-tip pen. Transparent dental cement (Super Bond C&B – Sun Medical) was then applied to cover the entire exposed skull surface and to glue a custom-made aluminum head-bar, placed behind lambda. To achieve widespread expression of jRCaMP1b over both hemispheres, the viral construct ssAAV-PHP.eB/2-hSyn1-NES_jRCaMP1b-WPRE-SV40p(A) (Viral Vector Facility, Zurich, Switzerland) was systemically injected into the retro-orbital sinus. The virus titer was 1.5 × 10¹³ vg/ml and a total volume of 150 μL was injected.

After the surgery, mice were recovered in a temperature-and humidity-controlled room, with food and water ad libitum for 3 weeks before recordings. The intact skull cranial window was implanted 1 month after injection and imaging sessions started after 1 week after the window implantation to allow full recovery after the surgical operation. On day 7 post-surgery, mice were acclimated to the setup and habituated to being handled (20-min habituation session). In the subsequent week, mice were habituated to resting with their heads fixed under the microscope in a dark environment with white noise (two sessions of 20 min each).

Wide-field imaging was performed in a dark room using a custom-made microscope. The excitation source for jRCaMP1b was a red-light beam emitted by diodes (595 nm LED light, M595L3 Thorlabs, New Jersey, United States), and the excitation band was selected using a bandpass filter (578/21 nm, Semrock, Rochester, New York, United States). The light beam was directed to the objective (TL2X-SAP, Thorlabs, New Jersey, United States) through a dichroic mirror (606 nm, Semrock, Rochester, New York, United States). The fluorescence signal of jRCaMP1b was collected through a band-pass filter (630/69, Semrock, Rochester, New York, United States) and focused onto the sensor of a high-speed complementary metal-oxide semiconductor (CMOS) camera (Orca Flash 4.0 Hamamatsu Photonics, NJ, United States) using a tube lens (500 nm). The camera captured images at a resolution of 512 × 512 pixels, covering a square field-of-view of 11.5 mm. The sampling rate was 40 Hz.

We utilized a custom-made platform comprising a square tube to position the animals beneath the microscope objective and secure the metal post that had been implanted on their heads. In this condition, the animals rest in the objective focal plane with no ability to move around or run.

Anesthesia induction was initiated at a concentration of 3–4% isoflurane. Following induction, the animals were transferred to the experimental setup, and the isoflurane concentration was reduced to 2%. Before imaging, anesthesia levels were stabilized, with a waiting period of 10–15 min after any adjustments to the isoflurane concentration. Longer waiting times were required before imaging in the awake state for anesthesia recovery. Each recording session had a duration of 120 min. During the deep and intermediate anesthesia states (isoflurane range 2–1%), the animals did not express any behavior (i.e., detectable movements). In the lighter anesthesia state (isoflurane 0.5%), mice started exhibiting small whisker twitches. When awake, a complete repertoire of facial and limb movements was observed. The mice were head-fixed, and their bodies were constrained within a square-sized framework. Trained mice were habituated to this positioning and did not show large movements.

Brain slices for histological evaluation were prepared as previously reported (Resta et al., 2022). Briefly, mice were perfused with 0.1 M PBS and 4% paraformaldehyde. Brain slices (100 mm thick) were prepared using a Vibratome (Vibratome Series 1500—Tissue Sectioning System), washed, and incubated with NeuN antibody (1:200, Sigma, ABN78) overnight at 4°C. After washing, slices were treated with a fluorescent Alexa Fluor 514 antibody (1:250, ThermoFisher, A-31558) and mounted on glass slides.

Images were acquired using a two-photon microscope (Thorlabs Bergamo) equipped with a Ti-Sa laser (Coherent Chameleon) tuned to 1,064 nm for jRCaMP1b and 950 nm for NeuN (Alexa Fluor 514 reporter) a × 20/0.95 NA water immersion Olympus objective. Emission filters (Semrock) were 520/35 nm and 630/69 nm for NeuN and jRCaMP1b channels, respectively.

Brain state modulation was achieved by changing isoflurane concentration. Specifically, the experiments began at the deepest level of anesthesia with 2% isoflurane. Then, the anesthetic concentration was gradually reduced to 1.75, 1.5, 1.0, and 0.5%, allowing 10 min for the neuronal activity to stabilize at each concentration. Based on previous work (Song et al., 2018; Dasilva et al., 2021; Torao-Angosto et al., 2021; Tort-Colet et al., 2023), we selected these percentages to study the deep, intermediate and light levels of anesthesia. In fact, high concentrations of isoflurane (2–1.75%) stabilize global rhythmic activity (averaged across the entire field of view) at approximately 0.1–0.2 Hz achieving “deep” anesthesia level. Slight adjustments to the isoflurane concentration (1.5%) result in global activity exhibiting higher upstate frequencies of 0.2–0.4 Hz, categorizing recordings in this state as a medium anesthesia level. Finally, the light anesthesia level is achieved at lower concentrations, with global rhythmic activity showcasing prominent oscillatory dynamics. A complete recovery of characteristic behaviors such as motion, whisking, grooming, and liking was expected before the awake imaging session.

To ensure the consistency of the field of view across sessions we performed offline image registration as previously described in Scaglione et al. (2022). Briefly, each frame of the fluorescence data was first downsampled with a factor of 4 resulting in a pixel size of about 90 μm and then offline registered by aligning each frame to two reference points (corresponding to bregma and lambda) that were previously marked on the dental cement during the surgery procedure. After the image registration across all days, subjects and groups, we kept only the regions in the field of view that were common to all sessions (Figure 2A). The more marginal regions were excluded from the analysis and thus the final area analyzed contains at most a very small portion of the lesion. To be able to compare changes in fluorescence across subjects and sessions, the raw image stacks, collected with the wide-field microscope, were expressed in percent changes (∆F/F₀) with respect to its mean session fluorescence value (F₀) for each individual pixel. To remove slow drifts, the ∆F/F₀ was then high-pass filtered (0.1 Hz) setting the bandwidth of the analyzed fluorescence signal from 0.1 to 20 Hz.

The normalized fluorescence signals from different levels of anesthesia, from iso 2% to awake, were then concatenated together into a single session (Figure 1B). In this way, a total of 6 sessions (2 recordings per animal) constitute the dataset analyzed with group ICA.

To be able to decompose the fluorescence signal in independent components (ICs) across subjects and sessions, we decided to employ the group ICA analysis, originally developed to make group inferences in the context of fMRI data. The main strength of this method lies in the extension of ICA from a single subject to a group of subjects. This is achieved by utilizing a series of data reduction steps that compresses data in the temporal domain, thereby lowering the computational requirements of ICA for the full dataset. In this way, the ICs found do not rely on any model of the neural dynamics since it uses ICA. Additionally, the same components are already generalized across subjects and sessions. Moreover, using this approach, once the ICs are found through ICA, the spatial representation of the component and its time course for a specific session can be expressed as a series of matrix multiplication steps (Calhoun et al., 2001), the back reconstructed components.

We used the GIFT toolbox¹ to carry out the group ICA analysis on the 6 recording sessions to extract the group ICA ICs (see Supplementary Table S1 for the list of options and parameters used to run the GIFT toolbox). ICASSO (Himberg et al., 2004) was then used to ensure the reliability of the ICs. In fact, ICASSO runs the ICA algorithm a given number of times by bootstrapping the data and or changing the initialization parameters of ICA between runs. Subsequently, the estimates obtained from each run are grouped based on mutual similarities through agglomerative clustering. A reliable estimate of the ICs is then obtained by using the centrotype of each cluster generated by the different ICA runs.

The number of ICs was selected as the number that minimizes the R-index which provides a measure for the compactness and separation of the clusters found when running ICASSO. In fact, while ICASSO can be used to estimate reliable ICs by selecting the centrotype of the cluster generated by each individual run of the ICA, as pointed out by Li et al. (2007), the compactness of the clustering performed during ICASSO can be used to validate the dimensionality or the order selection of the ICA. In fact, the stability of the ICs is closely related to the order selection. Li et al. (2007) show that by increasing the number of components, the compactness of the clustering deteriorates leading to overlapping clusters or, in the case of very high dimensions, the different Monte Carlo runs lead to IC estimates that do not display significant correlations making it difficult to define reliable clusters of single ICs.

Once the ICs were identified, those components whose maximum activation was mostly confined over the midline and or showed clear vasculature were discarded as the change in fluorescence signal along the midline can be contaminated by hemodynamic or movement artifacts.

To check that ICs extracted with group ICA are present in each session from each subject, we compared the back reconstructed components of each session, IC^gICA_i (i = 0…23), with the ICs extracted running ICA on each single session separately, IC^ICA_j (j = 0…23), together with ICASSO to extract the most reliable components. We then computed the Pearson correlation coefficient between the two sets of components. Thus, for each session a correlation matrix R is obtained whose elements r_i,j represent the correlation coefficient between the i-th back reconstructed component, IC^gICA_i, with the j-th single ICA component, IC^ICA_j. A correlation coefficient of 1 is obtained when two components have the same weights up to a multiplicative constant. Finally, columns of each correlation matrix are rearranged according to a rectangular assignment algorithm which associates the component on the back reconstructed ICA to at most one element of the single session ICA components based on the maximum value of the |R|.

Once the global ICs maps were obtained, the location of the pixel with the highest value in the map, independently of the sign, was marked. This location was used to name the component according to the Allen Brain Atlas. Finally, components were grouped into 5 different macro areas (first degree of division of the cerebral cortex according to the Allen Brain Atlas) (Figure 2A): (1) Posterior Parietal Association; (2) Retrosplenial; (3) Somatomotor, (4) Somatosensory, and (5) Visual cortical areas. It is important to note that when we are trying to associate the ICs to the anatomical areas defined by the Allen Brain Atlas that are defined by the cytoarchitectonic organization of the cerebral cortex. We are resourcing to the Allen Brain Atlas to keep the same nomenclature and to group cortical parcels into general macro areas that are well accepted and understood.

To quantify the contribution of the ICs to the overall fluorescence signal during the different stages of anesthesia, the time course of every IC in each session was back-reconstructed through group ICA. To identify timepoints where the component displayed a significant change with respect to its background in the time domain, a threshold was set to three times the standard deviation of the time course of the IC during the deepest state of anesthesia (iso 2%) t h = ± 3 ∗ σ i i s o 2 % . A binary vector for each IC was obtained by setting 1 every time that the component time course surpassed the threshold and zero otherwise. The sum of each of these binary vectors, N_IC,ST, was defined as the number of activations and was used as a proxy to assess a significant impact of the IC on the ∆F/F₀ signal. Therefore, the relative proportion of the number of activation N_IC,ST/∑_ICN_IC,ST represents the proportion of timepoints that an IC significantly impacted the fluorescence signal compared to all other ICs for that particular state ST. This number equals 1 for state ST, only if the time course of an IC passes its threshold for the considered state and all others do not.

To identify the most prominent components of each state, the relative proportion of activations for each component was compared to the expected proportion of activations. The expected proportion was determined by simulating 10,000 threshold crossings from the deepest state of anesthesia or the awake state. Specifically, we randomly sampled 5,000 threshold crossings with replacement from each state, creating a distribution of expected occurrences free from bias due to differences in crossing numbers between states and independent of the state. Subsequently, we used the expected proportion of activations as a threshold to identify prominent components. Finally, a component was considered prominent if the lower bound of its 95% confidence interval for the relative proportion calculated across subjects (N = 3) exceeded its expected proportion obtained through the aforementioned random sampling.

To study how components combined with each other to contribute to the overall fluorescence signal, for each session in each brain state we computed the Pearson correlation coefficient r between the binary vectors of the ICs. Assuming that each such vector has at least a zero and a one, a correlation coefficient of 1 implies that both vectors are the same and their underlying ICs are surpassing their threshold in the same timepoint meaning that both components are being coactivated. By contrast, a correlation coefficient of −1 implies that the two components are passing their respective thresholds in different time points. By computing this correlation between each pair of ICs a correlation matrix R is obtained for each brain state in each session.