In order to simultaneously monitor the activity of cortical neurons across a large area (~25 mm², or up to 36 mm² in some cases) of either the bilateral dorsal cortex (dorsal mount), or of both dorsal and lateral cortex across the right hemisphere (side mount) in awake, behaving mice, we chose to utilize 2p imaging of GCaMP6s fluorescence using an existing and commercially available 2p random access mesoscope (Thorlabs; 2p-RAM; Sofroniew et al., 2016). We chose this form of data collection for use in head-fixed mice because it allows for: (1) rapid scanning (i.e. it uses a resonant scanner coupled to a virtually conjugated galvo pair) over a large (5x5 mm fully corrected, or up to 6x6 mm imageable), field-curvature corrected field of view (FOV); (2) subcellular resolution in the z-axis to avoid region-of-interest (ROI) contamination by neuropil and neighboring cells (0.61 x 0.61 x 4.25 µm xyz point-spread function at 970 nm laser excitation wavelength); (3) correction for aberrations at all wavelengths between 900–1070 nm, to allow for combined imaging of multiple fluorophores. We achieved these three objectives in head-fixed mice that were able to walk freely on a running wheel (See Figure 1a, b, Video 1, and Video 6), to clearly view visual monitors on both sides, and to lick either of two water spouts (left, right) for 2-alternative forced choice responses, while we performed left, right, and rear videography to monitor each mouse’s face, pupil, and body.

The neuronal imaging methodology we chose (see below) is not the only one available. For example, another promising and recently developed alternative implementation, Diesel 2p, allows for similarly flexible and large FOV imaging at single-cell resolution, but with a significantly higher z-axis point-spread function (~8–10 µm; Yu et al., 2021). Despite this drawback, its use of dual scan engines allows for true simultaneous imaging of different cortical areas, while the 2p RAM mesoscope (Thorlabs) accomplishes this by rapid jumping (~1 ms) between FOVs. Both systems offer excellent corrected field curvatures on the order of +/-25 µm over ~5 mm.

Several primary design constraints needed to be met to achieve our goal of monitoring neuronal activity over a large portion of the mouse lateral and/or dorsal cortices. These ranged from aspects of the surgical preparation, mounting, and imaging, to neural recording (Figure 1—figure supplement 1, Videos 1 and 2, Protocols II-III). To increase the extent of the neocortex over which we could monitor neural activity, we developed two distinct headpost/cranial window preparations. In the first, the dorsal mount, the mouse is head-fixed upright on a cylindrical running wheel and the objective is vertical (Figure 1a). Here, both wings of the headpost are attached to support arms, and each support arm is mounted on two vertical 1” diameter posts secured with flexing feet for maximum stability. This preparation allowed for monitoring of neural activity over large regions of both cerebral hemispheres, from the posterior aspects of visual cortex to the anterior aspects of motor cortex, and laterally to the dorsal-most aspects of auditory cortex, depending on the placement of the 2p mesoscope objective (Figure 1d).

The second preparation, the side mount, required the head of the mouse to be rotated 22.5^o to the left (or right, not shown), with the microscope objective vertical or tilted 1–5 degrees to the right, so as to extend the 5x5 mm imaging FOV to include the auditory cortex and other neighboring ventral/lateral cortical areas (Figure 1b and e). For the side mount, the headpost had only a single, left wing so that the right side of the mouse’s face would not be occluded. The deep lateral/ventral extent of the side mount headpost prevented easy addition of a right headpost wing. Attaching the left wing to a single support arm supported by two 1” diameter vertical posts mounted with flexing feet, as in the dorsal mount preparation, was sufficient to minimize movement artifacts. A key additional difference to note here is that two small screws were used to attach the left wing of the side mount headpost to the support arm, thus giving it additional stability. The dorsal mount preparation used only one screw (the outermost hole) on each of the left and right side headpost wings. Thus, both preparations were secured by a total of two mounting screws, each countersunk into headpost arms. An important design feature is that the headpost wings fit into recessed rectangular slots machined into the support arms, facilitating stability.

We observed that mice adapt well to head tilt in the side mount preparation, and will readily whisk, walk or run, and learn to perform lick response tasks in this configuration. Keeping the microscope objective in a vertical or near-vertical orientation, and rotating the mouse’s head (instead of rotating the objective), significantly improved the manageability of the water meniscus between the objective and the cranial window. However, if needed, the objective of the Thorlabs mesoscope may be rotated laterally up to +/- 20^o for direct access to more ventral cortical areas, for example if one wants to use a smaller, flat cortical window that requires the objective to be positioned orthogonally to the target region. In general, setups such as Sutter’s MOM system and Thorlabs’ Bergamo microscope, which offer even greater degrees of microscope objective rotation, would allow enhanced versatility with our preparations.

Preliminary comparisons across mice indicated that side and dorsal mount mice showed a similar degree of behavioral variability both within and across sessions. An example of the relative ease with which mice adapted to the ~22.5 degree neck rotation can be seen in Video 2 and Video 10. It was in general important to make sure that the distance between the wheel and all four limbs was similar when using either preparation. In particular, careful attention must be paid to the positioning of the front limbs in the side mount mice so that they are not too high off the wheel. This can be accomplished by a slight forward angling of the left support arm. With dorsal mount mice, it was important to not place them too close to the wheel, in which case they would exert pressure on the headpost and support arms, leading to increased movement artifacts, or too far above the wheel, which can significantly reduce walking bout frequency.

Although it was in principle possible to image the side mount preparation in the same optical configuration without rotating the mouse (by rotating the objective to 20 degrees to the right), we found that the last 2–3 degrees of unavailable, yet necessary, rotation (our preparation is rotated 22.5 degrees left, which is more than the full available 20 degrees rotation of the objective), along with several other factors, made this undesirable. First, it was difficult or impractical to attach the horizontal light shield and to establish a water meniscus with the objective fully rotated. One could use gel instead of water (although, in our hands, gel was optically inferior), but without the horizontal light shield, light from the UV and IR LEDs can reach the photomultiplier tubes (PMTs) via the objective and contaminate the image or cause tripping of the PMT. Second, imaging the right pupil and face of the mouse was difficult under these conditions because the camera would need the same optical access angle as the objective, or would need to be moved down toward the air table and rotated up 20 degrees, in which case its view would be blocked by the running wheel and other objects mounted on the air table. A system of mirrors for imaging the animal could be used to overcome this problem, but would further complicate the current setup.

Cortical imaging in the novel side mount preparation was restricted to one hemisphere and a thin medial strip (~1 mm wide) of the contralateral hemisphere. This allowed for imaging from roughly the anterio-medial areas of V1 to medial and medio-lateral secondary regions of motor cortex (or slightly farther, if the anterior-posterior axis of the brain is oriented along the diagonal of the 5x5 mm FOV, or if one images to the full allowable but partially uncorrected 6x6 mm extent with ScanImage). This anterior-posterior reach was similar to that observed in the dorsal mount preparation. However, the side mount windowing and the accompanying mounting procedure significantly increased our ability to image neural activity in lateral (ventral) cortical aspects, including auditory, somatosensory, and association cortical regions (Figure 1e), notably without the need for substantial rotation of the objective.

To achieve 2p neuronal imaging from broad cortical regions, we created large and stable, custom (designed in AutoDesk Inventor) 3D-printed titanium (laser sintered powder with active cooling and shot-peened post-processing) headposts (Figure 1d, e and Figure 1—figure supplement 1e, f; Suppl Design Files; i.materialise.com and sculpteo.com) and glass (0.21 mm thick, Schott D263T) cranial windows (Figure 1, Figure 1—figure supplement 1a, c, d, e, and f ; Videos 1 and 2; Suppl Design files: https://github.com/vickerse1/mesoscope_spontaneous (copy archived at Vickers, 2024a); Labmaker.org for dorsal mount; TLC International for cutting, and GlasWerk for bending, for the side mount and early prototypes of the dorsal mount). We incorporated these custom made headposts and cranial windows into the dorsal mount (Figure 1a and d, Figure 1—figure supplement 1c upper, d upper left, e), and side mount (Figure 1b, e, Figure 1—figure supplement 1a, c, lower, d, lower left and right, and f, upper right) preparations. The dorsal mount preparation allowed for imaging over roughly the same extent of bilateral dorsal cortex as in Kim et al., 2016, although we were able to image up to ~25 mm² at a time, compared to the more typical maximum 2p FOV of ~1 mm². In addition, the area of bilateral cortex imageable with this preparation is comparable to that of many recent studies employing widefield 1-photon imaging, such as Musall et al., 2019. The side mount preparation allowed for imaging of an extent of right hemisphere comparable to that of Esmaeili et al., 2021, although they used through-the-skull widefield 1-photon imaging, while our cranial window preparation allows for either widefield 1-photon imaging, or single neuron resolution 2p imaging, in the same mouse.

To mimic the curvature of the brain and reduce tissue compression, which is a problem with flat coverslips with a diameter larger than ~3 mm, cranial windows were curved by heating pre-cut (TLC International) glass pieces over 9 or 10 mm bend-radius molds (GlasWerk). The bend radius (i.e. radius of half cylinder mold over which the melted glass is bent to achieve its curved shape) was fixed at 10 mm for Labmaker.org dorsal mount windows (Figure 1—figure supplement 1e, right; Video 1; early attempts with 11 or 12 mm bend radii failed for all but the largest adult male mice). For windows custom designed in collaboration with GlasWerk (Figure 1—figure supplement 1f, right; Video 2), the bend radius was set at either 9 mm, for a tight fit to ventral auditory areas, or at 10 mm, to enable simultaneous imaging of the entire preparation at a single focal depth. In our experience, successful cranial windows last between 100 and 150 days, or in some rare cases up to 300 days: https://github.com/vickerse1/mesoscope_spontaneous/blob/main/window_preparation_stability.pdf; Vickers, 2024b. The keys to this stability were even pressure across the surface of the window, complete enclosure of the glass-skull interface with flow-it, causing minimal damage to the dura during the craniotomy, and constant attention to the minimization of infection either across the surface of the skull, between the headpost and window surgeries, or around the external perimeter of the headpost, after the window surgery (See Protocol III and Figure 1—figure supplement 1).

The success rate for window implantation, for both dorsal mount and side mount preparations, was around ~65% (Figure 1—figure supplement 1g; see also Protocol III). However, with extensive practice this can be raised to be closer to 75–80%, especially if care is taken to both maximize the area of the craniotomy and to leave the dental cement attaching the headpost to the skull undamaged during drilling, so that the perimeter of the window fits fully inside the craniotomy, the window ‘floats’ or sits freely directly on the surface of the brain, and the headpost remains firmly attached to the skull. This was accomplished in later iterations of our designs and methods by making the side mount window slightly smaller at the anterior edge of M2, and by using an adjustable support arm (Figure 1—figure supplement 1b, lower) mounted on a small breadboard attached underneath the base of the stereotax in order to fix the mouse in its 22.5 degree left-rotated position during the window surgery so that it would not move or vibrate during drilling of the craniotomy. This also acted to minimize ‘break-throughs’ of the drill bit tip through the skull and into the brain.

The main causes of death during cranial window surgery were exsanguination during the surgery following a ruptured major blood vessel on the surface of the brain (either sagittal or lateral, usually), damage to the dura caused by the skull during the removal step of the craniotomy or by the sharp tip of a tool used to remove the skull, failure to properly pressurize the window across the entire cortical surface, or failure to completely seal the window to the skull fragment around its entire perimeter with flow-it (here, rapid UV light application during flow-it application can help). In some cases where the mouse survived the window surgery it was still considered a failure if a bleed re-erupted and occluded at least 30% of the window, or if glue occluded at least 30% of the window either above or below the surface of the window. In general, most bleeds cleared, on their own, by cerebrospinal fluid within a week if there was no direct clotting on the surface of the brain.

Changes in vasculature following cranial window surgery were usually minimal but could involve the following: (i) sometimes a vessel is displaced or moved during the window surgery, (ii) sometimes a vessel, in particular the sagittal sinus, will enlarge or increase its apparent diameter over time if it is not properly pressured by the cranial window, and (iii) sometimes an area experiencing window pressure that is too low will, over time, show outgrowth of fine vascular endings. The most common of these was (i), and (iii) is perhaps the least common. In general the vasculature is quite stable, and window preparations that we observed to be of high quality at ~1 week post-surgery showed minimal changes over the first ~150 days (see supplementary materials: https://github.com/vickerse1/mesoscope_spontaneous/blob/main/window_preparation_stability.pdf; Vickers, 2024b).

Within a field of view of 5x5 mm, curvature of the brain, especially in the mediolateral direction, is a significant problem. To partially compensate for brain curvature, online field-curvature correction (FCC) was applied in ScanImage (Vidrio Technologies, MBF Biosciences) during mesoscope image acquisition. In addition, complementary fast-z ‘sawtooth’ or ‘step’ corrections were applied when all ROIs were acquired at the same or different z-planes, respectively (see Sofroniew et al., 2016). Despite this, small uncorrected discrepancies in the depth of the imaging plane below the pial surface may have remained, especially along the long-axis of acquisition ROIs in the side mount preparation (i.e. mediolateral). It is likely that optical effects of the curvature of the glass partially compensated for these differences in targeted imaging depth (i.e. by refraction to normalize approach of the laser to the pial surface), although we did not quantify this effect. In cases where we imaged multiple small ROIs, nominal imaging depth was adjusted in an attempt to maintain a constant relative cortical layer depth (i.e. depth below the pial surface).

We estimate that we experienced ~200 μm of depth offset across 2.5 mm. If the objective is orthogonal to our 10 mm bend window and centered at the apex of its convexity, a small ROI located at the lateral edge of the side mount preparation would need to be positioned around 200 μm below that of an equivalent ROI placed near the apex, and would be at close to the same depth as an ROI placed at or near the midline, at the medial edge of the window. We determined this by examining the geometry of our cranial windows in CAD drawings (available at https://github.com/vickerse1/mesoscope_spontaneous/tree/main/cranial_windows), and by comparing z-depth information from adjacent sessions in the same mouse, the first of which used a large FOV and the second of which used multiple small FOVs optimized so that they sampled from the same cortical layers across areas.

The mouse was restrained by fixation of the headpost with two countersunk screws to either a fixed (Figure 1—figure supplement 1b, top) or adjustable aluminum support arm (custom; Figure 1a, b, Figure 1—figure supplement 1b, bottom) and mounting apparatus (Figure 1a and b). For the dorsal mount preparation, two headpost support arms, one on each side of the head, were used (Figure 1a, Figure 1—figure supplement 1c, top; outer screw of both headpost wings used) while for the side mount preparation, a single, left side support arm was sufficient (Figure 1—figure supplement 1, bottom; inner and outer screws in the left headpost wing were both used). It was important to fix the headpost mounting arms to the top of the airtable with 1-inch diameter vertical mounting posts (Thorlabs RS6P8, RSH4, PF175), making sure all joints were clean of debris and tightened securely, to reduce movement artifacts.

To further minimize movement artifacts, it was also important to mount the mouse at a distance from the surface of the running wheel that was not too low, which would allow the mouse to push up with its legs, or too high, so that part of the mouse’s body weight would be supported by the headpost. A final consideration to minimize movement artifacts was to ensure that proper, even pressure was maintained across the entire interface between the brain and the cranial window during window implantation, both by making the craniotomy large enough so that the entire window could be set freely on the brain surface, and by using the 3D-printed window stabilizer properly during the application of flow-it glue to attach the edge of the window to the skull (see Protocols II, III).

Both preparations were found to be relatively free of vibration and movement artifacts and allowed for micro-adjustments (i.e. pitch, yaw, and roll) of the mouse relative to both the objective and the running wheel (Figure 1a, b and c). Increased positioning flexibility of the support-arm assembly was achieved through either of two ways: (1) by positioning it on a distal base-attached ball-joint mount (Thorlabs, SL20 articulating base; not shown); (2) through the use of a custom adjustable headpost support arm that was proximally adjusted around a ball-and-joint assembly (Figure 1—figure supplement 1b, bottom). This adjustable support arm operated through the use of a custom wrench and reverse-threaded (left-handed) nut for initial coarse tightening, and a micro-clamp for rotational stabilization during tightening. Four micro-hex wrench driven set-screws, two on either side of the ball-and-joint fitting, were used to achieve the final, fully locked state for imaging. The adjustable support arm was found to be superior to the base-attached ball joint mount in allowing for iterative micro-positioning of the animal in relation to the running wheel, objective, and lick spouts after initial fixation (mounting) of the headstage to the support bar.

In order to perform 2p neuronal imaging, the large 2p-RAM water-immersion objective (~10x net magnification, 0.6 NA, ~1.3 kg, 12 mm diameter tip, 25.6 mm shaft) must be positioned between 2.2 and 2.8 mm (i.e. ~working distance minus window thickness) from the curved glass surface of the cranial window while maintaining a stable water meniscus. Providing a base for the water meniscus while also blocking incident light entry was achieved by creating a custom 3D printed plastic light shield (AutoDesk Inventor, MakerGear M3 printer, PLA) that was attached to the protruding, fitted rim of the headpost with 170 FAST CURE Sylgard (‘wok one’; Figure 1c and Figure 1—figure supplement 1c, d). A second light shield (‘wok two’) served to further block extraneous light entry into the objective and was attached to the base of the lower light shield (‘wok one’) by fitting of a U-shaped profile over a vertically protruding single edge profile along the perimeter of the first, lower light shield (‘wok one’; Figure 1c).

Together, these two custom-printed light shields prevented incident light (from the video stimulus monitor, the ultraviolet LEDs used for controlling the baseline and dynamic range of pupil diameter, and the infrared LEDs used to illuminate the mouse) from entering the imaging objective and thereby either contaminating the image or tripping the photomultiplier tubes (PMTs). Line-of-sight for the animal to the video stimulus monitor was retained, and this design allowed for free vertical and rotational (over a limited range) movement of the objective lens, which was contacted directly only by the water meniscus.

Direct left, right, and posterior camera angles, or ‘lines of sight’, were preserved to enable reliable recording and proper illumination of pupil diameter, whisker pad movement, and the movement of other body parts such as the nose, mouth, ear, paws, and tail (Figure 1a and b; see Videos 1 and 2). Placement of the rotated mouse in the side mount preparation near the left side of the running wheel allowed for the animal’s left and right fields of view to not be significantly obstructed. Adjustable positioning of two vertically and laterally offset conductance-based lick spouts for 2-alternative forced choice (2-AFC) task performance (lick left, lick right) and reward delivery was achieved with a rapidly translatable motorized linear stage (Figure 1c; 63.7ms total travel time over 7 mm; Zaber Technologies). This system allowed for rapid withdrawal of lick spouts between trials, and presentation of the lick spouts during the response period of each trial.

Examining neuronal activity at the single cell level and relating it to stimulus-evoked activity observed at the widefield level (Figures 2 and 3), as well as manipulating regions of the cortex through localized light delivery and optogenetics (see Video 11), required precise optical alignment across all three of these methods. First, a standardized cortical map needed to be fitted onto images of skull landmarks and the cortical vasculature by widefield multimodal sensory mapping (MMM; Figures 2 and 3a, b, and c, top). Then, a method was needed for transferring this cortical map onto the field of view of the mesoscope to both allow for online FOV targeting and for post-hoc assignment of each imaged neuron to a designated cortical area on the Allen common-coordinate framework (CCF v3.0) map for subsequent analyses (Figures 2 and 3c, bottom, d). Finally, the spatial coordinate system of the 1p opto-stimulation laser routed in through an auxiliary light-path of the mesoscope needed to be aligned to that of the 2p laser so that specific cortical subregions could be targeted for optogenetic inhibition by light activated, ChR2-mediated excitation of parvalbumin interneurons in Thy1-RGECO x PV-Cre x Ai32 mice (Video 11).

In order to align 2p maps of single neuron activity with functional 1p cortical maps determined through pseudo-widefield imaging (i.e. combined reflected and fluorescence light imaging with a standard ‘non-widefield’ format CCD camera), we designed a method for reliably aligning a standardized cortical map (i.e. the Allen CCF v3.0; we used either a 0 degree, https://github.com/vickerse1/mesoscope_spontaneous/blob/main/python_code/CCF_map_rotation/0deg/CCF_MMM.png; Vickers, 2024b, or 22.5 degree, https://github.com/vickerse1/mesoscope_spontaneous/blob/main/python_code/CCF_map_rotation/22deg/CCF_MMM.png; Vickers, 2024b, rotated CCF map outline created with the following code: https://github.com/vickerse1/mesoscope_spontaneous/blob/main/python_code/CCF_map_rotation/Rotate_CCF.py; Vickers, 2024b), to an image of the cortical vasculature pattern, which can be directly visualized with both widefield and 2p imaging techniques for each brain (see Protocol V). The epifluorescence light source (Excelitas) for the pseudo-widefield imaging, which was parfocal with the 2p imaging coordinate system, was controllable by a dial, shutter, and remote foot-pedal for optimal ease of positioning the objective in the water meniscus at a distance of ~2.2 mm from the headpost and cranial window without crashing into the preparation (Figure 1c and Figure 1—figure supplement 1d, top and bottom right). This allowed us to ‘drive’ the position of the 2p mesoscope FOV to the desired Allen CCF cortical area by moving the objective to a location where the observed vasculature pattern matched that of a processed image we created ahead of time on our 1-photon widefield imaging rig that combined vasculature, sensory responses, and skull landmarks from each mouse.

This technique required several modifications of the auxiliary light-paths of the Thorlabs mesoscope, and would also likely involve similar modifications in other comparable microscope setups, such as Diesel 2p (Yu et al., 2021). For switchable blue/green widefield imaging and 2p imaging in our original configuration (see (a) in https://github.com/vickerse1/mesoscope_spontaneous/blob/main/mesoscope_optical_path_opto_switching.pdf; Vickers, 2024b), we used a 469/35 Semrock excitation filter, 466/40 Semrock dichroic (1st, top cube), and mirror (2nd, bottom cube). For the combination of pseudo-widefield imaging and rapid, targetable optogenetic 1-photon (1p) stimulation with concurrent 2p imaging (see (b) in https://github.com/vickerse1/mesoscope_spontaneous/blob/main/mesoscope_optical_path_opto_switching.pdf), which we used with PV-Cre x Ai32 x Thy1-RGECO mice (Video 11), we established dual coupling of the broadband fluorescence light-source (Excelitas) used in the original configuration and a 473 nm laser driver (SF4C 473, Thorlabs), coupled via liquid light guide, to converging, dual input auxiliary light paths of the Thorlabs mesoscope via two in-series, magnetically secured, switchable filter cubes (DFM1T1 cube, Thorlabs; see https://github.com/vickerse1/mesoscope_spontaneous/blob/main/mesoscope_filterCube_schematic_dual_opto_Jan0320.jpg; Vickers, 2024b; the top cube can be switched to allow pseudo-widefield imaging of either GCaMP6s or Thy1-RGECO, and the bottom cube can be switched to allow either pseudo-widefield imaging or optogenetic stimulation).

The 473 nm (blue) laser driver was connected to a single open loop, high speed buffer (50 LD, Thorlabs), and targeted to the coordinate system of the 2p laser with a grid-calibrated, auxiliary galvo-galvo scanner (GVSM002, Thorlabs). This allowed for 2p imaging and blue 1p laser optogenetic stimulation to occur pseudo-simultaneously, because rapid, electronic control of the PMT1 and PMT2 shutters was able to limit interruption of image acquisition to a brief period during presentation of the optogenetic stimulus (~100 ms; see Video 11; see https://github.com/vickerse1/mesoscope_spontaneous/blob/main/mesoscope_optical_path_opto_switching.pdf). Here, replacement of the switchable mirror behind the objective, which blocked 2p imaging during pseudo-widefield imaging in our original configuration, with a dichroic that, when positioned in the light path, allowed both 920 nm (2p excitation) and 473 nm (opto-excitation) light to reach the mouse brain without requiring any additional slow switching of optical components.

Resonant scanners in 2p microscopes emit intense sound at the resonant mirror frequency, which is well within the hearing range of mice (~12.5 kHz emitted in the Thorlabs mesoscope). Because scanning precision requires the scanner to remain at a stable, elevated temperature, the scanner must remain on during the entire experimental session (i.e. up to ~2 hr per mouse). The unattenuated or native high-frequency background noise generated by the resonant scanner causes stress to both mice (Sadananda et al., 2008) and experimenters (Fletcher et al., 2018), and likely acts to prevent mice from achieving maximum performance in auditory mapping, spontaneous activity sessions, auditory stimulus detection, and auditory discrimination sessions/tasks.

To reduce this acoustic noise, we encased the resonant scanner and attached light path tubes with a custom 3-dimensional (3D)-printed assembly containing dense interior insulating foam. See Supplementary methods and materials for diagrams: https://github.com/vickerse1/mesoscope_spontaneous/blob/main/resonant_scanner_baffle/closed_cell_honeycomb_baffle_for_noise_reduction_on_resonant_scanner_devices.pdf; Vickers, 2024b, and for a text description: https://github.com/vickerse1/mesoscope_spontaneous/blob/main/resonant_scanner_baffle/closed_cell_honeycomb_baffle_methodology_summary.pdf; Vickers, 2024b; ‘3D scanning and honeycomb patterned nylon print’ (University of Oregon Innovation Disclosure #DIS-23–001, US provisional patent application UOR-145-PROV). It was critical to use 3D scanning of encased components in the design of the noise reduction shield so that the sound-reduction assembly would closely follow all of the surface contours and fit accurately in spaces with tight tolerances.

The result of this encasement was a large reduction in resonance scanner sound, from ~60 dB to ~5 dB measured at the head of the mouse (Bruel and Kjaer ¼” pressure filled microphone 4938 A-011; i.e. below the mouse hearing threshold at 12.5 kHz of roughly 15 dB; Zheng et al., 1999). By comparison, encasements designed using standard 3D-design and printing techniques (i.e. not based on a 3D scan) were, in our hands, only able to achieve a noise reduction of ~30 dB, to a level still audible to both mice and humans. This difference was due to the enhanced precision of encasement fit enabled by using the 3D-scanned map of the microscope’s surface contours.

Interpretation of the diversity of activity of thousands of neurons simultaneously identifiable with the mesoscope requires proper cortical areal localization, including that of areas anatomically or functionally distant from primary sensory cortices. Such mapping is not routinely possible directly on the 2p mesoscope when performing neuronal-level imaging, due to both the heterogeneity of responses within primary sensory cortices at the single neuron level, and to animal-to-animal variations in the spatial extent of cortical regions (de Vries et al., 2020; Bimbard et al., 2023). To facilitate the assignment of neurons to cortical areas, we sought to align the Allen Institute Common Coordinate Framework (CCF v3.0; Wang et al., 2020) to our 2p neuronal imaging results, using blood vessels, skull landmarks, and widefield imaging responses as intermediaries. We used an overlaid image of these features that we refer to as the ‘multimodal map’ (MMM) as on-the-fly guidance for FOV placement during 2p mesoscope imaging sessions, and then performed a precise post-hoc CCF alignment based on the MMM and vasculature patterns observed during 2p imaging to assign a unique CCF area identifier to each Suite2p-identified ROI/neuron during preprocessing stages of our data analysis (see https://github.com/vickerse1/mesoscope_spontaneous/tree/main/matlab_code for creation of the MMM, and https://github.com/vickerse1/mesoscope_spontaneous/tree/main/python_code for generation of rotated CCF map outlines and precise neuron assignment to CCF areas, which takes place in the following jupyter notebook: https://github.com/vickerse1/mesoscope_spontaneous/blob/main/python_code/mesoscope_pre_proc/meso_pre_proc_1.ipynb).

Alignment of the CCF to the dorsal surface of cortex and subsequent image-stack registration based on widefield imaging Ca²⁺ fluorescence data has been performed elsewhere with a variety of techniques. These include the use of skull landmarks (e.g. bregma and lambda; Musall et al., 2019), responses to unimodal sensory stimuli (Gallero-Salas et al., 2021), visual field mapping (Zhuang et al., 2017), and/or autocorrelation maps of spontaneous activity (Peters et al., 2021). Few studies, however, have performed such alignments with rotated cortex (i.e. imaging lateral portions of the cortex at an angle; but, see Esmaeili et al., 2021).

To summarize our approach, we used a technique where, for both the dorsal mount (Figure 2, Figure 1—figure supplement 2a) and side mount (Figure 3, Figure 1—figure supplement 2b) preparations, we first imaged 1p widefield GCaMP6s responses to unimodal passive sensory stimulation (e.g. visual, auditory, and somatosensory) to create a MMM consisting of multiple sensory area masks on top of the cortical vasculature. We then overlaid skull landmarks (e.g. bregma and lambda), imaged with reflected green light, onto the MMM. This intermediate overlay image, created using custom MatLab code to extract mean widefield dF/F sensory responses (https://github.com/vickerse1/mesoscope_spontaneous/blob/main/matlab_code/SensoryMapping_Vickers_Jun2520.m; Vickers, 2024b), followed by z-projection and selection masking techniques in Fiji/ImageJ, was used to guide selection of the FOV during 2p mesoscope imaging sessions. The Allen CCF was then warped onto the vasculature/skull/MMM overlay, and the mean 2p image was warped onto the MMM by vasculature alignment using custom code (see:https://github.com/vickerse1/mesoscope_spontaneous/blob/main/python_code/mesoscope_pre_proc/meso_pre_proc_1.ipynb for full initial processing steps of 2p data, or https://github.com/vickerse1/mesoscope_spontaneous/blob/main/python_code/mesoscope_preprocess_MMM_creation.ipynb; Vickers, 2024b for a short notebook containing only the relevant CCF alignment steps - as with all of our jupyter notebooks, these should be run in the ‘uobrainflex’ python environment: https://github.com/sjara/uobrainflex; Jaramillo, 2020). Finally, each neural ROI in the 2p image was assigned, post-hoc, to a position in the MMM x-y coordinate system and a corresponding CCF area based on the final overall alignment of the MMM, CCF, and 2p image. The following sections describe these steps in more detail.

To create the multimodal map, each mouse was put under light isoflurane anesthesia (~1.5%) after head-post surgery, but before implantation of the cranial window, and exposed to 5 min each of full-field visual (vertical and horizontal stationary grating patches; 0.16 cpd, 30 deg; Michaiel et al., 2019), tone-cloud auditory (a series of overlapping 30 ms duration tones randomly selected from a frequency range of 5–40 kHz and presented at 100 Hz; Xiong et al., 2015), and piezo-driven whisker deflection (or, in some cases, forelimb and trunk stimulation; see Gallero-Salas et al., 2021). For whisker deflection, a 1 s, 5 Hz burst of five 100 ms duration forward sweeps was used (i.e. each sweep consists of 100ms of forward movement and 100 ms of backward movement), consisting of posterior to anterior sweeps. The whisker deflector was a custom 3D-printed triangular polylactic acid (PLA) piece mounted on a 21-gauge needle and attached to a PL140.11 piezo-actuator (PI Ceramic) with epoxy glue. It was actuated with a Physik-Instrumente controller driven by a custom pulse sequence generated in Spike2 and delivered from an analog output of a CED Power 1401, with an inter-stimulus interval of ~10 s (Figure 2b, right, and Figure 3b, right; see also Videos 3–7).

Light anesthesia was used to minimize unwanted cortical activity due to spontaneous movements and arousal fluctuations, and to prevent the spread of cortical sensory responses to areas downstream of primary sensory areas. Imaging through the skull between the headpost and cranial window implantation surgeries was done to allow for coregistration of skull landmarks with vasculature and sensory responses, and in general yielded more contiguous, easily interpretable multimodal maps than the same widefield mapping done through the cranial window. Although the resulting overlay of vasculature and the multimodal sensory map was useful for cranial window placement, it was not strictly necessary. Additional MMM sessions performed through the cranial window were performed every 30–60 days or as necessary due to slight changes in vasculature. In general, vasculature was stable throughout the entire ~150–200 day lifespan of a successful cranial window preparation (see https://github.com/vickerse1/mesoscope_spontaneous/blob/main/window_preparation_stability.pdf; Vickers, 2024b), although in some cases slight increases in the diameter of major blood vessels (e.g. sagittal sinus), or outgrowth of fine arteriole endings, were observed.

Averaged, baseline-subtracted dF/F responses (Figures 2b and 3b) were thresholded, masked, outlined, and layered onto the blood vessel image (Figure 2a, c, Figure 3a, c, Figure 1—figure supplement 2a, b) along with bregma and lambda skull landmarks (Figure 1—figure supplement 1a) identified with 530 nm (green) reflected-light skull-imaging using a custom protocol/macro in Fiji (ImageJ; Figure 2c, Figure 3c, Figure 1—figure supplement 2a, b; Videos 3–7). This multimodal mapping procedure and alignment to the CCF and skull landmarks was repeated if significant changes in vasculature pattern occurred over the days/weeks of the experiment. More precise maps of visual cortical areas can be achieved through visual field mapping using a topographic stimulus consisting of a bar sweeping in azimuth or elevation (Garrett et al., 2014). This technique was not applied here, because our goal was global alignment of the Allen Institute cortex-wide CCF to our dorsal or side mount views, and not precise alignment to visual subareas per se.

The MMM was aligned to an overlay of the rotated CCF edges and region of interest (ROI) masks using custom Python code and the built-in function ‘PiecewiseAffineTransform’, given a user-supplied series of bounding-box and alignment points common to both images (Figure 2c (upper, part 1), 3 c (upper, part 1)). A second, similar alignment was then performed between the multimodal map and all neural ROI locations (i.e. Suite2p-identified neurons) relative to the 2p image plane; here, vasculature is inferred from the pattern of gaps (e.g. vascular ‘shadows’) in the spatial distribution of neurons (Figure 2c (lower, part 2), 3 c (lower, part 2)), and can be confirmed by parfocal pseudo-widefield (i.e. combined reflected and epifluorescence light) imaging directly on the 2p Thorlabs mesoscope. Note that in 2p imaging below the pial surface, the effective/apparent width of the vasculature, or its “shadow”, is significantly larger than that of the actual blood vessel, and its width increases with depth. The transforms resulting from each of these two alignments were applied in serial (i.e. multimodal map to CCF, then multimodal map to 2p neural image; Figures 2c and 3c) to overlay the neural image directly onto the cortical map, using outlines, and to assign a unique CCF area identifier (i.e. name and ID number) to each Suite2p-extracted, 2p-imaged neuron using CCF masks (Figures 2d and 3d). Note that, while the final assignment of CCF area identifiers to each Suite2p-identified neuron (Figure 2c (2), right) was performed post-hoc, the intermediate blood-vessel/MMM overlay image (Figure 2 (1) and (2), left) was used for online guidance of 2p mesoscope FOV selection during imaging sessions.

Previous 2p imaging studies have demonstrated that arousal and/or orofacial/body movement can explain a significant proportion of the variance in spontaneous and sensory evoked neural responses in visual and other restricted dorsal cortical regions (Musall et al., 2019; Stringer et al., 2019). Here, as a proof of principle, we examined the generality of this finding by simultaneously monitoring neuronal activity across broad regions of bilateral dorsal cortex, with our dorsal mount preparation, and both dorsal and lateral cortex across the right hemisphere, in our side mount preparation. Previous studies suggest that there may be both commonalities as well as heterogeneity in the effects of changes in arousal/movement on spontaneous and sensory-evoked responses across different cortical areas (McGinley et al., 2015; Musall et al., 2019; Stringer et al., 2019). For example, running typically enhances the gain of visually evoked responses in the mouse primary visual cortex (Niell and Stryker, 2010), while it significantly decreases that of evoked auditory responses in primary auditory cortex (Zhou et al., 2014; McGinley et al., 2015).

Assessing such potential differences between functionally distinct, and potentially distant, cortical areas required simultaneous imaging across multiple cortical regions while maintaining adequate imaging speed, quality, and resolution. To achieve this level of 2p imaging across several millimeters of cerebral cortex (e.g. from visual to motor or auditory cortical areas), we first sought to optimize the parameters of our mesoscope imaging methods and protocols.

Conventional 2p imaging using a preparation similar to our dorsal mount (“Crystal Ckull”; Kim et al., 2016) previously employed serial acquisition of ~1 x 1 mm FOVs (Kim et al., 2016). Other 2p Thorlabs mesoscope imaging studies have used acquisition protocols targeting z-stacks of 600x600 µm FOVs in barrel cortex (3 planes at 7 Hz, 1.17 µm/pixel; Peron et al., 2015), 900x935 µm FOVs in visual cortex (11 planes at ~1 Hz; Stringer et al., 2019), and three adjacent 600x1800 µm FOVs in CA1 hippocampus (Sun et al., 2023; bioRxiv). Here, with our dorsal and side mount preparations, we were able to routinely image between 2000 and 7600 neurons per session over up to ~25 mm² of dorsal cortex simultaneously at ~3 Hz with a resolution of 5 µm/pixel in GCaMP6s mice (combined total in both preparations for all large-FOV sessions: N=17 mice, n=91 sessions, ~350,000 neurons; Figure 1—figure supplement 2e, Video 8).

In order to achieve spatially broad and dense sampling, while monitoring changing activity in each neuron as frequently as possible, we typically tiled seven mediolaterally aligned 5000x660 µm FOVs (i.e. mechanical scanning along a long-axis oriented mediolaterally, combined with fast resonant-scanning across multiple short-axes oriented anterior-posterior) over either posterior-medial, posterior-lateral, or anterior dorsal, cortex at a pial depth of ~200–300 µm (cortical layers 2/3). Bidirectional scanning (with a scan-phase correction of between -0.9 and -1.0) and field-curvature correction were typically enabled, with a 1-5 ms ‘flyback’ and ‘frame’ time, which are defined in ScanImage as the time delay for retargeting of the laser from the end of one line or frame acquisition to the beginning of the next, respectively. Here, there is a trade-off between longer flyback times, which reduce imaging speed but minimize positioning errors at the beginning of each scan line or ROI, and shorter values, which accelerate overall imaging acquisition speed but increase these positioning errors.

We typically imaged GCaMP6s fluorescence in awake, behaving mice at an excitation wavelength of 920 nm with 60–95 mW power calibrated at the pia over 20–90 min sessions, with little to no bleaching or change in baseline fluorescence. Suite2p reliably extracted neurons with >0.9 classifier probabilities, and nearly completely (i.e. to less than 0.1 pixels) removed x-y movement artifacts with uncorrected, raw principal components of up to ~1.3 pixels (i.e. up to ~15% of a 20–40 µm, or 4–8 pixels, diameter neuron at an imaging resolution of 0.2 µm/pixel, reduced to less than 0.5 µm or 1–2% of the diameter of a neuron after combined rigid and non-rigid motion correction; see Video 11).

To further control for z-movement artifacts (whose correction is intractable post-acquisition under normal scanning conditions) we also occasionally used GPU-enabled online fast-z motion correction (42/52 large FOV side mount sessions, and 9/39 large FOV dorsal mount sessions, used online z-correction) with ScanImage (Vidrio Technologies, LLC). This was only possible, in our hands, during multi-frame imaging when all frames were acquired at the same z-level. Quality of post-hoc motion-corrected dF/F traces and correlations with behavioral variables did not appear, by eye, to be different under these conditions (Figure 4—figure supplement 1a vs b; see https://github.com/vickerse1/mesoscope_spontaneous/tree/main/online_fast_z_correction; Vickers, 2024b), as both types of sessions contained similar mixtures of neurons with both walk-bout dependent and independent activity, for example. Because the z-motion correction did not lead to large qualitative differences in neural activity and its relationship to behavioral primitives, both types of session are used interchangeably here.

In order to confirm neural region-of-interest (ROI) selection (i.e. each neuron as identified by a spatiotemporal footprint in Suite2p), increase scan speed, and correct our cortical layer 2/3 targeting for the effects of brain curvature, we also routinely imaged in random access mode with 4–6 small FOVs (660–1320 x 660 µm) positioned over visual (V1), somatosensory (S1), retrosplenial (RSpl), posterior parietal (PPC), and/or motor (MOs_pm, MOs_am, and/or MOs_al - ALM) cortices at independently controlled z-levels, with an imaging speed of 4–10 Hz and a resolution of 1.0–2.0 µm/pixel (see Protocol I; Figure 1—figure supplement 2d, e). These multi-FOV sessions yielded hundreds of high quality neurons per FOV, at a density roughly twice that of the lower resolution sessions (Figure 1—figure supplement 2c vs d).

Uncontrolled image motion due to the movement of behaving animals is a significant impediment to image quality (Pachitariu et al., 2016). We have taken significant care in our headpost design and fixation (see Figure 1a, b, Figure 1—figure supplement 1a–f, and Supplementary methods and materials) so that our preparations are highly stable, such that motion-corrected or ‘registered’ movies showed very little to no residual movement in either large FOV, low-resolution, or multiple small FOV, high-resolution sessions (Video 9; Video 11). Furthermore, movement-related neural activities across neighboring detected ROIs in large FOV, low-resolution sessions were correlated but non-identical upon detailed examination, and differed from dF/F Ca²⁺ fluorescence fluctuations detected in nearby non-ROI and blood-vessel regions of similar cross-sectional area (see Video 9; Video 11), thus strongly suggesting that it was non-artifactual. In theory, movement artifact signals should be relatively time-locked or nearly synchronous across cortical regions during vigorous walking or movement. However, we did not observe this in our preparations.

These findings, therefore, gave us confidence that neural activity changes detected during spontaneous movements of the mouse were due to real changes in Ca²⁺ fluorescence within each neuron, and not to movements of the brain causing the neuron to shift into or out of each Suite2p-defined ROI and leading to artifactual fluorescence transients (see Video 9; Video 11, which use blood vessels and inactive fluorophore/neuropil, respectively, as examples of the lack of significant movement-driven artifactual transients in GCaMP6s(-) structures in our preparations). In other studies, performed using the same setup and methods as those applied here, we have demonstrated that non-activity dependent fluorescent markers, such as the expression of a non-activity dependent fluorescent marker mCherry in cholinergic or noradrenergic axons, or autofluorescent blebs, exhibit no significant movement-related changes in fluorescence, indicating that movement artifacts are minimal in our preparations (Collins et al., 2023). It should be noted that axonal diameters are on the order of ~1 µm or less - this is significantly smaller than the diameter of the neuronal cell bodies (~15–30 µm) imaged here. Axons, therefore, are a more stringent test of the stability of our imaging preparations.

Recent work has shown that sorting fluorescence based 2p GCaMP cortical neuronal activity with an algorithm called Rastermap is a powerful method to reveal clusters of cells that exhibit similar patterns of activity (Stringer et al., 2023, bioRxiv). An example of such sorting of our data for a dorsal mount session of 3096 neurons is shown in Figure 4a (top). This method for one-dimensional nonlinear embedding works by sorting and clustering of ROIs (i.e. neurons) by similarity of neural activity patterns across multiple timescales, using k-means clustering, sorting by asymmetric similarity as defined by peak cross-correlations at non-negative timelags, and upsampling of individual cluster activities in principal components space to allow sorting within clusters (Stringer et al., 2023, bioRxiv). Here, sorting with Rastermap revealed clustering of neurons with clearly distinguishable activity patterns that exhibited heterogeneous relationships to the behavioral primitive arousal/movement variables of walk speed, whisker motion energy, and pupil diameter (Figure 4a, top, and b; Video 8).

The activity generated by these thousands of neurons was clearly not random, with sub-groups of cells exhibiting similar activity patterns. Many neurons exhibited activity that appeared to be coupled to walking/whisking (Figure 4a top, b). It has previously been shown that the first principal component of cortical neuronal activity is highly correlated with movement (Stringer et al., 2019). Indeed, sorting our neurons based on degree of activity loading onto their first two principal components (i.e. neurons at the top of the sort have the largest amount of activity accounted for by the first principal component of activity across the entire population) revealed several distinct patterns of activity. The first principal component, PC1 (Figure 4a, middle) demonstrated a clear relationship to walking/whisking, such that some neurons were strongly activated in association with movement (e.g. Figure 4a and c1 in PC1, middle, and c, top), while other neurons appeared to be deactivated during movement bouts (e.g. Figure 4a and c2 in PC1, middle, and c, bottom).

Closer examination revealed that some manually selected Rastermap and PC1 clusters were active primarily during walking (Figure 4a, top, d1, Figure 4a, middle, c1, and Figure 4, d1) and/or whisking, with either transient or sustained activity, while others were either active during low arousal periods (Figure 4a, middle, c2) or appeared to display a combination of ON/OFF transient whisking-related activity (Figure 4a, top, d2, and Figure 4d, right; see Figure 4—figure supplement 2). In the illustrated example, alignment of marked increases and decreases in bulk population neural activity to eight brief walking bouts and whisking bouts in the absence of walking (red and blue arrowheads, respectively; Figure 4a–c), both of which were accompanied by pupil dilations, was more clearly seen following ROI sorting by first and second principal components (Figure 4a middle and bottom; PC1 and PC2, respectively, Figure 4c) than in Rastermap sorting (Figure 4a, top). In particular, PC1 seemed to sort neurons by transient walk and whisk activity, while PC2 appeared to sort neurons according to a more sustained component of whisking activity.

Interestingly, separation of the top and bottom fifth of neurons in the example data set sorted by PC1 (Figure 4a middle, c1 and c2, respectively; Figure 4c top and bottom, respectively) showed large groups of neurons with clear positive (Figure 4c1, top) and negative (Figure 4c2, bottom) correlations with arousal/movement, as demonstrated by alignment of rasterized neural activity with walk and whisker transients, and local maxima in pupil diameter (see Figure 4b). These separate groups of ‘ON’ and ‘OFF’ neurons were not as well identified when sorting by PC2. Specifically, neurons with high PC1 values (Figure 4c1) showed elevated activity during periods of increased arousal/movement, while neurons with low PC1 values (Figure 4c2) showed depressed activity levels during these periods. Subsets of these neurons showed either elevated basal activity with a suppression of activity during periods of enhanced arousal and/or movement (Figure 4c2, top of Rastermap), or transient increased activity leading up to movement/arousal (Figure 4c2, bottom of Rastermap), followed by transient suppression (see Video 8).

Examination of individual ‘raw’ dF/F traces confirmed that these differences were not due to artifacts of z-scoring or superneuron averaging (Stringer et al., 2023, bioRxiv) in the PC and Rastermap sorted displays (not shown). The term ‘superneuron’ here refers to the fact that when Rastermap displays more than ~1000 neurons it averages the activity of each group of adjacent 50 neurons in the sort to create a single display row, to avoid exceeding pixel limitations of the display. Each single row representing the average activity of 50 neurons is referred to as a ‘superneuron’ (Stringer et al., 2023; bioRxiv).

Performing simple correlations between dF/F in each neuron with either whisker movements, walking, or pupil diameter revealed a broad range of values (Figure 5—figure supplement 2b; whisk –0.4 to +0.6, pupil –0.6 to +0.6, and walk –0.2 to +0.4; same mouse as Figure 4, but a different session). The population distributions of these correlations were highly skewed (positive skew for whisk and walk, negative skew for pupil diameter) with modes at small positive values (Figure 5—figure supplement 2a, b). Plotting the spatial location of neurons by correlation value within the cortical map area revealed a broad distribution across the dorsal cortex of neurons whose activity was either strongly negatively or positively (or in between) correlated with walking, whisking, and/or pupil diameter, inclusively across every CCF subregion imaged (Figure 5—figure supplement 2a–c).

At the local spatial scale (Figure 5—figure supplement 2a, colored dots indicate individual neurons), neurons whose GCaMP6s activity was strongly positively correlated with movement or pupil diameter could often be found near neurons whose activity exhibited a wide variety of correlations, from strongly positive to strongly negative. Plotting the average correlation value for neurons within a CCF region revealed modest yet statistically significant (see Figure 5—figure supplement 2c, legend) spatial heterogeneities (Figure 4e, Figure 5—figure supplement 1) that were evident both in each of two different single sessions (Figure 4e; Figure 5—figure supplement 2a, b) and across sessions (Figure 5—figure supplement 2c, mean of 8 sessions from the same mouse).

To examine the spatial distribution of different Rastermap groups, we created heatmaps of the density of cells in each CCF area that belong to a particular Rastermap group (i.e. percent of cells in each region belonging to that Rastermap group; Figure 4d; only areas with at least 20 cells were considered). A Rastermap group with increased activity in relationship to walking and whisking bouts (Rastermap cell group d1 in Figure 4a, top, and Figure 4—figure supplement 2a) had a CCF density distribution ranging from 0 to 25% and was concentrated in select motor, somatosensory and visual subareas (MOs, SSn, VISrl, and VISa; CCF neural density map; Figure 4d, left). Here, CCF density distribution refers to the full set of neural density percentages (i.e. percentage of neurons in each CCF area that belong to a particular Rastermap group) across all CCF areas present in a given preparation. A second example Rastermap group, d2 (Figure 4a, top, and Figure 4—figure supplement 2c) exhibited increases in fluorescence in relationship to whisking, but not as strikingly during bouts of walking. This cell group had a CCF density distribution ranging from 0 to 13% and was concentrated in select somatosensory and auditory subareas (SSun, SSb, and AUDd; Figure 4d, right).

Next, we compared, across all CCF areas, the proportion of neurons within each CCF area that exhibited large positive correlations with walking speed and whisker motion energy. In this example dorsal mount session, walk and whisk related neurons formed the largest relative proportion of cells in select CCF areas such as motor and somatosensory subareas (MOs, SSll, and SSn; Figure 4e left and right, respectively). Mean correlations for each CCF area were generally statistically significantly greater than zero (see Figure 4e, legend).

To extend the potential diversity of neural activity patterns during spontaneous behavior, and to examine the generality of the observation of movement/arousal related neuronal activity to the lateral cortex, we monitored neuronal activity in spontaneously behaving mice with our lateral window preparation. We imaged ~85 sessions in the side mount preparation (20–90 minutes each, including both large- and multi-FOV sessions) across 11 GCaMP6s mice, for a total of ~320,000 neurons (Figure 1—figure supplement 2e; Video 10). In an example imaging session (Figure 5) of a mouse implanted with our custom 3D-printed titanium side mount headpost (Figure 1e, Figure 1—figure supplement 1f) and 10 mm radius bend cranial window (Figure 1e, Figure 1—figure supplement 1a, d, f; Video 1), we were able to image 5678 neurons at 5x5 µm/pixel resolution over a combined 5.0x4.62 mm FOV (Figure 5). For the ScanImage online fast-z motion correction documentation corresponding to this example session, see the .csv files for 3056_200924_E235_1: https://github.com/vickerse1/mesoscope_spontaneous/tree/main/online_fast_z_correction; Vickers, 2024b. Note here that, although there were no neurons present in the primary auditory cortex in the example session in Figure 5, such neurons were present in other sessions from the same mouse (Figure 5—figure supplement 2d–f), and in other side mount preparation mice in general (e.g. Figure 3d).

As in the dorsal mount (Figure 4), Rastermap sorting of neurons across this single side mount example session (Figure 5) was performed in order to further explore the broad range of relationships between movement/arousal measures and neural activity. As in the dorsal mount preparation, Rastermap sorting showed non-random patterns of neuronal activity that were often related to movement/arousal measures. Here, we identify and further characterize both a cluster with increased activity in relationship to bouts of walking/whisking (Figure 5a, top, cell group d1; see Figure 5—figure supplement 1a, b), and another manually selected neural ensemble whose activity appeared, by eye, to be more related to whisking than walking, per se (Figure 5a, top, cell group d2; see Figure 5—figure supplement 1c, b). Interestingly, these two groups consisted of ~300 neurons each spread across the right lateral cortex, with some neurons nearby in the same CCF areas, and others in distant, functionally distinct regions (Figure 5d, left, right).

Sorting by the first principal component (PC1) of activity, as in our dorsal mount preparation (Figure 4a, middle), revealed that a majority of imaged neurons exhibited activity related to bouts of walking/whisking (Figure 5a, middle; walk bouts indicated by red arrowheads, whisk bouts indicated by blue arrowheads; see Video 10 and Video 11 ). The top fifth of these movement-activated neurons showed activity nearly completely dominated by walk-related signals (Figure 5a, middle, c1, and 5 c, top, (1)), while the bottom fifth showed activity negatively correlated with walking (Figure 5a, middle, c2, and 5 c, bottom, (2)). Sorting by the second principal component (PC2), on the other hand, appeared to show increases in activity more closely aligned to the onset of whisking movements (Figure 5a, bottom). Interestingly, some neurons in PC2 exhibited sustained inter-walk bout activation (Figure 5a, bottom).

As in our dorsal mount example session (Figure 5—figure supplement 2a–c), we examined the overall spatial pattern of correlations in a side mount example session, which in this case had 3897 recorded neurons (Figure 5—figure supplement 2d–f). Here, correlations of individual neuron activity with whisker motion energy, pupil diameter, and walk speed ranged from –0.6 to +0.6 (Figure 5—figure supplement 2d, e; whisk –0.4 to +0.6, pupil –0.4 to +0.4, and walk –0.2 to +0.5) with positively skewed distributions (except for pupil diameter) and modes at small positive values (Figure 5—figure supplement 2d, e). Strong heterogeneity of these correlations, as in the dorsal preparation example (Figure 5—figure supplement 2a–c), was evident at both local (Figure 5—figure supplement 2d, e, single session) and regional (Figure 5—figure supplement 2f, mean of 8 sessions from the same mouse) spatial scales.

Examining the regional variations in mean correlations between neural activity and walk speed (Figure 5—figure supplement 2f, bottom, in red), pupil diameter (Figure 5—figure supplement 2f, middle, in gray) and whisker motion energy (Figure 6—figure supplement 1f, top, in blue) showed that neurons with high correlations to arousal/movement tended to concentrate in M1 and somatosensory upper and lower limb (as one might expect for movement related activity) and dorsal auditory areas in this example session (dark red, gray, and dark blue, respectively), while neurons with lower correlations tended to concentrate in somatosensory barrel, nose, and mouth regions (light red, gray, and blue, respectively). As in the dorsal preparation (Figure 5—figure supplement 2c), the mean correlations across 8 sessions from this example side mount mouse (i.e. the grand mean for each CCF area, equal to the mean of all means for that area across all 8 individual sessions) were general statistically significantly greater than zero (see Figure 5—figure supplement 2f, legend).

From the above results, we can see that there are clear, but complex, relationships between behavior and neuronal activity across the dorsal and lateral cortex of the mouse. Principal component analysis of this activity suggests that one major factor is variations in behavioral state, as indicated by movement/arousal (see also Musall et al., 2019; Steinmetz et al., 2019; Stringer et al., 2019). Indeed, previous studies have demonstrated that the use of dozens or more of the top principal components of the motion energy of facial movement videos is beneficial in explaining variance in visual cortical activity (Stringer et al., 2019). However, beyond the first few principal components of facial movement, the precise behavioral meaning of the subsequent components are difficult to discern (Stringer et al., 2019).

Here, we took a different approach by using a multi-step semi-automated analysis pipeline to identify higher order behavioral motifs and to examine the relationship of these motifs to the diversity of patterns of neural activity observed in the cortex.

To test the feasibility of this approach, we first created pose estimates consisting of (x,y) positions of a set of user-labeled body-parts and joints from our left, right, and posterior mouse videos using DeepLabCut (DLC; Mathis et al., 2018; see Video 1). These pose estimate outputs were used for unsupervised machine learning extraction of behavioral motifs by an algorithm called ‘Behavioral segmentation of open field in DeepLabCut’ (‘B-SOiD’; Hsu and Yttri, 2021; see Video 10). To identify high-level patterns of spontaneous head-fixed mouse behavior, we next performed manifold embedding (‘UMAP’) and cluster identification (‘HDBSCAN’) of DLC pose tracking outputs in BSOiD. A random-forests machine classifier was trained with these clusters to be able to predict, or identify, behavioral motifs in any test session with the same set or subset of labeled body part pose estimates (i.e. x-y positions for a certain number of body parts with the same names), either within or between mice.

In an example side mount session (same session as Figure 5), we identified 16 BSOiD behavioral motifs that exhibited good coarse and fine alignment to the dynamics of behavioral arousal primitives (Figure 6a), including walking, and manually selected discrete Rastermap groups from the same session. Here, BSOiD was trained on a randomly selected 80% subset of four sessions from the same mouse across three days, including the target session itself, with minimum cumulative motif times set to 1% of total video time or ~100 s, and then used to label individual video frames from the entirety of the target session. We found that, in some cases, training on individual sessions resulted in poor identification of rare and or brief behaviors, such as walking (which occurs rarely in some sessions; not shown). Initial attempts to train BSOiD on sessions across mice failed, likely due to differences in their behavioral repertoires. However, we were able to identify a subset of repeating behavioral motifs across mice. Also, the number of motifs per session and per mouse were comparable and tended to exhibit state transitions on similar timescales.

Examination of four types of qualitatively identified behavioral epochs (‘twitch’, ‘whisk’, ‘walk’, and ‘pupil oscillate’) at a finer temporal scale showed good alignment of BSOiD motifs with fine-scale features of spontaneous head-fixed mouse behavior at the level of movement and arousal primitives (Figure 6a and b, Figure 6—figure supplement 1). For example, balancing wheel movements during ‘twitch’ seem to align well to flickering into BSOiD motifs 8 and 11, while flickering into motif 13 during ‘walk’ appears to correspond well to individual periods of high walk speed during the extended walking bout (see Figure 6—figure supplement 1b). The qualitative labels that best describe these behaviors (‘twitch’, Figure 6a–c; ‘whisk’, Figure 6a–c and Figure 6—figure supplement 1; ‘walk’, Figure 6a–c and Figure 6—figure supplement 1b; ‘pupil oscillate’, Figure 6a–c) did not appear to align with or be clearly explained by simple ‘first-order’ interaction effects of movement and arousal primitives (i.e. whisk/walk/pupil x ON/OFF), despite the fact that the above-mentioned, closer, albeit preliminary, analysis of fine-scale behavioral features revealed precise and reliable alignments. This suggests that these methods will yield novel and important insights into the organization of neurobehavioral alignment across cortex, because other factors that are currently poorly understood appear to be contributing to the alignments.

Alignment of the bulk activity of all neurons in each Rastermap group with BSOiD motif transitions for this example session (Figure 6a–c, colored, vertically oriented dashed boxes) showed differential alignment with the above-mentioned four high-level, manually identified qualitative behaviors (‘twitch’, ‘whisk’, ‘walk’, and ‘pupil oscillate’). Note that, although individual BSOiD motifs could occur across multiple qualitatively identified higher level behaviors, each behavior consisted of a repeating pattern of a restricted set of motifs. For example, motif 13 in this session seems to correspond to a syllable that might best be termed ‘symmetric and ready stance’. This tended to occur just before and after walking, but also during rhythmic wheel balancing movements that appear during the ‘oscillate’ behavior.

The occurrences of these behaviors, then, were clearly aligned with both behavioral movement/arousal primitives (Figure 6b) and the activity of putative, qualitatively segmented neural ensembles (Figure 6c–d) that displayed varied spatial distributions across defined CCF areas of the side mount preparation (Figure 6e).

For example, the qualitatively identified high-level behavior‘whisk’, which in this case corresponds to whisking in the absence of walking (Figure 6a and b), with relatively elevated pupil diameter, aligns with strong activity in Rastermap group 5 (Figure 6c–d, white dashed oval corresponding to Rastermap group #5, mean trace in blue), which has a higher density of neurons in retrosplenial, secondary visual, somatosensory barrel, and dorsal auditory CCF areas (Figure 6e, R5). On the other hand, the qualitatively identified high-level behavior ‘walk’, which in this case refers to neurons with strong, transient activity aligned to seven individual bursts of walking during an extended walking bout, aligns with strong activity in Rastermap groups 6 and 7 (Figure 6c–d, white dashed oval corresponding to Rastermap groups #6 and 7, mean trace in red), which has a higher density of neurons in primary and secondary motor cortices, upper and lower limb somatosensory cortices, and primary visual cortex.

Note that the qualitative behaviors identified in Figure 6 do not correspond directly to the Rastermap group labels used in earlier figures to point out possible alignment between neural activity and behavioral primitives. For example, the ‘whisk’ group in Figure 5d, right, corresponds to a subset of the neurons with strong neural activity aligned to the ‘twitch’ behavior in Figure 6c (R3 and R4), and the walk and whisk group in Figure 5d, left, corresponds to a subset of the neurons with strong neural activity aligned to the ‘walk’ behavior in Figure 6c (R6 and R7). Furthermore, these single session results are preliminary and are not intended, at this point, as a result that necessarily generalizes across sessions or across mice. Rastermap maximum neuron densities can be found in the corresponding figure legend, but we did not apply tests of statistical significance here for these reasons. However, with ~30 CCF areas, a uniform distribution of neurons in a given Rastermap group would be ~3.3% per area. One would expect, therefore, that some of the neural density values we observed, which ranged up to ~40% for area M2 in Rastermap group 7 corresponding to the ‘walk’ qualitative behavior, would be highly significant if they were to be observed consistently across sessions.

We expect that, in general and based both on our preliminary results and emerging trends in existing literature, that the predictive contribution of high level behavioral motifs relative to movement and arousal primitives will be greater both in lateral cortical areas, such as A1, in non-primary CCF areas further up the cortical hierarchy, such as secondary sensory cortices and M2 (Wang et al., 2023; bioRxiv), and over intermediate and longer timescales of activity aligned with purposeful and/or goal oriented movements organized at higher levels of the behavioral hierarchy (Mimica et al., 2023).

We developed novel headpost designs, surgical procedures, 3D-printed accessories, experimental protocols, and analysis pipelines that allowed us to routinely perform 2p imaging of up to ~7500 neurons simultaneously across ~25 mm² of either bilateral parieto-occipital (‘dorsal mount’) or unilateral temporo-parietal (‘side mount’) cortex in separate populations of awake, behaving mice, while simultaneously monitoring an array of global arousal state variables. Correlations between neural activity and arousal/movement were broadly, but heterogeneously, represented across both local and regional spatial scales. Preliminary analysis showed that this rich dataset could be used to segregate specific mappings between high-level, qualitatively identified behaviors (e.g. ‘twitch’, ‘whisk’, ‘walk’, and ‘pupil oscillate’; Figure 6a), each consisting of a persistent pattern of robustly identifiable, high-level behavioral motifs, low-level behavioral arousal, and movement primitives (i.e. walk, whisker, and pupil; Figure 6b), and spatially localized neural activity clusters (i.e. ‘Rastermap’ groups; Figure 6c–e). Further analysis of such mappings will both allow for the fine dissection of patterned behaviors and neural activity, and for the development of a powerful framework for the sophisticated prediction of performance in mice engaged in a variety of multimodal sensory discrimination tasks (see Hulsey et al., 2023; bioRxiv, and Wang et al., 2023; bioRxiv).

Taken together, these findings show the strong potential of our novel methods for further elucidation of the principles of pan-cortical neurobehavioral alignment at the level of densely sampled individual neurons. For example, the spatial distributions of densities of neurons in each Rastermap (Stringer et al., 2023, bioRxiv) group across CCF areas were partially overlapping, yet strongly dissociable (see Figures 4d—6e). This suggested that dynamic, recurring patterns of arousal-dependent reorganization might act to enable activity mode switching between various distributed, cortical functional ‘communities’. Each of these communities, then, could be specialized for different forms of activity related to various aspects of spontaneous behavior or task performance. This topic should be further explored with principled statistical techniques for recursive, top-down hierarchical community detection (Li et al., 2020).

In summary, we have shown here that it is feasible to monitor individual neuronal activities across broad expanses of the cerebral cortex and to perform neurobehavioral alignment of high resolution behavioral arousal state motifs and pan-cortical, activity-clustered neural ensembles in awake, behaving mice. Furthermore, our preliminary findings suggest a detailed alignment between both arousal/movement primitives and high-level behavioral motifs that appears to vary across broad regions of cortex (see Figure 6). These findings extend those of earlier studies that showed widespread encoding of behavioral state during spontaneous behavior in visual cortex (Stringer et al., 2019 ; imaging and Neuropixels recordings) and encoding of uninstructed movements during task performance both across dorsal cortex (Musall et al., 2019 ; widefield imaging across dorsal cortex, 2p imaging in restricted FOVs in primary visual and secondary motor cortex) and brain-wide during task performance (Steinmetz et al., 2019; Neuropixels recordings at low volumetric sampling density).

In particular, our findings are consistent with those of Stringer et al., 2019 in that we observe widespread, strong correlations of cortical neural activity with the behavioral state variables of movement (facial and locomotor) and pupil diameter. However, we imaged larger areas of cortex simultaneously, including medial, anterior, and lateral cortical areas, at a higher overall imaging frequency (~3 Hz vs. ~1 Hz). Although admittedly this imaging speed is still not fast enough to record fast cortical encoding of sensory information or decision making, these factors, taken together, may have contributed to our ability to observe enhanced heterogeneity across these non-primary cortical areas, including in neurons whose activities were strongly suppressed during periods of increased arousal/movement (see Figure 4c, 2). Higher concentrations of such neurons were evident over lateral regions of cortex such as primary auditory and somatosensory nose and mouth regions, and in anterior regions including secondary motor cortex.

We also showed that high-level behaviors, manually identified and aligned to clusters of repeating motifs by UMAP embedding (BSOiD; Hsu and Yttri, 2021), aligned well with patterns of neural activity identified by Rastermap corresponding to neural ensembles with non-uniform spatial distributions (Figure 6). This suggests that encoding of behavioral state across cortex may occur at multiple levels of the behavioral hierarchy, in addition to that of ‘low-level’ behavioral arousal/movement primitives such as pupil diameter, whisker motion energy, and walking speed (see Mimica et al., 2023).

Our finding that neurons with no correlation to arousal/movement, or with a large negative correlation, are also widespread across cortex, suggests that multiple mechanisms or pathways may exist in parallel to distribute information about behavioral state across the brain. Interestingly, as evidenced by our discovery with Rastermap of neural ensembles exhibiting diverse response kinetics and polarities (Figure 4a, top, d, Figure 5a, top, d, and Figure 6c–e), these pathways may be activated with different temporal dynamics simultaneously across various brain areas, thus making our approach for simultaneous recording of many areas one of the keys to potentially understanding the nature of their neuronal activity. The reason for this is that correlations between behavior and neural activity across different cortical regions appear to depend on the exact time since the behavior began (Shimaoka et al., 2018). Thus, the distribution of behavioral state dwell times must be the same across the recording epochs corresponding to all cortical areas in order to detect distributed patterns of correlations between behavior and neural activity. In practice this would seem to require the simultaneous recording of all the individual neurons across the dorsolateral cortex that are part of the neural ensembles identified by Rastermap, as we have done here - accurately piecing such ensembles together across multiple separate recordings might not be possible.

For this reason, directly combining and/or comparing the correlations between behavior and neural activity across regions imaged in separate sessions may not reveal the true differences in the relationship between behavior and neural activity across cortical areas, due to a ‘temporal filtering effect’, because the correlations between behavior and neural activity in each region appear to depend on the exact time since the behavior began (Shimaoka et al., 2018). In our view, this makes the simultaneous recording of multiple cortical areas essential for proper comparison of the dependence of their neural activity on arousal/movement, because only then are the distributions of behavioral state dwell times the same across the recording epochs corresponding to all cortical areas.

Additional experiments are warranted to uncover the mechanistic basis of such differences across large populations of cortical neurons associated with changes in activity level during periods of fluctuating arousal/engagement. For example, correlations between spiking-related Ca²⁺ activity and movement/arousal under 2p imaging, both as shown here and in previous studies (Musall et al., 2019; Stringer et al., 2019), and Neuropixels recordings (Steinmetz et al., 2019; Stringer et al., 2019), appear not to mirror those observed with widefield voltage imaging of membrane potential (Shimaoka et al., 2018; especially in VIS_lm and SS_b). This suggests the existence of a complex interaction between movement/arousal related changes in membrane potential and spiking that may vary across cortical areas.

In general, our preparations allow for the observation of diverse pan-cortical neural populations that likely communicate with each other in a transient and/or sustained manner during alternating periods of low and high movement/arousal and drive different patterns of global functional connectivity. Furthermore, the apparent variability of these neurobehavioral alignments suggests that interactions between cognitive and arousal/movement encodings may also occur during spontaneous behavior, in addition to during task performance (Musall et al., 2019), and that the degree of this interaction may depend on the cortical area.

A main advantage of our approach is the flexibility afforded by the combination of our two surgical preparations (see Protocols II, III), in terms of the sheer number of cortical areas spanning distant regions that can be recorded simultaneously at single cell resolution under 2p imaging. The interoperability of our preparations across widefield 1p and Thorlabs mesoscope 2p imaging rigs (see Protocol IV) allows for the straight-forward development of an experimental pipeline that first surveys widespread activity at fast timescales (widefield,~10–50 Hz) and then investigates the activity of large subregions in a serial manner at intermediate timescales (Thorlabs mesoscope,~1–10 Hz).

A drawback of our current preparations is that they require use of a large water-immersion objective with a relatively short working distance (~2.4–3.0 mm). This limits access for simultaneous electrophysiological recordings during mesoscale imaging, even though the same preparations could be used for simultaneous imaging and electrophysiological recordings on other acquisition rigs (i.e. widefield and/or standard 2p), used in parallel to the mesoscope with the same mice.

In tandem with our headpost, surgical, behavioral, and analytic protocols, this could enable simultaneous Neuropixels, whole-cell electrophysiology, and mesoscale single-cell resolution imaging experiments. Such an approach would allow integration of neural data from multiple brain depths and areas, as well as across multiple spatiotemporal scales. Thus, these techniques would allow modification of our current methods to allow examination of faster timescale cortical dynamics than we can currently capture with 2p imaging alone, such as those related to initial encoding of sensory information and decision making.

Another limitation of our current methodologies for single large FOV imaging, as presented here, is that we imaged neurons at a single cortical depth, and at an imaging rate that did not take full advantage of the speed of the Ca²⁺ indicator that we used (i.e. GCaMP6s; we estimate that imaging at ~10 Hz would sample all available information and avoid aliasing with a rise time of ~200ms and decay time of ~1.2 s; Chen et al., 2013). To address these limitations, we are expanding our analyses to sessions, which we have already recorded, acquired at ~5–10 Hz over 4–6 FOVs at multiple z-depths (e.g. V1, A1, M2, RSpl, and SS_m/_n cortical areas; Figure 1—figure supplement 2e), with resolutions ranging from 0.2 to 1 µm per pixel (Figure 1—figure supplement 2d).

Reliability of cortical area identification and alignment in our dorsal and side mount preparations is enhanced due to standardization of headposts and cranial windows, and the combined use of skull landmark and cortical vasculature alignment in a rigorous, semi-automated widefield multimodal sensory mapping protocol. This has allowed us to acquire an extensive, standardized neurobehavioral data set (~200 hours of spontaneous behavior with 30 Hz high-resolution face and body video from three cameras,~1,000,000 total neurons at ~3–10 Hz 2p acquisition rates imaged at ~1000 x 1,000 pixels across ~250 sessions, total combined across both preparations and both FOV imaging configurations in 17 mice; see Figure 1—figure supplement 2e).

We are currently working on modifying these preparations to allow combined 2p imaging and whole-cell electrophysiological recordings (i.e. using a Thorlabs Bergamo II microscope) and/or widefield imaging and high-density electrophysiological recordings (i.e. Neuropixels; Jun et al., 2017; Peters et al., 2021). This will allow us to use a single preparation in which each mouse can be used to acquire data across multiple spatiotemporal scales during spontaneous and task-engaged behavioral epochs. Such combined recordings, with either Neuropixels or whole cell recording, may soon be possible with mesoscale 2p imaging, either on the Diesel 2p or the Thorlabs 2p mesoscope (Sofroniew et al., 2016), by use of a recently developed air-immersion objective (Yu et al., 2022, bioRxiv). A similar mesoscale air-immersion 2p objective is also under development at Thorlabs.

In the future, we hope to expand our analyses of this extensive dataset to include identification of joint Rastermap/BSOiD (or keypoint-MOSEQ motif “syllable”; Weinreb et al., 2023; bioRxiv) transitions, and to perform further precision neurobehavioral alignment, using other methodologies such as hierarchical state-space models (Lindermann, S; https://github.com/lindermanlab/ssm; LindermanLab, 2018), and factorial HMMs (Ghahramani and Jordan, 1997). These methods will allow increased accuracy over multiple behavioral timescales and the ability to predict transition or change points between different behavioral and/or neural activity states.

We will also examine changes in pan-cortical functional connectivity between communities of neurons with similar activity relationships associated with changes in behavioral state with statistical techniques such as Vector Autoregressive Union of Intersections (VA-UoI; Balasubramanian et al., 2020; Ruiz et al., 2020; Bouchard and Pabhat Snijders, 2017) that allow us to rigorously test for and remove putative but ‘false positive’ connections based on coincidental activity correlations. Together, these analysis tools, when combined with the experimental techniques demonstrated here, will enable both the identification of patterns of neurobehavioral alignment between arousal/movement and neural activity that are strongly predictive of high levels of behavioral performance, and the direct probing control of these patterns with targeted optogenetic manipulations.

All experiments were approved by the University of Oregon Institutional Animal Care and Use Committee, under protocols 20–12 and 20–14 (IACUC IDs TR202300000012 and TR202300000014). To image cortical activity, we used both male and female, 8–50 week old, CaMKII-tTA x tetO-GCaMP6s (provided by Cris Niell, University of Oregon; JAX # 007004x024742), CaMKII-Cre x Ai148 (GCaMP6f; JAX #005359x030328), and CaMKII-Cre x Ai162 (GCaMP6s; JAX# 005359x031562) mice (Jackson Labs, Allen Institute). Neural activity was typically imageable for up to between 30 and 180 days post-windowing. For experiments with combined Ca²⁺ imaging and optogenetic inhibition we used PV-Cre x Ai32 x Thy1-RGECO mice (JAX # 017320x024109 x 030528 or 030527). Habituation of each mouse to the experimental setup was performed for 2–3 days prior to 2p data acquisition during standard protocol for spontaneous behavior.

See Protocols II and III, and Supplementary methods and materials.

See Supplementary methods and materials, Overall workflow, Protocol I. See Videos 1 and 2.

See Supplementary methods and materials, Overall workflow, Protocol I, and Multimodal mapping. See Videos 3–7.

See Supplementary methods and materials, ScanImage 2p acquisition. 2p imaging was performed with a Spectra Physics Mai Tai HP-244 tunable femtosecond laser and the Thorlabs Multiphoton Mesoscope. We controlled imaging with ScanImage (2018–2021; Vidrio) running in Matlab, with GPU-enabled online z-axis motion correction. Behavioral synchronization was achieved by triggering 2p acquisition with the first right body camera frame clock signal. Left and posterior cameras were triggered by the same signal, but were subject to variable LabView runtime-related delays on the order of up to ~10 ms and, on rare occaision, dropped frames. Exposure times for both left and right body cameras, as well as frame-clock times for all three body cameras, were recorded in Spike2 (CED Power1401) as continuous waveforms and events, respectively, to allow for temporal alignment of recorded videos.

We used the ScanImage function SI render to combine image strips from large field-of-view (FOV) sessions, and in some cases we used custom Suite2p Matlab scripts (courtesy of Carsen Stringer) to align small FOVs from the same or different z-planes in a single session. Rigid and non-rigid motion correction were performed in Suite2p, along with region of interest (ROI) detection and classifier screening of ROIs likely to be neurons. Rigid motion correction computes the shift between a reference image and each frame using phase-correlation, and non-rigid correction divides the image into subsections and calculates the shift of each block, or subsection, separately (Pachitariu et al., 2016; also: https://suite2p.readthedocs.io/en/latest/registration.html).

Fluorescence intensities were calculated as changes in fluorescence divided by baseline fluorescence (dF/F), where F was calculated, for each neuron, using an ~30–60 s (100–200 frame) rolling 10th or 15th percentile baseline of the neuropil subtracted mean ROI pixel intensity (F-0.7*Fneu). Traces of dF/F were then truncated to match the length of acquired behavioral data and resampled (python: scipy.signal.resample for upsampling, or numpy: slicing for downsampling), along with behavioral data, to 10 Hz for further analysis. 2p acquisition rates for the large FOV sessions ranged from roughly 1.6–4.5 Hz, with most sessions acquired at ~3 Hz. ROIs with a Suite2p classifier-assigned cell probability less than 0.5 were excluded from further analysis.

Behavioral data was acquired with a Power1401 in Spike2 (CED) at 5000 Hz and downsampled to 10 Hz without prior filtering, during preprocessing in Python, from real measurement rates of 100 Hz at the encoder for walk speed, and 30 Hz at the face/body cameras for pupil diameter and whisker motion energy.

An example DeepLabCut (DLC) (https://github.com/DeepLabCut/DeepLabCut; DeepLabCut, 2018; Mathis et al., 2018) annotated video with 66 labeled points is shown in Video 1 (bottom right). Video snippets were concatenated, cropped, truncated, and aligned to 2p frames with FFMPEG under the Linux subsystem for Windows (Ubuntu 20.04; Windows 10 Enterprise 2004) or Ubuntu 18.04 running on local network GPU clusters. In the example session shown in Figures 5, 6 and Figure 6—figure supplement 1, we analyzed the right Teledyne Dalsa camera video (30 Hz, 1020x760 pixel with 2x2 binning, 1” x 1.8” sensor, 90 min video, ffmpeg crop compression to 525.12 MB), which was natively aligned to the 2p acquisition by a common trigger with frameclock and exposure times recorded by a CED 1401 data acquisition device in Spike2 software (Cambridge Electronic Design), running 650 K iterations (i.e. below the default of 1.03 million, to avoid potential overfitting) with the resnet_50 model in DLC (see labeled video, Video 2). Some movies had noticeable sporadic motion blur (in particular of whiskers during whisking), small areas of overexposure, and brief periods of paw occlusion or mislabeled points during walking, grooming, and other brief periods of high motor activity. These difficulties could be remedied in the future by increasing sampling frame rate (although this generates larger file sizes), as well as further optimizing IR illumination. For the preliminary analysis contained here, we performed multiple reruns of DLC to achieve an acceptable quality standard, but we did not rigorously employ tracking, exclusion, or smoothing of DLC outputs.

Pose-tracking data outputs from DLC in the form of comma-separated value files were loaded into B-SOiD (https://github.com/YttriLab/B-SOID; Hsu et al., 2021; Hsu and Yttri, 2021) along with concatenated, cropped, and spatially downsampled (FFMPEG) right-camera videos (~20–90 min,~100 MB to 1 GB per video). B-SOiD uniform manifold approximation and projection (UMAP) embedding training was performed on between 1 and 4 sessions from the same mouse. Between 20 and 60 labeling space dimensions were typically assigned per session, with between 75 and 98% of features confidently assigned to between ~2 and 15 motifs per session. Minimum bout time was set at 900 ms, and minimum cumulative motif time per session was set at ~1–5 min.

Pixel-wise analysis of cumulative motion energy showed that the motifs identified by B-SOiD differed in terms of body-region localized basic movement patterns in a manner that was consistent with human-annotated categorization labels. Thresholding of cumulative motion energy by z-score further confirmed this and made the following kinematic patterns evident: (i) In asymmetric pose motifs, a clear region of low movement (dark pixels) is visible in the chest region between the left and right shoulders that is either absent or not as prominent in high-arousal and symmetric pose motifs; (ii) Enhanced nose and mouth movement was associated with high arousal walking but could also occur independently of walking; and (iii) Increased whisker movement was associated with high arousal walking but could also occur independently of walking.