Neuronal ensembles are groups of neurons with correlated activity associated with sensory, motor, and behavioral functions. To explore how ensembles encode information, we investigated responses of visual cortical neurons in awake mice using volumetric two-photon calcium imaging during visual stimulation. We identified neuronal ensembles employing an unsupervised model-free algorithm and, besides neurons activated by the visual stimulus (termed “onsemble”), we also found neurons that are specifically inactivated (termed “offsemble”). Offsemble neurons showed faster calcium decay during stimuli, suggesting selective inhibition. In response to visual stimuli, each ensemble (onsemble+offsemble) exhibited small trial-to-trial variability, high orientation selectivity, and superior predictive accuracy for visual stimulus orientation, surpassing the sum of individual neuron activity. Thus, the combined selective activation and inactivation of cortical neurons enhance visual encoding as an emergent and distributed neural code.

Animals and surgery

All experimental procedures were conducted in accordance with the US National Institutes of Health and Columbia University Institutional Animal Care and Use Committee and were similar to our previous study14. Mice were housed in a controlled environment under a 12 h dark-light cycle at room temperature of ~23 °C and ~50% of humidity. Mice had ad libitum access to food and water. Experiments were performed in 12 adult transgenic mice (Slc17a7-IRES2-Cre, JAX stock # 023527) crossed with TIGRE2.0 Ai162 (TIT2L-GC6s-ICL-tTA2, JAX stock # 031562) maintained in C57BL/6 J congenic background. Mice were anesthetized with isoflurane (1.5–2%) while maintaining body temperature at 37 °C. Dexamethasone sodium phosphate (0.6 mg/kg) and Enrofloxacin (5 mg/kg) were administered subcutaneously and Carprofen (5 mg/kg) intraperitoneally. A titanium head-plate was attached to the skull, a 4 mm diameter craniotomy opened (center at 2.1 mm lateral and 3.4 mm posterior from bregma), and a round coverslip was implanted and sealed. Animals received postoperative Carprofen for two days. Mice were allowed to recover for five days with food and water ad libitum, and their health was checked daily.

Visual stimulation

A custom-made MATLAB application (MouSee, 2023; doi:10.5281/zenodo.7765050) was used to display visual drifting gratings on an LCD monitor positioned 15 cm from the right eye at 45° to the long axis of the animal. In 10 out of 12 mice, only the blue channel of the monitor was used (red and green channels were disabled) to avoid light contamination in the photomultiplier (PMT). No discernible difference was observed between blue and black/white drifting gratings across experiments when the imaging objective was shielded. Visual stimulation started with a blue screen (mean of grating luminescence) followed by full sinusoidal gratings (100% contrast, 0.13 cycles/deg, 5 cycles/s) drifting in 8 directions selected randomly (0º, 45º, 90º, 135º, 180º, 225º, 270º, and 315º) presented for 2 s, and a blue screen inter-gratings lasting between 1 to 5 s. Drifting gratings were presented at least 6 times for each direction during a 5-min session (at least 48 trials per experiment).

Mouse facial recording

The mouse face was recorded using an infrared monochrome camera (DMK 21BU04.H, The Imaging Source) with a zoom lens (MVL7000, Navitar) and an infrared illuminator (AI4, Tendelux). Images were acquired at 30 frames per second and stored using IC Capture software (The Imaging Source). Whisking, blinking, and sniffing behaviors were measured using a custom-made MATLAB code (MoussionEnergy, 2023; doi:10.5281/zenodo.8422691).

Volumetric two-photon calcium imaging 

Imaging experiments were performed 5 days after head-plate implantation. Each mouse was head-fixed on a wheel under a two-photon microscope (a custom-modified Ultima IV, Bruker). Animals were acclimated to the head restraint for periods of 5 to 15 minutes for at least 2 days and exposed to visual stimulation sessions before recordings. The imaging setup was enclosed with a blackout screen to avoid light contamination into the PMT. An imaging laser (Ti:sapphire, λ = 920 nm, Chameleon Ultra II, Coherent) was used to excite GCaMP6s. The laser beam at the sample (30–60mW) was controlled by a high-speed resonant galvanometer scanning an XY plane (256 × 256 pixels) at 17.7 ms (frame period) covering a field of view of 452 × 452 μm using a 25X objective (NA 1.05, XLPlan N, Olympus). An electrically tunable lens (ETL) was used to change focus (z-axis) during the recording. Imaging was performed in three planes (30-50 μm apart) recorded consecutively at a depth of 150 μm to 250 μm from the pia, pausing ~10ms between planes for ETL focus to stabilize. Thus, we collected three frames, one per depth, every ~80ms for 5 min. Imaging and ETL were controlled by Prairie View software.

Ensemble activity detection

To identify ensembles, we used a custom MATLAB program(Xsembles2P, 2023; doi:10.5281/zenodo.8423311), built upon our previous work. First, we extracted the functional neuronal network from the raster matrix. Subsequently, we filtered the raster by removing spikes from neurons that are nonfunctionally connected in each column vector. Following this preprocessing step, hierarchical clustering was applied to all column vectors using Jaccard similarity and Ward linkage (as illustrated in Figure 1d). The optimal number of clusters is determined by identifying the maximum local contrast index. Each cluster was considered an ensemble if the similarity between their vectors was statistically significant (p < 0.05), determined through a z-test. This significance was assessed by comparing the average similarity within the cluster vectors being tested to the mean and standard deviation average similarity among the same number of the cluster vectors selected randomly over 1,000 iterations. Once significant ensembles were identified, their timestamps were located.