Distinct and Dynamic ON and OFF Neural Ensembles in the Prefrontal Cortex Code Social Exploration.

KEY RESOURCES TABLE CONTACT FOR REAGENT AND RESOURCE SHARING EXPERIMENTAL MODELS AND SUBJECT DETAILS METHOD DETAILS Virus injection GRIN lens implantation miniScope Social behavior test Novel object recognition test Social behavior test with NOR control In vivo calcium imaging Ablation Immunostaining Optogenetics Behavior annotation Calcium image processing Neural activities in three testing stages PCP effects on population calcium activities Identification of behaviorally tuned neurons Neural ensemble correlation analysis Neural ensemble consistency analysis Neural ensemble functional connectivity Direct-exploration tuned ensemble activity Neural ensemble overlap analysis Behavioral variable decoding QUANTIFICATION AND STATISTICAL ANALYSIS Statistics Sample size estimation Data inclusion and exclusion DATA AND SOFTWARE AVAILABLILTY CONTACT FOR REAGENT AND RESOURCE SHARING Further information and requests for resources and reagents should be directed to and will be fulfilled by the Lead contact, Dr. Yun Li ( yli30@uwyo.edu ).

EXPERIMENTAL MODEL AND SUBJECT DETAILS

All experiments were conducted in accordance with the guidelines of the Institutional Animal Care and Use Committee, the Intramural Research Program, National Institute on Drug Abuse, National Institutes of Health. All mice were wild type C57BL/6J at 3 – 4 months of age and 25 – 30 gram of weight. Mice were maintained in a regular light cycle (7:00am – 7:00pm) and provided with food and water ad libitum. All mice used in this study were males.

METHOD DETAILS Virus injection

We previously demonstrated that GFP protein driven by CamKII promoter in AAV1 virus is mostly expressed in excitatory neurons (i.e., only 1% GFP positive neurons were GABAergic neurons) once being injected into mouse prelimbic cortex ( Zhang et al., 2017 ). In this study, for calcium imaging experiments, we unilaterally injected AAV1.CamKII.GCaMP6f.WPRE.SV40 virus (University of Pennsylvania Vector Core, 500 nl with a titer of 2.76e13 GC/ml) into the dorsal region of the mouse prelimbic cortex, using the stereotactic coordinates (A/P: +1.9 mm, M/L: +0.3 mm, D/V: −1.7 mm). For ablation experiments, we bilaterally injected AAV1 viral mix (AAV1.CamKII0.4.Cre.SV40 and AAV1.mCherry.Flex.dtA, 500 nl in a ratio of 1:1, with a title of 4.496e13 GC/ml and 2.06e13 GC/ml, respectively) into the dorsal region of the mouse prelimbic cortex. For optogenetic experiments, we bilaterally injected AAV1 viral mix (AAV1.CamKII0.4.Cre.SV40 and AAV1.CAGGS.Flex.ChR2-tdTomato.WPRE.hGH, 500 nl in a ratio of 1:1, with a title of 4.496e13 GC/ml and 1.38e13 GC/ml, respectively), or AAV1 viral mix (AAV1.CamKII0.4.Cre.SV40 and AAV1.EF1a.DIO.eYFP.WPRE.hGH, 500 nl in a ratio of 1:1, with a title of 4.496e13 GC/ml and 2.9e13 GC/ml, respectively) as controls, into the dorsal region of the mouse prelimbic cortex. Mice were anaesthetized with 2% isoflurane in oxygen at a flow rate of 0.4 liter/min and mounted on a stereotactic frame (Model 962, David Kopf Instruments). A hole was drilled on the skull using a 0.5-mm diameter round burr with a high-speed rotary micro drill (19007–05, Fine Science Tools). A total of 500 nl of virus solution were injected at a rate of 25 nl/min using a micro pump and Micro4 controller (World Precision Instruments). The injection needle was kept in the parenchyma for 5 minutes after injection. The hole on the skull was then sealed with bone wax, and skin was sutured.

GRIN lens implantation

One or two weeks after AAV injection, a 1-mm diameter GRIN lens (GRINTECH GmBH) was directly implanted into the dorsal region of the mouse prelimbic cortex under anesthesia with ketamine/xylazine (ketamine:100 mg/kg and xylazine:15 mg/kg). In brief, a 1.1mm-diameter craniotomy was first made at the coordinates (A/P: +1.9 mm. M/L: +0.7mm). Before GRIN lens implantation, brain tissue above the prelimbic cortex was aspirated using a 30-gauge blunted needle, along a direction of 10° laterally-shifted angle into a depth of −1.8 mm. The needle was attached to a custom-constructed three-axis motorized robotic arm controlled by a MATLAB-based software ( https://github.com/liang-bo/AutoStereota ). The GRIN lens was then secured to the skull using dental cement (Metabond S380, Parkell). MiniScope Detailed information about the miniScope can be found in our previous study ( Barbera et al., 2016 ). Design files and part list can be found in the Github repository ( https://github.com/giovannibarbera/miniscope_v1.0 ). The miniScope consists of a blue LED (465 nm XLamp XP-E, Cree), a filter set (excitation filter, ET470/40, Chroma; dichroic mirror, FF495, Semrock; emission filter, EM525/50, Chroma), the relay optics (#83–605, #63–690, Edmund optics) and the CMOS sensor (MT9V022IA7ATM, Aptina). The miniScope housing was 3D printed (SLArmor Nickel-NanoTool, Proto Labs). The LED power was set to 0.1 – 0.3 mW at the miniScope focal plane. The image size was 400 × 400 pixels, acquired at 10-Hz frame rate. The pixel size of the acquired image was ~ 2.75 μm. Three to eight weeks after the GRIN lens implantation, the miniScope was mounted onto mouse head. After achieving the in-focus position for the entire field of view, the base of the miniScope was fixed on the skull using dental cement, and the main body of the miniScope was detached. Before each in vivo imaging experiment, the main body of the miniScope was connected and secured to the base on the mouse head through a side-locking screw. After data acquisition, the main body of the miniScope was detached from the base. We concurrently recorded calcium activities from hundreds of excitatory neurons in the dorsal prelimbic cortex, when mice freely performed the social behavior test ( Movie S1 ).

Social behavior test

We set up a social behavior test based on the previously established protocol known as “Crawley’s sociability and preference for social novelty test” ( Moy et al., 2004 ; Moy et al., 2007 ). Crawley’s sociability and preference for social novelty test was originally performed with a three-chamber apparatus. To facilitate in vivo calcium imaging, we modified the three-chamber apparatus by placing the subject mouse into one open square box (42 cm × 42 cm × 30 cm) with two small cylinder wire containers sitting at the opposite corners of a diagonal line (also see schematic illustration in Figure 1B ). The behavior test was composed of three consecutive 10-minute testing stages: “habituation”, “sociability”, and “social novelty”. During the “habituation” stage, the subject mouse freely explored the apparatus with two empty wire containers. During the “sociability” stage, an age- and weight-matched never-before-met same-sex (i.e., male) conspecific (stranger1) was introduced into one container while the other container remained empty. During the “social novelty” stage, a second age- and weight-matched never-before-met same-sex conspecific (stranger2) was introduced into the previous empty container during sociability stage while the first conspecific remained in the same location. “Sociability” was defined as propensity f or a subject mouse to spend time with a conspecific rather than with an empty container during the “sociability” stage. “Preference for social novelty” was defined as propensity for a subject mouse to spend time with a previously un-encountered conspecific rather than with a relatively familiar conspecific during the “social novelty” stage.

Novel object recognition test

A novel object recognition (NOR) assay was designed to examine mouse exploratory behavior to a non-social novel stimulus ( Leger et al., 2013 ). The NOR test contained two stages, a “familiarization” and a “testing” stage (also see schematic illustration in Figure S8B ). During the “familiarization” stage, mice freely explored two identical objects in a square box (42 cm × 42 cm × 30 cm) for 10 minutes. The two objects were placed at the same two corners of the diagonal line as that of the social behavior test. Following familiarization, mice were placed back to their home cages for 1 hour. During the “testing” stage, mice were re-introduced back into the same behavior apparatus – with one of the familiarized objects being replaced by a novel object. Mice explored these two objects for 10 minutes. Social behavior test with NOR control To further distinguish exploration of social salience versus exploration of novel stimuli, we performed the following social behavior test, incorporating two novel object controls (also see schematic illustration in Figure S8A ). The entire behavior test was composed of five consecutive 5-minute stages, including one “habituation” stage (H stage) and four “testing” stages (T1 to T4 stages). During the H stage, mice freely explored two empty wire containers (E1 and E2) in a square box (42 cm × 42 cm × 30 cm). E1 and E2 were placed at the opposite corners of a diagonal line. Following the H stage, a novel object (object1) was placed into the previously empty container E1 during the T1 stage. During the T2 stage, object1 was replaced by a novel conspecific (stranger1). During the T3 stage, a second novel object (object2) was placed into the previously empty container E2. During the T4 stage, object2 was replaced by a second novel conspecific (stranger2). In vivo calcium imaging A total of 18 mice were used for in vivo calcium imaging via miniScope during various behavior tests. These mice were unilaterally injected with AAV1 virus expressing calcium indicator GCamp6f mostly in excitatory neurons in the mPFC, and implanted with a 1-mm GRIN lens into the dorsal region of the prelimbic cortex. For the first batch of 9 mice, we performed in vivo calcium imaging with the NOR test first, followed by two social behavior tests. A week after the NOR test, the first social behavior test was performed in the absence of phencyclidine (PCP); while the second social behavior test was performed in the presence of PCP after another 5–14 days. PCP was intraperitoneally injected at a dose of 10 mg/kg, 30–50 minutes prior to the social behavior test. The objects in the NOR test and the strangers in the two social behavior tests were presented at the same locations (i.e., the same two corners of a diagonal line). With a second batch of 3 mice and a third batch of 6 mice, we performed social behavior tests in the absence of PCP multiple times (i.e., 4 times and 5 times, for the second and third batch of mice, respectively), a NOR test, and lastly a social behavior test in the presence of PCP. The timeline of the five social behavior tests (Test1 to Test5) was illustrated in Figure S4A . Specifically, one week after social behavior Test1, social behavior Test2 was performed with strangers presented at the same locations as in Test1; one week after social behavior Test2, social behavior Test3 was performed with strangers presented at different locations (i.e., two corners on the opposite diagonal line from the Test2); one to two weeks after social behavior Test3, social behavior Test4 and Test5 were performed on the same day (1 hour apart) with strangers presented at different locations. NOR test was performed between social behavior Test1 and Test4. After 3–14 days of social behavior Test4 and Test5, the social behavior test was performed again in the presence of PCP. PCP was intraperitoneally injected at a dose of 10 mg/kg, 30–50 minutes prior to the social behavior test. The second batch of 3 mice did not perform the social behavior Test3, while the rest of social behavior tests (i.e., Test1–2, Test4–5) were similar to those in the third batch of 6 mice. In all our social behavior tests, stranger mice were never-before-met conspecifics of the same gender (i.e., male) with matched age and weight. Ablation In the ablation experiment, we examined if focal ablation of excitatory neurons bilaterally in the mPFC would induce any abnormality in our social behavior test. After a pre-ablation social behavior test, we bilaterally injected AAV1 viral mix (AAV1.CamKII0.4.Cre and AAV1.mCherry.Flex.dtA) into the mPFC to express diphtheria toxin fragment A selectively in excitatory neurons in the mPFC. Five months after viral injection, a post-ablation social behavior test was performed with new pairs of strangers. Immunostaining After the post-ablation social behavior test, mPFC ablated mice and age-matched control mice were anesthetized with an overdose of ketamine (150 mg/kg) and xylazine (22.5 mg/kg), perfused with phosphate buffer solution (PBS) first and followed by a fixation buffer containing 4% paraformaldehyde (PFA) in PBS. Mice brains were dissected and post-fixed with 4% PFA in PBS overnight at 4°C. Fixed mice brains were sectioned via a Vibratome (LEICA VT1000S) and 30 μm coronal brain slices were collected. To evaluate the ablation efficiency, immunostaining of NeuN, a neuronal marker, was performed following a protocol for free-floating brain slice. In brief, floating brain slices were first washed in PBS (5 minutes for 3 times) and incubated in a blocking buffer (4% normal goat serum, 1% bovine serum albumin, 0.3% triton X-100 in PBS) with gentle rocking for two hours at room temperature. Brain slices were then incubated with a monoclonal mouse-anti-NeuN antibody (1:1000 diluted in blocking buffer, ab104224, Abcam) with gentle rocking overnight at 4°C. After washing in PBS (5 minutes for 3 times), brain slices were subsequently incubated with a secondary antibody (Alexa Fluor 488 conjugated donkey-anti-mouse IgG, 1:300 diluted in blocking buffer, Cat#715–546-150, Jackson ImmunoResearch) with gentle rocking for 1 hour at room temperature. After washing in PBS (5 minutes for 3 times), brain slices were mounted on slides for imaging. Fluorescent images were obtained through an Olympus VS120 scanner. For quantification, fluorescence intensity of NeuN staining was measured at the ablated region and an adjacent cortical region with a similar size. The ratio was calculated and compared between the ablated mice and controls.

Optogenetics

In the optogenetic experiment, we examined if transient disturbance of excitatory neural activity in the mPFC would induce any abnormalities in mice social exploration using our social behavior setup. We bilaterally injected AAV1 viral mix (AAV1.CamKII0.4.Cre and AAV1.CAGGS.Flex.ChR2-tdTomato) into the mPFC to express a light-gated cation-selective channel, channelrhodopsin-2 (ChR2), specifically in excitatory neurons in the mPFC. As a control, we bilaterally injected AAV1 viral mix (AAV1.CamKII.Cre and AAV1.EF1a.DIO.eYFP) into the mPFC. Two weeks after viral injection, customized optical fibers were bilaterally implanted into the dorsal region of the mouse prelimbic cortex (A/P: +1.9 mm, M/L: ±0.3 mm, D/V: −1.7 mm). Four months later, we first performed the social behavior test under a laser-ON condition. Specifically, following habituation, mice received a brief light stimulation from a 488-nm laser (10-ms long pulse at 20 Hz, laser ON for 20 seconds, ~ 7 mW at the fiber tip), right before the “sociability” and “social novelty” testing stages. One week later, we repeated the social behavior test with new pairs of strangers under the laser-OFF condition. Mice were connected with the optical fiber with the laser off and received a sham stimulation to mimic the auditory stimulation from the laser shutter (Vincent Associates). Specifically, following habituation, the shutter control was run for 20-seconds (10-ms long pulse, 20 Hz) right before the “sociability” and “social nove lty” testing stages.

Behavior annotation

The mouse social behavior test and NOR test were recorded using a top-view camera and a side-view camera (BFLY-PGE-12A2C-CS, Point Gray). The behavior images were synchronized with the calcium images by an external TTL signal from the miniScope controller. The behavior cameras had the same frame rate as the miniScope. Mouse behavioral activities were analyzed manually with frame-by-frame annotation based on an open-source MATLAB graphical user interface tool ( https://github.com/pdollar/toolbox ) ( Lin et al., 2011 ). Five distinct behavioral activities were annotated, including direct-exploration, proximity-exploration, irrelevant-exploration, ambulation and self-directed activities. Criterial for these annotated behaviors are listed below: (1) Direct-exploration labelled the scenario when the subject mouse proactively contacted the wire container through nose-poking, biting and paw-grabbing. (2) Proximity-exploration labelled the scenario when the head of the subject mouse was inside of the adjacent areas (within 5 cm of distance) of the wire container, but without any physical contact. (3) Irrelevant-exploration labelled the scenario when the subject mouse physically contacted walls of the apparatus through nose-poking or paw-touching. (4) Ambulation was defined as when the head of the subject mouse was physically outside of the adjacent areas of the wire container and moving about with measurable position (distance) changes. (5) Self-directed activities grouped several behavioral categories, including immobility, self-grooming and other self-directed subtle-movements. These behaviors were referred to as “self-directed activities” because mice presumably received less external information while performing these activities. Particularly, “immobility” was defined as when the subject mouse kept the same posture for more than 10 frames (1 second) at any place of the apparatus; “Self-grooming” was defined as previously reported ( Hong et al., 2014 ), when the subject mouse performed head-washing, nose- or face- grooming, head- or body- scratching, body-, leg-, paw- or genital- licking at any place of the apparatus; Other self-directed subtle-movement was defined as when the head of the subject mouse was physically outside of the adjacent areas of the wire container and performing subtle body posture changes without a measurable position (distance) change. Please also see Movie S2 for behavior annotation from a representative mouse. For each behavior, a binary behavior vector was constructed to represent the behavior periods (behavior is occurring: 1, behavior is not occurring: 0). The direct-exploration vector was further decomposed to direct-exploration of subject1 (H: E1; SB: S1; SN: S1) and direct-exploration of subject2 (H: E2; SB: E2; SN: S2). Similarly, the proximity-exploration vector was further decomposed to proximity-exploration of subject1 (H: E1; SB: S1; SN: S1) and proximity-exploration of subject2 (H: E2; SB: E2; SN: S2). Abbreviations H, SB and SN referred to habituation, sociability and social-novelty stage. E1 and E2 referred to the two empty containers, and S1 and S2 referred to stranger1 and stranger2, respectively.

Calcium image processing

Calcium images were processed and analyzed using custom scripts in MATLAB (Mathworks). Raw images were stabilized by determining the peak of the cross correlation with their neighboring images ( Guizar-Sicairos et al., 2008 ) to compensate for the brain position translations caused by mouse movements. To align the calcium images collected from different behavior tests, e.g. test 1 and test 2, we first calculated the maximum-projection image of all frames collected for each test. The maximum-projection image of test 1 was used as the reference image. We then used the MATLAB function “imregdemons ” (Image processing toolbox) to estimate the displacement field between any two maximum-projection images. We generated the registered image by applying the estimated displacement field to the maximum-projection image of test 2. Subsequently, we visually inspected the overlaid image containing both the reference and registered images with complementary colors to ensure that the estimated displacement field was appropriate ( Figure S3D ). We then applied the estimated displacement field to all the frames collected during the social behavior test 2. Cell identification was performed using the algorithm previously described ( Barbera et al., 2016 ). For each mouse, a cumulative cell map including all active neurons identified from all behavior tests was generated and manually verified. Calcium trace of each active neuron was extracted using detailed procedures below: (1) For a given neuron, a region of interest (ROI) was assigned with a circular region corresponding to the size of soma (diameter = 15 μm) and the centroid indicated the detected cell position. The minimum fluorescence of image-stack was used as the baseline fluorescence of the background (referred to as F b ). The fluorescence of ROI (F ROI ) was calculated as the averaged fluorescence after background subtraction from the raw image fluorescence (F raw ), i.e., F ROI = ROI , brackets indicate the average over ROI pixels. (2) A commonly used procedure ( Barbera et al., 2016 ; Chen et al., 2013 ; Kerlin et al., 2010 ; Pinto and Dan, 2015 ) was then applied to correct the potential contamination from out-of-focus neurons and neuropil fluorescence, i.e., F sig = F ROI − F con , where F sig was the true fluorescence signal, F con was the contamination fluorescence. F con was calculated over background subtracted images, using the averaged fluorescence from an annular zone (d1 = 20 μm, d2 = 30 µm) surrounding the soma ROI but avoiding other ROI regions. The slowly variable baseline due to background/neuropil activity was also removed in this procedure. (3) Each calcium trace ΔF/F was calculated as F sig /F 0 , where F 0 was the baseline fluorescence over the neuron ROI. (4) The baseline noise level for each calcium trace was estimated from the power density spectrum at a high-frequency band ( Pnevmatikakis et al., 2016 ). For each testing stage, active neurons were defined only if the maximum value of the calcium trace was 3 times greater than its baseline noise level. The calcium traces were smoothed with a 500-ms sliding window. For each social behavior test, a neuron needed to be active at least once during the three testing stages to be identified as active neuron. Neural activities in three testing stages We calculated the neural calcium activity changes in the presence of restrained social targets ( Figure 1F ). Data of social behavior test 1 (n = 18 mice) were used in this calculation. For each neuron, calcium activities at the three social behavior testing stages were compared first by the Kruskal-Wallis test. A neuron was defined as “unchanged” if the p-value of the Kruskal-Wallis test was greater than 0.05. If the p value of the Kruskal-Wallis test was less than 0.05, multiple comparisons with Tukey-Kramer correction were performed subsequently. A neuron was defined as “SB&SN increased” if neural activities at both the “sociability” and “social novelty” stages were significantly higher than that of the “habituation” stage (p < 0.05, two-side). A neuron was defined as “SB increased only”, if the neural activity at the “sociability” stage was significantly higher than that of the “habituation” stage (p < 0.05, two-side), while the neural activity at the “social novelty” stage was not significantly different from that of the “habituation stage” (p > 0.05, two-side). A neuron was defined as “SN increased only”, if the neural activity at the “social novelty” stage was significantly higher than that of the “habituation” stage (p < 0.05, two-side), while the neural activity at the “sociability” stage was not significantly different from that of the “habituation” stage (p > 0.05, two-side). Neurons defined as “SB&SN decreased”, “SB decreased only” or “SN decreased only” followed the same principle. PCP effects on population calcium activities To evaluate the PCP effect on calcium activities at the neural population levels, in vivo calcium imaging during a 10-minute habituation test was performed on a group of 5 mice injected with GCaMP6f virus and implanted with a GRIN lens, with the same apparatus set up as the habituation stage of the social behavior test as described above. These mice were subsequently intraperitoneally injected with 10 mg/kg PCP, and 30–50 minutes later, in vivo calcium imaging was performed again during a second 10-minute habituation test. As control, in vivo calcium imaging during the habituation tests with intraperitoneal saline injection was performed on a different group of three mice injected with GCaMP6f and implanted with a GRIN lens. Calcium images were processed using the same methods described above. Active neurons were identified and calcium traces were extracted. For each neuron, the averaged activity of the 10-minute recording before and after PCP or saline injection was calculated and denoted as C pre and C post , respectively. Averaged calcium activity of all the neurons from all mice were pooled together for comparisons between C pre and C post of PCP and saline injection, respectively. The Kolmogorov–Smirnov test was used to compare distributions of the averaged activity before and after PCP ( Figure S6D ) or saline injection ( Figure S6E ). We also calculated the activity change index by (C post − C pre )/ (C post + C pre ), for before and after PCP or saline injection, respectively ( Figure S6D-E ). The value of activity change index was bound between −1 and 1. Positive value indicated increased activity. Negative value indicated decreased activity.

Identification of behaviorally tuned neurons

For each testing stage, we identified behaviorally tuned neurons corresponding to each of the three annotated behaviors (i.e. direction-exploration, self-directed activities and irrelevant-exploration), following the same analytical steps as described below (also see Figure S2A-C ): (1) For any given neuron n, we first calculated the similarity between calcium trace (ΔF/F) C n and behavior vector B. Similarity was defined as normalized inner product of two vectors, 2B·C n /(|B| 2 +|C n | 2 ) ( Carrillo-Reid et al., 2015 ; Hamm et al., 2017 ). The value of similarity was bound between 0 and 1. A value of 1 meant these two vectors were identical; while a value of 0 meant they were completely different. (2) We then randomly shuffled the behavior epochs and calculated the similarity between the shuffled behavior vector and the calcium trace C n for a given neuron. We repeated this permutation process 5,000 times to generate a similarity distribution representing the similarity distribution histogram predicted by chance for a given neuron. (3) A neuron was defined as an “ON neuron” only if its actual similarity was greater than the 99.17% quantile of chance similarity distribution histogram generated by random shuffling. Consequently, ON neurons were those exhibiting significantly higher activities than that predicted by chance during the annotated behavior epochs. Conversely, a neuron was defined as an “OFF neuron” only if its actual similarity was lower than the 0.83% quantile of chance similarity distribution histogram generated by random shuffling. Therefore, OFF neurons were those exhibiting significantly lower activity than that predicted by chance during the annotated behavior epochs. Neurons with similarity falling between the 0.83% and 99.17% quantiles of chance similarity distribution histogram generated by random shuffling, were defined as “Other neurons” (also see Figure S2B-C ). For each mouse, we applied the calculations described above to all active neurons to identify the ON and OFF neurons correlated with each of the three annotated behaviors during different testing stages. We defined each group of ON neurons (or OFF neurons) as one “neural ensemble”. Consequently, we identified a total of 9 pairs of mPFC neural ensembles for each mouse, including Direct-ON H and Direct-OFF H , Self-ON H and Self-OFF H , Irrelevant-ON H and Irrelevant-OFF H during the H stage; Direct-ON SB and Direct-OFF SB , Self-ON SB and Self-OFF SB , Irrelevant-ON SB and Irrelevant-OFF SB during the SB stage; Direct-ON SN and Direct-OFF SN , Self-ON SN and Self-OFF SN , Irrelevant-ON SN and Irrelevant-OFF SN during the SN stage (see Figure 2 and Figure S2 ).

Neural ensemble correlation analysis

For each identified ON and OFF neuron, z-scored calcium activity was calculated by (C n – μ )/σ, where μ and σ were the mean and standard deviation of C n across H, SB and SN stages. For each ON and OFF neural ensemble, the averaged z-scored calcium activity was calculated by pooling the z-scored calcium activities from all identified ON and OFF neurons within the ensemble ( Figure 2A and Figure S2F ). We also calculated the averaged z-scored calcium activity across all behavior epochs (with an interval of −4 s to +4 s) for each ON and OFF neuron, and sorted all identified ON neurons based on the time of their maximal peak activities, and OFF neurons based on the time of their minimal activities ( Figure 2D ). To compare calcium dynamics around the behavior onset and the behavior ending, for each annotated behavior, we calculated the averaged z-scored calcium activity by pooling all ON or OFF neural ensembles tuned to the same behavior but identified at different testing stages ( Figure S2D ). We then compared the averaged calcium activity from all 18 mice, pre- (−2 s to 0 s) and post- (0 s to +2 s) behavior onset and ending for each annotated behavior ( Figure S2E ). To study the correlation between calcium activity of neural ensembles and specific behaviors, we calculated both (1) the Pearson correlation coefficient between the behavior vector and the averaged calcium activity of all neurons from a given neural ensemble (referred to as “group-correlation coefficient” in Figure 2E and Figure S3A ), and (2) the average of the individual Pearson correlation coefficients between the behavior vector and each individual neural activity (referred to as “individual-correlation coefficient” in Figure 2E and Figure S3A ).

Neural ensemble consistency analysis

We evaluated the reliability of ON and OFF neuron’s responses across all behavioral epochs ( Figure 2C ). For an ON neuron, we defined an “engaged” behavior epoch when the ON neuron exhibited activity 3 times higher than the baseline noise; we defined a “not engaged” behavior epoch when the ON neuron did not exhibit activity 3 times higher than the baseline noise. By contrast, for an OFF neuron, we defined an “engaged” behavior epoch when the OFF neuron did not exhibit activity 3 times higher than the baseline noise; we defined a “not engaged” behavior epoch when the OFF neuron exhibited activity 3 times higher than the baseline noise. We evaluated the consistency of ON and OFF neuron’s responses through calculating the percentage of “engaged” behavior epochs across all behavior epochs within each testing stage. We also calculated for any given behavior epochs, the percentage of ON or OFF neurons being “engaged” within each ON and OFF ensemble. Neural ensemble functional connectivity Functional connectivity within and between neural ensembles was evaluated by calculating the averaged pairwise Pearson correlation coefficient of the calcium traces within and between behaviorally tuned neural ensembles ( Figure 3B ). The connectivity of an ON ensemble tuned to a specific behavior was evaluated by the averaged pairwise Pearson correlation coefficient of the calcium traces within the ON ensemble. The connectivity of an OFF ensemble tuned to a specific behavior was evaluated by the averaged pairwise Pearson correlation coefficient of the calcium traces within the OFF ensemble. The connectivity between ON and OFF ensembles tuned to the same specific behavior was evaluated by the averaged pairwise Pearson correlation coefficient of the calcium traces between ON and OFF ensembles. “Other” neurons were defined as neurons not belonging to ON or OFF ensembles tuned to any of the three annotated behaviors. The connectivity of “other” neurons was evaluated by the averaged pairwise Pearson correlation coefficient of the calcium traces within these neurons. We also evaluated the connectivity between ON (or OFF) neurons and “other” neurons by the averaged pairwise Pearson correlation coefficient of the calcium traces between ON neurons (or OFF neurons) and “other” neurons. Direct-exploration tuned ensemble activity Calcium activity of Direct-ON and Direct-OFF ensembles during direct-exploration of the two targets at each testing stage was calculated ( Figure 8 ). The calcium trace of each neuron from a given neural ensemble was first normalized, using C norm = (C - C min )/(C max -C min ), where C max and C min were the maximum and minimum values of C, respectively. Subsequently, for each Direct-ON (or Direct-OFF) neuron, the average of the normalized calcium activity during direct-exploration on each of the two targets were calculated, respectively. Lastly, the averaged activity of all Direct-ON (or Direct-OFF) neurons for each mouse at each testing stage was calculated for quantification.

Neural ensemble overlap analysis

For a given pair of ensembles, A and B, the degree of observed overlap was calculated by the Sørensen–Dice index: 2N A∩B /(N A +N B ), with N A indicating the neuron number of ensemble A and N B indicating the neuron number of ensemble B, and N A∩B indicating the neuron number of overlaps between ensemble A and B. The overlap degree was 0 if both A and B were empty. The observed overlap was then compared to what would be predicted by chance. The probability of one neuron to be randomly assigned into ensemble A was N A /N, where N is the total imaged neuron number. Similarly, the probability of one neuron to be randomly assigned into ensemble B was N B /N. If both assignments were random and independent, the probability of one neuron to be simultaneously assigned into both ensemble A and B was N A N B /N 2 . Thus, the degree of overlap predicted by chance was calculated by 2N A N B /(N(N A +N B )). (A) Overlap within the same testing stage ( Figure 4C-D ): To evaluate the dynamic feature of neural ensembles, we split each 10-minute recording (one testing stage) into two 5-minute halves and re-identified ON and OFF neurons tuned to direct-exploration (Direct-ON and Direct-OFF neurons). We calculated the overlap between Direct-ON neurons of the two halves, and the overlap between Direct-OFF neurons of the two halves ( Figure 4C ). Next, to evaluate the reliability of our neuron identification process, we generated a union ensemble (Half1 ∪ Half2) by pooling together all the identified Direct-ON (or Direct-OFF) neurons from the two 5-minute halves and calculated the percentage of Direct-ON (or Direct-OFF) neurons originally identified based on the full-length 10-minute recording belonging to the union ensembles. (B) Overlap across different stages within one social behavior test ( Figure 4A-B , Figure S3C ): To clarify the across-stage overlaps between ON and OFF ensembles identified at different stages of a social behavior test, we defined across-stage overlap between ON-ON and OFF-OFF neural ensembles tuned to the same behavior as “ON-ON and OFF-OFF intra-behavior” overlap (e.g., neurons identified as Direct-ON at the “habituation” stage and remained as Direct-ON at the “sociability” or “social novelty” stage); across-stage overlap between ON-ON and OFF-OFF neural ensembles tuned to different behaviors as “ON-ON and OFF-OFF inter-behavior” overlap (e.g., neurons identified as Direct-ON at the “habituation” stage but Self-ON at the “sociability” stage); and across-stage overlap between ON-OFF neural ensembles tuned to the same behavior as “ON-OFF intra-behavior overlap” (e.g., neurons identified as Direct-ON at the “habituation” stage but Direct-OFF at the “soci ability” stage). The ON-ON intra- and inter-behavior overlaps between neural ensembles were illustrated in Figure 4A left panel. The OFF-OFF intra- and inter-behavior overlaps followed the same principle. The ON-OFF intra-behavior overlap was illustrated in Figure 4B left panel. For the quantification of ON-ON and OFF-OFF intra- and inter-behavior overlaps, we calculated the average of all the possible pairwise overlaps for each mouse (9 pairs per mouse for ON-ON or OFF-OFF intra-behavior overlap, and 18 pairs per mouse for ON-ON or OFF-OFF inter-behavior overlap, as illustrated in Figure 4A left panel). We also quantified ON-ON and OFF-OFF intra-behavior overlaps for each annotated behavior, by calculating the average of 3 possible pairwise overlaps for each mouse ( Figure S3C ). For quantification of ON-OFF intra-behavior overlap for each annotated behavior, we calculated the average of all the possible pairwise overlaps for each mouse (6 pairs per mouse for each annotated behavior, as illustrated in Figure 4B left panel). The corresponding overlap to be predicted by chance was calculated as described above. (C) Overlap between two different social behavior tests ( Figure 5 , Figure S4 and S5 ): We performed four different types of overlap comparisons for ON and OFF ensembles from all the social behavior tests (See In vivo calcium imaging section and Figure S4A ): (1) Social behavior Test1 versus Test2, performed on different days with social targets presented at the same locations; (2) Social behavior Test2 versus Test3, performed on different days with social targets presented at the different locations; (3) Social behavior Test4 versus Test5, performed on the same days with social targets presented at different locations; (4) Social behavior test in the absence of PCP versus social behavior test in the presence of PCP, performed on different days with social targets presented at the same locations. For the first three comparisons (i.e., Test1 versus Test2, Test2 versus Test3, Test4 versus Test5), (a) we first calculated for each annotated behavior, the ON-ON and OFF-OFF intra-behavior overlap across the same testing stages of the two tests (e.g., Direct-ON at the “sociability” stage of Test1 versus Direct-ON at the “sociability” stage of Test2, as illustrated in Figure S4B ). For quantification, there was only one pairwise comparison per mouse for each same-stage overlap of each annotated behavior ( Figure S4C-E ). (b) We then compared for each annotated behavior, the ON-ON and OFF-OFF intra-behavior overlap across same testing stages versus across different testing stages of the two tests (e.g., Direct-ON at the “habituation” stage of Test1 versus Direct-ON at the “sociability” stage of Test2, as illustrated in Figure S4F ). For quantification, we pooled all possible pairwise comparisons of same-stage overlaps (3 pairs per mouse) or different-stage overlaps (6 pairs per mice) for each annotated behavior. (c) Next, we grouped all three annotated behaviors to compare ON-ON and OFF-OFF intra-behavior overlap versus ON-ON and OFF-OFF inter-behavior overlap, both across same testing stages and across different testing stages of the two tests (as illustrated in Figure S5A ). Therefore, there were four types of ON-ON (or OFF-OFF) ensemble overlaps in this comparison: the overlap of “intra-behavior, same-stage”, the overlap of “intra-behavior, different-stage”, the overlap of “inter-behavior, same-stage”, and the overlap of “inter-behavior, different-stage”. Specifically, the ON-ON overlap of “intra-behavior, same-stage” was defined as the overlap of the two ON ensembles tuned to the same behavior identified in the same stages of the two tests (9 pairs per mouse). The ON-ON overlap of “intra-behavior, different-stage” was defined as the overlap of the two ON ensembles tuned to the same behavior identified in different stages of the two tests (18 pairs per mouse). The ON-ON overlap of “inter-behavior, same-stage” was defined as the overlaps of the two ON ensembles tuned to different behaviors identified in the same stages of the two tests (18 pairs per mouse). The ON-ON overlap of “inter-behavior, different-stage” was defined as the overlaps of the two ON ensembles tuned to different behaviors identified in different stages of the two tests (36 pairs per mouse). The four types of OFF-OFF overlaps were defined following the same principle. For quantification of the four types of overlaps, we calculated the average of all possible pairwise overlaps for each mouse ( Figure S5B-D ). (d) Lastly, we generated a union ON ensemble for each annotated behavior in each test, by pooling together all ON neurons tuned to the same behavior identified from all three testing stages. We also generated union OFF ensembles for each annotated behavior in each test using a similar strategy. We then compared the across-test ON-ON and OFF-OFF intra-behavior overlap (e.g., neurons identified as Direct-ON during Test1 and remained as Direct-ON during Test2), ON-ON and OFF-OFF inter-behavior overlap (e.g., neurons identified as Direct-ON during Test1 but Self-ON during Test2), and ON-OFF intra-behavior overlap among these union ensembles (e.g., neurons identified as Direct-ON during Test1 but Direct-OFF during Test2, See illustration in Figure 5A ). The across-test ON-ON intra-behavior overlap was defined as the overlap between ON ensembles tuned to the same behavior (3 pairs per mouse) identified in different tests; while the across-test ON-ON inter-behavior overlap was defined as the overlap between ON ensembles tuned to different behaviors (6 pairs per mouse) identified in different tests. The across-test OFF-OFF intra- and inter-behavior overlap followed the same principle. The across-test ON-OFF intra-behavior overlap was defined as the overlap between ON and OFF ensembles tuned to the same behavior (1 pair per mouse for each annotated behavior) identified in different tests. For quantification, we calculated the average of all possible pairwise overlaps for each mouse ( Figure 5A-C ). For the last comparison (i.e., social behavior tests in the absence and presence of PCP), we generated a union ON ensemble and a union OFF ensemble for each annotated behavior in each test as described above. We then compared the across-test ON-ON and OFF-OFF intra-behavior overlap, ON-ON and OFF-OFF inter-behavior overlap, and ON-OFF intra-behavior overlap among these union ensembles identified in the absence or presence of PCP (See illustration in Figure 7C ). All definitions and calculations were the same as described above in (d) of the across-test overlap comparisons. Behavioral variable decoding We performed binary decoding analyses to test if the mouse behavior variable (whether a behavior was occurring or not) could be reliably predicted using calcium activities of the identified ON and OFF neurons. Neural decoding is formulated as constructing a predictive model (classifier) f: C → b, where C is a set of calcium imaging derived features (calcium traces), and b is a behavior variable. For each mouse and a given behavior variable, five types of decoding tasks were conducted using five different types of calcium trace dataset: (1) calcium traces of all neurons; (2) calcium traces of the behavior tuned ON and OFF neurons; (3) calcium traces of the behavior tuned ON neurons and the randomly shuffled calcium traces of the behavior tuned OFF neurons; calcium traces of the behavior tuned OFF neurons and the randomly shuffled calcium traces of the behavior tuned ON neurons; (5) randomly shuffled calcium traces of the behavior tuned ON and OFF neurons. We used the widely used non-parametric decision tree classifier (Classification And Regression Trees, CART) ( Hastie et al., 2009 ). CART was implemented using the R package caret ( https://CRAN.R-project.org/package=caret ). Model parameters were tuned by the resampling-based method which only used the training dataset. Some behavioral variables were imbalanced. That is, the behavior only occurred on a small fraction of the total testing frames, thus one class outnumbers other classes by a substantial proportion. The down-sampling method ( Chawla, 2005 ) was used to correct the imbalance data problem. We used 10-fold cross-validation to assess classification performance. In the 10-fold cross-validation approach, the original dataset is randomly divided into 10 subsets. We then trained the model using 9 of the subsets (the training dataset) and tested the model using the 1 left-out subset (the test dataset). We repeated this process 10 times and calculate the average performance. Cross-validation provides a reliable assessment of model performance because the model never uses any cases from the test dataset. Model quality metric is the balanced accuracy, which is the arithmetic mean of sensitivity (true positive rate) and specificity (true negative rate). The balanced accuracy is in the range [0, 1]. The balanced accuracy = 1 represents perfect classification. The balanced accuracy = 0.5 represents the chance level. The decoding analysis was repeated for each stage of the social behavior test (test1), with a separate classifier trained for each decoding task. For the quantification of the decoding performance of each behavior variable ( Figure 3A ), we used the balanced accuracy of all three testing stages from 18 mice (54 cases).

QUANTIFICATION AND STATISTICAL ANALYSIS Statistics

All reported sample numbers represented biological replicates. All data were presented as mean ± SEM unless otherwise specified. In all the histograms in the figures, unless otherwise specified, dots indicated individual mice and bars represented group averages. In all the figures, **** represented p < 0.0001; ∗∗∗ represented p < 0.001; ** represented p < 0.01, * represented p < 0.05; ns represented not significant (p > 0.05). All statistical analyses were performed using Graphpad Prism (Graphpad) or MATLAB (Mathworks). Non-parametric tests were used in all of our statistical analysis, including Wilcoxon matched-pairs signed rank test, Friedman’s test and Dunn’s post-hoc test, Mann Whitney test, and Kolmogorov-Smirnov test. Pearson’s correlation was used for calculating correlation coefficient, and random permutations were used for determining chance level. All tests were two-sided, and statistical significance was defined as p < 0.05. Statistical details of all experiments can be found in Table S1 .

Sample size estimation

Our estimates of animal use for in vivo calcium imaging recording was based on past experiences and those presented in the literatures.

Data inclusion and exclusion

For in vivo calcium imaging experiments, we did not exclude any mice during our behavior data acquisition and analysis. Occasionally we could not reliably identify any behaviorally tuned neurons from a mouse in the absence or the presence of PCP, during a given testing stage for a given behavior, which would explain the reason that for neural ensemble related analyses, n numbers were less than 18, or not equal to the multiples of 18, in figure legends and in Table S1 .

DATA AND SOFTWARE AVAILABILITY

All datasets and custom MATLAB scripts will be available upon request.

EXPERIMENTAL MODEL AND SUBJECT DETAILS

All experiments were conducted in accordance with the guidelines of the Institutional Animal Care and Use Committee, the Intramural Research Program, National Institute on Drug Abuse, National Institutes of Health. All mice were wild type C57BL/6J at 3 – 4 months of age and 25 – 30 gram of weight. Mice were maintained in a regular light cycle (7:00am – 7:00pm) and provided with food and water ad libitum. All mice used in this study were males.

METHOD DETAILS Virus injection

We previously demonstrated that GFP protein driven by CamKII promoter in AAV1 virus is mostly expressed in excitatory neurons (i.e., only 1% GFP positive neurons were GABAergic neurons) once being injected into mouse prelimbic cortex ( Zhang et al., 2017 ). In this study, for calcium imaging experiments, we unilaterally injected AAV1.CamKII.GCaMP6f.WPRE.SV40 virus (University of Pennsylvania Vector Core, 500 nl with a titer of 2.76e13 GC/ml) into the dorsal region of the mouse prelimbic cortex, using the stereotactic coordinates (A/P: +1.9 mm, M/L: +0.3 mm, D/V: −1.7 mm). For ablation experiments, we bilaterally injected AAV1 viral mix (AAV1.CamKII0.4.Cre.SV40 and AAV1.mCherry.Flex.dtA, 500 nl in a ratio of 1:1, with a title of 4.496e13 GC/ml and 2.06e13 GC/ml, respectively) into the dorsal region of the mouse prelimbic cortex. For optogenetic experiments, we bilaterally injected AAV1 viral mix (AAV1.CamKII0.4.Cre.SV40 and AAV1.CAGGS.Flex.ChR2-tdTomato.WPRE.hGH, 500 nl in a ratio of 1:1, with a title of 4.496e13 GC/ml and 1.38e13 GC/ml, respectively), or AAV1 viral mix (AAV1.CamKII0.4.Cre.SV40 and AAV1.EF1a.DIO.eYFP.WPRE.hGH, 500 nl in a ratio of 1:1, with a title of 4.496e13 GC/ml and 2.9e13 GC/ml, respectively) as controls, into the dorsal region of the mouse prelimbic cortex. Mice were anaesthetized with 2% isoflurane in oxygen at a flow rate of 0.4 liter/min and mounted on a stereotactic frame (Model 962, David Kopf Instruments). A hole was drilled on the skull using a 0.5-mm diameter round burr with a high-speed rotary micro drill (19007–05, Fine Science Tools). A total of 500 nl of virus solution were injected at a rate of 25 nl/min using a micro pump and Micro4 controller (World Precision Instruments). The injection needle was kept in the parenchyma for 5 minutes after injection. The hole on the skull was then sealed with bone wax, and skin was sutured.

GRIN lens implantation

One or two weeks after AAV injection, a 1-mm diameter GRIN lens (GRINTECH GmBH) was directly implanted into the dorsal region of the mouse prelimbic cortex under anesthesia with ketamine/xylazine (ketamine:100 mg/kg and xylazine:15 mg/kg). In brief, a 1.1mm-diameter craniotomy was first made at the coordinates (A/P: +1.9 mm. M/L: +0.7mm). Before GRIN lens implantation, brain tissue above the prelimbic cortex was aspirated using a 30-gauge blunted needle, along a direction of 10° laterally-shifted angle into a depth of −1.8 mm. The needle was attached to a custom-constructed three-axis motorized robotic arm controlled by a MATLAB-based software ( https://github.com/liang-bo/AutoStereota ). The GRIN lens was then secured to the skull using dental cement (Metabond S380, Parkell). MiniScope Detailed information about the miniScope can be found in our previous study ( Barbera et al., 2016 ). Design files and part list can be found in the Github repository ( https://github.com/giovannibarbera/miniscope_v1.0 ). The miniScope consists of a blue LED (465 nm XLamp XP-E, Cree), a filter set (excitation filter, ET470/40, Chroma; dichroic mirror, FF495, Semrock; emission filter, EM525/50, Chroma), the relay optics (#83–605, #63–690, Edmund optics) and the CMOS sensor (MT9V022IA7ATM, Aptina). The miniScope housing was 3D printed (SLArmor Nickel-NanoTool, Proto Labs). The LED power was set to 0.1 – 0.3 mW at the miniScope focal plane. The image size was 400 × 400 pixels, acquired at 10-Hz frame rate. The pixel size of the acquired image was ~ 2.75 μm. Three to eight weeks after the GRIN lens implantation, the miniScope was mounted onto mouse head. After achieving the in-focus position for the entire field of view, the base of the miniScope was fixed on the skull using dental cement, and the main body of the miniScope was detached. Before each in vivo imaging experiment, the main body of the miniScope was connected and secured to the base on the mouse head through a side-locking screw. After data acquisition, the main body of the miniScope was detached from the base. We concurrently recorded calcium activities from hundreds of excitatory neurons in the dorsal prelimbic cortex, when mice freely performed the social behavior test ( Movie S1 ).

Social behavior test

We set up a social behavior test based on the previously established protocol known as “Crawley’s sociability and preference for social novelty test” ( Moy et al., 2004 ; Moy et al., 2007 ). Crawley’s sociability and preference for social novelty test was originally performed with a three-chamber apparatus. To facilitate in vivo calcium imaging, we modified the three-chamber apparatus by placing the subject mouse into one open square box (42 cm × 42 cm × 30 cm) with two small cylinder wire containers sitting at the opposite corners of a diagonal line (also see schematic illustration in Figure 1B ). The behavior test was composed of three consecutive 10-minute testing stages: “habituation”, “sociability”, and “social novelty”. During the “habituation” stage, the subject mouse freely explored the apparatus with two empty wire containers. During the “sociability” stage, an age- and weight-matched never-before-met same-sex (i.e., male) conspecific (stranger1) was introduced into one container while the other container remained empty. During the “social novelty” stage, a second age- and weight-matched never-before-met same-sex conspecific (stranger2) was introduced into the previous empty container during sociability stage while the first conspecific remained in the same location. “Sociability” was defined as propensity f or a subject mouse to spend time with a conspecific rather than with an empty container during the “sociability” stage. “Preference for social novelty” was defined as propensity for a subject mouse to spend time with a previously un-encountered conspecific rather than with a relatively familiar conspecific during the “social novelty” stage.

Novel object recognition test

A novel object recognition (NOR) assay was designed to examine mouse exploratory behavior to a non-social novel stimulus ( Leger et al., 2013 ). The NOR test contained two stages, a “familiarization” and a “testing” stage (also see schematic illustration in Figure S8B ). During the “familiarization” stage, mice freely explored two identical objects in a square box (42 cm × 42 cm × 30 cm) for 10 minutes. The two objects were placed at the same two corners of the diagonal line as that of the social behavior test. Following familiarization, mice were placed back to their home cages for 1 hour. During the “testing” stage, mice were re-introduced back into the same behavior apparatus – with one of the familiarized objects being replaced by a novel object. Mice explored these two objects for 10 minutes. Social behavior test with NOR control To further distinguish exploration of social salience versus exploration of novel stimuli, we performed the following social behavior test, incorporating two novel object controls (also see schematic illustration in Figure S8A ). The entire behavior test was composed of five consecutive 5-minute stages, including one “habituation” stage (H stage) and four “testing” stages (T1 to T4 stages). During the H stage, mice freely explored two empty wire containers (E1 and E2) in a square box (42 cm × 42 cm × 30 cm). E1 and E2 were placed at the opposite corners of a diagonal line. Following the H stage, a novel object (object1) was placed into the previously empty container E1 during the T1 stage. During the T2 stage, object1 was replaced by a novel conspecific (stranger1). During the T3 stage, a second novel object (object2) was placed into the previously empty container E2. During the T4 stage, object2 was replaced by a second novel conspecific (stranger2). In vivo calcium imaging A total of 18 mice were used for in vivo calcium imaging via miniScope during various behavior tests. These mice were unilaterally injected with AAV1 virus expressing calcium indicator GCamp6f mostly in excitatory neurons in the mPFC, and implanted with a 1-mm GRIN lens into the dorsal region of the prelimbic cortex. For the first batch of 9 mice, we performed in vivo calcium imaging with the NOR test first, followed by two social behavior tests. A week after the NOR test, the first social behavior test was performed in the absence of phencyclidine (PCP); while the second social behavior test was performed in the presence of PCP after another 5–14 days. PCP was intraperitoneally injected at a dose of 10 mg/kg, 30–50 minutes prior to the social behavior test. The objects in the NOR test and the strangers in the two social behavior tests were presented at the same locations (i.e., the same two corners of a diagonal line). With a second batch of 3 mice and a third batch of 6 mice, we performed social behavior tests in the absence of PCP multiple times (i.e., 4 times and 5 times, for the second and third batch of mice, respectively), a NOR test, and lastly a social behavior test in the presence of PCP. The timeline of the five social behavior tests (Test1 to Test5) was illustrated in Figure S4A . Specifically, one week after social behavior Test1, social behavior Test2 was performed with strangers presented at the same locations as in Test1; one week after social behavior Test2, social behavior Test3 was performed with strangers presented at different locations (i.e., two corners on the opposite diagonal line from the Test2); one to two weeks after social behavior Test3, social behavior Test4 and Test5 were performed on the same day (1 hour apart) with strangers presented at different locations. NOR test was performed between social behavior Test1 and Test4. After 3–14 days of social behavior Test4 and Test5, the social behavior test was performed again in the presence of PCP. PCP was intraperitoneally injected at a dose of 10 mg/kg, 30–50 minutes prior to the social behavior test. The second batch of 3 mice did not perform the social behavior Test3, while the rest of social behavior tests (i.e., Test1–2, Test4–5) were similar to those in the third batch of 6 mice. In all our social behavior tests, stranger mice were never-before-met conspecifics of the same gender (i.e., male) with matched age and weight. Ablation In the ablation experiment, we examined if focal ablation of excitatory neurons bilaterally in the mPFC would induce any abnormality in our social behavior test. After a pre-ablation social behavior test, we bilaterally injected AAV1 viral mix (AAV1.CamKII0.4.Cre and AAV1.mCherry.Flex.dtA) into the mPFC to express diphtheria toxin fragment A selectively in excitatory neurons in the mPFC. Five months after viral injection, a post-ablation social behavior test was performed with new pairs of strangers. Immunostaining After the post-ablation social behavior test, mPFC ablated mice and age-matched control mice were anesthetized with an overdose of ketamine (150 mg/kg) and xylazine (22.5 mg/kg), perfused with phosphate buffer solution (PBS) first and followed by a fixation buffer containing 4% paraformaldehyde (PFA) in PBS. Mice brains were dissected and post-fixed with 4% PFA in PBS overnight at 4°C. Fixed mice brains were sectioned via a Vibratome (LEICA VT1000S) and 30 μm coronal brain slices were collected. To evaluate the ablation efficiency, immunostaining of NeuN, a neuronal marker, was performed following a protocol for free-floating brain slice. In brief, floating brain slices were first washed in PBS (5 minutes for 3 times) and incubated in a blocking buffer (4% normal goat serum, 1% bovine serum albumin, 0.3% triton X-100 in PBS) with gentle rocking for two hours at room temperature. Brain slices were then incubated with a monoclonal mouse-anti-NeuN antibody (1:1000 diluted in blocking buffer, ab104224, Abcam) with gentle rocking overnight at 4°C. After washing in PBS (5 minutes for 3 times), brain slices were subsequently incubated with a secondary antibody (Alexa Fluor 488 conjugated donkey-anti-mouse IgG, 1:300 diluted in blocking buffer, Cat#715–546-150, Jackson ImmunoResearch) with gentle rocking for 1 hour at room temperature. After washing in PBS (5 minutes for 3 times), brain slices were mounted on slides for imaging. Fluorescent images were obtained through an Olympus VS120 scanner. For quantification, fluorescence intensity of NeuN staining was measured at the ablated region and an adjacent cortical region with a similar size. The ratio was calculated and compared between the ablated mice and controls.

Optogenetics

In the optogenetic experiment, we examined if transient disturbance of excitatory neural activity in the mPFC would induce any abnormalities in mice social exploration using our social behavior setup. We bilaterally injected AAV1 viral mix (AAV1.CamKII0.4.Cre and AAV1.CAGGS.Flex.ChR2-tdTomato) into the mPFC to express a light-gated cation-selective channel, channelrhodopsin-2 (ChR2), specifically in excitatory neurons in the mPFC. As a control, we bilaterally injected AAV1 viral mix (AAV1.CamKII.Cre and AAV1.EF1a.DIO.eYFP) into the mPFC. Two weeks after viral injection, customized optical fibers were bilaterally implanted into the dorsal region of the mouse prelimbic cortex (A/P: +1.9 mm, M/L: ±0.3 mm, D/V: −1.7 mm). Four months later, we first performed the social behavior test under a laser-ON condition. Specifically, following habituation, mice received a brief light stimulation from a 488-nm laser (10-ms long pulse at 20 Hz, laser ON for 20 seconds, ~ 7 mW at the fiber tip), right before the “sociability” and “social novelty” testing stages. One week later, we repeated the social behavior test with new pairs of strangers under the laser-OFF condition. Mice were connected with the optical fiber with the laser off and received a sham stimulation to mimic the auditory stimulation from the laser shutter (Vincent Associates). Specifically, following habituation, the shutter control was run for 20-seconds (10-ms long pulse, 20 Hz) right before the “sociability” and “social nove lty” testing stages.

Behavior annotation

The mouse social behavior test and NOR test were recorded using a top-view camera and a side-view camera (BFLY-PGE-12A2C-CS, Point Gray). The behavior images were synchronized with the calcium images by an external TTL signal from the miniScope controller. The behavior cameras had the same frame rate as the miniScope. Mouse behavioral activities were analyzed manually with frame-by-frame annotation based on an open-source MATLAB graphical user interface tool ( https://github.com/pdollar/toolbox ) ( Lin et al., 2011 ). Five distinct behavioral activities were annotated, including direct-exploration, proximity-exploration, irrelevant-exploration, ambulation and self-directed activities. Criterial for these annotated behaviors are listed below: (1) Direct-exploration labelled the scenario when the subject mouse proactively contacted the wire container through nose-poking, biting and paw-grabbing. (2) Proximity-exploration labelled the scenario when the head of the subject mouse was inside of the adjacent areas (within 5 cm of distance) of the wire container, but without any physical contact. (3) Irrelevant-exploration labelled the scenario when the subject mouse physically contacted walls of the apparatus through nose-poking or paw-touching. (4) Ambulation was defined as when the head of the subject mouse was physically outside of the adjacent areas of the wire container and moving about with measurable position (distance) changes. (5) Self-directed activities grouped several behavioral categories, including immobility, self-grooming and other self-directed subtle-movements. These behaviors were referred to as “self-directed activities” because mice presumably received less external information while performing these activities. Particularly, “immobility” was defined as when the subject mouse kept the same posture for more than 10 frames (1 second) at any place of the apparatus; “Self-grooming” was defined as previously reported ( Hong et al., 2014 ), when the subject mouse performed head-washing, nose- or face- grooming, head- or body- scratching, body-, leg-, paw- or genital- licking at any place of the apparatus; Other self-directed subtle-movement was defined as when the head of the subject mouse was physically outside of the adjacent areas of the wire container and performing subtle body posture changes without a measurable position (distance) change. Please also see Movie S2 for behavior annotation from a representative mouse. For each behavior, a binary behavior vector was constructed to represent the behavior periods (behavior is occurring: 1, behavior is not occurring: 0). The direct-exploration vector was further decomposed to direct-exploration of subject1 (H: E1; SB: S1; SN: S1) and direct-exploration of subject2 (H: E2; SB: E2; SN: S2). Similarly, the proximity-exploration vector was further decomposed to proximity-exploration of subject1 (H: E1; SB: S1; SN: S1) and proximity-exploration of subject2 (H: E2; SB: E2; SN: S2). Abbreviations H, SB and SN referred to habituation, sociability and social-novelty stage. E1 and E2 referred to the two empty containers, and S1 and S2 referred to stranger1 and stranger2, respectively.

Calcium image processing

Calcium images were processed and analyzed using custom scripts in MATLAB (Mathworks). Raw images were stabilized by determining the peak of the cross correlation with their neighboring images ( Guizar-Sicairos et al., 2008 ) to compensate for the brain position translations caused by mouse movements. To align the calcium images collected from different behavior tests, e.g. test 1 and test 2, we first calculated the maximum-projection image of all frames collected for each test. The maximum-projection image of test 1 was used as the reference image. We then used the MATLAB function “imregdemons ” (Image processing toolbox) to estimate the displacement field between any two maximum-projection images. We generated the registered image by applying the estimated displacement field to the maximum-projection image of test 2. Subsequently, we visually inspected the overlaid image containing both the reference and registered images with complementary colors to ensure that the estimated displacement field was appropriate ( Figure S3D ). We then applied the estimated displacement field to all the frames collected during the social behavior test 2. Cell identification was performed using the algorithm previously described ( Barbera et al., 2016 ). For each mouse, a cumulative cell map including all active neurons identified from all behavior tests was generated and manually verified. Calcium trace of each active neuron was extracted using detailed procedures below: (1) For a given neuron, a region of interest (ROI) was assigned with a circular region corresponding to the size of soma (diameter = 15 μm) and the centroid indicated the detected cell position. The minimum fluorescence of image-stack was used as the baseline fluorescence of the background (referred to as F b ). The fluorescence of ROI (F ROI ) was calculated as the averaged fluorescence after background subtraction from the raw image fluorescence (F raw ), i.e., F ROI = ROI , brackets indicate the average over ROI pixels. (2) A commonly used procedure ( Barbera et al., 2016 ; Chen et al., 2013 ; Kerlin et al., 2010 ; Pinto and Dan, 2015 ) was then applied to correct the potential contamination from out-of-focus neurons and neuropil fluorescence, i.e., F sig = F ROI − F con , where F sig was the true fluorescence signal, F con was the contamination fluorescence. F con was calculated over background subtracted images, using the averaged fluorescence from an annular zone (d1 = 20 μm, d2 = 30 µm) surrounding the soma ROI but avoiding other ROI regions. The slowly variable baseline due to background/neuropil activity was also removed in this procedure. (3) Each calcium trace ΔF/F was calculated as F sig /F 0 , where F 0 was the baseline fluorescence over the neuron ROI. (4) The baseline noise level for each calcium trace was estimated from the power density spectrum at a high-frequency band ( Pnevmatikakis et al., 2016 ). For each testing stage, active neurons were defined only if the maximum value of the calcium trace was 3 times greater than its baseline noise level. The calcium traces were smoothed with a 500-ms sliding window. For each social behavior test, a neuron needed to be active at least once during the three testing stages to be identified as active neuron. Neural activities in three testing stages We calculated the neural calcium activity changes in the presence of restrained social targets ( Figure 1F ). Data of social behavior test 1 (n = 18 mice) were used in this calculation. For each neuron, calcium activities at the three social behavior testing stages were compared first by the Kruskal-Wallis test. A neuron was defined as “unchanged” if the p-value of the Kruskal-Wallis test was greater than 0.05. If the p value of the Kruskal-Wallis test was less than 0.05, multiple comparisons with Tukey-Kramer correction were performed subsequently. A neuron was defined as “SB&SN increased” if neural activities at both the “sociability” and “social novelty” stages were significantly higher than that of the “habituation” stage (p < 0.05, two-side). A neuron was defined as “SB increased only”, if the neural activity at the “sociability” stage was significantly higher than that of the “habituation” stage (p < 0.05, two-side), while the neural activity at the “social novelty” stage was not significantly different from that of the “habituation stage” (p > 0.05, two-side). A neuron was defined as “SN increased only”, if the neural activity at the “social novelty” stage was significantly higher than that of the “habituation” stage (p < 0.05, two-side), while the neural activity at the “sociability” stage was not significantly different from that of the “habituation” stage (p > 0.05, two-side). Neurons defined as “SB&SN decreased”, “SB decreased only” or “SN decreased only” followed the same principle. PCP effects on population calcium activities To evaluate the PCP effect on calcium activities at the neural population levels, in vivo calcium imaging during a 10-minute habituation test was performed on a group of 5 mice injected with GCaMP6f virus and implanted with a GRIN lens, with the same apparatus set up as the habituation stage of the social behavior test as described above. These mice were subsequently intraperitoneally injected with 10 mg/kg PCP, and 30–50 minutes later, in vivo calcium imaging was performed again during a second 10-minute habituation test. As control, in vivo calcium imaging during the habituation tests with intraperitoneal saline injection was performed on a different group of three mice injected with GCaMP6f and implanted with a GRIN lens. Calcium images were processed using the same methods described above. Active neurons were identified and calcium traces were extracted. For each neuron, the averaged activity of the 10-minute recording before and after PCP or saline injection was calculated and denoted as C pre and C post , respectively. Averaged calcium activity of all the neurons from all mice were pooled together for comparisons between C pre and C post of PCP and saline injection, respectively. The Kolmogorov–Smirnov test was used to compare distributions of the averaged activity before and after PCP ( Figure S6D ) or saline injection ( Figure S6E ). We also calculated the activity change index by (C post − C pre )/ (C post + C pre ), for before and after PCP or saline injection, respectively ( Figure S6D-E ). The value of activity change index was bound between −1 and 1. Positive value indicated increased activity. Negative value indicated decreased activity.

Identification of behaviorally tuned neurons

For each testing stage, we identified behaviorally tuned neurons corresponding to each of the three annotated behaviors (i.e. direction-exploration, self-directed activities and irrelevant-exploration), following the same analytical steps as described below (also see Figure S2A-C ): (1) For any given neuron n, we first calculated the similarity between calcium trace (ΔF/F) C n and behavior vector B. Similarity was defined as normalized inner product of two vectors, 2B·C n /(|B| 2 +|C n | 2 ) ( Carrillo-Reid et al., 2015 ; Hamm et al., 2017 ). The value of similarity was bound between 0 and 1. A value of 1 meant these two vectors were identical; while a value of 0 meant they were completely different. (2) We then randomly shuffled the behavior epochs and calculated the similarity between the shuffled behavior vector and the calcium trace C n for a given neuron. We repeated this permutation process 5,000 times to generate a similarity distribution representing the similarity distribution histogram predicted by chance for a given neuron. (3) A neuron was defined as an “ON neuron” only if its actual similarity was greater than the 99.17% quantile of chance similarity distribution histogram generated by random shuffling. Consequently, ON neurons were those exhibiting significantly higher activities than that predicted by chance during the annotated behavior epochs. Conversely, a neuron was defined as an “OFF neuron” only if its actual similarity was lower than the 0.83% quantile of chance similarity distribution histogram generated by random shuffling. Therefore, OFF neurons were those exhibiting significantly lower activity than that predicted by chance during the annotated behavior epochs. Neurons with similarity falling between the 0.83% and 99.17% quantiles of chance similarity distribution histogram generated by random shuffling, were defined as “Other neurons” (also see Figure S2B-C ). For each mouse, we applied the calculations described above to all active neurons to identify the ON and OFF neurons correlated with each of the three annotated behaviors during different testing stages. We defined each group of ON neurons (or OFF neurons) as one “neural ensemble”. Consequently, we identified a total of 9 pairs of mPFC neural ensembles for each mouse, including Direct-ON H and Direct-OFF H , Self-ON H and Self-OFF H , Irrelevant-ON H and Irrelevant-OFF H during the H stage; Direct-ON SB and Direct-OFF SB , Self-ON SB and Self-OFF SB , Irrelevant-ON SB and Irrelevant-OFF SB during the SB stage; Direct-ON SN and Direct-OFF SN , Self-ON SN and Self-OFF SN , Irrelevant-ON SN and Irrelevant-OFF SN during the SN stage (see Figure 2 and Figure S2 ).

Neural ensemble correlation analysis

For each identified ON and OFF neuron, z-scored calcium activity was calculated by (C n – μ )/σ, where μ and σ were the mean and standard deviation of C n across H, SB and SN stages. For each ON and OFF neural ensemble, the averaged z-scored calcium activity was calculated by pooling the z-scored calcium activities from all identified ON and OFF neurons within the ensemble ( Figure 2A and Figure S2F ). We also calculated the averaged z-scored calcium activity across all behavior epochs (with an interval of −4 s to +4 s) for each ON and OFF neuron, and sorted all identified ON neurons based on the time of their maximal peak activities, and OFF neurons based on the time of their minimal activities ( Figure 2D ). To compare calcium dynamics around the behavior onset and the behavior ending, for each annotated behavior, we calculated the averaged z-scored calcium activity by pooling all ON or OFF neural ensembles tuned to the same behavior but identified at different testing stages ( Figure S2D ). We then compared the averaged calcium activity from all 18 mice, pre- (−2 s to 0 s) and post- (0 s to +2 s) behavior onset and ending for each annotated behavior ( Figure S2E ). To study the correlation between calcium activity of neural ensembles and specific behaviors, we calculated both (1) the Pearson correlation coefficient between the behavior vector and the averaged calcium activity of all neurons from a given neural ensemble (referred to as “group-correlation coefficient” in Figure 2E and Figure S3A ), and (2) the average of the individual Pearson correlation coefficients between the behavior vector and each individual neural activity (referred to as “individual-correlation coefficient” in Figure 2E and Figure S3A ).

Neural ensemble consistency analysis

We evaluated the reliability of ON and OFF neuron’s responses across all behavioral epochs ( Figure 2C ). For an ON neuron, we defined an “engaged” behavior epoch when the ON neuron exhibited activity 3 times higher than the baseline noise; we defined a “not engaged” behavior epoch when the ON neuron did not exhibit activity 3 times higher than the baseline noise. By contrast, for an OFF neuron, we defined an “engaged” behavior epoch when the OFF neuron did not exhibit activity 3 times higher than the baseline noise; we defined a “not engaged” behavior epoch when the OFF neuron exhibited activity 3 times higher than the baseline noise. We evaluated the consistency of ON and OFF neuron’s responses through calculating the percentage of “engaged” behavior epochs across all behavior epochs within each testing stage. We also calculated for any given behavior epochs, the percentage of ON or OFF neurons being “engaged” within each ON and OFF ensemble. Neural ensemble functional connectivity Functional connectivity within and between neural ensembles was evaluated by calculating the averaged pairwise Pearson correlation coefficient of the calcium traces within and between behaviorally tuned neural ensembles ( Figure 3B ). The connectivity of an ON ensemble tuned to a specific behavior was evaluated by the averaged pairwise Pearson correlation coefficient of the calcium traces within the ON ensemble. The connectivity of an OFF ensemble tuned to a specific behavior was evaluated by the averaged pairwise Pearson correlation coefficient of the calcium traces within the OFF ensemble. The connectivity between ON and OFF ensembles tuned to the same specific behavior was evaluated by the averaged pairwise Pearson correlation coefficient of the calcium traces between ON and OFF ensembles. “Other” neurons were defined as neurons not belonging to ON or OFF ensembles tuned to any of the three annotated behaviors. The connectivity of “other” neurons was evaluated by the averaged pairwise Pearson correlation coefficient of the calcium traces within these neurons. We also evaluated the connectivity between ON (or OFF) neurons and “other” neurons by the averaged pairwise Pearson correlation coefficient of the calcium traces between ON neurons (or OFF neurons) and “other” neurons. Direct-exploration tuned ensemble activity Calcium activity of Direct-ON and Direct-OFF ensembles during direct-exploration of the two targets at each testing stage was calculated ( Figure 8 ). The calcium trace of each neuron from a given neural ensemble was first normalized, using C norm = (C - C min )/(C max -C min ), where C max and C min were the maximum and minimum values of C, respectively. Subsequently, for each Direct-ON (or Direct-OFF) neuron, the average of the normalized calcium activity during direct-exploration on each of the two targets were calculated, respectively. Lastly, the averaged activity of all Direct-ON (or Direct-OFF) neurons for each mouse at each testing stage was calculated for quantification.

Neural ensemble overlap analysis

For a given pair of ensembles, A and B, the degree of observed overlap was calculated by the Sørensen–Dice index: 2N A∩B /(N A +N B ), with N A indicating the neuron number of ensemble A and N B indicating the neuron number of ensemble B, and N A∩B indicating the neuron number of overlaps between ensemble A and B. The overlap degree was 0 if both A and B were empty. The observed overlap was then compared to what would be predicted by chance. The probability of one neuron to be randomly assigned into ensemble A was N A /N, where N is the total imaged neuron number. Similarly, the probability of one neuron to be randomly assigned into ensemble B was N B /N. If both assignments were random and independent, the probability of one neuron to be simultaneously assigned into both ensemble A and B was N A N B /N 2 . Thus, the degree of overlap predicted by chance was calculated by 2N A N B /(N(N A +N B )). (A) Overlap within the same testing stage ( Figure 4C-D ): To evaluate the dynamic feature of neural ensembles, we split each 10-minute recording (one testing stage) into two 5-minute halves and re-identified ON and OFF neurons tuned to direct-exploration (Direct-ON and Direct-OFF neurons). We calculated the overlap between Direct-ON neurons of the two halves, and the overlap between Direct-OFF neurons of the two halves ( Figure 4C ). Next, to evaluate the reliability of our neuron identification process, we generated a union ensemble (Half1 ∪ Half2) by pooling together all the identified Direct-ON (or Direct-OFF) neurons from the two 5-minute halves and calculated the percentage of Direct-ON (or Direct-OFF) neurons originally identified based on the full-length 10-minute recording belonging to the union ensembles. (B) Overlap across different stages within one social behavior test ( Figure 4A-B , Figure S3C ): To clarify the across-stage overlaps between ON and OFF ensembles identified at different stages of a social behavior test, we defined across-stage overlap between ON-ON and OFF-OFF neural ensembles tuned to the same behavior as “ON-ON and OFF-OFF intra-behavior” overlap (e.g., neurons identified as Direct-ON at the “habituation” stage and remained as Direct-ON at the “sociability” or “social novelty” stage); across-stage overlap between ON-ON and OFF-OFF neural ensembles tuned to different behaviors as “ON-ON and OFF-OFF inter-behavior” overlap (e.g., neurons identified as Direct-ON at the “habituation” stage but Self-ON at the “sociability” stage); and across-stage overlap between ON-OFF neural ensembles tuned to the same behavior as “ON-OFF intra-behavior overlap” (e.g., neurons identified as Direct-ON at the “habituation” stage but Direct-OFF at the “soci ability” stage). The ON-ON intra- and inter-behavior overlaps between neural ensembles were illustrated in Figure 4A left panel. The OFF-OFF intra- and inter-behavior overlaps followed the same principle. The ON-OFF intra-behavior overlap was illustrated in Figure 4B left panel. For the quantification of ON-ON and OFF-OFF intra- and inter-behavior overlaps, we calculated the average of all the possible pairwise overlaps for each mouse (9 pairs per mouse for ON-ON or OFF-OFF intra-behavior overlap, and 18 pairs per mouse for ON-ON or OFF-OFF inter-behavior overlap, as illustrated in Figure 4A left panel). We also quantified ON-ON and OFF-OFF intra-behavior overlaps for each annotated behavior, by calculating the average of 3 possible pairwise overlaps for each mouse ( Figure S3C ). For quantification of ON-OFF intra-behavior overlap for each annotated behavior, we calculated the average of all the possible pairwise overlaps for each mouse (6 pairs per mouse for each annotated behavior, as illustrated in Figure 4B left panel). The corresponding overlap to be predicted by chance was calculated as described above. (C) Overlap between two different social behavior tests ( Figure 5 , Figure S4 and S5 ): We performed four different types of overlap comparisons for ON and OFF ensembles from all the social behavior tests (See In vivo calcium imaging section and Figure S4A ): (1) Social behavior Test1 versus Test2, performed on different days with social targets presented at the same locations; (2) Social behavior Test2 versus Test3, performed on different days with social targets presented at the different locations; (3) Social behavior Test4 versus Test5, performed on the same days with social targets presented at different locations; (4) Social behavior test in the absence of PCP versus social behavior test in the presence of PCP, performed on different days with social targets presented at the same locations. For the first three comparisons (i.e., Test1 versus Test2, Test2 versus Test3, Test4 versus Test5), (a) we first calculated for each annotated behavior, the ON-ON and OFF-OFF intra-behavior overlap across the same testing stages of the two tests (e.g., Direct-ON at the “sociability” stage of Test1 versus Direct-ON at the “sociability” stage of Test2, as illustrated in Figure S4B ). For quantification, there was only one pairwise comparison per mouse for each same-stage overlap of each annotated behavior ( Figure S4C-E ). (b) We then compared for each annotated behavior, the ON-ON and OFF-OFF intra-behavior overlap across same testing stages versus across different testing stages of the two tests (e.g., Direct-ON at the “habituation” stage of Test1 versus Direct-ON at the “sociability” stage of Test2, as illustrated in Figure S4F ). For quantification, we pooled all possible pairwise comparisons of same-stage overlaps (3 pairs per mouse) or different-stage overlaps (6 pairs per mice) for each annotated behavior. (c) Next, we grouped all three annotated behaviors to compare ON-ON and OFF-OFF intra-behavior overlap versus ON-ON and OFF-OFF inter-behavior overlap, both across same testing stages and across different testing stages of the two tests (as illustrated in Figure S5A ). Therefore, there were four types of ON-ON (or OFF-OFF) ensemble overlaps in this comparison: the overlap of “intra-behavior, same-stage”, the overlap of “intra-behavior, different-stage”, the overlap of “inter-behavior, same-stage”, and the overlap of “inter-behavior, different-stage”. Specifically, the ON-ON overlap of “intra-behavior, same-stage” was defined as the overlap of the two ON ensembles tuned to the same behavior identified in the same stages of the two tests (9 pairs per mouse). The ON-ON overlap of “intra-behavior, different-stage” was defined as the overlap of the two ON ensembles tuned to the same behavior identified in different stages of the two tests (18 pairs per mouse). The ON-ON overlap of “inter-behavior, same-stage” was defined as the overlaps of the two ON ensembles tuned to different behaviors identified in the same stages of the two tests (18 pairs per mouse). The ON-ON overlap of “inter-behavior, different-stage” was defined as the overlaps of the two ON ensembles tuned to different behaviors identified in different stages of the two tests (36 pairs per mouse). The four types of OFF-OFF overlaps were defined following the same principle. For quantification of the four types of overlaps, we calculated the average of all possible pairwise overlaps for each mouse ( Figure S5B-D ). (d) Lastly, we generated a union ON ensemble for each annotated behavior in each test, by pooling together all ON neurons tuned to the same behavior identified from all three testing stages. We also generated union OFF ensembles for each annotated behavior in each test using a similar strategy. We then compared the across-test ON-ON and OFF-OFF intra-behavior overlap (e.g., neurons identified as Direct-ON during Test1 and remained as Direct-ON during Test2), ON-ON and OFF-OFF inter-behavior overlap (e.g., neurons identified as Direct-ON during Test1 but Self-ON during Test2), and ON-OFF intra-behavior overlap among these union ensembles (e.g., neurons identified as Direct-ON during Test1 but Direct-OFF during Test2, See illustration in Figure 5A ). The across-test ON-ON intra-behavior overlap was defined as the overlap between ON ensembles tuned to the same behavior (3 pairs per mouse) identified in different tests; while the across-test ON-ON inter-behavior overlap was defined as the overlap between ON ensembles tuned to different behaviors (6 pairs per mouse) identified in different tests. The across-test OFF-OFF intra- and inter-behavior overlap followed the same principle. The across-test ON-OFF intra-behavior overlap was defined as the overlap between ON and OFF ensembles tuned to the same behavior (1 pair per mouse for each annotated behavior) identified in different tests. For quantification, we calculated the average of all possible pairwise overlaps for each mouse ( Figure 5A-C ). For the last comparison (i.e., social behavior tests in the absence and presence of PCP), we generated a union ON ensemble and a union OFF ensemble for each annotated behavior in each test as described above. We then compared the across-test ON-ON and OFF-OFF intra-behavior overlap, ON-ON and OFF-OFF inter-behavior overlap, and ON-OFF intra-behavior overlap among these union ensembles identified in the absence or presence of PCP (See illustration in Figure 7C ). All definitions and calculations were the same as described above in (d) of the across-test overlap comparisons. Behavioral variable decoding We performed binary decoding analyses to test if the mouse behavior variable (whether a behavior was occurring or not) could be reliably predicted using calcium activities of the identified ON and OFF neurons. Neural decoding is formulated as constructing a predictive model (classifier) f: C → b, where C is a set of calcium imaging derived features (calcium traces), and b is a behavior variable. For each mouse and a given behavior variable, five types of decoding tasks were conducted using five different types of calcium trace dataset: (1) calcium traces of all neurons; (2) calcium traces of the behavior tuned ON and OFF neurons; (3) calcium traces of the behavior tuned ON neurons and the randomly shuffled calcium traces of the behavior tuned OFF neurons; calcium traces of the behavior tuned OFF neurons and the randomly shuffled calcium traces of the behavior tuned ON neurons; (5) randomly shuffled calcium traces of the behavior tuned ON and OFF neurons. We used the widely used non-parametric decision tree classifier (Classification And Regression Trees, CART) ( Hastie et al., 2009 ). CART was implemented using the R package caret ( https://CRAN.R-project.org/package=caret ). Model parameters were tuned by the resampling-based method which only used the training dataset. Some behavioral variables were imbalanced. That is, the behavior only occurred on a small fraction of the total testing frames, thus one class outnumbers other classes by a substantial proportion. The down-sampling method ( Chawla, 2005 ) was used to correct the imbalance data problem. We used 10-fold cross-validation to assess classification performance. In the 10-fold cross-validation approach, the original dataset is randomly divided into 10 subsets. We then trained the model using 9 of the subsets (the training dataset) and tested the model using the 1 left-out subset (the test dataset). We repeated this process 10 times and calculate the average performance. Cross-validation provides a reliable assessment of model performance because the model never uses any cases from the test dataset. Model quality metric is the balanced accuracy, which is the arithmetic mean of sensitivity (true positive rate) and specificity (true negative rate). The balanced accuracy is in the range [0, 1]. The balanced accuracy = 1 represents perfect classification. The balanced accuracy = 0.5 represents the chance level. The decoding analysis was repeated for each stage of the social behavior test (test1), with a separate classifier trained for each decoding task. For the quantification of the decoding performance of each behavior variable ( Figure 3A ), we used the balanced accuracy of all three testing stages from 18 mice (54 cases).

Supplementary Material 1 2 3 4