GABAergic Restriction of Network Dynamics Regulates Interneuron Survival in the Developing Cortex.

LEAD CONTACT AND MATERIALS AVAILABILITY

Further information and requests for resources and reagents should be directed to and will be fulfilled by the Lead Contact, Natalia De Marco Garcia ( nad2018@med.cornell.edu ). This study did not generate new unique reagents.

EXPERIMENTAL MODEL AND SUBJECT DETAILS

All animal care and procedures were performed according to the Weill Cornell Medicine Research Animal Resource Center guidelines. Animals were housed in a controlled environment on a 12-hour light/dark cycle with food and water ad libitum . All animal lines have been described previously: Lhx6-Cre (JAX 026555), VGAT fl/fl (JAX 012897), Ai96 (RCL-GCaMP6s) (JAX 024106), 5HT3aR-EGFP (MMRRC 000273-UNC), Ai9 (RCL-tdT) (JAX 007909), 5Ht3aR-Cre (MGI:5435492, a gift from N. Heintz, Rockefeller University), Emx1 Cre (JAX 005628), SST Flp (JAX 028579), GABA A γ2 (Gabrg2 tm2lusc ) (JAX 016830). Mice of both sexes were used for analysis. Due to the young age of the mice, the influence of sex was not analyzed. For timed pregnancies, noon on the day of the vaginal plug was counted as E0.5. Since we did not detect significant statistical differences between Lhx6.VGAT +/+ and Lhx6.VGAT fl/+ mice in all of the analyses, we combined these genotypes in the control group. Lhx6.VGAT fl/fl mice are viable until P16 and do not show spontaneous seizures (0/300 mice) in the first two postnatal weeks. This agrees with other reports indicating that mouse mutants with developmental defects in interneuron number or VGAT deletion do not develop seizures until adulthood ( Cobos et al. 2005 ; Lin et al. 2018 ).

METHOD DETAILS Immunohistochemistry

Pups were anesthetized by hypothermia (< or equal to P6) or euthasol (>P7) and perfused via the left ventricle with 10 mL ice-cold PBS followed by 10 mL ice-cold 4% PFA. Brains were post-fixed in 4% PFA for 45 minutes on ice and equilibrated in 30% sucrose for > 16 hours at 4 C before being mounted in tissue-tek (Sakura). 20 μm sections were used for all counts except for ClCsp3 counts, which used 14 μm sections. After 3× 5 minute PBS washes, sections were blocked with 5% donkey serum, 0.03% Triton in PBS at room temperature for 1 hour. Sections were incubated with appropriate primary antibodies overnight (>16 hours) at 4 C: 1:1000 Rb-Sst (Abcam), 1:500 Rt-Sst (Millipore); 1:1000 Rb-Pv (Abcam), 1:1000 Gt-GFP (Rockland), 1:1000 Rt-GFP (nacalaitesque), 1:1000 Ch-RFP (Rockland), 1:250 Rb-ClCsp3 (Cell Signaling Technologies), 1:1000 Streptavidin Pacific Blue (Thermo Fisher). After more than 3× 5 minute PBS washes, sections were incubated with appropriate secondary antibodies at room temperature for 1 hour and stained for DAPI before mounting. Slides were mounted with Vectashield (hard set). Note that we found both somatostatin antibodies work most optimally on tissue post-fixed 20pA) and temporal characteristics (>100ms maximal width). We confirmed that GDP events were polysynaptic sIPSCs by recording at 0mV and abolishing them with GABA-A antagonist SR95531. For representative continuous cGDP traces where a series resistance test was applied every 10 seconds at the beginning of the episode, the series resistance test was blanked from 0 to 0.5 s. In Kir2.1 experiments, 1000 ms hyperpolarizing voltage steps (10pA) between −120mV and 0mV were applied and the maximum inward currents were measured. 5μM Barium Cl-(Sigma) was washed on in bath to abolish Kir2.1 currents. Following the recording session, pipettes were retracted and slices were post-fixed in 4% PFA and treated with fluorescence conjugated streptavidin labeling and immunohistochemistry. To post hoc identify MGE interneurons, we used their expression of tdTomato together with staining for Sst and/or Reelin to identify Sst cells. We were unable to use Pv as a marker as it is not expressed until ~P12. In addition, we used standard methods for classifications combined with unbiased dimensionality reduction and unsupervised hierarchical clustering to confirm the post-hoc immunohistochemical identity ( Anastasiades et al. 2016 ; Ferrante et al. 2017 ). Clustering generated three primary clusters FS, IB, and a third cluster of non-fast spiking adapting (NFS) neurons ( Figure S1E ) ( Ferrante et al. 2017 ; Krimer et al. 2005 ). Since immature FS interneurons may fail to exhibit narrow AP widths characteristic of the Pv class during the first postnatal week of development ( Goldberg et al. 2011 ), the NFS and FS clusters may contain misclassified interneurons. The difficulty in discriminating between immature FS-like(pPv) and NFS (Sst) interneurons at this age is exemplified by a rare case (1/13) in which an Sst interneuron was classified as FS in the clustering algorithm ( Figure S2E ). Since 92% of Sst neurons clustered in IB and NFS, we used those cluster groups as surrogates for putative Sst and pPv interneurons. For interneurons that did not successfully sort into one of the three FS, IB, or NFS clusters via Wards hierarchical clustering, their intrinsic properties were excluded from Table S2 . Optogenetic experiments were conducted using a Mightex LED arm attachment. ChR2-expressing cells were depolarized using a 5 ms pulse of 480 nm light. Light-evoked responses were confirmed to generate APs in a subset of experiments using K+ based pipette solution in the presence of 1μM tetrodotoxin (TTX) and 1mM 4-aminopyridine (4-AP). Cells were post hoc recovered to confirm interneuron identity.

Cranial Window Surgery and Image Acquisition

Cranial window surgery and 2-photon imaging were performed as previously described ( Che et al. 2018 ). Briefly, P6–7 pup was anesthetized on ice, a head plate was attached, and a piece of skull was removed without disturbing the pia. A glass coverslip was placed over the cranial window and secured. Following recovery from cranial window surgery (1-hour minimum), pups were head-fixed and imaged by FluoView FVMPE-RS multiphoton imaging system via a 25X (1.05 NA) water immersion lens (Olympus). Pups were un-anaesthetized and remained in resting state throughout the imaging session. LII/III was imaged at 150–200 μm below pia; LV was imaged at 350 m and below. For AAV.EF1a.Kir2.1.zsGreen and AAV.CAG.flex.JRGECOa1 experiments, we imaged first at two-photon excitation of 900 nm and detected GFP emission with a 495/540 nm filter. Once we identified the area infected with Kir2.1GFP, we shifted the two-photon excitation to 1040 nm to excite JRGECOa1 and detected the emission with a 575/645 filter (red channel). For dual GCaMP6s/TdTomato imaging, we concurrently imaged at two-photon excitation of 900 nm and separated the emission spectra with two filters (green signal a 495/540 nm band-pass filter and red signal a 575/645 filter). We shifted the 2-photon excitation wavelength from the GCaMP6s peak of 920 nm to 900 to reduce the excitation of tdTomato.

Calcium Imaging Analysis

Analysis was performed as previously described ( Che et al. 2018 ). For videos with virally expressed GCaMP6s and jRGECO1a, neuropil correction was performed on all videos ( Che et al. 2018 ). For chronic imaging, Lhx6.VGAT.Ai9 mice injected with AAV.ef1a.DIO.GCaMP6s with cranial window implants were imaged from P7 through P8, identifying the same region of the cortex on each day. Regions were aligned using vascular landmarks, both on the surface and visible deeper by nature of the Lhx6 promoter expression in endothelium. After a FOV was selected and imaged, a Z stack of 50 μm above and 50 μm below the FOV at 4 m interval was taken. In the tdTomato channel, each red cell in the P7 video was either located or deemed dead based on the matching P8 video. In addition, if a cell could not be found in the P8 video, it was searched in the z stack, as it was common to see cells moving out of the imaging plane due to the still rapid growth of the brain at this age. 6 to 23 dead cells were identified in FOVs containing 140–252 cells. Cell masks were made on the P7 video using methods previously described, and calcium dynamics were analyzed for each living versus dead cell ( Che et al. 2018 ). Assembly analysis (edited from Methods in the companion manuscript Modol et al) GCE detection: To detect population events we randomly circularly reshuffled each spike vector, hence maintaining within contour spike dependencies while randomizing the population and correcting for differences in baseline firing for different animals and ages. One thousand surrogate distributions were created. The spikes of each frame for these distributions were computed and the 99th percentile of the resulting “sum of spikes” vector was used as a statistical threshold. Peaks above the threshold that were at least separated by 7 frames were considered as synchronous calcium events. A population vector is a binary representation of the spiking activity of the event (peak +/− 3 frames), where a 1 is assigned when a contour has a spike within the interval and 0 otherwise. Spike distributions: To detect the calcium onsets of contours and quantify their individual contribution to synchronous network events, the averaged onset time of the first spike within an event was considered for all contours as well as the ratio of events that recruited a given. We then computed the probability density function of onsets and scaled them to the maximum participation rate per group. Statistical analyses were then applied to measure differences in the onset dynamics (see below). Identification of cell assemblies in GCE: Analyses were performed as described in ( Malvache et al., 2016 ). Cell assemblies were identified using a clustering algorithm based on GCE similarity for cell participation followed by a statistical test for cell participation in each GCE cluster. The GCE similarity metric was the squared Euclidian distance between columns of the normalized covariance matrix. This similarity metric allowed for a more efficient clustering. Unsupervised clustering of GCE was obtained by running the k-means algorithm on this metric with cluster numbers ranging from 2 to 19. Hundred iterations of k-means were run for each cluster number and the iteration that resulted in the best-averaged Silhouette value was kept. The silhouette value was computed as described previously ( Malvache et al., 2016 ). A random distribution of average silhouette values for each cluster was calculated by reshuffling cell participation across different GCE and applying the same algorithm. Clusters with average silhouette values exceeding the 95th percentile of the random case were considered as statistically significant. Each cluster was then associated to a cell assembly that comprised those cells that significantly participated to the GCE events within that particular cluster. Cell participation to a given cluster was considered statistically significant if the fraction of synchronous events in that cluster that activated the cell exceeded the 95th percentile of reshuffled data. If a cell was significantly active in more than one GCE cluster, it was associated to the one in which it participated the most (percentage wise). The overlap between assemblies was quantified by calculating the silhouette value of each cell (with the normalized hamming distance between each cell pair as a dissimilarity metric). A cell was significantly involved in a single assembly if its silhouette value was higher than expected by chance (95th percentile after reshuffling). GCEs were finally sorted with respect to their projection onto cell assemblies. A GCE was activating a given neuronal assembly if the number of cells recruited in that assembly was higher than expected by chance (95th percentile after reshuffling). Topological distribution of cell assemblies: For each cell assembly, the spatial silhouette value was computed using the pairwise anatomical distance between cells as a dissimilarity metric. To assess the statistical significance of spatial clustering for each cell assembly, its silhouette value was compared to the ones obtained after cell reshuffling (1000 iterations). Spatial clustering was considered significant if the silhouette value of the empirical assembly exceeded a certain percentile of the reshuffled silhouette value distribution. For each session, the percentile was set at (1 – 0.05/(number of clusters) to account for the inter-dependence of the comparisons (Bonferroni correction). T-distributed Stochastic neighbor embedding (t-SNE) Due to commonly associated shortcomings of the k-means algorithm, we used the t-distributed stochastic neighbor embedding method ( Van Der Maaten and Hinton, 2008 ), to visualize the data and also verify the concurrence with the k-means analysis. The t-SNE technique converts correlations between contour vectors to joint probabilities and tries to minimize the Kullback-Leibler divergence (KL divergence) between the joint probabilities of the low-dimensional embedding and the high-dimensional data. We first reduced the dimensionality of the data by considering the first 20 Principal Components of contours participation in GCEs. These components were then non-linearly mapped onto a two-dimensional space in a way that preserved both local and global structure of the data. The perplexity parameter (the number of nearest neighbors) was chosen in an iterative procedure that minimized the KL divergence. These values ranged between 15 and 25. We then color-coded each component according to the previous k-means analysis.

QUANTIFICATION AND STATISTICAL ANALYSIS

Statistics and graphs were generated by GraphPad Prism 8. Electrophysiology traces were generated from AxoGraph. The sample size (n), statistical test, p-values, and significance for each experiment are indicated in the legends. For all laminar density data, 2-way ANOVA was performed with Bonferroni correction for multiple comparisons. For all electrophysiology and morphology data, Mann-Whitney U tests were performed as some data did not follow the normal distribution or the sample size was too small to determine normality. For all calcium imaging data, normality was first determined by the Shapiro-Wilk test. Data that did not follow the normal distribution was analyzed by Mann-Whitney U tests, while normal data was analyzed by two-tailed unpaired t-test, with Welch’s correction if necessary. Statistical significance is indicated as *p < 0.5, **p < 0.01, ***p < 0.001, and ****p < 0.0001.

DATA AND CODE AVAILABILITY

This study did not generate any unique datasets or code. ADDITIONAL RESOURCES None.

LEAD CONTACT AND MATERIALS AVAILABILITY

Further information and requests for resources and reagents should be directed to and will be fulfilled by the Lead Contact, Natalia De Marco Garcia ( nad2018@med.cornell.edu ). This study did not generate new unique reagents.

EXPERIMENTAL MODEL AND SUBJECT DETAILS

All animal care and procedures were performed according to the Weill Cornell Medicine Research Animal Resource Center guidelines. Animals were housed in a controlled environment on a 12-hour light/dark cycle with food and water ad libitum . All animal lines have been described previously: Lhx6-Cre (JAX 026555), VGAT fl/fl (JAX 012897), Ai96 (RCL-GCaMP6s) (JAX 024106), 5HT3aR-EGFP (MMRRC 000273-UNC), Ai9 (RCL-tdT) (JAX 007909), 5Ht3aR-Cre (MGI:5435492, a gift from N. Heintz, Rockefeller University), Emx1 Cre (JAX 005628), SST Flp (JAX 028579), GABA A γ2 (Gabrg2 tm2lusc ) (JAX 016830). Mice of both sexes were used for analysis. Due to the young age of the mice, the influence of sex was not analyzed. For timed pregnancies, noon on the day of the vaginal plug was counted as E0.5. Since we did not detect significant statistical differences between Lhx6.VGAT +/+ and Lhx6.VGAT fl/+ mice in all of the analyses, we combined these genotypes in the control group. Lhx6.VGAT fl/fl mice are viable until P16 and do not show spontaneous seizures (0/300 mice) in the first two postnatal weeks. This agrees with other reports indicating that mouse mutants with developmental defects in interneuron number or VGAT deletion do not develop seizures until adulthood ( Cobos et al. 2005 ; Lin et al. 2018 ).

METHOD DETAILS Immunohistochemistry

Pups were anesthetized by hypothermia (< or equal to P6) or euthasol (>P7) and perfused via the left ventricle with 10 mL ice-cold PBS followed by 10 mL ice-cold 4% PFA. Brains were post-fixed in 4% PFA for 45 minutes on ice and equilibrated in 30% sucrose for > 16 hours at 4 C before being mounted in tissue-tek (Sakura). 20 μm sections were used for all counts except for ClCsp3 counts, which used 14 μm sections. After 3× 5 minute PBS washes, sections were blocked with 5% donkey serum, 0.03% Triton in PBS at room temperature for 1 hour. Sections were incubated with appropriate primary antibodies overnight (>16 hours) at 4 C: 1:1000 Rb-Sst (Abcam), 1:500 Rt-Sst (Millipore); 1:1000 Rb-Pv (Abcam), 1:1000 Gt-GFP (Rockland), 1:1000 Rt-GFP (nacalaitesque), 1:1000 Ch-RFP (Rockland), 1:250 Rb-ClCsp3 (Cell Signaling Technologies), 1:1000 Streptavidin Pacific Blue (Thermo Fisher). After more than 3× 5 minute PBS washes, sections were incubated with appropriate secondary antibodies at room temperature for 1 hour and stained for DAPI before mounting. Slides were mounted with Vectashield (hard set). Note that we found both somatostatin antibodies work most optimally on tissue post-fixed 20pA) and temporal characteristics (>100ms maximal width). We confirmed that GDP events were polysynaptic sIPSCs by recording at 0mV and abolishing them with GABA-A antagonist SR95531. For representative continuous cGDP traces where a series resistance test was applied every 10 seconds at the beginning of the episode, the series resistance test was blanked from 0 to 0.5 s. In Kir2.1 experiments, 1000 ms hyperpolarizing voltage steps (10pA) between −120mV and 0mV were applied and the maximum inward currents were measured. 5μM Barium Cl-(Sigma) was washed on in bath to abolish Kir2.1 currents. Following the recording session, pipettes were retracted and slices were post-fixed in 4% PFA and treated with fluorescence conjugated streptavidin labeling and immunohistochemistry. To post hoc identify MGE interneurons, we used their expression of tdTomato together with staining for Sst and/or Reelin to identify Sst cells. We were unable to use Pv as a marker as it is not expressed until ~P12. In addition, we used standard methods for classifications combined with unbiased dimensionality reduction and unsupervised hierarchical clustering to confirm the post-hoc immunohistochemical identity ( Anastasiades et al. 2016 ; Ferrante et al. 2017 ). Clustering generated three primary clusters FS, IB, and a third cluster of non-fast spiking adapting (NFS) neurons ( Figure S1E ) ( Ferrante et al. 2017 ; Krimer et al. 2005 ). Since immature FS interneurons may fail to exhibit narrow AP widths characteristic of the Pv class during the first postnatal week of development ( Goldberg et al. 2011 ), the NFS and FS clusters may contain misclassified interneurons. The difficulty in discriminating between immature FS-like(pPv) and NFS (Sst) interneurons at this age is exemplified by a rare case (1/13) in which an Sst interneuron was classified as FS in the clustering algorithm ( Figure S2E ). Since 92% of Sst neurons clustered in IB and NFS, we used those cluster groups as surrogates for putative Sst and pPv interneurons. For interneurons that did not successfully sort into one of the three FS, IB, or NFS clusters via Wards hierarchical clustering, their intrinsic properties were excluded from Table S2 . Optogenetic experiments were conducted using a Mightex LED arm attachment. ChR2-expressing cells were depolarized using a 5 ms pulse of 480 nm light. Light-evoked responses were confirmed to generate APs in a subset of experiments using K+ based pipette solution in the presence of 1μM tetrodotoxin (TTX) and 1mM 4-aminopyridine (4-AP). Cells were post hoc recovered to confirm interneuron identity.

Cranial Window Surgery and Image Acquisition

Cranial window surgery and 2-photon imaging were performed as previously described ( Che et al. 2018 ). Briefly, P6–7 pup was anesthetized on ice, a head plate was attached, and a piece of skull was removed without disturbing the pia. A glass coverslip was placed over the cranial window and secured. Following recovery from cranial window surgery (1-hour minimum), pups were head-fixed and imaged by FluoView FVMPE-RS multiphoton imaging system via a 25X (1.05 NA) water immersion lens (Olympus). Pups were un-anaesthetized and remained in resting state throughout the imaging session. LII/III was imaged at 150–200 μm below pia; LV was imaged at 350 m and below. For AAV.EF1a.Kir2.1.zsGreen and AAV.CAG.flex.JRGECOa1 experiments, we imaged first at two-photon excitation of 900 nm and detected GFP emission with a 495/540 nm filter. Once we identified the area infected with Kir2.1GFP, we shifted the two-photon excitation to 1040 nm to excite JRGECOa1 and detected the emission with a 575/645 filter (red channel). For dual GCaMP6s/TdTomato imaging, we concurrently imaged at two-photon excitation of 900 nm and separated the emission spectra with two filters (green signal a 495/540 nm band-pass filter and red signal a 575/645 filter). We shifted the 2-photon excitation wavelength from the GCaMP6s peak of 920 nm to 900 to reduce the excitation of tdTomato.

Calcium Imaging Analysis

Analysis was performed as previously described ( Che et al. 2018 ). For videos with virally expressed GCaMP6s and jRGECO1a, neuropil correction was performed on all videos ( Che et al. 2018 ). For chronic imaging, Lhx6.VGAT.Ai9 mice injected with AAV.ef1a.DIO.GCaMP6s with cranial window implants were imaged from P7 through P8, identifying the same region of the cortex on each day. Regions were aligned using vascular landmarks, both on the surface and visible deeper by nature of the Lhx6 promoter expression in endothelium. After a FOV was selected and imaged, a Z stack of 50 μm above and 50 μm below the FOV at 4 m interval was taken. In the tdTomato channel, each red cell in the P7 video was either located or deemed dead based on the matching P8 video. In addition, if a cell could not be found in the P8 video, it was searched in the z stack, as it was common to see cells moving out of the imaging plane due to the still rapid growth of the brain at this age. 6 to 23 dead cells were identified in FOVs containing 140–252 cells. Cell masks were made on the P7 video using methods previously described, and calcium dynamics were analyzed for each living versus dead cell ( Che et al. 2018 ). Assembly analysis (edited from Methods in the companion manuscript Modol et al) GCE detection: To detect population events we randomly circularly reshuffled each spike vector, hence maintaining within contour spike dependencies while randomizing the population and correcting for differences in baseline firing for different animals and ages. One thousand surrogate distributions were created. The spikes of each frame for these distributions were computed and the 99th percentile of the resulting “sum of spikes” vector was used as a statistical threshold. Peaks above the threshold that were at least separated by 7 frames were considered as synchronous calcium events. A population vector is a binary representation of the spiking activity of the event (peak +/− 3 frames), where a 1 is assigned when a contour has a spike within the interval and 0 otherwise. Spike distributions: To detect the calcium onsets of contours and quantify their individual contribution to synchronous network events, the averaged onset time of the first spike within an event was considered for all contours as well as the ratio of events that recruited a given. We then computed the probability density function of onsets and scaled them to the maximum participation rate per group. Statistical analyses were then applied to measure differences in the onset dynamics (see below). Identification of cell assemblies in GCE: Analyses were performed as described in ( Malvache et al., 2016 ). Cell assemblies were identified using a clustering algorithm based on GCE similarity for cell participation followed by a statistical test for cell participation in each GCE cluster. The GCE similarity metric was the squared Euclidian distance between columns of the normalized covariance matrix. This similarity metric allowed for a more efficient clustering. Unsupervised clustering of GCE was obtained by running the k-means algorithm on this metric with cluster numbers ranging from 2 to 19. Hundred iterations of k-means were run for each cluster number and the iteration that resulted in the best-averaged Silhouette value was kept. The silhouette value was computed as described previously ( Malvache et al., 2016 ). A random distribution of average silhouette values for each cluster was calculated by reshuffling cell participation across different GCE and applying the same algorithm. Clusters with average silhouette values exceeding the 95th percentile of the random case were considered as statistically significant. Each cluster was then associated to a cell assembly that comprised those cells that significantly participated to the GCE events within that particular cluster. Cell participation to a given cluster was considered statistically significant if the fraction of synchronous events in that cluster that activated the cell exceeded the 95th percentile of reshuffled data. If a cell was significantly active in more than one GCE cluster, it was associated to the one in which it participated the most (percentage wise). The overlap between assemblies was quantified by calculating the silhouette value of each cell (with the normalized hamming distance between each cell pair as a dissimilarity metric). A cell was significantly involved in a single assembly if its silhouette value was higher than expected by chance (95th percentile after reshuffling). GCEs were finally sorted with respect to their projection onto cell assemblies. A GCE was activating a given neuronal assembly if the number of cells recruited in that assembly was higher than expected by chance (95th percentile after reshuffling). Topological distribution of cell assemblies: For each cell assembly, the spatial silhouette value was computed using the pairwise anatomical distance between cells as a dissimilarity metric. To assess the statistical significance of spatial clustering for each cell assembly, its silhouette value was compared to the ones obtained after cell reshuffling (1000 iterations). Spatial clustering was considered significant if the silhouette value of the empirical assembly exceeded a certain percentile of the reshuffled silhouette value distribution. For each session, the percentile was set at (1 – 0.05/(number of clusters) to account for the inter-dependence of the comparisons (Bonferroni correction). T-distributed Stochastic neighbor embedding (t-SNE) Due to commonly associated shortcomings of the k-means algorithm, we used the t-distributed stochastic neighbor embedding method ( Van Der Maaten and Hinton, 2008 ), to visualize the data and also verify the concurrence with the k-means analysis. The t-SNE technique converts correlations between contour vectors to joint probabilities and tries to minimize the Kullback-Leibler divergence (KL divergence) between the joint probabilities of the low-dimensional embedding and the high-dimensional data. We first reduced the dimensionality of the data by considering the first 20 Principal Components of contours participation in GCEs. These components were then non-linearly mapped onto a two-dimensional space in a way that preserved both local and global structure of the data. The perplexity parameter (the number of nearest neighbors) was chosen in an iterative procedure that minimized the KL divergence. These values ranged between 15 and 25. We then color-coded each component according to the previous k-means analysis.

Supplementary Material 1 2 Video S1 (Related to Figure 1 ) . Example of spontaneous activity in an Lhx6.GCaMP6s control mouse in SSBF at P7 using two-photon calcium imaging. Movie is played back at 15fps. FOV is 509 × 509 μm. 3 Video S2 (Related to Figure 1 ) . Example of spontaneous activity in an Lhx6.VGAT fl/fl .GCaMP6s mouse in SSBF at P7 using two-photon calcium imaging. Movie is played back at 15fps. FOV is 509 × 509 μm. 4 Video S3 (Related to Figure 6 ) . Example of spontaneous co-activation of pyramidal cells and interneurons in an Lhx6.Ai9 control mouse in SSBF at P8 using two-photon calcium imaging. GCaMP6s was expressed by AAV.hsyn.GCaMP6s in pyramidal cells and AAV.EF1A.DIO.GCaMP6s in MGE-cINs, injected at P0. Movie is played back at 15fps. FOV is 509 × 509 μm. 5 Video S4 (Related to Figure S6 ) . Example of spontaneous co-activation of pyramidal cells in a control mouse in SSBF at P8 using two-photon calcium imaging. GCaMP6s was expressed by AAV.hsyn.GCaMP6s in pyramidal cells, injected at P0. Movie is played back at 15fps. FOV is 509 × 509 μm. 6 Video S5 (Related to Figure S6 ) . Example of spontaneous co-activation of pyramidal cells in an Lhx6.VGAT fl/fl mouse in SSBF at P8 using two-photon calcium imaging. GCaMP6s was expressed by AAV.hsyn.GCaMP6s in pyramidal cells, injected at P0. Movie is played back at 15fps. FOV is 509 × 509 μm. 7