Tightly-coupled inhibitory and excitatory functional networks in the early visual cortex

Intracortical inhibition plays a critical role in shaping activity patterns in the mature cortex. However, little is known about the structure of inhibition in early development prior to the onset of sensory experience, a time when spontaneous activity exhibits long-range correlations predictive of mature functional networks. Here, using calcium imaging of GABAergic neurons in the early ferret visual cortex, we show that spontaneous activity in inhibitory neurons is already highly organized into distributed modular networks before visual experience. Inhibitory neurons exhibit spatially modular activity with long-range correlations and precise local organization that is in quantitative agreement with excitatory networks. Furthermore, excitatory and inhibitory networks are strongly co-aligned at both millimeter and cellular scales. These results demonstrate a remarkable degree of organization in inhibitory networks early in the developing cortex, providing support for computational models of self-organizing networks and suggesting a mechanism for the emergence of distributed functional networks during development.


Introduction 25
Inhibition is crucial for shaping neural activity and response properties in mature cortex. GABAergic 26 interneurons have been implicated in a range of computations, including response gain, stimulus 27 discrimination, and network stabilization (for reviews, see Isaacson and Scanziani, 2011;Denève and 28 Machens, 2016; Ferguson and Cardin, 2020). In the columnar visual cortex, inhibitory neurons actively 29 shape response selectivity (Wilson, Scholl and Fitzpatrick, 2018) and are organized into functionally-30 specific networks with excitatory neurons (Wilson et al., 2017). However, relatively little is known about 31 inhibition in the early cortex prior to the onset of sensory experience. GABAergic inhibition is initially 32 absent in early development before progressively strengthening as the cortex matures (reviewed in Ben-33 Ari, 2002; Huang et al., 2007), raising the possibility that inhibition plays only a minor role in shaping 34 early patterns of cortical activity. 35 Recent work in the ferret visual cortex demonstrated that prior to the onset of visual experience, 36 excitatory activity is already highly structured, showing modular and distributed activity with long-range 37 correlations (Smith et al., 2018), reminiscent of the columnar stimulus-evoked activity found in the 38 mature cortex (Hubel and Wiesel, 1968;Blasdel and Salama, 1986 Inhibition could be weak or absent in the early cortex, or 49 be broad and unstructured, operating over a much larger 50 spatial scale than excitation. In contrast, early inhibition 51 might be organized and modular, but poorly aligned with 52 excitation. Finally, the highly organized and co-aligned 53 patterns of excitation and inhibition found in mature 54 animals (Wilson et al., 2017) might already be present in 55 the early cortex. Resolving this question is critical to both 56 constrain models of early network development-many 57 of which rely on structured intracortical inhibition-and 58 to understand the degree to which inhibition contributes 59 to early distributed patterns of cortical activity. 60 Here we address this by using wide-field and two-photon calcium imaging of spontaneous activity to 61 examine the structure of inhibitory networks in the developing ferret visual cortex. We first 62 demonstrate in vivo that already in the early cortex, GABAergic signaling exerts a strong inhibitory 63 effect. Next, by employing the inhibitory neuron-specific enhancer mDlx (Dimidschstein et al., 2016) 64 paired with the genetically encoded calcium indicator GCaMP6, we find that inhibitory neurons exhibit 65 modular patterns of activity that extend over several millimeters in the early visual cortex. These 66 patterns reveal long-range correlated networks with precise local organization, highly similar in 67 structure to those of excitatory networks. Furthermore, we find that inhibitory and excitatory networks 68 show a remarkable degree of spatial consistency and co-alignment on both global and cellular scales. 69 These findings clearly demonstrate the long-range and fine-scale organization of intracortical inhibition 70 in the developing cortex, and support the ability of large-scale cortical networks to self-organize through 71 precisely correlated local excitatory and inhibitory activity. 72

Results 73
Early spontaneous activity in inhibitory populations is modular and forms large-scale correlated 74 networks. 75 Over the course of development, GABAergic signaling switches from exerting depolarizing effects to 76 hyperpolarizing due to the maturation intracellular chloride concentrations (Ben-Ari, 2002). In the ferret, 77 Schematic showing potential arrangements of inhibitory and excitatory activity patterns in early cortex. Top row: Excitation is known to be highly modular and spatially structured. Middle row: Inhibition could be absent (a), broadly unstructured (b), or modular (c and d); if modular, the spatial arrangement of those modules could be shifted with respect to excitatory activity (c), or already well aligned as found in the mature cortex (d). Scale bar 1 mm. functional GABAergic synapses have been shown to be present as early as postnatal day 20 (P20) (Dalva,78 2010), but because these experiments utilized whole-cell recordings in cortical slices, it remains unclear 79 whether these synapses exert an inhibitory effect at this age. To address this, we expressed GCaMP6s in 80 excitatory neurons at P21-23 via AAV1-hSyn-GCaMP6s (Wilson et al., 2017) and imaged spontaneous 81 activity prior to and following direct application of the GABA(A) antagonist bicuculline methiodide (BMI)  82 to the cortex (Figure 2a-b). BMI resulted in a pronounced increase in the frequency (Figure 2c.   In the mature cortex, orientation preference maps exhibit smooth variation, punctuated by abrupt 106 discontinuities at orientation pinwheels and fractures (Bonhoeffer and Grinvald, 1991;Ohki et al., 2005), 107 an organization shared by inhibitory neurons in the mature ferret (Wilson et al., 2017). Notably, prior 108 work identified similar discontinuities in the patterns of distributed correlated activity (termed 109 correlation fractures) in excitatory neurons early in development prior to the emergence of orientation 110 maps (Smith et al., 2018). To determine if early inhibitory networks exhibit a similar fine-scale 111 organization, we calculated the rate of change in correlation patterns for all pixels within our field-of-112 view (Supplemental Figure 1). This analysis revealed the presence of spatially discrete fractures over 113 which inhibitory correlation patterns exhibit abrupt changes in structure, demonstrating a precise local 114 organization in the coupling to large scale inhibitory networks in the early visual cortex (Figure 3f). 115 Quantitative agreement between inhibitory and excitatory network structure in early cortex 116 The presence of modular activity patterns, together with long-range correlations exhibiting precise 117 local structure, suggests that inhibitory neurons may already be tightly integrated into functional 118 networks with excitatory cells in the early cortex. To begin to address this, we undertook a quantitative 119 comparison of excitatory and inhibitory networks assessed through spontaneous activity. We performed 120 wide-field imaging in animals expressing either hSyn.GCaMP6s, which is excitatory-specific in ferret 121 (Wilson et al., 2017), or the inhibitory-specific mDlx.GCaMP6s. We first quantified the size of modular 122 Wilcoxon rank-sum test), supporting the idea that the modular domains of local inhibitory activity 127 operate on the same scale as excitatory activity. 128 Next, we quantitatively examined the correlation patterns for excitatory and inhibitory networks 129 revealed by spontaneous activity. We find that correlations were similarly long-range, with equivalent 130 Prior work has shown that in locally heterogeneous network models that recapitulate the structure 145 of early excitatory networks, long-range correlations exhibiting both local anisotropy and pronounced 146 fractures strongly coincide with spontaneous activity patterns that reside in a low dimensional subspace 147 (Smith et al., 2018). Given the similarities in both local and long-range correlation structure, as well as 148 fracture strength between excitatory and inhibitory networks, we computed the dimensionality of 149 inhibitory spontaneous events and compared them to event-number-matched events in excitatory 150 neurons. We found that across animals, inhibitory and excitatory events tended to reside in similarly low 151 dimensional subspaces ( Example spontaneous inhibitory (mDlx.GCaMP6s, top) and excitatory (syn.jRCaMP1a, bottom) events recorded from same animal (I color axis: -2-2 z-score, E color axis: -1-1 z score). Example events were chosen based on pattern similarity. b. Highly similar correlation patterns in inhibitory and excitatory networks for corresponding seed points. c. Quantification of I vs. E correlation similarity for all seed points. Circles indicate seed points illustrated in (b). d. Similarity for all seed points in (c) falls within 95% confidence intervals of bootstrapped I vs I similarity. e. Within animal comparison of correlation strength for E and I (median and IQR within animal). Filled circles individually significant vs. surrogate (p < 0.01). f. Similarity of I vs E correlations is significantly greater than shuffle (median and IQR within animal). Shaded bars indicate IQR of bootstrapped I vs I similarity. g. Long-range correlations show significant network similarity. Correlation similarity remains significant vs. surrogate (red, 95% CI of surrogate, averaged across animals) for increasingly distant regions (excluding correlations within 0.4 -1.4 mm from the seed point).
To assess whether the large-scale alignment of excitatory and inhibitory networks observed above 184 extends to the cellular level, we performed simultaneous two-photon imaging of excitatory and 185 inhibitory neurons (P23-26). We co-injected AAVs expressing hSyn-GCaMP6s (excitatory) and mDlx-186 GCaMP6s-P2A-NLS-tdTomato (inhibitory), allowing us to distinguish inhibitory neurons in vivo through 187 tdTomato expression (Figure 6a, Supplemental Figure 3). We observed highly modular spontaneous 188 activity in both excitatory and inhibitory cells, with local populations showing tightly coordinated 189 patterns of activity across cell types (Figure 6b, Supplemental Figure 4). ( Figure 6e,f). Together, these results show that excitatory and inhibitory neurons are precisely organized 203 in the early visual cortex into the same spatially structured and locally correlated functional networks. 204

Discussion 205
By applying inhibitory interneuron-specific expression of fluorescent calcium sensors to the early 206 visual cortex, we were able to directly assess the structure of inhibitory networks in the developing 207 cortex and their integration with excitatory networks. We show that prior to eye opening and the onset 208 of reliable stimulus-evoked responses (Chapman, Stryker and Bonhoeffer, 1996), intracortical inhibition 209 is both present and highly organized, with inhibitory networks already displaying patterns of modular 210 and correlated spontaneous activity that span several millimeters. Inhibitory networks exhibit 211 quantitatively similar structure to excitatory networks, which together show precise alignment at both 212 local and global scales in the patterns of correlated spontaneous activity. Together, these findings 213 demonstrate the presence of tightly-coupled excitatory and inhibitory functional networks in the early 214 visual cortex. 215 In the mature ferret, both inhibitory and excitatory neurons are organized into a columnar map of 216 orientation preference, with orientation-specific subnetworks of functionally coupled excitatory and 217 inhibitory neurons present not only within iso-orientation domains but also near orientation pinwheels 218 ( it difficult to separate inhibitory neurons by subtype at these ages, our results suggest that a similarly 237 organized and integrated structure to that in the mature ferret may already exist in the developing 238 cortex. Future work, potentially leveraging inhibitory subtype-specific viral approaches ( Viral Injection. Viral injections were performed as previously described (Smith and Fitzpatrick, 2016). 399 Briefly we expressed GCaMP6s in inhibitory interneurons by microinjecting AAV1-mDLx-GCaMP6s-P2A- Anesthesia was induced with isoflurane (3.5-4%) and maintained with isoflurane (1-1.5%). 407 Buprenorphine (0.01mg/kg) and either atropine (0.2 mg/kg) or glycopyrrolate (0.01 mg/kg) were 408 administered, as well as 1:1 lidocaine/bupivacaine at the site of incision. Animal temperature was 409 maintained at approximately 37 °C with a water pump heat therapy pad (Adroit Medical HTP-1500, 410 Parkland Scientific). Animals were also mechanically ventilated and both heart rate and end-tidal CO2 411 were monitored throughout the surgery. Using aseptic surgical technique, skin and muscle overlying 412 visual cortex were retracted, and a small burr hole was made with a handheld drill (Fordom Electric Co.). 413 Approximately 1 µL of virus contained in a pulled-glass pipette was pressure injected into the cortex at 414 two depths (~200 µm and 400 µm below the surface) over 20 min using a Nanoject-II (World Precision  415 Instruments). The craniotomy was filled with 2% agarose and sealed with a thin sterile plastic film to 416 prevent dural adhesion. 417 Cranial window surgery. On the day of experimental imaging, ferrets were anesthetized with 3%-4% 418 isoflurane. Atropine was administered as in virus injection procedure. Animals were placed on a 419 feedback-controlled heating pad to maintain an internal temperature of 37 to 38 C. Animals were 420 intubated and ventilated, and isoflurane was delivered between 1 and 2% throughout the surgical 421 procedure to maintain a surgical plane of anesthesia. An intraparietal catheter was placed to deliver 422 fluids. EKG, end-tidal CO2, and internal temperature were continuously monitored during the procedure 423 and subsequent imaging session. The scalp was retracted and a custom titanium headplate adhered to 424 the skull using C&B Metabond (Parkell). A 6 to 7 mm craniotomy was performed at the viral injection 425 site and the dura retracted to reveal the cortex. One 4mm cover glass (round, #1.5 thickness, Electron 426 Microscopy Sciences) was adhered to the bottom of a custom titanium insert and placed onto the brain 427 to gently compress the underlying cortex and dampen biological motion during imaging. The cranial 428 window was hermetically sealed using a stainless-steel retaining ring (5/16-inch internal retaining ring, 429 McMaster-Carr). Upon completion of the surgical procedure, isoflurane was gradually reduced (0.6 to 430 0.9%) and then vecuronium bromide (0.4 mg/kg/hr) mixed in an LRS 5% Dextrose solution was delivered 431 IP to reduce motion and prevent spontaneous respiration. 432 Widefield epifluorescence and two-photon imaging. Widefield epifluorescence imaging was performed 433 with a sCMOS camera (Zyla 5.5, Andor; Prime BSI express, Teledyne) controlled by μManager (Edelstein 434 et al., 2010). Images were acquired at 15 Hz with 4 × 4 binning to yield 640 × 540 pixels (Zyla) or 2x2 435 binning and additional offline 2x2 binning to yield 512 x 512 pixels (Prime BSI). Two-photon imaging was 436 performed with a commercial microscope (Neurolabware) driven by an Insight X3 laser (Spectra 437 Physics). Imaging was performed at 920 nm (GCaMP) and 1040 nm (tdTomato), and fluorescence was 438 collected on separate PMTs using a 562 nm dichroic mirror and 510/84 nm (GCaMP) and 607/70 nm 439 (tdTomato) emission filters (Semrock). Images were collected at 796 x 512 pixels at 30 Hz. 440 Spontaneous activity was captured in 10-minute imaging sessions, with the animal sitting in a darkened 441 room facing an LCD monitor displaying a black screen. 442 Immunostaining and imaging. To confirm that tdTomato expression from AAV1-mDLx-GCaMP6s-P2A-443 NLS-tdTomato was specific to inhibitory neurons as expected from previous work (Wilson et al., 2017), 444 we performed immunostaining in a subset of animals. Following imaging, animals were euthanized and 445 transcardially perfused with 0.9% heparinized saline and 4% paraformaldehyde. The brains were 446 extracted, post-fixed overnight in 4% paraformaldehyde, and stored in 0.1 M phosphate buffer solution. 447 Using a vibratome, brains were tangentially sectioned along the surface of the imaging window (50 μm 448 steps). Slices were stained for GAD67 using Mouse anti-GAD67 (1:1000, Sigma Aldrich, MAB5406) and 449 Alexa 405 donkey anti-mouse (1:500, Abcam, AB175659) as described (Wilson et al., 2017). Imaging was 450 performed on a confocal microscope (Nikon C2). 451 Bicuculline application. To test whether GABAergic signaling exerted net inhibitory effects at the ages 452 examined in this study, we bath applied the GABA(A) antagonist bicuculline methiodide (BMI) directly to 453 early cortex at P21-23 in animals expressing AAV1-Syn-GCaMP6s-WPRE-SV40. To apply BMI, the cannula 454 covering the cranial window was removed and the exposed cortex was gently flushed with ACSF. After 455 collecting 10 minutes of baseline spontaneous activity, 10 μM BMI in ACSF was bath applied to the 456 cortex, and spontaneous activity was imaged for an additional 10 minutes. 457

Data analysis 458
Signal extraction for widefield epifluorescence imaging. Image series were motion corrected using rigid 459 alignment and an ROI was manually drawn around the cortical region of GCaMP expression. 460 Additionally, an ROI mask was manually drawn around blood vessels to remove vessel artefacts. The 461 baseline fluorescence (F0) for each pixel was obtained by applying a rank-order filter to the raw 462 fluorescence trace with a rank 70 (for excitatory data) or 190 (for inhibitory data) and a time-window of 463 30 s. The rank and time window were chosen such that the baseline faithfully followed the slow trend of 464 the fluorescence activity. The baseline corrected spontaneous activity was calculated as (F -F0)/F0 = Δ 465 F/F0. 466 Event detection. Detection of spontaneously active events was performed essentially as described 467 (Smith et al., 2018). Briefly, we first determined active pixels on each frame using a pixel-wise threshold 468 set to 3 s.d. above each pixel's mean value across time. Active pixels not part of a contiguous active 469 region of at least 0.01 mm 2 were considered 'inactive' for the purpose of event detection. Active frames 470 were taken as frames with a spatially extended pattern of activity (>50% of pixels were active). 471 Consecutive active frames were combined into a single event starting with the first high-activity frame 472 and then either ending with the last high-activity frame or, if present, an activity frame defining a local 473 minimum in the fluorescence activity. To assess the spatial pattern of an event, we extracted the 474 maximally active frame for each event, defined as the frame with the highest activity averaged across 475 Where H is the amplitude of the gaussian, x1 (y1) is the x (y) center, σx (σy) is the standard deviation of 508 the x (y) component, is the angular rotation of the gaussian, and y0 is the offset. Domain size was 509 calculated as the full-width at a tenth of the maxima (FWTM) of the minor axis of fitted gaussian: 510 = 2√2ln(10) 511 Strength of long-distance correlations. To determine the strength of correlations, we first identified 512 local maxima (minimum separation between maxima: 800 µm) in the correlation pattern for each seed 513 point. To assess the statistical significance of long-range correlations ~2mm from the seed point, we 514 compared the median correlation strength for maxima located 1.8-2.2mm away against a distribution 515 obtained from 100 surrogate correlation patterns. Surrogate correlation patterns control for 516 correlations that arise from finite sampling by eliminating most of the spatial relationship between 517 patterns (Smith et al., 2018). Surrogate correlation patterns were generated from spontaneous events 518 that were randomly rotated (rotation angle drawn from a uniform distribution between 0° and 360° with 519 a step size of 10°), translated (shifts drawn from a uniform distribution between±450 µm in increments 520 of 26 µm, independently for x and y directions) and reflected (with probability 0.5, independently at the 521 x and y axes at the center of the ROI). 522 Correlation pattern wavelength. To find the wavelength of the correlation patterns, we first centered 523 and averaged the local neighborhood (1500 μm radius) across all seed points. We then collapsed around 524 the angle to obtain the average correlation as a function of distance from the seed point and used spline 525 interpolation to fit the data. The wavelength of the resulting spline interpolation was estimated as the 526 distance to the first local maxima after 0. 527 Eccentricity of local correlation structure. To describe the shape of the local correlation pattern around 528 a seed point, we fit an ellipse (least-square fit) with orientation ϕ, major axis ς1 and minor axis ς2 to the 529 contour line at correlation=0.7 around the seed point. The eccentricity ε of the ellipse is defined as: 530 Where ε = 0 is a circle, and increasing values indicating greater elongation along the ellipse. 532 Dimensionality of spontaneous activity. We estimated the dimensionality deff of the subspace spanned 533 by spontaneous activity patterns by (Abbott, Rajan and Sompolinsky, 2011): 534 where λi are the eigenvalues of the covariance matrix for the N pixels within the ROI. As the value of the 536 dimensionality is sensitive to differences in detected event number, to estimate the distribution of the 537 dimensionality for each animal we calculated the dimensionality of randomly sub-sampled events (n=30 538 events, matched across animals, 100 simulations) and took the median of the distribution. 539 Comparison of inhibitory and excitatory correlation similarity. To compare the similarity between 540 inhibitory and excitatory correlation patterns within the same animal, we computed the second-order 541 correlation between patterns. For each seed point, we calculated the second order Pearson's correlation 542 between corresponding correlation patterns, while excluding pixels within a 400 μm exclusion radius 543 around the seed point to prevent local correlations from inflating the similarity between the two 544 networks. To get an estimate the upper bound of similarity within inhibitory networks, given a finite 545 sampling size, we randomly split the detected inhibitory events into two groups and separately 546 computed correlations and the second-order correlations between the halves (n simulations=100.) To 547 determine if the observed networks are more similar than chance, we calculated the similarity between 548 the excitatory network and a surrogate inhibitory network (surrogate events calculated as above, with 549 100 simulations and surrogate similarity calculated as median and IQR across simulations). To determine 550 if spontaneous correlations far from the seed point also maintain high degrees of similarity, we 551 systematically increased the size of the exclusion radii, calculating similarity only using data far from the 552 seed points. Exclusion radius size ranged from 400 to 1400 μm, in 200 μm steps. 553 Two-photon event detection and cellular correlations. Two-photon images were corrected for rigid in 554 plane motion via a 2D cross-correlation. Cellular ROIs were drawn using custom software (Cell Magic 555 Wand ( where Fneuropil was taken as the average fluorescence within 30 µm excluding all cellular ROIs and =0.6: 561 Activity was taken as ∆F/F0, where F0 was the baseline fluorescence obtained by applying a 60-s median 563 filter, followed by a first order Butterworth high-pass filter with a cutoff time of 60s. To compute 564 spontaneous correlations, we first identified frames containing spontaneous events, which were defined 565 as frames in which > 10% of imaged neurons exhibited activity > 2 s.d. above their mean. Cellular activity 566 on all event frames was then z-scored using the mean and s.d. of each frame, and pairwise Pearson's 567 correlations were computed across all neurons over all active frames. 568 The similarity of cellular correlation patterns for inhibitory and excitatory cells was computed by first 569 creating a spatially matched set of excitatory and inhibitory neurons. For each inhibitory cell in the field 570 of view, a spatially matched excitatory cell was identified by finding the closest excitatory cell (within 50 571 µm). Then, for every cell in the field of view (excitatory and inhibitory) pairwise correlations were 572 calculated with either inhibitory cells or the spatially matched excitatory cells. The similarity of these 573 correlation patterns was taken as the second order Pearson's correlation as above, while excluding cells 574 <200 µm from the seed neuron. Statistical significance was computed by comparing measured similarity 575 values against a shuffled distribution obtained by shuffling the pairwise correlations before computing 576 similarity (100 shuffles). 577 To compare the amplitude of events within a local area, activity traces for each cell were smoothed with 578 a 3-sample median filter and z-scored across all frames. We then computed the average amplitude of all 579 inhibitory (or excitatory) cells within a 75 µm radius for the maximally active frame of each event. 580 Quantification and statistical analysis. Nonparametric tests were used for statistical testing throughout 581 the study. Bootstrapping was used determine null distributions when indicated. Center and spread 582 values are reported as median and inter-quartile range, unless otherwise noted. Statistical analyses 583 were performed in MATLAB and Python, and significance was defined as p < 0.05. 584