Emerging Activity Patterns and Synaptogenesis in Dissociated Hippocampal Cultures

Cultures of dissociated hippocampal neurons display a stereotypical development of network activity patterns within the first three weeks of maturation. During this process, network connections develop and the associated spiking patterns range from increasing levels of activity in the first two weeks to regular bursting activity in the third week of maturation. Characterization of network structure is important to examine the mechanisms underlying the emergent functional organization of neural circuits. To accomplish this, confocal microscopy techniques have been used and several automated synapse quantification algorithms based on (co)localization of synaptic structures have been proposed recently. However, these approaches suffer from the arbitrary nature of intensity thresholding and the lack of correction for random-chance colocalization. To address this problem, we developed and validated an automated synapse quantification algorithm that requires minimal operator intervention. Next, we applied our approach to quantify excitatory and inhibitory synaptogenesis using confocal images of dissociated hippocampal neuronal cultures captured at 5, 8, 14 and 20 days in vitro, the time period associated with the development of distinct neuronal activity patterns. As expected, we found that synaptic density increased with maturation, coinciding with increasing spiking activity in the network. Interestingly, the third week of the maturation exhibited a reduction in excitatory synaptic density suggestive of synaptic pruning that coincided with the emergence of regular bursting activity in the network.


81
The goal of the present investigation is to quantify excitatory and inhibitory 82 synaptogenesis at different functional stages of development in dissociated hippocampal cell 83 cultures, ranging from sparsely connected networks that exhibit low levels of spiking activity to 84 densely connected mature networks that exhibit periodic bursting. To quantify synaptogenesis, 85 we evaluated and applied a novel automated, high-throughput approach, based on the spatial 86 correlation between presynaptic and postsynaptic structures. This approach does not significantly 87 depend on image intensity threshold and provides consistent estimation of noise components. We 88 validated our approach using synthetic images and then applied it to quantify excitatory and 89 inhibitory synaptogenesis during the first three weeks of maturation in dissociated hippocampal 90 cell cultures. We found that there is an increase in excitatory synaptic density during the first 91 weeks of maturation that coincided with increased spike activity at the network level. This was 92 . CC-BY 4.0 International license available under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which this version posted May 18, 2023. ;https://doi.org/10.1101https://doi.org/10. /2023 Page | 6 followed by a reduction in excitatory synaptic density towards the third week, suggesting that a 93 synaptic pruning phase occurs as these cultures develop that coincides with regular bursting 94 activity in the network. was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which this version posted May 18, 2023. ;https://doi.org/10.1101https://doi.org/10. /2023 Page | 7 non-neuronal cells such as glia.

148
To quantify network activity we calculated the mean firing rate (spikes/sec) and mean 149 burst rate (bursts/min) averaged across all electrodes in the MEA. Extracellular recordings were 150 filtered off-line by a digital filter (a Butterworth filter, second order band pass 300 Hz -1.5 kHz) 151 using Matlab software (MathWorks, Natick, MA, USA). The filtered output was used to detect 152 spikes, defined as negative deflections that exceeded five standard deviations of the filtered 153 signal. The multi-unit spike trains were saved in rasters as arrays of 0's (no spike) and 1's. To 154 detect bursts, the spike rasters were used as input to a leaky integrator with a time constant of 50 155 ms, a value close to the time constant of a hippocampal pyramidal cell (Staff et al. 2000). The

161
. CC-BY 4.0 International license available under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which this version posted May 18, 2023. ;https://doi.org/10.1101https://doi.org/10. /2023 Page | 9 In order to quantify synaptogenesis, cultures were fixed using 4% paraformaldehyde in 162 PBS and stained after 5, 8, 14 and 20 DIV. Excitatory and inhibitory synapses were stained in 163 separate cover-slip preparations. To identify excitatory synapses, neurons were triple stained to 164 label dendrites and excitatory pre-and post-synaptic terminals, using chicken-anti-MAP2 165 (Abcam 1.9µg/ml), rabbit anti-vGluT1 (Synaptic Systems, 10µg/ml), and mouse anti-PSD-95 166 (UC Davis/NIH NeuroMab Facility, 10µg/ml) respectively. To identify inhibitory synapses, 167 neurons were triple stained to label dendrites and inhibitory pre and post-synaptic terminals, 168 using chicken-anti MAP2 (Abcam 1.9 µg/ml), rabbit anti-vGAT (Synaptic Systems 10µg/ml), 169 and mouse anti-gephyrin (Synaptic Systems 10µg /ml). Binding was detected with Alexa Fluor 170 488-labeled, highly cross-adsorbed goat anti-chicken IgY, Alexa-647-labelled highly cross-171 adsorbed goat anti-rabbit IgG, and Alexa-594-labelled highly cross-adsorbed goat anti-mouse 172 IgG (1:500; Jackson). Cells were incubated in DAPI (300 nM for 2 min) to label nuclei, and 173 mounted in Cytoseal-60 (Thermo Scientific). Cover-slips were imaged with a 63×, 1.40 numerical aperture, oil immersion objective on a 177 laser scanning confocal microscope (Leica SP5 AOBS, in resonant scanner mode), with identical 178 illumination acquisition settings across staining conditions. We used 12-bit dynamic range and 179 highly-sensitive, linear HyD detectors for the pre-and post-synaptic puncta channels and 180 standard illumination settings were created to make use of the dynamic range for the typical 181 staining signals in this preparation. The image capture procedure is shown in Figure 1. For a 182 given age of the culture (5, 8, 14 or 20 DIV) and each synapse type (i.e., excitatory or 183 inhibitory), we fixed and stained cultures on two coverslips ( Fig 1A). In each coverslip we 184 . CC-BY 4.0 International license available under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which this version posted May 18, 2023. ;https://doi.org/10.1101https://doi.org/10. /2023 sampled 81 locations to capture images, each image measuring 51.2µm × 51.2 µm (1024 × 1024 185 pixels) (Fig 1B). We thus captured a total of 162 images for each synapse type, for a given age 186 within a culture. Sequential line scans were used to capture four separate channels of high 187 resolution images of dendrites (Alexa 594), presynaptic puncta (Alexa 647), post synaptic puncta 188 (Alexa 488) and nuclei (DAPI), using laser lines at 405 nm, 488 nm, 561 nm, and 633 nm, 189 respectively. Figure 1C 204 We performed an initial inspection prior to image analysis to discard images that were of 205 poor quality where the signal was indistinguishable from background. We used a criterion of 206 SNR < 0.5, as poor quality images. To remove distortion arising from the microscope's point 207 . CC-BY 4.0 International license available under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

Preprocessing and automated synapse detection
The copyright holder for this preprint (which this version posted May 18, 2023. ;https://doi.org/10.1101https://doi.org/10. /2023 spread function and to improve signal-to-noise ratio, images were pre-processed using Huygens 208 deconvolution software (Huygens v.4.5.1, SVI, Hilversum, The Netherlands) using maximum-209 likelihood estimation and signal to-noise ratio of 5. Subsequent image analysis involved 210 automated synapse quantification of series of images, implemented in batch-processing mode.

211
This was performed on a Windows 7 computer using a customized script written in ImageJ 212 (https://imagej.nih.gov/ij/index.html).

213
Synapse identification was applied to the maximum intensity projections from the three 214 z-planes, as our cell cultures are essentially a 2D monolayer of cells (neuropil thickness < 1 µm). 215 We identified a synapse as colocalized pre-and post-synaptic puncta with their centroids lying performed in a three-step procedure described in the following.

223
Step 1 involved the identification of putative pre-and post-synaptic puncta and dendrites 224 from the raw images (Fig 2A). For the puncta channels ( Fig 2A1-A2), we first applied a rolling-225 ball background subtraction to correct for unevenness in background illumination. We used a 226 rolling-ball radius of 4 pixels, which would produce a protected zone of 9-pixel-wide diameter 227 (450 nm) as sampled. We then used a 3×3 median filter, to remove point noise within the image.

228
To enable detection of low intensity signals while removing the low intensity noise component, 229 we employed a threshold at 45% of total intensity distribution which was below the mean 230 . CC-BY 4.0 International license available under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which this version posted May 18, 2023. ; https://doi.org/10.1101/2023.05.18.541345 doi: bioRxiv preprint Page | 12 intensity value commonly used (Schmitz et al. 2011;Schätzle et al. 2012;Harrill et al. 2015). We 231 then captured the pixel locations of all intensity-maxima exceeding this threshold, and expanded 232 each location by 2 pixels in all directions, creating 5-pixel wide (250 nm) puncta objects ( Fig   233   2B1-2). Since we used local intensity-maxima to detect putative synapses, the threshold value of 234 45% of total intensity distribution proved best for our samples as it dropped several background 235 local maxima thereby providing better signal to noise ratios. Next, we segmented the dendritic 236 image by thresholding at mean pixel intensity to detect the dendrites (Fig 2B3), and dilated the 237 dendritic mask by 2 pixels to capture all the colocalized pre-and post-synaptic puncta lying on 238 and in close proximity to the dendrites.

239
Step 2 was to estimate the total number of colocalized pre-and post-synaptic puncta on 240 the dendrites. To accomplish this, we performed a binary AND operation between the three 241 binary masks generated for the dendritic, pre-and post-synaptic puncta channels (Fig 2B1-3) and 242 counted the number (Fig 2E).

243
In step 3, we estimated the detection noise caused by random chance (or false positive) 244 puncta-colocalizations on the dendrites. We repeated the AND operation of step2 after disrupting 245 the spatial correlation between the pre-and postsynaptic puncta objects. This was achieved with 246 two independent methods: (a) randomizing locations of the pre-and post-synaptic puncta within 247 the respective images ( Fig 2C1-2); (b) shifting the original pre-and post-synaptic masks relative 248 to each other, i.e. spatial cross-correlation ( Fig 2D1-2). To establish the reliability of our noise 249 estimation procedure, the agreement between these noise estimates was assessed (see Fig 8). was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made  257 To evaluate the performance of our algorithm, we used it to detect signal and noise 258 components in simulated binary images of dendrites, pre-and post-synaptic puncta channels (Fig   259   4). To mimic a situation of our experimental data, we created images 1024 × 1024 pixels in size, 260 with 800 colocalized puncta on the dendrites (representing real synapses i.e. the signal) and 261 added pre-and post-synaptic noise components. The objects in each puncta channel were 262 depicted as 250 nm-wide circles (Fig 4A-B) and the dendrites were depicted as 5-pixel-wide 263 lines ( Fig 4C). The puncta objects representing real synapses (Fig 4D), were spatially correlated 264 along the dendrites and were separated by a distance less than 250 nm. We created a series of 265 images containing the original signal with different noise levels. This was done by adding 266 different amounts of 250 nm wide circular objects, at spatially random locations in the simulated 267 pre-and post-synaptic images. Using this procedure we obtained simulated image series with 268 known signal and noise components, i.e. in contrast to the measured data, we had access to the 269 gold standard. We then applied our algorithm to establish its performance using the known 270 signal-to-noise ratio.  273 We quantified both excitatory and inhibitory synaptic densities across four developmental progressive complexity were tested using the package ANOVA test until adding terms produced 290 no significant improvement in the model. We started with two random factors (electrodes nested 291 within cultures) and a mean value and ended with a cubic fit (the cubic fit was adopted because 292 the plots showed a clear increase in slope before decreasing again, which requires a third order 293 term to be modeled). We kept only slope as a random factor in the fit because the biology of the 294 models requires a zero intercept and adding higher order terms makes the model too flexible and 295 hard to fit and interpret. Significance was set at 0.01 to account for testing of multiple quantities 296 (4). To confirm the results, a nested bootstrap approach (random sampling of cultures with 297 replacement, followed by random sampling of electrodes with replacement within those cultures 298 the electrodes were selected once for all the DIV values to preserve the longitudinal sampling 299 . CC-BY 4.0 International license available under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

304
In this section we first present the typical network activity patterns that spontaneously 305 emerge during the first three weeks of culture maturation. Next we quantify synaptogenesis 306 during this period and validate our novel synapse detection algorithm. We especially determine 307 robustness of our detection procedure with respect to the presence of noise and detector threshold 308 selection. was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which this version posted May 18, 2023. ; https://doi.org/10.1101/2023.05.18.541345 doi: bioRxiv preprint was measurable spiking and bursting activity by 8DIV which increased significantly by 14 DIV, 323 followed by a reduction towards 20 DIV. The corresponding intracellular measurements confirm 324 the activity patterns observed in the MEA (Fig 5B). The bottom traces in Figure 5B show the  Simulated Images 343 We produced simulated images with known signal-to-noise ratios (SNR) (Methods, Fig 4) 344 to determine the performance of our synapse detection algorithm under realistic, noisy 345 . CC-BY 4.0 International license available under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made conditions. Figure 6 depicts the relationship between the real number of synapses, noise levels 346 and their estimates. We found that our detector reports a consistent estimate of number of 347 puncta-colocalizations (green solid line, Fig 6) for SNR values between 1.5 and 6, which is the 348 relevant range for our measured data set. Thus, in this range, the sensitivity of our detector is 349 rather stable, i.e. 90-92% (the ratio between the detected and true synaptic structures shown by 350 the green lines in Fig 6). In contrast, the error in the synaptic density estimation without noise 351 correction, due to false positive detections (Type I error), increases significantly (10-40%) with 352 decreasing SNR values in this range (purple line in Fig 6).

354
Using the confocal images, we quantified excitatory and inhibitory synaptogenesis in two series 355 of cultures that were plated at different cell densities: 650 cells/mm 2 and 850 cells/mm 2 (Fig 7A).

356
In both series, we found that excitatory synaptic density initially increased and reached a peak pruning. On the other hand, density of inhibitory synapses showed an increasing trend as the 361 cultures matured. Interestingly, the culture that was plated at sparser cell density showed onset 362 of synaptic pruning at a later stage in the development as compared to the culture plated at a 363 higher cell density (Fig 7B-C).

365
In contrast to the simulated images, signal and noise levels arising from random 366 colocalization are unknown in the experimental data. To compensate for this lack of knowledge 367 (in part), we estimated and compared the noise levels in the detector output with two independent 368 . CC-BY 4.0 International license available under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which this version posted May 18, 2023. ; https://doi.org/10.1101/2023.05.18.541345 doi: bioRxiv preprint methods: randomizing the locations of the binary puncta-objects in the masks, and spatially 369 shifting the original masks, i.e. spatial cross-correlation analysis. We found that these 370 independent estimates were in agreement. A representative example of our noise estimates is 371 shown in Figure 8. Figure    was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

394
In this work, we applied and evaluated a novel method for automated detection and 395 quantification of synapses in confocal images of neuronal networks and applied it to quantify 396 excitatory and inhibitory synaptogenesis in dissociated hippocampal cell cultures. Using both 397 simulated and experimental data sets, we established that the automated procedure provided 398 consistent estimates of synaptic density, and that these estimates were rather independent of 399 noise contamination of the images and detector threshold selection (Figs 6, 8 and 9). In the 400 experimental data, we found that excitatory synaptic density increased and reached a peak 401 around 8-14 DIV, and then declined towards 20 DIV, which might be suggestive of synaptic 402 pruning (Fig 7B,C). On the other hand, we found that the density of inhibitory synapses was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which this version posted May 18, 2023. ;https://doi.org/10.1101https://doi.org/10. /2023 Page | 20 synaptic densities. This is manifested at the functional level in the form of initially low spiking   (Table 1). Towards 20 DIV, however, there is a reduction in excitatory synaptic density 422 along with a parallel increase in inhibitory synaptic density that coincides with a reduction in 423 mean spike and burst rates (Table 1) at the network level. The spontaneous activity at this stage 424 is characterized by regular alternating periods of bursting and quiescence (Fig 5A).  Other groups have used neuron-glia cultures maintained in serum-based medium, which develop 431 dendritic spines [Van Huizen et al. 1985;Ichikawa et al. 1993;Schätzle et al. 2012;Harrill et al. 432 2015]. Furthermore, there are differences in seeded cell densities. In addition, different 433 techniques, i.e. electron microscopy and confocal microscopy were applied to assess 434 synaptogenesis. In spite of the differences in culture preparation and measurement technique, our 435 data suggesting synaptic pruning is in agreement with the electron microscopy results reported 436 by Van Huizen et al. (1985) and Ichikawa et al. (1993). Interestingly, both these studies used 437 . CC-BY 4.0 International license available under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which this version posted May 18, 2023. ;https://doi.org/10.1101https://doi.org/10. /2023 Page | 21 similar cell density as ours (~900 cells/mm 2 ). In contrast, the confocal microscopy results from However, as compared to our cell densities (650 cells/mm 2 and 850 cells/mm 2 ), both these 442 studies employed a much lower cell density: 40 cells/mm 2 and 315 cells/mm 2 respectively. In 443 conclusion, the applied culture density is the common difference between the studies that report 444 pruning and those that do not. These findings are not necessarily contradictory since our 445 experimental and modeling results demonstrate that the timing of the pruning phase in the 446 synaptic development critically depends on the network's cell density (Fig 7). We find that a low 447 cell density is associated with a delayed pruning phase. Thus in the studies that employed low 448 cell densities, e.g. Schätzle et al. (2012) andHarrill et al. (2015), the pruning phase could have 449 occurred outside the window over which the culture was observed. We would also like to point 450 out that we only observed and reported the synaptic densities at four discrete time points within a 451 3-week developmental period. Therefore, any significant changes in development of synaptic 452 densities that could have occurred in between these time points were not captured in this study.

453
One problem of using confocal studies to quantify synaptogenesis is the fact that the 454 staining procedures are not 100% specific for the target synaptic structures. For the synapse 455 detection this results in spurious staining, leading to a significant noise component in the images.

456
At the data acquisition level, uncertainty caused by this noise component can be reduced by 457 staining both pre-and post-synaptic terminals, and by using colocalization of these structures as 458 the synapse detection criterion. However, since both the pre-and post-synaptic terminal staining 459 include spurious results, their combination will still create noise in the form of random chance 460 . CC-BY 4.0 International license available under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which this version posted May 18, 2023. approach was to mitigate, as much as possible, the problems created by the noise components 473 and threshold effects (Figs 6, 8 and 9). We accomplished this by obtaining reliable estimates for 474 the noise component in our images (Fig 8), resulting in a highly sensitive detector (90-92%) over 475 a large range of SNRs (Fig 6). This reliable noise estimate made threshold selection a less critical 476 property (Fig 9), which enabled us to employ a lower intensity threshold value (i.e. below the 477 mean intensity) as compared with the mean pixel intensity that is commonly employed in 478 existing detection procedures (Schmitz et al. 2011;Schätzle et al. 2012;Harrill et al. 2015). Our 479 approach to synapse quantification is to use the spatial correlation of pre-and post-synaptic 480 puncta along a neuronal surface i.e., dendrites in our case. Our current method of performing the 481 AND operation on both the puncta and the surface masks would equally apply for other staining 482 . CC-BY 4.0 International license available under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which this version posted May 18, 2023. ;https://doi.org/10.1101https://doi.org/10. /2023 Page | 23 strategies in other types of cultures. For example, our approach could be extended to analyze 483 images where the masks capture dendritic spines as well.

484
One finding of particular interest is the presence of a pruning phase during the 485 development of the network (Fig 7). Albeit at a different timescale, several human studies 486 (Huttenlocher 1979;Huttenlocher 1986) as well as other primate studies (Bourgeois and Rakic 487 1993;Wolff et al. 1995;Mimura et al. 2003), have also shown evidence for initial

497
. CC-BY 4.0 International license available under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which this version posted May 18, 2023. ;https://doi.org/10.1101https://doi.org/10. /2023 Boyer C, Schikorski T, Stevens CF. Comparison of hippocampal dendritic spines in culture and 503 in brain. J Neurosci 18:5294-300, 1998.  was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which this version posted May 18, 2023. ;https://doi.org/10.1101https://doi.org/10. /2023 Page | 25 Danielson E, Lee SH. SynPAnal: software for rapid quantification of the density and intensity of 520 protein puncta from fluorescence microscopy images of neurons. PLoS One 9:e115298, 2014.    was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made      was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made    was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which this version posted May 18, 2023. ;https://doi.org/10.1101https://doi.org/10. /2023  which 81 multi-channel images were captured, arranged in a 9-by-9 grid. C) An example of one 613 of the 81 multi-channel images, captured from the cover-slip. This is a merged image formed by 614 superimposing four separate channels each capturing the pre-synaptic puncta (magenta), post-615 synaptic puncta (cyan), dendrites (green) and soma (grey). A zoomed in multi-channel image 616 comprising of D) excitatory and E) inhibitory pre-and post-synaptic puncta as well as dendrites 617 labeled with VGluT1/VGAT, PSD-95/Gephryn and MAP-2 respectively.

618
. CC-BY 4.0 International license available under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made  and post-synaptic puncta channels respectively, generated after implementing the following 627 steps: rolling-ball background subtraction, 3×3 median filtering, thresholding at 45% of the total 628 intensity distribution, identifying single-pixel local maxima, enlarging each maxima by 2 pixels 629 . CC-BY 4.0 International license available under a was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made   was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which this version posted May 18, 2023. ;https://doi.org/10.1101https://doi.org/10. /2023