Spontaneous retinal waves generate long-range horizontal connectivity in visual cortex

In the primary visual cortex (V1) of higher mammals, long-range horizontal connections (LHCs) are observed to develop, linking iso-orientation domains of cortical tuning. It is unknown how this feature-specific wiring of circuitry develops before eye opening. Here, we show that LHCs in V1 may originate from spatio-temporally structured feedforward activities generated from spontaneous retinal waves. Using model simulations based on the anatomy and observed activity patterns of the retina, we show that waves propagating in retinal mosaics can initialize the wiring of LHCs by co-activating neurons of similar tuning, whereas equivalent random activities cannot induce such organizations. Simulations showed that emerged LHCs can produce the patterned activities observed in V1, matching topography of the underlying orientation map. We also confirmed that the model can also reproduce orientation-specific microcircuits in salt-and-pepper organizations in rodents. Our results imply that early peripheral activities contribute significantly to cortical development of functional circuits. Highlights Developmental model of long-range horizontal connections (LHCs) in V1 is simulated Spontaneous retinal waves generate feature-specific wiring of LHCs in visual cortex Emerged LHCs induce orientation-matching patterns of spontaneous cortical activity Retinal waves induce orientation-specific microcircuits of visual cortex in rodents Significance statement Long-range horizontal connections (LHCs) in the primary visual cortex (V1) are observed to emerge before the onset of visual experience, selectively connecting iso-domains of orientation maps. However, it is unknown how such tuning-specific wirings develop before eye-opening. Here, we show that LHCs in V1 originate from the tuning-specific activation of cortical neurons by spontaneous retinal waves during early developmental stages. Our simulations of a visual cortex model show that feedforward activities from the retina initialize the spatial organization of activity patterns in V1, which induces visual feature-specific wirings of V1 neurons. Our model also explains the origin of cortical microcircuits observed in rodents, suggesting that the proposed developmental mechanism is applicable universally to circuits of various mammalian species.


Introduction
determines organization of the orientation map in V1, we suggest that spontaneous retinal activity 139 determines the patterns of both orientation map and activity pattern in V1, by generating horizontal 140 wirings that connect iso-domains of the underlying orientation map. In this simple scenario, 141 topographies of spontaneous V1 activity and underlying orientation maps must be correlated. Further, 142 once LHCs develop, silencing the feedforward activity in a retina or LGN 33 cannot eliminate the 143 correlated activity in V1, as observed in ferrets. 144 To validate this model, we removed all feedforward drive from the retina in our model and 145 simulated activities by randomly driving the V1 network with developed LHCs. Following the analysis in 146 a previous study 33 , we selected reference points at arbitrary locations in V1 and computed the Pearson 147 coefficient of correlation for spontaneous activities between the reference and other locations across 148 cortical space (Fig. 5e, left, see also Supplementary Video 2, 3). We observed strong matching 149 between the activity correlation map and underlying orientation map, even though the V1 circuit does 150 not receive inputs from the feedforward pathway. As observed in ferrets 33 , correlation between the 151 activity correlation map and orientation map was significantly higher than in the controls where two 152 maps were randomly rotated (Fig. 5f, two-tailed paired t-test with randomly aligned controls; data 153 maps: n = 100, *p = 1.1×10 -35 , cat model: n = 100, *p = 2.23×10 -308 ; monkey model: n = 100, *p = 154 1.90×10 -240 ). We also repeated this for randomly chosen reference points and confirmed the statistical 155 significance of the correlation (Fig. 5g, Mann-Whitney U-test; data maps: n = 8, p = 0.02, cat model: n = 156 367, *p = 1.17×10 -41 ; monkey model: n = 318, *p = 4.94×10 -47 ). These results suggest that the Development of orientation-specific horizontal connections without a periodic map 161 So far, we have shown that our model provides an explanation for how LHCs develop spontaneously in 162 V1 of higher mammals with columnar orientation maps. Next, we show that our model further explains 163 how orientation-specific LHCs also develop in rodents V1 with salt-and-pepper organizations 35 (Fig.   164   6a). It is notable that the key assumption of our model is that retinal waves co-activate V1 neurons of 165 similar tuning to develop LHCs between them, and that this mechanism works regardless of the spatial 166 organization of orientation preference in V1 (Fig. 6b). 167 To validate our prediction in salt-and-pepper type organizations of V1, we implemented a model 168 V1 circuit with salt-and-pepper organization. Using mouse retinal-mosaic data 36 and the same 169 developmental model for cats and monkeys, we confirmed that cortical neurons of similar orientation tuning tend to fire in correlated patterns. As a result, similar to the V1 model with a periodic orientation 171 map, we found that LHCs with significant orientation-specificity developed (Fig. 6c, Cuzick's test for 172 trend; initial: n = 970, p = 0.15; developed by retinal waves: n = 551, *p < 2.23×10 -308 ; developed by 173 random activity: n = 970, p = 0.09). This result demonstrates that orientation-specific LHCs can also 174 emerge from the correlated activity induced by retinal waves, even in salt-and-pepper type 175 organizations 26 , as observed in layer 2/3 horizontal microcircuits of rodent V1 35 . Notably, observations 176 that orientation-specific microcircuits appear to develop even in dark rearing conditions with no visual 177 experience may also support our model. Our model predicts that these microcircuits can develop in 178 dark rearing conditions, because spontaneous retinal waves can contribute under this condition 37,38 179 ( Fig. 6d and   On the other hand, our model explains the emergence of correlated activity patterns in V1, the 201 topographic matching to orientation map, and how these processes are carried out before the onset of 202 any visual experience. Our result proposes a simple but powerful model of the developmental 203 mechanism underlying the origin of spontaneous activity patterns in V1 and its correlation to orientation 204 tuning maps, to complete the scenario.

205
Observations that support our developmental model were also reported in previous studies. 206 Durack et al. (1996) 44 found that initial clustering of LHCs in ferret V1 coincides with, but does not 207 precede, the development of orientation preference 44 . This implies that the development of LHCs may 208 "reflect", rather than "seed" the structure of orientation maps. In addition, orientation-specific 209 microcircuits of V1 in mice begin appear to emerge after development of retino-cortical projections and 210 orientation tuning of V1 neurons 45 . These results also suggest that, after feedforward pathway has been 211 developed to induce cortical orientation tuning, retinal waves drive the development of orientation-212 specific horizontal connections.

213
The previous study reported that clustered horizontal connections are observed even after  For the simulation of our retina-V1 model, we used ON/OFF-type RGC mosaics data from mammals.

274
Simulations shown in the main results are based on cell body mosaics of the cat 50 ( Fig. 3-5) and 275 mouse 36 (Fig. 6). We also provided simulation results based on monkey receptive field mosaics 30 .

276
Considering that differences in cortical tuning organization, modeling, and simulation schemes differ 277 between higher mammals (cat, monkey) and rodents (mouse); from here, we describe the framework 278 for simulating higher mammalian visual cortex and later describe specifications for the rodent cortex where and denote the cell index. We also measured the density of OFF RGC, and from that, we 283 defined a representative spacing so that an ideal hexagonal RGC lattice with spacing 284 would have the same cell density as the data OFF mosaic. We computed in the same manner. information, we first augmented the data RGC mosaic with surround RGC padding and an AC lattice.

291
Note that added cells were only used to aid wave propagation, and we did not consider them in further 292 simulations.

293
First, to provide a retinal region sufficient for the wave to propagate, we padded the data RGC 294 mosaic with synthesized hexagonal lattices of ON and OFF RGCs, each with spacing and .

295
Note that with this spacing, a synthesized lattice has the same cell density as the data mosaic. We 296 gave the padded lattice a circular boundary of radius 3000 µm (limited by GPU memory) to reduce 297 potential propagation direction bias. with AC input and becomes active for 1 s upon reduction of such inhibitory input. The described 320 connectivity is modeled as follows: We chose the dendritic radius of ON RGC and AC considering the size of the cell dendritic arbors experimentally measured by Akrouh  The response curve of V1 neurons was modeled as a nonlinear sigmoid kernel with parameters 1 383 and 1 . , Here, , 1 ( ) is the response of the ℎ cortical neuron at wave time , where activation of feedforward input here, but horizontal input or direct cortical stimulus is allowed in later simulations.

388
Parameter details are provided in Table 1.
Here, we define the learning threshold �������� as the running average of the sampled 404 responses of a cell during the learning steps. The term represents how fast the threshold changes 405 during the learning steps and the term , the learning rate, denotes how quickly the weight update is 406 done. We assumed that the resource for a single connection is limited by limit .

407
We simulated the development of RGC-V1 feedforward connections using a post-processed 408 retinal wave dataset, 15 epochs in total. At each epoch, we shuffled wave order and iterated over all 409 waves one by one, applying the learning rule. All the parameter details are provided in Table 1.

Initialization of the V1 horizontal network model 412
After the RGC-V1 feedforward development was complete, we froze the feedforward connection 413 weights and simulated development of the V1 horizontal connection network by retinal waves. Initially, 414 we horizontally wired V1 cells with random weights.
Here, 1 represents the synaptic connection weight from the ℎ to ℎ cortical site ( ≠ ), 417 where 1 is drawn from a random uniform distribution (0,1). After setting the random connections, 418 we normalized each V1 cell's outgoing connection weight sum by init 1 and finished the initialization 419 step. Parameter details are provided in Table 1. Then, peak responses of V1 neurons were sampled and used as a response profile for a 427 covariance rule-based weight update of the V1 horizontal network.
Here, we define the learning threshold �������� as the running average of peak responses of a 432 V1 neuron over learning steps. Here, 1 represents how fast the threshold changes during the 433 learning steps and 1 , the learning rate, denotes how quickly the weight update is done. We assumed 434 that resource for a single connection is limited by limit 1 .

435
wave dataset (permuted dataset as control case), 30 epochs in total; at each epoch, we shuffled the 437 wave order and iterated all the waves one by one, applying the learning rule. All the parameter details 438 are provided in Table 1.
Here, the horizontal bar denotes averaging over all spontaneous activity images and 468 denotes the standard deviation of activity over all images at location 33 .
Then, we assessed the group trends using nonparametric Cuzick's test for the trend of Matching between the spontaneous activity correlation map and orientation map 542 To quantify the alignment between the cortical activity correlation pattern and orientation map, for a given reference point , we measured Pearson's correlation coefficient between the orientation 544 similarity map ( ) and activity correlation map ( ) of reference ( denotes pixel locations).
Here, is the number of pixels. To test for statistical significance of the spatial correlation