The cortico-thalamic loop attunes competitive lateral interactions across retinotopic and orientation preference maps

In the early visual system, corticothalamic feedback projections greatly outnumber thalamocortical feedforward projections. Extensive experimental and modeling work has been devoted to the functional impact of the feedforward pathway, but the role of its denser feedback counterpart remains elusive. Here, we propose a novel unifying framework where thalamic recurrent interactions and corticothalamic feedback act in a closed-loop fashion to attune multiple stimulus representations. At each position of the visual field, the loop puts into competition local representations of the stimulus in thalamus and cortex through direct excitation of narrow topologically-aligned portions of the thalamus, accompanied with peri-geniculate nucleus mediated broad inhibition suppressing the topological surround. We built a detailed conductance-based spiking model incorporating retinal input, lateral geniculate nucleus, peri-geniculate nucleus, primary visual cortex, and all the relevant intra-areal and feedback pathways. For the first time we perform comparative analyses between model configurations with completely or locally inactivated cortico-thalamic feedback, as in the experimental preparations. The model mechanistically explains (i) the existence of intra-thalamic surround suppression, (ii) the sensitivity of thalamic neurons to orientation tuning, (iii) the cortex-dependent center-surround opponency in thalamic cells, (iv) the cortical increase of size and orientation selectivity, (v) the cortically enhanced competition between cross-oriented domains within the hypercolumn, and (vi) the selective suppression of cortical functional connectivity. Our results integrate decades of experimental and theoretical research, supporting the hypothesis that cortico-thalamic loop exerts competitive influence between neighboring regions in the thalamus and cortex, complementing the lateral intra-V1 interactions in center-surround contextual modulation.


34
The classic framework to understand visual cortical function has relied dominantly on the feedfor- From the left: Retina is modeled as a layer of phenomenological X retinal ganglion cells (RGC) of the two functional subgroups with center-surround opponency (ON-center, white inner disk, OFF-center, dark inner disk). Each RGC excites (red arrow 1) one LGN cell determining its receptive field. Both ON-and OFF-center LGN cells excite (red arrow 2) PGN cells (blue filled circles) and, in return, PGN cells inhibit (blue arrow 2) both ON-and OFF-center LGN cells. PGN cells also inhibit each other (blue arrows 3). In the primary visual cortex (V1), the afferent connections of excitatory (red circles) and inhibitory (blue circles) cells are formed by sampling connections from a Gabor probability distribution centered at the retinotopic position of the given cortical neuron and overlaid onto the sheets of ON-and OFF-center LGN cells (red cone and arrow 4). Note that, in reality, LGN thalamocortical connections make collaterals into PGN, but for clarity we separated them into 2, 3, and 4. Positive Gabor subfields are overlaid on ON-center and negative on OFF-center sheets. V1 excitatory-to-inhibitory and inhibitory-to-excitatory connections (5) implement a push-pull connectivity (Troyer et al. 1998). The slow recruitment of lateral connections incorporates distance-dependent delays (6). Cortical feedback excitatory connections to the thalamus (red arrow) are formed by sampling connections from Gaussian probability distributions overlaid on PGN (7) and LGN  and temporal (c) frequency, size (d), and orientation (e). Each row details a given integration stage in the early visual system (from top to bottom rows: RGC, LGN, PGN, V1). When available, experimental population averages (empty squares and circles with error bars) were used, otherwise, multiple single-cell recordings (various filled symbols) were used. The original firing rates are reported on the top of the vertical axis, using the same color code. The number of recorded model cells is reported in the lower right of column a (green), and, when available from experimental studies, in the other plots (black). When available, experimental data having multiple stimulus feature variations in the same study were used over data from single feature studies. We resorted to macaque data when data for cat was unavailable (see the stimuli section of Methods for experimental data sources). LGN. In the feedforward-only model configuration (dashed), the LGN was dominated by excitatory conductance. The inhibitory conductance increased beyond 0.48°stimulus radius, corresponding to the PGN increase in excitatory inputs from the LGN. In the full model configuration (solid), the LGN was dominated by inhibitory conductance. (d) Excitatory-to-inhibitory balance in the LGN. In the feedforward-only model configuration (dashed), the LGN conductance ratio is dominated by excitation and becomes balanced only at large sizes. In the full model configuration (solid), it is overall more balanced and dominated by inhibition. the full model, are consistent with in-vivo experimental studies. Murphy and Sillito (1987) recorded 138 LGN X-type cells from cats with either intact or ablated primary cortical areas and found a signifi-139 cant reduction (-35.5%) of the average suppression index between control and ablated cortex con-  To get a mechanistic understanding of these results, we used our model to extract population-157 averaged input conductance tuning curves of LGN (Figure 3c). In the feedforward-only configura-158 tion, the excitatory conductance grew with the stimulus size, while inhibitory conductance started 159 to grow only with larger stimuli, but then increased much faster following the increase in PGN firing. of 1 excitatory-dominated regimes. We showed that, in the feedforward-only model configuration, 166 LGN cells are characterized by an excitatory-dominated regime for small sizes (EICB=0.91±0.08), 167 which shifted towards inhibition at large sizes (EICB=0.62±0.12). This confirms that the inhibitory 168 contribution from PGN required large-sized stimuli to be effective at inhibiting LGN response. In 169 the full model, the corticothalamic feedback raises PGN inhibition, leading to a further shift toward 170 inhibition (small: EICB=0.45±0.05, large: EICB=0.38±0.11).

171
Cortical studies have shown that both excitatory and inhibitory conductances are size tuned in 172 V1 neurons, supporting the view that inhibition-stabilization is the underlying mechanism for such 173 conductance tuning (Ozeki et al., 2009). In contrast, in LGN, our simulations predict that this is not 174 the case. Instead, both excitatory and inhibitory conductances increase with stimulus size, while 175 the size-tuning of the spiking response is mediated through stimulus-size-dependent changes to 176 excitatory vs. inhibitory balance. Our model thus predicts a different mechanism of size tuning 177 generation in LGN in comparison to V1. 178 These results suggest that intra-thalamic connectivity is sufficient per se to foster competition 179 between spatially offset LGN neurons through the ubiquitous mechanism of short-range excitation 180 and longer-range inhibition. Cortical feedback further enhances this competition. However, the 181 cortex is also sensitive to other stimulus features, such as orientation, which may be projected 182 onto the thalamic circuitry, as we will see in the following section.

183
Cortical feedback introduces a thalamic bias for stimulus orientation 184 Although orientation selectivity is considered a hallmark of thalamocortical convergence, weak sen-185 sitivity of cell responses to stimulus orientation has been reported for cat LGN cells (Daniels et  perimental data based on trial-averaged LGN firing rate responses. Similar effects of cortical de-197 afferentation are found for the orientation bias and the ratio of preferred to non-preferred orien-198 tations (Figure 4a). Mean orientation bias significantly decreased (-15.7%, paired t-test, p=0.0007) 199 from 1.27±0.14, in the feedforward-only configuration, compared to 1.13±0.07, in the "full" model.

200
These results are in line with Vidyasagar and Urbas (1982), who tested orientation selectivity in 201 cat LGN cells, but using moving bars, and reported that biases of LGN X-cells changed significantly 202 between the two configurations, with a 14.5% decrease of mean orientation bias from 1.83±0.55 203 to 1.74±0.62 (paired t-test, p<0.001).

204
The orientation tuning of cat LGN cells in the presence of intact V1 has been tested also with

214
In the feedforward-only model configuration, there was no significant response preference to-215 ward any orientation in either LGN groups (Figure 4c, blue lines). However, in the "full" model 216 (green lines), the two groups of thalamic cells responded differently to the orientation protocol. 217 We found statistically marginal change due to cortical feedback in the Pinwheel-LGN group (Welsh 218 t-test, p=0.097). In contrast, a highly significant change in orientation preference emerged (Welsh 219 t-test, p=0.0071) in the ISO-Domain-LGN group (Figure 4c). 220 A mechanistic explanation can be provided by comparing the excitatory and inhibitory conduc-221 tance tuning curves of the two LGN cell groups in the feedforward-only and full model configu-222 rations (Figure 4d). The corticothalamic feedback significantly raised (Welsh t-test, p=0.0001) the 223 excitatory conductance of the ISO-domain-LGN group, preferentially at the orientation matching 224 that of the co-registered cortical domain (by convention, 0 degrees). It also raised the inhibitory con-225 ductance preferentially for cross-oriented stimuli. Similar biases, but of much smaller magnitude, 226 were present in the conductances recorded from the Pinwheel-LGN group. We hypothesize that 227 this residual orientation selectivity in the Pinwheel-LGN group is due to remnant non-uniformity in 228 the orientation preferences represented around a pinwheel. The excitatory-to-inhibitory conduc-  Vidyasagar and Urbas (1982) showed a significant mean reduction of orientation bias ratio (n=94, drifting bars of 15x0.1 deg, moving at 5 deg/s). Our model also exhibited such reduction when measured using drifting gratings. Right: Orientation selectivity indexes measured using drifting gratings by Naito et al. 2007 (only available for intact condition). Our full model matches the OSI measured in cats when measured using an identical stimulus. (b) Setup of the virtual orientation tuning experiment. Two groups of model LGN cells were selected based on the orientation preference of their cortical feedback afferent inputs (1: from a cortical pinwheel, and 2: from a "0 degree" cortical ISO-domain). (c) For the LGN group receiving input exclusively from cortical pinwheels, the trial-averaged firing orientation tuning curves (top) show no significant selectivity changes between the full (green, shaded SEM) and "feedforward-only" model (purple) configurations. For the LGN group receiving selective inputs from 0 degrees-oriented ISO-preference cortical domains, the trial-averaged firing orientation tuning curves (top) showed selectivity in the full model (green), significantly different from the nonselective units recorded in the feedforward-only model configuration (purple). (d) The trial-averaged synaptic conductance tuning curves for both group 1 and 2 (dashed) confirm the absence of selectivity in the feedforward-only model configuration. The conductance tuning curves of the full model show tuned conductances for both groups. (e) Excitatory-to-inhibitory balance. In the feedforward-only model configuration (dashed), the LGN conductance input ratio is balanced for both groups. In the full model configuration (green), only for the Domain-LGN group, global input conductance is inhibitory dominated, becoming more balanced only when integrating the orientation-biased cortical input. area of retinotopically co-registered cortex.

239
Cortical feedback enhances center-surround opponency in LGN cells 240 The experimental studies and our modeling presented in the previous two sections hint at the 241 possibility that the functional influence of the corticofugal pathway onto its thalamic target could 242 depend on the spatial relationship between thalamic and cortical receptive fields (RFs). In fact, the 243 seminal work by Tsumoto et al. (1978) showed that spatial overlap between the RFs of V1 cells LGN-V1 RFs. However, the mechanism underlying such opponency remains unclear. The constraints outlined by anatomical and functional studies of both intra-thalamic and cor-251 ticothalamic connectivity, which we introduced in our model, imply the spatial impact of the disy-252 naptic inhibitory pathway combining V1→PGN→LGN was broader than the direct monosynaptic 253 excitation of V1→LGN connections. We, therefore, hypothesized that a stimulus that engages over-   (Figure 5b). 268 Preferred stimulus size was defined as the grating patch size eliciting maximal response in the 269 given LGN cell (peak of the size tuning-curve) recorded in the intact (control) condition. We then 270 measured the percentage change of the peak response of LGN neurons (n=45) during cortical in-271 activation. The response changes were averaged for three grating patch size ranges: (i) smaller-272 than-preferred stimulus sizes, (ii) preferred stimulus size, and (iii) larger-than-preferred stimulus 273 sizes, When we inactivated the overlapping cortical group (Figure 5c), the recorded LGN responses 274 showed a significant decrease for less-than-preferred sizes (-29.3%, empty bars; Wilcoxon pair 275 test, p=0.001), together with smaller decreases for preferred (-16.7%) and for larger-than-preferred 276 sizes (-9.1%). In contrast, inactivation of the non-overlapping cortical group lead to a significant in-277 crease in response (+26.5%) for larger-than-preferred sizes (Figure 5d; Wilcoxon matched pair test, 278 p=0.002), as well as smaller but positive changes for preferred (+13.2%) and smaller-than-preferred   Figure 5c, filled bar) below control levels for less-than-preferred sizes (Wilcoxon pair test, 284 p=0.003). We found a smaller but statistically significant decrease for preferred and a statistically LGN cells (dashed circle), while in the second condition they were non-overlapping with (b, still in proximity to) the recorded LGN RFs. The distance between overlapping and non-overlapping locations was chosen such that the effective corticothalamic direct excitation (red cone) was flanked by a disynaptic indirect inhibition (blue annulus). The overlapping cortical location was inactivated in a, c, e, while the non-overlapping cortical location was inactivated in b, d, f (black cross over the corresponding red circle). (c,d) Population summary histogram of the mean percentage change in LGN cell responses between intact and locally-inactivated model cortex (empty bars), vs. Jones et al. 2012 data (filled bars). The signs of the changes in the model were opposite for the overlapping and non-overlapping conditions (negative in c and positive in d) and their magnitude decreased (c) or increased (d) with the grating stimulus radius (abscissa in degrees), in agreement with the experimental data. (e) Excitatory-to-Inhibitory conductance balance (EICB) is reduced for the overlapping inactivation (full model in green, local cortical inactivation in purple). (Inset) Trial-averaged mean excitatory (red) and inhibitory (blue) synaptic conductance tuning curves of model LGN cells. When the overlapping cortical location is inactivated, only the excitatory conductance is reduced appreciably (dashed), whereas no or minimal change is observed in the inhibitory conductance. (f) The EICB is increased only at large sizes during non-overlapping inactivation. (Inset) When a non-overlapping cortical site is inactivated, the inhibitory conductance is lower (dashed) while the excitatory conductance profile is unchanged. sizes (Wilcoxon pair test, p=0.003), together with a smaller increase for preferred and for smaller-

292
We could dissect our model to gain a mechanistic understanding of the experimental results for 293 the overlapping and non-overlapping configurations. We hypothesized that the connectivity con-  (Figure 3). 313 The Cortical viewpoint 314 In the previous sections, we have shown how cortical activity exerts influence over the thalamus 315 through the cortico-thalamic feedback connectivity, and hence recurrently reshapes its own input, 316 and consequently its activity. We will now explore the hypothesis that the thalamo-cortical loop 317 allows the cortex to up-or down-regulate its own functional selectivity in a stimulus-dependent 318 fashion.

319
The loop enhances cortical selectivity to stimulus size and orientation 320 A series of electrophysiology studies showed that the activity of LGN neurons that further project to 321 V1 is tuned to the size of the stimulus with larger than preferred stimuli suppressing the response  Responses were more suppressed for larger-than-preferred sizes (>1.5 deg) in the full (green) compared to the feedforward-only (purple) model configurations. (b) Normalized trial-averaged orientation tuning curves for cortical cells. Responses were lower for stimuli orthogonal to the preferred orientation in the full model (green) vs the feedforward-only (purple) model configurations. (c) Excitatory-to-inhibitory size conductance ratio. The closing of loop reduced the relative impact of inhibition, with E/I ratio values closer to 0.5 for the full (green) vs feedforward-only (purple) model configurations. Inset shows the corresponding trial-averaged mean excitatory (red) and inhibitory (blue) synaptic conductance size tuning curves of model V1 cells in the full (solid) vs feedforward-only (dashed) configurations. (d) Excitatory-to-inhibitory orientation conductance ratio. Although the presence of the closed loop altered the response for orthogonal stimuli, there was only a minor change in excitatory/inhibitory balance. In the inset, trial-averaged mean excitatory (red) and inhibitory (blue) synaptic conductance orientation tuning curves of model V1 cells.
orientations cross-oriented with the preferred one.

338
In order to compare cortical responses in the two configurations, we measured the "feedforward-339 only" (purple) and "full" (green) trial-averaged normalized firing rates (Figure 6ab). For the size tun-340 ing protocol (Figure 6a), in the "feedforward-only" configuration the suppression index was low (SI: 341 0.11±0.13), while in the "full" model, the presence of the cortico-thalamo-cortical feedback path-342 way resulted in a significantly higher values of the suppression index (SI: 0.32±0.12, Welch t-test, 343 p=0.0052). For the orientation tuning protocol (Figure 6b), in the feed-forward-only configuration, 344 the suppression for non-preferred stimuli was significantly lower than in the full model (-8.8%,

345
Welch t-test p=0.0071). 346 We explored the underlying relationship between excitatory and inhibitory synaptic conduc-  359 Importantly, the reduction of suppression measured for "feedforward-only" vs "full" configura-  In summary, our model shows that the cortico-thalamic loop enhances suppression for non-369 preferred (both in terms of orientation preference and size) stimuli, facilitating (i) competition be-  374 The loop enhances competition within the hypercolumn but not in the surround 375 So far we have studied the cortical response to stimulus size and orientation independently. We 376 thus next proceeded to study whether the cortico-thalamic loop engages the short vs. long-range 377 cortical circuits differentially and whether these lateral interactions depend on the orientation pref-378 erence of the neurons. To do so, we extended our analysis to also include (i) cells spatially offset 379 from the cortical location retinotopically aligned with the stimulus center, and (ii) cells with cross-380 oriented preference to that of the stimulus. In the previous section, we found that the cortico-381 thalamic feedback facilitates competition between the representation of nearby stimuli through 382 enhanced surround suppression. Here we found that this competition is restricted to the hyper-383 column aligned with the stimulus center and the strength of these competing influences is inde-384 pendent of the orientation preference of the modulated neurons. 385 We analyzed two topologically defined domains, "Center" and "Surround" (Figure 7a) neurons were partitioned into ISO-group containing cells with orientation preference matching 392 the stimulus orientation, and the CROSS-group containing cells with orthogonal preference. By 393 convention, the grating orientation was set at the preferred orientation 0 degrees. In order to 394 compare the responses of all groups, we reported the normalized firing rate in the "full" (green) 395 and "feedforward-only" (purple) model. 396 In the center, the ISO group mean responses were more suppressed for large sizes (>1.5 deg) 397 in the "full" model (green) compared to the "feedforward-only" configuration (purple), as already 398 shown in Figure 6. Interestingly, both neurons in the ISO and CROSS groups exhibited a similar 399 increase in surround suppression in the "full" relative to the "feedforward-only" condition, in line 400 with the fact that the additional surround suppression in the "full" condition is mediated by the LGN  (Figure 7c), but this change was absent in neurons located in the Surround-domain (Figure 7d). 409 These results showed that the cortico-thalamic loop reinforces locally the competition between  Figure 7. (c) Cortical DII peak amplitude (solid curve) of ISO-oriented cells in the Center, triggered by spikes in the Surround. In the "feedforward-only" configuration (purple), the DII is reaching its peak maximum at 0.5 deg of stimulus radius. For larger stimuli, the DII saturates. In the "full" model (green), the DII peak amplitude is overall lower, with a corresponding peak at 0.5 deg stimulus radius, but also remains significantly lower than in the "feedforward" configuration for larger stimuli. The DII tuning roughly followed the firing tuning (dashed). "full"-loop model configuration relative to the "feed-forward" model configuration (Figure 8 cf). In 451 both ISO-group conditions (Figure 8 c,d, green lines), the DII tuning curve for the "full"-model con-

457
In summary, the cortico-thalamic loop maintained, relative to the "feed-forward" configuration, 458 the facilitatory interactions between neural populations in the Center and Surround, whose func-459 tional preference matched the orientation of the stimulus, but only for stimulus sizes that were 460 confined to CRF (as defined by the peak of the size tuning curve). But, the facilitating interactions 461 were suppressed once stimuli started invading the surround. In contrast, for neurons whose pref-462 erence did not match the stimulus orientation, a suppression (relative to the "feed-forward" con- The early sensory pathways integrate numerous reverberating loops between the cerebral cortex 477 and thalamus (Jones, 1985). While the feedforward visual sensory pathway from LGN to V1 has 478 been studied extensively, the understanding of corticothalamic feedback is limited, due to the ex-  (Figure 3a), while also strengthening locally functional contrast in 503 the cortically encoded feature maps (Figure 4cd, Figure 5cd). The model also describes the mecha-504 nism by which the cortico-thalamo-cortical loop supports increased stimulus feature selectivity by 505 increasing the relative difference between levels of activation of local neural populations represent-506 ing orthogonal stimulus orientations (Figure 6 and Figure 8). Particularly, the model indicates that 507 the cortico-thalamic loop may enhance competition between feature-selective domains coexisting 508 within the same hypercolumn but not beyond (Figure 7). The model suggests also that the loop 509 selectively suppresses cooperative facilitation between ISO-preference cortical domains (Figure 8).  (Born et al., 2021). Interestingly, in this work, a wide inhibitory feedback coupling kernel was used 538 to reproduce feedback-enhanced surround suppression and sharpening of LGN receptive fields. Here, we consider that the selectivity required to categorize visual stimuli and the flexibility to cope 545 with the combinatorial explosion of possible configurations of visual features need to be faced in 546 the early stages of visual processing (Barlow and Levick, 1969). Both goals can be achieved in a 547 system that supports selectivity emergence through recurrent competition (Edelman, 1993), and 548 flexibility through synergistic interactions (Kauffman et al., 1995). We propose that the corticotha-

571
Here we provide the description of the principles used to construct the simulated cortical and 572 thalamic regions, the various connection pathways within and between these regions, and the 573 cellular and synaptic properties. We also detail the experimental protocols, the procedures for the 574 collection and analysis of data, and the software stack used to develop and simulate the model.

575
The full code listing and parameters are available on the project page on GitHub (https://github.
with being the amplitude of response, and the radius of the RF component. These parameters  The transmission delays between processing layers we adopted were taken from several converg- m, corresponding to 0.5 degrees considering V1 magnification factor around the area centralis.

786
The relative spatial position of cortical and thalamic receptive fields were also taken into account,  Iterative procedure to find a single parameter set 831 We systematically presented the stimuli described above to our model and heuristically found a 832 single parameter set that fitted qualitatively and quantitatively (see next section for the analysis).

833
To assess model stability, we then searched the parameter space around two crucial parameters: where and denote the conductance and equilibrium potential for each of the two synaptic 877 types (see above Table 1). A biophysical forward-modelling scheme for the LFP was adapted from we averaged across all extracted chunks. This STA was then analyzed to extract its features. We 894 identified the presence of a negative peak as absolute STA minimum ("trough") and characterized 895 its lag, amplitude and duration. The lag was measured as (ms) difference between reference time 896 and negative absolute minimum time. The amplitude was measured as the absolute minimum 897 value (mV). And the duration was measured as the interval between the first two neighboring local 898 maxima next to the absolute minimum (ms). We repeated this procedure for each stimulus size  In this case, the standard error of the mean was also computed.