Local and thalamic origins of ongoing and sensory evoked cortical correlations

Thalamic inputs of layer 4 (L4) cells in sensory cortices are outnumbered by local connections. Thus, it was suggested that robust sensory response in L4 emerges due to synchronized thalamic activity. In order to investigate the role of both inputs in generation of cortical synchronization, we isolated the thalamic excitatory inputs of cortical cells by optogenetically silencing cortical firing. In anesthetized mice, we measured the correlation between isolated thalamic synaptic inputs of simultaneously patched nearby L4 cells of the barrel cortex. In contrast to correlated activity of excitatory synaptic inputs in the intact cortex, isolated thalamic inputs exhibit lower variability and asynchronous spontaneous and sensory evoked inputs. These results were further supported in awake mice when we recorded the excitatory inputs of individual cortical cells simultaneously with the local field potential (LFP) in a nearby site. Our results therefore indicate that cortical synchronization emerges by intracortical coupling.

2 Thalamic inputs of layer 4 (L4) cells in sensory cortices are outnumbered by local connections. Thus, it was suggested that robust sensory response in L4 emerges due to synchronized thalamic activity. In order to investigate the role of both inputs in generation of cortical synchronization, we isolated the thalamic excitatory inputs of cortical cells by optogenetically silencing cortical firing. In anesthetized mice, we measured the correlation between isolated thalamic synaptic inputs of simultaneously patched nearby L4 cells of the barrel cortex. In contrast to correlated activity of excitatory synaptic inputs in the intact cortex, isolated thalamic inputs exhibit lower variability and asynchronous spontaneous and sensory evoked inputs. These results were further supported in awake mice when we recorded the excitatory inputs of individual cortical cells simultaneously with the local field potential (LFP) in a nearby site. Our results therefore indicate that cortical synchronization emerges by intracortical coupling.
Thus, a cortical mechanism that actively decorrelates synaptic inputs could improve coding 14,15 . Since spiking mechanisms of cortical cells are thought to be highly reliable 16,17 , noise correlations in spiking are likely to reflect correlated membrane potential fluctuations. Indeed, ongoing and sensory evoked synaptic activities in nearby cortical cells are correlated both in time and magnitude 7,6,[18][19][20] . In primary sensory cortices, layer 4 (L4) cells receive the majority of their synaptic inputs from neighboring cortical cells 21,22 . However, they are also strongly driven by feedforward thalamic inputs 23,24 . Therefore, correlated activities between cells in L4 could either arise from common cortical noise or inherited directly from shared thalamic inputs.
In support of the first view, several studies reported that trial to trial variability of sensory evoked cortical response strongly depends on the instantaneous state of cortical activity at the time of stimulation [25][26][27][28]9,5 . It was also shown that both ongoing and evoked activities can be modulated by the animal's behavior and neuromedulators 18,19,[29][30][31] . State dependent modulation of noise correlation was revealed both when using paired intracellular recordings 18 , or when the membrane potential was simultaneously recorded with nearby local field potential (LFP) 32,33 .
Furthermore, the cortex shows slow ongoing oscillations in membrane potential after isolation from adjacent tissue 34 and when thalamus was pharmacologically inactivated 35 . Another study showed that a large component of the covarying response in the thalamus and cortex of the somatosensory system is independent of stimulus properties 36 . Taken together, these studies strongly suggest that noise correlation results from variation in cortical activity.
Alternatively, L4 variability could be inherited directly from thalamic inputs.
In line with this view, it was shown that silencing cortical firing had a negligible effect on the variability of membrane potential response of L4 cells to repeated visual stimuli 4 . Additionally, Bruno and Sakmann 37 proposed that the convergence of inputs from a large number of synchronous thalamic cells strongly drive L4 cells, obviating the need for cortical mechanisms such as recurrent cortical amplification to explain noise correlations.
In this study, we optogenetically silenced the cortex [38][39][40] while simultaneously performing whole cell and LFP recordings in awake mice and dual intracellular recordings in anesthetized mice. This enabled us to study the contribution of thalamic and cortical excitatory synaptic inputs to the subthreshold correlated ongoing and sensory evoked activities in the barrel cortex. Our experiments show that cortical synchrony is not inherited from thalamic inputs but rather depends on recurrent cortical activity.

Barrel cortex amplifies thalamic inputs
Layer 4 (L4) cells in sensory cortices are strongly driven by feedforward thalamic inputs 23,24 . Yet the role of these inputs in the generation of correlated cortical activity was never directly tested. In order to determine the contribution of thalamic inputs to subthreshold correlation between L4 cells, we used an optogenetic approach to silence cortical firing while recording isolated thalamic synaptic inputs of cortical cells. To this end we employed Gad/PV Cre transgenic mice crossed with a ChR2 reporter strain ( Fig. 1a and see Methods). Surface illumination of the somatosensory cortex (S1) with a blue LED (470 nm, ~7mW, LED-ON condition) activated GABAergic cells (Fig. 1b [38][39][40][41] . The latter also receives remote inputs from higher cortical areas [41][42][43][44][45] . This allowed us to estimate the amplification of these inputs by recurrent cortical circuits. For each cell we averaged the whisker evoked excitatory synaptic current (EPSC), both during intact cortical activity and when the cortex was silenced (Fig. 1d).
Thalamic contribution in response to principal whisker (PW) stimulation varied considerably in individually recorded L4 and L5 cells (Figs. 1d and 1e). The mean relative thalamic contribution was larger in L4 than in L5 cells (0.46±0.06 and 0.19±0.07, respectively, p=0.0198 Mann Whitney test, see the depth profile in Figure   1e and the right traces in Figure 1d), probably reflecting the greater innervation of L4 compared to L5 by thalamic fibers 41 . Importantly, the relative contribution of thalamic inputs were indistinguishable when we stimulated either the PW or the adjacent whisker (AW, Fig. 1f p=0.55, n=18, Wilcoxon signed rank test). Hence, thalamic contribution to total response is unrelated to the optimality of the stimulus, similar to findings in the visual and auditory cortices 39,40 .
We next verified that our manipulation allowed us to correctly isolate thalamic inputs. We first ruled out the possibility that the reduction in synaptic response was caused by shunting inhibition. Indeed, in contrast to a shunting effect, in some cells a prominent reduction in the response was recorded while no change in input resistance was measured (Fig. 1g, left example), whereas in others the response was unaffected although input resistance was reduced (Fig. 1g, right example). No significant correlation was found between the thalamic fraction and the measured change in input resistance (Fig. 1h, population data. A small trend exists, but it cannot explain the 5 reduction of the response by shunting, as it shows minimal attenuation for cells in which input resistance was clearly reduced). In addition, by recording thalamic single units we also ruled out the possibility that our manipulation altered the firing of VPM cells due to cortico-thalamic feedback connections (Figs. 1h). Finally, illumination of the cortex 100 ms before whisker stimulation had no effect on the evoked currents ( Supplementary Fig. 1), excluding the possibility that the reduction in EPSC is due to slow extrasynaptic activation of GABA(B) receptors 46,47 . These results indicate that the isolated thalamic synaptic inputs were not affected by cortical inactivation.
In order to study the contribution of thalamic inputs to the correlations between individual cells in thalamic recipient cortical layers, we performed simultaneous in-vivo whole-cell recordings from pairs of nearby excitatory neurons in L4 (Euclidean distance <200µm; Fig. 2a) of anesthetized mice (some of the cells presented in Fig. 1e were recorded as pairs). We analyzed only pairs of cells for which both cells received direct thalamic inputs, as evident from the reduction but not full loss of their response to whisker stimulation (Fig. 2b). We found that thalamic inputs could be substantially different for simultaneously recorded nearby cells (example in Fig. 2b). This was quantified by calculating the similarity index (SI) of To reveal the contributions of thalamic inputs to cortical synchronized ongoing activity in L4, we compared the correlations between the excitatory synaptic currents in each pair when cortical firing was intact (LED OFF) to that calculated when cortical firing was silenced (LED ON). Whereas excitatory currents were highly synchronized in the intact cortex ( Fig. 3b and Supplementary Fig. 2

Trial to trial correlation between layer 4 cells of whisker evoked activity is not determined by thalamic inputs
We next examined the cortical and thalamic contributions to trial to trial variability of whisker evoked EPSCs of L4 cells. In the visual system, inactivation of cortical firing had a negligible effect on trial to trial membrane potential variability 4 , suggesting that cortical variability is dominated by thalamic inputs. In contrast, we found that in the barrel cortex, optogenetic silencing of local firing profoundly reduced the variability of the whisker evoked EPSCs (Figure 4a-c). This was evident both in the standard deviation and in the coefficient of variation (CV). On average the standard deviation of peak EPSPs was reduced by 62%±22% following cortical silencing (Fig. 4b, p=0.0003, n=17, z=-3.6, Wilcoxon signed rank test). This trend remained even after normalization by the mean (Fig. 4c, reduction of 14%±29% in CV, p=0.029, n=17, z=-2.2, Wilcoxon signed rank test).
The larger trial to trial variability of intact cortex compared to inactivated cortex suggests that it is strongly influenced by recurrent cortical circuits. Therefore, we examined the cortical and thalamic origins of the sensory evoked excitatory trial to trial correlation (TTC EE , noise correlation) between pairs of cells that received direct thalamic inputs. Figure  to ongoing activity, we can conclude that the TTC EE of sensory response between cortical cells is not determined by direct thalamic inputs but rather depends on recurrent cortical activity.

Ongoing and sensory evoked correlated cortical activities in awake animals do not emerge from thalamic inputs
Both ongoing and evoked activities in cortical cells are regulated by animal behavior 18,19,29,30,48 . As so, we wished to confirm our result in awake mice. Naïve animals were head-fixed, and were not trained to perform a task. Similar to our Whole cell recordings in awake animals were made simultaneously with local field potential (LFP) recordings using an additional glass pipette that was placed in proximity to the recorded cell in L4 (Fig. 6a, <200μm). Inactivation of cortical firing reduced the amplitude of the averaged LFP response to whisker stimulation (Fig. 6b, two examples of simultaneous cell-LFP recordings). Population analysis showed that thalamic contribution varied also for LFP recordings (Fig. 6c). On average, the contribution of thalamic inputs to the two signals was similar (Fig. 6c,  To reveal the contribution of shared thalamic inputs to cortica1 synchrony during ongoing activity in awake mice, we calculated the correlation between membrane potential and the LFP when cortical firing was intact and when firing was optogenetically silenced. Importantly, we found that the magnitude of the correlation coefficient (CC) between the LFP signal and the electrical activity of the recorded cells was independent of the recording mode method (Fig. 6d, cc=-0.45 at current clamp and cc=0.47 at voltage clamp). Hence, voltage-clamp recordings capture the 9 functional correlations that exist between subthreshold activity of individual cells and the LFP signal. We interleaved trials in which cortical firing was intact with trials in which cortical firing was silenced and calculated the correlation between EPSCs of L4 cells and the LFP signals during ongoing activity. The example paired recording in Figure 6e shows that the thalamic contribution for both signals is nearly one, as evident from the negligible change in the average responses of both signals to whisker stimulation when the cortex was illuminated (Fig. 6e, inserts). Yet, optogenetic silencing of cortical firing during ongoing activity drastically reduced the correlation between the excitatory current of the recorded cell and the nearby LFP signal (from 0.68 to 0.13). Silencing cortical firing in 8 similar recordings reduced the correlation between the LFP and the excitatory current during ongoing activity from 0.31±0.08 to 0.03±0.03 ( Fig. 6f; n=8, p=0.01, z=-2.8, Wilcoxon signed rank test). We can therefore conclude that cortical synchrony during ongoing activity in L4 of the barrel cortex in awake mice is not driven by thalamic inputs.
Next, we examined the effect of cortical silencing on the trial to trial correlation between intracellular excitatory current and a nearby LFP signal in response to vibrissa stimulation (TTC EL ). Similar to anesthetized mice, cortical silencing reduced the response variability of the individual cells and of the LFP (Supplementary Fig. 7), suggesting that noise correlations are affected by recurrent cortical circuits. Indeed, the example in Fig. 6g shows that the response to vibrissa stimulation varied considerably from trial to trial for both signals. Sorting the cellular responses from the smallest to the largest, while plotting the corresponding LFP signal, reveals a clear correlation between the two signals (r 2 =0.85, p=0.00001).
Silencing cortical firing by light reduced the TTC EL (Fig. 6h to r 2 =0.6 p=0.005). On average, no significant change in TTC EL was found between the two conditions ( Fig.   6i, TTC EL =0.47±0.1 and 0.41±0.1 for LED OFF and LED ON respectively, p=0.57, z=-0.56, n=7 Wilcoxon signed rank test). As in anesthetized animals, the TTC EL in the intact cortex was not correlated to that found during cortical silencing (r 2 =10 -5 , p=0.99). Taken together, noise correlation between single cell and the population LFP response in the intact cortex cannot be inferred from the measured correlations of the isolated thalamic inputs, implying that cortical recurrent connections determine the degree of synchronization in L4 of awake mice.

Discussion
In this work, we investigated the role of thalamic inputs in shaping the synaptic correlations between neighboring cells in thalamic recipient cortical layers during ongoing and sensory evoked activities. To address this question, we optogenetically silenced cortical firing in anesthetized and awake mice in order to isolate the thalamic excitatory inputs of intracellularly recorded cells in L4 and L5.
Specifically, in awake mice we examined the effect of silencing cortical firing on the correlation between excitatory inputs of individually recorded cortical cells and a nearby LFP signal whereas in anesthetized mice we simultaneously recorded the excitatory inputs of nearby pairs of neurons. Our results show that synchronized activity during ongoing activity emerges from intracortical inputs, rather than being driven by direct thalamic inputs. Trial to trial sensory evoked correlation ('noise correlation') in response to vibrissa stimulation during intact cortical firing is poorly related to the noise correlation in the thalamic inputs, indicating that it is also a product of intracortical recurrent activity.

The contribution of thalamic inputs to sensory response of cortical cells
To isolate the thalamic input of cortical cells we illuminated the barrel cortex of transgentic mice expressing ChR2 in GAD+ cells while recording excitatory currents. A similar approach was used in recent studies of the auditory and visual cortices of anesthetized mice [38][39][40] where ChR2 was expressed in PV+ cells. Due to the high level of expression in these transgenic mice, it is reasonable to assume that neurons that did not exhibit direct activation by light are excitatory cells. In anesthetized mice we found that thalamic inputs contributed on average about 46% of the total excitatory input of layer 4 cells, which is slightly higher than previously reported in the visual cortex (~30%) 38 and roughly the same as in the auditory cortex of mice (41%) 40 . Our estimate is slightly lower than in the rat auditory cortex where cortical firing was pharmacologically silenced (61%) 49 Fig. 1e), but for most cells amplification was prominent. It is possible that this discrepancy reflects differences across species (mice in our study and rats in ref#37).
To the best of our knowledge, our study is the first that compared the amplification of thalamic inputs by recurrent cortical circuits across anesthetized and awake mice. In both conditions, the contribution of thalamic input to the total excitatory input varied with the recording depth (Figs. 1e and 5d). Higher contribution of thalamic input was found in cells that were recorded in L4 compared to L5 cells.
The higher contribution of thalamic inputs in L4 is expected from the dense innervation of L4 by thalamic inputs 41,44,52 . Importantly, the contribution of thalamic input in L4 and L5 of anesthetized mice was very similar to the contribution of these inputs to the same layers in awake mice (Fig 5e).
Similar to previous studies of the visual and auditory cortices [38][39][40] , cortical silencing showed that the contribution of thalamic inputs was invariant to the optimality of stimulation. We demonstrated it comparing the contribution of thalamic inputs when independently stimulating the PW and one of the AWs within the same cells (Fig 1f). This suggests that the local circuitry of a cortical cell amplifies its thalamic inputs in a particular manner for each cell, regardless of the feature that activates this cell.
Recurrent cortical activity engaged quite rapidly, roughly at the same time of the onset of thalamic input, as evident when the average response to whisker stimulation in intact cortex was compared to the time course of thalamic input alone under cortical silencing (Fig. 1d, 2b, 5c and 6b). Naively, one would expect that recurrent inputs will be delayed by a few milliseconds with respect to the onset of thalamic input, as it involves at least one additional synapse. However, this does not necessarily need to be the case. Assuming that we sampled the cortical population randomly, the thalamic input would arrive to some cells relatively early while to others it would arrive later, The earliest firing in L4 can be as short as 5 ms following whisker stimulation 53 , therefore, local cortical cells should provide input to other cells roughly at the same time or even before the onset of thalamic input of cells that are not the 'primer' cells. Indeed, the latency of the response under cortical silencing in some of our recorded cells was clearly delayed by a couple of milliseconds relative to the onset of the response when the cortex was intact (Fig. 6b, cell #W8). Hence, recurrent cortical activity engages rapidly in the somatosensory cortex. Rapid amplification of thalamic inputs is evident also in the studies of the visual and auditory cortices 38,40 .

Origins of cortical synchrony during ongoing and sensory evoked activities
The suggesting that cortical synchrony emerges due to intracortical inputs [25][26][27][28]9,5 . In contrast to the findings of Sadagopan and Ferster 4 , our experiments show that the cortex adds substantial variance to that which originates from the thalamic inputs, as the standard deviation of the response was significantly larger during intact cortical firing compared to that measured when the cortex was silenced (Fig. 4). Direct measurements of trial to trial correlation indicate that they are not determined by thalamic inputs. In awake mice, the correlation between excitatory inputs of L4 cells and nearby LFP signal, when the cortex was intact, was highly variable across the population. This was observed both for ongoing activity (Fig. 6f) and for the evoked response (Fig. 6i). A large range of correlations was also measured between the excitatory currents in anesthetized mice (Figs. 3c and 4e). In other studies of the barrel cortex in awake mice 18,19 , nearby cells, recorded in layer 2/3, exhibited a much narrower range of correlation during ongoing activity and on average it was higher than what we report in this study of L4. This discrepancy probably reflects laminar differences in the strength of correlations. Indeed, the correlated variability of extracellularly recorded neurons in upper cortical layers of the visual cortex of awake monkeys was found to be significantly higher than between cells located in the granular layer (i.e., layer 4) 55 . 13 The wide range of correlation strengths between individually recorded cells and the nearby LFP in awake animals, as well as between the paired intracellular recordings in anesthetized animals, is reminiscent of the large diversity of network coupling that was recently reported in the visual cortex using extracellular recordings 56  In conclusion, we found that synaptic correlations of nearby cortical cells in L4 and L5 during ongoing and sensory evoked activities are poorly related to their thalamic excitatory inputs. Moreover, the contribution of thalamic inputs varies considerably across the population. The functional role of the asynchronous nature of thalamic inputs and diversity in thalamic contribution in L4 is unclear. An intriguing possibility is that such connectivity may smooth the population response curve to a wide range of stimuli, allowing better encoding of sensory inputs.

Animals:
All procedures involving animals were reviewed and approved by the Weizmann Institutional Animals Care Committee. Animal surgeries and in vivo recordings were performed as previously described 21 . Briefly, recordings were made on young adult mice of either sex (9-16 weeks old) housed up to 5 in a cage with a 12/12h dark/light cycle. Two strains were used, GAD-CRE mice (JAX #010802) and PV_CRE mice (JAX #008069) crossed with a ChR2 reporter strain (JAX # 012569).
Since cortical firing was silenced similarly in both strains, data were pooled from both types.

Anesthetized animal preparation:
For intracellular recording from the barrel cortex, after initial anesthesia with ketamine (90 mg/kg; i.p.) and xylazine (2 mg/kg; i.p.), a tracheotomy was made and the animals were mounted in a stereotaxic device and artificially respirated with a mixture of halothane (0.5-1%) and oxygen-enriched air. The scalp and fascia were removed and a metal headplate was mounted over the left hemisphere using dental cement (Lang dental) and VetBond (3M). A craniotomy (∼1 mm in diameter) was made above the barrel cortex (centered 1.

Awake animal preparation:
21 Animals underwent the implantation of a head bar to allow awake head-fixed recordings as follows: Following initial anesthesia in an induction chamber containing a mix of isoflurane and oxygen enriched air, animals were mounted in a stereotaxic device, and kept deeply anesthetized, monitored by checking for lack of reflexes and pace of breathing. Area of incision was treated with lidocaine and cleaned with iodine and 70% ethanol. The skullcap was exposed and cleaned. The Following a recovery period (4-7 days) animals were anesthetized in an induction chamber containing a mix of isoflurane and oxygen enriched air, animals were then mounted in a stereotaxic device, and kept deeply anesthetized. The silicon glue covering the skull over the barrel cortex was removed and a craniotomy was performed exposing the barrel cortex and leaving the dura intact. The brain was then covered in an agar layer (2% w/v) held in place with silicon glue and the animal was returned to the cage for a recovery period (1-2 hours). The animal was then returned to the set and head-fixed for the electrophysiological recordings.

Cortical patch recordings:
Borosilicate micropipettes were pulled to produce electrodes with a resistance For simultaneous LFP recordings a patch pipet was inserted to a recording depth of 400µM. The signal was band passed at 0.1-300Hz before being digitized at 10 kHz.

Thalamic extracellular recordings:
Extracellular recordings were performed using Juxta electrodes filled with patch solution with a resistance of 20-30 MΩ. The craniotomy was centered 1.5mm lateral and 1.5mm posterior of the bregma over the ventral posteromedial nucleus (VPM) at a depth of 3.6mm. Signals were amplified using Axoclamp-700B amplifier, low passed at 3 kHz and digitized at 10 kHz.

Cortical silencing:
In order to activate ChR2, a LED light source at 460nm (Prizmatix Opt-LED-460) was coupled to a bare optical fiber (200μm dia., 0.22NA; ThorLabs M25L05) placed above the cortex. The LED was driven by an analog output from our acquisition system (National Instruments) for one second. The intensity of the light was around 7mW at the tip of the fiber. As Li et al. 22

Whisker stimulation and protocols:
Whiskers were trimmed to a length of 10-20 mm. When single whisker stimulation was given either the principal whisker (PW) or adjacent whisker (AW) were inserted into 21G needle attached to a galvanometer servo motor with a matching servo driver and controller (6210H, MicroMax 677xx, Cambridge Technology Inc.). The displacement was measured off-line using an optical Whisker stimulation was delivered without and with LED illumination, which stared 300ms before the whisker was stimulated and the light was turned on for 1sec.
Trials with LED stimulation alone were also delivered and they were used to correct drifts in voltage recordings, if occurred. These trials were pseudo-randomly delivered and were either 3s or 5s long with 2s inter-trial intervals. Each condition was repeated at list 6 times.

Data analysis:
The recordings were analyzed using custom software written in MATLAB