Chrna2-Martinotti Cells Synchronize Layer 5 Type A Pyramidal Cells via Rebound Excitation

Martinotti cells are the most prominent distal dendrite–targeting interneurons in the cortex, but their role in controlling pyramidal cell (PC) activity is largely unknown. Here, we show that the nicotinic acetylcholine receptor α2 subunit (Chrna2) specifically marks layer 5 (L5) Martinotti cells projecting to layer 1. Furthermore, we confirm that Chrna2-expressing Martinotti cells selectively target L5 thick-tufted type A PCs but not thin-tufted type B PCs. Using optogenetic activation and inhibition, we demonstrate how Chrna2-Martinotti cells robustly reset and synchronize type A PCs via slow rhythmic burst activity and rebound excitation. Moreover, using optical feedback inhibition, in which PC spikes controlled the firing of surrounding Chrna2-Martinotti cells, we found that neighboring PC spike trains became synchronized by Martinotti cell inhibition. Together, our results show that L5 Martinotti cells participate in defined cortical circuits and can synchronize PCs in a frequency-dependent manner. These findings suggest that Martinotti cells are pivotal for coordinated PC activity, which is involved in cortical information processing and cognitive control.


Author Summary
Cognitive functions and information processing are linked to the coordination of neuronal events and activities. This coordination is achieved through the synchronization of neuronal signals within subnetworks. Local networks contain different types of nerve cells, each of them playing distinct roles in the synchronization mechanism. To understand how synchronization is initiated and maintained, we have identified one of the key players using genetic strategies; we have identified a subtype of nicotine receptors uniquely expressed in cortical Martinotti cells. Because of their architecture and connection properties, Martinotti cells are able to synchronize ongoing activity of unconnected pyramidal cells (PCs). We show that this mechanism only applies to one subtype of PCs, thereby demonstrating that Martinotti cell inhibition is not spread randomly. By testing optimal firing patterns of Martinotti cells, we are able to coordinate the firing of this specific PC subtype over longer periods of time, showing how one unique interneuron is contributing to information processing. a1111111111 a1111111111 a1111111111 a1111111111 a1111111111

Introduction
Martinotti cells, ubiquitous to the cortex [1], are the most prominent cross-laminar interneuron subtype forming synapses in layer 1 onto the distal dendrites of cortical pyramidal cells (PCs) [1][2][3]. Despite this close structural relationship, the role of Martinotti cell inhibition is not clear. Studies identifying Martinotti cells by various markers have found different morphologies and microcircuit connectivity depending on the cortical layer in which their cell bodies reside [2]. In general, the division of neocortical interneurons into either parvalbumin-, somatostatin (SOM)-, or 5HT3aR-expressing cells [4][5][6] has been helpful for dissecting neural functionality; yet, these groups can be further subdivided and show partial overlap between interneuron markers. Martinotti cells are a subclass of SOM+ cells [7,2,4], and several combinations of transgenic lines have been created to try to genetically and morphologically isolate Martinotti cells [8][9][10]. For example, the SOM-cyclization recombinase (Cre) mouse line marks layer 1-projecting Martinotti cells with cell bodies in both layer 5 (L5) and layer 2/3 (infragranular and supragranular layers) but also labels non-Martinotti cells in layer 4 [10]. Although electrophysiologically, SOM+ Martinotti cells are often referred to as low-threshold spiking (LTS) neurons [3] or slow-inhibitory interneurons [11], early studies have shown up to four different firing patterns for Martinotti cells [2,9].
Functionally, cortical SOM+ interneurons have been suggested to provide a "blanket of inhibition" [12], a dense and nonspecific spread of inhibition on nearby PCs. Whether Martinotti cells are capable of generating such indiscriminate inhibition when firing simultaneously in large groups has not been tested. Martinotti cells that reside in the main cortical output L5 provide frequency-dependent disynaptic inhibition (FDDI) on neighboring PCs [13,14], an inhibitory mechanism that synchronizes two or more PCs by one or a few Martinotti cells [15]. Synchronized activities in the cortex have been reported in vivo [16] as well as in vitro, where slow oscillations appear to be initiated in L5 [17]. Moreover, computational studies suggest that Martinotti cell activity can synchronize L5 PC spiking through distal inhibition [18]; however, this has not been tested experimentally. It is intriguing that distal dendrite-targeting interneurons, generating attenuated inhibitory currents, can affect PC spike time output. Here, we genetically targeted the L5 Martinotti cell population using a nicotinic acetylcholine receptor α2 subunit (Chrna2)-Cre mouse line to investigate how Martinotti cell inhibition can synchronize L5 PC firing. Our results show that Chrna2-Cre-labeled L5 Martinotti cells were preferentially and reciprocally connected with thick-tufted PCs. Furthermore, we found that short burst firing of L5 Martinotti cells was able to reset L5 PC spiking and that controlling Martinotti cell activity to rapid bursts repeated in a slow rhythm was the most efficient inhibition to synchronize unconnected PCs. Finally, we show that L5 PC microcircuits could synchronize their own action potentials (APs) when coupled by L5 Martinotti cells and that inhibition was crucial for PC synchronization over prolonged periods.
The spike-frequency adaption shows how firing frequency of Chrna2-Cre/R26 tom cells decreases as a function of time (20.16 ± 2.44 Hz at 416 ms, n = 36 cells; Fig 1H, right). In summary, both morphological and electrophysiological characteristics of infragranular Chrna2-Cre/R26 tom cells (~97%) are similar to those reported in previous studies of Martinotti cells [2,13,21], thus we hereafter refer to these L5-specific Chrna2-Martinotti cells as MCs α2 .

MCs α2 Are Reciprocally Connected with Thick-Tufted Type A PCs
Previous studies have speculated that SOM+ cells (including Martinotti cells) predominantly contact specific subpopulations of PCs [21,22]. Thus, we patched pairs of a MC α2 and its neighboring PC in L5 ( 60 μm). Next, we categorized PCs into type A and type B cells based on morphological and electrophysiological criteria [23]. Cells with a large cell body, thicktufted basal dendrites with apical dendrites extensively branching in layer 1, burst-regular spiking, responding with large AHPs, prominent hyperpolarization sags, and pronounced rebound ADP were classified as type A PCs (Fig 2A, left). Cells with small soma, thin-tufted basal dendrites with limited spreading apical dendrites, absence of AHP or ADP, and small hyperpolarization sags were classified as type B PCs (Fig 2A, right). We expected L5-specific MCs α2 to be locally connected to PCs [2,13] and found, amongst morphologically reconstructed pairs of patched cells, that 77% of type A PCs-MCs α2 were connected (n = 7/9), while none of the patched type B PCs-MCs α2 were connected (n = 0/9). Out of the paired MC α2 -type A PC recordings, 55% (n = 5/9) of pairs were reciprocally connected. Paired recordings of PCs and MCs α2 revealed that high-frequency stimulation (70 Hz) of type A PCs generated excitatory postsynaptic potentials (EPSPs), or an occasional spike, in MCs α2 (example shows 12 repetitions from one patched type A PC-MC α2 pair; Fig 2B,   The reconstruction of a typical type A PC showing a thick-tufted dendrite (scale bar = 40 μm) and its response to a 500-ms-long depolarizing (100 pA) and hyperpolarizing (−60 pA) stimulus. Right: A representative type B PC with a thin-tufted apical dendrite (scale bar = 40 μm) and its current clamp response (as for left). Note the deeper AHP (following a depolarizing current pulse), the more prominent sag (during a hyperpolarizing current pulse), as well as the pronounced rebound ADP (following a hyperpolarizing current pulse) in the type A PC compared to type B PC (see arrows). (B) Type A PCs can excite postsynaptic MCs α2 (inset) and generate facilitating EPSPs (left, n = 7/9 pairs, 12 repetitions from one example pair are shown) when stimulated with high frequency (70 Hz), whereas type B PCs do not trigger EPSPs in MCs α2 (right, n = 0/9 pairs, 12 repetitions). Inset shows experimental setup. (C) Typical MC α2 discharges (top) to a 500-ms-long ( Additionally, inhibitory postsynaptic potentials (IPSPs) were generated in type A PCs (IPSP amplitude: −1.08 ± 0.12 mV; example shows 12 repetitions from one patched MC α2 -type A PC pair; Fig 2C,  Optogenetic Activation of MCs α2 Shows Frequency-Dependency of MC-PC Inhibition We next investigated the influence of MC inhibition on PCs when simultaneously activating a large group of MCs α2 in Chrna2-Cre mice (1-2 mo old) previously injected with floxed Channelrhodopsin-2 (ChR2; Fig 3A). Compared to electrical stimulation of single MCs α2 , light activation of MC α2 groups produced IPSPs in type A PCs with a higher mean amplitude (from −0.96 ± 0.05 mV to −1.41 ± 0.04 mV), a smaller mean time to peak (from 29.53 ± 1.24 ms to 20.54 ± 0.97 ms), and a decreased mean half decay time (from 63. We also recorded from type B PCs (n = 12 cells), but no IPSPs were observed in type B PCs in response to blue light stimulation of ChR2+ MCs α2 (Fig 3B, right). We next tested different stimulation frequencies (2,5,15,25,40, and 70 Hz) [13,18] for ChR2+ MCs α2 to investigate the role of MC α2 firing frequency on IPSP amplitude in type A PCs (n = 12 cells; Fig 3C). We found a nonlinear relationship between IPSP amplitude and MC α2 stimulation frequency in which, at higher frequencies (>15 Hz), IPSPs summed into smooth compound IPSPs, probably due to the depressing synaptic properties of the MC α2 -to-PC connection [13]. To further characterize the frequency-dependency of the MC α2 -PC IPSPs, we stimulated MCs α2 with continuous light, which generated accommodating firing in MCs α2 (Fig 3D, Fig 3D and 3E, S3 Data). This suggests that large compound IPSPs can be generated in type A PCs when MCs α2 fire at high frequencies and that the compound IPSP amplitude mainly depends on the firing frequency during the first 300 ms (Fig 3D and 3E).

Bursts of MCs α2 Can Reset Type A PC Firing
Martinotti cells have been shown to provide FDDI [13,24,15] onto PCs. To examine whether L5-specific MCs α2 generate FDDI, we patched pairs of type A PCs with cell bodies next to MCs α2 and provided high-frequency (70 Hz) current injection to one PC (Fig 4A, top). This led to an early EPSP in the other PC, presumably due to monosynaptic PC-PC connections, followed by a delayed inhibition (FDDI, amplitude: −1.02 ± 0.23 mV, time to peak: 57.16 ± 2.43 ms, half decay time: 95.34 ± 5.64 ms, n = 12 cells; Fig 4A, Fig 4C, S4 Data). Current clamp recordings from MCs α2 showed that the green light generated strong hyperpolarization of MCs α2 and that, subsequently, bursts of rebound spikes were generated in HaloR-expressing MCs α2 upon light termination (Fig 4D, top and S4A and S4B Fig, S5 Data). HaloR-activation for 500 ms consistently evoked one or more rebound APs with varying frequency in MCs α2 that could not be blocked by the I h blocker ZD7288 (20 μM, S4C Fig), similar to I h -independent rebound APs in distal dendrite-targeting X98 cells [9]. Additionally, MCs α2 and type A PCs were patched in the presence of carbachol (10 μM) to further examine how MCs α2 could modulate L5 PC spontaneous firing. Carbachol depolarized PCs and MCs α2 by 10 to 15 mV and did not result in a specific oscillatory frequency as seen with high concentrations of carbachol;  large IPSCs in type A PCs (107.40 ± 3.54 pA, n = 3 cells; Fig 4E, top, S4 Data) and the absence of IPSCs in type B PCs (n = 3 cells; Fig 4E, bottom). Together, these results show that a rapid burst of MC α2 APs can abruptly halt the firing of type A PCs and also reset PC firing by temporal coupling rebound APs of PCs while leaving type B PC firing unaffected. Therefore, we only aimed for type A PCs for the remainder of the study.

Bursts of MC α2 APs Synchronize Firing of Type A PCs at Slow Frequencies
It is unknown how PCs synchronize their firing, although computational studies have suggested a role for distal dendrite-targeting interneurons in synchronizations [11,18,29]. Thus, we only patched unconnected type A PCs and recorded spontaneous firing, in the presence of carbachol, from two PCs simultaneously (n = 24 cells) while optogenetically stimulating the MC α2 population at various frequencies (2,5,15,25,40, and 70 Hz). To identify the frequency of MC α2 activity that best temporally aligns unconnected, randomly firing, type A PCs, we recorded repeats of 4-s sweeps (2 s light-off, 2 s light-on). When pairwise superimposing simultaneous recordings from two type A PCs, we observed that light stimulation of 2 Hz or 15 Hz (n = 24 cells; 12 black and 12 grey PC spike trains; Fig 5A) created a rhythmical firing pattern of type A PCs highlighted by kernel density estimates, which show the distribution of APs over time (orange traces). Although mean power spectral density plots from both 2 Hz and 15 Hz MC α2 stimulation revealed peaks around 2 Hz (1.99 ± 0.09 versus 1.87 ± 0.14 Hz, n = 24 cells), other frequencies tested (5,25,40, and 70 Hz) did not result in any clear peaks ( Fig 5B, top and S7 Fig, S7 Data). This indicates that MCs α2 preferentially give rise to slow frequencies in a group of type A PCs. However, flat mean coherence plots of pairwise analyzed PCs did not show any correlation between simultaneously recorded type A PCs in specific frequency bands plotted up to 20 Hz. This suggests that type A PCs as a population can produce an oscillatory firing rhythm, but individual cells are mostly out of phase and not synchronized with each other (n = 24 cells; Fig 5B, bottom). To generate complete synchrony between type A PCs, we hypothesized that a rapid burst of MC α2 activity could reset/align type A PC spiking (3-4 APs as seen in Fig 4D and S6A Fig) and a slow rhythm could maintain in-phase synchronous firing. To test this, we patched two unconnected type A PCs and stimulated MCs α2 with 15-Hz bursts every 500 ms (2 Hz). This stimulation protocol resulted in high AP synchronization ( Fig 5C) directly after MCs α2 were paused. Mean power spectral density (peak at 2.02 ± 0.04 Hz, n = 24 cells) and mean coherence examination showed that type A PCs followed MC α2 stimulation frequency and were pairwise aligned in that frequency (Fig 5D, S7 Data), suggesting synchronized firing of type A PCs at slow frequencies. This shows that MCs α2 have means to both initiate and maintain prolonged type A PC synchronous firing.

Slow Frequency Burst Stimulation of MCs α2 Is Synchronizing Type A PC Spike Timing via Minimally Depressing IPSPs
Next, we sought to quantify the synchrony (provide a synchrony index [30]) between type A PC firing when MCs α2 were activated in bursts of 15 Hz. A representative recording of two simultaneously captured type A PCs is shown in Fig 6A, where initially unsynchronized type A PCs aligned during MC α2 stimulation (orange rectangles highlight synchronized APs). In the absence of MC α2 stimulation, the mean cross-correlograms of pairwise analyzed recordings showed only low magnitude peaks, while light stimulation organized firing of both type A PCs in cohorts every 500 ms (n = 24 cells; Fig 6B). A 3-fold increase in the synchrony index could be extracted from the cross-correlograms when MCs α2 were light stimulated (control: 0.21 ± 0.03, burst stimulation: 0.61 ± 0.04, n = 12 dual recordings, n = 24 cells, p < 0.0001; Fig 6C, S8 Data).
Thus, we conclude that delivery of MC α2 inhibition in bursts of 15 Hz indeed synchronized type A PC firing. This is likely due to burst firing repeated in slow frequency, causing inhibition in type A PCs with little depression compared to continuous 15 Hz stimulation of MCs α2 showing apparent synaptic depression (examples in grey, mean in black, red dashed lines for visual guidance, Fig 6D). Interestingly, the 15-Hz continuous light stimulation revealed that bursting PCs can switch firing patterns from burst-spiking into single-spiking [25,18]   To test if type A PC circuits can self-synchronize their firing through Martinotti cell activation, we designed a closed-loop system (optical feedback inhibition [31]) for paired recordings that Chrna2-Cre mice previously injected with floxed ChR2. Kernel density estimates showed increased co-occurrences of type A PC APs during optical feedback inhibition (Fig 7A). Moreover, because these experiments involved a leading and a following type A PC, it was possible to calculate the statistical dependency between the spike trains of two PCs and to express this with a mutual information index (see Materials and Methods). This index gives an estimate of how well one signal can predict the other and is helpful to interpret to what extent one PC can drive another, e.g., via recurrent or feedforward inhibition. In controls, the mutual information index was low (2 ± 1, n = 12 dual recordings), whereas turning on the optical feedback inhibition resulted in immediate auto-alignment of type A PC APs with an increased index (11 ± 5, n = 12 dual recordings, p < 0.05; Fig 7B, S9 Data). Venn diagrams show the mutual information as the degree of overlap between two circles, representing each PC train as entropy (Fig 7B, inset). The overlap demonstrates the predictive value (mutual dependency) Synchronization by Chrna2-Martinotti Cells between a known PC train and a following PC train. When shifting one spike train relative to the other, the incremental mutual information index plot (mutual information index as a function of time lag, Fig 7B) showed that the mutual dependency was largest around 0-ms lag, suggesting high synchronization of the two PCs directly when coupled by optical feedback inhibition. Peaks around ± 400-600 ms indicate that this activity-dependent inhibition causes repeated synchronization every 400-600 ms.

Discussion
We found that MCs α2 were exclusively synaptically connected to large, thick-tufted PCs, often referred to as L5B PCs [32] or type A PCs [23]. Different PC morphologies seem to be associated with different connectivity patterns in the brain, e.g., large, thick-tufted PCs are usually synonymously named subcerebral projection neurons or pyramidal tract neurons, whereas thin-tufted PCs, or type B PCs [23] are callosal projection neurons or intratelencephalic neurons [26][27][28]. Our in vitro preparation could not define PCs according to connectivity patterns; however, based on the extensive branching of the distal dendrites and the large triangularshaped cell bodies, we find it likely that the type A PCs correspond to the thick-tufted PCs [22] and are probably subcortically projecting [26][27][28]. Thick-tufted type A PCs can further be described by firing properties as single-spiking or burst-spiking [25][26][27][28]. Typically, at nearthreshold potentials, bursting cells respond with two or more bursts, of two or more APs, generated in quick succession with short interspike intervals [25]. Burst properties of PCs disappear with increasing current injections [25] and may be dependent on the size of the dendritic tree [33]. In addition to morphological variances, such as a smaller soma compared to type A PCs, type B PCs had characteristic electrophysiological differences. Our pair-recordings between type A PCs and MCs α2 confirmed that type A PCs provided facilitating synaptic responses in MCs α2 . We also found depressing synaptic connections from MCs α2 to type A PCs [13], while no type B PC connectivity with MCs α2 was observed. Lack of IPSPs in type B PC was not likely due to shunting of inhibition, as voltage clamp recordings also failed to find synaptic connectivity between MCs α2 and type B PCs. However, due to difference in thin-versus thick-tufted morphology, it is possible that the internal solution creates less dialysis of the chloride ion Clconcentration in type B PCs compared to type A PCs. Therefore, perforated patch recordings would be needed to firmly rule out the possibility of shunting of inhibition. Still, we found that FDDI, a Martinotti cell-dependent feature [13,24,15] was relayed by MCs α2 and consequently was specific for type A PCs. This is in agreement with a previous study showing FDDI between thick-tufted PCs but not between corticocallosally projecting cells [22].
Distal inhibition by individual MCs is important for shaping local dendritic voltage-activated responses. FDDI combined with dendritic depolarization has shown that MCs can attenuate back-propagating AP-activated Ca 2+ spike firing and thereby reduce burst firing of PCs [34]. On the network level, collective and precisely timed Martinotti cell activity can further be potent enough to affect somatic spike generation. Here, we first used HaloR to examine if blocking MC α2 activity could eliminate the appearance of FDDI in thick-tufted PCs [13,24,15]. This led to the observation that on the termination of green light MCs α2 fired bursts of HaloRinduced rebound spikes, inhibiting PCs and subsequently causing hyperpolarization-induced rebound APs that could reset PC firing. The HaloR-induced rebound in MCs α2 is a methodological artifact and has little physiological relevance; however, it is interesting to speculate whether Martinotti cells receive inhibition that could generate rebound spikes. Recently, the vasoactive intestinal peptide (VIP) interneuron has been shown to densely inhibit Martinotti cells [35]. The high connection probabilities between VIP cells and Martinotti cells [36] suggest that VIP cells could provide strong hyperpolarization in Martinotti cells for the possible generation of rebound excitation. Rebound spikes have been previously demonstrated to occur in entorhinal cortex neurons in vivo and are attributed to play a role in generating grid cell fields that usually arises when grid cells fire synchronized [37,38]. A similar role could be applicable to rebound spikes in the neocortex, where a "blanket of inhibition" [12] evolves through the synchronized spread of inhibition, serving to coordinate PC firing. Thereby, VIP cells appear to make "holes in the blanket of inhibition" [35] by inhibiting Martinotti cells [35,36]. In other words, VIP cell activity might regionally disrupt coordinated PC firing while local Martinotti cell activity could reset and rescue PC synchronous firing.
Second, focusing on the combined activity of ChR2-expressing MCs α2 , our data show that bursts of MCs α2 were able to reset type A PC firing and, if repeated, could synchronize PC activity. In a computer model, oscillatory inhibition of the distal PC dendrite at 10-20 Hz, presumably by LTS SOM+ Martinotti cells, was shown to control L5 PC firing [18]. Our findings support that 15 Hz firing of MCs α2 can align type A PC firing but also show that 15 Hz firing in short bursts more reliably synchronizes PCs compared to continuous 15 Hz firing. In other computational models, the importance of a beta rhythm in regulating gamma oscillations and intercortical signaling has been demonstrated and, furthermore, that the beta frequency is regulated by cholinergic modulators [11,29,39]. In this respect, the exclusive expression of the alpha 2 cholinergic receptors in MCs α2 is noteworthy and may suggest a specific role for MCs α2 in transmitting the modulatory action of cholinergic signaling. Cholinergic modulation of LTS cells has been suggested to generate beta oscillatory activity (beta2) in L5 of the primary auditory cortex [40]. These oscillations were insensitive to the muscarinic antagonist atropine but sensitive to the nicotinic receptor antagonist d-Tubocurarine [40]. Thus, computational and experimental studies indicate that the beta rhythm is important for network properties [40,41]; however, beta activity in bursts repeated in slow frequency has not been reported previously. At this slow frequency MC α2 -PC inhibition shows minimal depression, similar to the minimal depression of slow firing SOM+ interneurons defined by their green fluorescent protein expression in a transgenic mouse (GIN-cells) [42], and therefore, a combination of rapid bursting and slow rhythmical inhibition seems most effective to synchronize PCs.
Genetic targeting and optical feedback inhibition are a potent technique to study how PCs can drive a population of interneurons by their innate rhythm. A previous study used a closedloop system to optogenetically produce feedback inhibition onto PCs from parvalbumin + interneurons [31]. Sohal et al. used synthetic excitatory post synaptic currents (EPSCs, dynamic clamp) in a single PC triggering parvalbumin+ interneuron excitation with light [31]. Differently, in our study, we depolarized optically stimulated MCs α2 in the presence of carbachol and measured synchronization of simultaneously recorded PCs using mutual information. Analogously to our optogenetic stimulation, gap junctions could provide a physiological mechanism for the synchronization of interneuron populations [42][43][44]. Berger et al. have shown the existence of electrical coupling between L5 MCs [15]. It will be interesting in the future to explore the existence of gap junctions between MCs α2 .
The Chrna2-Cre/R26 tom mouse line simplifies identification and characterization of L5 Martinotti cells. MCs α2 are morphologically and electrophysiologically homogenous, further evincing the specificity of our marker. The dense axonal plexus observed in L1 and the near absence of Cre+ cell bodies in L2 (2.4%) in Chrna2-Cre/R26 tom mice also indicate that Cre + cells are, in fact, L5 MCs. Still, we found a high proportion of Cre+ cells in Chrna2-Cre/ R26 tom mice that were not labelled with the antibody against SOM, and this could be due to extra-somatic location of the peptide. So far, SOM-Cre is the most widely used transgenic mouse line for targeting MCs together with the GIN mouse [8,45], but still SOM-Cre has been shown to label all cell layers [46]. Here, we provide a layer-specific, single genetic marker for MCs across the cortex and confirmed their inhibitory nature using single cell RT-PCR. Although we did not explicitly block optogenetically evoked, inhibitory postsynaptic currents of MCs α2 (e.g., with Gabazine), Martinotti cell dendritic inhibition in vivo has been shown to be GABA A -mediated [34]. The specific expression of Chrna2 in inhibitory L5 MCs α2 raises questions of how important the α2 subunit is for cholinergic inputs. Several cortical interneurons express nicotinic acetylcholine receptors (nAChRs) [47][48][49], suggesting cholinergic modulation of inhibition in the cortex, most likely from the basal forebrain [50]. Cortical LTS cells, such as Martinotti cells [51,52], are excited by acetylcholine via nicotinic receptors and alter cortical circuit processing [53]. Cholinergic input is most likely mediated by additional nicotinic subunits that together form high affinity receptors for acetylcholine [54]. Several candidate subunits exist, but perhaps the more promising ones, judged from their specific expression in cortical L5, response to nicotine, and known co-expression with α2 subunits, include α6-nAChRs, β2-nAChRs, and β4-nAChRs [55][56][57]. The focus of our work has been on the functionality of Martinotti cells, not the nAChR subunits; however, earlier studies of cholinergic subunits can provide potential clues to Martinotti cell function. A deletion of α2-nAChRs has shown a normal phenotype but altered responses during nicotine-associated behaviors [58]. Deletion of α2-nAChRs has also shown reduced nicotine-induced hippocampal LTP in the temperoammonic path, most likely via oriens-lacunosum moleculare (OLM) interneurons [59,19]. Interestingly, Chrna2 is expressed in OLM cells, which target the distal dendrites of hippocampal PCs in a comparable manner as MCs α2 target the distal dendrites of cortical PCs. In some similarity to the suggested role for LTS cells in directing the flow of information in the cortex [53], OLM cells have been suggested to gate internal and external signals to the hippocampus [19]. In addition, MCs α2 might modulate cortical states, because SOM+ interneurons have recently been implied to be involved in transitions between UP and DOWN states [60]. Furthermore, studies of β2-nAChR KO mice have suggested a role for β2-containing nAChR in restricting cortical UP states and might be interesting for future studies on how nAChR are distributed in cortical interneurons such as Martinotti cells [13,61,62].
Our preparation did not examine cortical UP and DOWN states; instead, we depolarized neurons with a low concentration of the cholinergic agonist carbachol. As the IPSP amplitude generated by MCs α2 is dependent on the membrane potential of the postsynaptic PCs, this illustrates how MC α2 inhibition (amplitude of IPSPs) could alter in a state-dependent manner, thereby exerting a state-dependent modulation of PC excitability.
In summary, we report the identification of a marker specific for L5 Martinotti cells projecting to layer 1. These Martinotti cells were synaptically connected to large, thick-tufted PCs with prominent AHP and ADP, demonstrating a distinctive microcircuit between one type of interneuron and one subtype of PCs. Furthermore, we demonstrate that Martinotti cell-mediated inhibition can initiate and also maintain synchronous firing between PCs. We also show that this inhibition is frequency dependent and, when repeated in beta bursts, can continuously align firing of PCs. Lastly, using a closed-loop system in which PCs auto-synchronized their firing, we show that Martinotti cells were able to bridge the communication between unconnected PCs via activity-dependent inhibition. Thus, via their feedback and feedforward connections, Martinotti cells are important for regulating thick-tufted type A PC output in L5, most likely altering voltage-dependent dendritic properties and actively influencing somatic spike generation and synchronization.

Ethics Statement
All experiments were approved by the Swedish Animal Welfare authorities and followed Uppsala University guidelines for the care and usage of laboratory animals (ethics permits C132/13 and C135/14). Efforts were made to minimize the numbers of animals used.

CLARITY
The CLARITY procedure followed a standard protocol [63]. In summary, 2-3-mo-old Chrna2-Cre/R26 tom mice (n = 2) were transcardially perfused with 20 ml of ice-cold 1x PBS solution followed by 20 ml of a hydrogel monomer solution consisting of 4% acrylamide, 0.05% bis-acrylamide, 0.25% VA-044 initiator, and 4% paraformaldehyde in PBS. Brains were quickly dissected and placed in hydrogel monomer solution for 3 d at 4˚C. Prior to polymerization of the hydrogel monomer solution, samples were placed in a desiccation chamber attached to a vacuum pump. With the sample lid ajar, air was removed from the chamber for 10 min and replaced with nitrogen gas, after which the sample lid was tightly shut. The hydrogel monomer solution was polymerized by heating the samples to 37˚C for 3 h in a water bath whilst shaking. Embedded tissue was extracted from the gel, and brains were sliced to 3-mm coronal sections using a brain matrix. Passive clearing of slices was achieved by repeated, 3-d washes in a 4% Sodium Dodecyl Sulphate (SDS) sodium borate buffer (200 mM, pH 8.5) solution at 45˚C on a shaker plate for 6 wk. SDS was removed from the samples by incubating in PBST 0.1 (1x PBS and 0.1% Triton X-100) on a shaker plate for two consecutive 1-d washes.
Patch-clamp data were analyzed with custom routines in MATLAB. APs were triggered by 500-ms depolarizing current injections from 10-100 pA. The first fired AP in response to minimal current injection was analyzed for AP amplitude (peak to AHP voltage), threshold (where the change in membrane potential exceeds 20 mV/ms), half-width (halfway between threshold voltage and peak), and first spike latency (time between stimulus onset and the AP threshold of the first spike). AHPs were analyzed for magnitude (AP threshold-minimum of voltage trough between the first and the second AP in a spike train). Spike rate was calculated as the number of APs per 1,000 ms. Spike-frequency adaptation was measured as the inverse of the mean of the last three interspike intervals (steady-state frequency) divided by the inverse of the first interspike interval (maximum frequency) in response to 100-pA current injections and subtracted from 100% (no adaptation). In spike-frequency adaptation plots, the reciprocal of consecutive interspike intervals is shown for each AP versus the time after onset of the current pulse.
ChR2 stimulation frequencies (2,5,15,25,40, and 70 Hz) and HaloR-evoked hyperpolarization (5, 10, 100, 250, 500, and 1000 ms) were applied in randomized order to avoid statistical dependencies between cases. ChR2-triggered IPSPs and FDDI were measured as amplitude, time to peak, and half decay time, for which the onset was defined as the time at which the potential exceeded three times the standard deviation of the preceding baseline. HaloR-evoked hyperpolarization amplitudes were quantified as the difference between resting membrane potential and the peak of hyperpolarization. Rebound APs were quantified by number, maximum frequency, and duration (time of last minus time of first rebound AP). Burst-spiking PCs were distinguished from single-spiking PCs by obtaining the interspike intervals, at which burst-spiking PCs showed an increased amount of short ( 20 ms) interspike intervals at nearthreshold potentials.

Imaging and Tracing
MCs α2 were identified by cortical tdTomato-expression in Chrna2-Cre/R26 tom mice that were perfused as previously described [19], and 20-μm-and 60-μm-thick coronal slices were imaged using a Zeiss LSM 510 Meta confocal microscope. Cells targeted by electrophysiology experiments were routinely filled with biocytin and stained with streptavidin-488 nm for post-hoc analysis. Images were collected on a Zeiss LSM 510 Meta confocal microscope and stacked and 2D-stichted using ImageJ 1.50a (NIH), where the color palette was adjusted for consistency (tdTomato-red, biocytin/ChR2-green). Soma detection and Neurite tracings were done semi-automated with NeuronJ 1.4.2 (ImageJ Plugin) or fully automated with Imaris 8.1 Fila-mentTracer (Bitplane) using "Autopath" in the algorithm settings and the threshold mechanism to correct for over-/under-sampled tracings following the image intensity.

Optogenetics
Chrna2-Cre/R26 tom mice (1-2 mo old, n = 33) were anesthetized with Isofluran (1%-4%) and placed on a heat pad, with the head fixed with a nose holder and ear bars in a stereotaxic frame (Stoelting Co.). The skin was cleaned with iodine and opened with a straight incision, and the bregma was identified using small amount of peroxide. The coordinates for bilateral virus injection were as follows: AP: −2.46 mm, ML: +/−4.00 mm, and DV: 2.00/2.50 mm. We used bilateral injections to obtain the maximum number of slices containing the primary auditory cortex, with preserved dendritic trees of the type A PCs in layer 1 (on average two slices, 300 μm thick, per hemisphere) per animal. A small hole was drilled in the skull using a dental micro drill, causing minimal bleeding during the process. Viral vectors (pAAV-EF1a-double floxed-hChR2(H134R)-EYFP-WPRE-HGHpA, University of Pennsylvania Vector Core Facility) in solution (6.2 x10 12 / 1.6×10 13 particles/ml) of 0.50-1.00 μl were slowly infused (0.10 μl/ min) into the auditory cortex at two depths (2.00 and 2.50 mm) using a Hamilton 10-μl syringe and an infusion/withdraw pump (World Precision Instruments/KD Scientific). After infusion, the needle was left in place for 1-5 min to allow complete diffusion of the virus. Next, the scalp was rehydrated with saline and sutured with 4-5 stitches and local anesthesia (a drop of Lidocaine/Marcaine) applied onto the sutured skin before the mouse was allowed to wake up. The animal was monitored and kept warm until fully awake (moving and starting to eat and drink water). Mice were killed after approximately 3-6 wk for in vitro electrophysiological experiments and/or histological procedures.

Optical Feedback Inhibition
We modified our dynamic clamp system [30] running the Real Time Application Interface for Linux-based (RTAI) from the Politecnico di Milano Institute-Dipartimento di Ingegneria Aerospaziale (Mantegazza, http://www.rtai.org/) on a Dell Precision T1500 with a Quad-Core Intel Core i7 with 2.80 Ghz, 5.8 Gigabyte memory, and a National Instruments DAQ card (NI PCI-6251). Routines for data acquisition were programmed in GNU-C using the Linux Control and Measurement Device Interface (COMEDI). The membrane potentials of the patched cells were acquired in 20 kHz, and APs were detected based on threshold (>−20 mV) in real time, triggering the LED (CoolLED pE-1) via 3-ms (ChR2 experiments) or 500-ms (HaloR experiments) TTL pulses.

Data Analysis
Matlab (version 2013a, MathWorks) was used for data analysis. APs from MCs α2 and PCs were detected based on threshold (>−20 mV). PC spike trains were transformed into a series of 0 (no spike) and 1 (spike), with 0.1-ms precision (binning). Accordingly, kernel density estimates are probability densities and were computed on population data (all patched PCs) with 0.1 "bandwidth" ("ksdensity" command in Matlab). The kernel density estimates show the distribution of APs over time and highlight increased (peaks) and decreased (valleys) co-occurrences of spikes. Power spectral density analysis of the binary spike series was made using Welch's method ("pwelch" command in Matlab) to find the frequency components with highest power. Coherence was calculated pairwise (i.e., for simultaneously patched PCs, the similarity between the binary sequences of the two PCs was calculated) and plotted as mean over all recordings ("cohere" command in Matlab) to investigate the dependence of two cells as a function of frequency. Cross-correlograms were calculated using the "coeff" option of the cross-correlation command in Matlab (to scale the cross-correlation values from −1 to 1 and prevent dependency of the cross-correlation on the number of spikes) and then smoothed by a moving average filter with a span of 10 ms to find the functional dependence between the APs of simultaneously recorded cells over time. We displayed cross-correlations over a lag range of ±1 s. The synchrony index was defined as the maximum peak of the normalized cross-correlograms between −50 ms and 50 ms, as previously described [30], with 0 ≙ no synchronization and 1 ≙ full synchronization. To investigate statistical dependence, mutual information MI (PCleft, PCright) was determined on the binary spike data (binning = 20 ms) by calculating the distributions of the spike trains (univariate distribution of each spike train separately as well as bivariate distribution of both spike trains) and expressing them as entropies H(PCleft) and H(PCright), meaning how "diverse" the spike trains were. The sum of the two entropies H (PCleft) and H(PCright) minus the joint entropy H(PCleft, PCright) quantified the conditional entropy, i.e., the mutual information MI(PCleft, PCright). The MI was formulated as a mutual information index with a scaling factor of 1,000 [30].

Statistical Analysis
All statistical analysis was performed using R version 3.2.3 (Foundation for Statistical Computing, Vienna, Austria). Data are reported as mean ± standard error of the mean (SEM) and plotted as bar plots or box plots. Data larger than q3 + 1.5 Ã (q3-q1) or smaller than q1-1.5 Ã (q3-q1), with q1 and q3 denoting the 25th and 75th percentiles (see box plots), were considered as outlier and discarded. Statistical comparisons were determined using two-tailed Student's paired t test, and to account for multiple comparisons, the data were analyzed using ANOVA and post-hoc test with Tukey correction ( Ã ≙ p < 0.05, ÃÃ ≙ p < 0.01, ÃÃÃ ≙ p < 0.001, and ÃÃÃÃ ≙ p < 0.0001). The order of stimulations of different frequencies (e.g. 2,5,15,25,40,70 Hz) was systematically varied to avoid statistical dependencies between the timing of recordings and the frequency investigated. (XLSX) S1 Movie. Tomato+ cells of Chrna2-Cre/R26 tom mice visualized across cortical areas. A series of images from adult (2 months old) Chrna2-Cre/R26 tom mouse cortex (coronal slice, 1300 μm thickness) after CLARITY processing is shown. Please note the second band of tomato+ cells highlighted in the stratum oriens of hippocampus [19] and the dense axonal arborisation in stratum lacunosum-moleculare, highlighted as a grey dense mass.