Intrinsic Circuits in the Lateral Central Amygdala

Abstract Network activity in the lateral central amygdala (CeL) plays a crucial role in fear learning and emotional processing. However, the local circuits of the CeL are not fully understood and have only recently begun to be explored in detail. Here, we characterized the intrinsic circuits in the CeL using paired whole-call patch-clamp recordings, immunohistochemistry, and optogenetics in C57BL/6J wild-type and somatostatin-cre (SOM-Cre) mice. Our results revealed that throughout the rostrocaudal extent of the CeL, neurons form inhibitory connections at a rate of ∼29% with an average amplitude of 20 ± 3 pA (at −40 mV). Inhibitory input from a single neuron is sufficient to halt firing in the postsynaptic neuron. Post hoc immunostaining for protein kinase Cδ (PKCδ) in wild-type mice and paired recordings in SOM-Cre mice demonstrated that the most common local connections were PKCδ(−) → PKCδ(−) and SOM(+) → SOM(+). Finally, by optogenetically activating either SOM(+) or SOM(−) neurons, we found that almost all neurons in the CeL were innervated by these neuronal populations and that connections between like neurons were stronger than those between different neuronal types. These findings reveal a complex network of connections within the CeL and provide the foundations for future behavior-specific circuit analysis of this complex network.


Introduction
The amygdala has long been known to play a crucial role in processing innate emotions, particularly fear (Klüver and Bucy, 1939;Weiskrantz, 1956;Sah et al., 2003). In Pavlovian fear conditioning, an associative learning paradigm widely used to study amygdala function, subjects learn to associate a neutral sensory stimulus [the conditioned stimulus (CS)], with an aversive one (the un-conditioned stimulus; LeDoux, 2000). Following learning, the previously neutral CS now evokes a defensive response (i.e., freezing of movement or flight; Gross and Canteras, 2012). A converging body of evidence has established the amygdala as a central player in fear conditioning where the basolateral amygdala (BLA) and the central amygdala (CeA) are the key sites involved in the acquisition and expression of fear (LeDoux, 2000;Sah et al., 2003;Duvarci and Pare, 2014). The BLA has been extensively studied with respect to its cell types, intrinsic circuits, and extrinsic connections (LeDoux, 2000;Sah et al., 2003;Duvarci and Pare, 2014), while the CeA has received considerably less attention and the intrinsic circuits within this nucleus are less well understood.
The CeA is a GABAergic nucleus (McDonald and Augustine, 1993;Sun and Cassell, 1993) that is anatomically divided into the lateral sector of the CeA (CeL) and the medial sector of the CeA (CeM), with substantial unidirectional connections between the CeL and the CeM (Mc-Donald, 1982;Grove, 1988;Jolkkonen and Pitkänen, 1998). Neurons in both regions also make extensive local connections (McDonald, 1982;Sun and Cassell, 1993;Jolkkonen and Pitkänen, 1998), with local glutamate excitation of CeL neurons evoking IPSCs in neighboring neurons (Lopez de Armentia and Sah, 2004). Recent studies have divided CeL neurons into distinct populations based on the expression of immunohistochemical markers, electrophysiological properties, and synaptic connections Haubensak et al., 2010;Li et al., 2013). Of these, one population expresses protein kinase C␦ [PKC␦(ϩ)], and these neurons are predominantly described as late-firing (LF) neurons, exhibiting a substantial delay to action potential (AP) initiation in response to depolarizing somatic current injections. Following fear conditioning, these neurons respond to the CS with a reduction in activity and have therefore been called Ce-L OFF cells Haubensak et al., 2010). A second population of CeL neurons, which is largely separate from the PKC␦(ϩ) population, expresses somatostatin [somatostatin-positive (SOMϩ); Li et al., 2013]. These neurons receive direct synaptic input from the lateral amygdala, which is potentiated following auditory fear conditioning (Li et al., 2013). Electrophysiologically, PKC␦(Ϫ) neurons which are predominantly SOM(ϩ), have been described as either LF or regular spiking (RS). Following fear conditioning, PKC␦(Ϫ) neurons respond to the CS with an increase in activity, and have therefore been called CeL ON neurons (Ciocchi et al., 2010;Haubensak et al., 2010), which likely also correspond to SOM(ϩ) neurons (Yu et al., 2016). PKC␦(Ϫ) neurons inhibit PKC␦(ϩ) neurons, which in turn project to the CeM (Haubensak et al., 2010).
This organization has led to one model in which fear expression is mediated by CS-related information driving PKC␦(Ϫ) neurons, presumably SOM(ϩ) neurons, in the CeL via excitatory input from the BLA and thalamus. These neurons in turn inhibit PKC␦(ϩ) neurons, resulting in disinhibition of the CeM and the expression of fear Haubensak et al., 2010). However, some SOM(ϩ) neurons in the CeL also project to the periaqueductal gray (PAG;Penzo et al., 2014), and CSdriven activity of these neurons also contributes to fear expression (Tovote et al., 2016). Moreover, recent studies have reported that neurons in the CeL are also involved in feeding (Cai et al., 2014), and pain (Han et al., 2015). Neurons engaged during feeding and pain responses are also part of the PKC␦ and SOM population, indicating that the intrinsic circuitry of the CeL is complex, and the strength, identity, and physiologic role of individual local connections are not fully understood. In this study, we provide a detailed investigation of local circuits in the CeL.

Animals
All studies were approved by the University of Queensland Animal Ethics Committee, and experiments were conducted in accordance with the Australian Code of Practice for the Care and Use of Animals for Scientific purposes. Adult (6 -15 weeks old) male wild-type C57BL/6J mice were used for electrophysiology experiments. Where stated, we also used both male and female mice (8 -12 weeks old) from a somatostatin-IRES-cre mouse line (SOM-Cre; C57BL/6J background; Sst tm2.1(cre)Zjh ) that was acquired from The Jackson Laboratory. These mice express cre recombinase under the SOM promoter, thereby allowing selective targeting of SOM(ϩ) neurons using cre-dependent viral constructs (described below). Mice were genotyped by the Australian Equine Genetics Research Center.
Data were acquired and analyzed using AxoGraphX software (AxoGraph). Brain slices were continuously perfused with oxygenated aCSF (34°C; 3-4 mL/min), and recording electrodes (4 -6 M⍀; glass capillaries, Harvard Apparatus; PC-10 Electrode Puller, Narishige) were filled with a KMeSO 4 -based internal solution (KMeSO 4 135 mM, NaCl 8 mM, HEPES 10 mM, MgATP 2 mM, GTP 0.3 mM, phosphocreatine 7 mM, EGTA 0.2 mM, and biocytin 0.2%, pH 7.2 with KOH, osmolarity 295 mOsm/kg) unless otherwise stated, in which case a CsMeSO 4 -based internal solution was used (CsMeSO 4 135 mM, NaCl 8 mM, HEPES 10 mM, MgATP 2 mM, GTP 0.3 mM, phosphocreatine 7 mM, and spermine 0.1 mM, pH 7.2 with CsOH, osmolarity 300 mOsm/kg). In some experiments GABA (10 mM) was added to the KMeSO 4 -based internal solution to avoid any run down of responses due to wash out during wholecell recordings (Apostolides and Trussell, 2013), although no difference in response was observed when using GABA internal solutions. No corrections were made for junction potentials. The pairs of neurons chosen for recordings were located within 50 -100 m of each other in the coronal plane and 10 -40 m in the rostrocaudal plane. To probe for connections during paired recordings, one cell was held in current-clamp mode and injected with a 5 ms, 600 -700 pA current pulse to evoke an AP. Meanwhile, the second (postsynaptic) neuron was held in voltage-clamp mode at Ϫ40 mV, well away from the chloride reversal potential (approximately Ϫ73 mV), given that neurons in the CeL are known to be GABAergic, forming inhibitory synapses (Sun and Cassell; Lopez de Armentia and Sah, 2004;Haubensak et al., 2010;Li et al., 2013). This protocol was repeated for at least 20 (but not Ͼ50) episodes, and sweeps were averaged for analysis. The same was then done in the opposite direction. Only connections with an amplitude of Ͼ5 pA were considered to be connected. Finally, in pharmacology experiments, bicuculline (10 M; Sigma-Aldrich) or CNQX (10 M; Tocris Bioscience) were bath applied to the slice.

Firing properties
APs were evoked using current injections applied in increments of 20 pA from Ϫ60 to 240 pA. AP threshold, amplitude, delay, half-width, rise time, and spike accommodation were analyzed off-line (described below). Spike accommodation was measured as the difference in AP frequency over at least eight APs at twice threshold. Although the two main firing types we observed ultimately had significantly different AP onsets, we used the absence or presence of spike accommodation to classify these firing types, as AP onset varied with small changes in holding membrane potential.

Data analysis Electrophysiological properties
Resting membrane potential (R m ) was recorded on-line immediately after break-in, whereas input resistance (R i ) was measured off-line as R i ϭ dVm/l, where dVm is the change in membrane potential in response to a Ϫ20 pA (800 ms) current injection (l). For connections, decay was measured by fitting the average IPSC by a sum of two exponentials (simplex sum of squared errors) to calculate a weighted time constant: w ϭ ͑t 1 ·a 1 ϩ t 2 ·a 2 ͒ / ͑a 1 ϩ a 2 ͒ .
Onset delay was calculated as the difference between the time of the presynaptic AP peak and the time of IPSC onset (time at 5% of peak). For firing properties, AP threshold was measured as the membrane potential at the start of the fast-rising phase. AP amplitude was measured from the threshold to peak, and delay was measured as the duration from the start of the current injection to the start of the fast-rising phase of the first AP.

Statistical tests
Datasets were tested for normality using the Shapiro-Wilks test. In the cases where a subset of the population was tested (e.g., drug application), we based our choice of statistical test on whether or not the overall dataset was normally distributed. We used parametric tests (t tests) when the data followed a normal distribution, whereas nonparametric tests (Wilcoxon and Mann-Whitney tests) were used for datasets that were too small to reliably test for or did not follow a normal distribution. Two-tailed tests were used unless otherwise stated, and differences were considered to be significant at p Ͻ 0.05.

Labeling for immunohistochemical characterization
For characterization of CeA neurons, mice were anesthetized by intraperitoneal injection of pentobarbitone sodium (3250 mg/kg; Virbac) and transcardially perfused with 40 ml of a 1% sodium nitrite solution (phosphate buffer, 0.1 M), followed by 40 ml of 4% paraformaldehyde (PFA; in 0.1 M phosphate buffer). Brains were then removed and left in 4% PFA at room temperature overnight and washed (3 ϫ 15 min, PBS 0.1 M) before sectioning (50 -60 m sections). Brains were placed in 30% sucrose for 48 h and sectioned using a sliding microtome (model SM200R, Leica). Coronal subsections (50 m) were then stained for PKC␦ using a mouse anti-PKC␦ antibody (1: 500; BD Biosciences), for SOM using a rabbit anti-SOM antibody (1:1000; Millipore Bioscience Research Reagents/ Millipore), and for NeuN using a chicken anti-NeuN antibody (1:1000; Millipore; 72 h at room temperature). In the case of virus-injected animals, fluorescence was amplified using either a rabbit anti-red fluorescent protein antibody (1:1000; Abcam) or chicken anti-green fluorescent protein (1:1000; Life Technologies). Sections were then washed and incubated with mouse-fluorophore 647 (for PKC␦; 1:2000; Invitrogen), rabbit-fluorophore 488 (for SOM; 1:2000; Invitrogen), rabbit-fluorophore 568, or chicken-fluorophore 488 (for fluorescence-enhanced sections; 1:2000; Invitrogen). Brain sections used for counts were immunolabeled for NeuN to allow reliable identification of mature neurons, and only NeuN(ϩ) neurons were counted. Cell counts were made in both the right and left hemispheres, but, because these were not significantly different, the data were pooled for each bregma location.

Post hoc labeling of recorded neurons
Alexa Fluor 568 (1 ng/ml internal solution) was added to the internal recording solution, and images of dendritic morphology were taken during recordings to correctly identify the presynaptic and postsynaptic cells after recovery of recorded neurons. Following electrophysiolog- CeA, central amygdala, which is divided into the CeL (orange) and the central medial amygdala (CeM, in white). Arrows show dorsal and medial orientation. Scale bar, 1 mm. Bottom panels show closeups of the CeL in 50 m sections that were stained for NeuN (to stain somas of neurons, white fluorescence), PKC␦ (green fluorescence), and SOM (red fluorescence). Scale bar in bottom left square, 100 m. For clarity, the merged panels represent the merging of PKC␦ and SOM only. The CeL is outlined in the bottom panel, and this outline was defined both by landmarks visible in bright field (data not shown), and the presence of PKC␦(ϩ) somas. PKC␦(ϩ) fibers can typically be seen in the CeM. The locations of both the BA and the CeM are also labeled in the merged panels, and note that by 1.80 mm the CeM is no longer present. The inset in the lower right corner of the far right merged panel shows a closeup of the most common cells types: PKC␦(ϩ)/SOM(Ϫ) (white arrowhead) and SOM(ϩ)/PKC␦(Ϫ) neurons (yellow arrowhead; scale bar, 10 m; PKC␦ green fluorescence, SOM red fluorescence, NeuN blue fluorescence). B, Only NeuN(ϩ) neurons were counted to ensure that only mature neuronal cells were taken into account. Of these, 48 Ϯ 5% were PKC␦(ϩ)/SOM(Ϫ) (mean n ϭ 83 Ϯ 19 neurons/1.0 ϫ 10 Ϫ3 mm 3 ), and 38 Ϯ 3% were SOM(ϩ)/PKC␦(Ϫ) (mean n ϭ 66 Ϯ 14 neurons/1.0 ϫ 10 Ϫ3 mm 3 ). These two populations were largely distinct as only 1 Ϯ 0.5% of neurons were PKC␦(ϩ)/SOM(ϩ) (mean n ϭ 2 Ϯ 0.3 neurons/1.0 ϫ 10 Ϫ3 mm 3 ), and 12 Ϯ 2% NeuN(ϩ) cells were PKC␦(Ϫ)/SOM(Ϫ) (mean n ϭ 20 Ϯ 4 neurons/1.0 ϫ 10 Ϫ3 mm 3 ). The dotted line on the graph indicates 50% and bregma-specific ical recordings, slices were fixed in 4% PFA (in 0.1 M phosphate buffer) for either 1 h at room temperature or overnight at 4°C, and then washed for 3 ϫ 15 min in 0.1 M PBS. Slices were then placed in blocking solution (1% BSA, 0.05% saponin, and 0.05% sodium azide) for 1-2 h at room temperature before incubation with an Alexa Fluor 555-bound streptavidin (overnight at room temperature; 1:2000 in blocking solution; Life Technologies). Slices were then washed (3 ϫ 15 min, 0.1 M PBS), mounted (DABCO), and imaged using either an upright fluorescent microscope (5ϫ and 20ϫ; Zen Software, Zeiss) or spinning disk confocal microscope (20ϫ and 40ϫ water-immersion objective, model #CSU-W1, Yokogawa; Slidebook software). All images were analyzed using FIJI (ImageJ). For protein PKC␦ staining, slices were subsequently embedded in 4% agarose and subsectioned (50 m sections; VT1000S vibratome, Leica) before being incubated with the PKC␦ mouse-antibody (72 h at room temperature; 1:500; BD Biosciences). Sections were then washed and incubated with mouse-fluorophore 647 (1:2000; Invitrogen), and the nuclei of the cells stained with DAPI, before being mounted and imaged as described above. Although PKC␦ clearly labeled somas, the somatostatin antibody did not deliver reliable post hoc staining, as a result of which we focused on PKC␦ for postrecording labeling experiments.
Animals were quarantined for 48 h then allowed to recover for at least 4 weeks postinjection. Brain slices were prepared as described above for electrophysiological experiments, and cells were only recorded well within the spread of the virus to ensure that nonfluorescent neurons were indeed SOM(Ϫ) rather than simply not infected. To verify the expression of channelrhodopsin (ChR2) and to activate ChR2 in infected cells, an LED system (470 nm, 1.4 mW; pE-2 LED System, CoolLED) attached to the microscope (via the rear C-mount port) was used. A prolonged light pulse (100 ms) was used to verify that cells expressed functional ChR2. In the case of AAV2/5-EF1␣.dflox.hChR2(H134R)-mCherry experiments, for example, neurons were considered SOM(ϩ) if they were both fluorescent and displayed a prolonged depolarization in response to prolonged light stimulation (470 nm, 100 ms), whereas a SOM(Ϫ) neuron was not fluorescent and showed no excitation to the light pulse. A light pulse of 2 ms (n ϭ 57 neurons) or 1 ms (n ϭ 10 neurons) was used to evoke responses in the CeL.

Characterization of neurons in the central lateral amygdala Immunohistochemical characterization
Neurons in the CeL have been separated based on the expression of a range of neuropeptides and markers that include PKC␦, SOM, corticotropin-releasing factor, oxytocin receptors, enkephalin, and others (Cassell and Gray, 1989;Haubensak et al., 2010). Of these, the two most highly expressed and clearly distinct neuropeptides are PKC␦ and SOM (Haubensak et al., 2010;Li et al., 2013). Immunostaining of brain sections from four locations posterior to bregma (Ϫ1.20, Ϫ1.40, Ϫ1.60, and Ϫ1.80 mm; Ϯ0.05 mm; Fig. 1A, top diagrams) shows that PKC␦ labeling within the amygdala was specific to the CeL, whereas SOM expression was also present outside the central amygdala. In the CeL, 48 Ϯ 5% of neurons expressed PKC␦, and 38 Ϯ 3% SOM (Fig. 1A), with the two populations largely nonoverlapping, and dual-labeled [PKC␦(ϩ)/SOM(ϩ)] neurons accounting for only 1.5 Ϯ 0.5% of neurons. The remaining neurons (13 Ϯ 2%) were negative for both markers. It was notable that whereas the proportions of PKC␦(ϩ)/SOM(Ϫ) and PKC␦(Ϫ)/SOM(ϩ) neurons were similar between bregma Ϫ1.40 and Ϫ1.60 mm, the difference between the total numbers of the two cell types changed at bregma Ϫ1.20 and Ϫ1.80 mm, the rostral and caudal limits of the CeL (Fig. 1B).

Electrophysiological properties
Based on their response to somatic current injections, three general types of CeL neurons have previously been described, with the two major types being LF neurons, which show a significant delay before onset of the first AP (ϳ100 -200 ms), and early-spiking (ES) neurons (also described as regular-spiking; AP onset, ϳ50 ms). A third, smaller population of low-threshold bursting neurons has also been described (Dumont et al., 2002;Lopez de Armentia and Sah, 2004;Haubensak et al., 2010;Li et al., 2013;Hou et al., 2016). We characterized the firing properties of 151 CeL neurons. However, while classifying neurons we found that AP onset varied with changes in holding potential, whereas the presence of spike frequency accommodation was more reliable. Using this measure, neurons were classified either as nonaccommodating (NA), where AP frequency remained relatively consistent (ϳ17 Hz), or accommodating (Ac), where there was clear spike frequency adaptation (AP 1-2 , ϳ32 Hz; AP 7-8 , 13 Hz; p Ͻ 0.001 Wilcoxon matched-pairs test; Fig.  2C). The large majority of our neurons were nonaccommodating (n ϭ 80 neurons; Fig. 2A) or accommodating (n ϭ 59 neurons; Fig. 2B). Nonaccommodating neurons also had a significantly longer mean onset compared with that of accommodating neurons (Table 1), and these neurons generally corresponded to the LF and ES types (Haubensak et al., 2010;Amano et al., 2012). Thus, for consistency we have termed these LF-NA and ES-Ac neurons. Apart New Research from resting membrane potential, which was significantly more depolarized in ES-Ac neurons, other membrane properties such as input resistance, threshold potential, AP amplitude, rise time, and half-width did not differ significantly between LF-NA and ES-Ac neurons (Table 1).
In the remaining 12 neurons (8%; Fig. 3), we found a distinct stuttering firing type that resembled that of some interneurons in the BLA (Woodruff and Sah, 2007;Sosulina et al., 2010;Spampanato et al., 2011). These neurons were easily distinguishable due to their distinctive firing pattern, with bursts of high-frequency APs (ϳ60 Hz;  (Table 1) and distinct firing pattern, large fast afterhyperpolarizations (as indicated by the red arrowhead) are also typical of this firing type. Inset shows a closeup of a spontaneous EPSP (sEPSP) in green. B, Overlay of the first AP of a stuttering (red), LF-NA (black), and ES-Ac (blue) neurons. The AP rise time and half-width of S neurons were significantly faster than those of LF-NA and ES-Ac neurons (Table 1). C, sEPSPs in S neurons were significantly more numerous than in LF-NA and ES-Ac neurons during the hyperpolarizing steps of this protocol. Numbers shown are the total counted over the Ϫ60, Ϫ40, and Ϫ20 pA current injections (B; S vs LF-NA: p ϭ 0.001, unpaired t test; S vs ES-Ac: p Ͻ 0.0001, unpaired t test). D, Example biocytin recovery of an S neuron, which was PKC␦(Ϫ) (top inset, yellow arrowhead indicates the soma of the S neuron). Scale bars: 20 m; top inset, 10 m; bottom inset, 5 m. This neuron displayed an extensive axon with inset showing a closeup of the axon in the dotted white square. E, Percentage of firing types for recovered neurons that were PKC␦(ϩ) (n ϭ 8) or PKC␦(Ϫ) (n ϭ 17). F, Shows total percentage of each firing type. Fig. 3A). Moreover, these neurons had significantly briefer APs with a half-width of 0.6 Ϯ 0.04 ms compared with 1.1 Ϯ 0.03 ms in ES-Ac neurons and 1.2 Ϯ 0.03 ms in LF-NA (Table 1; Fig. 3B). Stuttering neurons also displayed a higher frequency of spontaneous synaptic events compared with LF-NA and ES-Ac neurons (Fig. 3C). For stuttering neurons, we were unable to recover the entire cell; however, dendrites were filled, and visible, and showed that, unlike LF-NA and Es-Ac neuron, stuttering neurons were aspiny.
Twenty-five recorded neurons were successfully recovered with biocytin and labeled for PKC␦. Of these, PKC␦(ϩ) neurons (n ϭ 8) were either LF-NA or ES-Ac at equal incidence (50%), whereas PKC␦(Ϫ) neurons (n ϭ Figure 4. Neurons in the central lateral amygdala form local connections. A, Paired recordings were performed in the CeL, the location of which is shown in a diagram of a coronal slice (left). Middle, A bright-field image (300 m slice) of the area within the orange rectangle: the border of the CeL is clearly defined by visible fiber bundles, and the right panel shows the approximate outline of the three main amygdala regions: BLA, CeL, and CeM. In reality, the CeL extends slightly more ventrally than outlined here; however, we aimed to keep recordings within the outlined area to ensure that we did not mistakenly record from CeM neurons. B, C, Example traces of IPSCs, which were on average 20 Ϯ 3 pA, from a unidirectional connection (B) and a bidirectional connection (C). In each case, "cell 1" was current clamped and given a short current injection (5 ms, 600 -700 pA, illustrated in black directly under each current trace) to elicit one AP, while "cell 2" was voltage clamped at Ϫ40 mV. The protocol was then repeated in the opposite direction: from cell 2 to cell 1. Example average traces (black) and representative traces from single episodes (gray) are shown.  17) were more likely to be LF-NA (ϳ59%) than ES-Ac (ϳ23%). As previously described using Golgi methods (McDonald, 1982;Cassell and Gray, 1989), the majority of CeL neurons resembled medium-spiny neurons (see Fig.  5). Stuttering neuron somas that were successfully recovered and stained (n ϭ 3) were all PKC␦(Ϫ) (Fig. 3D,E). These results show that PKC␦-expressing (48%), and SOM-expressing (38%) neurons are the major cell types in the CeL, with very few neurons expressing both markers (1.5%). These neurons have one of two firing properties, LF-NA or ES-Ac. We also identified a previously unrecognized population of stuttering neurons (8%) that express neither PKC␦ or SOM (see below).

Local inhibitory connections
To determine the nature of local connections between neurons in the CeL, paired whole-cell recordings were made in acute coronal slices of wild-type mice (Fig. 4A). A total of 152 pairs were tested, of which 45 (29%) were connected. This was a monosynaptic connection with an onset latency of 0.85 Ϯ 0.06 ms after the AP peak and a high release probability (failure rate, 23 Ϯ 3%), which is consistent with a monosynaptic connection (Fig. 4B,C). At a holding potential of Ϫ40 mV, the IPSC had a mean amplitude of 20 Ϯ 3 pA, a 10 -90% rise time of 1.7 Ϯ 0.1 ms, and a decay time constant of 19.2 Ϯ 1.5 ms. Connections were predominantly unidirectional (n ϭ 42 of 45 connected pairs; Fig. 4B), with only 3 connected pairs displaying bidirectional connectivity (Fig. 4C,D). Apart from the stuttering cells, these neurons resembled medium-spiny neurons, (Fig. 5A-C), and spine density did not differ significantly between presynaptic and postsynaptic neurons (Fig. 5B); nor were differences observed in soma diameter, soma volume, number of primary dendrites, number of nodes, or total dendrite length ( Table 2). Recordings were made throughout the rostrocaudal extent of the CeL, and the resulting map of connected and unconnected pairs revealed no obvious location preference (Fig. 5D).
Neurons in the CeL are predominantly GABAergic, and in our connected pairs the IPSC reversal potential was Ϫ72 mV, which corresponds to the calculated chloride reversal potential (approximately Ϫ73 mV; Fig. 6A). Application of the GABA A receptor (GABA A -R) antagonist, bicuculline (10 M) blocked these IPSCs ( Fig. 6B; n ϭ 5 paired recordings), confirming that they were GABA A -Rmediated chloride currents. In current clamp, these connections were hyperpolarizing, with a mean amplitude of Ϫ1.1 Ϯ 0.3 mV (n ϭ 17), which is sufficient to halt firing in the postsynaptic cell ( Fig. 6C; n ϭ 5 paired recordings), and in some cases this inhibition was followed by a rebound increase in spike probability (Fig. 6D). These results demonstrate that neurons throughout the CeL form local inhibitory connections at a relatively high rate, which are capable of shaping the activity of the postsynaptic cell.

Distinct connection patterns exist between local CeL neurons
To determine the identity of recorded pairs, recovered neurons were processed using immunohistochemistry. As No PKC␦(ϩ) ¡ PKC␦(Ϫ) connections were found. Connected cells displayed a variety of discharge properties (Fig. 7F), with the most common connections being either LF-NA ¡ LF-NA (ϳ26%; n ϭ 5 of 19 paired recordings) or ES-Ac ¡ LF-NA connections (ϳ21%; n ϭ 4 of 19 paired recordings). Although less common, we also found ES-Ac ¡ ES-Ac connections (ϳ10%; n ϭ 2 of 19 paired recordings). Stuttering neurons were always presynaptic (n ϭ 3), with two connections to LF-NA neurons and one to an ES-Ac neuron.
These results show that local CeL connections occur between a variety of immunohistochemically and electrophysiologically distinct neuronal types with the most common connection between PKC␦(Ϫ) neurons. Given that ϳ75% of PKC␦(Ϫ) neurons are SOM(ϩ) (Fig. 1), we turned to a SOM-Cre mouse line to reliably identify and selectively activate SOM(ϩ) neurons in vitro. It was important to confirm that neurons considered to be PKC␦(Ϫ) were not false negatives due to protein washout during whole-cell recordings. To label SOM(ϩ) neurons, we injected an adeno-associated virus containing a DIO-td-tomato vector (AAV-DIO-tdTom) into the CeL of SOM-Cre mice (Fig.  8). SOM-tdTom and PKC␦ labeling in the CeL revealed proportions of these markers that were similar to those in wild-type mice (Fig. 8A,B; n ϭ 3 mice; at bregma, Ϫ1.40 to Ϫ1.60 mm). We also determined the firing properties of SOM(ϩ) and SOM(Ϫ) neurons (Fig. 8C). In agreement with recordings in wild-type mice, SOM(ϩ) neurons were mostly LF-NA (ϳ81%; n ϭ 13 of 16 neurons; ES-Ac: ϳ19%; n ϭ 3 of 16 neurons), whereas the SOM(Ϫ) neurons were mostly ES-Ac (ϳ65%; n ϭ 11 of 17 neurons; LF-NA: ϳ29%; n ϭ 5 of 17 neurons). Notably, the one stuttering neuron found in these recordings was SOM(Ϫ). Values are the mean Ϯ SEM. Four connected pairs (total of eight neurons) were recovered, and their morphologies were analyzed. When these properties were compared between presynaptic and postsynaptic neurons, no significant differences were observed (Mann-Whitney test).

Population-driven inhibition is greater between like neurons
Somatostatin-positive neurons As described above, paired recordings in coronal brain slices from both wild-type, and SOM-Cre mice show that connections were most frequent between somatostatin expressing PKC␦(Ϫ) neurons. However, previous studies indicate that the inhibition of SOM(Ϫ) neurons by SOM(ϩ) cells not only exists, but plays a key role in fear expression (Li et al., 2013;Hou et al., 2016). Such a motif is also suggested by inhibition of PKC␦(ϩ) neurons by PKC␦(Ϫ) neurons (ON neuron ¡ OFF neuron;Haubensak et al., 2010). One possibility for our low incidence of SOM(ϩ) ¡ SOM(Ϫ) connections is that we are sampling local connections (ϳ50 -100 m apart) in the coronal plane, and SOM(ϩ) ¡ SOM(Ϫ) connections may be more common among "distal" (i.e. Ͼ100 m) connections. To address this, we injected an AAV-containing DIO-channelrhodopsin-mCherry into the CeL of SOM-Cre mice (Fig. 9A,B) to directly activate SOM(ϩ) terminals.
Whole-cell recordings were made from SOM (ϩ) and SOM(Ϫ) neurons, and synapses made by SOM(ϩ) neurons were activated optically. All SOM(Ϫ) cells received input from SOM(ϩ) neurons with a mean IPSC of 162 Ϯ 24 pA (n ϭ 15; holding voltage, Ϫ40 mV; Fig. 9C). Next, paired recordings were made using a Cs-based internal solution, allowing voltage clamping of cells at the ChR2 reversal potential (ϳ0 mV) to test for SOM(ϩ) ¡ SOM(ϩ) connections. In this configuration, all SOM(Ϫ) and Ϫ40 to Ϫ90 mV) and average currentϪvoltage (I-V) curve of local IPSCs (right, n ϭ 5 paired recordings). This I-V curve is typical of a chloride current: a linear I-V relationship (r 2 ϭ 0.98) that reverses here at Ϫ72 mV, close to the theoretical reversal potential (ϳ73 mV). B, Local IPSCs were also blocked by the GABA A receptor antagonist bicuculline (10 M); example traces with aCSF in black and bicuculline in red (left). IPSCs were completely blocked by bicuculline (right, mean IPSC aCSF: 24.7 Ϯ 5.4 pA; mean IPSC bicuculline: 1.7 Ϯ 0.5 pA; n ϭ 5 paired recordings; p ϭ 0.03, one-tailed Wilcoxon test; dotted line joins data points from the same neuron). C, Overlay of 10 example traces from a connected pair where a short positive current injection (5 ms, 600 -700 pA) was applied to the presynaptic cell to fire one AP at t ϭ 0 s (top trace). Meanwhile, the postsynaptic cell was also in current-clamp mode, and current was injected such that the cell fired continuously (bottom trace). A single AP in the presynaptic cell evoked an IPSP that was sufficient to stop the postsynaptic cell from firing. Bottom histogram shows the number of APs fired in the above trace over time, in 50 ms bins. D, The spike probability was significantly lower in the 200 ms following inhibition onset compared with preinhibition (mean spike probability before inhibition, 0.14 Ϯ 0.02; mean spike probability during inhibition, 0.02 Ϯ 0.01; p ϭ 0.02, paired t test), and in most cases increased when the postsynaptic cells recommenced firing (mean spike probability before inhibition, 0.14 Ϯ 0.02; mean spike probability after inhibition, 0.2 Ϯ 0.02; p ϭ 0.01, paired t test). Each color represents data points from the same neuron (n ϭ 5 pairs). SOM(ϩ) neurons received large IPSCs when SOM(ϩ) terminals were activated [SOM(Ϫ) ϭ 22 neurons; SOM(ϩ) ϭ 10 neurons; Fig. 9D-F]. IPSCs in response to SOM(ϩ) terminal activation were fully blocked by bicuculline (10 m, n ϭ 5, Fig. 9G), reversed at approximately Ϫ67 mV (n ϭ 4), and were able to halt firing in the postsynaptic cell. From this cohort, 10 SOM(Ϫ) neurons were recovered, of which 5 were PKC␦(ϩ), showing direct SOM(ϩ) ¡ PKC␦(ϩ) and SOM(ϩ) ¡ PKC␦(Ϫ) connections (Fig. 9H). While all neurons received input from SOM neurons in the CeL, overall input to SOM(ϩ) neurons was significantly larger than to SOM(Ϫ) neurons (Fig. 9I). This difference is consistent with our paired recordings where five of eight SOM(ϩ) ¡ SOM(ϩ) pairs were connected, but none of the SOM(ϩ)/SOM(Ϫ) pairs were (n ϭ 16 pairs). In the course of these recordings, it was clear that, using SOM as a neuronal marker, a wide variety of connections are present in the CeL. Thus, for example, in one SOM(Ϫ) ¡ SOM(Ϫ) single connected pair (illustrated in Fig. 9J), both cells also received input from local SOM(ϩ) neurons.
Together with our connected paired recordings, these results are consistent with the presence of SOM(Ϫ) ¡ SOM(ϩ) and SOM(Ϫ) ¡ SOM(Ϫ) connections within the CeL. Furthermore, they suggest that, as with SOM(ϩ) neurons, a high proportion of CeL neurons receive inhibitory local connections from SOM(Ϫ) neurons, and with inhibition within the population being stronger than that between populations.

Discussion
The CeA is generally considered to be the main output nucleus of the amygdalar complex and is divided into the lateral and medial sectors. It contains GABAergic neurons that have been divided into several distinct populations using immunohistochemical and electrophysiological markers. These cells form local, as well as long-range connections, and different cell types have been associated with distinct functional roles (McDonald, 1982;Sun and Cassell, 1993;Jolkkonen and Pitkänen, 1998;Haubensak et al., 2010;Li et al., 2013). Here, using whole-cell paired recordings and optogenetics, we characterized neurons of the CeL and their intrinsic connections. We find that neurons in the CeL are extensively interconnected, with local connections apparent between all types of neurons, but strongest between like neurons. Moreover, we describe a new type of neuron in the CeL with distinct firing properties. These results highlight the complex intrinsic circuits within the CeL and suggest that particular cell groups identified using current methods, rather than mediating specific behaviors, participate in a range of different circuits.

Local networks in the CeL
Consistent with previous studies, we found that PKC␦ and SOM labeled two separate populations of neurons in the CeL (ϳ48% and ϳ38%, respectively), with very little overlap (ϳ1-2%), that account for 88% of the total cell population. In response to current injection, these neurons show two types of discharge patterns, late firing (LF-NA) and early spiking (ES-Ac), and their overall incidences (ϳ52% and ϳ39% respectively) were comparable to those previously described in the mouse (Haubensak et al., 2010;Hou et al., 2016). While SOM(ϩ) neurons were mostly LF-NA (ϳ81%) and SOM(Ϫ) neurons (largely PKC␦ expressing) were more likely to be ES-Ac (ϳ65%), these electrophysiological properties could not be used to separate the two populations. A smaller number of neurons (ϳ12%) were PKC␦(Ϫ) and SOM(Ϫ). These neurons may express CRF or one of the other peptides that are known to be present in CeL neurons (Cassell and Gray, 1989;Haubensak et al., 2010).
A small number of neurons (ϳ8%), had faster action potentials and a stuttering phenotype, with bursts of highfrequency AP discharge. This type of neuron has not been previously reported in the mouse CeL, although a similar "fast-spiking" neuron has been described in rare cases in the CeL and CeM of the guinea pig and cat (Martina et al., 1999;Dumont et al., 2002). These neurons were PKC␦(Ϫ) in wild-type mice, and the one stuttering neuron in SOM-Cre mice was SOM(Ϫ), suggesting that they may reflect a distinct PKC␦(Ϫ)/SOM(Ϫ) population. Although the role of this particular type of neuron is not clear, paired recordings showed that stuttering neurons were always presynaptic, and in cases where we had successful recovery of dendrites they had an aspiny morphology, different from that of the typically recovered CeL neurons. This, together with its fast-spiking properties, suggests the presence in the CeL of a local interneuron-like cell as opposed to the principal-type neurons typically found in the CeL.
Paired recordings demonstrated that neurons in the CeL were connected with an incidence of ϳ29%. In these recordings, we find that at the local level (ϳ50 -100 m in coronal slices), the most common connection was unidirectional and between two PKC␦(Ϫ) or two SOM(ϩ) cells.
we found that nearly all cells received a large input from both cell types. This difference in connectivity indicates that neurons make long-range connections within the CeL, perhaps in the rostrocaudal plane.
For the SOM neurons, using paired recordings, the monosynaptic connection had a mean amplitude of ϳ20 pA (at Ϫ40 mV), whereas when SOM neurons were transduced with ChR2, the optically driven IPSC had a mean amplitude of ϳ160 pA, showing that on average approximately eight SOM(ϩ) neurons innervate each SOM(Ϫ) neuron. In paired recordings, the IPSC had rapid rise times, suggesting that these contacts were likely to be somatic, or close to the soma (Delaney and Sah, 2001), which is consistent with the ability of these connections to halt spiking.

The CeL and behavior
The role of the CeL in cued fear expression is clear: a large body of data supports a model whereby conditioned stimulus-mediated disinhibition of CeM output drives conditioned fear Haubensak et al., 2010;Li et al., 2013). However, it remains unclear how the high level of CeL connectivity (both intra-CeL and extra-CeL afferents) can be reconciled with the increasing number of important behaviors in which CeL activity has been implicated. For example, fear expression has also been suggested to require activation of the parabrachial nucleus (PB) input to the CeL (Han et al., 2015;Sato et al., 2015), and yet this PB ¡ CeL circuit has also been implicated in appetite suppression (Carter et al., 2013;Cai et al., 2014). Meanwhile, other CeL circuits have been shown to underlie the switch between innate and conditioned fear (Isosaka et al., 2015), and anxiety generalization (Botta et al., 2015). Last, as well as forming local inhibitory connections (Li et al., 2013), SOM(ϩ) neurons are also projection neurons that target the PAG (Penzo et al., 2014), and this CeA ¡ PAG projection is engaged in mediating defensive behaviors (Tovote et al., 2016). We have shown that these neurons are also highly interconnected both within and between distinct neuronal populations. Our results suggest that within the CeL, neither cytosolic markers (PKC␦ and SOM) nor their electrophysiological properties alone can be used to identify cells engaged in particular behavioral roles.
The physiologic role, if any, of SOM and PKC␦ are not known; however, they clearly label separate populations of neurons in the CeL. Developmentally, the CeL has a striatal origin (Medina et al., 2011), and SOM and PKC␦, rather than specifying different populations that mediated different functional roles, should be thought of as lineage markers. We suggest that PKC␦-expressing and SOMexpressing neurons form heterogeneous populations of neurons, with different populations contributing to different behavioral outcomes. Understanding the flow of information through the CeA and its outputs, in a behaviorally specific and relevant manner, will be a challenge for future experiments. Similarly, it will be important to take these additional local circuits into account in further investigations of the CeL circuitry, particularly when judging the effects of pharmacological treatments during in vivo studies.