Cell-type-specific control of basolateral amygdala neuronal circuits via entorhinal cortex-driven feedforward inhibition

The basolateral amygdala (BLA) plays a vital role in associating sensory stimuli with salient valence information. Excitatory principal neurons (PNs) undergo plastic changes to encode this association; however, local BLA inhibitory interneurons (INs) gate PN plasticity via feedforward inhibition (FFI). Despite literature implicating parvalbumin expressing (PV+) INs in FFI in cortex and hippocampus, prior anatomical experiments in BLA implicate somatostatin expressing (Sst+) INs. The lateral entorhinal cortex (LEC) projects to BLA where it drives FFI. In the present study, we explored the role of interneurons in this circuit. Using mice, we combined patch clamp electrophysiology, chemogenetics, unsupervised cluster analysis, and predictive modeling and found that a previously unreported subpopulation of fast-spiking Sst+ INs mediate LEC→BLA FFI.


Introduction
The ability of animals to learn to associate sensory stimuli with outcomes plays an important role in their survival. Specifically, animals learn to associate environmental stimuli with particular valences (rewarding or aversive outcomes). The BLA is key in the formation of these associations (Duvarci and Pare, 2014;Janak and Tye, 2015). Following a small number of presentations of a stimulus and valence information, such as a reward or threatening foot-shock, BLA excitatory PNs undergo plastic changes. For example, after a few trials BLA neurons fire selectively to novel odors that are informative about the outcome in an associative olfactory learning task (Schoenbaum et al., 1999). This plasticity is critically important as future presentations of the stimulus can then elicit robust firing in subsets of BLA PNs that output to downstream regions to guide goal-directed behavior (Beyeler et al., 2016;Janak and Tye, 2015;Quirk et al., 1995).
Accumulating evidence demonstrates that this learning induced-plasticity is input-specific (Kim and Cho, 2017;McKernan and Shinnick-Gallagher, 1997;Nabavi et al., 2014); however, the circuit mechanisms underlying this input specific plasticity remain unknown. Importantly, GABAergic inhibition regulates the ability of BLA PNs to undergo LTP with plasticity unable to occur in the presence of intact inhibition (Bissière et al., 2003;Krabbe et al., 2018). Further, local dendritic targeting GABAergic INs expressing Sst exert strong control over BLA learning processes in vivo: Sst + INs are normally inhibited for the duration of the sensory stimulus-valence pairing and their activation impairs learning (Wolff et al., 2014). In acute slice preparations, FFI exerts this inhibitory control over BLA plasticity (Bazelot et al., 2015;Bissière et al., 2003;Tully et al., 2007). In cerebral cortex and hippocampus, FFI is mediated by perisomatic targeting, PV + INs to ensure the temporal fidelity of PN action potential (AP) output (Glickfeld and Scanziani, 2006;Pouille and Scanziani, 2001;Tremblay et al., 2016). However, in BLA, FFI plays a dual role regulating PN AP output (Lang and Paré, 1997) and gating their plasticity (Bazelot et al., 2015;Bissière et al., 2003;Tully et al., 2007). Which IN population mediates FFI to control plasticity in BLA remains unclear: perisomatic targeting PV + INs are candidates as they mediate FFI in cerebral cortex and hippocampus (Glickfeld and Scanziani, 2006;Pouille and Scanziani, 2001;Tremblay et al., 2016), but anatomical studies suggest that Sst + INs may play this role in the BLA (Smith et al., 2000;Unal et al., 2014). Further, the role of Sst + INs in regulating BLA learning in vivo (Wolff et al., 2014) suggests a role in regulating BLA plasticity via FFI.
The LEC receives extensive innervation from the olfactory bulb (Igarashi et al., 2012) and piriform cortex (Johnson et al., 2000), and neurons in this area respond to odorants (Leitner et al., 2016;Xu and Wilson, 2012). Restricted firing in LEC suggests that it may play a role in modulating odor-specific, experience-and state-dependent olfactory coding (Xu and Wilson, 2012). LEC stimulation exerts an inhibitory effect on the olfactory input from the olfactory bulb to BLA and, to a lesser extent, piriform cortex (Mouly and Di Scala, 2006). Given that BLA is involved in encoding the motivational significance of olfactory cues in associative learning (Schoenbaum et al., 1999) and that the LECfiBLA synapse is plastic (Yaniv et al., 2003), the LECfiBLA circuit could be involved in plasticity during olfactory learning. In addition to its role in olfactory processing, the LEC plays an important role in multimodal sensory processing (Keene et al., 2016;Tsao et al., 2013) and the BLA is a major target of LEC efferents (McDonald, 1998). Finally, the LEC projects along the perforant pathway to engage the hippocampal trisynaptic circuit and receives input to its deep layers from CA1 (Neves et al., 2008). In turn, it is these deep layer LEC neurons that innervate the BLA (McDonald, 1998;McDonald and Mascagni, 1997). Taken together, this raises the possibility that the LECfiBLA circuit could be involved in sensory-valence learning across sensory modalities or for more multimodal natural stimulus information under hippocampal influence. Here we address the circuit mechanisms underlying BLA dependent learning by studying the role of distinct populations of inhibitory interneurons in mediating LECfiBLA circuit motifs.
Due to the relative scarcity of cortical afferent synapses onto BLA PV + INs (Smith et al., 2000), the role of BLA Sst + INs in regulating learning in vivo (Wolff et al., 2014), and the importance of FFI circuits in gating BLA LTP (Bazelot et al., 2015;Bissière et al., 2003;Tully et al., 2007), we hypothesized that Sst + INs provide FFI onto local BLA PNs. We tested this hypothesis by pairing patch clamp recordings in BLA with unsupervised cluster analysis, predictive modeling, and chemogenetic manipulations and found that a previously unreported subpopulation of fast spiking (FS) Sst + INs mediate FFI in the LECfiBLA circuit.

LEC afferents preferentially drive disynaptic FFI in BLA
To examine how LEC afferents, a major cortical source of polysynaptic inhibition to BLA (Lang and Paré, 1997), engaged BLA neuronal circuitry, we prepared acute horizontal slices of mouse brain containing both BLA and LEC and recorded synaptic responses of BLA PNs ( Figure 1A, B). A single stimulation of LEC elicited evoked EPSCs (EPSCs) and IPSCs (IPSCs) in PNs ( Figure 1B). EPSCs were blocked by glutamate receptor antagonists DNQX (20 mM) and D-APV (50 mM) but not by the GABA A receptor antagonist gabazine (gbz; 5 mM) ( Figure 1C, D; p = 6.72 Â 10 À4 , Kruskal-Wallis test). In contrast, IPSCs were blocked by either gbz or DNQX/APV ( Figure 1C, D; p = 6.45 Â 10 -4 , Kruskal-Wallis test), consistent with a monosynaptic glutamatergic nature of the EPSCs and a polysynaptic GABAergic nature of the IPSCs. Additionally, IPSC onset was delayed relative to EPSC onset by~3 ms ( Figure 1E; p = 0.023, paired t-test) consistent with prior reports of polysynaptic inhibitory circuits in BLA (Arruda-Carvalho and Clem, 2014;Hübner et al., 2014;Lucas et al., 2016). The two main archetypal circuit motifs that mediate polysynaptic inhibition are feedforward and feedback inhibition (Tremblay et al., 2016). In order to distinguish these motifs, we recorded from BLA PNs in response to a stimulus train of five pulses to the LEC at 20 Hz. If eIPSCs were the result of FFI alone, we would expect to see little to no AP activity in BLA PNs in response to stimulation. However, if the eIPSCs were a result of combined feedforward and feedback inhibition, we would expect to see the PNs fire in response to LEC stimulation. Recording from BLA PNs in current clamp (À60 mV), we found they rarely fired in response to stimulation ( Figure 1F,G). These data suggest LEC stimulation preferentially drives FFI in BLA PNs.
To determine whether our experimental setup recapitulates LEC suppression of BLA PNs, as found in vivo (Lang and Paré, 1997), we tested the effect of gbz on the firing of BLA PNs in response to LEC stimulation ( Figure 1F-H). However, GABA A receptor antagonism might increase the firing of BLA PNs by bringing them closer to AP threshold via membrane depolarization. To control for GABA A receptor blockade induced depolarization of the recorded PNs, we maintained the membrane voltage (V m ) of the PNs at À60 mV throughout these experiments ( Figure 1H; p=0.38, Wilcoxon signed rank test). gbz mediated blockade of FFI led to a significant increase in PN firing in response to LEC stimulation ( Figure 1F,G; p=0.031, Wilcoxon signed rank test). These findings confirm that the horizontal slice preparation contained a functionally complete LECfiBLA circuit supporting its use for the study of FFI.

Cell type specificity of the LECfiBLA circuit
We performed minimal stimulation experiments in the LECfiBLA circuit using whole cell patch clamp electrophysiology to examine putative unitary synaptic events between neurons (Gabernet et al., 2005;Kumar and Huguenard, 2001). Briefly, stimulation intensity was tuned to a threshold intensity where LEC stimulation elicits eEPSCs in BLA neurons with a~50% success rate and an all-or-none amplitude (Figure 2-figure supplement 1); at lower stimulation intensities EPSCs do not occur. As these EPSCs likely represent the response of the neuron to the activation of a single LEC neuron, this method provides a reliable measure of the unitary EPSC (uEPSC) between the stimulated and recorded neurons (Gabernet et al., 2005;Kumar and Huguenard, 2001). To identify dendritic-targeting Sst + and perisomatic-targeting PV + INs in acute brain slices, we generated Sst-tdTomato and PV-tdTomato mouse lines by crossing Sst-ires-Cre and PV-ires-Cre mice to Ai9 tdTomato reporter line. Sst-tdTomato mice showed reliable labeling of Sst + cells and low overlap with PV + cells and PV-tdTomato mice showed reliable labeling of PV + cells and low overlap with Sst + cells (
Since minimal stimulation likely reports the response of a BLA neuron to a single upstream LEC neuron (Gabernet et al., 2005;Kumar and Huguenard, 2001), we can use threshold stimulation intensity as an indirect measure of convergence of LEC projection neurons onto different BLA cell types. If threshold stimulation is lower for one BLA neuronal population compared to the others, it would follow that convergence of LEC inputs onto that cell type is likely greater relative to the other populations as it requires the activation of fewer LEC neurons. We found that threshold intensity was lower for the Sst + INs compared to other cell types ( Figure 2F; p = 3.33 x 10 À4 , Kruskal Wallis test).
Taken together, these data demonstrate that LEC stimulation leads to large and fast unitary currents in BLA Sst + and PV + INs and small and slow unitary currents in BLA PNs. Finally, although Sst + and PV + INs display equivalent uEPSCs, the finding that Sst + INs have a lower threshold stimulation intensity compared to PV + INs and PNs suggest that LEC afferents may have a greater functional convergence onto BLA Sst + compared to PV + INs.

A fast spiking phenotype distinguishes BLA Sst + INs targeted by LEC afferents
Though little is known about BLA Sst + INs, they appear to have diverse electrophysiological properties ex vivo (Krabbe et al., 2018;Sosulina et al., 2010) and responses to stimuli in vivo (Krabbe et al., 2018;Wolff et al., 2014), suggesting Sst may be expressed by a broad range of GABAergic IN subtypes, similar to cerebral cortex (Tremblay et al., 2016). Only a subset of Sst + INs responded to cortical stimulation raising the question whether these cells represented a distinct cell type. To address this hypothesis, we ran an unsupervised cluster analysis using Ward's method (Ward, 1963) based on 15 membrane properties from 105 Sst + INs. Applying Thorndike's procedure (Thorndike, 1953) suggested two distinct clusters (Groups I and II). A large majority of Sst + INs that responded to cortical stimulation clustered into Group I (14/16 cells) whereas a large majority of non-responsive Sst + INs clustered into Group II (17/19 cells) ( Figure 3A). To better understand what membrane properties best distinguished Group I and II Sst + INs, we used decision tree analysis (Breiman et al., 1984;Therneau and Atkinson, 1984) which returned maximum firing rate, hyperpolarization induced sag, and AHP latency as the most salient parameters for cluster separation (Figure 3-figure supplement 1). To reduce potential model overfitting with the decision tree analysis, we used an additional predictive modeling technique, the random forest method (Liaw and Wiener, 2002). Like decision tree analysis, the random forest method returned maximum firing rate and hyperpolarization induced sag in addition to AP halfwidth as the most salient parameters for defining BLA Sst + INs ( Figure 3B,C).
Having observed two subpopulations of Sst + INs, we compared membrane properties across Group I and II Sst + INs and PV + INs ( Figure 3D-F, Figure 3-source data 1). These data show that Group I Sst + and PV + INs differed from Group II Sst + INs in a subset of membrane properties such as maximum firing rate and hyperpolarization induced sag. Further, all three IN subtypes differed in AP halfwidth with PV + INs firing the fastest APs and Group I Sst + INs firing faster APs than Group II Sst + INs (see Figure 3-source data 1 for detailed statistics on membrane properties). The     In order to compare the morphology of the IN populations, we included biocytin in the patch pipette in a subset of recordings to allow for post hoc visualization of the different IN subtypes. Using these biocytin-filled neurons, we created reconstructions of the somatic and dendritic morphology of the INs ( Figure 3G; see Figure 3-figure supplement 2 for Group II Sst + IN data which were not included in formal analyses due to low number of recovered morphologies [n SstII = 3, 2]). When we compared the somatic and dendritic morphology of PV + and Group I Sst + INs, we found no significant differences in any measurement ( Figure 3H-J, see Figure 3-source data 2 for statistical analysis).
Taken together, the membrane properties of BLA Sst + and PV + INs reveal two distinct subpopulations of Sst + INs that are readily distinguished at the biophysical level by their FS phenotype and at the functional circuit level by synaptic responses to cortical stimulation. However, the IN subtypes do not appear to have any readily observable differences in their somatic and dendritic morphology. For clarity, we refer to the Group I and II Sst + INs as FS and non-fast spiking (nFS) Sst + INs, respectively. Source data 1. Differences in active and passive membrane properties among BLA Sst + and PV + INs. Source data 2. Table of statistical analyses used in Figure 3. Source data 3.  Figure 3F. Source data 6. Table of morphological data for Figure 3 and Supplements.  Probing the LECfiBLA circuitry suggests distinct functional feedforward/feedback roles for IN subtypes BLA Sst + INs have lower threshold stimulation intensity compared to PV + INs and PNs ( Figure 2F). Since these data suggest a higher rate of convergence of LEC afferents onto Sst + compared to PV + INs, we wanted to test the hypothesis that LEC input to BLA may preferentially recruit Sst + over PV + INs. To do this, we recorded from BLA Sst + and PV + INs in current clamp at rest and stimulated LEC with five pulses at 20 Hz. The stimulation intensity was set to the empirically derived threshold stimulation for BLA PNs (defined as the median PN threshold stimulation; 273.00 nC; Figure 2F) allowing us to determine how BLA INs respond when LEC activity is sufficient to ensure that PNs receive input. We found that stimulation led to robust spiking in BLA Sst + INs whereas PV + INs rarely fired ( Figure 4A-C; p=0.016, Mann-Whitney U test).
To probe the underlying mechanisms of the preferential spiking of Sst + INs, we used two additional measures: resting V m (V rest ) and subthreshold EPSP amplitude. We found that V rest was more depolarized in Sst + compared to PV + INs ( Figure 4D; p = 0.034, unpaired t-test). Because individual INs displayed variability in the number of APs/stimulus, each IN would have a different number of subthreshold EPSPs. Therefore, we compared both the distribution of all subthreshold EPSP amplitudes and the mean cellular amplitudes across Sst + and PV + INs. Regardless of the analysis method, we found that evoked subthreshold EPSPs were larger in Sst + compared to PV + INs ( Figure 4E; all subthreshold EPSP events: p = 2.19 x 10 À21 , Kolmogorov-Smirnov test; cellular subthreshold EPSP: p = 0.030, Mann-Whitney U test). Thus, the data show that two distinct mechanisms underlie the recruitment of BLA Sst + INs by LEC afferents: Sst + INs were more depolarized at rest compared to PV + INs, and, despite the concomitant decrease in the driving force through glutamate receptors, the magnitude of evoked subthreshold EPSPs was greater in Sst + compared to PV + INs. Finally, when we classified the Sst + INs from this experiment ( Figure 3A), we found that all but one (7/8 cells) were FS Sst + INs, consistent with our finding that LEC afferents appear to selectively target FS cells among the Sst + INs.
Taken together with the data in Figures 2 and 3, these data show that LEC stimulation leads to synaptic responses in BLA FS Sst + INs, PV + INs, and PNs but not in nFS Sst + INs. Although FS Sst + and PV INs have equivalent unitary events following LEC stimulation, the responses of these cell types to LEC input diverge in important ways. Specifically, FS Sst + INs have a lower threshold stimulation intensity compared to PV + INs, and FS Sst + INs have greater subthreshold EPSP amplitudes and are more likely to fire in response to PN threshold stimulation of LEC compared to PV + INs. These findings suggest that LEC afferents to BLA have a greater functional convergence onto FS Sst + INs compared to PV + INs and raise the question whether they are involved in BLA FFI.
In neocortical circuits, PV + INs mediate FFI whereas Sst + typically mediate feedback inhibition (Tremblay et al., 2016); nevertheless, our data point to a role for Sst + INs in BLA FFI, consistent with prior hypotheses that BLA Sst + INs mediate FFI and PV + INs mediate feedback inhibition (Duvarci and Pare, 2014). Thus, we wanted to test for responses of Sst + and PV + INs to local BLA PNs to probe for potential roles in feedback inhibition. To do this, we recorded PN-IN pairs in BLA, drove AP firing in PNs and compared uEPSC responses to the AP across IN subtypes ( Figure 4F-H). We found responses in all IN subtypes (8/10 FS Sst + , 5/5 nFS Sst + , 8/12 PV + INs with uEPSCs in response to PN APs) and found that BLA PV + INs responded to PNs with larger amplitude uEPSCs compared to either FS or nFS Sst + INs ( Figure 4H; p=0.025, Kruskal-Wallis test). Further, we found no difference in the latency or jitter of the uEPSCs across cell types (Figure 4-figure supplement 1). These data are consistent with a larger role for PV + INs in BLA feedback inhibition relative to Sst + INs.
Finally, we wanted to look at the level of spontaneous excitation onto the different IN subtypes as different levels of spontaneous glutamatergic activity could lead to different levels of basal excitability for the INs in the LECfiBLA circuit ( Figure 4I,J). We recorded spontaneous EPSCs (sEPSCs) across BLA IN subtypes. We found that PV + INs had larger amplitude sEPSCs than either Sst + IN subtype and that PV + and FS Sst + INs had more frequent sEPSCs than nFS Sst + INs ( Figure 4K   feedback inhibition, they both receive greater levels of spontaneous excitatory input compared to nFS Sst + INs.
To assess the role of Sst + and PV INs in BLA FFI, we stimulated LEC and recorded evoked EPSCs and IPSCs in BLA PNs using Sst-hM4Di and PV-hM4Di mice before and after bath application of CNO (10 mM). To determine the effect of IN inactivation on FFI we quantified two measures: evoked IPSC amplitude and inhibition-excitation balance (I/E balance). We defined I/E balance as the evoked IPSC amplitude for each cell normalized to its evoked EPSC amplitude. We examined this measure in addition to evoked IPSC amplitude to control for potential differences across cells with regards to the number of excitatory inputs to BLA activated by LEC stimulation. Consistent with the disynaptic nature of LEC-driven FFI in BLA (Figure 1), LEC stimulation elicited an IPSC that was significantly delayed relative to the EPSC in BLA PNs in both Sst-hM4Di and PV-hM4Di mice in control conditions and in the presence of CNO ( Figure 5B, see Figure 5-source data 1 for detailed statistical analysis). Importantly, these data indicate that hM4Di expression or CNO application does not alter the ability of LEC stimulation to drive FFI. When we perfused CNO to inactivate the different IN subtypes, CNO reduced IPSC amplitude and I/E balance in PNs from Sst-hM4Di mice by 30.2% and 40.2% respectively but had no effect on either measure in PNs from PV-hM4Di mice ( Figure 5C,D; p Sst-IPSC = 0.040, p Sst-IE = 0.013, Wilcoxon signed rank tests; p PV-IPSC = 0.35, p PV-IE = 0.85, paired t-tests). Finally, demonstrating the effects of Sst + and PV + IN inactivation were specific to FFI, bath application of CNO had no effect on eEPSCs in either mouse line ( Figure 5-figure supplement 1).
To control for off-target effects of CNO (Gomez et al., 2017), we repeated the experiments in wildtype (WT) littermates of Sst-hM4Di and PV-hM4Di mice. In WT littermates, LEC stimulation elicited the delayed EPSC-IPSC pairing consistent with FFI ( Figure 5B, see Figure 5-source data 1 for detailed statistical analysis), and CNO application had no effect on LEC-driven excitation or FFI in BLA ( Figure 5E, Figure 5-figure supplement 1, see Figure 5-source data 1 for detailed statistical analysis). Taken together with our data demonstrating that LEC afferents selectively synapse onto FS Sst + INs among the Sst + IN subtypes (Figures 2-3), these data show that FS Sst + , but not PV + or nFS Sst + , INs mediate LEC-driven FFI in BLA.

Discussion
Our data show that a previously unreported subpopulation of fast spiking Sst + INs mediates LECdriven FFI in BLA ( Figure 6). Although LEC afferents synapsed onto both PV + and FS Sst + INs (Figures 2 and 3), only FS Sst + INs fired in response to LEC activity ( Figure 4); conversely, PV + INs received stronger synaptic input from local BLA PNs than both populations of Sst + INs. Finally, inactivation of Sst + , but not PV + , INs led to a reduction in FFI onto BLA PNs ( Figure 5). Our findings are consistent with prior experiments in BLA indicating that Sst + INs form the anatomical circuit underlying FFI (Duvarci and Pare, 2014;Smith et al., 2000;Unal et al., 2014). In summary, these data illuminate a circuit mechanism for the feedforward inhibitory control mediated by entorhinal afferents onto BLA PNs (Lang and Paré, 1997). The role of BLA FFI in gating plasticity at afferent synapses onto PNs (Bazelot et al., 2015;Bissière et al., 2003;Tully et al., 2007) and of local Sst + INs in gating BLA-dependent learning in vivo (Wolff et al., 2014) suggests a potential role for this circuit in olfactory and other forms of sensory-valence learning (Keene et al., 2016;Kitamura et al., 2017;McDonald, 1998;Mouly and Di Scala, 2006;Schoenbaum et al., 1999;Tsao et al., 2013;Xu and Wilson, 2012).
BLA Sst + INs are a heterogenous population of dendritic targeting INs with diverse firing properties (Krabbe et al., 2018;Muller et al., 2007;Sosulina et al., 2010;Wolff et al., 2014). This is similar to what is observed in cerebral cortex (Tremblay et al., 2016). Accordingly, our cluster analysis and predictive modeling techniques revealed two distinct subpopulations of Sst + INs that could be distinguished based on membrane properties, most notably maximum firing rate, hyperpolarization induced sag, and AP halfwidth (Figure 3). These membrane properties suggest differential expression of HCN and K v 3 channels in the Sst + IN subpopulations (Poolos et al., 2002;Rudy and McBain, 2001). Consistent with this, K v 3.2 is expressed in a large subpopulation of BLA Sst + INs (McDonald and Mascagni, 2006).
How IN heterogeneity maps onto specific functions within the BLA is critical for understanding survival circuits. We observed that our cluster analysis based on intrinsic properties separated the populations of Sst + INs based on whether they displayed a synaptic response to LEC stimulation: 87.5% of Sst + INs with a synaptic response were classified as fast spiking and 89.5% of Sst + without a synaptic response were non-fast spiking. Although canonically PV + INs are thought to be synonymous with FS INs, our data indicate that, at least in BLA, this is not the case. In BLA, our data reveal two functionally distinct populations of FS INs can be separated based on PV or Sst expression. Further, taken together with recent descriptions of FS Sst + INs in cortex (Large et al., 2016;Ma et al., 2006;Naka et al., 2019;Nigro et al., 2018), we argue that a fast spiking phenotype is insufficient to classify an IN solely as PV + . Indeed, despite similar physiological phenotypes, in neocortex, FS Sst + INs have distinct functions from PV + INs. For example, whereas a well characterized circuit role of PV + INs is their control of thalamocortical driven FFI onto layer 4 PNs (Tremblay et al., 2016), a recent elegant study demonstrated that neocortical FS Sst + receive their input from layer 4 PNs and provide strong feedback inhibition to these PNs that is capable of silencing sensory evoked responses in the PNs (Naka et al., 2019). In cerebral cortex and hippocampus, a major circuit role of perisomatic PV + INs is to regulate AP output of PNs via FFI (Glickfeld and Scanziani, 2006;Pouille and Scanziani, 2001;Tremblay et al., 2016). Similarly in BLA, PV + INs provide strong regulation of local PN activity (Andrási et al., 2017;Veres et al., 2017;Woodruff and Sah, 2007b). Nevertheless, anatomical data indicates that BLA PV + INs receive more synaptic input from local PNs as compared to afferent sources (Smith et al., 2000). Our data demonstrating that PV + INs receive larger uEPSCs from local PNs compared to Sst + INs ( Figure 4) and other recent work showing that PV + INs fire APs more readily in response to local PN activation compared to cholecystokinin expressing INs (Andrási et al., 2017) provide a physiological compliment to the anatomy findings. Thus, PV + INs are positioned to provide feedback and lateral inhibition to control BLA spiking output whereas FS Sst + INs provide feedforward inhibition to regulate PN responses to sensory afferents to BLA. Together, these inhibitory networks would work in concert to fine-tune patterns of BLA activity to allow animals to learn about and react to stimuli. Our results illuminate an important point: archetypal circuit motifs that reappear across brain regions may be realized by different types of neurons in these brain areas depending on the functional role of the local circuit motif.
Although our data indicate that BLA PV + INs do not fire in response to LEC stimulation, afferent activation does lead to postsynaptic responses in these INs (Figures 2 and 4). When considered with our finding that local PNs preferentially target this IN subtype (Figure 4), this finding suggests an intriguing possibility for PV + IN functioning. Specifically, if afferents drive specific subsets of BLA PNs to fire, this afferent driven subthreshold depolarization of PV + INs would prime them to be more likely to fire in response to the subsequent PN activity. As BLA PV + IN firing is sufficient to prevent AP firing in postsynaptic PNs (Andrási et al., 2017;Veres et al., 2017;Woodruff and Sah, 2007b), this motif could provide an important source of lateral inhibition to select for activity in specific subsets of BLA PNs in response to sensory stimuli. Indeed, in vivo electrophysiological and Ca 2+ imaging studies in behaving animals indicate that sensory stimuli drive activity in specific subpopulations of BLA PNs rather than widespread activity across BLA PNs (Beyeler et al., 2016;Grewe et al., 2017;Schoenbaum et al., 1999).
Do the findings of the LECfiBLA circuit extend to other afferents to BLA (e.g. thalamic and other cortical inputs)? A recent study (Lucas et al., 2016) shows that PV + INs mediate polysynaptic inhibition in the lateral but not basal subnucleus of the BLA following stimulation of cortical (external capsule) or thalamic (internal capsule) fiber tracts to BLA. As stimulation of these fiber tracts leads to activation of local PNs and feedback inhibition (Szinyei et al., 2000), their results could be indicative of PV + INs mediating feedback inhibition in the lateral subnucleus of the BLA, as has been suggested previously (Duvarci and Pare, 2014). The horizontal slice preparation does not allow for unambiguous discrimination between the lateral and basal subnuclei of the BLA; therefore, we cannot determine the subnucleus within BLA where we conducted our recordings. Since the majority of LEC afferents target the basal subnucleus of the BLA (McDonald, 1998), it is conceivable that we recorded predominantly from neurons in this BLA subnucleus. If this were the case, this would provide an alternative explanation for the apparent discrepancy between our data and the data reported by Lucas and colleagues (Lucas et al., 2016); namely, that PV + INs provide FFI in the lateral subnucleus and Sst + INs provide FFI in the basal subnucleus. However, BLA PV + INs of both the lateral and basal subnuclei of the BLA receive relatively little input from a variety of cortical sources (Smith et al., 2000). Instead, cortical inputs target PV -/calbindin + INs throughout the BLA (Unal et al., 2014). Interestingly, the major population of PV -/calbindin + INs in BLA are Sst + INs (McDonald and Mascagni, 2002). Taken together these data suggest that our findings that FS Sst + INs mediate BLA FFI will likely hold across cortical inputs to the BLA. Nevertheless, future experiments will be necessary to test this premise.
What might be the functional role of BLA FS Sst + IN mediated FFI? In BLA, FFI plays a critical role in exerting control over the plasticity of PNs (Bazelot et al., 2015;Bissière et al., 2003;Tully et al., 2007). BLA Sst + INs primarily target the dendrites of PNs (Muller et al., 2007;Wolff et al., 2014). One function of dendritic targeting INs is to regulate the local Ca 2+ flux in dendrites (Chiu et al., 2013;Miles et al., 1996;Müllner et al., 2015). Recent work in cerebral cortex and BLA demonstrates that dendritic disinhibition via Sst + INs inhibition is critical in gating cortical pyramidal neuron plasticity ex vivo and in gating learning in vivo via regulation of learning dependent changes in pyramidal neuron activity (Adler et al., 2019;Williams and Holtmaat, 2019;Wolff et al., 2014). Specifically, sensory stimuli inhibit Sst + IN firing in BLA (Wolff et al., 2014), and further decreasing Sst + IN activity facilitates learning whereas increasing Sst + IN activity impairs learning (Wolff et al., 2014). Taken in the context of these findings, our data suggest an intriguing potential circuit mechanism explaining how this pattern of activity in vivo may lead to learning. Sensory stimulus mediated inhibition of Sst + INs would lead to a reduction in FFI and consequent dendritic disinhibition which could allow for plastic changes in the BLA PN to determine appropriate behavioral responses to the sensory stimulus in the future. Future experiments should test the hypothesis that Sst + IN mediated FFI of BLA PNs gates BLA plasticity and learning.
Taken together, the data presented in this study demonstrate the existence of a previously unreported population of FS Sst + INs in BLA. This population of INs is biophysically and functionally distinct from PV + and other Sst + INs. We show that these FS Sst + INs provide FFI at the corticoamygdalar synapse in contrast to nFS Sst + and PV + INs that do not engage in this circuit motif. Our data highlight the importance of probing IN heterogeneity within the context of local circuits as a key aspect for understanding circuit functioning. As FFI gates synaptic plasticity in BLA (Bazelot et al., 2015;Bissière et al., 2003;Tully et al., 2007), our data raise the possibility that the FFI mediated by this novel BLA population of FS Sst + INs serves as the underlying circuit mechanism for learning in the survival circuits of the amygdala.

Contact for reagent and resource sharing
Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Molly M Huntsman (molly.huntsman@ucdenver.edu).

Experimental model and subject details
All experiments were conducted in accordance with protocols approved by the Institutional Animal Care and Use Committee at the University of Colorado Anschutz Medical Campus. Slice electrophysiology experiments were conducted on mice aged postnatal days 35-70. Immunohistochemistry experiments were conducted on mice aged postnatal days 60-120. Experiments were conducted regardless of the observed external genitalia of the mice at weaning. To track estrous, vaginal swabs were collected from mice with vaginas that were used in experiments. The following mouse lines were used in the experiments: C57Bl/6J (Jackson Lab #000664), Sst-tdTomato, PV-tdTomato, Sst-hM4Di, PV-hM4Di, and wildtype littermates of the Sst-hM4Di and PV-hM4Di mice. Sst-tdTomato, PV-tdTomato, Sst-hM4Di, and PV-hM4Di mouse lines were generated by crosses of Sst-ires-Cre (Jackson Lab #013044) and PV-ires-Cre (Jackson Lab #008069) lines with either the Rosa-CAG-LSL-tdTomato-WPRE (Ai9; Jackson Lab #007905) or the R26-hM4Di/mCitrine (Jackson Lab #026219) lines. Please see Key Resources Table for more details on strain information.

Electrophysiology
Slices were placed in a submerged slice chamber and perfused with ACSF heated to 32-37˚C. Slices were visualized using a moving stage microscope (Scientifica: Uckfield, UK; Olympus: Tokyo, Japan) equipped with 4Â (0.10 NA) and 40Â (0.80 NA) objectives, differential interference contrast (DIC) optics, infrared illumination, LED illumination (CoolLED, Andover, UK), a CoolSNAP EZ camera (Photometrics, Tuscon, AZ, USA), and Micro-Manager 1.4 (Open Imaging, San Francisco, CA, USA). Whole cell patch clamp recordings were made using borosilicate glass pipettes (2.5-5.0 MW; King Precision Glass, Claremont, CA, USA) filled with intracellular recording solution. For voltage clamp experiments on FFI a cesium methanesulfonate (CsMe) based intracellular solution was used (in mM: CsMe, 120; HEPES, 10; EGTA, 0.5; NaCl, 8; Na-phosphocreatine, 10; QX-314, 1; MgATP, 4; Na 2 GTP, 0.4; pH to 7.3 with CsOH; osmolality adjusted to approximately 290 mOsm/kg). For all remaining voltage clamp experiments and for all current clamp experiments, a potassium gluconate based intracellular solution was used (in mM: potassium gluconate, 135; HEPES, 10; KCl, 20; EGTA, 0.1; MgATP, 2; Na 2 GTP, 0.3; pH to 7.3 with KOH; osmolality adjusted to approximately 295 mOsm/kg). A subset of the potassium gluconate recordings were supplemented with 0.2-0.5 % biocytin to allow for post-hoc morphological analysis. Access resistance was monitored throughout the experiments and data were discarded if access resistance exceeded 25 MW or varied by more than ± 20%. No junction potential compensation was performed. Data were acquired with a Multiclamp 700B amplifier and were converted to a digital signal with the Digidata 1440 digitizer using pCLAMP 10.6 software (Molecular Devices, Sunnyvale, CA). Data were sampled at 10 kHz and lowpass filtered at 4 kHz. Offline, current data were filtered using a 3 rd order Savistky-Golay filter with a ± 0.5 ms window after access resistance was assessed. Mean traces were created by first aligning all events by their point of maximal rise (postsynaptic currents) or by threshold (APs) and then obtaining the mean of all events; mean subthreshold EPSPs were not aligned prior to averaging.

Cell-type identification Principal Neurons (PNs)
PNs were targeted based on their large, pyramidal-like soma. Recordings were terminated if the physiology of the neuron was inconsistent with BLA PNs (e.g. high membrane resistance, narrow AP halfwidth, large and fast spontaneous EPSCs).

Interneurons (INs)
INs were targeted based on fluorescence in the Sst-tdTomato, PV-tdTomato, SST-hM4D i , and PV-hM4D i mouse lines. A 470 nm LED was used to identify mCitrine + INs in SST-hM4D i and PV-hM4D i mouse lines, and a 535 nm LED was used to identify tdTomato + INs in the SST-tdTomato and PV-tdTomato mouse lines (CoolLED, Andover, UK).
Pharmacology DNQX, D-APV, and gbz were purchased from Tocris Biosciences (Bristol, UK) and CNO was purchased from Enzo Life Sciences (Farmingdale, NY). DNQX stock was made at 40 mM and diluted to a final concentration of 20 mM in ACSF; D-APV stock was made at 50 mM and diluted to a final concentration of 50 mM in ACSF; GBZ stock was made at 25 mM and diluted to a final concentration of 5 mM in ACSF; and, CNO stock was made at 10 mM and diluted to a final concentration of 10 mM in ACSF. All stocks were stored at À20˚C and CNO was used within one month of making the stock solution.

Electrophysiology experimental design FFI, voltage clamp
The LEC was stimulated using a bipolar stimulating electrode (FHC, Inc, Bowdoin, ME, USA). Evoked EPSCs (V hold = À70 mV) and IPSCs (V hold = 0 mV) were recorded from BLA PNs in response to LEC stimulation. To assess the effects of different drugs on the EPSCs and IPSCs, ACSF containing DNQX, D-APV, gbz, and/or CNO was perfused onto the slice for five minutes prior to and continuously during the experiment. Effects of DNQX/APV and gbz on EPSCs and IPSCs were recorded using in an unpaired design where some PNs were recorded under control conditions and in the presence of DNQX/APV and gbz (given sequentially with time for washout) whereas others were recorded under control conditions in the presence of DNQX/APV or gbz. Effects of CNO on EPSCs and IPSCs were examined with a paired design where all PNs were recorded in both control conditions and in the presence of CNO.

FFI, current clamp
Membrane voltage of BLA PNs was recorded in response to 5 stimulations of the LEC at 20 Hz. I hold was adjusted such that V rest of the PNs was approximately À60 mV. To assess the role of GABA A receptor mediated inhibition on PN AP firing, gbz was perfused onto the slice for five minutes prior to and continuously during the experiment.

Current injections
Membrane voltage of BLA neurons was recorded in current clamp in response to a series of square hyperpolarizing and depolarizing current injections. Prior to initiation of the series of current injections, V m of the BLA neurons was adjusted to approximately À60 mV. Each cell was subjected to two series of 600 ms square current injections: À100 pA to +100 pA at 10 pA intervals and À200 pA to +400 pA at 25 pA intervals. The data collected in these experiments were used to determine active and passive membrane properties of the neurons.

Minimal stimulation
EPSCs (V hold = À70 mV) were recorded in voltage clamp in BLA PNs, Sst + INs, and PV + INs in response to LEC stimulation. Stimulation intensity was adjusted such that LEC stimulation resulted in recorded EPSCs having a success rate of approximately 50% and an all-or-none amplitude response.

Recruitment of BLA INs by LEC afferents
The median stimulation intensity necessary to observe putative uEPSCs in BLA PNs (273 mA Â ms) was used as the empirically derived PN threshold stimulation intensity. Membrane voltage responses of BLA Sst + and PV + INs were recorded in current clamp in response to 5 stimulations of the LEC at 20 Hz at the empirically derived PN threshold stimulation intensity.

Paired PN-IN recordings
Paired recordings were made between BLA PNs and nearby Sst + or PV + INs. We used a 2.5 nA, 2 ms current injection to drive a single AP in the PN (I hold adjusted such that V m » -60 mV. BLA PN APs were repeated at 0.25 Hz and the response of the IN (V hold = -70 mV) was recorded.
Spontaneous EPSC recordings sEPSCs were recorded for 5 min in INs (V hold = À70 mV) with no drugs in the bath.

Pharmacological effects of CNO on membrane potential
Membrane voltage of mCitrine + neurons in BLA was recorded in the presence of 20 mM DNQX, 50 mM D-APV, and 5 mM GBZ. To assess the effects of CNO on membrane potential, baseline V m was allowed to stabilize and was recorded for 3 min in the absence of CNO. Following recording of baseline V m , ACSF containing 10 mM CNO was perfused onto the slice and V m was recorded for an additional 10 min. V m was separated into 30 s bins. DV m was defined as the difference in V m between the mean V m during the 3 min of baseline recordings and the mean V m during the last 3 min of CNO application.

Definitions of electrophysiological parameters Evoked EPSC/IPSC detection and amplitude
EPSCs (V hold = -70 mV) were defined as negatively deflecting postsynaptic events that exceeded the mean baseline current (500 ms before stimulation) by 6 Â the median absolute deviation of the baseline current and that occurred within 20 ms of the end of the electrical stimulus artifact. IPSCs (V hold = 0 mV) were defined as positively deflecting postsynaptic events that exceeded the mean baseline current by 6 Â the median absolute deviation of the baseline current and that occurred within 20 ms of the end of the electrical stimulus artifact. To ensure that the detected EPSCs/IPSCs were related to the stimulus, we subsampled 25% of the sweeps in the experiment (or 5 sweeps if the experiment consisted of <20 sweeps) and found the maximal peak negative (EPSC detection) or positive (IPSC detection) deflection from baseline in the 20 ms after the stimulus artifact in those sweeps. Then, we found the median peak time for those sweeps and repeated the analysis over all sweeps in the experiment with a detection threshold of 6 Â the median absolute deviation of the baseline current and with a window set to ± 5 ms (EPSC detection) or ± 7.5 ms (IPSC detection) around the median peak time. The amplitude of each EPSC and IPSC was defined as the difference between the peak amplitude of the detected EPSC or IPSC and the mean baseline current for that sweep. The EPSC or IPSC amplitude for each cell in an experiment was defined as the mean of the amplitudes recorded from that cell (EPSC/IPSC successes only; failures were not included in EPSC/IPSC amplitude calculation). Where EPSC failure amplitude is reported on a per sweep basis, it was defined as the maximal negative deflection from the current trace within ± 5 ms of the mean current peak time for that cell. If all sweeps in an experiment were EPSC/IPSC failures, the mean EPSC/IPSC amplitude was defined as the maximal negative (EPSC) or positive (IPSC) deflection in the mean current trace that occurred within 20 ms of the end of the electrical stimulus artifact.

Success rate
Success rate was defined as 100% Â nsuccessful sweeps ntotal sweeps where successful sweeps were defined as sweeps where an EPSC was detected.
I/E balance I/E balance was defined as the ratio of IPSC to EPSC amplitude recorded in the same neuron.
EPSC/IPSC 20-80% risetime 20-80% risetime was defined as the time it took an EPSC or IPSC to reach 80% of its peak amplitude from 20% of its peak amplitude. 20-80% risetime was calculated for each sweep unless obscured by a spontaneous event. The risetime for each cell was defined as the mean of all risetimes recorded from that cell.

EPSC/IPSC latency and jitter
Latency of EPSCs and IPSCs was defined as the time between the end of the electrical stimulation or the peak of the PN action potential (in paired recordings) and the point of 20% rise for an EPSC or IPSC as calculated for the 20-80% risetime. The EPSC or IPSC latency for each cell was defined as the mean of the latencies recorded from that cell. The EPSC or IPSC jitter for each cell was defined as the standard deviation of the latencies recorded from that cell.

EPSC/IPSC t Decay
EPSC t Decay was determined using a single exponential fit, f t ð Þ ¼ Ae Àt=t . IPSC t Decay was defined as the weighted time-constant of IPSC decay. Briefly, a double exponential fit, f t ð Þ ¼ A 1 e Àt=t1 þ A 2 e Àt=t2 , was used to obtain the parameters to determine the weighted time-constant where t weighted ¼ t 1 A 1 þ t 2 A 2 ð Þ= A 1 þ A 2 ð Þ. t Decay was calculated using the mean EPSC or IPSC trace for a cell.
uEPSC detection and amplitude uEPSCs were defined as negative current deflections recorded in the IN that exceeded a detection threshold of 6x the median absolute deviation of the baseline current and occurred within 3ms of the PN AP peak during paired recordings. The amplitude of each uEPSC was defined as the difference between the peak amplitude of the detected uEPSC and the mean baseline current for that sweep. The uEPSC amplitude for each cell in an experiment was defined as the mean of the amplitudes recorded from that cell (successes only). sEPSC detection and amplitude sEPSCs were detected by a combined template and threshold method. Briefly, a template was made by subsampling 10% of local negative peaks exceeding at least 5 Â the median absolute deviation of a rolling baseline current (50ms prior to the peak). The template current was then truncated from its 20% rise point through the end of the decay time constant for the template current. Next, all local negative peaks exceeding 5 Â the median absolute deviation of a rolling baseline current (50ms prior to the peak) were collected. The template current was then scaled to each individual putative sEPSC peak and each sEPSC peak was assigned a normalized charge integral relative to the template. Finally, a normalized charge integral cutoff was chosen to exclude obvious noise/non-physiological events below a certain normalized charge integral. sEPSC amplitude was defined as the difference between the peak amplitude of each detected current and its corresponding baseline current. sEPSC for each cell was defined as the median peak amplitude for that cell. sEPSC frequency sEPSC frequency for each sESPC was defined as the inverse of the interevent intervals of the sEPSCs. The frequency measure for each neuron was defined as the median of the sEPSC frequencies for that cell.

Membrane resistance
Membrane resistance was defined as the slope of the best fit line of the I-V plot using the À100 pA to +100 pA (10 pA steps) series of current injections. Mean voltage response to each current injection step was defined as the difference between baseline mean membrane voltage (100 ms prior to current injection) and the mean membrane voltage during the 100 ms period from 50 ms after the start of the injection to 150 ms after the start of the current injection. This 100 ms window was chosen to allow for measurement of the change in V m after the membrane had charged and prior to any potential HCN channel activation. The I-V plot was constructed using all current steps below rheobase.

Maximum firing rate
Maximum firing rate was defined as the inverse of the inter-spike interval (ISI) during the first 200 ms of the most depolarizing current injection step before attenuation of AP firing was observed. Max FR was calculated using the À200 pA to +400 pA (25 pA steps) series of current injections.

AP threshold
AP threshold was defined as the voltage at which dV dt exceeded 20 V/s. AP threshold was calculated at the rheobase sweep of the -200 pA to +400 pA (25 pA steps) series of current injections.

AP amplitude
Amplitude of the AP was defined as the voltage difference between the peak of the AP and its threshold potential. AP amplitude was calculated at the rheobase sweep of the À200 pA to +400 pA (25 pA steps) series of current injections.

AP halfwidth
AP halfwidth was defined as the time between the half-amplitude point on the upslope of the AP waveform to the half-amplitude point on the downslope of the AP waveform. AP halfwidth was calculated at the rheobase sweep of the À200 pA to +400 pA (25 pA steps) series of current injections.
After-hyperpolarization potential (AHP) magnitude AHP magnitude was defined as the difference between the most hyperpolarized membrane voltage of the AHP (occurring within 100 ms after AP threshold) and AP threshold. AHP magnitude and latency data were calculated at the rheobase sweep of the À200 pA to +400 pA (25 pA steps) series of current injections. DAHP data were calculated at the rheobase + 50 pA sweep of the À200 pA to +400 pA (25 pA steps) series of current injections.

AHP latency
AHP latency was defined as the time from AP threshold and the peak of the AHP.

DAHP
DAHP was defined as the difference between the first and last AHP (DAHP ¼ AHP last À AHP first ).

AP phase plot
The AP phase plot was obtained by plotting the rate of change of the mean AP for each cell from the rheobase sweep of the À200 pA to +400 pA (25 pA steps) series of current injections as a function of the corresponding membrane voltage.
Latency to first AP AP latency was defined as the time from the initiation of the current injection to the peak of the first AP. AP latency was calculated at the rheobase sweep of the À200 pA to +400 pA (25 pA steps) series of current injections.

Firing rate adaptation ratio (FR adaptation)
Firing rate adaptation was defined as the ratio of the first and the average of the last two ISIs, such that Firing rate adaptation ¼ ISIfirst meanISIlast 2ISI . Firing rate adaptation was calculated at the rheobase +50 pA sweep of the -200 pA to +400 pA (25 pA steps) series of current injections.

AP broadening
AP broadening was defined as the ratio of the AP halfwidths of the first two APs (Broadening ¼ halfwidthsecond halfwidthfirst ). AP broadening was calculated at the rheobase +50 pA sweep of the -200 pA to +400 pA (25 pA steps) series of current injections.

AP amplitude adaptation
AP amplitude adaptation was defined as the ratio of the AP amplitude of the average of the last three APs and the first AP, such that AP amplitude adaptation ¼ meanAmplitudelast 3APs Amplitudefirst . AP amplitude adaptation was calculated at the rheobase +50 pA sweep of the -200 pA to +400 pA (25 pA steps) series of current injections.

Membrane decay t
Membrane decay t was determined by using a single exponential fit, f t ð Þ ¼ Ae Àt=t , to fit the change in V m induced by a À100 pA sweep in the À100 pA to +100 pA (25 pA steps) series of current injections.

Hyperpolarization-induced sag
Hyperpolarization-induced sag was calculated using the equation, VminÀVss VminÀVbl Â 100%, where V min was defined as the most hyperpolarized membrane voltage during the current injection, V ss was defined as the mean steady-state membrane voltage (last 200 ms of the current injection), and V bl was defined as the mean baseline membrane voltage (100 ms prior to current injection). Hyperpolarization-induced sag was measured from the -200 pA current injection.

Rebound spikes
Rebound spikes were defined as the number of APs in the 500 ms following the À200 pA current injection.

APs per stimulus
The number of APs per stimulus was defined as the number of APs occurring within 50 ms of the stimulus.

V rest
V rest was defined as V m (I hold = 0 pA) during a 500 ms baseline prior to LEC stimulation during the experiments testing the recruitment of BLA INs by LEC afferents.

Subthreshold EPSP amplitude
Subthreshold EPSP amplitude was defined as the maximal, non-stimulus artifact, voltage deflection within 40 ms after LEC stimulation.

Histological validation of PV + and Sst + IN identity
To perform the PV and Sst immunostaining, mice (n = 3 Sst-tdTomato mice, 3 PV-tdTomato mice) were sacrificed and transcardially perfused with ice cold 4% PFA (with 1.5% picric acid and 0.05% glutaraldehyde) followed by 30% sucrose protection. After the brain sank, coronal BLA slices of 30 mm thickness were obtained. Before application of blocking solution, slices were incubated for 30 minutes at room temperature with iFX-enhancer (Invitrogen, ThermoFisher Scientific, Waltham, MA, USA). After blocking, the slices were incubated with either rat anti-Sst antibody (1:100, MAB354, Millipore, Burlington, MA, USA) or guinea pig anti-PV antibody (1:500, 195004, Synaptic Systems, Gö ttingen, Germany) for at least 48 hours at 4˚C. Then, a secondary antibody of either donkey antirat or donkey anti-guinea pig Alexa Fluor 488 (1:500, Jackson ImmunoResearch) was applied overnight at 4˚C. The slices were imaged using a confocal laser scanning microscope (TCS SP5II, Leica Application Suite, Leica Biosystems, Buffalo Grove, IL, USA) with 10 Â 0.40 NA and 20 Â 0.70 NA dry objectives to determine the neuron identity.

Analysis of somatic and dendritic morphology
To determine morphological characteristics of biocytin filled neurons, stitched images were imported to Neurolucida (MBF Bioscience, Williston, VT, USA) to perform tracing. All analysis including sholl analysis (50 mM rings), dendrite branching, etc. was performed from traces using Neurolucida Explorer (MBF Bioscience).

Quantification and statistical analysis Statistical analyses
All data analysis (except decision tree and random forest analyses) were performed offline using custom written MATLAB code. Normality of the data was assessed using the Anderson-Darling test. For assessment of whether a single group differed from a normal distribution centered around zero, a one-sample t-test was used. For a test between two groups, a paired or unpaired t-test was used where appropriate. For tests between two groups of non-normal data, a Mann-Whitney U or Wilcoxon signed-rank test was used where appropriate. For tests between three or more groups of normal data with one independent variable, a one-way ANOVA was used with Tukey's post-hoc test to examine differences between groups. A Kruskal-Wallis test was used to examine differences between three or more groups of non-normal data with one independent variable. A Mann-Whitney U test was used as a post-hoc test following a significant result in a Kruskal-Wallis test and was corrected for multiple comparisons using the FDR method (Curran-Everett, 2000). The critical significance value was set to a = 0.05 or was set to a FDR-corrected value (a FDR ) for multiple comparisons. All statistical tests were two-tailed. Unless otherwise stated, experimental numbers are reported as n = x, y where x is the number of neurons and y is the number of mice. Statistical parameters are reported in the Results section and figure legends display p values and sample sizes.

Unsupervised cluster analysis
Unsupervised cluster analysis using Ward's method (Ward, 1963) was used to classify Sst + INs. Briefly, this method involves plotting each neuron in multidimensional space where each dimension corresponds to a given parameter. For our data, we plotted each Sst + IN in 15-dimensional space (where each dimension corresponds to a z-score transformation of one of the 15 membrane properties obtained from the current injection experiments; values were z-score transformed so that parameters with large values, for example membrane resistance, would not influence the cluster analysis more than those with small values, for example halfwidth). From here, the analysis proceeded along n -one stages where n is the number of Sst + INs. At stage n = 1, the two closest cells in the 15dimensional space are grouped together. At subsequent stages, the closest cells are grouped together until only one group of all objects remains. We determined the final number of clusters by using the Thorndike procedure (Thorndike, 1953) where large distances between group centroids at one cluster stage relative to other stages are indicative of significant differences between groups (see Figure 3A, inset).

Decision tree analysis
Recursive partitioning analyses (Breiman et al., 1984) were conducted in R (1.01.136). The R package rpart (Therneau and Atkinson, 1984) was used for recursive partitioning for classification. Decision trees were plotted using rpart.plot package, and random seeds were set using the rattle package. Predicted classes were determined from the unsupervised cluster analysis and were used to determine important discriminating parameters. The model was internally cross validated with a nested set of subtrees and the final model was selected as the subtree with the least misclassification error. Individual decision trees can suffer from overfitting. Therefore, the input parameters of the model were independently cross validated by bootstrapping 500 subsamples (without replacement and using random seeds). Each tree was pruned by choosing the complexity parameter with the lowest cross validated error. The mean correct prediction of the test set classifications (n = 20% of sample) by the pruned tree generated by modeling the training sets (n = 80% of sample) was 82.9%.

Random forest analysis
Supervised classification random forest was employed to reduce potential model overfitting. Random forest classification was conducted using the randomForest package in R (Liaw and Wiener, 2002). Random forest classification employs random subsampling of both parameters and bootstrapped subsamples of the dataset (with replacement). 10,000 decision trees were generated, and the modal tree was used as the final classification model. The out of bag estimate of the error rate of the model was 11.43%. An additional cross validation step was conducted by bootstrapping 80% of the initial sample to create the random forest model and tested against a subsample (n = 20% of sample; bootstrapped without replacement). The mean accuracy after 20 random forest runs was 90.7%. Gini impurity was calculated to determine parameter importance. Visualization of the frequency of individual samples falling within the same node across one run of the classification algorithm was obtained by calculating a proximity matrix for the samples and plotting it against the first two principle components.

Data display
Data visualizations were created in MATLAB and Adobe Illustrator. After analysis was completed, Neurolucida traces of soma and dendrites were thickened by 7 pixels (Figure 3) or 3-7 pixels (Figure 3-figure supplement 2) in Adobe Photoshop to improve visibility in figures. Normal data are presented as the mean ± s.e.m. Non-normal data are presented as the median with error bars extending along the interquartile range.