ASIC-dependent LTP at multiple glutamatergic synapses in amygdala network is required for fear memory

Genetic variants in the human ortholog of acid-sensing ion channel-1a subunit (ASIC1a) gene are associated with panic disorder and amygdala dysfunction. Both fear learning and activity-induced long-term potentiation (LTP) of cortico-basolateral amygdala (BLA) synapses are impaired in ASIC1a-null mice, suggesting a critical role of ASICs in fear memory formation. In this study, we found that ASICs were differentially expressed within the amygdala neuronal population, and the extent of LTP at various glutamatergic synapses correlated with the level of ASIC expression in postsynaptic neurons. Importantly, selective deletion of ASIC1a in GABAergic cells, including amygdala output neurons, eliminated LTP in these cells and reduced fear learning to the same extent as that found when ASIC1a was selectively abolished in BLA glutamatergic neurons. Thus, fear learning requires ASIC-dependent LTP at multiple amygdala synapses, including both cortico-BLA input synapses and intra-amygdala synapses on output neurons.

The amygdala is a complex structure in the mid-temporal lobe that plays a key role in fear learning and emotional processing [1][2][3] . Genetic variation in ACCN2, the human ortholog of the acid-sensing ion channel-1a subunit (ASIC1a) gene, is associated with both panic disorder and alterations in human amygdala structure and function 4 . Several lines of evidence implicate acid-sensing ion channels (ASICs) in the rodent amygdala in fear-like behavior. First, ASIC1a-null mice display impaired conditioned fear behavior and reduced innate fear 5,6 . Second, restoring ASIC1a in the basolateral complex of the amygdala (BLA) of ASIC1a-null mice rescues the fear memory impairments 7 . Third, overexpression of ASIC1a increases acquired fear-related behavior 8 . Finally, ASICs contribute to long-term potentiation (LTP) at cortico-BLA synapses, and this synaptic plasticity is critical for associative fear learning and memory 9 .
As proton-activated Na + -permeable channels, ASICs belong to the degenerin/epithelial Na + channel superfamily [10][11][12] . To date, seven ASIC subunits (ASIC1a, 1b, 2a, 2b, 3, 4, and 5; a and b refer to splice variants), encoded by five genes (ASIC1 to ASIC5), have been identified 13,14 . At least ASIC1a, ASIC2a, and ASIC2b are expressed in the central nervous system 11,[15][16][17] . Of these, ASIC1a is the obligatory subunit of the ASIC 5,18 . In the brain, ASICs consist of homotrimeric ASIC1a and heterotrimeric ASIC1a/2a and ASIC1a/2b channels 19,20 . Although ASICs are expressed widely in the brain, ASIC1a and ASIC2 are particularly enriched in the amygdala 9,20 . The amygdala comprises a heterogeneous collection of nuclei, including the BLA, the central amygdala (CeA), and intercalated cell masses (ICMs) 1,3,21 . Fear conditioning is known to cause widespread synaptic plasticity in the amygdala [22][23][24][25] . Despite the importance of ASICs in the cortico-BLA for fear learning, it is not clear whether the distribution of ASICs varies among different cell types within amygdala circuits, or whether synaptic plasticity in other cell types depends on ASICs and such plasticity also plays a role in fear learning.
To address these questions, we measured ASIC currents in different types of amygdala neurons in acutely-prepared brain slices from young adult mice. We found that ASIC expression was highly cell type-and region-specific, whereas ASIC gating kinetics and subunit composition were similar. Notably, ASICs were highly expressed in glutamatergic principal neurons (PNs) of the BLA receiving cortical and thalamic inputs, and low-threshold bursting (LTB) gamma-aminobutyric acid (GABA)-ergic neurons in the medial sector of the CeA (CeM), which is the output station of the amygdala circuitry. Analysis of synaptic plasticity revealed that the abundance of ASICs in postsynaptic neurons correlated with, and contributed to, LTP of glutamatergic synapses and fear-like behavior. Furthermore, ASIC-dependent LTP at both cortico-BLA and intra-amygdala glutamatergic synapses was required for the formation of fear memories.

Results
Cell type-specific expression of ASICs in the BLA. The amygdala is composed of several synaptically interconnected nuclei 2,3,26 . The borders of individual nuclei in a coronal section of the mouse brain were outlined, and neurons were identified using the neuronal marker NeuN (Fig. 1a, left). To differentiate between GABAergic and non-GABAergic neurons, we used knock-in (KI) mice expressing the green fluorescent protein (GFP) in glutamate decarboxylase 67 (GAD67)-positive neurons 27 (Fig. 1a, middle). Consistent with previous studies 2,21,26-28 , we found that the BLA consisted of a great majority (approximately 80%) of glutamatergic PNs and a minority of sparsely GABAergic interneurons (INs), whereas the CeA and ICMs were composed of mostly (approximately 90%) GFP-expressing cells (Fig. 1a, right).
To investigate the functional expression of ASICs, we used fast application of proton (H + ) to excised nucleated patches from recorded neurons (Fig. 1b). As previously described 19,29 , this method allows rapid application of proton, an almost ideal space-clamp, and a reliable measurement of current density. Solution exchange could be achieved in less than 200 μs (see Methods) with our fast application system. We found that BLA-PNs generated regular and accommodating action potential (AP) trains, in response to depolarizing current pulses in the current clamp (Fig. 1b, top). Nucleated patches from BLA-PNs exhibited ASIC-like, transient inward currents (Fig. 1b, bottom), in response to a submillisecond switch of extracellular pH from 7.4 to 5. The peak amplitude of the current depended on the pH value of the applied solution (Fig. 1c).
The ICM MV and ICM MD are clusters of GABAergic neurons between the BLA and CeA, receiving input from the BLA and sending projections to the CeL and CeM (Fig. 1a) 39 . On the other hand, the lateral ICM (ICM L ) is located on the dorsal-lateral side of the BLA, projecting inhibitory outputs to the BLA (Fig. 1a) 39 . We found that all ICM neurons showed similar electrophysiological properties (Supplementary Table S4) and exhibited modest ASIC currents similar to CeL neurons (ICM MV , 54 ± 8 pA/pF, n = 5; ICM MD , 44 ± 9 pA/pF, n = 6; ICM L , 54 ± 8 pA/pF, n = 5; Fig. 2k, l, m, and n).
In summary, the above functional mapping of ASIC current densities in diverse cell types of the amygdala network showed that ASICs were highly expressed in the major cell types of input (BLA) and output (CeM) regions of the amygdala (Fig. 2o).
Collectively, ASIC1a deletion caused a highly significant reduction in the magnitude of LTP in BLA-PNs and CeM-LTB neurons (Fig. 3e), both of which exhibited the highest ASIC current density, whereas this deletion had little effect on LTP in CeM-ES and CeL neurons that expressed relatively low levels of ASICs. This suggests that the level of ASIC expression contributed to the extent of LTP induction. In support of this idea, we found that the percentage of reduction in the LTP magnitude (Δ LTP%) in ASIC1a -/littermates, compared to WT mice, positively correlated with the ASIC current density observed in the postsynaptic neuron (Pearson's R = 0.528; P < 0.001, F-test; Fig. 3f). Thus, the differential expression of ASICs among amygdala neurons contributed to LTP at both cortico-BLA and intra-amygdala glutamatergic synapses.
ASIC1a deletion in GABAergic neurons selectively impaired LTP at BLA-CeM neuron synapses. Next, we investigated whether LTP at cortex→BLA-PN synapses was normal if ASICs were deleted in GABAergic neurons only. By crossing ASIC1a fl/fl mice with GAD65-Cre transgenic mice, we generated GABAergic neuron-specific conditional deletion of ASIC1a in their offspring. Indeed, normal ASIC currents were detected in BLA-PNs (Fig. 4a, top), whereas ASIC currents were eliminated in CeM neurons (Fig. 4b, top) in mice (GAD65 Cre/+ ; ASIC1a fl/fl ) with specific ASIC1a deletion in GABAergic neurons. In line with this finding, we also found that the application of HFS robustly induced LTP at cortex-BLA-PN synapses in mutant mice (GAD65 Cre/+ ; ASIC1a fl/fl ), which had the specific ASIC1a deletion in GABAergic neurons (225 ± 51%; n = 9 cells, 5 animals; P < 0.01, Wilcoxon signed-rank test; Fig. 4a, bottom). On average, the magnitudes of LTP at cortex-BLA-PN synapses were similar in WT mice and mutants (Fig. 4c). Finally, we tested whether selective deletion of ASIC1a in GABAergic cells, including amygdala output neurons, eliminated LTP in these cells. We found that HFS failed to induce significant LTP (115 ± 14%; n = 9 cells, 5 animals; P = 0.30, Wilcoxon signed-rank test; Fig. 4b, bottom) at BLA-CeM-LTB neuron synapses in mutants (GAD65 Cre/+ ; ASIC1a fl/fl ), similar to no LTP in mice with pan-neuronal ASIC1a deletion ( Fig. 3e; P = 0.60, Wilcoxon rank-sum test), but in contrast to marked LTP induction at the same synapse in WT mice (Fig. 4d).
Deleting ASIC1a in GABAergic neurons impaired conditioned fear learning. The impairment of fear learning seen in ASIC1a-null mice could be prevented by specific transgenic expression of ASIC1a in the amygdala 7 . The finding that disrupting ASIC1a in mice resulted in deficits in LTP at cortex→BLA-PN synapses has prompted the notion that ASIC-dependent LTP at the sensory inputs to the amygdala plays a critical role in conditioned fear learning 9 . In light of the present finding of ASIC-dependent LTP in the amygdala output neurons, we further examined whether deleting ASIC1a in GABAergic neurons also affected conditioned fear behavior. Using mice with either pan-neuronal ASIC1a deletion or a specific Pavlovian fear conditioning was used to test the role of ASICs in learned fear 1,2 . Mice were placed into the conditioning box (context A) and received five trials of tone-foot shock pairings. As illustrated (Fig. 5a), an initially neutral, auditory pure-tone conditioned stimulus (CS) was paired with a noxious unconditioned stimulus (US, an aversive footshock). During the training on day 1, we found that pan-neuronal ASIC deletion mice (Nestin Cre/+ ; ASIC1a fl/fl ) showed decreased CS/US-induced freezing responses, relative to WT littermates (Fig. 5b), suggesting impaired fear learning in ASIC1a-deletion mice. Interestingly, mice with specific ASIC1a deletion in GABAergic neurons (GAD65 Cre/+ ; ASIC1a fl/fl ) also exhibited impaired learning behavior similar to that found in pan-neuronal ASIC1a-deletion mice (Fig. 5b). On day 2, context-and cue-dependent fear memory was tested. We found that presenting either the training context or the cue (tone) elicited normal high-level freezing responses in WT mice (Fig. 5c, d). However, mice with either pan-neuronal ASIC1a deletion or specific ASIC1a deletion in GABAergic neurons showed similar, but significantly lower, freezing levels (Fig. 5c, d).  Finally, we tested the behavioral impact of selective deletion of ASIC1a in glutamatergic PNs of the BLA, using the Cre/loxP system. We found that bilaterally injecting an adeno-associated viral vector serotype 8 (AAV8), encoding GFP-Cre under the control of the CaMKIIα promoter (AAV8-CaMKIIα-GFP-Cre), into the BLA of ASIC1a fl/fl mice specifically eliminated ASIC currents in virus-transfected BLA-PNs, but not in non-transfected neurons (Fig. 5e) or neurons transduced with the fluorophore alone (Fig. 5f). Moreover, ASIC1a fl/fl mice that received the viral vector encoding GFP-Cre in the BLA exhibited lower freezing levels than mice injected with a viral vector encoding GFP alone during conditioning (Fig. 5g) and the recall test (Fig. 5h). In summary, these results indicate that ASIC1a expression in GABAergic neurons was as important as ASIC1a expression in glutamatergic neurons in supporting conditioned fear behavior.

Discussion
In this study, we characterized ASIC function in the amygdala at cellular, synaptic, and behavioral levels. Using ASIC1a -/and ASIC2 -/mice, we found that ASIC1a and ASIC2 were co-expressed in the amygdala neurons. Moreover, ASIC expression level was highly variable and cell type-specific. In support of the notion that ASIC is a synaptic modulator, the extent of LTP correlated positively with the ASIC current density in the postsynaptic neuron. Importantly, ASIC1a deletion in GABAergic neurons only eliminated LTP of intra-amygdala glutamatergic synapses onto CeM output neurons and reduced both contextual and cued fear memory to the same extent as that found for ASIC1a deletion in all neurons or selectively in BLA glutamatergic PNs. Thus, fear learning requires ASIC-dependent LTP at multiple amygdala synapses, including both cortico-BLA input synapses, as well as intra-amygdala synapses on output neurons.
Our recent study shows differential expression of ASICs in different cell types within hippocampal networks 19 . Somatic ASIC current density of oriens lacunosum-moleculare (O-LM) cells in the CA1 region, a classical type of dendrite-targeting IN, is 6 times greater than that of fast-spiking basket cells (BCs) in the dentate gyrus, a major class of soma-targeting IN. Similarly, ASIC currents in dendrites of O-LM cells evoked by local acid puffing are approximately 6-fold greater than those in BC dendrites at remote distances (up to 100 μm) from the soma 19 . Nevertheless, there is a general concern whether somatic ASICs are equally representative of those in the synaptic membrane. In this study, a detailed characterization of ASICs was made from patches excised from somata rather than dendrites. Therefore, it is important to be aware of this potential pitfall in the interpretation of the correlation between the extent of LTP and the ASIC current density in the postsynaptic neurons. Finally, people should keep in mind that the extent of ASIC-dependent LTP might be sensitive to the patterns of induction paradigms used at synapses. In this study, we used the LTP induction protocol for each synapse type according to previous studies (see Methods).
It is intriguing to note that GAD65 Cre/+ ; ASIC1a fl/fl mice show a similar extent of fear learning/memory deficit as Nestin Cre/+ ; ASIC1a fl/fl mice. The CeM is the final output station projecting to the brainstem areas 42 , which are responsible for freezing expression. The selective disruption of synaptic plasticity at the BLA→CeM-LTB synapses, but not cortex→BLA-PN synapses, in GAD65 Cre/+ ; ASIC1a fl/fl mice indicates a central role of the CeM in fear expression. Nevertheless, we are aware of the important role of BLA-INs in controlling fear learning 43 . Thus, we cannot exclude the possibility that ASIC1a deletion in GABAergic INs in GAD65 Cre/+ ; ASIC1a fl/fl mice causes abnormal inhibitory control in the BLA, given that ASIC expression levels are high in most BLA-IN subtypes. Finally, we speculate that similar changes in LTP induction can be observed in cortex-BLA-PN or BLA-CeM-LTB pathway induced by selective deletion of ASIC1a in glutamatergic neurons using the AAV-Cre/loxP system. However, if it is not the case, it implies that deletion of ASICs during early embryonic stages may disrupt synapse formation or alter the synaptic function during the development. If restoring ASIC1a with AAV-ASIC1a can rescues LTP in these two pathways in ASIC1a-null mice, it can exclude the essential role of ASIC during development.
The high level of ASIC expression in the primary input (i.e., BLA) and output (i.e., CeM) neurons of the fear circuitry is in accordance with the behavioral observation that ASIC1a deletion leads to deficits in fear acquisition and expression. In terms of fear expression, at least two types of CeM neurons project to two distinct brainstem sites: the periaqueductal gray and dorsal vagal complex 42 . With respect to brainstem-projecting neurons, whether there is a correspondence between the cell type (i.e., LTB versus ES) and their brainstem site remains unknown. On the other hand, information transfer from the lateral amygdala to the CeM is flexibly gated, depending on the specific pattern of environmental cues confronting the animal 44 . It is thought that the CeL and ICMs fulfill this function, as they receive glutamatergic inputs from the BLA and send GABAergic projections to the CeM 3 . As described above, ASICs are modestly expressed in CeL and ICM neurons. The role of ASICs in these neurons and their relative contribution to the acquisition of fear conditioning and expression of conditioned fear require further studies, for which cell type-specific gene disruption, as demonstrated here, would prove useful.
The gating properties of ASICs described here including desensitization and recovery time course (Supplementary Fig. S3) are similar to those of CA1 pyramidal neurons, which contain a mixture of homomeric ASIC1a, heteromeric ASIC1a/2a, and heteromeric ASIC1a/2b channels 15,19 . PcTX1 is known to inhibit ASIC1a homomers and ASIC1a/2b heteromers, but not ASIC1a/2a heteromers 33  homomers) in ASIC2 -/mice, which is approximately 20% of the total current in WT mice, exhibits a rapid cumulative reduction in responses to successive applications of proton. This phenomenon is termed tachyphylaxis 45 . Since only ASIC1a homomer exhibits prominent tachyphylaxis 45 , we speculate that PcTX1-sensitive component is primarily mediated by ASIC1a homomers. The difference in ASIC current densities in various cell types reflects the variation in the expression of ASIC subunits. In our study, since no difference in gating properties was found in ASIC currents recorded from all cell types ( Supplementary Fig. S3), the differential expression of ASICs in each cell type may involve the coordinated regulation of ASIC subunit expression, without altering the relative contribution by ASIC1a homomer, and ASIC1a/2b and ASIC1a/2a heteromers.
Two recent studies showed that ASICs participate in synaptic transmission in the amygdala and nucleus accumbens where they contribute to approximately 5% of excitatory postsynaptic currents 9,46 . ASICs are Na + -and Ca 2+ -permeable channels that are activated by extracellular acidosis 29 . We speculate an important role of proton-ASIC signaling during intense presynaptic stimulation. During basal transmission, synaptic pH drop is rapidly buffered by a variety of homeostatic mechanisms 47 . Massive extracellular pH reductions may occur during intense activity such as fear conditioning. The lower pH can generate larger ASIC currents, thereby boosting N-methyl-d-aspartate (NMDA) receptor function 18 . These effects could explain the involvement of protons and the requirement of ASICs for LTP induced by HFS. In this scenario, the strength of proton-ASIC signaling should be critical in determining the relative contribution of ASICs to learning, memory, and synaptic plasticity. Indeed, several lines of evidence support this view. First, ASIC1a and ASIC2 expression in the hippocampus is relatively low, compared to other brain regions 5,20 . Consistent with this, ASIC1a -/-, ASIC2 -/-, and ASIC1a/2 -/mice display normal spatial learning and memory 20,48 . Second, ASIC currents during synaptic transmission are not detected in the hippocampus. In keeping with this, LTP at CA3-CA1 synapses can be readily induced in ASIC1a -/mice 48 . Third, the observed effects of ASICs on synaptic plasticity, learning, and behavior are found in the amygdala and nucleus accumbens, the brain regions where ASICs are highly enriched 9,46 . Fourth, increasing or decreasing pH buffering capacity can bidirectionally modulate the magnitude of LTP at cortex→BLA-PN synapses 9 . Finally, in our study, the reduction in the magnitude of LTP at different glutamatergic synapses in ASIC1a -/mice, relative to WT controls, correlates positively with the abundance of ASICs in postsynaptic neurons (Fig. 3f). Collectively, our study also supports the view that postsynaptic ASIC is a positive regulator of associative fear learning and memory at both the synaptic and behavioral levels. Such knowledge could be of importance for future drug discovery and development.
A recent study finds that disrupting ASIC1a leads to an increase in spine density and changes in glutamate receptor function in nucleus accumbens, although the underlying mechanisms remain unknown 46 . Thus, it is worth to note that our study did not exclude the possibility that ASICs are involved in regulation of synapse structure and function in the amygdala neurons. Further experiments are needed to investigate whether ASIC is required for LTP in a cell-autonomous fashion. Several outstanding questions remain to be answered, including why ASIC is important for LTP at some certain synapses, and why ASICs are so enriched in several BLA-IN subtypes, but not in FSINs. Furthermore, it will be of interest to determine whether ASIC mediates or modulates synaptic plasticity in BLA-INs. Finally, the striking differences in the ASIC expression level between non-FSINs and FSINs raise general questions regarding the roles of ASIC in amygdala INs in fear learning and memory.

Methods
Brain slice preparation. Coronal brain slices (300 μm thick) were prepared from mice (19-75 days old for AISC characterization; 1-4 months old for LTP experiments and 3-5 months old for behavioral tests) of either sex on C57BL/6 genetic background, including WT, GAD67-GFP KI, ASIC1a -/-, and ASIC2 -/mice, using a vibratome (DTK-1000, Dosaka, Kyoto, Japan), as described previously 49 . Animals were killed by decapitation, in accordance with national and institutional guidelines, and all procedures were approved by the Animal Care and Use Committee of National Yang-Ming University. Floxed ASIC1a (ASIC1a fl/fl ) mice were crossed with Nestin-Cre and GAD65-Cre transgenic mice to generate ASIC1a gene deletion in their offspring 48 . ASIC2 -/mice 50 were obtained from Jackson Laboratory (Bar Harbor, ME, USA). To facilitate cell type identification, we used GAD67-GFP KI mice, in which GFP expression was driven by the GAD67 promoter 27 . Slices were sectioned in ice-cold cutting solution containing (in mM): 87 NaCl, 25 NaHCO 3 , 1.25 NaH 2 PO 4 , 2.5 KCl, 10 glucose, 75 sucrose, 0.5 CaCl 2 and 7 MgCl 2 . Following sectioning, slices were incubated in the cutting solution (oxygenated with 95% O 2 /5% CO 2 ) in a holding chamber at 34 °C for 30 min, and then at room temperature until used. During experiments, individual slices were transferred to a submersion recording chamber and were continuously superfused with oxygenated artificial cerebrospinal fluid (CSF) containing (in mM): 125 NaCl, 25 NaHCO 3 , 1.25 NaH 2 PO 4 , 2.5 KCl, 25 glucose, 2 CaCl 2 and 1 MgCl 2 . The recording temperature was 25 ± 3 °C. Neurons were visualized under the guidance of infrared differential interference contrast (DIC) optics and by their green fluorescence with epifluorescence illumination. Nucleated patch recordings. Nucleated patch recordings were made, as described previously 49,51 , using an Axopatch 200B amplifier (Molecular Devices, Sunnyvale, CA, USA). Pipette capacitance was compensated. Signals were low-pass filtered at 5 kHz (four-pole Bessel) and sampled at 10 kHz using a Scientific RepoRts | 5:10143 | DOi: 10.1038/srep10143 digitizer (Digidata 1440 A; Molecular Devices). Pulse sequences were generated by a Digidata 1440 A via pClamp 10.2 (Molecular Devices). Minor and major axes of nucleated patches were measured. It was assumed that nucleated patches were approximately ellipsoid 52 , and the membrane surface area was calculated using the following formula: The total membrane capacitance was determined from the surface area, using the value (1 μF/cm 2 ) of specific membrane capacitance 52,53 . Fast application of H + on nucleated patches, isolated from identified neurons, was performed, as described previously 19 . Fast application experiments were started 1-2 min after the patches were excised. Double-barreled application pipettes were fabricated from theta glass capillaries (2 mm outer diameter, 0.3 mm wall thickness, 0.12 mm septum, Hilgenberg GmbH, Malsfeld, Germany) and mounted on a piezoelectric-based solution switching system (MXPZT-300, Siskiyou, OR, USA). The time necessary for the complete exchange of solutions was determined using an open patch pipette by switching between Na + -rich and 10% Na + -rich solutions, which was 186 ± 14 μs (n = 3) by measuring 20-80% rise time of the junction potential change. ASIC currents evoked by 1 s pulses of H + were applied every 20 s, except for some pharmacological experiments where short pulses (50 ms) were used.
To Immunohistochemistry. For immunofluorescent labeling, brains were removed from GAD67-GFP KI mice which were transcardially perfused with 4% paraformaldehyde in 0.1 M phosphate buffered saline (PBS). Brains were further dehydrated in 15% sucrose for 24 h, and 30% sucrose for 24 h. Cryostat sections (20 μm thick) were rehydrated with PBS. Following a PBS wash, sections were permeabilized with 0.3% PBST (0.3 % triton-X-100 in PBS) for 30 min and incubated in 10% normal goat serum (NGS) for 2 h to block non-specific binding. To stain NeuN, which is a neuron-specific protein, sections were incubated in primary mouse anti-NeuN antibody (1:400; MAB377; Merck Millipore, Billerica, MA, USA) with 5% NGS in 0.3% PBST for 24 h at 4 °C. Following a PBS wash, sections were incubated in secondary antibody (goat anti-mouse Alexa Fluor 594; Life Technologies, Grand Island, NY, USA) with 2% NGS in 0.3% PBST for 2 h at 4 °C. Following a PBS wash, sections were mounted in VECTASHIELD HardSet Mounting Medium (Vector Laboratories, Burlingame, CA, USA) and viewed under a Zeiss Microscope Axio Observer A1 (Zeiss, Oberkochen, Germany).
Fear conditioning. Fear conditioning was performed in two different contexts. On day 1, mice were placed into the conditioning box (context A) for 2 min habituation and then received five trials of tone (80 dB, 20 s)-foot shock (0.6 mA, 2 s) pairings. A foot shock was given 2 s before the end of a tone. There was 1 min observation time between each pairing. To evaluate fear learning, we recorded the freezing time during habituation and between each trial. The context-dependent fear memory test was performed 24 h later by re-exposing the mice for 5 min to context A. One hour later, mice were placed into context B for 2 min, which was followed by three trials of 20 s tone plus 1 min observation time to evaluate the tone-dependent fear memory. We recorded the freezing time to context or tone during the total observation time. Freezing is defined as a lack of movement (>2 s) associated with a crouching posture, except for heartbeat and respiration. The freezing time is expressed as a percentage of the total observation time.
Virus injections. Mice (2-4 months old) were anaesthetized with isoflurane and placed into a stereotaxic frame (Stoelting, Wood Dale, IL, USA), with the mouths and noses covered by the anaesthetizing mask with constant air flow containing 1.5% isoflurane (air flow rate: 4 mL/min). The mice were placed over a homeothermic blanket (Panlab Harvard apparatus, Barcelona, Spain) to keep their body temperature constant (36 °C), with their eyes protected by ophthalmic gel during securing. The skull was surgically exposed using scissors and drilled over the desired coordinates. AAV8-CMV-GFP was generated in the laboratory of our collaborator Dr. Min-Hong Tai (National Sun Yat-Sen University, Taiwan). The transgene (pAAV-CMV-GFP), packaging (pLT-RC08), and helper (pHGTI-adeno1) constructs were gifts from Dr. Jeng-Shin Lee (Harvard Gene Therapy Initiative, Harvard Medical School, Boston, MA, USA). AAV2/8 was serotyped with AAV8 capsid proteins. The AAV8-CaMKIIα-GFP-Cre was purchased from University of North Carolina Vector Core (Chapel Hill, NC, USA). The virus was delivered through the craniotomy bilaterally using a 10 μL NanoFil syringe (World Precision Instruments, Sarasota, FL, USA) and a 35-gauge beveled metal needle. The injection volume (0.5 μL at each location) and flow rate (0.1 μL/ min) were controlled with a nanopump Controller (KD Scientific, Holliston, MA, USA). The needles were left in position for 10 min after injection. Injection coordinates, relative to Bregma, were: anteriorposterior, -2 mm; medial-lateral, ±3.4 mm; dorsal-ventral, -5 mm and -5.1 mm. Mice were allowed to recover for 3 weeks, following injection.
Hierarchical clustering analysis. For cell classification, unsupervised clustering 34,55,56 was performed, using squared Euclidean distances and Ward's method 36 . Electrophysiological properties of BLA INs and CeL neurons were tested for uniformity in their distributions. Variables with a nonuniform distribution (see Supplementary Fig. S1a, S1b, S1c, S1d, S2c, S2d, S2e, and S2f) were used for subsequent unsupervised clustering. Hierarchical clustering was operated, as follows. First, each neuron was transformed into a four-dimensional data point with variables. Before clustering, variables were first normalized into the range (0, 1) by performing Min-max normalization. The distance between data points represented the dissimilarity between them; closer data points have higher similarity. Next, all the data points were clustered by the following iterative procedure. First, each data point was assigned to a cluster; every cluster therefore contained only one data point. Second, the two closest clusters were merged into one cluster, such that there would be one fewer cluster. Third, the distance between the new cluster and each of the old clusters was determined. Fourth, steps two and three were repeated, until there was only one cluster left. Ward's method linkage rules 36 minimize the error sum of squares of any pair of cluster in step three. The pair of clusters with a minimum between-cluster distance was merged. The hierarchical clustering analysis was carried out using Free Statistics Software v.1.1.23-r7. (Wessa, P. Free statistics software, available at: http://www.wessa.net/) Data analysis and statistics. Data were analyzed and fitted with Clampfit 10.2 (Molecular Devices) and Prism 5.01 (GraphPad, La Jolla, CA, USA). The accommodating ratio is the maximal ratio of the mean of the last five inter-spike intervals (ISIs) divided by the mean of the first five ISIs under minimal current step injection (<400 pA) which could generate>10 APs per second. The spike delay is the latency of the first AP upon 1 s near-threshold depolarizing (rheobase) current step injection. The ISI ratio is calculated as the latter ISI divided by the previous ISI. The maximal coefficient of variation (CV) of ISI ratios was calculated from spike trains evoked by 1 s current step injection. The maximal mean firing rate is the maximal number of APs that could be generated by 1 s current step injection. Input resistance is defined by the ratio of steady-state voltage change/1 s hyperpolarizing current (-25 pA or -50 pA). For classification of BLA GFP-expressing cells, four features consisting of the accommodating ratio, maximal CV of ISI ratio, delay, and maximal mean firing rate were used to identify different cell types ( Supplementary Fig. S1). The desensitization time constant of ASIC current was obtained by fitting currents with the function shown below: A denotes the peak amplitude of current, τ represents the desensitization time constant, and C denotes the amplitude of steady-state current. The pH-response curve was fitted with the function, as shown below: 50 where A is the constant for the maximal effect, c denotes the concentration, EC 50 represents the half-maximal effective concentration, and n denotes the Hill coefficient.
For reversal potential measurements, data points of I-V relations were fitted with second-order polynomials, from which the interpolated potentials were calculated. The theoretical reversal potential of sodium channels (E Na ) was calculated, according to the Nernst equation:  57 . Values indicate mean±s.e.m. (standard error of mean); error bars in figures also represent s.e.m. Statistical significance among groups was tested using the nonparametric Kruskal-Wallis test. Where significant, pairwise comparisons, using the Wilcoxon rank-sum test, were carried out for each pair of groups. Statistical significance of linear regression was tested using the F-test. All tests were performed at the significance level (P) as indicated. The s.e.m. of reversal potentials were obtained by analyzing data of individual experiments separately.