GABAergic microcircuitry of fear memory encoding

Abstract

The paradigm of fear conditioning is largely responsible for our current understanding of how memories are encoded at the cellular level. Its most fundamental underlying mechanism is considered to be plasticity of synaptic connections between excitatory projection neurons (PNs). However, recent studies suggest that while PNs execute critical memory functions, their activity at key stages of learning and recall is extensively orchestrated by a diverse array of GABAergic interneurons (INs). Here we review the contributions of genetically-defined INs to processing of threat-related stimuli in fear conditioning, with a particular focus on how synaptic interactions within interconnected networks of INs modulates PN activity through both inhibition and disinhibition. Furthermore, we discuss accumulating evidence that GABAergic microcircuits are an important locus for synaptic plasticity during fear learning and therefore a viable substrate for long-term memory. These findings suggest that further investigation of INs could unlock unique conceptual insights into the organization and function of fear memory networks.

Introduction

Fundamental to survival is the ability to learn from threatening experiences to avoid future harm. Fear conditioning is an influential model of this process that has delivered key insights into basic mechanisms of memory at the molecular, cellular and systems levels of the mammalian brain. In addition, it has informed our understanding of an array of conserved behavioral responses, such as avoidance, flight and aggression, that are underlying factors in numerous psychiatric conditions. In recent years, investigators have opened exciting new frontiers for establishing how these responses are orchestrated at the neural circuit level and, by extension, how the brain represents emotional information.

By most accounts, memories are formed as a result of enduring changes in neural processing (Herry and Johansen 2014, Clem and Schiller 2016, Bocchio et al. 2017, Krabbe et al. 2018, Lucas and Clem 2018, Maddox et al. 2019, Ressler and Maren 2019, Grundemann 2020, Sun et al. 2020, Yousuf et al. 2020). During Pavlovian fear conditioning, this occurs when an otherwise innocuous conditioned stimulus (CS, e.g. auditory tone) is encountered in conjunction with an innately aversive unconditioned stimulus (US, e.g. foot-shock), leading to the formation of a neural trace of CS-US pairing. When reencountered on its own, the CS has the potential to reactivate this trace and trigger preparation for imminent threat. For successful recall, therefore, the CS and US must be processed into signals that are amenable to encoding by neural circuits, and in a manner that facilitates their association at the level of individual neurons and/or assemblies. In addition, to encode a memory of this experience, a persistent change in the intensity or pattern of CS-evoked activity is required to transform the neural representation of a neutral CS into one that elicits a defensive behavioral response. A range of plausible mechanisms, including changes in synaptic efficacy, connectivity, or neuronal excitability, could underlie sculpting of neural pathways to generate such effects. Memory formation might therefore be understood by identifying the specific neurons undergoing these changes and establishing the functional consequences of their plasticity within the networks in which they are embedded.

In brain regions implicated in fear conditioning, such as the amygdala, hippocampus and prefrontal cortex, these networks are primarily populated by glutamatergic excitatory projection neurons (PNs) and GABAergic INs (INs). Largely for historical reasons, PNs figure more prominently in our conceptual understanding of memory. This is in part because PNs out-number GABAergic INs by roughly 4 to 1, and as a consequence they are more amenable to electrophysiological recordings, as well as over-represented in histological and biochemical preparations. By comparison, not only are GABAergic INs far less abundant, but compared to PNs they are more heterogeneous in terms of their gene expression, morphology, connectivity, and electrophysiology (Kepecs and Fishell 2014, Wamsley and Fishell 2017, Huang and Paul 2019). While selective analysis of GABAergic subtypes once proved intractable, however, methodological barriers are vanishing due to the proliferation of viral and genetic technology with exquisite cell-type specificity.

Just as GABAergic INs have eluded detailed investigation, PNs have also attracted relatively more attention in fear conditioning because early studies suggested that they have properties favorable to memory encoding. For example, compared to GABAergic INs, PNs were found to exhibit higher stimulus-specificity and narrower receptive fields, potential computational advantages in cue encoding and discrimination (Azouz et al. 1997, Swadlow and Gusev 2002, Usrey et al. 2003, Niell and Stryker 2008). Because they project broadly throughout the brain, PNs are also ideally positioned to relay their computational output to downstream effectors, whereas GABAergic INs give rise to few if any long-range projections. Finally, PNs express Hebbian synaptic plasticity, which many consider to be the most viable cellular mechanism for memory formation. Hebbian long-term potentiation (LTP) relies on subcellular processes that are largely lacking in GABAergic INs, including NMDA-receptor-mediated Ca²⁺ entry, activation of Ca²⁺/ calmodulin-dependent protein kinase II (CaMKII), and growth of dendritic spines (Mahanty and Sah 1998, Malenka and Bear 2004, Matta et al. 2013, Andersen et al. 2017). However, evidence suggests that many GABAergic INs are more stimulus-selective than previously assumed (Hirsch et al. 2003, Runyan et al. 2010, Runyan and Sur 2013, Li et al. 2015, Khan et al. 2018), and like PNs they express a range of synaptic plasticity mechanisms that might be conducive to trace formation (Mahanty and Sah 1998, Perez et al. 2001, Woodin et al. 2003, Pelkey et al. 2005, Chen et al. 2009, Le Duigou and Kullmann 2011, Huang et al. 2013, Owen et al. 2013, Cohen et al. 2016, Friend et al. 2019), raising the important question of how they contribute to memory processing.

Clarifying the role of GABAergic INs requires contemplation of factors unique to this cell class, foremost among these being that INs signal predominantly through local synaptic connections. In order to influence brain networks and ultimately modulate behavior, GABAergic INs must therefore cooperate with neighboring PNs. Inhibitory connections onto PNs play important roles in gain control and rhythmic entrainment of PNs, which are essential for regulating the dynamic range and temporal patterning of PN activity (Karnani et al. 2014, Bocchio et al. 2017). However, this is not the whole story, because the existence of synaptic connections between INs paradoxically endows these cells with the capacity to generate disinhibition (i.e. inhibition of inhibition) (Pfeffer et al. 2013, Pi et al. 2013, Letzkus et al. 2015). Indeed, studies suggest that disinhibitory interactions between INs play important roles in stimulus gating at the level of PN firing (Letzkus et al. 2011, Lee et al. 2013, Pi et al. 2013, Xu et al. 2013, Courtin et al. 2014, Wolff et al. 2014, Krabbe et al. 2019, Xu et al. 2019, Cummings and Clem 2020), and they may rely on organization of IN subtypes into functional hierarchies based on interconnectivity (Pfeffer et al. 2013). The degree to which these hierarchies vary across cytoarchitectural boundaries or within functional subnetworks of the brain remains unclear, but is a potential source of additional complexity in the functional logic of microcircuits. Nonetheless, it is clear that GABAergic INs participate in an assortment of inhibitory and disinhibitory circuit motifs with largely unknown consequences for stimulus coding.

In this review, we discuss emerging evidence that genetically-defined GABAergic INs make important contributions to cue processing and memory formation in fear conditioning. We devote particular attention to how these cell types interact with both PNs and INs to orchestrate phasic changes in PN firing underlying memory encoding and retrieval, as well as the impact of learning-induced IN plasticity on the function of memory-related networks.

Functional specialization of INs

While they share many molecular and anatomic features, GABAergic INs are perhaps best known for comprising discrete subclasses identifiable based on gene expression, morphology and anatomical connectivity. These subclasses originate, to a large extent, from embryonic progenitor populations in the ganglionic eminences, after which they rely on genetic programming and site-specific cues to complete their migration throughout the forebrain and subsequent terminal differentiation (Bandler et al. 2017). At different stages of this process, the expression of transcription factors, peptides and calcium-binding proteins effectively delineates INs that will share unique functional attributes. While no single gene commands their functional destiny, INs expressing different markers tend to exhibit key differences in microcircuit properties.

Because the contribution of a single neuron to a network computation is determined to a large extent by its input and output, one important way that INs can be functionally differentiated is by their specific pattern of synaptic connections. The synaptic targeting of parvalbumin (PV-INs) and somatostatin-expressing INs (SST-INs) is a canonical example of such specificity, wherein axon terminals from these subtypes preferentially contact the somatic versus dendritic compartments of excitatory PNs, respectively (DeFelipe et al. 2013). This anatomical bias positions PV-INs to exert profound influence over PN spiking, an effect even more pronounced among a subset of PV-INs, the chandelier cells, that target the axon initial segment (Veres et al. 2014). By contrast, Martinotti-type SST-INs synapse predominantly onto dendritic segments of cortical PNs, where they can modulate synaptic and dendritic physiology, but have a more indirect effect on spiking (Yavorska and Wehr 2016). Dendritic bias similarly characterizes SST-IN populations in the frontal cortex (Ali et al. 2020), hippocampus (Katona et al. 1999) and basolateral amygdala complex (Muller et al. 2007). Models emphasizing these features risk oversimplifying the function of PV- and SST-INs, particularly given the extreme molecular and anatomical diversity of SST-INs (Yavorska and Wehr 2016), but have been important in advancing a theoretical perspective of GABAergic function based on functional circuit motifs.

Along with postsynaptic targeting, the source from which an IN derives its excitatory input plays an important role in dictating its function. Within cortical networks, dendritic arbors of INs adopt a variety of morphologies extending across laminar and columnar dimensions, enabling them to sample specific categories of axon terminals. In addition to this spatial organization, the proportion of dendritic inputs originating from local versus long-range axons determines whether they engage in feedback versus feedforward inhibition of local PNs, respectively. In sensory cortex, feedforward motifs are a hallmark of PV-INs, particularly in thalamocortical processing (Bruno and Simons 2002, Cruikshank et al. 2007, Cruikshank et al. 2010), whereas cortical SST-INs are thought to participate more readily in feedback as well as lateral inhibition of intracortical excitation (Tan et al. 2008, Yu et al. 2019). However, these specializations do not necessarily apply to every system or functional subpopulation.

For example, while PV-INs in the lateral nucleus of the amygdala receive potent excitation from thalamic and cortical afferents, and in turn generate feedforward inhibition, those populating the basal nucleus are predominantly driven by local excitatory inputs (Smith et al. 2000, Lucas et al. 2016). Interestingly, SST-INs appear to play roles complementary to those of PV-INs in the basolateral complex, where they engage in feedback and feedforward inhibition in the lateral and basal nuclei, respectively (Smith et al. 2000, Guthman et al. 2020, Unal et al. 2020). In agranular frontal areas, PV-INs and cholecystokin (CCK)-positive basket cells are strongly activated by afferent pathways from the amygdala, hippocampus and mediodorsal thalamus (Delevich et al. 2015, McGarry and Carter 2016, Canetta et al. 2020, Cummings and Clem 2020, Liu et al. 2020). However, in addition to eliciting PV-IN spiking, claustral afferents generate overwhelming feedforward inhibition of frontal cortex in part through recruitment of neuropeptide Y (NPY)-containing INs (Jackson et al. 2018). Because NPY-INs co-express SST, this implies that subsets of SST-INs may specialize in different circuit motifs. In addition, prefrontal afferents generate relatively strong excitation of SST-INs as a whole, sufficient to drive spiking of these cells during stronger stimulus regimes (McGarry and Carter 2016, Cummings and Clem 2020). Therefore, while neuropeptide expression is a useful tool for differentiating feedback versus feedforward motifs, it is a not a universal marker of these processing modules.

Due to the hyperpolarizing effect of GABA on excitatory PNs, particularly in the mature brain, conceptual thinking about INs has historically centered around their role in constraining PN activity. Operating through feedforward and feedback mechanisms, as well as through related motifs like lateral inhibition, INs strictly regulate the number of PNs that fire in response to a given stimulus, the duration of their activity, and even the precise timing of individual action potentials. Failure of these mechanisms can disrupt the delicate balance of excitation to inhibition (E:I) that PNs require to operate within a meaningful range. In addition to enforcing such stability, dynamic adjustment of E:I is sometimes required to prohibit or enable specific forms of processing according to changing behavioral demands. To this end, INs play time variant roles in gating the flow of excitation through neural pathways and networks. While PN synapses are critical for these effects, accumulating evidence indicates that INs are also extensively connected with one another, providing additional substrates for complex and sometimes counterintuitive outcomes of GABAergic transmission.

Homotypic interactions among INs are mediated by both chemical and electrical synapses, which can reinforce tight coupling of membrane potentials among a sparse population of cells and occur preferentially between genetically-defined subclasses (Connors and Long 2004, Hestrin and Galarreta 2005) (although this specificity has been challenged (Hatch et al. 2017)). By synchronizing neuronal firing, both chemical and electrical synapses facilitate the entrainment of gamma oscillations by PV-INs and may play a similar role in SST-INs during beta/ slow-gamma activity (Bartos et al. 2007, Hu and Agmon 2015, Chen et al. 2017). Oscillations organize signaling within connected networks of PNs to fulfill a requirement in information coding for the precise timing of action potentials (Duzel et al. 2010, Likhtik and Paz 2015, Bocchio et al. 2017). As such, the prevalence of different oscillation frequencies is thought to reflect different modes of network operation, which may ultimately depend on the relative activity of genetically-defined INs. However, this is not the only way that cooperation between INs can lead to the reallocation of network resources.

During the past decade, precise targeting and manipulation of INs has revealed extensive interconnectivity between different subtypes (Pfeffer et al. 2013, Pi et al. 2013, Letzkus et al. 2015). The mere existence of these connections suggests that INs can through heterotypic inhibitory interactions generate disinhibition. Behavioral conditions may in part dictate which INs dominate these interactions, for example due to differences in the recruitment of different subtypes by the available stimuli. In addition, several studies indicate that interconnectivity varies in strength, which may contribute to a functional hierarchy that specifies which IN subtypes mediate disinhibition as well as the specific GABAergic populations that they control (Pfeffer et al. 2013, Pi et al. 2013, Walker et al. 2016, Cummings and Clem 2020).

A canonical example of a cell class that specializes in disinhibition is the neocortical vasointestinal peptide-expressing IN (VIP-IN), which inhibits SST- and PV-INs more strongly than it does PNs (Pi et al. 2013). Salient behavioral stimuli, such as those conveying feedback about environmental rewards and threats, elicit phasic firing of VIP-IN in the auditory cortex and basolateral amygdala (Pi et al. 2013, Krabbe et al. 2019). VIP-IN activation in turn amplifies the responsiveness of PNs to auditory cues, which may facilitate cellular plasticity mechanisms underlying cue learning. A similar role in learning has been proposed for the more enigmatic inhibitory cell types that occupy neocortical layer 1, where release of acetylcholine leads to disinhibition of underlying PNs (Letzkus et al. 2011). However, these well-known examples of disinhibitory control likely represent the tip of the iceberg, particularly given emerging reports that the most abundant IN subtypes, PV- and SST-INs, suppress the firing of other GABAergic populations in behaving mice (Xu et al. 2019, Cummings and Clem 2020, Dudok et al. 2021). Disinhibition fundamentally expands the computational repertoire of these cells by enabling their activation, in addition to inactivation, of downstream elements, bolstering their capacity to rapidly reorganize large local networks.

Network inhibition and disinhibition in fear memory processing

By virtue of their dense and complex wiring, INs have a range of mechanisms at their disposal for modulating the rate and temporal patterning of PN activity, which are critical for the processing of fear associative cues (Lucas and Clem 2017, McKenzie 2017, Artinian and Lacaille 2018, Lamsa and Lau 2019). One of the more consequential impacts of local GABAergic transmission is that it restricts suprathreshold responses to afferent excitation through feedforward, feedback and lateral inhibition of PNs, imposing a blanket of silence on the majority of the network. However, a small number of PNs that escape suppression can participate in stimulus coding and are an eligible substrate for experience-dependent plasticity. According to computational theory, this reflects precisely the type of network that is optimized for storage of many independent associations while avoiding catastrophic interference (Marr 1971, Kanerva 1988). Indeed, empirical studies support the idea that fear memory is encoded by a discrete PN ensemble, commonly referred to as an “engram”, whose size and composition has important consequences for memory strength and specificity (Semon 1921, Josselyn et al. 2015, Tonegawa et al. 2015, Denny et al. 2017, Josselyn and Tonegawa 2020)

A number of strategies have been developed to visualize and manipulate engram cells by utilizing regulatory genetic elements of immediate early genes (IEGs), such as c-fos and arc, to drive the expression of fluorescent reporters along with chemo- or optogenetic actuators (Barth et al. 2004, Reijmers et al. 2007, Guenthner et al. 2013, Kawashima et al. 2013, Denny et al. 2014). Experiments utilizing these “tagging” approaches have suggested that only a small proportion of neurons distributed across the brain are active during memory encoding, and that reactivation of these populations accounts for behavioral correlates of memory expression such as freezing (Liu et al. 2012, Ramirez et al. 2013, Denny et al. 2014, Tanaka et al. 2014, Cai et al. 2016, DeNardo et al. 2019, Lacagnina et al. 2019). Allocation to a engram population is in part determined by properties intrinsic to an individual PN, such that cells that are more excitable due to differences in membrane conductances are more likely to become integrated into the engram (Han et al. 2007, Han et al. 2009, Zhou et al. 2009, Yiu et al. 2014, Cai et al. 2016, Park et al. 2016, Rashid et al. 2016). However, recent studies indicate that microcircuit interactions between INs and PNs also play an important role in specifying the size of an engram population as well as its segregation from other engrams.

During learning, competition amongst PNs sculpts memory traces by constraining the total number of neurons active at any time (Josselyn and Frankland 2018, Rao-Ruiz et al. 2019). In the hippocampus, this is largely attributed to lateral inhibition, a phenomenon first observed in electrophysiological recordings (Sloviter and Brisman 1995, Hirase et al. 2001) and later implicated in memory allocation by in vivo optogenetic manipulations. For example, photostimulation of a small population of dentate granule cells suppresses the activity of neighboring granule cells during contextual fear conditioning, while silencing a similar proportion of granule gyrus cells disinhibits the firing of surrounding populations (Stefanelli et al. 2016). Correspondingly, photoinhibition of dentate gyrus SST-INs during learning increases the size of memory-related ensembles and improves recall of context fear, while stimulation of SST-INs produces opposite effects (Stefanelli et al. 2016). A comparable constraint on engram size exists in the lateral amygdala, where reduction of PV-IN activity during auditory fear conditioning increases the number of PNs recruited by CS-US pairing (Morrison et al. 2016).

Aside from modulating engram size, recruitment of INs at the time of learning might prevent certain information from reaching PNs that might otherwise corrupt the learning process. For example, in vivo imaging experiments indicate that SST-INs in the dorsal hippocampus are robustly activated by aversive foot shocks, responses that are attributable septal cholinergic input (Lovett-Barron et al. 2014). Pharmacogenetic inhibition of SST-INs, but not PV-INs, prevents contextual fear acquisition and increases US-elicited activity of PNs, consistent with the idea that SST-IN activity is required to exclude the integration of US features into the context representation, which could impair conditioning (Fanselow et al. 1993). Indeed, US-responsive SST-INs presumably comprise oriens lacunosum moleculare (OLM) cells, which are known to gate dendritic activity of CA1 PNs in response to CA3 and entorhinal inputs (Leão et al. 2012). However, the role of OLMs in learning may vary as a function of their molecular phenotype or position along the hippocampal dorsoventral axis, as activation of OLMs expressing the nicotinic receptor α2 subunit in the intermediate hippocampus impairs fear memory acquisition (Siwani et al. 2018).

Beyond initial learning, regulation of engram size by surround inhibition may have important consequences for memory specificity and flexibility. In the lateral amygdala, auditory fear conditioning promotes the transient accumulation of parvalbumin-positive axon terminals contacting lateral amygdala PNs (Rashid et al. 2016), potentially augmenting surround suppression by the newly formed engram. During this period of heightened inhibition, conditioning with a novel CS generates a second engram that overlaps considerably with the first, causing the two memories to become linked at the cellular and behavioral levels. The regulation of engram reactivation by surround inhibition may also have important consequences for memory reconsolidation, a process in which memory is susceptible to updating with new information following its retrieval-induced destabilization. For example, experiments that perturb competition between PNs result in aberrant recruitment of CA1 populations during recall of cocaine-place memories (Trouche et al. 2016). These newly activated cells become incorporated into the engram, resulting in the loss of a reward-like response to the drug-paired context. Thus, evidence suggests that dynamic competition between PNs, mediated by IN-PN interactions, determines not only the size but also the composition, specificity and stability of memory-related ensembles.

In addition to direct monosynaptic interactions, INs modulate PN activity indirectly through disynaptic disinhibition. Precisely-timed relief from inhibition can theoretically serve many functions during memory processing, such as gating PN firing, synchronizing PN ensembles, or modulating the excitatory potential of incoming stimuli (Bartos et al. 2007, Kepecs and Fishell 2014, Letzkus et al. 2015, Lucas and Clem 2017, Artinian and Lacaille 2018, Krabbe et al. 2018). Genetically-defined IN subpopulations were first implicated in such effects in auditory fear conditioning, where they exert control over PN processing in the basolateral amygdala as well as auditory cortex (Letzkus et al. 2011, Wolff et al. 2014). During conditioning, a majority of basolateral amygdala PV-INs exhibits increased firing during CS presentation, which is in turn associated with reduced firing of SST-INs (Wolff et al. 2014). Synaptic connections between PV- and SST-INs account for these inversely correlated firing patterns, which ultimately result in disinhibition of PN dendrites. However, at US onset both PV- and SST-IN firing are suppressed, relieving inhibition along the entire somatodendritic axis. Two distinct forms of disinhibition thus amplify CS-evoked potentials as well as US-evoked firing of PNs, which may facilitate Hebbian strengthening of CS inputs underlying memory formation. Consistent with this notion, optogenetic manipulation of PV- or SST-IN responses leads to bidirectional changes in learning (Wolff et al. 2014). More recent work extends the circuit diagram underlying US-evoked disinhibition by implicating shock-responsive VIP-INs in suppression of PV- and SST-INs (Krabbe et al. 2019), which similar to their counterparts in the neocortex are the principal target of VIP-INs (Pfeffer et al. 2013, Pi et al. 2013).

In primary auditory cortex, a similar disinhibitory motif is activated by an aversive US and contributes to encoding of fear responses to a complex auditory CS (frequency-modulated sweeps) (Letzkus et al. 2011). In this paradigm, foot shock elicits firing among a majority of layer 1 INs through basal forebrain release of acetylcholine acting on nicotinic receptors. Activated layer 1 INs in turn inhibit PV-INs in layer 2/3, which mediate perisomatic inhibition of neighboring PNs. PV-mediated disinhibition augments PN responses to the CS, whereas preventing such disinhibition attenuates memory formation (Letzkus et al. 2011). Although the involvement of specific layer 1 subtypes in these effects remains unclear, a similar role has been ascribed to auditory cortex VIP-INs in processing of aversive air puffs in a go-no-go task (Pi et al. 2013). This reinforces the idea that disinhibition mediated by IN-IN interactions is an important mechanism by which salient events gate stimulus processing and plasticity of PNs to enable new associative learning.

At the retrieval stage, studies of dorsomedial PFC (dmPFC, prelimbic and cingulate area 1) suggest that PN disinhibition is also a critical process underlying memory expression. Early work in the rat showed that PN firing correlates with CS presentation and is required for conditioned freezing, suggesting that dmPFC is an important locus underlying memory expression (Corcoran and Quirk 2007, Burgos-Robles et al. 2009). However, complex modulation of dmPFC INs by a fear-associated CS was first demonstrated by electrophysiological recordings in the mouse, where CS trials simultaneously activate a subset of INs while inhibiting another (Courtin et al. 2014). Through the use of optogenetics, it was established that CS-inhibited INs largely comprise PV-INs, whose photoexcitation during retrieval reduces memory expression. Meanwhile, PV-IN silencing, even in the absence of the CS, elicits defensive freezing that is associated with disinhibition of BLA-projecting PNs as well as phase resetting of local theta oscillations (Courtin et al. 2014), which are thought to be an important causal mechanism in fear expression (Seidenbecher et al. 2003, Dejean et al. 2016, Karalis et al. 2016).

The above results provide compelling evidence for the involvement of disinhibition in memory expression but leave undefined the mechanism underlying suppression of PV-IN firing. One possibility suggested by prior studies of US-related disinhibition (Letzkus et al. 2011, Pi et al. 2013, Wolff et al. 2014) is that an unidentified population of INs, when activated by the CS, might inhibit PV-INs through GABAergic connections. Indeed, we recently demonstrated that auditory fear conditioning is associated with an increase in CS-related activity of prelimbic SST-INs (Cummings and Clem 2020), which may correspond to CS-activated units in previous electrophysiological recordings (Courtin et al. 2014). Photoinhibition of SST-INs attenuates conditioned freezing, indicating that SST- and PV-INs exert opposing control over memory expression (Courtin et al. 2014, Cummings and Clem 2020). Conversely, photoactivation of SST-INs elicits freezing in conditioned mice, an effect that is negated by co-activation of PV-INs and associated with recruitment of a distributed fear-related network (Cummings and Clem 2020). Interestingly, disinhibition orchestrated by prefrontal SST-INs also underlies the expression of learned social fear, in which exposure to a conspecific reminder of threat triggers suppression of PV-IN firing (Xu et al. 2019). Convergent evidence thus indicates that, as in memory acquisition, retrieval of fear memory depends on interactions between IN subtypes that mediate phasic disinhibition of PNs. Less clear, however, is how associative disinhibition might be encoded during learning, or whether acquisition of CS responses involves IN plasticity. Below we discuss emerging evidence that fear conditioning modifies the function of inhibitory microcircuits, potentially closing the loop on their involvement in all stages of memory, including its long-term storage.

Inhibitory plasticity in memory formation

Initial evidence that GABAergic INs undergo plasticity during fear learning came predominantly in the form of molecular and structural changes related to GABAergic signaling. For example, fear conditioning is associated with reduced amygdala expression of GABA receptor subunits and GABA synthetic enzymes (Heldt and Ressler 2007, Lin et al. 2011), changes that are reversed upon extinction (Chhatwal et al. 2005, Heldt and Ressler 2007, Lin et al. 2009, Lin et al. 2011). In addition, it has now been shown that fear conditioning decreases the density of GABAergic axon terminals contacting PNs in the CA1 region of the hippocampus (Donato et al. 2013) as well as the basolateral amygdala (Rashid et al. 2016, Kasugai et al. 2019). Such effects imply that by sculpting the substrates underlying IN transmission, learning may induce persistent changes in the function of inhibitory in addition to excitatory circuits. Indeed, for nearly as long as the phenomenon of LTP has been investigated within glutamatergic pathways (Bliss and Lomo 1973), it has been established based on electrophysiological recordings that GABAergic microcircuits share a capacity for activity-dependent plasticity (Buzsaki and Eidelberg 1982, Kullmann et al. 2012).

While potentiation of synaptic responses by electrical stimulation does not necessarily imply that natural experience engages similar processes, a series of studies have revealed that auditory fear conditioning induces changes in excitatory synaptic transmission within the amygdala that are highly reminiscent of electrically-induced plasticity. Potentiation of auditory responses in the lateral nucleus develops rapidly during training, is contingent upon CS-US pairing, and mimics the effects of LTP induction within the same pathway (Rogan and LeDoux 1995, Rogan et al. 1997). In acute brain slices from conditioned animals, whole-cell recordings indicate a persistent strengthening of transmission at synapses formed by subcortical and cortical afferents onto lateral amygdala PNs (Tsvetkov et al. 2002, Zhou et al. 2009, Clem and Huganir 2010, Clem and Huganir 2013, Hong et al. 2013, Nabavi et al. 2014, Namburi et al. 2015, Kim and Cho 2017), which is accompanied by increases in probability of glutamate release (Tsvetkov et al. 2002, Zhou et al. 2009), ratio of AMPA- to NMDA-receptor-mediated excitatory postsynaptic currents (EPSCs) (Clem and Huganir 2010, Clem and Huganir 2013, Hong et al. 2013, Nabavi et al. 2014, Namburi et al. 2015, Kim and Cho 2017), and rectification of AMPA receptor-mediated EPSCs (Clem and Huganir 2010, Clem and Huganir 2013, Hong et al. 2013). Although effects of conditioning on inhibitory circuits of the lateral amygdala are far less clear, the emerging picture is of parallel adjustments in excitation and inhibition.

In addition to downregulating GABAergic molecular markers (Heldt and Ressler 2007, Lin et al. 2011), electrophysiological recordings indicate that auditory fear conditioning reduces spontaneous inhibitory postsynaptic currents (IPSCs) in lateral amygdala PNs, providing the first functional correlate of these molecular effects (Lin et al. 2011). Interestingly, reduction of IPSC frequency is reversed by extinction, but returns following shock-evoked reinstatement, an effect that is accompanied by GABA receptor internalization. As noted above, we recently demonstrated that a sparse but morphologically complex population of PV-INs mediates feedforward inhibition onto lateral amygdala PNs in response to activation of cortical and subcortical afferents, which convey auditory CS input to this nucleus (Lucas et al. 2016). Following conditioning, input and output synapses of lateral amygdala PV-INs exhibit decreased AMPA and GABA release probability, respectively. This effectively diminishes feedforward inhibition within CS pathways, which could potentially amplify conditioned responses or facilitate their generalization, for example by increasing the size of fear-related PN ensembles. In contrast, increased excitatory transmission was observed after learning in basal amygdala PV-INs, which do not contribute to feedforward inhibition but might influence memory consolidation through entrainment of gamma oscillations (Kanta et al. 2019). Thus, genetically homogeneous INs that populate different circuit motifs may be uniquely affected by fear learning, with potentially important functional consequences.

Precisely how IN plasticity modulates amygdala CS responses remains unclear, but calcium imaging experiments suggest some possibilities. During auditory CS-US pairing, CS-evoked ensemble activity is transformed, becoming more similar but not identical to the US representation (Grewe et al. 2017). Among the cellular level changes are PNs that exhibit increased firing but also many whose cue activity decreases after learning. In addition, consistent with previous electrode recordings (Quirk et al. 1995), many PNs that are not responsive to the CS before conditioning become responsive after CS-US pairing. These outcomes are not easily explained by Hebbian LTP of excitatory PN inputs but could be an indirect result of IN plasticity leading to inhibition and disinhibition of PNs (more about this below). The precise role of INs in these population-level effects will be difficult to dissect, however, due to our poor understanding of their wiring, but also because they are not the only major source of inhibition within the basolateral complex. Several intercalated cell groups that border this area give rise to GABAergic projections targeting amygdala PNs. These molecularly distinct populations process sensory information and exhibit experience-dependent plasticity, making it likely that they contribute to amygdala response transformations (Asede et al. 2015, Asede et al. 2021). Similar plasticity has been observed in GABAergic projection cells of the central amygdala (Li et al. 2013, Penzo et al. 2014, Penzo et al. 2015), which underlines the importance of GABAergic plasticity in fear conditioning, but is beyond the scope of the present review.

With more complex stimuli, auditory fear conditioning relies on processing within primary auditory cortex (Weible et al. 2014). As in the amygdala, INs in the cortex play important roles in shaping CS and US responses, so it is natural to anticipate that learning under these conditions might involve IN plasticity. Unlike the amygdala, however, the cortex contains an entire layer (layer 1) devoid of excitatory cell bodies and populated by unique and poorly characterized INs. Recently, genetic access was gained to a subset of these cells that express neuron derived neurotrophic factor (NDNF) (Abs et al. 2018), which allowed for selective monitoring of their activity using calcium imaging. Similar to Martinotti-type SST-INs, NDNF-INs form inhibitory synapses onto the apical dendrites of underlying PNs, where they regulate the occurrence of dendritic spikes. Following conditioning, NDNF-INs become more responsive to a CS paired with a co-terminating US, but not one that was explicitly unpaired. Thus, a component of memory encoding in the auditory cortex may involve an upregulation of dendritic inhibition. Interestingly, NDNF-INs are reciprocally connected with SST-INs but not PV- or VIP-INs, suggesting that these cell types may form mutually opposing inhibitory modules. However, the functional role of either population in memory expression remains to be examined.

While a role for mPFC in fear memory formation was initially disputed (Morgan and LeDoux 1995), it has been well-established that it plays a critical role in expression of auditory fear memories (Sotres-Bayon and Quirk 2010), as described above. Recently, however, we showed that conditioning is associated with increased glutamate release probability at excitatory inputs onto prelimbic SST-INs, an effect that correlates with a higher CS-evoked activity (Cummings and Clem 2020). In addition, optogenetic interrogation of SST- and PV-INs revealed changes in relative transmission at basolateral amygdala afferent connections onto these cells as well as their GABAergic outputs onto local populations, suggesting that learning induces a broader scope of synaptic plasticity affecting both SST- and PV-INs. These changes appear to favor the recruitment of SST-INs as well as their suppression of PV-IN activity, suggesting a mechanism for encoding of cue-related PN disinhibition. Indeed, selective interference with SST-IN plasticity during conditioning impairs subsequent memory retrieval.

Although this establishes a role for dmPFC inhibitory plasticity in memory formation, the results raise a number of important questions. For example, because only some downstream targets of the dmPFC are activated during memory retrieval, how can potentiation of SST-IN activity account for such selective effects? The conventional view would imply that SST-INs mediate indiscriminate inhibition of local cell populations. However, experiments suggest the frontal cortex contains distinct subnetworks of PNs that are preferentially modulated by inhibition versus disinhibition upon stimulation of genetically-defined INs (Garcia-Junco-Clemente et al. 2017). Such functional motifs could arise from selective wiring of PNs with GABAergic subtypes, such as SST- versus PV-INs, or varying strength of transmission in these connections. A related question that remains to be examined is whether CS-activated SST-INs constitute a valence-specific subpopulation or otherwise operate as a functionally specialized ensemble, analogous to excitatory engram cells. This seems particularly relevant given a recent report that opioid treatment increases SST-IN→PV-IN transmission in the prelimbic cortex and may therefore enhance PN disinhibition (Jiang et al. 2021). In a study awaiting peer review, we utilized intersectional genetic tagging to provide some initial answers to these questions. The results indicate that SST-IN ensembles preferentially activated by fear learning exhibit engram-like properties, including higher excitatory drive, reactivation during memory retrieval, and a unique capacity to elicit freezing (Cummings et al. 2021). In addition, however, these cells interact differentially with learning-activated PNs, which they inhibit more weakly than their non-activated neighbors. This suggests that selective recruitment of fear networks by SST-INs may be orchestrated through joint mono- and disynaptic control of functionally discrete PN populations, which are either inhibited or disinhibited depending on their unique pattern of GABAergic input.

More heavily implicated in context fear associations, the dorsal hippocampus has also been identified as a locus for SST-IN plasticity during memory formation. In the first study, SST-INs were examined as a potential substrate for structural plasticity in the APP/PS1 model of Azheimer’s disease (Schmid et al. 2016), for which an influential hypothesis about cognitive dysfunction involves dysregulation of E:I balance (Palop et al. 2007). SST-INs in APP/PS1 mice exhibit morphological abnormalities including progressive axon loss and reduced turnover of dendritic spines, an anatomical structure more commonly associated with PNs. Interestingly, acquisition of context fear is associated with an increase in spine density of SST-INs in the CA1 region of wildtype, but not APP/PS1 mice. Learning-related spine gain is dependent on cholinergic signaling, which in the hippocampus arises from the diagonal band of the medial septum. More recently, another study demonstrated that SST-INs exhibit an increase in spontaneous EPSC frequency after contextual fear conditioning (Artinian et al. 2019), potentially confirming that newly generated dendritic spines are indeed functional (Schmid et al. 2016). Both SST-IN plasticity as well as context fear depend on metabotropic glutamate receptor 1 (mGlu1) signaling and cell-autonomous activation of the mammalian target of rapamycin 1 (mTORC1) pathway. These data add to growing evidence that SST-INs participate in memory formation and confirm that mechanisms governing their ex vivo LTP play analogous roles in experience-dependent plasticity (Vasuta et al. 2015, Artinian et al. 2019).

Finally, although initial learning is a critical stage for plasticity, it is widely recognized that memories remain susceptible to change even after they are successfully encoded. An important way that pathological fear memories transform with time is through a loss of precision, compromising the discrimination of fear-related cues from non-threatening stimuli. This process, termed over-generalization, has been modeled in contextual fear conditioning, where a mechanism underlying memory specificity is the growth of synaptic connections between dentate gyrus and PV-INs in hippocampal CA3 (Guo et al. 2018). Downregulation of actin-binding LIM protein 3 (ABLIM3) in dentate granule cells precedes this anatomical growth under normal conditions and, when artificially induced, it confers specificity to remote memories in aging mice. Other work suggests that inhibitory plasticity set in motion by fear learning may support emergent network states that linger after training to influence memory consolidation as well as the specificity of future learning (Karunakaran et al. 2016, Rashid et al. 2016). Aside from these promising results, very little is known about the contribution of INs to later stages of memory, or whether their recruitment and plasticity might influence processes like systems consolidation or memory updating. Given an increasingly powerful toolset for investigating INs, these may be fruitful areas for exploration.

Concluding remarks

As methods for monitoring and manipulating neuronal subsets advance, so too will the complexity of our theoretical understanding. Accurate models of memory encoding will require that we grapple with nonconventional roles of inhibition within cortical and limbic microcircuits, which go beyond familiar processes like gain control and rhythmic oscillations. Illuminating the function of inhibition and disinhibition, precisely-timed and spatially-restricted, promises not only to improve our understanding of how memories are written but also our comprehension of how brain networks are organized to aid their processing and retrieval. Local INs represent an ideal target in these endeavors because they control broad swaths of their local network, but as we peer deeper into this circuitry, an important challenge will be accounting for the compounding effects of various GABAergic motifs, which can reverberate through many synaptic connections and ultimately throughout the brain. Attributing causality in this landscape to specific circuit elements will require increasingly greater precision in our mechanistic analysis.

Figure 1. Inhibitory circuit motifs implicated in fear memory.

Circuits implicated in conditioning involve interneurons governing learning-related network dynamics as well as those exhibiting plasticity in response to training. At the retrieval stage, inhibitory circuit activity is implicated in cue-related memory expression. OLMα2-INs: oriens lacunosum moleculare alpha-2 nicotinic acetylcholine receptor-positive interneurons. Ndnf1-INs: neuron-derived neurotrophic factor 1-expressing interneurons. L1-INs: unspecified cortical layer 1 interneurons.

Figure 2. Circuit mechanisms of memory encoding by prelimbic SST-INs.

A, Auditory fear conditioning results in potentiation of excitatory synaptic inputs onto SST-INs as well as cue-evoked SST-IN activity (Cummings and Clem 2020). In conditioning to both auditory and social stimuli, SST-INs suppress PV-IN firing via monosynaptic inhibitory connections, leading to disinhibition of fear-related PNs (Cummings and Clem 2020; Xu et al. 2019). B, Potential mechanism for selective interactions between cue-responsive SST-IN and PN ensembles. Differential inhibition of fear-relevant and -irrelevant PN populations by SST-INs, together with PV-IN-mediated disinhibition, could enable selective recruitment of fear-related outputs from the prelimbic cortex through joint mono- and disynaptic control (Cummings et al. 2021). Inset depicts remote brain regions that activated upon photoexcitation of prelimbic SST-INs (yellow) and therefore represent potential targets for fear-related prelimbic PNs (Cummings and Clem 2020). PL = prelimbic cortex, CPu = caudate putamen, DG = dentate gyrus, PVT = paraventricular thalamus, LHb = lateral habenula, MD = mediodorsal thalamus, NAc = nucleus accumbens, BLA = basolateral amygdala, DM = dorsomedial thalamus, vCA1 = ventral cornu ammonis 1, vlPAG = ventrolateral periacqueductal gray.