The ins and outs of inhibitory synaptic plasticity: neuron types, molecular mechanisms and functional roles

Abstract

GABAergic interneurons are highly diverse and their synaptic outputs express various forms of plasticity. Compelling evidence indicates that activity-dependent changes of inhibitory synaptic transmission play a significant role in regulating neural circuits critically involved in learning and memory and circuit refinement. Here, we provide an updated overview of inhibitory synaptic plasticity with a focus on the hippocampus and neocortex. To illustrate the diversity of inhibitory interneurons, we discuss the case of two highly divergent interneuron types, parvalbumin-expressing basket cells and neurogliaform cells, which support unique roles on circuit dynamics. We also present recent progress on the molecular mechanisms underlying long-term, activity-dependent plasticity of fast inhibitory transmission. Lastly, we discuss the role of inhibitory synaptic plasticity in neuronal circuits’ function. The emerging picture is that inhibitory synaptic transmission in the CNS is extremely diverse, undergoes various mechanistically distinct forms of plasticity, and contributes to a much more refined computational role than initially thought. Both the remarkable diversity of inhibitory interneurons and the various forms of plasticity expressed by GABAergic synapses provide an amazingly rich inhibitory repertoire that is central to a variety of complex neural circuit functions, including memory.

Introduction

Synaptic plasticity is a fundamental phenomenon that refers to the ability of synapses to change their efficacy as a result of transient neuronal activity. It is believed to represent a major cellular event sculpting circuit dynamics and underlying learning and memory (Takeuchi et al., 2014). Studies in animal models (e.g., rodents) established molecular mechanisms and neuronal circuits involved in this phenomenon (Luscher et al., 2000; Herring & Nicoll, 2016). Most of the information on synaptic plasticity was initially obtained from excitatory synapses, but many experimental and computational studies more recently established that inhibitory synaptic plasticity (ISP) is also a major player in neuronal circuit mechanisms underlying memory processes (Hennequin et al., 2017) as well as maintenance of circuit excitability (Nelson & Turrigiano, 2008). Identifying cellular and molecular mechanisms underlying ISP is important for several reasons: (i) ISP is expected to critically control the computation of cortical pyramidal neurons (Kullmann et al., 2012; Hennequin et al., 2017), (ii) it may promote stability of network activity following the formation of excitatory engrams (Barron et al., 2017), which are defined as ensembles of neurons involved in storing and recalling memory, (iii) it is likely to be involved in brain disease, e.g., synaptic remodeling after stroke – (Berger et al., 2019), or may be exploited for therapeutic intervention (Di Lazzaro et al., 2018).

This review article focuses on GABAergic inhibitory synapses and ISP as a central contributor to fundamental circuit functions. We first provide a brief account of GABAergic neuron diversity in the cerebral cortex and hippocampus, and discuss how molecular properties at specific GABAergic synapses may confer distinct capacity for plasticity. Next, we review key molecular mechanisms underlying different forms of ISP, and speculate on how this diversity in mechanisms for ISP may support the wide palette of sophisticated contribution of inhibitory synapses to circuit computations. Finally, we discuss emerging hypotheses about the role of ISP. Thus, we bring together recent experimental and theoretical studies, and discuss how GABAergic neuron diversity and ISP heterogeneity contribute to neural circuit computations and complex brain functions such as learning and memory.

GABAergic neuron diversity

Cortical GABAergic neurons are diverse in their developmental origin, gene expression, morphology, function and connectivity (Gupta et al., 2000; Klausberger & Somogyi, 2008; Tremblay et al., 2016; Pelkey et al., 2017; Fishell & Kepecs, 2019; Huang & Paul, 2019). This diversity is probably even more pronounced in the human cerebral cortex, where GABAergic cell types without apparent homology with rodent cortical cell types have been found (Boldog et al., 2018). Combined information of dendritic and axonal patterns, molecular markers and functional activities of neurons is useful to determine cell types. Consistent with this, multiparametric methods have been endorsed to classify GABAergic neurons (Petilla Interneuron Nomenclature et al., 2008; DeFelipe et al., 2013). GABAergic cells are eminently target specific, selectively innervating subcellular domains of postsynaptic cells. For example, axo-axonic interneurons make synapses exclusively on the axon initial segment of cortical pyramidal cells (Somogyi et al., 1983); basket cells (BCs) target preferentially the somata and proximal dendrites of postsynaptic neurons (Thomson et al., 1996; Tamas et al., 1997), Martinotti and neurogliaform cells (NGFCs) target the dendrites of postsynaptic cells (Tamas et al., 2003; Wang et al., 2004), some neocortical interneurons preferentially target other interneurons (Pfeffer et al., 2013). Functional specialization of inhibitory neurons provides subtle regulation of cortical networks. For example, cortical NGFCs provide feed-forward inhibition of distal dendrites of postsynaptic pyramidal neurons (Tamas et al., 2003) and also elicit presynaptic inhibition of transmitter release (Olah et al., 2009). A division of labor amongst interneuron types in governing network activity is well known in the hippocampus and cortex (Kawaguchi & Kubota, 1997; Klausberger & Somogyi, 2008; Tremblay et al., 2016). Recent advances in high-throughput single-cell transcriptomics (scRNAseq) provided a new quantitative genetic framework to elucidate GABAergic neuron diversity (Poulin et al., 2016; Shekhar et al., 2016; Paul et al., 2017; Que et al., 2019). Given this diversity, it is possible that different GABAergic neuron types display different forms of ISPs, an idea that for now is supported by a limited amount of experimental data and will need further work (Horn & Nicoll, 2018; Schulz et al., 2018).

Interneurons synapse specialization: molecular diversity of synapses from BC and NGFC

Here we will compare two GABAergic neuron types that are placed at the extremes of the GABAergic neuron diversity spectrum, namely, parvalbumin expressing (PV+) BCs and NGFCs of neorcortex and hippocampus. We will summarize molecular differences at the inhibitory synapses established by NGFCs and PV+ BCs, and discuss hypotheses about how these properties may endow synapses from these neuron types with distinct capacity for plasticity.

PV+ BCs mediate fast, phasic inhibition (Hefft & Jonas, 2005), in contrast NGFCs evokes volume transmission leading to slow inhibition (Capogna and Pearce, 2011). Information regarding synaptic-associated molecules expressed by PV BC+ and NGFCs is relatively abundant. By discussing their similarities and differences we can provide testable hypotheses about how molecular specificity may inform on interneuron-type specific signaling pathways underlying ISP. PV+ interneurons are estimated to be numerous (e.g, about 14% of CA1 hippocampal interneurons - (Bezaire & Soltesz, 2013) and elicit robust inhibitory responses in targeted principal neurons (Hu & Jonas, 2014), thus, they are likely to represent a significant group of interneuron mediating ISP in experimental conditions in which the presynaptic neuron is not identified (Vogels et al., 2013). However, certain experimental settings allow for the specific identification of the inhibitory connection. For example, long-term potentiation (LTP) of cortical PV+ BC-mediated synaptic inhibition is elicited by visual deprivation (Maffei et al., 2006). Spike timing dependent plasticity (STDP) has also been observed at this type of connection: short delay between spiking of a cortical PV+ BC and a principal neuron elicits long-term depression (LTD), whereas longer delays evoke LTP (Holmgren & Zilberter, 2001). More recently, cortical PV+ interneuron-mediated inhibition shows STDP that contributes to auditory map remodeling (Vickers et al., 2018). In contrast to PV+ interneurons, much less is known on ISP mediated by NGFCs. In the hippocampus, this interneuron type displays marked synaptic depression evoked by a train of presynaptic stimuli at theta frequency which has a recovery time constant of about 10 minutes (Karayannis et al., 2010). Furthermore, the injection into a postsynaptic NGFC in vitro of a firing pattern recorded in a NGFC in vivo displays shorter-term retrograde synaptic depression lasting about 1 minute (Li et al., 2014a). As for long-term ISP, the molecular properties we discuss below suggest that synapses established by NGFCs have the machinery for changing their synaptic efficacy. In support of this possibility, recent findings have implicated this interneuron subtype in memory processes in the cortex (Abs et al., 2018).

Salient features of hippocampal PV+ BCs and NGFCs are illustrated in Figure 1. Morphological properties of PV+ BCs and NGFCs are notably different. PV+ BCs have multiple dendrites that often cross layers (Gulyas et al., 1999; Tukker et al., 2013), suggesting heterogeneous inputs. In contrast, the dendrites of NGFCs are compact and arranged in a stellate fashion around the soma (Vida et al., 1998), pointing to inputs from a more restricted set of afferent pathways. The axon of PV+ BC shows extensive arborization, and generates a divergent inhibitory output restricted to the perisomatic domain of postsynaptic targets (Sik et al., 1995). The axon of NGFC is also extensive but it branches profusely and usually targets exclusively the dendritic domain of postsynaptic neurons in the hippocampus (Vida et al., 1998; Price et al., 2005), as in the neocortex (Tamas et al., 2003). Molecular features of PV+ BCs and NGFCs are also quite different. A subpopulation of BCs expresses PV with an average concentration of 10 μM at the soma in hippocampal dentate granule cell layer. Because PV is a slow calcium buffer, it affects the time course of intracellular calcium transients in terminals after an action potential, and hence regulate short-term synaptic plasticity favoring synaptic depression (Eggermann & Jonas, 2011). In contrast, both neocortical and hippocampal NGFCs express an array of marker proteins, including α-actinin2, neuropeptide Y, neuronal nitric oxide synthase (nNOS), the transcription factor COUP-TFII and the extracellular matrix protein reelin (Price et al., 2005; Fuentealba et al., 2008; Olah et al., 2009; Fuentealba et al., 2010; Szabadics et al., 2010). The presence of nNOS supports the idea that NGFCs synapses are plastic and this plasticity may have physiological significance (Makara et al., 2007). When hippocampal NGFCs generate a theta rhythm-associated activity their synapses display inhibitory short-term synaptic plasticity onto pyramidal neurons. This phenomenon, called “firing-induced suppression of inhibition” (FSI) requires backpropagation of action potentials, postsynaptic calcium influx through L-type calcium channels, nNOS activity and NO retrograde release, and activation of NO-sensitive guanylyl cyclase receptors at presynaptic terminals (Li et al., 2014a). FSI indirectly increases the amplitude of excitatory postsynaptic potentials (EPSPs), and may enhance spatial and temporal summation of excitatory inputs to NGFCs, thereby regulating their inhibition of pyramidal cells (Li et al., 2014a). Neocortical PV+ BCs also show NO-dependent short-term plasticity, but its time scale is markedly different from the one exhibited by hippocampal NGFCs. Somatic depolarization of pyramidal cells in layer 5 triggers Ca²⁺-dependent retrograde release of NO which diffuses to PV+ BC axon terminals and elicits a persistent increase of GABA release (Lourenco et al., 2014).

Figure 1. Interneuron diversity: neurogliaform and parvalbumin expressing basket cells.

Left. Reconstruction of a neurogliaform cell (NGFC, soma and dendrites, red; axon, green) and a CA1 pyramidal neuron (soma and dendrites, black) recorded from a rat in vitro. Salient features detected in NGFCs are listed below the image. The picture is taken from Price et al, J. Neurosci., 28(27):6974–6982, 2008(Price et al., 2008). Right. Reconstruction of a parvalbumin expressing basket cell (BC, soma and dendrites, black; axon, gray) and a CA1 pyramidal neuron (soma and dendrites, blue) recorded from a rat in vitro. Key features observed in BCs are listed below the image. The picture is taken from Foldy et al, Nature Neurosci.,13(9):1047–1049, 2010 (Foldy et al., 2010).

Distinct synaptic organization for PV+ BC and NGFC

The synaptic organization of PV BCs and NGFCs is markedly different. A single neocortical NGFC axon has a release site density comparable to that of five or six BC axons (Olah et al., 2009), suggesting that GABA released from NGFC axons can reach synaptic but also non-synaptic receptors. Moreover, axon boutons of NGFCs are often (Olah et al., 2009), but not always (Tamas et al., 2003; Price et al., 2008; Fuentealba et al., 2010), found 1–5 μm away from target dendrites, a surprisingly long distance compared to the 10–20 nm typically detected at conventional PV+ BC synapses (Tukker et al., 2013). The release of GABA from neocortical NGFCs can also inhibit the release of glutamate or GABA from axon terminals located remotely from NGFC release sites (Olah et al., 2009). Therefore, it has been suggested that NGFCs can mediate volume transmission (Olah et al., 2009) wherein a widespread, prolonged, low-level GABA transient is produced by a dense array of NGFC release sites (Capogna & Pearce, 2011). Furthermore, at the postsynaptic level, data obtained with high resolution replica immunogold labeling indicate that all CA1 pyramidal cell somatic inhibitory synapses contain the α1, α2, β1, β2, β3 and γ2 subunits along with the adhesion molecule neuroligin-2 (NL-2) (Kerti-Szigeti & Nusser, 2016), whereas neocortical NGFCs synapses show also high level of GABA_A receptors GABA_ARs) containing δ subunits (Olah et al., 2009). The presence of the δ subunits in GABA_A receptors at synapses is unique to NGFCs, as receptors containing this subunit are typically located extrasynaptically (Farrant & Nusser, 2005; Belelli et al., 2009). More recently, hippocampal NO-synthase expressing NGFCs have been found to activate postsynaptic α5-GABA_A receptors which strongly contribute (50–80%) to the inhibitory synaptic conductance (Schulz et al., 2018). Importantly, the inhibition of dendritic NMDA spikes via α5-GABA_A receptors-containing synapses provides a mechanistic basis for the powerful control of NMDA receptor-dependent burst firing and synaptic plasticity in CA1 pyramidal cells by dendrite-targeting interneurons (Schulz et al., 2018). Conversely, α5-GABA_A receptors supply a negligible contribution to perisomatic inhibition elicited by fast-spiking PV interneurons (Schulz et al., 2018). Whether synaptic junctions formed by PV+ BCs or NGFCs also differ in the organization of postsynaptic molecular components, such as scaffolding proteins (e.g., gephyrin, collybistin), adhesion proteins (e.g., NL-2), kinases (e.g., PKC, ERK1/ERK2) and proteases (e.g., calpain) for GABA_AR phosphorylation/dephosphorylation, is currently not known. These molecular components are pivotal for localization, stability and regulation of baseline as well as plastic GABAergic signaling (Chiu et al., 2019).

A distinctive functional feature is that action potentials of PV+ BC have fast kinetics that evoke GABA_AR-mediated inhibitory postsynaptic responses with a time constant of 5–10 ms (Cobb et al., 1995), whereas NGFCs have broader spikes that evoke slow postsynaptic inhibitory potentials (time constant can be > 30 ms) (Tamas et al., 2003). Upon repetitive stimulation, inhibitory responses evoked by both PV+ BCs and NGFCs show marked depression (Hefft & Jonas, 2005; Karayannis et al., 2010).

Overall, PV+ BCs synapses onto their target neurons have been convincingly shown to express pre- and postsynaptic molecular motifs apt to confer fast neurotransmission proper of this GABAergic neuron type. These motifs include the tight “nanodomain” coupling between Ca²⁺ channels and release sensors for exocytosis, promoting efficacy and temporal precision of neurotransmitter release, as well as shortening the synaptic delay (Bucurenciu et al., 2008). Furthermore, a subpopulation of PV+ interneurons in the neocortex and hippocampus uses the synaptotagmin 2 isoform as a release sensor protein for neurotransmitter release (Kerr et al., 2008), in contrast to the most common synaptotagmin 1 expressed in principal cells (Geppert et al., 1994). The synaptotagmin 2 isoform has the fastest Ca²⁺ binding kinetics among synaptotagmin family proteins, a feature that is likely to contribute to fast neurotransmission. Conversely, direct measurements of presynaptic Ca²⁺ transients in hippocampal NGFCs reveal slow decaying Ca²⁺ transients at axonal boutons (Price et al., 2008), consistent with the slow kinetics of neurotransmission elicited by this neuron type.

The spiking patterns of PV+ BCs and NGFCs detected in vitro and in vivo are also quite characteristic. PV+ BCs show remarkable high action potential frequency (> 10 Hz) (Sik et al., 1995), whereas hippocampal NGFCs fire maximally up to 10 Hz during theta network oscillations in vivo (Fuentealba et al., 2010) under basal conditions without applying any stimulation. Neocortical and hippocampal NGFCs also display a so-called barrage firing (Sheffield et al., 2011; Suzuki et al., 2014; Chittajallu et al., 2020), a persistent spiking activity occurring up to several minutes after cells’ stimulation, which can reach frequencies up to 130 Hz. Because of prominent frequency-dependent synaptic depression, it is still unclear how much inhibition is provided to postsynaptic targets during NGFC barrage firing.

At the circuit level, a major difference between PV+ BCs and NGFCs is that PV+ BCs mediate both feedforward (FF) and feedback (FB) inhibition (Hu et al., 2010), whereas NGFCs only mediate FF inhibition (Capogna & Pearce, 2011). This has been well characterized in hippocampus. CA1 afferent glutamatergic axons (e.g. from entorhinal cortex or Schaffer collaterals) elicit FF inhibition via parallel activation of CA1 pyramidal cells, PV+ BCs and NGFCs. CA1 pyramidal cells driven by the same glutamatergic axons in turn activate PV+ BCs, but not NGFCs, recruiting circuit-specific FB inhibition. The speed of both FF and FB inhibition elicited by PV+ BCs is high; the latency of disynaptic inhibition under physiological conditions is less than 2 ms (Miles, 1990). Hippocampal NGFCs evoke much slower FF inhibition onto CA1 pyramidal neurons and influence the temporal integration of incoming excitatory signals on a longer time scale (Price et al., 2008). Hippocampal PV+ BCs powerfully inhibit spiking of postsynaptic neurons due to the perisomatic location of its GABAergic synapses, but also elicit rebound spiking after postsynaptic hyperpolarization that promote network synchronization activity in vitro (Cobb et al., 1995). Rebound spiking has also been observed in vivo (Adhikari et al., 2012). Neocortical NGFCs inhibit the dendrites of postsynaptic neurons and reduce dendritic Ca²⁺ signals in vitro (Perez-Garci et al., 2006) and in vivo (Abs et al., 2018). This neuron type presynaptically inhibits the release of glutamate from nearby excitatory terminal via a GABA_B receptor mediated mechanism (Olah et al., 2009). In addition, PV+ BCs have specific functions in microcircuits. FF inhibition by CA1 hippocampal PV+ BCs sharpens the window for temporal summation of EPSPs and action potential initiation in principal neurons (Pouille & Scanziani, 2001), and broadens the dynamic range of activity in principal neuron ensembles (Pouille et al., 2009). Conversely, lateral FB inhibition promotes a “winner-takes-all” mechanism, enforcing spiking in principal cells with the strongest input, and inhibiting activity in the weaker principal cells (de Almeida et al., 2009). PV+ BCs could also contribute to other circuit functions such as sparsification of activity (Pernia-Andrade & Jonas, 2014), pattern separation (Leutgeb et al., 2007), and grid-to-place code conversion (de Almeida et al., 2009). For a review on the role of PV+ BC in these events see also Hu et al, Science 2010 (Hu et al., 2010). Hippocampal PV+ BCs usually enhance their spiking activity during network oscillations (Klausberger & Somogyi, 2008). Experimental manipulation of PV+ interneuron activity showed that these neurons modulate several phenomena such as the shape of place fields and the phase precession in CA1 pyramidal neurons (Royer et al., 2012), the gain of sensory responses (Lee et al., 2012), the regulation of learning (Donato et al., 2013). Much less is known on the functional role of NGFCs. Hippocampal NGFCs preferentially spike phased-locked to theta network oscillations (Fuentealba et al., 2010). In addition, activity of NGFCs is increased during fear memory retrieval in an auditory associative fear learning test. This effect is specific to NGFC, as the activity of another group of dendrite-targeting interneurons, those expressing SST, remains unaltered (Abs et al., 2018). Based on the distinct molecular and structural features, it is conceivable that synapses established by PV+ BCs and NGFCs onto principal neurons engage distinct mechanisms of ISP, which can mediate long-term redistribution of synaptic inhibition impinging on different compartments of pyramidal neurons. Recent experimental evidence appears to support this prediction (Horn & Nicoll, 2018).

Molecular mechanisms of Inhibitory Synaptic Plasticity

In addition to the molecular heterogeneity of inhibitory neurons and their synapses, there is also remarkable diversity in the mechanisms underlying ISP (for more extensive reviews, see (Gaiarsa et al., 2002; Castillo et al., 2011; Luscher et al., 2011; Kullmann et al., 2012; Maffei et al., 2017; Chiu et al., 2019). Although such diversity imposes an additional challenge to the study of ISP, some mechanistic principles have been identified. As for excitatory synaptic plasticity, ISP can be expressed presynaptically as changes in GABA release, or postsynaptically as changes in GABA receptor number or function (Figure 2). In this section, we summarize recent advances and emerging mechanisms on long-term, activity-dependent strengthening and weakening of fast inhibitory transmission mediated by GABA_ARs –i.e. I-LTP and I-LTD, respectively. While we will focus on data from synapses of the rodent neocortex and hippocampus, other brain areas sharing similar mechanisms of ISP will be cited in order to highlight generalizability.

Figure 2. Molecular mechanisms underlying activity-dependent forms of ISP.

Left. Presynaptic ISP commonly involves activation of postsynaptic glutamate receptors (GluR), e.g. NMDAR and mGluR-I, and mobilization of retrograde signals that induce long-lasting changes in GABA release, by targeting a presynaptic receptor (R). Glutamate released from neighboring terminals, by activating glutamate receptors (e.g. NMDARs) on GABAergic terminals, can also induce presynaptic ISP. Coincident presynaptic and postsynaptic activity, via voltage-gated calcium channels (VGCC), can contribute to the mobilization of retrograde signals and/or modulation of downstream signaling in the GABAergic terminal. Right. Postsynaptic ISP can be induced by coordinated presynaptic and postsynaptic activity, including subthreshold changes in membrane potential (ΔVm), and GluR (NMDAR or mGluR-I) activation. Such activities set in motion a metabolic cascade of events that results in exo/endocytosis, lateral diffusion and phosphorylation/dephosphorylation of GABA_ARs

Presynaptic forms of ISP

Presynaptic I-LTP and I-LTD are widely expressed throughout the brain (Castillo et al., 2011; Castillo, 2012). Here, GABAergic terminals integrate diverse signals to induce long-lasting changes in GABA release by a (poorly understood) mechanism that may involve changes in the release machinery, presynaptic Ca²⁺ influx, as well as presynaptic structural changes (Castillo, 2012; Yang & Calakos, 2013; Atwood et al., 2014; Monday et al., 2018). Induction is commonly mediated by retrograde signals mobilized following transient, repetitive activation of nearby excitatory synapses – in a form of heterosynaptic plasticity – or repetitive firing of the postsynaptic neuron. Presynaptic Ca²⁺ elevations, via voltage-gated calcium channels (VGCCs) or presynaptic NMDA receptors (pre-NMDARs), activate metabolic cascades that may play a permissive, modulatory or instructive role in the induction of presynaptic ISP (Castillo, 2012; Monday et al., 2018). Diverse chemical messengers act as retrograde signals (Regehr et al., 2009), and a number of these messengers mediate ISP. Chief among them are endocannabinoids (eCBs), brain-derived neurotrophic factor (BDNF) and nitric oxide (NO). In some cases, ISP can be induced by presynaptic signals alone –i.e. in the absence of retrograde signals (Castillo et al., 2011).

The best-characterized form of presynaptic ISP is probably the eCB-mediated I-LTD. In many brain regions, repetitive activation of glutamatergic inputs triggers eCB mobilization from the postsynaptic cell to the presynaptic terminal, where they bind to G_i/o-coupled type 1 cannabinoid receptors (CB₁) to suppress GABA release in a long-term manner (Castillo et al., 2012). Typically, eCB-mediated I-LTD is triggered by postsynaptic activation of group I metabotropic glutamate receptors (mGluR-I), leading to the production of diacylglyercol (DAG) by phospholipase C (PLC). Diacylglycerol lipase (DGL) converts DAG to the major eCB 2-AG, which is released from the postsynaptic cell and travels back across the synapse to activate presynaptic CB₁ receptors (Heifets & Castillo, 2009; Kano et al., 2009). eCB-mediated I-LTD has been reported in several areas of the rodent brain, including the hippocampus (Chevaleyre & Castillo, 2003), amygdala (Marsicano et al., 2002; Azad et al., 2004), dorsal striatum (Adermark et al., 2009), hypothalamus (Crosby et al., 2011), and visual cortex (Jiang et al., 2010; Sun et al., 2015). In the hippocampus and basolateral amygdala (Freund et al., 2003; Vereczki et al., 2016), CCK+ (regular-spiking), but not in PV+ (fast-spiking) interneurons, selectively express CB₁ receptors (Freund et al., 2003). However, this dichotomy is less clear in other brain areas (Younts & Castillo, 2014). Theta-burst firing of CA1 pyramidal neurons for a few minutes, by raising intracellular calcium and mobilizing eCBs, is sufficient to induce hippocampal I-LTD at both somatic and dendritic inhibitory synapses (Younts et al., 2013).

BDNF/TrkB-mediated I-LTP is induced by activity-dependent release of BDNF from either axon terminals or dendrites (Edelmann et al., 2014), and is typically observed in immature circuits (Castillo et al., 2011). There is good evidence that dendritically released BDNF mediates I-LTP by activating presynaptic TrkB receptors in the hippocampus (Gubellini et al., 2005; Sivakumaran et al., 2009) and visual cortex (Inagaki et al., 2008). This ISP is initiated by intracellular Ca²⁺ rise via NMDARs or VGCCs, or calcium release from intracellular stores. As in eCB-mediated I-LTD (Heifets et al., 2008), input-specificity of BDNF-mediated I-LTP may derive from a requirement for coincident presynaptic interneuron activity to enhance TrkB signaling (Liu et al., 2007; Edelmann et al., 2014). The effects of BDNF/TrkB signaling on GABAergic transmission and plasticity are particularly complex (Lu et al., 2014), given that BDNF can increase the number of GABAergic terminals (Marty et al., 2000; Hong et al., 2008; Jiao et al., 2011), but can also modulate the expression of the chloride transporter KCC2 (see below), and the surface expression, localization and function of GABA_ARs.

NO-mediated I-LTP of GABA release has been reported in neocortex and other brain areas. In layer 5 pyramidal neurons of the somatosensory cortex, postsynaptic Ca²⁺ rise following repetitive firing activates nitric oxide synthase (NOS) and mobilizes NO to induce presynaptic I-LTP (Lourenco et al., 2014). NO readily permeates through the membrane and presumably stimulates presynaptic guanylate cyclase (GC), thereby augmenting cGMP levels to enhance GABA release via an unknown mechanism. The NO-dependent increase in GABA release is selective for perisomatic inhibition from PV+ but not SST+ interneurons (Lourenco et al., 2014) indicating cell-type specificity for this form of ISP. In several other brain areas (Castillo et al., 2011), NO-mediated I-LTP is also induced by repetitive activation of glutamatergic inputs and NMDAR-mediated postsynaptic Ca²⁺ rise, which activates NOS. However, evidence for this heterosynaptic mechanism of ISP at hippocampal and neocortical synapses is lacking.

Retrograde signaling is a highly regulated process (Iremonger et al., 2013), allowing for multiple points of modulation and cross-talk with several signaling systems. For example, theta-burst stimulation in layer 2/3 neurons of somatosensory cortex induces I-LTD that requires BDNF-TrkB signaling, but not mGluR-I activation (Zhao et al., 2015). Activation of presynaptic, G_i/o-coupled type 2 dopamine receptors (D₂R) act synergistically with CB₁ downstream signaling to induce I-LTD in the prefrontal cortex (Chiu et al., 2010), suggesting additional layers of modulatory complexity in presynaptic terminals. Lastly, in CA1 pyramidal neurons, STDP-induced I-LTD requires coactivation of (presumably presynaptic) G_i/o-coupled M₂-type muscarinic acetylcholine receptors and eCB signaling (Ahumada et al., 2013).

ISP can be induced homosynaptically in the absence of retrograde signaling (Castillo, 2012). This is the case of fast-spiking interneurons in layer 2/3 of the mouse visual cortex where repetitive activity of these neurons can induce presynaptic I-LTP that likely relies on presynaptic calcium influx via P/Q type calcium channels (Sarihi et al., 2012). Adding to the diversity of presynaptic mechanisms for ISP, there is evidence that neuromodulatory systems can also induce ISP. A good example can be found at PV+ to CA2 pyramidal cell synapses, where activation of presynaptic G_i/o-coupled delta-opioid receptors, likely acting on the release machinery, induces a form of presynaptic I-LTD reminiscent of eCB-dependent I-LTD (Piskorowski & Chevaleyre, 2013). Other presynaptic metabotropic receptors (Atwood et al., 2014) may also contribute to the induction of ISP.

Postsynaptic forms of ISP

In many cases activity-dependent ISP relies on postsynaptic modifications such as changes in the properties or number of GABA_ARs expressed at the synapse (Luscher et al., 2011; Vithlani et al., 2011). Here, we briefly summarize the induction and expression mechanisms underlying this form of plasticity. Postsynaptic ISP can be induced by repetitive firing of the postsynaptic neuron, coordinated activity of the GABAergic interneuron and the postsynaptic neuron, or heterosynaptically by repetitive activation of nearby glutamatergic synapses (Gaiarsa et al., 2002; Castillo et al., 2011; Lourenco et al., 2014; Chiu et al., 2019). A common theme is that Ca²⁺ rise either via VGCCs or NMDARs sets in motion a metabolic cascade of events that leads to changes in GABA_AR function or number. In layer 5 pyramidal neurons of rat visual cortex, repetitive firing from a depolarized membrane potential, induces I-LTD at somatic inhibitory synapses, whereas cell firing during slow membrane voltage oscillations induces I-LTP. Interestingly, while I-LTD is mediated by L-type calcium channels, I-LTP is mediated by R-type calcium channels, suggesting that the relative amount of calcium entry through these channels determines the polarity of this ISP (Kurotani et al., 2008). As mentioned above, timing of pre and postsynaptic activity is also important for the polarity of ISP (Holmgren & Zilberter, 2001), although the relationship between timing and sign of plasticity differs from that reported for excitatory synapses (Bi & Poo, 2001). Pairing of presynaptic burst activity and subthreshold postsynaptic depolarization also induces I-LTP at PV+ and layer 4 pyramidal neuron synapses in the developing visual cortex (Maffei et al., 2006; Wang & Maffei, 2014). This plasticity is occluded by visual deprivation, a manipulation that potentiates inhibitory transmission, suggesting that I-LTP can occur in vivo.

Glutamatergic activity can induce heterosynaptic, postsynaptic ISP by activating ionotropic (typically NMDARs) or mGluRs. (Ouardouz & Sastry, 2000). This form of postsynaptic I-LTP has recently been reported at dendritic synapses between SST+ interneurons and layer 2/3 pyramidal neurons of the mouse medial prefrontal cortex (Chiu et al., 2018). I-LTD induction in the CA1 area of the hippocampus relies on a similar mechanism (Lu et al., 2000). Earlier studies (Aizenman et al., 1998) documented a form of NMDAR-dependent I-LTP requiring postsynaptic calcium rise in DCN neurons induced by high frequency stimulation of glutamatergic inputs.

The expression mechanisms of postsynaptic ISP are diverse. Increases or decreases in channel function can occur as a result of GABA_AR phosphorylation by multiple kinases, including, CaMKII, PKC, PKA and Src (Vithlani et al., 2011). GABA_ARs are further controlled by constitutive cycling, regulating insertion and removal, as well as lateral diffusion at the synaptic membrane surface (Luscher et al., 2011; Vithlani et al., 2011; Maynard & Triller, 2019; Pizzarelli et al., 2019; Tomita, 2019). The number of GABA_ARs at the cell surface is regulated by endocytosis and exocytosis. Evidence that postsynaptic I-LTP is due to GABA_AR insertion in the plasma membrane comes from the observation that intracellular loading of botulinum toxin or tetanus toxin, which blocks vesicular exocytosis and insertion of new receptors into the plasma membrane, reduces this potentiation in layer 5 pyramidal neurons in the rat visual cortex (Kurotani et al., 2008). The activation of NMDARs by either exogenous agonists or endogenous glutamate potentiates GABAergic synapses from SST+, but not PV+, interneurons in pyramidal neurons of layer 2/3 of prefrontal cortex (Chiu et al., 2018). Furthermore, intracellular cellular loading of a peptide that interferes with endocytosis, not only blocks I-LTD but uncovers I-LTP in layer 5 pyramidal neurons (Kurotani et al., 2008).

The molecular mechanisms involved in postsynaptic GABA_AR plasticity have been largely investigated in heterologous systems and cultured neurons where the circuit architecture is not preserved and where “plasticity” is commonly induced using various chemical methods (Vithlani et al., 2011; Petrini et al., 2014; Maynard & Triller, 2019; Pizzarelli et al., 2019; Tomita, 2019). While informative, these studies may not capture the nuances of synaptic inhibition and activity-dependent ISP (i.e. I-LTP, I-LTD) in vivo, including the remarkably diverse interneuron subtypes, and the unique postsynaptic sub-cellular compartments (e.g. soma, axon, proximal and remote dendrites) (Chiu et al., 2019). While new molecular players continue to be discovered (Uezu et al., 2016; Martenson et al., 2017) and the nanoscale molecular structure of the GABAergic postsynaptic density is better understood (Pennacchietti et al., 2017; Crosby et al., 2019), more work is required to elucidate the molecular basis underlying GABA_AR regulation in postsynaptic ISP, especially in preparations where the circuit architecture is preserved. This includes the precise role of scaffolding and auxiliary proteins in receptor trafficking, and the postsynaptic molecular determinants of ISP heterogeneity.

Postsynaptic GABAergic plasticity can occur independently of direct receptor regulation. In hippocampal neurons, coincident pre- and postsynaptic activity alters the activity of the K⁺/Cl⁻ co- transporter KCC2, resulting in long-lasting changes in the reversal potential of GABAergic synaptic currents (Woodin et al., 2003). In this case, plasticity is induced by the activation of postsynaptic VGCCs. This mechanism should affect all inhibitory synapses, in a neuron type-independent manner. This plastic change in GABAergic synaptic strength, which is dependent on coincident pre- and postsynaptic spiking, should set the level of inhibition in accordance to the temporal pattern of postsynaptic excitation (Woodin et al., 2003). ISP induced by regulation of the reversal potential of GABA currents has powerful effects on the ability of pyramidal neurons to generate action potentials (Saraga et al., 2008), providing powerful control over activity propagation in neural circuits.

Inhibitory plasticity and circuit function

A growing body of evidence highlighted the contribution of inhibitory neurons to a variety of neural circuit functions (Maffei, 2017). Different populations of GABAergic neurons are activated during behaviors (Kepecs & Fishell, 2014). However, very little is currently known about the function of GABAergic ISP.

Several circuit computations are thought to depend on inhibitory circuits (Kepecs & Fishell, 2014). Experimental results and computational/theoretical approaches strongly suggest that GABAergic circuits are engaged in establishing the flow of signals in cortical circuits. A prominent example is lateral inhibition, a process by which the flow of signals is directed by dampening the activation of neighboring neurons (Kayser & Miller, 2002). GABAergic inhibition is also thought to contribute to adjusting the excitability of principal neurons through subtractive or divisive normalization (Pouille et al., 2009; Silver, 2010; Mejias et al., 2014; Seybold et al., 2015; Bhatia et al., 2019): the modulation of neurons input/output function that determines the sensitivity of neurons responsiveness to small changes in incoming input (gain). Differences in tuning of responses to stimuli between principal neurons, relatively narrowly tuned to incoming inputs, and GABAergic inhibitory neurons, mostly broadly tuned (but see (Moore & Wehr, 2013)), contribute to adjusting the magnitude of principal neurons responsiveness to incoming activity (Cardin et al., 2007).

Experimental work showed that somatodendritic targeting inhibitory neurons are primarily involved in this process (Cardin et al., 2008; Pouille et al., 2013; Keller et al., 2018), though a contribution of perisomatic targeting neurons to gain modulation has also been reported (Atallah et al., 2012; Lourenco et al., 2020). These apparently opposing findings raise the possibility that distinct inhibitory neuron groups could contribute to the same computational function through different mechanisms, and that the recruitment of specific neuron groups or mechanisms may depend on local circuit connectivity or network state (Mejias & Longtin, 2014).

At the network level, inhibitory neurons are thought to contribute to network synchronization and to the generation of oscillatory patterns of activity (Vierling-Claassen et al., 2010; Avoli, 2019). The connectivity of neural circuits and the distinct firing properties of different neuron groups (e.g. fast spiking GABAergic neurons versus regular spiking principal cells) can shape network oscillations by the alternation of neurons activation (Cardin et al., 2009; Drexel et al., 2017; Panthi & Leitch, 2019). As specific oscillatory rhythms are associated with distinct cognitive functions, alterations in the coupling between excitatory and inhibitory neurons can results in changes in networks states (Francavilla et al., 2018; Pala & Petersen, 2018).

The theoretical framework established to better investigate these functions tends to consider inhibition as a uniform system that is broadly connected and primarily recurrent (van Vreeswijk & Sompolinsky, 1996; Vogels & Abbott, 2009; Kanashiro et al., 2017; Huang et al., 2019). This approach is supported by reports of widespread connectivity of different groups of inhibitory neurons (Fino & Yuste, 2011; Fino et al., 2013) suggesting that GABAergic inhibition may act by modulating circuits broadly without specificity. It is also supported by studies showing GABAergic neurons broad tuning to incoming stimuli (Cardin et al., 2007; Li et al., 2015a; Li et al., 2015b; Hayashi et al., 2018; Li et al., 2019), a property that renders them not particularly sensitive to the granular features of an input.

However, there is evidence that the tuning of inhibitory neurons to incoming stimuli is not always broader than that of principal cells (Moore & Wehr, 2013; Camillo et al., 2018; Khan et al., 2018), and can be modified by experience (Dorrn et al., 2010; Cai et al., 2018). While it is reliably found that inhibitory neurons connect broadly to principal neurons in cortical circuits, specificity can be achieved through selective targeting of neuronal compartments by distinct groups of neurons with distinct molecular identities (Somogyi et al., 1983; Thomson et al., 1996; Tamas et al., 2003; Wang et al., 2004; Chamberland & Topolnik, 2012; Paul et al., 2017), or by activity-dependent changes in inhibitory synapses efficacy that depend on the level of activity of the postsynaptic neuron (Wang & Maffei, 2014; Chiu et al., 2018; Lourenco et al., 2020). Furthermore, inhibitory synapses can change their efficacy in response to shifts in circuit excitability (Maffei et al., 2004; Maffei et al., 2006), a diverse array of patterns of activity (Woodin et al., 2003; Chevaleyre et al., 2007; Younts et al., 2013; Petrini et al., 2014; D’Amour J & Froemke, 2015), can display distinct mechanisms depending on the identity of the presynaptic neuron (Chiu et al., 2018) and the level of activity of the postsynaptic neuron (Vogels et al., 2013; Wang & Maffei, 2014) supporting the interpretation that inhibition and ISP contribute to sculpting circuit activity and function in a more refined fashion than previously anticipated.

These experimental findings, together with novel approaches to neural network models that incorporate feedforward inhibition (Troyer et al., 1998; Miller, 2003; Miska et al., 2018), differential effects of distinct population of inhibitory neurons (Vierling-Claassen et al., 2010; Hertag & Sprekeler, 2019; Wilmes & Clopath, 2019) and plasticity rules for inhibitory synapses (Vogels et al., 2011; Wilmes & Clopath, 2019), provide new insights into the contributions of inhibition and ISP to circuit function.

Inhibition and ISP as mechanisms to stabilize circuit excitability

One of the better characterized roles of changes in efficacy of inhibitory synapses is that of stabilizer (or homeostatic regulator) of circuit excitability (Turrigiano et al., 1998). Experimental work showed that the efficacy of GABAergic inhibitory synapses can change in response to network activity in a compensatory fashion: acting to preserve excitatory neurons excitability in the absence of circuit activity (Kilman et al., 2002). A similar effect was observed in response to manipulation of sensory drive in rodents, whereby loss of visual drive induced two distinct forms of plasticity of GABAergic synapses at soma-targeting and dendrite-targeting inhibitory neurons (Maffei et al., 2004). The idea that ISP can participate in the maintenance of circuit excitability was further confirmed in a variety of different cortical circuits and its implications have been further explored experimentally (House et al., 2011; Campanac et al., 2013; Graupner & Reyes, 2013; Xue et al., 2014; Froemke, 2015) and with theoretical/computational approaches (van Vreeswijk & Sompolinsky, 1996; Vogels & Abbott, 2009; Lo et al., 2015; Rubin et al., 2017).

The hypothesis emerging from this work is that balanced excitation and inhibition is crucial for neural circuit function and that ISP plays an important role in this balancing process. The displacement from balance is thought to be needed for processing incoming inputs, and subsequent induction of ISP would bring the system back to a stable set point (Nelson & Turrigiano, 2008). Under these premises, changes in inhibitory synaptic efficacy are induced to adjust inhibitory tone and oppose shifts in network activity (Vogels et al., 2011). Overall, this principle is conceptually quite similar to the proposed role for homeostatic plasticity (Turrigiano & Nelson, 2000). Nevertheless, several neural network models implement a stabilizing form of ISP by adapting Hebbian-like rules (Vogels et al., 2013), and approach that has the advantage of modulating the induction of ISP depending on the level of circuit activity, as well as on the timing of pre- and postsynaptic activity (Vogels et al., 2013; Wang & Maffei, 2014). Models including ISP provide a substantial advancement from models that do not include it, as they do not require as much parameter adjustment, and can recapitulate a variety of experimental results. However, even these models are not fully capable of reflecting the complex, self-generating dynamics of a neural network and are not yet capable to account for the diversity of forms of ISP of which spike timing-dependent forms are just a subset (Chevaleyre et al., 2007; Kullmann et al., 2012; Wang & Maffei, 2014).

GABAergic inhibitory synaptic plasticity, learning and memory

The initial reports that GABAergic synapses are plastic, suggested that ISP contributes to cortical circuit refinement (Komatsu & Iwakiri, 1993; Maffei et al., 2004; Maffei et al., 2006) and possibly learning (Kano, 1995; Buzsaki, 1997; Nugent et al., 2007). Since these pioneering studies, experimental work demonstrated that changes in inhibitory synaptic efficacy are associated with adaptation of olfactory responses in drosophila (Das et al., 2011), with changes in sensory experience (Maffei et al., 2004; Maffei et al., 2006; House et al., 2011; D’Amour J & Froemke, 2015; Cai et al., 2018), fear conditioning (Letzkus et al., 2011; Xu et al., 2014; Lucas et al., 2016) and exposure to drugs of abuse (Nugent et al., 2007; Dacher & Nugent, 2011). The specific mechanisms by which ISP contributes to learning and memory are only beginning to being unraveled. Experimental work is currently focused on determining the mechanisms of ISP (Maffei, 2011; Kullmann et al., 2012; Younts & Castillo, 2014; Chiu et al., 2019) and the role of ISP in information transfer (Lourenco et al., 2020). Computational work has only recently begun to introduce ISP in network models (Vogels et al., 2011; Mongillo et al., 2018) and theoretical ideas are being developed regarding its functional role (Barron et al., 2017; Hennequin et al., 2017; Sprekeler, 2017; Mongillo et al., 2018).

The efficacy of inhibitory synapses is sensitive to changes in sensory inputs (Maffei et al., 2004; Maffei et al., 2006; Dorrn et al., 2010; Takesian et al., 2012; Kotak et al., 2013; Gainey et al., 2016). This is especially prominent during windows of heightened sensitivity to changes in sensory drive (Hensch, 2005b; Sanes & Kotak, 2011; Li et al., 2014b; Cai et al., 2018). A model tested the possibility that gradient-like changes in inhibitory tone may contribute to the experience-dependent development of ocular dominance in rodent visual cortex (Toyoizumi et al., 2013). This idea was based on results showing that the postnatal maturation of inhibitory synaptic drive is essential for regulating onset and duration of the critical period for visual cortical plasticity (Hensch, 2005a). While this effect is thought to depend on the maturation of PV+ inhibitory neurons (Katagiri et al., 2007), inhibitory tone could also potentially be modulated by changes in synaptic transmission from different inhibitory neuron types, including NGFCs which signal through both synaptic and volume transmission (Olah et al., 2009)

Recent theoretical work suggests that ISP also plays a central role in learning and memory (Vogels et al., 2011; Wilmes & Clopath, 2019). A current working hypothesis is that ISP is induced to rebalance the network following plastic changes at excitatory synapses (Vogels et al., 2011; Sprekeler, 2017). Potentiation of glutamatergic synapses is associated with memory formation and the emergence of memory engrams (Fig.3A) (Ryan et al., 2015; Tonegawa et al., 2015; Bocchio et al., 2017). Nevertheless, the formation of multiple memory engrams that can partially overlap, may saturate the dynamics range of neural circuits, leading to saturation of storage capacity (Rashid et al., 2016; Mongillo et al., 2017). Recent experimental work proposed that ISP may act to protect stored memories by contributing to inhibitory engrams that rebalance circuit activity (Fig. 3B) (Koolschijn et al., 2019). The definition of inhibitory engram is currently ambiguous: it is not specified whether it is a counter-engram that depends on ISP at inhibitory synapses onto the principal neurons participating in the memory engram, or a group of coactivated inhibitory neurons counterbalancing the excitability of the engram circuit, but not directly connected to the engram neurons. In network models, the inhibitory engram mediated by ISP would stabilize network excitability while preserving relative differences at excitatory connections (Abbott & Nelson, 2000; Renart et al., 2003). This would preserve memories and maintain the dynamic range of circuit excitability so that it is permissive to the formation and storage of new memories. Conceptually, the role inhibitory engrams is an upgraded version of the well-established idea that the primary role of inhibition is to keep circuit excitability in check. In this case ISP would be induced to prevent overactivation of a subcircuit of excitatory neurons (Barron et al., 2017). Instead of providing widespread control on excitability, ISP would act in a more restricted fashion.

Figure 3. Inhibitory plasticity as a mechanism for circuit reconfiguration.

Simplified diagram of how inhibitory plasticity can increase the dynamic range of neural circuits. Black and grey: pyramidal neurons; yellow: soma targeting inhibitory neurons; blue: dendrite targeting inhibitory neurons. A. Plasticity at excitatory synapses is thought to facilitate the activation of memory storage units (engrams). Green shaded areas and thick lines indicate the site of LTP at excitatory connections. The darker neurons highlight an example engram, while the surrounding faded neurons are part of the circuit but do not participate in the excitatory engram. B. In this diagram, the excitatory engram shown in A is paired with an inhibitory engram triggered by ISP. As the inhibitory engram is not defined to be connection-specific (Hennequin et al., 2017), the area of influence of the inhibitory engram is indicated by the red shaded area. Greed shaded areas and thick lines indicate LTP at excitatory synapses. Neurons not participating in either the excitatory or the inhibitory engrams are shown in lighter colors. C. The diagram shows how connection-specific LTP and LTD at excitatory and inhibitory synapses can expand the dynamic range of the circuit, facilitating the accommodation of multiple engrams. Large green shaded areas and thick black lines indicate LTP at excitatory synapses; small green shaded areas and thin black lines indicate LTD at excitatory synapses. Large red shaded areas and thick yellow and blue lines indicate LTP at inhibitory synapses; small red shaded areas and think yellow and blue lines indicate LTD at inhibitory synapses. In this diagram both excitatory and inhibitory plasticity are connection-specific. Cooperative interactions between excitatory and inhibitory plasticity are taken into account to indicate the site of LTP and LTD at both populations of synaptic connections (see (Wang & Maffei, 2014; Hennequin et al., 2017)).

A caveat of the inhibitory engram idea stems from its reliance of a temporal sequence in which the induction of excitatory plasticity precedes changes in inhibitory synaptic efficacy (Hennequin et al., 2017): plasticity at excitatory synapses is induced during learning, shifting the balance between excitation and inhibition, then ISP is induced to restore balance (Vogels et al., 2011). Experimental evidence in favor of such a sequence of events has been reported in some cortical circuits (Kuhlman et al., 2013). Nevertheless, changes in inhibition can be induced before or together with excitatory plasticity at recurrent synapses (Wang & Maffei, 2014; Gainey et al., 2018). Firing rates of inhibitory neurons recover more rapidly than those of excitatory neurons following brief sensory deprivation (Hengen et al., 2013), indicating that the temporal sequence suggested by the inhibitory engram hypothesis is not generally applicable. Furthermore, the baseline ratio of excitatory and inhibitory charge onto cortical pyramidal neurons is shifted toward inhibition (Tatti et al., 2017), suggesting that disinhibition may be needed for excitatory plasticity to be induced. Finally, the ratio of excitation and inhibition is circuit and experience-dependent (Tatti et al., 2017; Bridi et al., 2019), adding complexity to how ISP may affect learning. These data are consistent with theoretical work suggesting that sparse cortical activity depends on inhibition being dominant, and with experimental results indicating that disinhibition of cortical circuits is permissive for learning (Froemke et al., 2007; Letzkus et al., 2011; Kuhlman et al., 2013). This alternative framework suggests that depression of inhibitory synaptic transmission facilitates the induction of excitatory plasticity correlated with the formation of associations between conditioned and unconditioned stimuli. In support of this idea, a study demonstrated that excitatory and inhibitory plasticity cooperatively interact in an anticorrelated fashion, with potentiation of inhibition facilitating depression of convergent excitatory synapses, and depression of GABAergic synapses enabling potentiation of glutamatergic input onto the same postsynaptic target (Wang & Maffei, 2014). As this crosstalk is connection-specific and relies on postsynaptic activity (Maffei et al., 2006; Wang & Maffei, 2014), it has sufficient resolution to modulate single cortical connections instead of acting as a broad balancing factor.

Finally, it is unclear whether long term ISP would be necessary if its role is to restore the balance between excitation and inhibition. Homeostatic synaptic plasticity can be a slow process effectively restoring stability over the course of several hours to a few days (Kilman et al., 2002; Lambo & Turrigiano, 2013; Whitt et al., 2014; Glazewski et al., 2017). Such time course is compatible with the idea that changes in inhibitory synaptic transmission restore circuit excitability to a balanced state. However, ISP in the form of I-LTP or I-LTD is typically induced rapidly, is long-lasting (Woodin et al., 2003; Chevaleyre et al., 2007; Nugent et al., 2007; Wang & Maffei, 2014), and recruits complex post translational mechanisms that can lead to structural and functional changes in inhibitory connectivity (Tyagarajan et al., 2011; Petrini et al., 2014; Ghosh et al., 2015). These effects are comparable to what is typically reported for long term plasticity at excitatory synapses (Herring & Nicoll, 2016).

It has been proposed that ISP can act as a form of metaplasticity (Chevaleyre & Castillo, 2004; Xu et al., 2014), a set of changes that alter the state of a principal neuron modulating its responsiveness to incoming activity (Abraham, 2008). This hypothesis is compatible with the experimental results discussed above, but does not fully explain the duration of the changes in synaptic strength, unless we consider ISP as a direct contributor to memory storage. Indeed, a long lasting state change in a neuron, whether excitatory or inhibitory, may serve as memory signature shaping how this neuron (or circuit) will be activated by future stimuli. Instead of contributing to rebalance, or protect, the memory engram activated by excitatory plasticity, long lasting changes in GABAergic synapses and their plasticity would expand the dynamic range of circuit plasticity, thus being a central component of the memory trace (Fig 3C).

Conclusions

Inhibitory synaptic transmission is extremely diverse and undergoes various mechanistically distinct forms of activity-dependent plasticity. Such diversity endows neural circuits with a broad repertoire for inhibitory control of neural circuits. By sampling inputs widely, as GABAergic neurons are often less narrowly tuned than their excitatory counterpart, and connecting broadly to many excitatory neurons, inhibitory circuits have the capacity of influencing cortical computations quite significantly. Their plasticity - which can result in changes in GABA release, or in synapse-specific or connection-specific changes in GABAergic responses - the diversity of neuron types, postsynaptic target regions and cellular mechanisms available to change synaptic efficacy, make them fundamental contributors to complex and sophisticated processes including learning, memory and other cognitive functions.

Abbreviations:

2-AG

2-arachidonoylglycerol

BC

basket cell

BDNF

brain-derived neurotrophic factor

CB₁

type 1 cannabinoid receptors

CCK

cholecystokinin

DAG

diacylglyercol

DGL

diacylglycerol lipase

D₂R

type 2 dopamine receptors

eCB

endocannabinoid

EPSP

excitatory postsynaptic potential

ERK

extracellular signal-regulated kinase

FB

feedback

FF

feedforward

FSI

Firing Induced Suppression of Inhibition

GABA

gamma amino butyric acid

GC

guanylate cyclase

ISP

inhibitory synaptic plasticity

KCC2

potassium chloride co-transporter

LTD

long-term depression

LTP

long-term potentiation

mGluR-I

group I metabotropic glutamate receptors

NGFC

neurogliaform cell

NL-2

neuroligin-2

NMDA

N-methyl-D-aspartate

nNOS

neuronal nitric oxide synthas

NO

nitric oxide

PLC

phospholipase C

PV

parvalbumin

PKC

protein kinase C

scRNAseq

single cell RNA sequence

STDP

Spike timing dependent plasticity

TrkB

tropomysin related kinase B

VGCC

voltage-gated calcium channel