Activity manipulations in cultures of dissociated hippocampal or neocortical neurons generally affect excitatory and inhibitory synapses in opposite directions. After a prolonged period of activity blockade, excitatory synapses get strengthened and inhibitory synapses are weakened, and synaptic changes are in opposite directions when activity is enhanced (Turrigiano et al., 1998; Kilman et al., 2002; Hartman et al., 2006; Swanwick et al., 2006). Therefore, changes in excitation and inhibition cooperate to compensate for the change in activity level. For inhibitory synapses, changes in mIPSC amplitude are most commonly reported, reflecting changes in synaptic strength. Sometimes they are accompanied by changes in mIPSC frequency, which could either reflect a change in the number of synapses or a change in release properties. Dissociated cultures provide excellent experimental access and are therefore well-suited for studying underlying mechanisms of homeostatic plasticity. However, the artificial environment in which neurons grow in culture may affect synaptic maturation (Wierenga et al., 2006; Rose et al., 2013) and consequently cellular or synaptic mechanisms of plasticity. Cellular mechanisms that were identified to mediate the changes in inhibitory synapses after activity manipulations include: changes in number of postsynaptic receptors (Kilman et al., 2002; Swanwick et al., 2006; Saliba et al., 2007; Peng et al., 2010; Rannals and Kapur, 2011) or scaffolding proteins (Vlachos et al., 2012; study in slice cultures) on the postsynaptic side, and changes in presynaptic release probability (Kim and Alger, 2010), presynaptic vesicle loading (De Gois et al., 2005; Hartman et al., 2006; Lau and Murthy, 2012), or GABA-producing enzymes (Peng et al., 2010; Rannals and Kapur, 2011) on the presynaptic side. Only in a few cases, changes in the number of inhibitory synapses were reported (Hartman et al., 2006; Peng et al., 2010). Homeostatic changes of inhibitory synapses could be induced in a cell autonomous fashion (Peng et al., 2010), or required a change in activity of the entire neuronal network (Hartman et al., 2006), emphasizing that there are multiple mechanisms of homeostatic plasticity at inhibitory synapses. In particular, distinct mechanisms could exist for activity-dependent downregulation and upregulation of inhibitory synapses.

In contrast to dissociated cultures neurons in more intact tissue, such as acute slices or organotypic cultures, make more specific connections and form structured networks. This network configuration makes the interpretation of synaptic changes more complex. In slices that were submitted to activity manipulations, changes in inhibition have been observed opposite to (Marty et al., 2004; Karmarkar and Buonomano, 2006; Kim and Alger, 2010) as well as in conjunction with (Buckby et al., 2006; Echegoyen et al., 2007) changes in excitation. It was also shown that different types of homeostatic mechanisms have different time courses (Karmarkar and Buonomano, 2006) and that different subsets of inhibitory synapses can respond differently. For instance, the presence of cannabinoid receptors in a subset of inhibitory synapses renders them selectively receptive to changes in endocannabinoid levels induced by inactivity (Kim and Alger, 2010). In another example, inactivity differentially affected somatic and dendritic inhibitory inputs on CA1 pyramidal cells. Interestingly, both types of synapses showed reduction in the number of presynaptic boutons and upregulation of release probability, but the functional end-effect on inhibitory input to the postsynaptic cells was different (Chattopadhyaya et al., 2004; Bartley et al., 2008). This emphasizes that simple in vitro homeostatic rules for scaling inhibitory synapses get complicated in more complex networks. In addition, other factors such as different cell (glia) types or the extracellular environment in more intact tissue potentially influence homeostatic plasticity compared to dissociated cultured cells.

Typically, when studying activity-dependent or homeostatic changes in vivo, sensory deprivation is used as experimental paradigm to lower activity levels in the primary sensory cortex (e.g., whisker trimming, monocular deprivation, or retinal lesion). While in vitro activity manipulations by pharmacological means affect the activity of all neurons in equal amounts, sensory deprivation in vivo will affect different neurons in the circuitry differentially. Therefore, in vivo responses of inhibitory synapses to changes in activity vary widely and strongly depend on the specific cell types, cortical layer, and specific circuitry (Maffei et al., 2004; Maffei and Turrigiano, 2008; Chen et al., 2011). Furthermore, it is well-known that inhibition in sensory cortex areas undergoes important developmental changes (Hensch, 2005), which means that the same deprivation paradigm can have different effects on inhibitory synapses depending on the postnatal period that is considered (Chattopadhyaya et al., 2004; Maffei et al., 2006; Maffei et al., 2010). An emerging theme from the in vivo studies is that inhibitory synapses can respond rapidly to sensory deprivation. It was shown that inhibitory axons in cortical layer 2/3 reduce the number of boutons within the first 24 h after a retinal lesion or monocular deprivation (Chen et al., 2011; Keck et al., 2011). Over longer periods, inhibitory axons in the barrel cortex were shown to sprout and form new axonal branches after whisker plucking (Marik et al., 2010). Interestingly, the reduction of inhibition was often found to precede adaptive changes of the excitatory circuitry (Marik et al., 2010; Keck et al., 2011). The rapid downregulation of inhibition might serve to render the local circuit more permissive for excitatory plasticity to occur (Ormond and Woodin, 2011; Gambino and Holtmaat, 2012). In two recent studies it was shown that inhibitory synapses that are located on spines (presumably next to an excitatory synapse) showed much higher turnover rates compared to inhibitory synapses on shaft after visual deprivation (Chen et al., 2012; vanVersendaal et al., 2012). It will be interesting to see whether direct cross talk of the two types of synapses exists.

In conclusion, there is a large amount of compelling evidence for activity-dependent adaptations in inhibitory synapses in vitro as well as in vivo. The precise expression mechanisms significantly vary between different preparations and experimental paradigms.

Most of the studies that were mentioned above were performed in excitatory axons and it is not entirely clear to what extent the results are also valid for inhibitory axons. Important observations have been made in live imaging studies of inhibitory axons. Presynaptic terminals in inhibitory axons were shown to be dynamic structures in vitro and in vivo. Inhibitory boutons can appear, disappear, and reappear over the course of several minutes to hours (Kuhlman and Huang, 2008; Marik et al., 2010; Keck et al., 2011; Fu et al., 2012; Schuemann et al., 2013), and the same has been shown for clusters of pre- or postsynaptic proteins at inhibitory synapses (Dobie and Craig, 2011; Chen et al., 2012; Kuriu et al., 2012; vanVersendaal et al., 2012). Bouton dynamics are comparable in vitro and in vivo and likely reflect physiological processes. Interestingly, these dynamic changes were shown to be affected by network activity and mediated, at least in part, by activation of GABA receptors (Fu et al., 2012; Kuriu et al., 2012; Schuemann et al., 2013). This could represent a mechanism by which the synaptic activity of inhibitory synapses may regulate their own stability using GABA as a feedback signal.

New inhibitory synapses can form rapidly by the appearance of a bouton at locations where the inhibitory axon is in close contact with a dendrite, without the involvement of dendritic protrusions (Wierenga et al., 2008; Dobie and Craig, 2011). This finding indicates an important contrast with the formation of excitatory synapses, in which new synapses are usually formed by the outgrowth of dendritic protrusions. It also emphasizes the important role of crosstalk between neighboring boutons within inhibitory axons for synapse formation. Nascent inhibitory synapses recruit release machinery proteins and synaptic vesicles on the presynaptic side and receptors and scaffolding molecules on the postsynaptic side within a few hours (Wierenga et al., 2008; Dobie and Craig, 2011; Kuriu et al., 2012; Schuemann et al., 2013). Interestingly, simultaneous translocations of pre- and postsynaptic proteins over several micrometers were observed in cultures (Dobie and Craig, 2011; Kuriu et al., 2012) and it will be interesting to see if such movement of inhibitory synapses can also occur in slices or in vivo. Together, these observations reveal the dynamic nature of inhibitory axons and strongly suggest that the exchange of presynaptic material between existing and emerging boutons within the axonal shaft plays an essential role in the activity-dependent formation, maintenance and plasticity of inhibitory synapses.

In general, it is not clear if molecular differences exist between excitatory and inhibitory axons, other than the neurotransmitter that is produced and loaded into synaptic vesicles. For instance, the extent or regulation of dynamic exchange between boutons could be different in these two types of axons. The protein composition of the release machinery at excitatory and inhibitory presynaptic terminals is surprisingly similar, although small difference have been reported (Gitler et al., 2004; Kerr et al., 2008; Kaeser et al., 2009; Grø, 2010; Zander et al., 2010; Boyken et al., 2013; Bragina et al., 2013). It is currently not known if some of these differences have consequences for plasticity or presynaptic dynamics within axons. Furthermore, it is not known if there are differences between axons of the various inhibitory cell types (Ascoli et al., 2008; Klausberger and Somogyi, 2008). However, there is a clear difference between excitatory and inhibitory axons in the expression of specific cell adhesion molecules at excitatory and inhibitory synapses.

The observation that inhibitory boutons appear at specific, predefined locations along the axon (Sabo et al., 2006; Wierenga et al., 2008; Schuemann et al., 2013), strongly suggests that something is marking these locations prior to bouton formation (Shen and Bargmann, 2003; Shen et al., 2004). Inhibitory axons are characterized by their tortuous and highly branched morphology and they are in close contacts with many nearby dendrites. In fact, it was shown that inhibitory axons have substantially larger overlap with the dendritic trees of their potential target neurons than expected from chance, whereas this is not the case for excitatory axons (Stepanyants et al., 2004). This suggests that inhibitory axons possibly search for or are attracted by dendrites during development. Contacts between dendrites and inhibitory axons could be maintained by guidepost cell-adhesion molecules, even without inhibitory synapses present (Shen and Bargmann, 2003; Shen et al., 2004). Their presence would mark the location of a postsynaptic dendrite and therefore a potential spot for an inhibitory synapse.

Cell adhesion molecules are transmembrane proteins, which play a role in recognition of synaptic partners during the initial contact and provide specificity of synaptic connections (Meijers et al., 2007; Wojtowicz et al., 2007). In addition, cell adhesion molecules have been shown to play a role in the process of synaptic maturation following the initial contact, in the recruitment of synaptic proteins as well as in maintaining proper synaptic function throughout the lifetime of the synapse (Dalva et al., 2007; Krueger et al., 2012; Thalhammer and Cingolani, 2013). Cell adhesion molecules can also play an active role in the process of synapse disassembly (O'Connor et al., 2009). In conclusion, cell adhesion molecules are an essential part of synapses and synaptic plasticity most likely involves regulation of cell-adhesion molecules. Here we discuss how synaptic adhesion could be regulated in an activity-dependent manner (Figure 2) and we summarize current knowledge of cell adhesion molecules that are specific for inhibitory synapses.

Postsynaptic neuroligins and their presynaptic partners neurexins are transmembrane cell adhesion molecules that have been established as important synaptic regulators (Südhof, 2008; Siddiqui and Craig, 2011; Krueger et al., 2012). When expressed in non-neuronal cells, neurexins as well as neuroligins can induce the formation of synapses in co-cultured neurons (Graf et al., 2004; Kang et al., 2008). This suggests that neurexins and neuroligins function in the initial assembly of synaptic connections. However, knock out studies showed that they are not strictly required for synaptogenesis, but they play a crucial role in the proper assembly and functional maturation of synapses (Varoqueaux et al., 2006). Neuroligin-2 localizes specifically to the postsynaptic membrane of inhibitory synapses (Varoqueaux et al., 2004; Chubykin et al., 2007) and has been shown to be a regulator of inhibitory synapse formation and function (Varoqueaux et al., 2006; Chubykin et al., 2007; Poulopoulos et al., 2009). Interestingly, a recent report suggested that the preferential localization of neuroligin-2 at inhibitory synapses can be contributed to the low abundance of β-neurexin1 in inhibitory axons (Futai et al., 2013), suggesting that the presynaptic axon determines specificity of cell adhesion molecules at inhibitory synapses. Mice lacking neuroligin-2 show impairments in inhibitory synaptic transmission and exhibit anxiety-like behavior and increased excitability (Blundell et al., 2009; Gibson et al., 2009; Jedlicka et al., 2011). Interestingly, although neuroligin-2 is present at all inhibitory synapses, only perisomatic synapses were affected in the absence of neuroligin-2 (Gibson et al., 2009). Recently, two adhesion molecules were found to show specific interactions with neuroligin-2 at inhibitory synapses. MDGA1 inhibits the interaction between neuroligin-2 and neurexins and therefore specifically suppresses the inhibitory synaptogenic activity of neuroligin-2 (Lee et al., 2013; Pettem et al., 2013). IgSF9 specifically localizes at inhibitory synapses on inhibitory neurons, where it binds to neuroligin-2 via the scaffolding protein S-SCAM (Woo et al., 2013). These findings raises the possibility that neuroligin-2 serves different functions at different inhibitory synapses, depending on its interactions with other cell adhesion molecules.

Leucine-rich repeat (LRR) proteins have received considerable research attention recently. The members of the subfamily of Slitrk (Slit and Trk-like) proteins are involved in synapse formation and has been linked to several neurological disorders (Aruga and Mikoshiba, 2003; Takahashi and Craig, 2013). Slitrk3 has been shown to be present at the postsynaptic side of inhibitory synapses and it can induce the formation of inhibitory synapses through its interaction with the presynaptic tyrosine phosphatase receptor PTPδ (Takahashi et al., 2012; Yim et al., 2013). Here, the specificity for inhibitory synapses is dictated by the postsynaptic slitrk3, as it was shown that presynaptic PTPδ can interfere with other synaptic organizing molecules to promote formation of excitatory synapses (Yoshida et al., 2011, 2012). The slitrk3 knock out mouse has no gross defect in brain morphology, but shows decreased expression of inhibitory markers (Takahashi et al., 2012). Accordingly, these mice have an increased susceptibility for seizures and sometimes display spontaneous seizures. Interestingly, not all inhibitory synapses were equally affected by the loss of slitrk3. In the hippocampal CA1 region, specifically inhibitory synapses in the middle of the pyramidal layer were lost (Takahashi et al., 2012). It will be interesting to examine whether specificity of inhibitory synapses correlates with different subsets of pre- or postsynaptic neurons types or function.

Members of the closely related subfamily of leucine-rich transmembrane proteins (LRRTMs) have also been implicated in excitatory synapse formation and plasticity (Linhoff et al., 2009; Ko et al., 2011; de Wit et al., 2013; Siddiqui et al., 2013), but so far no LRRTM that is specific for inhibitory synapses has been identified.

Semaphorins are well-known as (repulsive) axon guidance molecules acting through rearrangements of the cytoskeleton in the growth cone. They play an important role in the early development of the brain (Pasterkamp, 2012). Some semaphorins are also expressed later in development and have been implicated in the formation and plasticity of neuronal connections (Sahay et al., 2005; Morita et al., 2006; Paradis et al., 2007; O'Connor et al., 2009; Ding et al., 2012; Mizumoto and Shen, 2013). Knocking down the membrane-bound semaphorin Sema4D was shown to specifically reduce the number of inhibitory synapses, while excitatory synapses were not affected (Paradis et al., 2007). Furthermore, application of soluble Sema4D was able to increase the density of GABAergic synapses within 30 min in rat hippocampal neurons (Kuzirian et al., 2013). These new inhibitory synapses became functional within 2 h and could restore normal levels of activity in an in vitro model for epilepsy (Kuzirian et al., 2013). The effect of sema4D on inhibitory synapses depends on the plexinB1 receptor (Kuzirian et al., 2013). It was earlier shown that activation of plexinB1 by sema4D can induce opposing responses on the cytoskeleton, depending on different interacting proteins (Basile et al., 2004; Swiercz et al., 2008; Tasaka et al., 2012), but the intracellular pathway used for inhibitory synapse formation is not known. Sema4D is a membrane-bound protein, but the protein can also be cleaved (Basile et al., 2007; Zhu et al., 2007). It was recently shown that extracellular cleavage of sema4D occurs in neurons, but does not interfere with its synaptogenic properties at inhibitory synapses (Raissi et al., 2013).

There are many other cell adhesion molecule proteins and with continued research on inhibitory synapses, it is expected that more of them will be found to play a role at inhibitory synapses. Here we just mention a few that have been reported at inhibitory synapses.

Neural cell adhesion molecule (NCAM) has been reported to be important for the maturation of perisomatic inhibitory synapses in the cortex (Pillai-Nair et al., 2005; Brennaman and Maness, 2008; Chattopadhyaya et al., 2013). NCAM acts through activation of Fyn kinases and possibly recruits other adhesion molecules (Chattopadhyaya et al., 2013). Interestingly, it was recently reported that also members of the ephrin family, ephrinA5 and EphA3, can affect inhibitory synapses and they require NCAM for their action (Brennaman et al., 2013). In vivo, NCAM is polysialylated (NCAM-PSA) in an experience-dependent manner and developmental downregulation of NCAM-PSA was shown to coordinate maturation of perisomatic inhibitory synapses in the visual cortex (Di Cristo et al., 2007).

Several components of the dystrophin-associated glycoprotein complex (DGC), such as dystroglycan, dystrophin, and dystrobrevin, are also specifically located at a subset of inhibitory synapses (Knuesel et al., 1999; Brünig et al., 2002; Lévi et al., 2002; Grady et al., 2006), but the function of this complex at inhibitory synapses is not well understood. The DGC could be linked to postsynaptic neuroligin-2 via the scaffolding protein S-SCAM (Sumita et al., 2007) and to presynaptic neurexins (Sugita, 2001). Interestingly, a synaptic guanine exchange factor SynArfGEF has been identified that specifically co-localizes at inhibitory synapses, which could be involved in the downstream signaling pathway of the DGC (Fukaya et al., 2011), but its exact function remains to determined.

Integrins are receptors for extracellular matrix proteins, soluble factors, and counter-receptors on adjacent cells and they have an intracellular link to actin filaments via adaptor proteins (Hynes, 2002; Harburger and Calderwood, 2009). Integrins have been implicated in activity-dependent synaptic changes (Chavis and Westbrook, 2001; Chan et al., 2003) and in homeostatic scaling of excitatory synapses (Cingolani et al., 2008). At glycinergic inhibitory synapses in the spinal cord, postsynaptic β1 and β3 integrins have been reported to regulate glycine receptor stabilization at the postsynaptic membrane, with the two integrins acting in opposing directions (Charrier et al., 2010).

Finally, the cell adhesion molecule neurofascin has been shown to regulate the formation of a specific subset of inhibitory synapses on the axon initial segment of principal neurons (Ango et al., 2004; Burkarth et al., 2007; Kriebel et al., 2011).

Brain-derived neurotrophic factor (BDNF) is a secreted neurotrophin that has been shown in many different preparations to promote the formation and maturation of inhibitory synapses by presynaptic modifications (Vicario-Abejón et al., 1998; Huang et al., 1999; Marty et al., 2000; Yamada et al., 2002; Gottmann et al., 2009). Only excitatory neurons produce BDNF (Gottmann et al., 2009; Park and Poo, 2013) and BDNF is released from principal neurons in an activity-dependent manner (Kolarow et al., 2007; Kuczewski et al., 2008; Matsuda et al., 2009), which makes BDNF an attractive candidate molecule to regulate activity-dependent inhibitory synapse formation (Liu et al., 2007). Interestingly, the availability of postsynaptic BDNF signaling in individual neurons was shown to affect the number and strength of inhibitory synapses specifically onto the affected neurons (Ohba et al., 2005; Kohara et al., 2007; Peng et al., 2010). These cell-autonomous effects indicate the potential for BDNF in mediating changes in inhibitory synapses with high synaptic specificity. In excitatory axons, BDNF was shown to reduce the anchoring of synaptic vesicles at presynaptic terminals and thereby increase the exchange of vesicles between boutons (Bamji et al., 2006). It is currently not known if BDNF has a similar effect in inhibitory axons.

Neuregulin1 is a neurotrophic factor, which exists in various membrane-bound and diffusible isoforms. Mutations (both loss-of-functions and gain-of-function) in neuregulin1 have been linked to schizophrenia (Mei and Xiong, 2008). The main receptor for neuregulin1, ErbB4, is specifically expressed in interneurons (Vullhorst et al., 2009; Fazzari et al., 2010) and is located at postsynaptic densities of excitatory synapses in interneuron dendrites as well as at inhibitory axon terminals. An important role for neuregulin1 is the regulation of excitatory input onto interneurons through postsynaptic ErbB4 (Fazzari et al., 2010; Wen et al., 2010; Ting et al., 2011). Presynaptic ErbB4 can enhance GABA release from inhibitory synapses (Woo et al., 2007; Fazzari et al., 2010) and may affect the number of synapses made by inhibitory axons (delPino et al., 2013). In addition to ErbB4, neuregulin1 isoforms can also activate other receptors resulting in downregulation of postsynaptic GABA_A receptors (Yin et al., 2013). This suggests that neuregulin1 has multiple actions on inhibitory synapses depending on the isoform and activated receptors.

Fibroblast growth factors (FGFs) are secreted signaling glycoproteins, which exert their effect by binding to FGF receptor tyrosine kinases (FGFR). In the brain, FGF signaling is important for several developmental processes, including patterning of different brain structures and neurogenesis (Dono, 2003; Reuss and von Bohlen und Halbach, 2003). In addition, FGFs have been implicated as target-derived presynaptic organizers (Umemori et al., 2004). FGF7 is of particular interest, as it localizes specifically to inhibitory synapses in the hippocampal CA3 region, where it is secreted from the postsynaptic membrane and organizes presynaptic release properties (Terauchi et al., 2010). FGF receptors have been shown to directly interact with adenosine A2A receptors (Flajolet et al., 2008), which are important for GABA release (Cunha and Ribeiro, 2000) as well as for GABA uptake from the synaptic cleft (Cristóvão-Ferreira et al., 2009). In this way, FGFR and A2A receptors may act together to regulate GABAergic transmission in the hippocampus.

Studies with neuronal and astrocyte co-cultures and astrocyte-conditioned medium have shown that astrocyte-released factors are crucial for synapse formation and plasticity (Elmariah et al., 2005; Christopherson et al., 2005a; Hughes et al., 2010; Crawford et al., 2012). For instance, thrombospondins, oligomeric proteins of the extracellular matrix produced by astrocytes (Christopherson et al., 2005b; Eroglu et al., 2009) are involved in the formation of glutamatergic synapses and the pro-inflammatory cytokine TNFα, coming from glia, (Stellwagen and Malenka, 2006) plays a role in homeostatic plasticity of these synapses. In addition, a different and so far unidentified, protein is secreted by astrocytes, which has been found to increase the density of inhibitory synapses in cultured neurons (Elmariah et al., 2005; Hughes et al., 2010).

A special secreted factor is the inhibitory neurotransmitter GABA itself. It is well-established that synapse formation does not depend on neurotransmitter release (Verhage, 2000; Harms and Craig, 2005; Schubert et al., 2013). However, the development and maturation of inhibitory synapses are influenced by their neurotransmitter GABA (Li et al., 2005; Huang and Scheiffele, 2008; Huang, 2009; Lau and Murthy, 2012). It was shown that individual axons of parvalbumin-positive basket cells are sensitive to their own GABA release (Chattopadhyaya et al., 2007; Wu et al., 2012) and that the amount of GABA release per vesicle can be regulated by activity (Hartman et al., 2006; Lau and Murthy, 2012). Inhibitory boutons are less dynamic in axons in which GABA release is impaired (Wu et al., 2012) or when GABA receptors are blocked (Schuemann et al., 2013), strongly suggesting that GABA is used as an important activity sensor for regulating activity-dependent presynaptic changes at inhibitory synapses. Both GABA_A and GABA_B receptors have been implicated in mediating this regulation (Fu et al., 2012; Schuemann et al., 2013), but the precise molecular mechanisms remain unknown.