Synaptic Competition Sculpts the Development of GABAergic Axo-Dendritic but Not Perisomatic Synapses

The neurotransmitter GABA regulates many aspects of inhibitory synapse development. We tested the hypothesis that GABAA receptors (GABAARs) work together with the synaptic adhesion molecule neuroligin 2 (NL2) to regulate synapse formation in different subcellular compartments. We investigated mice (“γ2 knockdown mice”) with an engineered allele of the GABAAR γ2 subunit gene which produced a mosaic expression of synaptic GABAARs in neighboring neurons, causing a strong imbalance in synaptic inhibition. Deletion of the γ2 subunit did not abolish synapse formation or the targeting of NL2 to distinct types of perisomatic and axo-dendritic contacts. Thus synaptic localization of NL2 does not require synaptic GABAARs. However, loss of the γ2 subunit caused a selective decrease in the number of axo-dendritic synapses on cerebellar Purkinje cells and cortical pyramidal neurons, whereas perisomatic synapses were not significantly affected. Notably, γ2-positive cells had increased axo-dendritic innervation compared with both γ2-negative and wild-type counterparts. Moreover heterologous synapses on spines, that are found after total deletion of GABAARs from all Purkinje cells, were rare in cerebella of γ2 knockdown mice. These findings reveal a selective role of γ2 subunit-containing GABAARs in regulating synapse development in distinct subcellular compartments, and support the hypothesis that the refinement of axo-dendritic synapses is regulated by activity-dependent competition between neighboring neurons.


Introduction
During development neurotransmission regulates synapse formation and guides the selective assembly of circuitry. Activity mediates competition between converging inputs, through which more active synapses are stabilized and less active synapses are eliminated [1][2][3][4]. For instance, in the cerebellum, an imbalance in synaptic activity removes surplus climbing fibers innervating individual Purkinje cells (PCs) [5][6][7]. However, central synapses differ highly in their structural and molecular organization [8], and it is unknown if synapse competition is a general feature of CNS development.
In brain circuits, synapse heterogeneity is exemplified by the numerous types of GABAergic synapses that target distinct subcellular domains (somatic, dendritic or axonal) of principal neurons [9][10][11]. How these selective connections are generated during brain development and how their number is controlled is only partially understood [12,13]. GABA signaling itself coordinates inhibitory synapse development and activity-dependent regulation of synapse density in neuronal compartments [14,15]. In one study, reducing GABA synthesis in neocortical interneurons resulted in deficits in perisomatic synapse formation around pyramidal cells [16]. Conversely, loss of GABA A receptors (GABA A Rs) from cerebellar PCs in GABA A R a1 knockout mice affected axo-dendritic synapses made by stellate cells, but not perisomatic synapses established by basket cells [17]. The interpretation of this result was complicated, however, because PCs express transiently a3-GABA A Rs at a time when perisomatic synapses form [18]. Nevertheless, these findings imply that GABAergic activity has a selective effect on inhibitory synapse formation in separate types of neuron and/or different neuronal compartments.
To establish the importance of GABAergic signaling for synapse formation in different neuronal populations, synapse organization could be examined in genetically modified neurons that have reduced sensitivity to GABA. Ideally, to study synapse development in vivo neurotransmission should be silenced in only a subset of neurons, in order to directly compare the effects on synapse formation with neighbouring neurons that show intact GABA signaling. Moreover, mutations should not compromise animal survival during the postnatal period of intense synaptogenesis. For example, mice with total knockout of the GABA A R c2 subunit gene die in the first postnatal week [19], making it impossible to study how GABA A Rs influence later brain development. Here, we describe a new mouse line, GABA A R c2 knockdown (c2 KD), that has a strongly reduced expression of the c2 gene throughout the brain during development. Despite this, c2 KD mice survive until their third postnatal week, thus covering postnatal synaptogenesis. Remarkably, brains of c2 KD mice have a mosaic expression of the c2 subunit gene, resulting in a strong imbalance of GABAergic activity in neighbouring neurons. Thus this mouse line allowed us to study the role of GABA in postnatal brain development and synaptogenesis.

Generation of c2 KD mice
The targeting vector was designed such that, by gene targeting in embryonic stem cells, the native GABA A R c2 subunit gene (gabrg2, gene reference number ENSMUSG00000020436; www. ensembl.org/Mus_musculus/geneview) had an insertion of a modified c2 cDNA, flanked by loxP sites, in exon1 (Fig. 1A). The genomic DNA containing the mouse c2 subunit gene was obtained on a Bacterial Artificial Chromosome (BAC) by screening a mouse 129 BAC library (BAC Mouse ES release I, BAC4921, Genome Systems Inc, USA). As the basis for the targeting vector, an approximately 9 kb SpeI/SalI fragment containing exon 1 was subcloned into the BamHI site of pBluescript (Stratagene). Into this 9 Kb fragment, a SalI site was placed by in vitro mutagenesis into the 59UTR region of exon 1, 177 bp 59 (upstream) of the start-of-translation-ATG codon [20]. Into this SalI site, we placed a cassette containing an HA-epitope tagged c2 I77 subunit cDNA and SV40 polyadenylation sequence, followed by an frt-flanked neomycin resistance gene (Fig. 1A) [21]. The entire c2-neomycin cassette was flanked by loxP sites (Fig. 1A). The targeting vector was linearized with Not I, and electroporated into mouse R1 embryonic stem cells (strain 129/Sv). About 800 G418-resistant (Geneticin) ES cell colonies were screened for homologous targeting by Southern blot analysis. Three ES cell colonies had a homologous targeting event. One clone was expanded and the frt-flanked neomycin resistance cassette was removed through transient expression (electroporation) of enhanced FLP (eFLP) recombinase. After confirmation of removal of the neomycin resistance gene (by hybridizing genome DNA of Flptransfected colonies with a neomycin probe and looking for negative lanes), the targeted ES cells were microinjected into C57BL/6 blastocysts to generate chimeras (by Dr. Frank Zimmermann, University of Heidelberg, Germany). After generation of F1 mice, the genotyping was done by PCR across the 59loxP site.
The mice used in the present study were generated from heterozygous breeding pairs at our respective institutions. The animals were genotyped by PCR analyses (Fig. 1C) of genomic DNA tail biopsies using the primer pairs: Pr1 59-CTG CTT CTC TCA TTT GCC TTC CTG TGT  ACA TCT CTG-39  Pr2 59-GCT GAT GAT TTG ATG CCG GCT CCC CCC  ACC TGC CTC-39 All

In Situ Hybridization
In situ hybridization to mouse brain sections with [ 35 S]-labeled oligonucleotide probes was performed as described [22]. Nonperfused brains were removed and frozen on dry ice. Sections (14 mm) were cut on a cryostat, mounted onto poly-L-lysine-coated slides, and dried at room temperature. Sections were fixed in 4% formaldehyde, washed in phosphate-buffered saline (PBS), and dehydrated into 95% ethanol for storage at 4uC. Before hybridization, sections were removed from ethanol and allowed to air dry. Probes (0.3 pmol/ml) were 39 end labelled using terminal deoxynucleotidyl transferase (Roche Diagnostics, Germany) and a 100:3 molar ratio of a 35 SdATP (250 mCi/ml; Perkin Elmer, UK) to oligonucleotide. Labelled probe, dissolved in hybridization buffer, was applied to sections. Hybridization buffer contained 50% formamide/46 SSC/10% dextran sulphate (16 SSC: 0.15 M NaCl, 0.015 M Na-citrate). Hybridization was at 42uC overnight. Sections were washed with 16 SSC at room temperature for 5 min, 16SSC at 65uC for 40 min, 0.16SSC for 1 min at room temperature, 70% ethanol for 1 min at room temperature before 95% ethanol dehydration. Images were generated from four to six-week exposures to Kodak Biomax MR X-ray film (Eastman Kodak, Rochester, NY). To assess nonspecific labeling of the sections, each labeled oligonucleotide was hybridized to brain sections with a 100-fold excess of unlabeled oligonucleotide. Oligonucleotide sequences were:

Antibody characterization
The primary antibodies used in the present study are listed in Table 1. Polyclonal antibodies against the a1 and c2 subunits of GABA A Rs reveal on Western blots single bands of 50 and 43-48 kDa. This labelling is abolished by competition with the respective antigens [23,24]. Labelling specificity has also been verified in brain sections of knockout mice lacking the corresponding GABA A R subunit [19,25].
The rabbit anti-NL2 antiserum recognizes a single band of 105 kDa in Western blots of rat and mouse brain homogenates [26]. Immunolabelling is abolished by preabsorption with the peptide antigen, and no bands are visible in Western blots of NL2 knockout mouse brain homogenates. Moreover, the antiserum does not cross-react with NL1 or NL3 in transfected cells.
The monoclonal antibody against a-dystroglycan (a-DG; clone VIA4-1) recognizes a single band of approximately 156 kDa in Western blots of skeletal muscle lysate [27]. In neurons, mAbVIA4-1 gives a punctate labelling that colocalizes with GABA A Rs at postsynaptic specializations and is abolished by genetic deletion of dystroglycan (ref. [28] and unpublished observations).
Antibodies against carbonic anhydrase 8 (Car8), a selective marker of Purkinje cells, recognize a single band of 35 kDa in mouse cerebellar homogenates [18]. When applied to immunofluorescence on parasagittal brain sections, the antibodies strongly label Purkinje cells, and the immunoreactivity is abolished by preabsorption with the immunogen.
Mouse monoclonal anti-calbindin reacts specifically with calbindin (28 kDa) in immunoblots of brain homogenates of different species, and does not cross-react with calretinin or other known calcium-binding proteins [29]. No labelling is visible in brain sections obtained from calbindin D-28k knockout mice [30]. The rabbit anti-calretinin antiserum has been previously characterized by Western blotting of sturgeon brain extracts [31], in which the antiserum recognized a single protein band of the appropriate molecular weight. Mouse monoclonal anti-parvalbumin stains the 45Ca-binding spot of parvalbumin (12 kD and IEF 4.9) in a two-dimensional immunoblot [32]. This antibody has also been characterized extensively by immunohistochemistry [29,33].
The monoclonal GAD-6 antibody has been characterized via Western blot of rat brain homogenates and found to recognize selectively GAD65 but not GAD67 [34]. Epitope deletion studies have demonstrated that the GAD-6 antibody recognizes an epitope located between amino acids 475 and 571 of the Cterminus of GAD65 [35,36].
The polyclonal guinea pig anti-mGluR1a antibody recognizes a single band of 145 kDa in Western blots of cerebellar homogenates [37]. Specificity of the immunolabeling has also been verified on samples from mGluR1a null mice [38]. The mouse monoclonal antibody SMI 311 was kindly provided by Dr. Rita Garbelli (Besta Neurological Institute, Milan). This antibody recognizes the 50 and 200 kDa components of nonphosphorylated neurofilaments [39], and has been previously used to label cortical neurons and their dendritic profiles [40,41].

Immunofluorescence
For detection of postsynaptic molecules (NL2, GABA A Rs, a-DG), postnatal mice (P7-P20) were anesthetized with i.p. ketamine-xylazine 1:1 (0.1 ml/kg) and decapitated. The brains were excised and cut manually in either sagittal (cerebellum) or coronal (cerebral hemispheres) slabs, that were fixed by immersion in ice-cold formaldehyde (4% in 0.1 M phosphate buffer, PB, pH 7.4) for 20-30 min. Alternatively, mice were perfused with 4% formaldehyde in PB, and their brains were postfixed for 3 hours. Tissue slabs were cryoprotected in sucrose, sectioned with a cryostat, and the sections were collected on gelatin-coated slides. Following a blocking step in normal goat or donkey serum (3% in PBS with 0.5% Triton X-100), the sections were incubated overnight with combinations of two or three primary antibodies. The sections were then washed and incubated with the appropriate secondary antibodies, raised either in goat or in donkey, conjugated to one of the following fluorophores: Alexa 488 (Molecular Probes, Eugene, OR), Alexa 568, or the cyaninederived Cy3 and Cy5 (Jackson Immunoresearch, West Grove, PA). The sections were rinsed again and coverslipped with Dako fluorescence mounting medium (Dako Italia, Italy).

Confocal microscopy and data analysis
The sections were analyzed with a laser scanning confocal microscope (Zeiss LSM5 Pascal) using the multichannel acquisition mode to avoid fluorescence crosstalk. Quantitative analyses were performed on a minimum of three mice per group. Synaptic structures were analyzed on images acquired with a 6100 oilimmersion objective (1.4 numerical aperture) at a magnification of 8.1610 23 mm 2 /pixel, and the pinhole set at 1 Airy unit. The images were processed with the image-analysis program Imaris (release 4.2; Bitplane, Zurich, Switzerland). After segmentation, synapse density was quantified with NIH Image J software (http:// rsb.info.nih.gov/nih-image) as described in detail previously [42]. The number of perisomatic and axo-dendritic synapses was determined by counting manually synaptic clusters at the surface of PCs labeled for carbonic anhydrase 8 (Car8) or pyramidal neurons labeled with a monoclonal antibody against the panneuronal neurofilament marker SMI 311. Pyramidal neurons were identified by their typical morphology (triangular shaped cell body, apical dendrite and multiple basal dendrites). Heterologous contacts between GABAergic axon terminals and PC dendritic spines were quantified in confocal images after immunofluorescence labeling with selective markers [18]. Data are expressed as the number of contacts per surface of GABAergic boutons. Confocal imaging of the cerebellum of GABA A R a1 knockout mice was performed in sections used for our previous study [18].

Electrophysiology
Cerebellar PCs were recorded in acute cerebellar slices obtained from postnatal mice (P15-P18) prepared as previously described [43]. Whole-cell voltage-clamp recordings were performed at room temperature using the Multiclamp 700B/Digidata1440A system (Molecular Devices, Sunnyvale CA, USA) at pipette holding voltage of 270 mV. PCs were visually identified using an upright Olympus BX51WI microscope (Olympus, Japan) equipped with Nomarski optics. Patch pipettes, pulled from borosilicate glass capillaries (Hilgenberg, Malsfeld, Germany), showed 4 to 5 MV resistance when filled with high-chloride intracellular recording solution containing (in mM): 126 KCl, 4 NaCl, 1 MgSO 4 , 0.02 CaCl 2 , 0.1 BAPTA, 15 Glucose, 5 HEPES, 3 ATP and 0.1 GTP (pH 7.3 with KOH and osmolarity 290 mosmol l 21 ). Slices were continuously perfused with an extracellular solution containing (in mM): 125 NaCl, 25 NaHCO 3 , 25 glucose, 2.5 KCl, 1.25 NaH2PO 4 , 2 CaCl 2 , and 1 MgCl 2 (pH 7.4 when bubbled with 95% O 2 -5% CO 2 ). Kynurenic Acid (1 mM) was added to the extracellular solution to prevent glutamatergic events. Currents were sampled at 50 kHz, filtered at 3 kHz and stored in a computer hard drive. GABAergic IPSCs were analysed using the Clampfit 10.0 detection module (Molecular Devices, Sunnyvale CA, USA) that exploits a sliding template-based algorithm. The threshold for detection was set at 5 times the standard deviation of the baseline noise. For each recording, averaged traces are obtained from at least 50 synaptic events.
The rise time of GABA-elicited currents was estimated as the time needed for a 10 to 90% increase of the peak current response. The decaying phase of currents was fitted with exponential function in the form: where Ai are the fractions of the respective components.
In the case of analysis of normalized currents, SA i = 1. Deactivation time course was fitted with a sum of two exponentials (n = 2) for wild-type (WT) and WT-like cells, and with a monoexponential function (n = 1) for KD cells. Data are expressed as mean 6 SEM, and unpaired Student's t-test was used for data comparison.

Manipulation of the gabrg2 gene produces a strong knockdown of its expression
The c2 KD mouse line arose accidently. We had intended to generate a mouse line with a conditional gabrg2 allele, provisionally termed ''c2 switch'', that would serve as an elaboration of our method of making subsets of neurons selectively sensitive to zolpidem [44]. By homologous recombination in mouse embryonic stem cells, a c2 cDNA encoding the zolpidem-insensitive c2I77 version was placed into the 59UTR of the native c2F77 gene followed by a polyadenylation signal (Fig. 1A). In the resulting mice, the endogenous c2 promoter and regulatory regions were expected to drive the expression of the inserted c2I77 reading frame in the pattern of the native gene, causing all neurons to be insensitive to zolpidem and b-carbolines [21,44]. The c2I77 cDNA was also flanked with loxP sites to allow its removal by Cre recombinase in selected neuronal types, so restoring expression of the original c2 gene encoding zolpidem-sensitivity. A problem became apparent, however, when we produced homozygous mice for the modified c2 allele. Mice homozygous for the c2 subunit switch allele (c2 KD mice) were normal at birth, but soon developed several abnormalities, including reduced growth ( Fig. 2A,B), hunched posture, hyperactivity, impaired grasping and righting reflex, and died around P20. Mutant mice were not obtained at the expected Mendelian frequency, and the breeding of c2 KD mice was extremely time consuming due to drastically reduced reproduction rates.
We evaluated the expression of the c2 subunit using in situ hybridization on brain sections obtained from postnatal c2 KD mice, heterozygous and WT littermates. Normally, the c2 subunit gene is transcribed from before embryonic stage 14, and has a sustained strong expression throughout embryonic and postnatal development in most regions of the CNS [45]. c2 KD mice had a dramatically reduced level of c2 transcripts during brain development (Fig. 2C). However, there were no changes in the expression of the c1 and c3 subunit genes, suggesting that the c2 KD phenotype is not compensated by increased expression of either the GABA A R c1 or GABA A R c3 subunits (Fig. 2D). Similarly, there were no changes in mRNA hybridization signals for other major GABA A R subunits, such as a1, b2 and d (Fig. 2E). As the c2 subunit is required for the postsynaptic accumulation of GABA A Rs [46][47][48][49][50], the severe phenotype and reduced life span of c2 KD mice are likely due to a strong decrease of synaptic GABAergic transmission (see below). We did not examine why the cDNA insertion disrupted c2 gene expression; other than the original report mapping the c2 gene's transcriptional start sites and proximal promoter [20], nothing more has become known about how this gene is regulated. Nevertheless, although we could not use the mice for their intended purpose of allowing Creinducible zolpidem sensitivity in particular brain regions [44], the line did offer an excellent opportunity to study the role of GABA in postnatal brain development and synaptogenesis.

Normal brain assembly in c2 KD mice
The brains of c2 KD mice seemed correctly assembled. Nissl staining revealed that the brains of mutant and control littermates were morphologically similar, although c2 KD brains were slightly smaller, consistent with the reduced size of mutant mice (Fig. 3A). In particular, cortical layering appeared normal in c2 KD mice, as also supported by normal expression of neocortical lamination markers seen by in situ hybridization (not shown). In addition, no sign of heterotopia or neuronal cysts were visible in the cerebral and cerebellar cortices. Using antibodies against nonoverlapping subtypes of GABAergic interneurons, we found a general trend towards increased densities of interneurons in the hippocampus and sensorimotor cortex of c2 KD mice, which however reached significance only for the subgroup of parvalbumin-positive neurons (Fig. 3B). In the cerebellar cortex, double labeling for parvalbumin and calbindin revealed a normal density of PCs and molecular layer interneurons (Fig. 3B). These data indicate that reduced signaling through c2-containing GABA A Rs has no major effects on neuronal differentiation and interneuron survival. The increased density of parvalbumin-positive interneurons in cortical areas may be a compensatory mechanism to counteract the strong decrease of GABAergic inhibition.

Mosaic expression of the c2 subunit in neurons of c2 KD mice
We used immunofluorescence to investigate the distribution of the c2 subunit in brains of c2 KD mice. Labeling for the c2 subunit was punctate, suggesting synaptic localization (Fig. 4A-F). However, in all regions analyzed there was a noticeable reduction in the density of c2-positive puncta as compared with the WT situation (Fig. 4G). Remarkably, the c2 subunit appeared to have a mosaic expression, resulting in the presence of c2-positive and c2negative neurons co-existing in the same areas. This was particularly evident in the cerebellar cortex, where more than 60% of PCs were c2-negative, as determined by the absence of immunolabeled puncta outlining the cell body (Fig. 4F). Notably, the percentage of c2-negative PCs was constant from P7 (64%, n = 67 cells), when perisomatic synapses are initially assembled, to P20 (63%, n = 127 cells), suggesting that the majority of PCs do not express the c2 subunit during the entire period of postnatal development. Moreover, the reduced expression of the c2 subunit was paralleled by a similar decrease of puncta immunolabeled for the a1 subunit (Fig. 5C), indicating that loss of the c2 subunit caused a disruption of postsynaptic GABA A Rs.
Patch-clamp recordings were performed on PCs to ascertain how downregulation of the c2 subunit affects GABAergic synaptic transmission. We recorded spontaneous inhibitory postsynaptic currents (sIPSCs) from cerebellar acute slices of c2 KD and WT mice aged P15-P18. The majority of PCs exhibited a markedly reduced synaptic activity as compared with WT cells (Fig. 4H). Indeed, we observed a strong reduction in sIPSC amplitude from 136.3630.3 pA (n = 11) in WT mice to 10.961.2 pA (n = 17) in KD cells (Fig. 4I). Similarly, a strong reduction in the frequency of sIPSCs was observed in KD neurons (WT: 5.2561.00 Hz; KD: 0.1560.04 Hz; Fig. 4J). An additional analysis aimed at studying the kinetic properties of synaptic currents upon knock down of the c2 subunit revealed that sIPSCs recorded from KD cells exhibited slower deactivation and onset (Fig. 4K). The t mean of current deactivation increased from 8.260.9 ms in WT cells to 17.861.0 ms in KD cells (Fig. 4L) and, similarly, the 10-90% sIPSC rise time increased from 0.7360.06 ms in WT to 3.3060.41 ms in KD cells (Fig. 4M). Notably, two PCs recorded from c2 KD slices had sIPSC properties (amplitude: 56.5613.4 pA; frequency: 3.160.17 Hz; t mean : 8.90260.41 ms) that approached those of WT, suggesting that they contained c2positive clusters (Fig. 4N). Therefore, loss of the c2 subunit causes a dramatic decrease of inhibitory postsynaptic currents in PCs, and the few remaining currents have considerably slower kinetics compared with the WT situation.

Synaptic competition determines axo-dendritic innervation patterns in cerebellar PCs
The data so far indicate that mosaic expression of the c2 subunit causes a strong imbalance in GABAergic activity in neighboring neurons. This situation is ideal for studying the importance of GABAergic signaling for synapse development. We have shown previously that deletion of the GABA A R a1 subunit does not affect the postsynaptic localization of NL2 and a-DG in PCs [18]. Similarly, we found that NL2 and a-DG clustered at postsynaptic sites facing GAD65-positive boutons in PCs lacking the c2 subunit ( Fig. 5A-C). We estimated the density of NL2 and GAD65positive structures in the molecular layer of c2 KD mice. Compared to WT, NL2 puncta and GAD65 terminals were decreased respectively by 34% (mean 6 SEM puncta/1000 mm 2 : 79.962.9 in WT; 52.764 in KD; p = 0.0012; unpaired t-test; n = 3 mice per group) and 36% (mean 6 SEM puncta/1000 mm 2 : 68.561.6 in WT; 4461.4 in KD; p = 0.0001; unpaired t-test; n = 3 mice per group), indicating that knockdown of the c2 subunit results in a similar decrease of the density of pre-and postsynaptic structures. We then used antibodies against NL2 and/or a-DG to compare synapse organization in neighboring PCs that were either c2-positive or c2-negative. The results were compared with the situation in WT littermates. The density of perisomatic synapses was similar in c2-positive, c2-negative and WT PCs (Fig. 5D), substantiating the idea that the development of perisomatic synapses in PCs does not depend on GABAergic activity levels [17]. We then analyzed axo-dendritic synapses, using antibodies against Car8 to label selectively PC dendrites (Fig. 5C). Given that the development of axo-dendritic synapses is influenced by GABAergic signaling [17,18], our prediction was that the strong imbalance in the expression of the c2 subunit would cause a reduction of GABAergic innervation in c2-negative PCs, as well as increased connectivity in c2-positive PCs. Indeed, the density of DG-positive synapses was markedly reduced in c2-negative dendrites. Conversely, synapse density was significantly higher in c2-positive dendrites compared with both c2-negative and WT PCs (Fig. 5E). These data suggest that the number of axo-dendritic synapses that are established during development is strongly influenced by the level of GABAergic activity in the postsynaptic neurons.
Mosaic expression of the c2 subunit prevents the formation of heterologous synapses in c2 KD cerebella Mouse models that lack GABAergic transmission in all PCs (global a1 knockout mice [17,18]; PC-Da1 mice [51]; PC-Dc2 mice [49]) are characterized by reduced axo-dendritic innervation and the presence of heterologous synapses in which several PC spines are contacted by an unusually large GABAergic terminal. To see if in c2 KD mice GABAergic axons are similarly attracted by improper postsynaptic targets, we performed double labeling for GAD65 to identify GABAergic terminals and mGluR1a to label PC spines ( Fig. 6A; see also ref. [18]), and made a direct comparison with global a1 knockouts. In c2 KD mice GABAergic boutons were significantly smaller than in global a1 knockouts, being close in size to those of WT animals (Fig. 6B). Moreover, in both c2 KD and WT littermates GABAergic terminals were rarely found in close apposition with spines, whereas in a1 knockouts the number of heterologous contacts was significantly higher (Fig. 6C). These data suggest that heterologous synapses are an aberrant phenotype that occurs in situations in which all PCs are rendered silent to GABAergic transmission.
Target zone-specific differences in GABAergic synapse development are not restricted to the cerebellar cortex The results obtained in PCs indicate that competition mediated by synaptic GABA A Rs sculpts the development of axo-dendritic, but not perisomatic inhibitory synapses. To understand whether this is a general principle in GABAergic synapse development, we extended our analysis to other neuronal circuits. We initially asked whether the absence of the c2 subunit affects postsynaptic clustering of NL2 in different types of neuron and different types of GABAergic synapse. We frequently observed NL2-positive, c2negative clusters in the hippocampus and neocortex (Fig. 7A), as well as in several other brain regions (not shown). These punctate structures were present in the neuropil and also around cell bodies. Triple labeling for the c2 subunit, NL2 and GAD65 revealed that NL2 clustered at presumed GABAergic synapses lacking postsynaptic GABA A Rs (Fig. 8A). These data extend our previous observations in PCs [18] and indicate that GABA A Rs are not required for postsynaptic accumulation of NL2 at perisomatic and axo-dendritic GABAergic synapses. Interestingly, c2 KD mice had a lower density of NL2-positive clusters in synaptic layers in both CA1 (stratum radiatum) and sensorimotor cortex (layer V; Fig. 7B). This is reminiscent of the situation in the cerebellum (see above), and suggests that silencing of synaptic GABAergic transmission decreases the number of GABAergic synapses in dendritic domains (see below).
We then analyzed the organization of perisomatic and axodendritic synapses in pyramidal neurons of sensorimotor cortex layer V. This region was selected because c2-positive and c2negative pyramidal cells were clearly discernible using antibodies against the c2 subunit and NL2 (Fig. 8A). We found no significant differences in the density of NL2 clusters outlining the profile of c2-positive, c2-negative and WT pyramidal cells, suggesting that perisomatic innervation of cortical neurons is not regulated by synaptic GABA A Rs (Fig. 8B). We then quantified synapse density along the dendrites of pyramidal neurons labeled with an antibody against SMI 311 (Fig. 8C). In c2-negative dendrites there was a significant decrease in the density of NL2-positive clusters compared with both c2-positive and WT dendrites (Fig. 8D). Dendrites that were positive for the c2 subunit had a slightly elevated density of GABAergic contacts compared with the WT situation, however this difference was not significant (Fig. 8D). It should be noted, however, that labeling for SMI 311 did not fill the entire dendritic arborization of pyramidal neurons, therefore our analysis was restricted to the more proximal dendritic domains, where synapse number appears to be less influenced by activity levels [18]. These data indicate that in neocortical neurons, like in PCs, perisomatic and axo-dendritic synapses have different dependencies on GABAergic transmission.

Discussion
To investigate the importance of GABAergic signaling in synapse development, we have taken advantage of an engineered gabrg2 mouse allele that strongly reduced the expression of synaptic GABA A Rs. The c2 subunit is essential for postsynaptic aggregation of GABA A Rs [46], and its deletion dramatically affects inhibitory postsynaptic currents (refs. [44,[47][48][49] and Fig. 4H-J). Global ablation of the c2 subunit in mice causes perinatal lethality [19,46], thus preventing in vivo analyses of synapse differentiation. In contrast, the c2 KD mice reported here survive until their third postnatal week, when synaptogenesis has reached an advanced stage in most brain regions, making these mice useful to study GABA's developmental role during the peak period of synaptogenesis.

Knockdown of the c2 subunit strongly downregulates GABAergic synaptic currents
The strongly reduced GABAergic synaptic activity observed in c2 KD mice is consistent with previous studies in which the c2 subunit was deleted in neurons by Cre-mediated recombination [44,48,49]. In both cases, the residual sIPSCs had small peak amplitudes and slow decay time constants. This effect is most likely due to spillover of synaptically released GABA onto lowconductance a/b GABA A Rs [52]. Given the key role of the c2 subunit for the synaptic localization of GABA A Rs [46,53], it can be speculated that in c2 KD cells the activation of a/b-containing receptors dispersed in the perisynaptic and extrasynaptic membrane would be delayed by the time needed for GABA to diffuse outside the synaptic cleft. As a consequence, the macroscopic sIPSC onset and decay kinetics would be delayed. It cannot be excluded that loss of the c2 subunit also affects the gating properties of GABA A Rs [54], with a direct impact on synaptic current kinetics.
A few PCs in c2 KD slices had synaptic currents similar to those of WT, as predicted by the co-existence of c2-positive and c2negative cells revealed by immunofluorescence (Fig. 4N). These cells had somewhat lower amplitudes and frequencies compared to the average WT values, suggesting that even c2-positive PCs may Figure 5. Postsynaptic GABA A Rs determine axo-dendritic but not perisomatic innervations patterns in cerebellar PCs. A, Perisomatic synapses in PCs of a c2 KD mouse. a-DG (green) co-localizes precisely with the c2 subunit (red) in a c2-positive PC (c2+) and also outlines the profile of a c2-negative PC (c22). B, Upper panel: NL2 (red) co-localizes precisely with the c2 subunit (blue) in a c2-positive PC (c2+) and also outlines the profile of two c2-negative PCs (c22). Lower panel: triple labeling shows that NL2 clusters opposite GAD65-positive boutons (green) in both c2-positive and c2-negative PCs (arrows). C, Confocal images of PC dendritic profiles after triple labeling for GABA A R a1 (blue), a-DG (green) and Car8 (red). DG co-localizes with GABA A R clusters in PCs of WT mice as well as in c2-positive PCs of c2 KD mice (the superposition of the three fluorescent channels results in white clusters). The lower panel shows a c2-negative dendritic profile, where a-DG clusters are not associated with GABA A Rs (triangles). D, The density of perisomatic synapses is similar in c2-positive, c2-negative and WT PCs (c2+ vs c22, p = 0.6920; c2+ vs WT, p = 0.7312; c22 vs WT, p = 0.9230; unpaired t-test; n = 5 mice per group). E, The density of a-DG clusters is lower in c2-negative dendrites and higher in c2-positive dendrites compared with the WT situation (**, p = 0.0064, ***, p = 0.0003; unpaired t-test; n = 5 mice per group). Scale bars: A,B = 10 mm. C = 5 mm. doi:10.1371/journal.pone.0056311.g005 Figure 6. Absence of heterologous synapses in the cerebellum of c2 KD mice. A, Representative confocal images after double labeling for mGluR1a (a marker of PC spines) and GAD65. In global a1 knockout mice (upper panel), large GABAergic axon terminals make multiple contacts with PC spines (arrows). In c2 KD (middle panel) and WT mice (lower panel), GAD65-positive terminals have a smaller size and are less frequently found in close apposition with spines. B, Distribution of GAD65-positive boutons in the molecular layer of total a1 knockouts, c2 KD and WT mice based on size. The average area (6 SEM) is indicated. In total a1 knockout mice, GABAergic axon terminals are significantly larger than in the other groups (a1 KO vs WT, p = 0.0016; a1 KO vs c2 KD, p = 0.0008; c2 KD vs WT, p = 0.9348; unpaired t-test; n = 4 mice per group). C, The density of heterologous contacts between GAD65-positive terminals and PC spines is significantly higher in global a1 knockout mice compared with both c2 KD and WT mice (a1 KO vs WT, p = 0.03; a1 KO vs c2 KD, p = 0.03; c2 KD vs WT, p = 0.8836; unpaired t-test; n = 3 mice per group). Scale bar: 2 mm. doi:10.1371/journal.pone.0056311.g006 express lower-than-normal levels of the c2 subunit. Verification of this assumption would require data from a larger population of PCs, which has been hampered so far due to the limited availability of mutant mice. However, our data clearly demonstrate that expression of the c2 subunit was sufficient to rescue the deficit in the formation of axo-dendritic synapses in c2-positive PCs (see below).
Knockdown of the c2 subunit has no major effects on brain development Surprisingly, impaired GABA signaling in c2 KD mice did not interfere with normal brain assembly and cortical lamination. This might appear in contrast with many studies indicating that GABA is an important regulator of cell proliferation, neuroblast migration and neuronal differentiation [55][56][57][58][59][60][61][62]. However, the residual expression of a/b GABA A Rs in neurons lacking the c2 subunit (see above) leaves open the possibility that GABA might exert nonsynaptic effects. On the other hand, the absence of major neurodevelopmental defects in c2 KD mice is consistent with other investigations that have revealed a largely normal brain architecture in mice with null mutations in key genes of the GABA pathway [63][64][65][66]. It remains possible that c2 KD brains present subtle defects. For example, the increase in the population of parvalbumin-positive interneurons that we observed in the Figure 7. NL2 clusters at synapses lacking c2-GABA A Rs in cortical and hippocampal circuits. A, Confocal images of sensorimotor cortex (layer V) and hippocampal CA1 (stratum radiatum) after double labeling for NL2 (red) and GABA A R c2 (green). Note that in WT NL2 puncta co-localize extensively with c2-positive structures (arrows). In contrast, in c2 KD mice many puncta are labeled for NL2 but not for the c2 subunit (triangles). B, Reduced density of NL2-positive clusters in cortical and hippocampal neuropil of c2 KD mice as compared with WT (***, p,0.0001,** p = 0.0031; unpaired t-test; n = 3 mice per group). Scale bar: 3 mm. doi:10.1371/journal.pone.0056311.g007 Lower panel: triple labeling shows that NL2 clusters opposite GAD65-positive boutons (green) in both c2positive and c2-negative pyramidal cells (arrows). B, The density of perisomatic synapses is similar in c2-positive, c2-negative and WT pyramidal neurons (c22 vs WT, p = 0.0748; c2+ vs WT, p = 0.8187; c2+ vs c22, p = 0.0602 unpaired t-test; n = 3 mice per group). C, Confocal images of dendritic profiles after triple labeling for GABA A R c2 (green), NL2 (red) and SMI 311 (blue). NL2 co-localizes with GABA A R clusters in pyramidal neurons of WT mice as well as in c2-positive pyramidal neurons of c2 KD mice. The lower panel shows a c2-negative dendritic profile, where NL2 clusters are not associated with GABA A Rs (triangles). D, The density of NL2 clusters is significantly lower in c2-negative dendrites compared with the other two groups, whereas no difference was found between c2-positive and WT dendrites (c22 vs WT, p = 0.0008; c22 vs c2+, p = 0.04; c2+ vs WT, p = 0.83; unpaired t-test; n = 3 mice per group). Scale bars: A = 10 mm. C = 3 mm. doi:10.1371/journal.pone.0056311.g008 hippocampal CA1 and sensorimotor cortex (Fig. 3B) may influence the function and plasticity of cortical circuits [67][68][69][70].
c2-GABA A Rs are not essential for postsynaptic clustering of NL2 An important goal of our study was to determine how c2-GABA A Rs regulate the developmental assembly of GABAergic synapses. Ideally, to visualize synapses by immunohistochemistry, one should co-stain sections with antibodies directed against both pre-and postsynaptic markers (Fig. 5B, 8A). To distinguish between c2-positive and c2-negative neurons, however, we were forced to use a brief-fixation protocol that has been optimized for the detection of postsynaptic molecules [18,42]. Therefore, quantification of GABAergic synapses was mainly based on labeling for NL2 and/or a-DG as markers of the postsynaptic specialization.
The c2 subunit is a crucial organizer of GABAergic synapses and may stabilize postsynaptic receptor aggregates by directly interacting with other transmembrane proteins [50,53,71]. One hypothesis is that the developmental assembly of GABAergic synapses depends on an activity-dependent link between GA-BA A Rs and NL2, although it is unclear whether these proteins interact directly [15,72]. Previous studies have indicated that NL2 clusters at postsynaptic sites lacking GABA A Rs in PCs of mutant mice [18,51]. Similarly, we found here that NL2 clusters faced GAD65-labeled boutons in neurons lacking c2-containing GA-BA A Rs (Fig. 5B, 8A and data not shown), indicating that GABA A Rs are not essential for recruiting NL2 to the postsynaptic specialization. A recent study, however, found that clustering of NL2 at axo-axonic synapses of CA1 hippocampal neurons largely depends on a2-containing GABA A Rs [73], suggesting that NL2-GABA A R interactions may be synapse-specific. On the other hand, studies in NL2 knockout mice have indicated that NL2 contributes to stabilize postsynaptic GABA A Rs, at least in specific types of inhibitory synapses, and is required for normal GABAergic transmission [72,[74][75][76][77]. Collectively, the data indicate that neither NL2 nor GABA A Rs are essential for the formation of morphologically-recognizable inhibitory synapses; however, NL2 and GABA A Rs interact in a synapse-specific manner to organize postsynaptic specializations and determine synaptic properties [50,78]. One mechanism by which NL2 seems to regulate the maturation of GABAergic synapses is a direct interaction with gephyrin and collybistin, which promotes the formation of a postsynaptic scaffold onto which GABA A Rs are tethered [76].

Different regulation of perisomatic and axo-dendritic synapses
In all regions of the c2 KD brain analyzed, there was a decrease in the density of NL2-positive puncta in synaptic layers (Fig. 7), suggesting that silencing of GABAergic transmission perturbs axodendritic synapse development. This was confirmed by highresolution analyses on the dendrites of cerebellar PCs (Fig. 5C,E) and cortical pyramidal neurons (Fig. 8C,D), that revealed a decreased density of GABAergic postsynaptic structures in neurons lacking the c2 subunit. These data are consistent with previous studies on cultured neurons that demonstrated that c2subunit containing GABA A Rs are essential for normal GABAergic innervation [79,80]. However, we found that the number of perisomatic postsynapses was not affected by loss of synaptic GABA A Rs, providing strong support to the idea that perisomatic and axo-dendritic synapses have different dependencies on GABAergic activity levels. Li et al. (ref. [79]) reported a modest reduction of perisomatic innervation (24-29% compared with a 53% reduction of GABA A R cluster density) of cortical neurons after in utero electroporation of c2 shRNAs. The slight discrepancy between our results (no significant effect on perisomatic synapses) and those of Li et al. (modest reduction of perisomatic innervation) could be possibly explained by differences in the sensitivity of the immunolabeling procedure, or by differences in the quantification method, that was based on immunolabeling for the presynaptic vesicular GABA transporter (VGAT) in Li et al. [79] and on labeling for NL2 in the present investigation. In support of our observations, a recent study [73] has also shown that deletion of a2-GABA A Rs does not affect perisomatic innervation in CA1 pyramidal neurons, although compensation by the a1 subunit has to be taken into account in this case.
In contrast with the fixed situation of perisomatic synapses, the development of axo-dendritic synapses was sensitive to differences of GABAergic activity among neighboring cells. Our data clearly demonstrate that PCs lacking the c2 subunit are disadvantaged for axo-dendritic synapse formation or stabilization, whereas their neighbors expressing the c2 subunit increased the number of inhibitory postsynaptic sites compared with WT PCs. A similar situation has been reported in cultured neurons, where knockdown of the c2 subunit [79] or the palmitoyltransferase GODZ [80] caused a disruption of GABA A R clusters and selectively impaired GABAergic innervation.
Most likely, a similar activity-dependent process was responsible for the absence of heterologous synapses on spines in c2 KD mice. Heterologous synapses are abundant in situations in which there is a uniform suppression of GABAergic transmission in all PCs [17,18,51]. However, there is probably no advantage in maintaining heterologous synapses in the competitive environment of the c2 KD cerebellum. In other words, expression of the c2 subunit in at least some PCs resulted in increased axo-dendritic connectivity and was sufficient to avoid the formation and/or maintenance of heterologous synapses. These observations also indicate that axo-dendritic innervation does not depend on a hardwired process based on exclusive molecular interactions, but results from a mechanism of selection among potential synaptic partners. In normal conditions, GABA signaling serves to determine the density of synapses within the dendritic arborization of individual PCs. When the preferred connections are silenced, however, synapses can form with alternative partners, including dendritic spines.
The differences between perisomatic and axo-dendritic synapses could be related to their different roles in neuronal networks. Perisomatic synapses are known to control neuronal output very efficiently and are involved in neuronal synchronization [69,[81][82][83]. This requires that their number is strictly determined during development, and scarcely influenced by activity levels. Conversely, synapses on dendrites regulate glutamatergic inputs and calcium signals and exhibit a higher degree of activity-dependent plasticity [84][85][86]. Under this assumption, the level of plasticity at maturity is the main determinant of the relative importance of activity versus molecular cues in the assembly of inhibitory synapses during development.
The downstream pathways linking postsynaptic GABA A Rs to synapse maturation are presently unclear. Obviously, loss of GABA A Rs not only affects inhibitory neurotransmission, but also impairs molecular interactions within synaptic complexes. However another study has demonstrated that knockdown of GABA synthesis in cortical interneurons inhibits the ability of GABAergic axons to establish synapses with the appropriate targets, supporting the idea that GABAergic synapses are stabilized by an activity-dependent mechanism [16]. During cerebellar development, GABAergic synapses made by molecular layer interneurons are characterized by presynaptic miniature currents (preminis) that depend on the activation of presynaptic GABA A Rs and enhance neurotransmitter release [87]. The authors have proposed that depolarization due to preminis and autoreceptor activation produces a feedback loop that maintains a high release probability at recently formed synapses. Combined with our present findings, this suggests that axodendritic synapses could be stabilized by a mechanism involving the combined activation of GABA A Rs located at both pre-and postsynaptic sites. This hypothesis could be tested by analyzing how a selective ablation of GABA A Rs from molecular layer interneurons affects synapse development.
In conclusion, our findings reveal a remarkable selectivity in the way that synaptic activity determines the stoichiometry of synaptic connections in distinct subcellular compartments. Interestingly, the resilience of perisomatic synapses observed during development matches the situation in the aging brain, when synapses located on dendritic domains are significantly reduced while those located on the cell body are relatively unaffected [88]. Similarly, perisomatic synapses are spared, if not potentiated, in some forms of intractable epilepsies, and may contribute to the generation of pathological network activity [89,90]. Thus, understanding the different dependency of perisomatic and dendritic synapses on activity levels may be relevant for deciphering brain disorders that arise from altered GABAergic activity or changes in the excitatory/inhibitory balance.