Disinhibition of olfactory bulb granule cells accelerates odour discrimination in mice

Granule cells are the dominant cell type of the olfactory bulb inhibiting mitral and tufted cells via dendrodendritic synapses; yet the factors regulating the strength of their inhibitory output, and, therefore, their impact on odour discrimination, remain unknown. Here we show that GABAAR β3-subunits are distributed in a somatodendritic pattern, mostly sparing the large granule cell spines also known as gemmules. Granule cell-selective deletion of β3-subunits nearly abolishes spontaneous and muscimol-induced currents mediated by GABAA receptors in granule cells, yet recurrent inhibition of mitral cells is strongly enhanced. Mice with disinhibited granule cells require less time to discriminate both dissimilar as well as highly similar odourants, while discrimination learning remains unaffected. Hence, granule cells are controlled by an inhibitory drive that in turn tunes mitral cell inhibition. As a consequence, the olfactory bulb inhibitory network adjusts the speed of early sensory processing.

G ranule cells (GCs) are key players for the early processing of odour information in the olfactory bulb (OB). These interneurons are exclusively GABAergic 1 and account for the second layer of inhibition within the OB, dominating the OB circuitry [2][3][4][5] . GCs are involved in synchronization and establishment of slow temporal patterns of mitral cells (MCs) [6][7][8] , contrast enhancement for spatial representations of odours [8][9][10] and sharpening of activity onset in the OB network 11,12 . These are features of recurrent and lateral inhibition in the OB, mediated by the reciprocal synapse established between the GCs and MCs [13][14][15] . This synapse has been characterized at molecular, physiological and behavioural levels 11,[16][17][18][19] .
While the function of GCs with regard to MC inhibition is well studied on different levels, it remained less clear how GC activity itself is regulated and how it affects odour discrimination behaviour. Previous work described that GC activity is refined by inhibition mediated by monosynaptic inputs from Blanes cells, large stellate-shaped GABAergic interneurons present in the GC layer (GCL) that fire persistently on glomerular stimulation [20][21][22][23] . In situ hybridization studies have shown that GCs express all subunits necessary to generate both tonic and phasic inhibition 24,25 , although GCs express only the b3-subunit of the b-subunit subfamily 7,25,26 . GABA A receptors (GABA A Rs) containing the b3-subunit are major players in the coordination of network oscillations and are involved in synaptic inhibition 27,28 . Genetic deletion of the b3-subunit impairs odour discrimination behaviour and learning in mammals 7 . However, using mice globally lacking b3-subunits it is difficult to attribute the observed behavioural phenotype to a specific mechanism, cell type or brain region. Any cell-type-specific or even OB-specific conclusions from global b3-knockouts are limited because mitral and tufted cells in the OB express the b3-subunit as well 25,26 and b3 is expressed in other brain regions potentially involved in odour learning, discrimination and decision-making.
How does inhibition of GCs affect MC inhibition in the main OB? Where does this inhibition take place in the GC? How does GC inhibition affect odour discrimination speed and accuracy? We addressed these questions using a conditional mouse line with loxP sites flanking the exon 3 of the GABRB3 gene 29 to selectively abolish GABA A R-dependent inhibition in GCs by viral expression of Cre-recombinase in the GCL 19 . Our immunohistochemistry data show a clustered distribution of b3-subunit-containing GABA A Rs in the somata and dendritic stems of GCs representing GABAergic synapses. At the physiological level, the phasic inhibition mediated by these synapses strengthens MC inhibition. Finally, we use the go/no-go operant-conditioning paradigm to show that specific deletion of the b3-subunit in the GCL accelerates odour discrimination times (ODTs) for monomolecular odours and highly similar mixtures, but leaves discrimination learning unaffected.

Results
Clustered distribution of b3-containing GABA A Rs in GCs. We first examined the expression pattern of the b3-subunit in horizontal sections of the OB using a b3-selective antibody. Consistent with previous results 7, 30 , immunolabelling was found in all layers of the OB (Fig. 1a), but most prominently in the external plexiform layer (EPL, Fig. 1b). Viral labelling of GCs with membrane-bound green fluorescent protein (mGFP) reveals b3-subunit immunolabelling throughout the spatial domain occupied by the GCs (Fig. 1a). To visualize the expression and distribution of the b3-subunit within GCs, single GCs were labelled by sparsely expressing mGFP in the GCL ( Supplementary  Fig. 1A,B). To achieve this, rAAV-DIO-mGFP together with 1:1,000 diluted rAAV-Cre particles were stereotaxically delivered to the GCL of wild-type mice. After 3 weeks of expression, the OB tissue was stained with antibodies directed against the b3-subunit. Using the three-dimensonal (3D)-immunohistochemistry approach 31 , a strategy to selectively visualize proteins expressed within a structure of interest (illustrated in Supplementary  Fig. 1C-F), we found a clustered distribution of the b3-subunit in GCs (Fig. 1c-h). These clusters show a preferred localization at the soma and dendritic shaft (Fig. 1c), with only low abundance in the distal dendritic tree of GCs (Fig. 1c,h). Magnified 3D reconstructions of the regions indicated in Fig. 1c further illustrate this distribution pattern ( Fig. 1d-g). The two apical regions shown in panels d and e reveal dense b3-immunoreactive clusters; yet, none of these are positioned within the apical GC dendrite. Consistent with their localization in the EPL, these clusters represent inhibitory synaptic contacts of the MC lateral dendrites, known to abundantly express b3-subunits. In contrast, the two regions selected from a distal part of the primary dendrite (F) and the proximal dendrite at the junction to the soma (G) contain multiple b3-immunoreactive clusters (Fig. 1f,g). b3-positive immunoreactivity was mostly absent from GC gemmules ( Fig. 1c-h). Identical results were obtained in a total of five cells taken from three mice analysed with this approach. In summary, these results predict that inhibitory inputs at the soma and dendritic shaft produce a rather global inhibition of GCs, while a local, gemmule-specific inhibition appears unlikely. Furthermore, the clustered distribution is consistent with a synaptic localization of the b3-subunit-containing GABA A Rs in GCs.
To further support a synaptic localization of the b3-subunitcontaining GABA A Rs, we co-labelled the postsynaptic scaffolding protein gephyrin and the presynaptic vesicular GABA transporter VIAAT ( Fig. 2 and Supplementary Figs 2 and 3). Each of these stains exhibits a clustered distribution of the immunosignals, while the overlay of VIAAT and b3-signals reveals a lack of overlap yet close apposition of VIAAT clusters next to b3-clusters. In contrast, b3 and gephyrin signals widely overlap (Fig. 2a).
Pixel intensity-based colocalization analysis of the VIAAT and b3-immunosignals revealed no overlap (Fig. 2b, left; Pearson's correlation coefficient o0.2). The correlation found in the lowintensity range supports the notion that VIAAT and the b3-subunit are located in close proximity, with a few pixels of both signals overlapping because of the limited resolution of confocal microscopy. The same analysis of b3and gephyrin immunosignals revealed a high degree of colocalization of the clusters (Fig. 2b, right; Pearson's correlation coefficient 40.8), suggesting that virtually all b3-containing GABA A Rs are inserted in the inhibitory postsynaptic matrix delineated by gephyrin.
These observations were further confirmed by detailed 3D analyses ( Supplementary Fig. 3) of a small area of the cell body and dendritic shaft (Fig. 2c). The presynaptic vesicle cluster (visualized using the glow over palette) is located outside but in close proximity to the GC membrane (transparent grey). The postsynaptic b3 (cyan) and gephyrin (yellow) signals are located within the GC. This is further illustrated in the small panels showing the individual clusters, their pairwise and triple overlays. The reverse situation was found in the EPL. Here the b3 immunosignal was located outside of the labelled GC gemmule with an opposed cluster of inhibitory vesicles inside the GC gemmule (Fig. 2d). These results, found in three independent experiments, demonstrate that b3 immunosignals in the EPL predominantly reflect the postsynaptic component of dendrodendritic synapses in MCs, while in the GCL b3 immunosignals report postsynaptic sites in GCs.
Taken together, these data demonstrate the synaptic localization of b3-subunit-containing GABA A Rs in GCs, predicting that inhibition of GCs will be exclusively synaptic.
GC layer-specific deletion of the b3-subunit. To abolish b3-subunit expression in GCs, we took advantage of a mouse line harbouring loxP sites flanking the exon 3 of the GABRB3 gene 29 . Specific deletion of the b3-subunit in GCs was achieved by stereotaxic delivery of rAAV-Cre particles into the GCL. Immunohistochemistry against Cre-recombinase revealed that infected cells were almost exclusively present in the GCL sparing the MCL and other OB layers (Fig. 3a). Western blot analysis showed a significant decrease in the b3-subunit in the OB (Fig. 3b, n ¼ 3 mice in each group, two samples per mouse, t-test, P ¼ 0.016). To probe the efficiency of GC infection, we determined the colocalization of Cre-recombinase (all infected cells), 4,6-diamidino-2-phenylindole (DAPI; nuclei of all cell types) and NeuN (neurons; Fig. 3c). DAPI and NeuN were (a) Immunostaining using a b3-subunit-selective antibody (red hot colour palette) in horizontal OB slices infected with rAAV-mGFP (grey). The images represent single confocal frames. From left to right, GCL, MC layer (MCL), EPL and glomerular layer (GL). Scale bar, 200 mm. (b) Abundance of b3-subunit immunosignal in the EPL and GL relative to the GCL. Representative image used to quantify the grey-scale intensity values (here using the rainbow colour palette). See Methods for further details. Ratios relative to GCL in EPL (1.33 ± 0.04; n ¼ 16; ***Po0.001), GL (0.82±0.04; n ¼ 16; **Po0.01; Bonferroni's multiple comparison test). The mean ratios±s.e.m. Scale bar, 58 mm. (c) Distribution of b3-subunit clusters (red hot colour palette) within a single GC (grey). Images represent maximum-intensity projections of a confocal stack extending over 70 image frames, after frame-by-frame multiplication as illustrated in Supplementary Fig. 1C-F. Grey-scale levels adjusted for display. Scale bar, 0.25 mm. (d-g) 3D reconstructions of regions shown in c. 3D reconstruction procedure is illustrated in Supplementary Figs 2 and 3. The cartesian axes x (blue), y (red) and z (white) show the rotation of the reconstructed region from origin. The width of the image corresponds to B12.5 mm in d-f and 25 mm in g. Panels: mGFP, membrane-labelled GC; GABRB3, all immunosignals in the respective volume; mGFPxGABRB3, immunosignals residing inside the GC; merge, d,e show all GABRB3 immunosignals surrounding the GC; f,g show GABRB3 immunosignals residing within the GC. (h) Distribution of b3-subunit clusters (red hot colour palette) within the distal dendritic tree (grey). The images represent maximum-intensity projections of a confocal stack extending over 124 image frames. Grey-scale levels adjusted for display. Scale bar, 10 mm.
colocalized in 68 ± 2% (n ¼ 3 mice) of the cells, approximately reflecting the ratio of neurons and glia. DAPI and Cre were colocalized in 42 ± 2%, while NeuN and Cre-recombinase overlapped in 63±3% of the cells, suggesting that b3-deletion occurred in B63% of the GCs.
The deletion of the b3-subunit in GCs was further examined by b3-immunostaining of GABARB3 lox/lox and control tissue 3 weeks after injecting rAAV-Cre designed to yield a dense expression of Cre in a large fraction of GCs ( Supplementary  Fig. 4A). Immunostaining with b3-antibody (Fig. 4a) was quantified in the GCL and ECL, and an expression ratio was calculated (Fig. 4b). The ratio approach was chosen to normalize b3-expression in the GCL to that in the EPL because most of the b3 signals in the EPL originate from MCs and thereby serve as an internal control. The decrease in the expression ratio (Fig. 4b) confirms a reduction of the b3-subunit in the GCL. The extent of the ratio decrease is consistent with the overall efficiency of viral infection (see Fig. 3). To further assess changes in b3-expression on the level of individual GCs, rAAV-Cre (diluted 1:1,000) was co-expressed with rAAV-DIO-mGFP, thereby producing a b3-subunit knockout in individually resolvable GCs. We found a strongly reduced expression 21 days after virus injection ( Fig. 4c and Supplementary Fig. 4B). This time course is consistent with our previous experience in deleting ionotropic glutamate receptor subunits 19 . Furthermore, this approach provided additional evidence for the specificity of the antibody used against the b3-subunit and the 3D immunohistochemistry approach.
In summary, these results show that our strategy is suitable to achieve an efficient reduction in b3-subunit expression within the GCL not affecting neurons of other OB layers. Although the vast majority of cells in the GCL are GCs, neurons other than GCs 22,32 may get transduced with AAV-Cre particles. Hence, rAAV-Creinfected mice are referred to as GABRB3 DGCL .
b3-deletion in GCs reduces GABA A R-mediated mIPSCs. Deletion of the b3-subunit, a member of the essential b-subunits 25 , has been shown to abrogate synaptic GABA A R-mediated Control cells exhibited a low mIPSC frequency and, on deletion of the b3-subunit, mIPSCs occurred only rarely ( Fig. 5f; n ¼ 14 control cells and n ¼ 12 knockout cells; t-test; P ¼ 0.001). Furthermore, local application of 100-mM muscimol using a 50-ms puff (Fig. 5a, lower panels) revealed a strong reduction in the GABA A R current in knockout cells ( Fig. 5g; n ¼ 8 control cells and n ¼ 14 knockout cells; t-test; Po0.0001). This is consistent with the observed strong reduction in the number of b3-positive clusters (see above) and previously published results obtained in global b3-subunit knockout mice 7 . The remaining clusters had much smaller intensities and therefore contained a strongly reduced number of b3-subunits. The latter will cause a decrease in the mIPSC amplitudes towards the noise level of the recordings, giving rise to an apparent reduction in mIPSC frequency. Variability in receptor turnover, viral infection rate and incubation time after viral injection may explain that GABA A R-mediated currents were abolished in only 50% of the cells, while the remaining cells showed small currents.
In conclusion, these results provide evidence for a strong decrease in functional GABA A -Rs at inhibitory input synapses of GCs in GABRB3 DGCL mice.

Phasic inhibition of GCs controls strength of MC inhibition.
We next aimed at evaluating the physiological consequences of b3-subunit abrogation in GCs on recurrent inhibition of MCs. To accomplish this, whole-cell current-clamp recordings were established from MCs in 300-mm-thick acute horizontal OB sections at 34°C. Single-current injections with increasing amplitude (see Methods) were applied until MCs fired an action potential (AP). Immediately after AP firing, a strong hyperpolarization of the MC membrane potential was observed, reflecting recurrent inhibition of the MC (Fig. 6a). The amplitude of the inhibitory post-synaptic potential (IPSP) was increased in MCs recorded in GABRB3 DGCL mice compared with control ( Fig. 6a; t-test; P ¼ 0.017; n ¼ 11), while the IPSP decay time constant was not significantly changed ( Fig. 6a; t-test; P ¼ 0.07; n ¼ 11). Thus, recurrent inhibition elicited by a single AP in MCs was stronger in the absence of GABA A R in the majority of surrounding GCs.
MCs have been shown to respond to odour stimuli with brief bursts of APs 12 . To test frequency dependence of recurrent inhibition, 20 consecutive APs were elicited by applying suprathreshold currents (control cells: 654.5 ± 74.3 pA, n ¼ 11, GABRB3 DGCL : 809.1±63.9 pA, n ¼ 11; t-test; P ¼ 0.13). The IPSP amplitude at the end of the train was greatly increased in MCs of GABRB3 DGCL mice ( Fig. 6b; t-test; P ¼ 0.001; n ¼ 11). The IPSP decay time constant was not significantly changed ( Fig. 6b; t-test; P ¼ 0.12; n ¼ 11). When compared with single AP firing in controls, the decay time constant was significantly increased (t-test; P ¼ 0.005; n ¼ 11), while in MCs of GABRB3 DGCL mice it remained unchanged (t-test; P ¼ 0.87; n ¼ 11). Furthermore, we analysed the IPSP amplitudes during the train of APs (Fig. 6c). We found only a small increase in MCs of control mice while the IPSP amplitude strongly increased in MCs of GABRB3 DGCL mice (two-way analysis of variance (ANOVA); P ¼ 0.0017; n ¼ 11). Moreover, using the same stimulation thresholds we found that control MCs showed fewer failures in AP generation compared with MCs from GABRB3 DGCL mice ( Fig. 6d; Cumulative-Gaussian F-test, Po0.0001; n ¼ 10 control cells and n ¼ 11 GABRB3 DGCL cells). These observations strongly suggest that GC inhibition is an important mechanism regulating MC firing properties and tuning the amount of inhibition that MCs can recruit on stimulation. b3-deletion in the GCL accelerates odour discrimination. The observation that recurrent inhibition of MCs is strongly increased in GABRB3 DGCL mice and that increased inhibition accelerates odour discrimination 19 leads us to hypothesize that mice lacking the b3-subunit in GCs would require less time to discriminate odours. To test this, we used a go/no-go operant conditioning odour discrimination paradigm 34 illustrated in Fig. 7a. Briefly, a trial is self-initiated by the mouse inserting its head into the odour port and thereby breaking the infrared light beam. On odour application the mouse retracts its head when an unrewarded (S À ) odour is applied or stays in the odour port and starts licking until the water reward is applied at the end of the trial. The corresponding state of the infrared light beam is shown in the lower panel for three individual example trials. After sampling hundreds of trials, a mean fraction of the beam being disrupted at any given time point can be calculated and a bootstrapping statistical procedure be used to determine the time point at which the S þ and S À curves differ significantly, defined as the ODT (Fig. 7b, see Methods for a more detailed description). Mice require B1,000 trials to achieve 490% accurate discriminations; we refer to this learning process as odour discrimination learning.
ODT is determined only after a stable performance at 490% correct identifications has been reached.
In the first experiment we applied our previously established paradigm 19 . The OBs of mice were stereotaxically injected with rAAV-eGFP (sham group, n ¼ 8) or with rAAV-Cre (GABRB3 DGCL group, n ¼ 8). First, mice were trained to discriminate between the monomolecular odourants amyl acetate (AA) and ethyl butyrate (EB). Both sets of mice readily acquired the discrimination task and reached a discrimination accuracy exceeding 90% within B1,000 trials (Fig. 7c). Second, a 60/40% binary mixture of these odourants was used to generate highly similar stimuli. Discrimination accuracy initially dropped to a chance level but increased to 490% after B600 trials (Fig. 7c). The learning curves of control and GABRB3 DGCL knockout mice were indistinguishable for both monomolecular odourant discrimination and binary mixture discrimination (Fig. 7c,d; two-way ANOVA; P ¼ 0.29 for simple odours and P ¼ 0.77 for binary mixtures; n ¼ 8). ODTs were measured when performance exceeded 90% (Fig. 7d). We found that monomolecular odourant discrimination and mixture discrimination were both accelerated in mice lacking the b3-subunit in GCs ( Fig. 7e; t-test; monomolecular odourants P ¼ 0.016; binary mixtures P ¼ 0.018; n ¼ 8). Thus, in contrast to previous molecular perturbations affecting MC inhibition, abrogating GC inhibition resulted in accelerated discrimination independent of the similarity of odour representations in the OB.
In the second experiment we aimed at further controlling for unspecific effects of surgery, Cre-recombinase expression and other potentially confounding variability factors. To achieve this, we applied a longitudinal design consisting of sessions including discrimination in non-injected wild-type (n ¼ 9) and GABRB3 lox/lox (n ¼ 9) mice followed by viral perturbation and subsequent sessions comparing control and GABRB3 DGCL mice. This allowed us to follow behavioural parameters longitudinally in every single mouse. For task habituation, mice were first trained to discriminate the simple odours cineol (C) and eugenol (E) until accuracy exceeded 80%. Subsequently, mice rapidly learned to discriminate monomolecular odourants AA/EB and highly similar binary mixtures thereof with high accuracies (Fig. 8a). The learning curves of the two groups ( Fig. 8a; two-way ANOVA; simple odours P ¼ 0.4195 and binary mixtures P ¼ 0.4709; n ¼ 9) and steady-state discrimination accuracy were indistinguishable ( Fig. 8b; two-way ANOVA; simple odours P ¼ 0.42 and binary mixtures P ¼ 0.47; n ¼ 9). ODTs determined from the last 300 trials were indistinguishable between the groups, measuring B285 and 317 ms for monomolecular  Supplementary Fig. 4B). Only a small number of faint b3-immunosignals (marked with arrows) are detectable 3 weeks after viral perturbation (representative for three GABRB3 lox/lox mice co-injected with rAAV-Cre (1:1,000 diluted in 1 Â PBS) and rAAV-DIO-mGFP (mixed in a 1:1 ratio with diluted rAAV-Cre). Cartesian axes x(blue), y (red) and z (white), image size 235 mm Â 235 mm, pixel size ¼ 0.229 mm, 90 image frames covering a depth of 13.3 mm.
odourants and binary mixtures, respectively (Fig. 8c). The increased discrimination time for binary mixtures is consistent with our previous results 19,34 .
Next, all mice were stereotaxically injected with rAAV-cre (Fig. 8, arrowhead and vertical dashed line), and after recovery from surgery the same cohort of mice performed a second session of behaviour using the same test odour pairs. Discrimination of AA versus EB started at high discrimination accuracy consistent with the mice remembering the odour pair from previous training sessions and reached saturated performance already after less than 200 trials (Fig. 8a,b). Binary mixture discrimination started at 65% accuracy and reached saturating accuracy after 600 trials, suggesting that mice also remembered this odour discrimination,    ARTICLE although not as efficiently as the monomolecular odourant discrimination (Fig. 8b). The learning curves and discrimination accuracies were indistinguishable between the two groups ( Fig. 8a; two-way ANOVA; P ¼ 0.95 for simple odours and P ¼ 0.08 for binary mixtures; n ¼ 9). The ODTs for monomolecular odourants and binary mixtures were shorter in this compared with the pre-injection session, possibly because of continued training (two-way ANOVA; P ¼ 0.0015 for simple odours and P ¼ 0.0024 for binary mixtures; n ¼ 9). However, GABRB3 DGCL mice revealed significantly faster ODTs of both monomolecular odourants as well as binary mixtures ( Fig. 8c; two-way ANOVA; simple odours P ¼ 0.046 and binary mixtures P ¼ 0.008; n ¼ 9), confirming the results of the first discrimination experiment. The motivation of mice to perform the task was controlled over the entire experiment by measuring the inter-trail interval (ITI). No changes were observed either before or after the surgery ( Fig. 8d; two-way ANOVA; simple odours P ¼ 0.083 and binary mixtures P ¼ 0.11; n ¼ 9) and neither between groups (two-way ANOVA; simple odours P ¼ 0.183 and binary mixtures P ¼ 0.078; n ¼ 9). Our conclusions thus far rest on using two monomolecular odourants and their binary 60/40 mixtures; however, accelerated odour discrimination may be confounded by ceiling effects and may not apply to stimuli of different similarities. To test this, we first trained a new batch of mice (n ¼ 8 controls, n ¼ 8 GABRB3 DGCL ) to discriminate increasing dilutions of the binary mixtures (Fig. 9). After pretraining with cineol versus eugenol, mice rapidly learned to discriminate monomolecular odourants AA and EB. After learning binary mixture discrimination at our standard odour concentration of 1%, they equally well learned to discriminate 0.1% of the mixture. However, starting with a dilution of 0.001%, the learning curve was less steep and required longer to reach our target criterion of 490% discrimination accuracy. At 10 À 5 %, mice performed at only 75% accuracy even after 900 trials, suggesting that even under non-saturated behavioural performance the difference in discrimination time persisted. The learning curves and discrimination accuracies did not differ between control and GABRB3 DGCL mice while ODTs were significantly reduced in all situations (Fig. 9). These results indicate that increased inhibition accelerates odour discrimination independent of the similarity and difficulty of the discrimination task. To further corroborate this result, we tested discrimination of carvone stereoisomers and cineol versus eugenol, both monomolecular odourants and their binary mixtures ( Supplementary Fig. 5). Learning the discrimination of carvone and cineol versus eugenol mixtures took more time while 90% discrimination accuracy could be reached. However, ODTs were reduced in GABRB3 DGCL mice in all situations, further supporting the prediction that the accelerated odour discrimination is a general feature of increased MC inhibition.
In conclusion, increased MC inhibition, caused by disinhibition of GCs, accelerates odour discrimination independent of odour similarity.

Discussion
In this study we investigated the cellular and behavioural role of GC inhibition. In contrast to previous studies examining a global b3-subunit knockout 7 , we deleted the GABRB3 gene 29 selectively in the GCL. This 'genetic' disinhibition of GCs increased the inhibition of MCs and accelerated odour discrimination. ODT was decreased independently of the similarity of odourant representations in the OB. These observations suggest that GCs are inhibited by a persistent GABAergic input and/or an odourevoked feedforward signal. This inhibitory drive may serve as a mechanism to regulate the strength of MC inhibition and thereby adjust the time required to discriminate odours.
Our study provides the second example of a targeted molecular GC perturbation, resulting in a gain-of-function phenotype in odour discrimination. Similar to our previous study deleting GluA2 in the GCL 19 , we found increased MC inhibition and accelerated discrimination after deleting b3-subunits. Hence, with two different molecular perturbations converging on the same cellular and behavioural outcome, we suggest that increased MC inhibition, in general, decreases ODT while reduced MC inhibition increases ODT 19 . Therefore, the relative strength of GC inhibition fine-tunes MC inhibition and ODT. The level of inhibition could affect gamma cycle frequency and with it the speed of transmitting a certain number of synchronous MC APs towards higher brain areas.
Why would the OB generate changes in ODT rather than cortical areas rendering discrimination decisions? We suggest that a certain level of MC inhibition is required to produce accurate odour discrimination decisions. Strong MC inhibition will reach this level of inhibition faster than weak MC inhibition (see Fig. 8b, Abraham et al. 19 ; this Figure shows that the time course of inhibition matches the differences in ODT observed for molecular perturbations). Therefore, the time course of inhibition build-up in MCs may define a substantial fraction of the ODT. Downstream brain areas directly concerned with decision-making may contribute to ODT too; however, these contributions are predicted to work on a shorter timescale.
Our results show a prominent link between MC inhibition and ODT; yet, odour discrimination learning appears to be unaffected by the strength of MC inhibition. This lack of an effect could reflect limitations in the odourant space sampled with our behavioural assay and the difficulty of the discriminations assessed. However, we think that this is unlikely because we used a wide range of odourants, binary mixtures and dilutions. Furthermore, forebrain-specific deletion of GluA2 accelerated discrimination learning while deletion in the OB did not 35 . These results suggest that discrimination learning may primarily occur in cortical olfactory areas; yet, effects of OB inhibition may become evident when pushing the difficulty of discrimination tasks further.
NATURE COMMUNICATIONS | DOI: 10.1038/ncomms9950 ARTICLE processing time in the OB network to achieve pattern separation. In contrast to the expectation that discrimination of dissimilar stimuli may work without inhibition of MCs, we show that this is not the case. Consistently, a tendency for accelerated ODT was also found after deletion of GluA2 receptors in GCs (see Fig. 5d in ref. 19). Our colocalization analysis of VIAAT, gephyrin and b3-subunit clusters in GCs demonstrates that GABA A Rs in GCs are located synaptically and hence mediate phasic rather than tonic inhibition of GCs. This conclusion is inconsistent with the proposal that tonic, extrasynaptic inhibition dominates over phasic, synaptic inhibition in GCs 36 . Their conclusion was based on using furosemide as an antagonist of GABA A Rs containing the b3and a4or a6-subunits (the latter not expressed in GCs 25 ), independent of the presence of g2or d-subunits 37,38 . Because we have demonstrated a synaptic distribution of the b3-subunit in GCs, extrasynaptic receptors would have to be homomeric a4-channels. However, such GABA channels have not yet been described to exist in vivo 39,40 . The findings of Labarrera et al. 36 may be explained with the inhibition of KCC2 (ref. 41) by furosemide, resulting in decreased GABAergic inhibition due to accumuated intracellular chloride and increased AP firing (see Fig. 4 (ref. 36)). Alternatively, our imaging approach may lack the resolution and sensitivity to detect single GABA A R channels distributed at a low density across the plasma membrane of GCs. Hence, a tonic component in addition to phasic inhibition may exist 25,40 ; however, its dominant impact 36 seems highly unlikely according to our results.
Several cell types known to exist in the OB 42 may provide inhibitory input to GCs, for example short-axon cells or Blanes cells. The latter were shown by Pressler and Strowbrige 22 to fire persistently on glomerular stimulation and to induce a GABAergic synaptic current in GCs. Hence, MC APs may spread via collateral axons and activate Blanes cells that in turn inhibit GCs. Depending on the connectivity of the Blanes cells, a single such cell could potentially inhibit a large number of GCs and thereby decrease MC inhibition in a manner precisely tuned to the onset of the odour stimulus. Such a circuit predicts that a MC may even disinhibit itself, assuming appropriate wiring, a function that could contribute to contrast enhancement. In conclusion, our physiological and behavioural findings are consistent with an inhibitory drive on GCs as suggested by Pressler and Strowbridge 22 . Yet, we cannot exclude that GCs are also inhibited by other cells of the OB or by long-range GABAergic neurons projecting from areas outside the OB, similar to ChAT-and GAD-positive neurons in the horizontal limb of the diagonal band 43 .
Our electrophysiology data suggest that sparse connectivity between GCs and MCs 44,45 might be a consequence of GC inhibition. This is consistent with the observation that a single MC AP elicits a similar degree of inhibition compared with 20 consecutive APs in the unperturbed OB (see Fig. 6a,b, compare control IPSP amplitudes). In contrast, when inhibition of GCs is nearly abolished after b3-subunit deletion, MC inhibition increases strongly with the number of APs, indicating that the mechanism of inhibiting GCs determines the frequencydependent strength of inhibition of MCs (Fig. 6c).
GC excitatory inputs originating from MC dendrites may coincide with inhibitory inputs, a mechanism proposed to suppress multiple spiking of GCs 46 . We observed that in wild type, the decay time of the IPSP after 20 consecutive APs increases about four times when compared with that elicited by a single AP, while after b3-subunit deletion this relationship is lost (Fig. 6a,b, compare decay time constants). This observation supports the hypothesis that GC inhibition might be an important mechanism to define the kinetics of MC inhibition. Strong hyperpolarization of the GC may trigger a long-lasting T-type Ca 2 þ current 11 that may prolong GABA release from GC gemmules.
In this study we established a longitudinal behavioural design that allows comparisons of individual mice before and after deletion of the target gene. We found that the overall variability of the derived parameters is smaller and that a larger number of controls can be achieved in one experiment. By using a second set of wild-type mice, the line harbouring the floxed allele can be directly compared with wild type and the effect of the injected virus construct on wild type can be assessed.
Acute targeted genetic perturbation. Male, 21-day-old mice were stereotaxically injected using a BENCHMARK stereotax. The rAVV chimeric vectors (1:1 ratio of AVV1 and AAV2 capsid proteins) were injected in the GC layer of the OB, as previously described 19 . Briefly, the rAAV-Cre vector contained the coding sequence for Cre recombinase, N terminally fused to the haemagglutinin-tag and the SV40 nuclear localization signal in a rAAV plasmid backbone (pAM) containing the 1.1-kb cytomegalovirus enhancer/chicken b-actin promoter, the woodchuck post-transcriptional regulatory element and the bovine growth hormone polyA. For immunohistochemistry, rAAV-Cre (1:1,000 diluted in 1 Â PBS) was co-injected with rAAV-DIO-mGFP 47 (1:1). This strategy was used to obtain single GCs labelled with mGFP. The rAAV-DIO-mGFP construct contains the mGFP sequence in an inverted orientation (3 0 to 5 0 direction) with loxP and lox2272 sites flanking it on each side, intercalated in opposing orientations (loxP4lox22724 3 0 mGFP 5 0 oloxPolox2272). Cre-recombinase expression then yields an inversion of the mGFP sequence and expression of mGFP. For voltageclamp experiments of mIPSC recordings of GCs, rAAV-Cre was co-injected with rAAV-stopflox-Clomeleon 48 in GABRB3 flox/flox mice to identify Cre-infected cells.
Immunohistochemistry. To specifically analyse the expression and distribution of the b3-subunit, 3D immunohistochemistry was performed. In short, 21-day-old wild-type mice were stereotaxically injected with rAAV-Cre diluted 1:1,000 in 1 Â PBS together with rAAV-DIO-mGFP mixed in a 1:1 ratio (for sparse expression) to identify single GCs. After 3-4 week expression, the animals were transcardially perfused using 4% paraformaldehyde (PFA). The brains were removed and free-floating 50-mm-thick horizontal slices of the OB were made using a vibratome. Antigen retrieval was made by incubating OB slices in 25-mM trisodium citrate (pH ¼ 7.4) in a warm water bath at 60°C during 30 min. This temperature did not change fluorescence properties of GFP. Afterwards, the slices were incubated for 45 min at room temperature in vehicle buffer (5% normal goat serum, 0.5% bovine serum albumin fraction V and 0.3% Triton X-100, in 1 Â PBS, pH ¼ 7.4) with 0.1% cold fish gelatin (blocking solution). The antibody against the b3-subunit (Novus Biologicals, cat. #: NB300-199) was carried in vehicle (1:1,000 dilution) and incubated overnight at 4°C. After the third wash in vehicle (10 min each wash step), the alexa dye-conjugated secondary antibody (1:1,000 in vehicle; Life Technologies cat. #: A-21244) was incubated for 90 min at room temperature. The slices were washed three times in 1 Â PBS, pH ¼ 7.4, and mounted on coverslips using self-made moviol. All incubation steps were performed under gentle incubation. For the triple labelling, the antibodies against VIAAT (SYSY, cat. # 131004) and gephyrin (kind gift from Professor Jochen Kuhse) were used.
Quantification of b3-expression in the OB. Single-frame confocal images (1,476.2 mm Â 1,476.2 mm with a pixel size of 1.44 mm) were used without additional processing. From each layer, squares with 100 Â 100 pixels were selected and the average grey value in the GCL, EPL and GL was quantified (n ¼ 4, 4 regions per experiment). For each experiment the average grey value of the GCL was used to normalize the pixel intensity of the EPL and GL.
Imaging. Images were acquired in a confocal microscope (Leica SP5, Leica, Germany) using a Â 63 glycerol immersion objective (numeric aperture ¼ 1.3). For high-resolution image acquisition, 150-nm z steps were performed, with 1,024 Â 1,024 or 2,048 Â 2,048 pixels in the xy plane, with a field of view of 234.32 mm Â 234.32 mm. 3D reconstructions were performed using the mGFP signal to do anatomical reconstructions of entire cells, cell bodies and dendrites. To analyse the cellular localization of the b3-subunit, VIAAT and gephyrin, the antibody signal was excised by image multiplication with the mGFP channel using the 3D immunohistochemistry approach 31 .
Quantitative western blot analysis. Western blots were made from whole-OB protein extracts isolated in ice-cold artificial cerebrospinal fluid (see electrophysiology section). In total, OBs from six mice were used (three control and three rAAV-Cre injected) and two samples per mouse were analysed. The nuclear fraction was discarded by centrifugation at 800g, 20 min at 4°C. The supernatants were loaded in self-made 5-16% gradient gels for electrophoresis 49 . Nitrocelulose membranes were used to blot the proteins. b3-subunit (Novus Biologicals, cat. #: NB300-199) was detected by using a goat anti-rabbit horseradish peroxidaseconjugated antibody (Sigma-Aldrich, cat. # P7899). For actin detection, a primary antibody horseradish peroxidase conjugated was used (Sigma-Aldrich, cat. # A3854). The peroxidase reaction was made using enhanced chemoluminescence (Thermo Scientific Pierce ECL Plus substrate, cat. # 32132). X-ray films were scanned in a conventional scanner (Canon). The grey images were further processed in ImageJ. Square selections of regions of interest were made, keeping the area constant across all measurements. Pixel intensities were used to calculate the ratios of actin and b3-subunit between wild-type and GABRB3 DGCL subjects.
Quantification of Cre expression efficiency. To access the total number of nuclei (DAPI, Sigma-Aldrich cat. # D9542-1MG), total number of neural cells (NeuN, Millipore, cat. # MAB377) and total number of Cre-positive cells (Cre, Novagen, cat. # 69050), 4% PFA-fixed 50-mm-thick horizontal OB slices were used, that were obtained from mice sterotaxically injected with rAAV-Cre used in behaviour (n ¼ 3). Stacks of images (200 Â 200 mm in the xy plane and 0.15 mm in z plane) of randomly selected regions in the GCL were acquired in a confocal microscope (Leica SP5). From each mouse, three to five slices were used and in each slice eight to ten regions were imaged. To count nuclei, maximal intensity projections of stacks of 10 images were used, using the ImageJ cell counter built-in function.
The amount of current injected in MCs to fire a single AP was defined as threshold current. Cells that did not fire 20 consecutive APs at the maximal current injection possible using our amplifier were excluded from statistical analysis.
Behavioural assessment. Behaviour experiments were conducted using a go/nogo operant conditioning scheme as previously published 19,34 . Male mice were kept separated in macrolon cages type II in a reverse light cycle (12-12 h), in a temperature-and humidity-controlled environment. Behaviour experiments were conducted in a dark room, corresponding to the night period. Behavioural pretraining, during which mice explored the set-up and learned that water can be obtained by licking at a metal tube placed in the odour port, started 3-4 weeks after the surgery and 2-3 days before the animals were kept under water restriction (by periods no longer than 12 h). The weight of the animals was strictly monitored and kept 485% of the initial body weight. The pre-training took 3-5 days and the behavioural training took 4-8 weeks. All animals tested were able to successfully learn the behaviour paradigm. Eight-channel semi-automated olfactometers 50 (Knosys, Washington) and data acquisition were controlled through customprogrammed software, written in Igor Pro 6 (Wavemetrics Inc.). All odours were diluted to 1% in mineral oil and further diluted 1:20 in air in the olfactometers. Test odours (AA, EB, cineol, eugenol and binary mixtures thereof) were used to determine the reaction times. Odours were presented in a randomized manner, however, disabling more than five consecutive trials of the same stimulus. Bias towards any of the odours presented was avoided by counterbalancing two groups of mice.
The ODT was determined as described previously 34 and is illustrated in Fig. 7a,b. A mouse self-initiates a trial by poking its head into the sampling port and breaking an infrared light beam. Mice need to wait 0.5 s (brown bar) at the sampling port for the odour onset to ensure that the odourant is present at the final valve at the appropriate concentration. Then, the odour is applied for 2 s. Mice stereotypically retracted their heads from the sampling port when unrewarded stimuli were presented (S À ); however, they kept their heads inside the sampling port when rewarded stimuli were presented (S þ ) and immediately started licking at the reward tube to receive the water reward at the end of the trial (2.5 s total trial duration). Mice needed to lick during at least three 500-ms time bins (out of four in the 2-s trial time) to score a correct response. The sampling pattern (time when the head remained inside the sampling port) of S À stimuli was used to determine the ODT. Alternatively, the onset of licking for S þ trials could be used to determine ODT. After reaching 490% discrimination accuracy, an average curve was calculated from each 300S À and S þ trials, and bootstrapping statisitics 19,34 was used to determine the first time point at which both curves differed significantly (P ¼ 0.05; first detect reaction time). This time was defined as the ODT. Hence, the ODT refers to a state after which mice have learned to discriminate an odour pair with high accuracy. In contrast, discrimination learning refers to the time when mice require to learn the discrimination of an odour pair from a naive state to highly accurate discrimination (490% correct responses).
In the first experiment, GABRB3 lox/lox mice were used only by testing shaminjected (rAAV-eGFP) versus rAAV-Cre-injected mice. The second experiment NATURE COMMUNICATIONS | DOI: 10.1038/ncomms9950 ARTICLE was designed with WT animals (no floxed alleles) versus GABRB3 lox/lox animals. First, all animals went through the established behaviour protocol to see whether the presence of the loxP sites per se would change reaction times. In this experiment, before the test odours, cineol and eugenol were used for task habituation until the animals reached 480% odour discrimination accuracy. After this first behaviour round, all animals were injected with rAAV-Cre (same production batch) and reaction times were again accessed.
Data analysis. Immunohistochemistry data were analysed in ImageJ. The procedure is illustrated in Supplementary Figs 2 and 3. First, the data were filtered using a two-dimensonal (2D) anisotropic diffusion filter. To reduce noise, a mask at the centre (180°) of the Fourier frequency domain of the images was applied by setting the pixel intensity to zero in each image, within a radius of 5 pixels. Afterwards, a Fourier-based bandpass filter was used to smoothen the signal. Colocalization analysis was performed with the imageJ colocalization finder plugin, and 3D visualizations were generated using the ImageJ 3D viewer plugin. Snapshots of the 3D visualizations are presented in the figures.
Electrophysiology and behaviour data analysis were performed using self-made custom software written in Igor Pro6 (Wavemetrics Inc).
Data are reported as mean ± s.e.m. t-test refers to the unpaired two-sided Student's t-test unless otherwise noted. For each experiment, the number of samples tested was sufficiently big to match the statistical requirements for population representation and robust statistical analysis. Each experimental procedure was repeated at least three times to evaluate the coherency of results across experiments, and data were pulled together in the statistical analysis.