Introduction

The ability to recognize and react to environmental stimuli is essential for the survival of animals. Insect olfactory systems allow the sensing and integration of volatile signals to identify food resources and avoid potential threats (Wilson, 2013). Although CO₂ is a component of the atmosphere, local changes in ambient CO₂ levels can induce context-dependent behavioral responses (Faucher et al. 2006; Hansson & Stensmyr, 2011). Walking Drosophila show robust CO₂-induced avoidance (Suh et al. 2004; Kwon et al. 2007; Faucher et al. 2006), while in more aroused states such as flight or foraging, CO₂ becomes an attractive olfactory cue (van Breugel et al. 2018) which is linked to dopamine levels affecting mushroom body output neurons (Lewis et al. 2015; Sachse & Beshel 2016). At the circuit level, prolonged CO₂ exposure leads to olfactory habituation, mediated by plasticity of GABAergic interneurons in the antennal lobe (Das et al. 2011; Olsen & Wilson, 2008; Nagel & Wilson, 2016; Sachse et al. 2007; Semelidou et al. 2018). Interestingly, Drosophila avoidance behavior directly correlates with the activity of CO₂-specific olfactory receptor neurons (MacWilliam et al. 2018), raising the question of how CO₂-induced sensory activity is transformed into central valence representations and modulated to support context-dependent behavior (Frechter et al. 2019).

The peripheral CO₂ circuit follows the canonical organization of olfactory sensory channels. A single population of ORNs (ab1C) co-express the gustatory receptors Gr21a and Gr63a and projects ipsilaterally to the V glomerulus in the ventral antennal lobe (Suh et al. 2004; Kwon et al. 2007; Vosshall & Stocker 2007). CO₂ sensation is transmitted to higher brain centers through two glomerular PNv types: a bilateral cholinergic PN (PNv^bi) projecting to both the mushroom Body (MB) and lateral Horn (LH), and a LH specific relay via a unilateral PN (PNv^uni) (Grabe et al. 2016; Lin et al. 2013; Sachse et al. 2007; Tanaka et al. 2012). Although distinct spatial LH domains encoding opposite hedonic values and odor intensity have been identified (Frechter et al. 2019; Das Chakraborty et al. 2022), glomerular PNs provide synaptic input to a diverse set of LH neurons making the identification of central brain representation and context-dependent modulation of a given sensory channel difficult (Dolan et al. 2019; Scheffer et al. 2020).

Here we describe the identification of a novel CO₂ relay channel, morphologically and functionally separate from the canonical glomerular circuit via a polarized olfactory interneuron, LN23. The LN23 pathway mediates a robust CO₂-induced aversion largely unaffected by glomerular circuit plasticity but allows early multi-modal integration. Furthermore, the parallel pathways of glomerular and extraglomerular PNs converge onto a small group of LH interneurons as the core domain for CO₂ valence coding in the central brain.

Material and Methods

Drosophila Stocks

VT031497 (PNv^bi Gal4 and the alternative line: R53A05; BDSC#38859; BDSC: Bloomington Drosophila Stock Center), NP7273 (PNv^uni Gal4), R44A02 (LN23 Gal4; BDSC#50196) were used to target neurons PNv^bi, PNv^uni and LN23 respectively. LN23-lexA (BDSC#53644) and PNv^bi-lexA (VT031497 lexA) were used in sybGRASP (BDSC#93197 / reverse direction BDSC#64315) and tGRASP (BDSC#79040/reverse direction BDSC#79039) experiments to express the postsynaptic GFP fragment. Gr21-Gal4 (BDSC#57600) was used to target ORNs that project specifically to the V-glomerulus. To target LNv1 and LNv2, we used NP1227P from the Max Planck Institute and NP2426-P from the Kyoto Stock Center. The Gal4 driver line BDSC #40444 was used for l2LN19. UAS-shibire^ts (Gift from Barry Dickson) was used to inhibit neurons at restrictive 31°C. UAS-ReaCh (BDSC#53741) was used in optogenetic experiments to artificially activate specific neuronal subsets which are labelled by the Gal4 lines. CantonS and w1118 wild-type flies were used for control experiments. The Gr63a¹ (BDSC#9941) mutant stock was used in behavior experiments as a control for flies which are unable to detect CO₂. In addition to our main LN23 line (BDSC#50196) an additional line also labeling LN23 (BDSC#48408) was used for optogenetics. UAS-CaLexA (BDSC#66542) was used in combination with the Gal4 lines to label active neurons. To further activate or silence specifically active neurons, a CaLexA stock without GFP was used (BDSC#66543) in combination with lexAop-ReaCh or lexAop-shi^ts respectively. UAS-DenMark::cherry was used to label dendrites of neurons while UAS-syb::GFP and UAS-syt::GFP specifically labeled presynaptic terminals. To confirm the location of active T-bars in LN23 and PNv^bi neurons, we also applied UAS-brp::GFP (BDSC#36291 and BDSC#36292).

Immunohistochemistry

Flies were dissected in PBS (phosphate buffer solution) and brains were fixed in 2% PFA (Paraformaldehyde). After the fixation brains were washed 3 times with 0.3% PBT (phosphate buffer with TritonX) and then blocked with goat serum. After 1 hour of blocking the first AB (Antibody) was applied and the brains were kept in a 4°C fridge overnight. The next day brains were washed again 3 times before applying the second AB. After the staining protocol was complete, brains were mounted and subsequently scanned with a confocal microscope. Anti-Ncad rat (1:10, DHSB) was used to label all neuropiles in the fly brain in all experiments. For CaLexA experiments anti GFP rabbit (1:5000, Invitrogen) was used additionally to enhance the signal strength, according to Masuyama et al. (2012). GRASP experiments, DenMark and labeling of presynaptic terminals (syt::GFP and syb::GFP) didn’t require any antibodies besides anti-Ncad since the endogenous signal was sufficient and changes after CO₂ incubation should not be distorted. For MCFO, first Antibodies were Anti-Ncad rat (1:10) or Anti-ollas rat (1:2000), Anti-flag mouse (1:2000) and Anti-HA rabbit (1:2000). As second antibodies anti-rat 647 (1:500), anti-mouse (highly cross-absorbed, 1:300) and anti-rabbit (1:500) were applied.

MCFO (Multicolor flip-out)

To label single cells in a usually rather broad expressing Gal4-line, we took advantage of a method established by Nern et al. 2015, called MCFO. They recombined four constructs, consisting of an UAS-site, followed by a stop-cassette to suppress transcription and one of four different transmembrane proteins, which can, if expressed, be labeled via antibodies. These four independent constructs can also be expressed together in the same cell, which leads to a mixture of colors and increased probability of one cell getting labeled in one distinct color. The stochastic excision of the stop cassette is achieved by a heat sensitive enzyme, whose activation is triggered by incubating the flies at 37°C in a water bath.

GRASP

To visualize synaptic contacts two different GRASP tools were used. SybGRASP is activity dependent because one fragment of GFP is located on vesicle membranes. As vesicles fuse with the presynaptic membrane, which is dependent on incoming action potentials, the large GFP fragment is exposed in the synaptic cleft and can reconstitute with the small GFP fragment located on the postsynaptic membrane (Macpherson et al. 2015). tGRASP, or targeted GRASP (Shearin et al. 2018), is activity independent and the two GFP fragments are bound to pre and postsynaptic specific proteins (Cacophony and Telencephalin respectively). A comparison of these two GRASP techniques allowed us to make assumptions about changes of synaptic activity and changes to synaptic density. No antibodies against GFP were used since the endogenous expression was sufficient. GFP intensities were quantified with FIJI, which is measured as the mean gray value of a defined region of interest (ROI) above a threshold. To subtract the mean gray value of the background noise, intensity was measured in another glomerulus (same size of ROI) which was not showing any signal. Each datapoint in the diagrams represents one quantified glomerulus. The same applies to intensity measurements for the following two sections.

DenMark::cherry / Syt:GFP, syb::GFP

UAS-syt::GFP and UAS-syb::GFP were used to label presynaptic terminals of target neurons. Synaptotagmin (syt) and synaptobrevin (syb) are proteins of the vesicle membrane and therefore a Gal4 driven expression of the GFP tagged proteins shows strong signal in the presynaptic terminals. UAS-DenMark (Dendritic Marker) is a somatodendritic marker (cherry) to label dendrites. GFP/Cherry intensities were quantified with FIJI.

CaLexA

CaLexA (Masuyama et al. 2012) was used to label active cells. It is based on the NFAT transcription factor, which is calcium responsive. As Ca⁺ enters a cell, a chimeric lexA-VP16-NFAT, enters the nucleus where expressed lexA can bind to lexAop inducing the expression of GFP. The stronger and longer the excitation lasts, higher GFP intensities are expected. GFP levels were quantified in FIJI as it was described in the section about GRASP. There was no selection for gender and all flies tested were 2 days old at the start of the experiments. In the experimental condition where flies were kept under the elevated level of 5% CO₂ the duration of the exposure was 3 days. We also used a CaLexA construct where expressed lexA is not expressing GFP but instead driving the expression of a temperature sensitive shibire protein to silence neurons. Further we also used this stock in combination with lexA driven expression of ReaCh to specifically further activate physiologically active neurons. 99% spermidine was used as an inhibitor to GR21a receptors. 5µl were put on a small piece of cotton which was then transferred to a vial containing the flies which were used in the experiments after 3 days of exposure.

Behavioral experiments

For behavioral experiments, flies were kept in total darkness, in 25°C incubators in vials containing fly food. Flies which were used in optogenetics were raised on fly food containing all-trans-retinal to enhance their response to light exposure (Vilinsky et al. 2018). All flies were age matched and tested when they were 3-4 days old. Since there was no observed difference between male and female flies, both sexes were used in all behavioral experiments. As we did not observe any behavioral artifacts caused by either green or red LED light we opted for the green LEDs, since they elicited a stronger response in the flies (compared to Inagaki et al. 2014).

Flies were immediately transferred to empty vials which can be connected to the CO₂ or optogenetic setup after they were taken out of the incubators. The whole procedure was also done in darkness (red light was used to be able to handle the flies, which is barely visible for the flies) to minimize visual distraction affecting the fly’s behavior. The whole apparatus was shifted sideways, and the flies were gently moved to the elevator segment of the setup. The elevator was lowered to a position short above the decision point and the flies were given 60 seconds rest. After the resting period either a green LED light or CO₂ flow was activated and the elevator was lowered to the decision point.

The animals were given 20 seconds to move inside the behavioral setup and make a choice. Afterwards the elevator was pulled up again, so that the flies remained in either the test tube, control tube or the central elevator area. Each of these populations was counted and a preference index (PI) calculated, defined as the percentage of flies located in the control arm (air-filled /dark) of the setup, minus the percentage of flies located in the test arm (CO₂ filled/light). Since in some of the tested populations, a few animals remained in the central elevator part of the arena, their decision was defined as neutral. Their number was divided by two and added to both the test and control population for the PI. Each datapoint in the diagrams represents the PI of a group of flies (n=15-40).

Hemibrain analyses

Connectivity data were obtained from the neuPrint interface (neuPrint, Janelia Research Campus; https://neuprint.janelia.org/) built on the Hemibrain connectome (Scheffer et al. 2020). It was used to study and model the neurons involved in the V-circuit. The main goal was to identify synaptic connectivity between the different layers of information processing within the circuit. Finally, the putative circuit was reconstructed in 3D and PN/LN output zones were identified. Neuronbridge.janelia.org was used to gather 3D images from l2_PNm17 and sm_PNm1. FlyWire.ai was a tool to quickly get information about the predicted neurotransmitter identities of neurons.

Statistics

All diagrams and statistical analyses were done by using GraphPad Prism 9 software. Normally distributed data were compared with a student’s t-test, whereas datasets that did not meet normality assumptions were analyzed with a Mann–Whitney U test. Behavioral data spanning multiple days were tested with a one-way ANOVA. Statistical significance is indicated in the figures by asterisks, with p < 0.05 (*), p < 0.01 (**), p < 0.001 (***), and p < 0.0001 (****).

Results

Glomerular circuit dynamics of CO₂ perception

The V glomerulus at the ventral AL surface is innervated by a single population of about 40 CO₂-sensitive olfactory receptor neurons (ORNv) expressing the two gustatory receptors GR21a and GR63a (Kwon et al. 2007; Jones et al. 2007) (Fig.1A-B). Each ORNv axon forms 3-5 terminal branches tightly packed in a non-overlapping fashion with other ORN terminals (Fig. 1, C-C”), providing sensory input to glomerular inter– and projection neurons (Fig1D-G’) (Grabe et al. 2016).

Figure 1.

To identify relay neurons of the V glomerulus with different physiological properties we screened a collection of Gal4 lines using the CaLexA activity reporter to compare intracellular Ca²⁺ levels under ambient CO₂ conditions (Koyama et al. 2021; Masuyama et al. 2012) (Fig. 1H-O). Gal4 lines for ORNv show strong CaLexA reporter activity (Fig.1H), in line with a high default activity of CO₂-receptive sensory neurons due to the absence of GABA_B-mediated presynaptic inhibition (MacWilliam et al. 2018; Olsen & Wilson, 2008; Root et al. 2008). In contrast to elevated Ca²⁺ levels in ORNv, the two uniglomerular PN classes of the V glomerulus, unilateral V_l2PNs (PNv^uni) and bilateral V_ilPNs (PNv^bi, Fig. 1D) do not show any significant level of CaLexA activation under ambient air conditions (Fig. 1I, J). Two main classes of local interneurons, LNv1 and LNv2 display moderate levels of CaLexA activity at atmospheric CO₂ levels (Fig. 1K, L). A strong CaLexA-GFP signal, comparable to ORNv can be observed for l2LN23 (LN23), a less characterized olfactory class of sparsely innervating local interneurons (Fig.1F, F’, M and N) (Liou et al. 2018). Fluorescence signal of Ca²⁺-induced GFP expression quickly decreases in LN23 following sensory deprivation in adult flies via removal of both antennae (Fig.1O). Furthermore, an extended exposure to the ORNv inhibitor spermidine (MacWilliam et al. 2018) results in a reduced CaLexA activity not only in ORNv, but also LN23, indicating a preferred physiological coupling of sensory neurons to LN23 interneurons under ambient CO₂ conditions (Fig. 1O).

A predominant relay of CO₂-induced sensory activity by LN23 is supported by the modified activity-dependent GRASP technique (Macpherson et al. 2015). Expression of the presynaptic syb::GFP^1-10 in ORNv and the postsynaptic GFP¹¹ in AL neurons revealed a significantly stronger GRASP signal for ORNv➔LN23 connectivity compared to the ORNv➔PNv^bi synaptic activity (Fig.2F and G). Similarly, the activity-independent tGRASP (Shearin et al. 2018) is most prominent for ORNv➔LN23 (Fig.2I and J), which is in line with connectome data (neuPrintExplorer, hemibrain; Scheffer et al. 2020; Schlegel et al. 2021). In summary, ORNv activity at ambient CO₂ levels seems to be mainly represented by the LN23 class of local interneurons and less by glomerular PN relay neurons, thereby identifying a new circuit component for the ORNv-induced behavioral aversion.

Figure 2.

Cell type plasticity of glomerular relay neurons following chronic CO₂ stimulation

We next determined how the overall activity of V glomerulus relay– and interneurons changes following long-lasting sensory stimulation. Adult Drosophila melanogaster avoids ambient levels of CO₂ in the T-maze choice assay (Faucher et al. 2006; Suh et al. 2004) (Fig. 2A) but shows behavioral habituation following chronic CO₂ exposure (Das et al. 2011, Fig. 2B, C). The reduction of CO₂-induced aversive responses is associated with an increase in the synaptic volume of the V glomerulus (Fig.2D; Sachse et al. 2007 and Das et al. 2011). At the ORN level, mosaic analysis revealed an extension in the surface area and size of individual synaptic boutons of ORNv axons (Fig.2E). Compared to ambient CO₂ levels, prolonged elevated CO₂ (5%) sensation results in neuron-type specific changes of CaLexA activity (Fig. 1H’-N’) (Masuyama et al. 2012). While PNv^bi neurons showed a strong increase in Ca²⁺ levels following CO₂ habituation (Fig.1I’), whereas no CaLexA-dependent GFP expression was observed in PNv^uni (Fig.1J’), indicating different physiological properties of the two main glomerular CO₂ relay channels (Lin et al. 2013). In addition to chronic CO₂ stimulation, short pulses of elevated CO₂ also did not induce Ca^2+-mediated-GFP expression in PNv^uni (Fig.1O). No significant changes in CaLexA-GFP levels could be detected for LN23 neurons following prolonged CO₂ exposure compared to the already high expression at ambient CO₂ conditions (Fig 1M’, N’, O).

As GABAergic local interneurons are associated with olfactory habituation (Das et al. 2011; Sadanandappa et al. 2013) we compared changes in CaLexA activation in LNv1 and LNv2 (Das et al. 2011, Sachse et al. 2007). While an increase in the Ca²⁺ reporter signal could be observed for both LN types following chronic CO₂ exposure, the activity of LNv1 is significantly higher compared to LNv2 (Fig.1K’ and Ĺ). In line with the cell-type specific CaLexA expression, previous connectome analysis showed significant differences of LN input for CO₂ relay neurons: (neuPrintExplorer, hemibrain): While both PNv^uni and LN23 do not receive major LN-type input, PNv^bi gains GABAergic modulation by lLN1 neurons (“bc” subtype, 16% of input) and a group of multiple lLN2 neuron types (X, F, P, 19%) (Schlegel et al. 2021; Bates et al. 2020). These results indicate cell type specific physiological changes of glomerular neurons following CO₂-induced behavioral habituation, in which PNv^bi shows dynamic changes in glomerular connectivity while both PNv^uni and LN23 neurons maintained their characteristic activity seen in ambient CO₂ conditions.

We next tested how the cell-type specific changes in neuronal activity during habituation impact their synaptic connectivity within the V glomerulus. A strong increase of activity-dependent sybGRASP levels for ORNv input onto PNv^bi could be detected following chronic CO₂ stimulation (Fig.2F’). In contrast, no significant changes in sybGRASP levels were observed for ORNv input onto LN23 (Fig.2G’). Of note, compared to ambient CO₂ conditions, the standard deviation for ORNv➔LN23 sybGRASP levels was reduced in flies chronically exposed to CO₂, indicating that sensory input onto LN23 also increases in the context of habituation (Fig.2H). In addition, activity-independent tGRASP expression showed that ORNs increase their synaptic connections to the two main relay neurons, PNv^bi and LN23 (Fig.2I’, J’), which is in line with morphological changes of ORN axon terminals described above (Fig.2D, E). Following prolonged CO₂ exposure, the PNv^bi class displayed increased sybGRASP levels on PNv^uni and LN23 (Fig.2L-M’’’), while no significant changes could be observed for LN23 on PNv^bi (Fig.2N-N’’’). Taken together, CO₂-induced ORNv activity seems to be relayed via two distinct glomerular output channels dependent on the stimulus strength: LN23 neurons indicate default ambient CO₂ levels while PNv^bi signals at conditions of extended elevated CO₂.

Identity, innervation patterns and polarity of second-order CO₂ sensitive neurons

To test if the difference in CaLexA activity might be related to the organization of presynaptic sites within CO₂ relay neurons we visualized the active zones using syt::GFP/brp::GFP transgenes together with the dendrite marker DenMark::mCherry (Kittel et al. 2006, Nicolai et al. 2010; Wagh et al. 2006). PNv^bi, PNv^uni and LN23 differ significantly in their expression and localization of presynaptic proteins within the dendritic compartment (Fig.3A-C). While PNv^bi shows prominent dendritic syt::GFP localization throughout the glomerulus (Fig.3C), no syt::GFP signal can be detected within PNv^uni dendrites (Fig.3A), indicating different cellular polarity of the two main glomerular PNv classes. Despite the significant expression level, syt::GFP appeared more scattered within LN23 dendrites and less restricted to the glomerulus neuropil compared to PNv^bi (Fig.3B). In fact, single optical sections allow the localization of the active zone marker Brp::GFP in LN23 dendrites outside the glomerular neuropile (Fig.3F) but in PNv^bi confined to the V glomerulus (Fig.3G). In line with these results hemibrain EM data lists enriched presynaptic site in LN23 dendrites outside the V glomerulus in contrast to the restriction PNv^bi presynaptic sites to the glomerular neuropil (Fig.3I) (https://neuprint.janelia.org/; Schlegel et al. 2021; Bates et al. 2020).

Figure 3.

Following chronic sensory stimulation, PNv^bi showed a significant increase in the density of presynaptic sites while no significant change in the amount presynaptic syt::GFP labeling of LN23 dendrites can be observed (Fig.3D). Like the subcellular organization at ambient CO₂ conditions, presynaptic sites of PNv^bi dendrites are uniformly distributed throughout the V-glomerulus (Fig.3D, last image), while most presynaptic sites of LN23 dendrites are found outside of the glomerular boundaries (Fig.3D, middle image). In contrast to CO₂ relay neurons PNv^bi and LN23, no dendritic localization of presynaptic marker transgenes could be detected for PNv^uni (Fig.3D, first image), neither under ambient nor chronic CO₂ condition, further supporting a differential synaptic profile of the two PN classes. All syt::GFP and DenMark::cherry intensities were quantified using ImageJ (Fig.3H). An additional presynaptic marker syb::GFP revealed a similar expression pattern: while PNv^bi showed a strong signal within the glomerular boundaries, which increases following CO₂ incubation, LN23 dendrites show a robust GFP signal adjacent to the neuropil border of the V glomerulus (Fig.3E). As both, PNv^bi and PNv^uni are described as being excitatory, cholinergic neurons (Bates et al. 2020; Lin et al. 2013; Schlegel et al. 2021; Tanaka et al. 2012) we investigated the neurotransmitter identity of LN23. Anti-GABA immunostaining does not label LN23 cell bodies (Supplemental Fig.3A), excluding an inhibitory function similar to canonical interglomerular LNs. These results revealed not only major differences in neuronal organization among the 3 classes of CO₂ relay neurons, but also a distinct response profile in intraglomerular plasticity of dendro-dendritic connections. CO₂ input diverges into two more robust feed-forward channels (PNv^uni towards LH and LN23 within the AL) while the dynamic PNv^bi pattern indicates intraglomerular processing.

Developmental profile of the inter-glomerular CO₂ relay neuron

Having established the polarity of LN23, our next objective was to investigate the development of this neuron, specifically focusing on how interglomerular relay neurons are integrated into stereotyped antennal lobe circuits. While the glomerulus-specific assembly of olfactory receptor neuron (ORN) axons with projection neuron (PN) dendrites has been extensively documented (Hong & Luo 2014), less is known about the mechanisms underlying the incorporation of relay neurons into these circuits. At the beginning of pupal development, PN dendrites innervate the early precursor of the adult AL and occupy distinct spatial domains before ORN axons enter the AL by around 20h APF (Jefferis et al. 2001; Jefferis et al. 2004). LN23 neurons can be traced back to the late larval stage with a cellular morphology related to SEZ (subesophageal zone) interneurons (Truman & Bate, 1988; Winding et al. 2023, Fig. 4). During early pupal stages, larval LN23 precursor neurons retract from the SEZ to accumulate at the ventral region of the AL precursor neuropile (Fig4A-C’). Here LN23 processes target a region which is devoid of dendrites of uniglomerular PNs (Fig4F, F’). Before ORN axon arrival, LN23 neurons establish a separate dendritic branch, which extends dorsally before bending back to the posterior AL and merges with the main dendritic domain (Fig. 4D-E’).

Figure 4.

Loss of the secreted molecule Wnt5, which affects PN dendrite patterning (Hing et al. 2020; Wu et al. 2014), leads to a continued growth of the dorsal LN23 branch along PN axon tracts which exit the AL towards higher olfactory brain centers (Fig.4I-K). These results indicate that LN23 neurons display an intrinsic polarity similar to uniglomerular PNs, but instead of relaying sensory information from AL glomeruli to the LH/MB, a Wnt5-dependent mechanism seems to prevent AL exit and target the presynaptic process to the extraglomerular region in the ventral AL. In wild-type, ORN axons innervate the developing V glomerulus only at the ipsilateral AL, in contrast to the majority of bilateral ORN classes (Fig.4F-H) (Couto et al. 2005). To gain insight into the mechanisms which support glomerulus-specific ORNv input on LN23, we analyzed candidate genes which are expressed on growing ORNs. Mutations in the protocadherin-type cell adhesion molecule Flamingo (fmi), which is highly enriched on ORN axons in the developing V glomerulus, affect their synaptic targeting. Here, fmi mutant ORNv axons are no longer restricted to the glomerular boundaries but extend into the posterior AL next to the output domain of LN23 (Suppl. Fig.4A, B), indicating cellular interactions to restrict class-specific axon-dendrite recognition within the glomerular domain.

In summary, despite being morphologically classified as a local interneuron, the LN23 class shows cellular and developmental features similar to those of olfactory projection neurons: a highly polarized neuronal morphology in the adult olfactory system is established by a glomerulus-specific dendrite targeting for singular ORN input and a default axonal extension along olfactory PN tracts rerouted by external signals (Komiyama & Luo, 2006).

LN23 mediates CO₂-induced aversion

To test the role of the different CO₂ relay neurons on Drosophila behavior, we expressed the shibire^ts transgene to block synaptic transmission in a cell type specific manner (Chen et al. 1991; Kitamoto 2001). Neuronal silencing of each of the two glomerular PNs led to a reduction of innate CO₂ avoidance, with stronger effects observed for PNv^uni compared to PNv^bi (Fig. 5A). While CO₂ avoidance can still be observed following the neuronal silencing of each PN class, shibire^ts induced silencing of LN23 fully abolished aversion behavior towards CO₂ (Fig. 5A). To test for cell type specificity of the behavioral phenotype following LN23 silencing, we used the Ca²⁺ dependent LexA activation of the CaLexA construct for shi^ts expression and observed a reduction of the innate CO₂ aversion (Fig. 5B) (Masuyama et al. 2012). To increase lexA levels we chronically exposed flies to high levels of CO₂. Interestingly, increased CaLexA mediated shi^ts expression via chronic exposure not only suppressed behavioral aversion but led to a significant level of attraction towards CO₂ (Fig.5B). These results indicate a prominent function of LN23 in CO₂ valence coding: Blocking LN23 activity is sufficient to fully suppress CO₂-induced behavioral aversion and together with the GABAergic suppression of PNv^bi during habituation an attraction towards CO₂ is observed, which phenocopies the silencing of ORNv (Jones et al. 2007; Kwon et al. 2007, MacWilliam et al. 2018; Suh et al. 2004).

Figure 5.

To further determine if these relay neurons are sufficient for behavioral aversion, we optogenetically activated each neuron type and measured the behavioral response (Inagaki et al. 2014; Klapoetke et al. 2014; Pulver et al. 2009) (see M&M). Light-induced activation of ORNv (GR21a-Gal4) expressing UAS-ReaCh led to strong aversion (Fig. 5C). Similarly, optogenetic activation of LN23 is sufficient to trigger behavioral aversion via two independent Gal4 driver lines (LN23 Gal4 and the alternative line we used: VT48408-Gal4; Fig.5C) or by Ca2⁺ induced LexA activation (Fig.5D). In contrast, no change in light-induced activation for both PNv^bi (two independent driver lines, see Methods) nor PNv^uni could be observed (Fig.5C). These results support a predominant role of LN23 in CO₂-induced avoidance behavior. Compared to the more dynamic activity of PNv^bi following acute and chronic sensory stimulation, LN23 shows little changes in intrinsic activity, representing ORNv activity independent of changes in CO₂ levels or temporal pattern. Based on these findings we propose a ground state of ORNv/LN23 mediated aversion induced by ambient CO₂ default state, which is modified either by the parallel PN channels or downstream of LN23.

Third-order neurons postsynaptic to LN23 have a putative role in valence coding

To gain further insights into the unique organization of the LN23 pathway compared to the glomerular relay channel, we analyzed the postsynaptic candidate neurons within the posterior AL. Two PN types, l2_PNm17 and sm_PNm1 are receiving a large proportion of their synaptic input (27% and 18%, respectively) from LN23 (Fig.6A, neuPrint, Janelia Research Campus; https://neuprint.janelia.org/; Schlegel et al. 2021). Like the two PN classes of the V glomerulus, PNm1 and PNm17 show distinct relay properties: While input into PNm1 is mostly restricted to the LN23-CO₂ channel, dendrites of PNm17 extend into adjacent VP1 and VP3-5 glomeruli to receive direct input from thermo– and humidity-receptor neurons (neuPrint, Janelia Research Campus; https://neuprint.janelia.org/; Frank et al. 2015; Marin et al. 2020). Both types of non-olfactory relay neurons target multiple distinct brain areas adjacent to the canonical olfactory LH, the posterior lateral protocerebrum (PLP), superior clamp (SCL) and the superior lateral protocerebrum (SLP) (Fig6 A, B, C) (Ito et al. 2014). To visualize the trajectories of the PNs, we used MCFO (Aso et al. 2014) to create single cell clones, which show that PNm1 and PNm17 axons turn close before entering the LH to target adjacent brain regions (Fig. 6B, C). Of note, compared to glomerular PNv^uni and PNv^bi, PNm1 and PNm17 receive only minor input from GABAergic LNs (neuPrint, Janelia Research Campus; https://neuprint.janelia.org/), indicating reduced gain control properties of the extra-glomerular ORN➔LN23➔PNm pathway (Olsen & Wilson, 2008).

Figure 6.

Optogenetic activation of each of the two PNm class neurons downstream of LN23 induce opposite behavioral responses: While activity of the “CO₂-only” class PNm1 results in strong aversion, the “multi-modal” PNm17 initiates behavioral attraction (Fig.6D). Furthermore, CO₂-induced aversive behavior could be attenuated following shi^ts-mediated silencing of PNm1 but was not changed by silencing PNm17 (Fig.6 E, F). Interestingly, FlyWire.ai (Dorkenwald et al. 2024; Eckstein et al. 2024; Schlegel et al. 2024; Zheng et al. 2018) revealed distinct neurotransmitter identities for the two PNm classes, GABA expression for PNm1, while PNm17 shows a cholinergic synaptic profile. Together with the behavioral responses following optogenetic activation this suggests a circuit motif in which the default activity of the ORNv➔LN23➔PNm1 channel by ambient CO₂ sensation blocks a central brain region important for the generation of behavioral attraction (Dolan et al. 2019).

Connectivity in higher brain centers and lateral communication in the antennal lobe

As the glomerular PNs and the extra-glomerular LN23➔PNm channels target adjacent central brain regions we asked whether the two parallel CO₂ pathways converge onto common downstream neurons. The hemibrain dataset (https://neuprint.janelia.org/; Plaza et al. 2022; Scheffer et al. 2020) identified three main types of glutamatergic LH neurons (FlyWire.ai; Dorkenwald et al. 2024; Eckstein et al. 2024; Schlegel et al. 2024; Zheng et al. 2018) as PNv or PNm targets: LH^PD5c1 as the primary target of PNv^bi (Fig.6 G), LH^PV9b1 downstream of PNm1 (Fig.6 H) and LH^CENT4 as the major postsynaptic partner of PNv^uni (Fig.6 I). All 3 LH target neurons of the parallel CO₂ pathways are directly interconnected (Fig. 8): While LH^PV9b1 not only receives input from PNm1 but also from PNv^bi, LH^PD5c1, the main downstream target of PNv^bi, is also targeted by LH^CENT4, the primary postsynaptic partner of PNv^uni. This core motif of CO₂ perception receives additional input from other sensory modalities. LH^PD5c1 gains input from “food-sensing” glomeruli (DM4, DC4/DP1, and VM7) via class-specific cholinergic PNs, indicating a neuronal integrator of CO₂ and food-related cues (Couto et al. 2005; Dumenil et al. 2025; Min et al. 2013). In addition, LH^CENT4 is also a main target for l2PN^VP4+VL1, mediating the perception of dryness and pyrrolidine-type odorants, thereby adding another layer of multi-modal sampling to the LN23 ➔ PNm1 pathway (Fig.8) (Li et al. 2022; Choi et al. 2022; Ni 2021). Interestingly, the convergence of the glomerular CO₂ and food channels is not only observed for LH neurons downstream of AL relay pathways but seems to occur already at the level of AL glomeruli. Hemibrain describes with l2LN19 another class of polarized local interneuron (Fig.7A, B) with enriched postsynaptic sites at “food-sensing” glomeruli DM4, DC4/DP1, and VM7 (Fig.7A) as well as targeted output onto PNv^bi and PNv^uni in the V-glomerulus (Fig.7B), but merely with LN23 (Fig.7C) (https://neuprint.janelia.org/; Schlegel et al. 2021; Bates et al.2020). Accordingly, optogenetic activation of LN19 led to an aversive behavioral response (Fig.7D), while inhibition had no effect (Fig.7E).

Figure 7.

Figure 8.

Based on these distinct connectivity patterns and behavioral response profiles of individual relay neurons we are proposing the following circuit mechanism for CO₂ valence coding (Fig.8): The activity of CO₂-responsive sensory neurons (ORNv, black) is relayed by parallel PN channels and converge onto two central brain target neurons of opposite intrinsic valence identity. The glomerular PNs target the ventral domain of LH^PD5 to trigger behavioral aversion and extraglomerular PNs activated via the lateral LN23 relay converge onto LH^PV9 which mediates attraction. While PNm17 supports LH^PV9 activity, GABAergic PNm1 blocks LH^PV9 through which the negative valence state of LH^PD5 triggers avoidance. This core aversion pathway (PNv^bi/PNv^uni/PNm1), without prominent GABAergic modulation, directly represents the activity of CO₂ sensory neurons, is modified either via the activation of the dorsal LH^PD5 domain by fruit odor PNs (PN^DM/VM/DC) or by an increase of PNm17 through non-olfactory sensory neurons to support a context-dependent behavioral response repertoire.

Discussion

In Drosophila, the activation of a single class of CO₂ receptive sensory neurons induces avoidance behavior, which is modified by multiple external and internal stimuli (Bräcker et al. 2013; Faucher et al. 2006; Dissegna et al. 2021; Suh et al. 2004; Turner and Ray, 2009). How does the Drosophila brain transform ORNv input into flexible behavioral responses towards CO₂? Here we identified a novel CO₂ relay channel with unique morphological and physiological properties compared to the canonical glomerular PN pathway. First, instead of direct transmission of sensory input by glomerular PNs in the anterior AL towards higher brain centers, the ORNv induced activation of the local interneuron LN23 relays peripheral CO₂ sensation to an extraglomerular neuropil of the posterior AL generating two parallel output channels with different target areas. Second, in contrast to the high level of neuronal plasticity observed for PNs within glomerular circuits, LN23 analysis revealed very little dynamic changes under different physiological conditions. This might be due to distinct intrinsic features of PNs and LN23 such as the expression of ion channels or receptor molecules but could also be explained by the absence of glomerular GABAergic neurons at the extraglomerular CO₂ pathway. Third, following the manipulation of neuronal activity, only LN23, but no individual PNs are able to induce an aversive behavioral response similar to ORNv. Finally, the extraglomerular CO₂ pathway includes a structural segregation of opposite valence-encoding elements, in which the “CO₂-only” neuron induces aversion while activation of the parallel “multi-modal” channel determines behavioral attraction. These findings are consistent with a designated circuit motif encoding a default behavioral state of avoidance even under atmospheric CO₂ levels. As the “CO₂-only” relay neuron PNm1 has a predicted GABAergic identity (FlyWire.ai; Dorkenwald et al. 2024; Eckstein et al. 2024; Schlegel et al. 2024; Zheng et al. 2018), the previously reported ORNv induced avoidance (MacWilliam et al. 2018; Das et al. 2011) could be explained with the inhibition of a premotor area by the LN23 channel.

The interplay between the extraglomerular default channel and the glomerular dynamic pathway in encoding behavioral avoidance was demonstrated in the context of habituation. While the increased GABAergic inhibition of glomerular PNs following chronic CO₂ sensation reduces innate aversion, the simultaneous silencing of LN23 not only eliminates all aversive response features but reverses CO₂ valence to trigger attraction. Interestingly, the reduced aversion through inhibition of PN activity in the context of habituation seems partially compensated by increasing PNv^bi synaptic plasticity. We observed a strong gain of presynaptic sites of glomerular PNv^bi dendrites onto LN23, which might maintain a balanced level of aversion activity for local CO₂ levels following habituation (Das et al. 2011; Sachse et al. 2007). Although both PNv^uni and PNv^bi are excitatory cholinergic projection neurons (Tanaka et al. 2012; Lin et al. 2013; Bates et al. 2020; Schlegel et al. 2021), they differ in their cellular organization and behavioral function. In contrast to PNv^bi, PNv^uni dendrites lack presynaptic sites, therefore specifies a feed-forward channel similar to LN23. Based on these results we are proposing complementary function of the three relay neurons in mediating flexible CO₂ behavioral response patterns: While LN23 encodes default avoidance under ambient conditions, supported by PNv^uni as a sensor for elevated local levels of CO₂, PNv^bi defines the main component of experience-dependent modulation by LNs and neurons of the MB calyx (CA) (Bräcker et al. 2013).

Interestingly, developmental analysis revealed a polarized growth pattern of LN23 that resembles projection neurons more than typical local interneurons (Jefferis et al. 2004; Komiyama & Luo 2006). LN23 originates from a larval interneuron located in the suboesophageal zone (SEZ), which undergoes remodeling into an adult olfactory interneuron during early pupal stages (Truman & Bate, 1988). We hypothesize that the secreted Wnt5 protein, a critical factor for initial projection neuron dendrite patterning, supports the targeting of LN23 processes to the posterior antennal lobe (AL) region. Like the genetically defined recognition mechanisms that guide glomerulus-specific ORN-PN connectivity, growing ORNv axons appear to target LN23 dendrites as synaptic partners. Since LN23 dendrites are organized into a defined glomerular input and extraglomerular output zone, we propose Flamingo-dependent expression on ingrowing ORN axons to prevent sensory input to LN23 outside the glomerular domain.

The divergence of CO₂-induced ORN activation into parallel relay channels allows for early convergence of additional sensory information. This includes not only the integration of non-olfactory cues by the positive valence branch PNm17 downstream of LN23 (Frank et al. 2015; Li et al. 2022), but also the sampling of food-odor-sensitive glomeruli by the polarized interneuron LN19 for V-glomerulus input (Couto et al. 2005; Dumenil et al. 2025; Min et al. 2013). The convergence of different parallel CO₂ channels onto a common set of lateral horn (LH) interneurons suggests a potential hub for motor action control (Dolan et al. 2019). The results presented in this study support a conceptual model in which the default state of behavioral aversion, induced through inhibition of an LH motor hub by the “CO₂-only” ORNv→LN23→PNm1 pathway, is modified by the inclusion of additional glomerular and extraglomerular projection neuron classes to allow approach behavior. Further studies on the physiological architecture of the CO₂ core circuit within and downstream of the LH may reveal deeper insights into the multi-modal processing underlying flexible, context-dependent behaviors.

Additionally, despite the conserved molecular receptor identity of CO₂ sensory neurons among insects, species-specific behavioral responses to local CO₂ concentrations are evident (Turner & Ray 2009). In many species, such as mosquitoes, elevated CO₂ levels signal potential food sources and elicit an attractive response (Tauxe et al. 2013; Siju 2009; Ghaninia 2007). The identification of a dedicated CO₂ aversion channel in D. melanogaster opens new avenues for comparative research into circuit motifs and provides insight into the neural basis of behavioral flexibility across insect evolution (Oram & Card 2022).

Figure supplements

Supplementary Figure 1.

Supplementary Figure 3.

Supplementary Figure 4.

Data availability

Quantified GFP intensity in glomeruli (image-derived numerical data) for GRASP and CaLexA experiments. Quantified DenMark::cherry signal intensity for the respective experiment. Quantified volume of glomeruli after CO2 exposure. Calculated a performance index from behavioral experiments (CO2 and optogenetics). Used data that can be accessed on https://neuprint.janelia.org/ and https://flywire.ai/. All data points are visible in the respective figures.