Extremely Sparse Olfactory Inputs Are Sufficient to Mediate Innate Aversion in Drosophila

Innate attraction and aversion to odorants are observed throughout the animal kingdom, but how olfactory circuits encode such valences is not well understood, despite extensive anatomical and functional knowledge. In Drosophila melanogaster, ~50 types of olfactory receptor neurons (ORNs) each express a unique receptor gene, and relay information to a cognate type of projection neurons (PNs). To examine the extent to which the population activity of ORNs is required for olfactory behavior, we developed a genetic strategy to block all ORN outputs, and then to restore output in specific types. Unlike attraction, aversion was unaffected by simultaneous silencing of many ORNs, and even single ORN types previously shown to convey neutral valence sufficed to mediate aversion. Thus, aversion may rely on specific activity patterns in individual ORNs rather than the number or identity of activated ORNs. ORN activity is relayed into the brain by downstream circuits, with excitatory PNs (ePN) representing a major output. We found that silencing the majority of ePNs did not affect aversion, even when ePNs directly downstream of single restored ORN types were silenced. Our data demonstrate the robustness of olfactory aversion, and suggest that its circuit mechanism is qualitatively different from attraction.


Introduction
Olfactory circuits in insects and mammals exhibit striking functional and anatomical similarities [1]. Functionally, olfactory inputs are high dimensional, corresponding to the activation patterns of~50 types of olfactory receptor neurons (ORNs) in flies and~1000 ORN types in mice. Among these input channels, a minority are "specialists", responding specifically to one odorant or one category of odorants [2][3][4]. Most input channels, however, are "generalists" where one odorant activates multiple types of ORNs and one ORN type responds to multiple odorants [5]. Anatomically, ORNs of the same type express the same receptor, and project to the same glomerulus in the antennal lobes in flies [6][7][8] or olfactory bulbs in mice [9,10]; each projection neuron (PN) in flies or mitral/tufted cell in mice then relays information from a single glomerulus to higher brain centers for associative learning and innate behavior [11][12][13][14]. Such similarities in organization across diverse phyla suggest that convergent evolution might have produced a particular neural anatomy that suits olfactory processing.
What aspects of ORN activity are required for higher brain centers to direct olfactory aversion and attraction [15]? In C. elegans each sensory neuron is preferentially linked to aversion or attraction [16,17]. However, unlike insects and mammals, worms have much fewer olfactory sensory neurons, each of which expresses multiple receptor genes [18]. It is thus unclear whether the conclusions from worms are applicable to olfactory systems with more neurons and neuronal types. Here, we exploit the genetic tools in Drosophila melanogaster to probe the causal relationship between ORN activity and innate olfactory aversion.
Similar to C. elegans, aversion can be induced in flies when single "specialist" ORN types are activated by specific repellents, such as CO 2 [3] or geosmin [4], emitted by other stressed flies or toxic microbes, respectively. The corresponding ORNs are thus believed to stereotypically convey aversion. However, how "generalist" ORN activities are relevant for aversion remains controversial. Flies tend to avoid odorants at higher concentrations, regardless of chemical compositions [19], and increased odorant concentration activates more ORN types [5]. Three classes of hypotheses are consistent with these observations, the conceptual elaborations of which can be found in a recent review on intensity coding in mammalian olfaction [20]. First, for aversion, downstream circuits may extract a population metric such as the number of activated ORN classes (Hypothesis I). Alternately, a downstream aversion circuit may be hardwired to ORNs expressing specific receptors. In this model, aversion-specific ORNs with low and non-specific odorant affinity may signal only when any odorant reaches sufficiently high concentrations (Hypothesis II). In a third scenario, specific activity patterns within individual ORNs may be interpreted by higher brain centers as signals of aversive cues, while the number or identity of activated ORNs is incidental. For example, as odorant concentration increases, aversion could be encoded by elevated firing rate or reduced latency in any "generalist" ORNs.
The identity of activated ORNs have been correlated with aversion and attraction in larval [21] and adult flies [22]. However, these studies implicitly adopted Hypothesis II and could not causally test other possibilities. Meanwhile, attempts to establish causal relations were limited by the fact that ORNs can be divided into two major classes expressing either odorant receptor (OR) family [23] or ionotropic receptor (IR) family of receptors [24]. Previous studies only manipulated either OR+ or IR+ ORNs, while outputs from the other class complicated the interpretation of results [25,26]. To overcome these limitations, we applied a genetic strategy of blocking all ORN activity, and then restoring activity in ORNs expressing specific receptors. Our results indicate that aversion does not require the overall ORN activity pattern, and that a single ORN type can convey aversion. Moreover, our data demonstrate that aversion is intact after inactivating a large subset of ePNs that express GH146-GAL4, including those directly post-synaptic to the active ORNs, long considered to be the major output route from the antennal lobe. Our data constrain future models of valence coding in the olfactory circuit.

Olfactory chemotaxis assays
The four-quadrant behavioral arena was 16.5 cm by 16.5 cm, and 1 cm deep [39]. It was placed inside a 33°C box in complete darkness. The airflow was filtered and saturated with water, entered each quadrant at a rate of 40 mL/min, and left through the central hole in the arena floor. One branch of airflow was controlled by solenoid valves through the LabView software (National Instruments), so that for each trial air passed directly into one quadrant for 2.5 min, and then was switched to bubble through 5 mL of water (for vinegar and acetic acid) or paraffin oil (for all the other odorants tested) containing an odorant with a specified concentration. The odorant source was replenished for each experiment.
Flies were raised at 25°C, collected within 2-7 days of eclosion, and deprived of food for~1 day in a vial with a wet Kimwipe. All tests were carried out in the morning peak activity window. The flies were pre-incubated at 33°C for 5 min, loaded in the arena in groups of 15-20 through a central hole in the bottom glass, tested with the aforementioned 2.5-min-air-5-minodor paradigm, and then discarded. The "permissive temperature" control flies were raised and starved at 18°C, and then tested at 25°C. We used females throughout this study.
Illumination, tracking (30 frames/s), and pre-processing of data followed published protocols [40]. All analyses were performed with Matlab (Mathworks). The algorithm faithfully kept identities of individual flies, until they bumped into each other or walked onto the reflective edge of the arena. Each continuous segment of fly positions thus corresponded to the unambiguous trajectory of one individual. We only excluded traces that had not moved during the entirety of their extent, which corresponded to dead flies or reflections. The final dataset consisted of short trajectories, each consisting a fly's x-y positions over consecutive frames.
In each frame, we then counted the number of flies in each quadrant. To calculate the preference index (PI), we added the counts from all the frames within a specific period, and then used the equation in Fig 1A. Based on a previous study [39], the PI for aversion was calculated with data between 2.5-5 min of the 5-min odorant period, during which aversion reaches steady state; the PI for attraction was calculated with data between 1-3 min, because attractive behavior appears to be desensitized after that.

Aversion is resistant to broad ORN inactivation
To quantify olfactory behavior in adult flies, we delivered innately attractive or aversive odorants to a single quadrant of an arena and air to the other three quadrants [39], and simultaneously tracked 15-20 flies in each trial ( Fig 1A). We defined Preference Index (PI) as a comprehensive indicator of olfactory behavior (Fig 1A, see Methods for details), where a value of 1 means that flies never leave the odorant quadrant, a value of 0 means that flies respond neutrally to the odor and air quadrants, and a value of-1 means that the flies invariably avoid the odorant quadrant.
We first assessed the impact of ORN silencing on attraction to apple cider vinegar. ORNs can be divided into two classes expressing either an odorant receptor (OR) or an ionotropic receptor (IR). We used Orco-GAL4 [23] to target all the OR+ ORNs, and expressed UAS-shibire ts1 (Orco > shi ts1 ), which acutely abolishes synaptic transmission at its restrictive temperature [27]. Compared to the robust attraction in control flies, blocking OR+ ORNs abolished attraction to vinegar (Fig 1B 1 ); as another control, Orco > shi ts1 did not affect attraction at the permissive temperature where shi ts1 does not block synaptic release (Fig 1B 2 ). Since acetic acid, the major component of vinegar, preferentially activates the Ir8a-expressing subset of IR + ORNs [8], we also examined Ir8a > shi ts1 flies [28], and found that they were also less attracted to vinegar (Fig 1B). Thus, normal attraction to vinegar requires both OR+ and acidsensing IR+ ORNs.
We then tested the roles of OR+ and acid-sensing IR+ ORNs in aversion, using E2-hexenal and acetic acid, two representative chemicals preferentially activating OR+ and IR+ ORNs respectively. In contrast to attraction, aversion was not affected by blocking OR+ ORNs or acidsensing IR+ ORNs (Fig 1C and 1D).

Abolishing aversion with pan-ORN inactivation
As aversion to acetic acid and E2-hexenal appeared resistant to the silencing of subsets of ORNs, we hypothesized that ORNs expressing different receptors redundantly mediate aversion. To test this hypothesis, we expressed shi ts1 with the pan-ORN driver Pebbled-GAL4 [29] to block synaptic transmission in all ORNs. However, Pebbled > shi ts1 caused lethality, likely due to shi ts1 expression outside the olfactory system. We thus used an intersectional approach to selectively target ORNs by combining Pebbled-GAL4 with ey-FLP [32], a recombinase preferentially expressed in the eye-antennal disc during development [30], as well as two copies of UAS-FRT-stop-FRT-shi ts1 (Pebbled AND ey > shi ts1 ) [31]. Here the FRT-flanked stop signal is only removed by FLP in ey+ neurons (Fig 2A). Visualized using UAS-FRT-stop-FRT-CD8::GFP [32], the intersection between Pebbled and ey specifically labeled ORNs as well as sensory neurons projecting to antennal mechanosensory and motor center and suboesophageal ganglia ( Fig 2B). In the antennal lobe, every glomerulus was GFP+, indicating that we indeed targeted all ORN types ( Fig 2B). As predicted, pan-ORN inactivation (Pebbled AND ey > shi ts1 ) abolished aversion to E2-hexenal and acetic acid at the restrictive but not the permissive temperature for shi ts1 (Fig 2C and 2D). For technical reasons, we could not perform the equivalent experiments for attraction, as the Pebbled-GAL4, ey-FLP control flies already showed an attraction deficit (data not shown).
To exclude the possibility that such a loss of aversion was caused by blocking the non-olfactory neurons in Pebbled AND ey > shi ts1 , we rescued aversion by restoring ORN outputs in the "pan-ORN inactivation" background. To do this, we introduced Orco-GAL80 [34] to prevent GAL4 from expressing shi ts1 in OR+ ORNs (Fig 2A). To test the effectiveness of the suppression of Pebbled-GAL4 by Orco-GAL80, we used a GFP tagged with neuronal synaptobrevin (nsyb) [33], which distinguishes glomeruli better than the membrane-tagged CD8::GFP. All anterior glomeruli ( Fig 2E) lost GFP expression in the presence of Orco-GAL80, while most posterior glomeruli maintained GFP expression (Fig 2F, compare to Fig 2E), consistent with the fact that OR+ ORNs predominantly project to anterior glomeruli whereas IR+ ORNs project to posterior glomeruli [8]. Orco-GAL80 restored aversive responses in the background of pan-ORN inactivation (Pebbled AND ey NOT Orco > shi ts1 ) to control levels (Fig 2C and 2D).
Thus, aversion is only abolished with pan-ORN inactivation. Either OR+ ORNs (Fig 1C  and 1D) or OR-ORNs (Fig 2C and 2D) are sufficient to mediate aversion at wild-type levels. This is in stark contrast to attraction, which is sensitive to the inactivation of each subset of ORNs (Fig 1B).

Restoring outputs from specific ORNs rescues aversion
Pan-ORN inactivation offered an anosmic baseline upon which we could test the sufficiency of single ORN types in aversion. To restore synaptic transmission in ORNs expressing specific receptors, we generated GAL80 flies with the enhancers for Or22a, Or42b, and Or85a so that we could compare our results to a previous study [25], and we used Ir64a-GAL80 [26] to represent IR+ ORNs.
As with Orco-GAL80, we first validated the suppression of Pebbled-GAL4 by these GAL80 transgenes. Ir64a-GAL80 and Or22a-GAL80 suppressed Pebbled expression in the DC4+DP1m and DM2 glomeruli respectively (Fig 3A-3D), consistent with previous reports [6,26]. In the other two cases (Fig 3E-3H), Or85a-GAL80 suppressed DM3 and DM4 in addition to the predicted DM5, and Or42b-GAL80 suppressed DL1, DM2, and V in addition to the predicted DM1 [6]. Thus, this strategy afforded the possibility of restoring function to a variety of ORN subsets, ranging from small groups to single types.
We next combined pan-ORN inactivation with GAL80-meditated restoration and examined aversive responses. This scheme only allows synaptic transmission in very few or single types of ORNs (Fig 3I). Despite silencing all but a few ORNs, sparse ORN outputs was sufficient to mediate aversion to various extents (Fig 3J and 3K). In the cases of Or85a-GAL80 in E2-hexenal aversion (Fig 3J, compared to Fig 2C) and Ir64a-GAL80 in acetic acid aversion ( Fig  3K, compared to Fig 2D), PIs were restored to control levels.
Of particular interest was Or42b-GAL80. Since this transgene serendipitously restored activity to ORNs innervating several glomeruli including DM2 (Fig 3H), we compared its PI to that of Or22a-GAL80, which only targets ORN innervating DM2 (Fig 3D). Or22a-and Or42b-restoration did not differ in their aversion to acetic acid (Fig 3K), yet aversion to E2-hexenal was weaker in the latter (Fig 3J). It is possible that slightly different levels of DM2 restoration by Or22a-and Or42b-GAL80s accounts for this phenotypic difference, but it seems more likely that the additionally restored channels in the latter case signal positive valence, or at least counter the DM2-mediated aversion.

Aversion does not require GH146+ excitatory projection neurons (ePNs)
Having tested the effects of manipulating ORN inputs to the antennal lobe on attraction and aversion, we next explored the role of antennal lobe output neurons. ePNs constitute the major output route, over 2/3 of which are labeled by GH146-GAL4 [42]. Intriguingly, echoing the aversion/attraction dichotomy at the beginning of our study, silencing two thirds of the ePNs abolished attraction to vinegar (Fig 4A), but did not affect aversion to E2-hexenal or acetic acid (Fig 4B and 4C).
Given that a subset of ORNs were sufficient to mediate aversion, we reasoned that ePNs that do not express GH146 might account for the robust aversion seen in GH146 > shi ts1 flies. We therefore utilized aversion mediated by a single ORN type as a more stringent test for ePN functions in relaying aversive information. Each ePN receives direct input from a single glomerulus, so we first asked whether blocking output directly downstream to a single restored ORN type affects aversion. To do this, we identified an enhancer trap line [37], 853-GAL4, which only labels ePNs innervating DM2 and DA1, in addition to expression in other brain areas (data not shown). To make this GAL4 line specific to ePNs, we intersected it with GH146-FLP [32] to restrict reporter expression to DM2 and DA1 ePNs ( Fig 4D 1 and 4D 2 ). Surprisingly, combining the inactivation of DM2 and DA1 ePNs with the inactivation of all ORNs except those projecting to DM2 (Fig 4E 1 ) did not have any effect on aversion (Fig 4F and 4G). ORNs. (B 1 , B 2 ) Visualizing the intersection between Pebbled-GAL4 and ey-FLP with UAS-FRT-stop-FRT-CD8::GFP. Single image slices are shown in (B 1 ), and the projection of the whole brain in (B 2 ). (C) Aversion to E2-hexenal is abolished by Pebbled AND ey > shi ts1 at the restrictive (C 1 , n 3) but not the permissive (C 2 , n 4) temperature for shi ts1 , and rescued by Orco-GAL80 (C 1 ). (D) Aversion to acetic acid is also abolished by the same manipulation as in C 1 , and rescued by Orco-GAL80 (n 4). (E, F) Orco-GAL80 suppresses Pebbled-GAL4 expression in a large subset of glomeruli, visualized with UAS-nsyb::GFP.
doi:10.1371/journal.pone.0125986.g002 Each image is representative of 5 brains of the same genotype. (I) A representative scheme of using Or22a-GAL80 to restore inputs to DM2 glomerulus in the pan-ORN inactivation background, same legend as Fig 2A. (J, K) Aversion in Pebbled AND ey > shi ts1 flies to E2-hexenal (J, n 3) and acetic acid (K, n 3) is recued by GAL80s, and Or42b-GAL80 causes less aversion to E2-hexenal than Or22a-GAL80 (J). The stars right below each bar indicate statistical significance as compared to the non-GAL80 control.
doi:10.1371/journal.pone.0125986.g003 Furthermore, aversion generated by restoring DM2 ORNs in pan-ORN inactivation was not affected even when we also blocked all GH146+ ePNs [36], including those innervating DM2 (Fig 4E 2 , 4F, and 4G). Aversion is thus unaffected by GH146+ ePN inactivation, even when the sensory inputs derive from only one type of ORNs. These data suggest that aversion uses antennal lobe output neurons not expressing GH146.
We have compared Or22a-and Or42b-restoration in the pan-ORN inactivation background above (Fig 3J), and concluded that the additional glomeruli rescued by Or42b-GAL80 most likely counter DM2-mediated aversion in Or22a-GAL80. We tested whether such antagonism depends on GH146+ ePNs. If GH146+ ePNs relay positive valence from the additional glomeruli restored with Or42b-GAL80 to higher brain centers, where the negative signal from DM2 gets canceled, blocking GH146+ ePNs would eliminate the difference between Or22a-and Or42b-restored PIs. On the contrary, we observed that, in the GH146 inactivation background, Or42b-GAL80 still restored less aversion than Or22a-GAL80 (Fig 4H). The antagonism thus does not require GH146 + ePNs. Rather, it could take place before information leaves the antennal lobe, mediated by local interneurons; alternatively, the additional glomeruli restored with Or42b-GAL80 could send attractive information through a yet-to-be-identified output channel (see Discussion).

Aversion coding by ORNs
Previous works examined the effects of ORN-specific restoration of Orco in Orco mutants [25,43], where inputs from IR+ ORNs were neglected. Our GAL80 restoration strategy represents the first test of the behavioral role of single ORN types by blocking all the other ORNs. Since we assessed GAL80 efficacy using suppression of GFP expression, one caveat is that we might have overlooked mild suppression by GAL80 in some glomeruli; variable strengths of different OR drivers may also subtly affect the firing of the restored ORNs. However, the predominant effect of our GAL80 strategies should be the restoration of outputs in the annotated ORNs.
With our new results (summarized in Fig 5), we can revisit the hypotheses regarding aversion coding by ORNs (see Introduction). Aversion is not only unaffected by broad ORN inactivation (Fig 1C and 1D), but also requires no more than a single ORN type in the extreme cases (Fig 3J and 3K). Hypothesis I of a "population code" for aversion is thus unlikely to be true.
The aversion restored by Or22a-GAL80 (Fig 3J and 3K) is particularly pertinent to the rejection of "hardwired channels" in Hypothesis II. In a previous report [25], rescuing Orco expression with Or22a-GAL4 has a neutral, if not slightly attractive, effect on olfactory behavior. In a correlative study [22], DM2 is a strong indicator of attraction rather than aversion. Taken together with our results, it thus appears that aversion can be mediated by an ORN type capable of conveying positive, neutral, or negative valence.
Although the valence conveyed by a single ORN type might depend the "context" such as the acitivty in other ORNs, it is worth noting that different ORN types might still preferentially convey aversion or attraction. Or22a-GAL80 restores the activity of DM2 ORNs (Fig 3C  and 3D), while Or42b-GAL80 restores the activity of DM2, DM1, DL1, and V ORNs (Fig 3G  and 3H), and the latter caused less aversion in response to E2-hexenal (Fig 3J). E2-hexenal does not activate DL1 ORNs [5], and V ORNs consist a well-established specialized channel for CO 2 [3]. DM1 ORNs thus most likely accounts for the countering of DM2-mediated aversion. Moreover, in response to acetic acid, which does not activate DM1 [44], Or42b-GAL80 did not cause less aversion than Or22a-GAL80 (Fig 3K). These observations are consistent with the notion that DM1 ORNs are biased towards attraction [25]. We note that such effects are unlikely caused by ephaptic coupling between ORNs [45], because Or22a+ and Or42b + ORNs are housed in different sensilla, and in general shi ts1 only targets the synaptic outputs from ORNs in the antennal lobe without affecting the electric properties of ORN cell bodies in the antenna.
Given the high redundancy of aversion coding by ORNs, and the potential of coding opposite valences by a single ORN type, we favor the third scenario in introduction, where exact activity pattern within individual ORNs is more relevant than the identity or number of activated ORNs. Single types of ORNs have such coding capability, as an ORN responds to different odorants with very different dynamics [46], and it has been demonstrated that mice can use the temporal information from a single glomerulus in discriminative learning [47]. To test specific hypotheses in this class, one could perturb the temporal profile of ORN activity by manipulating receptors or ion channels and observe the change in olfactory behavior, for which our single-ORN-type strategy can be used as a paradigm.
Finally, we caution that more conditions need to be tested before our conclusions can be generalized. For example, we can't exclude the existence of "generalist" ORN channels hardwired for attraction that we have not yet examined through either a broader odorant panel, or further utilization of our GAL80 strategy. Another caveat is that sparse ORN restoration in the background of pan-ORN inactivation might complicate interpretation: in the absence of lateral inhibition from the rest of the ORN population, the restored ORN outputs might achieve an artificially high level never possible in an intact circuit, driving an aberrant form of aversion. Finally, there is also a distinction between "innate" and "hardwired" behavior: if our pan-ORN silencing induced re-organization of the circuit in favor of aversion, we could be observing a behavior considered "innate" in the sense that it requires no training, but not "hardwired" because circuit plasticity induced by our genetic manipulation was required.

Aversive output from antennal lobe
The lack of effect on aversion after silencing GH146+ ePNs (Fig 4B and 4C) parallels that of broad ORN inactivation. The aversion here can be similarly interpreted as conveyed by other PNs not expressing GH146. Granted that typical ePNs dominate antennal lobe output in numbers, recent works suggested other neurons suitable for forwarding olfactory information to higher brain centers [48][49][50][51]. Given the functional studies mostly focused on GH146+ ePNs, more extensive genetic manipulations of antennal lobe output neurons will be necessary to identify alternative output routes and expand our framework of understanding olfactory coding.

The aversion/attraction dichotomy
In contrast to aversion, attraction is very sensitive to broad ORN or ePN inactivation (Fig 5). Furthermore, in our pilot experiments for this study, we could not identify a concentration of acetic acid or E2-hexenal nearly as attractive as vinegar. Both observations appear consistent with a population code for attraction. Intuitively, the identity of the odorant is coded by the activity pattern in all ORNs, and blocking half of the ORN types activated by vinegar does more than merely reducing the valence by half-the resulting input pattern is unlikely to correspond to any attractant at all. By the same token, pure chemicals are unlikely to exactly mimic the activity pattern induced by natural odorant mixtures, and are thus less attractive.
It is worth noting that, in a previous study on the execution of olfactory behaviors [39], aversion and attraction show distinct turning kinematics, and aversion-specific motor-related neurons were identified. The sensory and motor dichotomies thus suggest that, despite the apparent symmetry between aversion and attraction and their comment origin in ORN activation, they are likely relayed or deciphered by distinct downstream circuits, and processed with different logic. Future models should treat aversion and attraction as such, and should account for the observation that extremely sparse ORN inputs are sufficient for aversion to general repellents.