Most primary olfactory neurons have individually neutral effects on behavior

  1. Tayfun Tumkaya
  2. Safwan Burhanudin
  3. Asghar Khalilnezhad
  4. James Stewart
  5. Hyungwon Choi
  6. Adam Claridge-Chang  Is a corresponding author
  1. Institute for Molecular and Cell Biology, A*STAR, Singapore
  2. Program in Neuroscience and Behavioral Disorders, Duke NUS Graduate Medical School, Singapore
  3. Department of Medicine, National University of Singapore, Singapore
  4. Department of Physiology, National University of Singapore, Singapore

Abstract

Animals use olfactory receptors to navigate mates, food, and danger. However, for complex olfactory systems, it is unknown what proportion of primary olfactory sensory neurons can individually drive avoidance or attraction. Similarly, the rules that govern behavioral responses to receptor combinations are unclear. We used optogenetic analysis in Drosophila to map the behavior elicited by olfactory-receptor neuron (ORN) classes: just one-fifth of ORN-types drove either avoidance or attraction. Although wind and hunger are closely linked to olfaction, neither had much effect on single-class responses. Several pooling rules have been invoked to explain how ORN types combine their behavioral influences; we activated two-way combinations and compared patterns of single- and double-ORN responses: these comparisons were inconsistent with simple pooling. We infer that the majority of primary olfactory sensory neurons have neutral behavioral effects individually, but participate in broad, odor-elicited ensembles with potent behavioral effects arising from complex interactions.

Editor's evaluation

Olfactory coding is still an open question in neuroscience. Therefore, this paper is of potential interest to a broad audience of neuroscientists. It undertakes a thorough investigation of how olfactory sensory neurons drive avoidance or attraction in flies and also addresses how combinations of active ORNs can become behaviorally meaningful. It has great potential value for clarifying how animals map sensory input to valence.

https://doi.org/10.7554/eLife.71238.sa0

Introduction

Animals interact with their environment using motor functions that are guided by information that enters the brain from multiple sensory systems. These diverse sensory inputs are thought to interact with each other, with previously stored information, and with the internal physiological state of the animal to elicit a more or less appropriate behavioral response. Two central problems of neuroscientific research are (1) how individual sensations influence behavior and (2) how multiple streams of sensory information are reconciled into meaningful behavior. The Drosophila olfactory system is an effective model to address these critical questions (Couto et al., 2005; Eisthen, 2002; Wang et al., 2003), facilitated by powerful genetic approaches, the ability to handle large sample sizes, and the numerical simplicity of Drosophila neural systems. Flies detect odors with their antennae and maxillary palps, which together contain ~1300 olfactory-receptor neurons (ORNs) (Lai et al., 2008). The odor-response profile of each adult ORN is determined by one of ~45 possible receptor types (Fishilevich and Vosshall, 2005; Gomez-Diaz et al., 2018). ORNs sharing the same receptor type converge on a glomerulus in the antennal lobe, where they synapse with local interneurons (LNs) and projection neurons (PNs) (Couto et al., 2005; Gao et al., 2000). Innervating throughout the antennal lobe and connecting multiple glomeruli, the LNs facilitate both excitatory and inhibitory interactions between glomeruli (Groschner and Miesenböck, 2019). This modified information is relayed by the PNs to higher brain centers, namely mushroom bodies and the lateral horn (Lai et al., 2008; Wang et al., 2014; Wong et al., 2002). The distinct nature of the ORN types allows us to consider each type as a single channel of information input. Mapping how ORNs steer behavior would inform a broader understanding of how sensory circuits influence behavioral output.

Odor-induced activity in ORNs can trigger approach and avoidance behaviors, collectively referred to as behavioral ‘valence’ (Knaden et al., 2012). At least some ORN-driven behaviors appear to follow simple rules: a subset of receptors respond specifically to single odorants, and their ORNs individually drive innate valence (Ache and Young, 2005; Grabe and Sachse, 2018; Haddad et al., 2010; Haverkamp et al., 2018; Stensmyr et al., 2012; Suh et al., 2007). These acutely tuned, strongly valent ORN classes include neurons tuned to danger (e.g. toxic odorants) and pheromones. Given the direct relationship between such odors and valence, these ORN types and their associated downstream pathways have been termed ‘labeled lines’ (Grabe and Sachse, 2018; Hildebrand and Shepherd, 1997; Kurtovic et al., 2007). The existence of labeled lines proves that at least some olfactory behaviors follow simple ORN-activity rules.

Unlike labeled lines, many other olfactory receptors are broadly tuned to respond to many odorants, and most pure odorants evoke responses across many ORN classes (Hallem and Carlson, 2006). As most olfactory behavior relies on activity in ORN groups, there is the outstanding question of how individual channels contribute to an odor’s overall valence. It is not known how much more complex multi-ORN valence is compared to the relative simplicity of labeled-line behavior. Earlier studies have looked at whether multi-glomerular olfactory valence could be explained by statistical models of ORN or PN activity patterns. Depending on the type of experiments, some found that valence could be explained by simple rules, for example weighted summation of larval ORN activity (Kreher et al., 2008). Other studies found no relationship between single-glomerulus properties and odor-evoked behavior, or invoked more complex models of antennal-lobe function (Badel et al., 2016; Knaden et al., 2012; Kuebler et al., 2012; Meyer and Galizia, 2011). Due to the many–many relationship of most odorants and ORNs, using natural odors to isolate single-ORN valence effects is challenging (Haddad et al., 2010; Knaden et al., 2012; Semmelhack and Wang, 2009; Thoma et al., 2014; Turner and Ray, 2009). One study overcame this challenge by activating single-ORN types optogenetically (Bell and Wilson, 2016). Using eight attractive ORN types the researchers found that two-way ORN valence combinations follow either summation or max-pooling; this supports the idea that olfactory valence arises from simple rules. Thus, both simple mechanisms (labeled lines, summation) and complex inter-channel interactions have been invoked to explain olfactory valence, but their relative importance remains controversial.

The present study had two primary aims: to map which single ORN types drive valence; and to examine the extent to which simple pooling rules govern ORN–valence combinations. To do so, we measured the valence coding of the primary olfactory system by optogenetically stimulating single ORN classes. In the wild, olfaction typically occurs in windy environments, and is influenced by hunger state, so we explored whether single-type-ORN valence is similarly contingent on these factors (Bell and Wilson, 2016; Sengupta, 2013). We activated pairs of ORN types to investigate how their combinations influence valence, and built statistical models of ORN interactions. All the results indicate that ORN–valence computations are complex.

Results

An optogenetic behavior assay reports on sensory valence

We investigated which single classes of primary chemosensory neurons can elicit preference behavior, and the capability of the present experimental system to replicate prior studies. Generating Drosophila lines that express the red-light-shifted channelrhodopsin CsChrimson (Chr) in different receptor neurons, we tested whether individual neuronal types drive attraction or avoidance. Flies were presented with a choice between light and dark environments in a wind- and light-induced self-administration response (WALISAR) apparatus (Figure 1A). We tested the validity of this approach with flies expressing the Chr channel under the control of the Gr66a-Gal4 driver line, which labels bitter-taste-sensing gustatory-receptor neurons (Moon et al., 2006) and has been previously reported to drive avoidance when artificially activated (Aso et al., 2014; Shao et al., 2017). In a sequence of 12 trials using three light intensities (14, 42, and 70 μW/mm2) and two airflow states (on/off), the experimental Gr66a > Chr flies displayed robust light avoidance (Figure 1B–F; Figure 1—figure supplement 1B–1D; Figure 1—figure supplement 2). To benchmark the assay against a published method, we compared present data with prior results (Shao et al., 2017); the standardized effect size of Gr66a-neuron avoidance at the highest intensity (Cohen’s d = –2.70 at 70 μW/mm2) was almost as large as that of the previously reported valence response (d = –3.63; Figure 1—figure supplement 3A). This replication of aversive Gr66a-cell activation confirmed that this study’s optogenetic-choice apparatus could be used to measure negative valence mediated by sensory neurons.

Figure 1 with 3 supplements see all
An optogenetic behavior assay reports on sensory valence.

(A) Side view schematic (top) showing air flow path and overview (bottom) of the WALISAR assay. Individual flies were placed in chambers and given a choice between no light or red-light illumination. Air flow was used in some experiments. (B) A representative path of a fly displaying strong light avoidance. The white line represents the location of a Gr66a-Gal4> UAS-CsChrimson fly in the chamber throughout an experiment. The light preference of each fly was calculated by how much time it spent in the illuminated zones after the initial encounter with light (yellow arrows). (C-E) Trace plots representing the mean location of controls (w1118; Gr66a-Gal4 and w1118; UAS-CsChrimson), and test flies (Gr66a-Gal4 > UAS-CsChrimson) throughout an experiment at 14, 42, and 70 μW/mm2 light intensities. The blue and orange ribbons indicate 95% CIs for the control and test flies, respectively. (F) An estimation plot presents the individual preference (upper axes) and valence (lower) of flies with activatable Gr66a+ neurons in the WALISAR assay. In the upper panel, each dot indicates a preference score (wTSALE) for an individual fly: w1118; Gr66a-Gal4 and w1118; UAS-CsChrimson flies are colored blue; and Gr66a-Gal4 > UAS-CsChrimson flies are in orange. The mean and 95% CIs associated with each group are shown by the adjacent broken line. In the bottom panel, the black dots indicate the mean difference (ΔwTSALE) between the relevant two groups: the valence effect size. The black whiskers span the 95% CIs, and the orange curve represents the distribution of the mean difference.

To benchmark attraction behavior, we used an Orco-Gal4 driver line that labels ~70% of all ORNs (Larsson et al., 2004; Wang et al., 2003). Others have reported Orco-neuron valence results: one study showed no behavioral effect Suh et al., 2007; another found attraction in the presence of a wind cue only Bell and Wilson, 2016. In our experiments, at two higher light intensities (42 and 70 μW/mm2), Orco > Chr flies exhibited pronounced attraction even in still air (Figure 2G; Figure 2—figure supplement 1). The valence was typically stronger than that reported in prior studies (e.g. d = 0.56 in still air in this study versus d ≤ 0.10), establishing assay sensitivity for attraction (Figure 1—figure supplement 3B). Changing the temporal sequence of the 12 trials had negligible effects on Orco-neuron positive valence, suggesting that valence is not greatly susceptible to order bias, for example, due to habituation (Figure 2—figure supplement 1). Driver and responder controls typically had very similar distributions (Figure 2—figure supplement 2). Together with the Gr66a + results, these data indicate that WALISAR is a valid, sensitive assay for measuring the valence of chemosensory circuits.

Figure 2 with 2 supplements see all
A small minority of ORN classes individually drive valence.

(A) An effect-size plot of the valence screen of 45 single-ORN types (and Orco neurons). Each dot represents a mean wTSALE difference of control (N≅ 104) and test (N≅ 52) flies, whisker indicate 95% CIs. The shades of red represent the three light intensities. Valent ORNs are shaded with magenta (aversion) or green (attraction). (B) The histogram (gray bars) of the median ORN ΔwTSALE ratios, and the mixture model that is fit to the data by empirical Bayes method. The blue line represents the null distribution, while the magenta and green curves represent ORNs with negative and positive valences, respectively. The black line represents the overall distribution of the effect sizes. (C) The signed posterior probability that each valence score is a true behavioral change, plotted against the respective effect size (median effect across the three light intensities). (D-E) The olfactory valence of the Orco and four single-ORNs were tested in female flies (the male fly data is replotted from Panel A for comparison). The valence responses are represented as the mean difference (ΔwTSALE) of control (N≅ 104) and test (N≅ 52) flies, along with 95% CIs. Color key is the same as above. (F-G) Estimation plots show the optogenetic preference (upper panel) and valence (lower panel) of flies with activatable Or67d and Orco neurons across three light intensities. The dots in the top panels represent single flies, while the broken lines indicate the mean and 95% CIs. The differences between the pairs of test and control groups are displayed in the bottom panel, where the whiskers are 95% CIs and the curves are the distributions of the mean difference.

Only one-fifth of ORN types drive valence

We aimed to estimate what proportion of single ORN types elicit valence when activated alone. Using available Gal4 lines, each driving expression in a single ORN type, we assessed the optogenetic valence of 46 receptor classes (Figure 2A). To separate the valence effects from noise, we analyzed data from ~7176 flies with the empirical Bayes method (Figure 2B–C). The empirical Bayes analysis identified 10 valent classes: six ORNs elicited attraction and four elicited aversion (Figure 2A). The hits included six ORN classes with identified ligands. Four are considered labeled lines: Or56a, the receptor for the aversive odorant geosmin; Gr21a, the receptor for the aversive odorant carbon dioxide; Or67d, the receptor for the pheromone 11-cis-vaccenyl acetate; and Or47b, which senses the pheromone palmitoleic acid (Davis, 2007; Jones et al., 2007; Kurtovic et al., 2007; Lin et al., 2016; Stensmyr et al., 2012; Suh et al., 2004; van der Goes van Naters and Carlson, 2007). Additionally, Or83c mediates attraction to farnesol, an odorant produced by some ripe fruits (Ronderos et al., 2014), while Or42b mediates attraction to vinegar (Semmelhack and Wang, 2009). Given that this screen successfully recaptured six ORN types already known to be involved in ecologically relevant valence functions, we consider that the screen was valid and sensitive. Furthermore, the majority of the hits being ORNs with already-established valence implies that most ORNs are not singly valent.

To contextualize the screen’s outcome, we conducted a literature review, tabulating prior and current results in ORN valence (Supplementary file 1), from 16 studies including this one (Bell and Wilson, 2016; Chin et al., 2018; Dweck et al., 2013; Faucher et al., 2006; Gao et al., 2015; Hernandez-Nunez et al., 2015; Jung et al., 2015; Knaden et al., 2012; Mathew et al., 2013; Poon et al., 2010; Ronderos et al., 2014; Semmelhack and Wang, 2009; Stensmyr et al., 2012; Suh et al., 2007; Suh et al., 2004). Although methodological diversity precludes a formal, quantitative meta-analysis (Borenstein et al., 2009; Tumkaya et al., 2018), it is clear that—for many ORNs—a consensus on valence is lacking. For example, the ORN screen showed that two additional, pheromone-responding ORNs (Or88a and Or65a) were not Empirical-Bayes hits (Chin et al., 2018; van der Goes van Naters and Carlson, 2007); however, no prior single-ORN data have shown these ORN types to individually drive valence (Supplementary file 1). It is possible that the behavioral effects of Or88a and Or65a depend on the presence of other cues or the activation of other receptors.

We only used male flies in the screen. Because odor responses in female flies might differ—especially for pheromonal receptors—we checked for possible sex differences in five receptor classes: Orco cells and four pheromone-responsive ORNs (Figure 2D–E). Although male flies exhibited strong responses to Orco-neuron activation, females showed no response. Surprisingly, this lack of response in females turned into a strong attraction when they were starved (Figure 2—figure supplement 2). Only one pheromone-receptor class showed sexual dimorphism: activation of Or47b neurons (sensors of an aphrodisiac pheromone) was attractive to males, while females were indifferent.

Together, these results indicate that in isolation, most ORN classes do not drive valence. The presence of six known valent ORN types in the 10 hits, and the predominance of neutral ORNs suggest that most olfactory channels influence behavior only when activated in concert as part of an odor-evoked ensemble. It should be noted, however, that this initial analysis was performed on flies measured in still air—an abnormal condition in the wild.

Wind does not amplify single-ORN valences

It has been previously reported that wind is essential for the optogenetic valence of Orco neurons (Bell and Wilson, 2016). We thus aimed to test the hypothesis that wind amplifies ORN valence, possibly eliciting valence in some otherwise non-valent ORN classes. We tested this in the same flies by also measuring the valences of 46 ORN types in the presence of airflow (Figure 3A). From each fly, we used the paired wind–no-wind responses to calculate wind-specific effect sizes, ∆∆, for each light intensity and each ORN type (Figure 3B). With a lone exception (wind rendered Or59b valence more aversive in the lowest light intensity only), an empirical-Bayes model found that the wind effect sizes were indistinguishable from noise (Figure 3C–D). Contradicting our hypothesis, this result indicates that wind has essentially no impact on single ORN type-elicited behavior in walking flies. This result also generalizes the finding that, in either still or windy conditions, only a minority of ORN classes individually drive valence.

Wind does not amplify single-ORN valences.

(A) The results of ORN valence assays in the presence of airflow. The red dots represent the mean wTSALE differences of control (N ≅ 104) and test (N ≅ 52) flies with the 95% CIs. (B) The differences between the effect sizes of air-on and air-off experiments (ΔΔwTSALE). The magenta shaded box indicates the sole empirical Bayes hit, Or59b, that showed an increase in aversion at the lowest light intensity only. (C) A statistical mixture model was fitted to the ΔΔwTSALE scores. The grey bars are the response histogram. The magenta and green curves represent the effect sizes that differ from the null distribution (blue line). The black line represents the overall distribution of the ΔΔwTSALE scores. (D) The signed posterior probabilities of the ΔΔwTSALE scores being true behavioral changes versus their median effect sizes. The overall probability of true wind effects was nearly zero [p(∆∆wTSALE) ≅ 0.0].

Hunger has a limited effect on single-ORN valence

Chemosensory behaviors—for example gustatory and olfactory responses—are influenced by an animal’s internal energy state. Low internal energy, for example, can sensitize food odor-responsive ORNs and drive foraging (LeDue et al., 2016; Sengupta, 2013; Zhou et al., 2010). We hypothesized that starvation would thus increase the single-ORN valence behavior, especially for attractive ORN types. To test this hypothesis, we assayed ~7176 starved flies for their optogenetic ORN valence (Figure 4A), and compared their behavior with that of the fed flies described above. Surprisingly, starvation did not enhance attraction. On the contrary, it reduced the responses triggered by three otherwise attractive ORNs: Or42b, Or47a, and Or83c (Figure 4B). Starvation also shifted the otherwise neutral valences of Or85f and Or49a, two receptors involved in sensing wasp odors (Ebrahim et al., 2015), into aversion (Figure 4B). Thus, hunger reduced the positive valence of a few pheromone- and food-sensing-ORNs, while slightly increasing the aversiveness of predator-sensing ORN classes. Overall, hunger does not have a broadly amplifying effect on single-ORN valence.

Starvation affects valence responses for five ORN classes.

(A) An ORN valence screen for starved animals. The red dots indicate the mean wTSALE differences between control (N ≅ 104) and test (N ≅ 52) flies. (B) The mean differences between fed and starved flies across ORNs. The magenta shading indicates the ORNs that are affected by starvation according to the Empirical-Bayes analysis. (C) A histogram of the median ΔΔwTSALE ratio distribution. The blue line represents the null distribution, the magenta and green curves represent the distributions of negative and positive valences that are separated from the null distribution, and the black line represents the overall ΔΔwTSALE distribution. (D) The signed posterior probability of the ΔΔwTSALE scores being true behavioral changes are plotted against their median ratios.

ORN-valence combinations follow complex rules

Because the ORN screens showed that most single ORN types do not drive valence individually, it would appear that activity across multiple sensory channels simultaneously is required to drive most (non-labeled-line) olfactory behavior. One hypothesis of combinatorial odor valence holds that ORN-combination behaviors arise from simple two-way pooling rules: summation and max-pooling (both widely used in neural-net construction) (Bell and Wilson, 2016; Goodfellow et al., 2016).

To address this hypothesis, we asked how single-ORN valences are combined when two ORN classes are activated. We crossed eight ORN driver lines (three positive, two negative, and three neutral) so as to generate seven, two-way ORN combinations (henceforth ‘ORN-combos’). Compared to their constituent single ORNs, the ORN-combos elicited distinct valences (Figure 5A, B and F). We modeled the combination effect sizes with three pooling functions: summation, max-pooling, and min-pooling (Figure 5C–E). Strikingly, regression showed that none of the three functions could account for a large proportion of ORN-combo valence: summation, min-pooling, and max-pooling all had coefficients of determination of ~0.2 or lower (Figure 5—figure supplement 1A-C). Furthermore, Bland-Altman method-comparison plots revealed wide limits of agreement (LoAs) between the observed and predicted ORN-combo valences by all three models: summation [SD1.96–0.45, 0.28], max-pooling [SD1.96–0.34, 0.13], min-pooling [SD1.96–0.1, 0.35] (Figure 5—figure supplement 1D, E, F). This analysis thus demonstrates that none of these three simple pooling rules are major predictors of how two-ORN odor valence emerges from single-ORN valence.

Figure 5 with 4 supplements see all
ORN-valence combinations follow complex rules.

(A) Valence responses of the single-ORN lines used to generate ORN-combos (replotted from Figure 1). The dots represent the mean valence between control (N≅ 104) and test (N ≅ 52) flies (∆wTSALE with 95% CIs). The shades of red signify the three light intensities. (B) The valence responses produced by the ORN-combos in the WALISAR assay in three light intensities. (C–E) ORN-combo valences as predicted by the summation (C), max pooling (D) and min pooling (E) models. (F) Three positive (green), two negative (magenta), and three neutral (gray) ORNs were used to generate seven ORN two-way combinations. (G–I) Scatter plots representing the influence of individual ORNs on the respective ORN-combo valence. The red (G), maroon (H), and black (I) dots indicate ORN-combos at 14, 42, and 70 μW/mm2 light intensities, respectively. The horizontal (β1) and vertical (β2) axes show the median weights of ORN components in the resulting combination valence. (J) Euclidean distances of the ORN-combo β points from the diagonal (summation) line in panel J. The average distance increases as the light stimulus intensifies: 0.14 [95CI 0.06, 0.23], 0.20 [95CI 0.06, 0.34], and 0.37 [95CI 0.20, 0.53], respectively. (K) The β weights of the ORN-combos from the multiple linear regression are drawn as the signed distances of each ORN-combo from the diagonal line over three light intensities. The ORN weightings change magnitude and, in a few cases, the dominant partner changes with increasing optogenetic stimulus.

To generalize this analysis, we built multiple-linear regression models of ORN-combo associations (Figure 5—figure supplement 2). In these models, as the effect sizes are standardized on both individual ORNs and combos, the estimated β weights indicate the relative contribution strength of each of the two ORN classes. We drew scatterplots of the medians of β values for the three light intensities (Figure 5GHI). If the combination valences arose from summation, we would expect the β points to cluster along the diagonal (equal contribution); if combination valences followed max- or min-pooling, we would expect points clustering along the axes. However, the β points were dispersed, indicating a diversity of pooling rules. Moreover, as the light intensity increased, the β points shifted further away from the diagonal (Figure 5J) and, in some cases, flipped dominance (Figure 5K). For these ORN-type pairs, increasing intensity is associated with two phenomena: one of the two ORNs tends to become more dominant; and the combination rules are different at different activity levels. Thus, the interactions between ORN pairs vary depending on receptor identity and stimulus intensity.

Single ORN information predicts odor behavior poorly

The variable interactions between pairs of ORN types support the idea that olfactory valence is determined by complex dynamics in multiple layers of downstream circuits. So we anticipated that models using ORN activity would not be strongly predictive of odor valence. To explore this, we drew on data from two previous studies that used a panel of 110 odorants: one used the odor panel to make physiological recordings from 24 ORN types—of which 23 were tested for behavioral valence in this study (Hallem and Carlson, 2006); the other study measured behavioral valence for all 110 odors (Knaden et al., 2012). We adopted a partial-least-squares discriminant analysis (PLS-DA), to reduce the 23-dimensional feature space into fewer latent variables (LVs) with internal correlation (Figure 5—figure supplement 3A). Various models based on various numbers of LVs could partially predict odor preference. A multiple linear regression (MLR) model with eight LVs had the best performance, with an adjusted coefficient of determination (R2adj) of 0.23. While the two-dimensional space defined by the first two LVs supports a partial separation of aversive and attractive odors (Figure 5—figure supplement 3B), the low R2adj score indicates that linear combinations of the available ORN-activity patterns are only weakly predictive of valence. The poor predictiveness of both the MLR model and a support-vector regression model were further confirmed by cross-validation (Supplementary file 2 - Sheet 1). For non-linear models, performance was limited by the relatively small sample size: two models with non-linear kernels (polynomial and radial basis function) had high error rates and learning curves indicating overfit (Supplementary file 2 - Sheet 1, Figure 5—figure supplement 4). Lastly, we asked whether the models could be improved with incorporation of the optogenetic valence data. Although valence-weighting improved the R2adj of the MLR model from 0.23 to 0.33, the predictive performances of the models were not improved (Supplementary file 2 - Sheet 2). This further supports the hypothesis that linear combinations of ORN activity can only account for a minority of odor preference.

Discussion

Most ORN classes have no individual effect on valence

This study’s primary goal was to identify the proportion of ORNs that can drive valence behavior individually. The screen results indicate that 10/45 of single ORN-types have the ability to influence locomotor preference. Thus, most ORN types are not individually valent. Of the 10 valent ORN types, 6 have previously known roles in valence: two sensors each for pheromones, chemical-threat odorants, and food-aroma molecules. So the past two decades of Drosophila olfaction research have already identified a large proportion of strongly valent ORN types. The screen identified four novel valent ORNs (85d, 59 c, 35 a, 47 a) with ligands and ecological roles yet to be determined. Around 60% of pure odorants are valent (Knaden et al., 2012), while at least three of the known valent ORN classes (those for geosmin, CO2, and cVA) are specialized to bind specific ligands exclusively. This disconnect between a preponderance of valent odorants (~60% of odorants) and the scarcity of broadly tuned, valent ORN types ( ≤ 15%) implies that most olfactory valence arises from combined activity in multiple ORN classes (Mathew et al., 2013; Parnas et al., 2013). This idea is also supported by the Orco activation result showing that broad activation across ~70% of ORNs drives strong attraction.

Wind and hunger effects on ORN valence are minor

Another goal of this study was to ask whether two contextual factors—wind and hunger—would increase olfactory valence. Prior studies observed that the Orco neurons are either not optogenetically valent or that airflow is essential for valence (Bell and Wilson, 2016; Suh et al., 2007). However, our experiments showed that olfactory neurons elicit valence without airflow, and that wind has little to no amplifying effect (Figures 1C and 2). Note also that even though Drosophila larvae behavioral tests are routinely performed in still air, they display both olfactory behavior and optogenetic ORN valence (Bellmann et al., 2010; Hernandez-Nunez et al., 2015; Mathew et al., 2013). Thus, the overall evidence indicates that Drosophila ORNs influence valence behavior even without wind. A second factor, hunger, has also been reported to increase olfactory attraction (Gelperin, 1971; Ko et al., 2015). However, in the single-ORN screen, starvation had modest effects on valence (Figure 4). If anything, valence was somewhat lower for several attractive ORNs (Figure 4B). There are at least two plausible explanations as to why starvation did not affect many single-ORN valences. First, the difference in olfactory responses between fed and starved animals might only be pronounced for weaker olfactory stimuli: one study showed that the ethyl acetate (EA) response difference between fed and starved flies declines as the EA concentration increases (Chakraborty, 2010). So optogenetic activation might be too strong to observe the starvation effect. Second, ORNs likely have less influence in isolation. Co-activated glomeruli modulate each other via lateral inhibition and excitation (Groschner and Miesenböck, 2019; Huang et al., 2010; Shang et al., 2007; Wilson, 2008). As most odors activate multiple ORN types, any hunger effect might require these lateral signals. One study found that five vinegar-responsive ORNs are modulated by hunger only when activated in concert, but not when they are activated in subsets (Root et al., 2011). In our results, a notable exception was the hunger switch of Orco valence in females (Figure 2—figure supplement 2). The absence of hunger amplification in single-ORN valence—and its presence in female Orco valence—also suggests that the potentiating effects of hunger on olfactory attraction might only operate on multi-ORN stimuli like Orco activation and natural odors. That neither wind nor hunger increased valence of single-ORNs verifies that the majority of single receptor types, on their own, do not convey valence information. Together, the single-ORN results support the idea that most odor-guided locomotion arises from the broad activation of ORNs simultaneously.

How is ORN information combined?

Many individual odorants bind multiple olfactory receptors, and most natural odors are complex blends that activate receptors broadly, such that odors typically activate multiple ORN classes. A number of groups have constructed statistical models of the relationship between initial layers of olfactory systems and their eventual valent locomotion (Badel et al., 2016; Bell and Wilson, 2016; Kreher et al., 2008; Kuebler et al., 2012; Kundu et al., 2016; Meyer and Galizia, 2011; Mohamed et al., 2019; Riffell et al., 2009; Thoma et al., 2014). The existence of labeled lines indicates that, at least for some odorants and their receptors, ORNs can have a deterministic effect on valence. In a model of Drosophila larval olfaction, the weighted summation of ORN activity in just five of 21 receptors could be used to predict odor valence (Kreher et al., 2008). In adults, an optogenetic study of ORNs reported that combination valences could be explained by summation and/or max-pooling (Bell and Wilson, 2016), further supporting a direct relationship. Summation and max-pooling correspond to special cases of two-way combination weightings of (0.5, 0.5) and (0, 1), respectively; the present analysis found little support for such equal or all-or-none weightings. Rather, the ORN combinations have diverse, intermediate valence weights within the weight space. Moreover, these weights shift as the stimulus intensity increases, and appear to even swap dominancy at different intensities. These features suggest complex ORN-channel interactions.

Possible models of ORN combination

The variability of β values across a range of stimuli indicates that (1) single-ORNs can contribute a dynamic range of weights when co-activated, (2) the experiment captured particular instances of ORN activation from these many possible neural response mechanisms, and (3) that many ORN combinations likely share downstream targets. Therefore, we can cautiously infer that the underlying neural network is multifactorial (Qiu et al., 2021) and perhaps context-specific. Several kinds of complex architectures and dynamics could contribute to multi-ORN valence. The increasing dominance of one ORN observed in some of the pairs suggests competitive, antagonistic interactions between channels, such as those mediated by broad lateral synaptic inhibition (Olsen and Wilson, 2008) or ephaptic inhibition between neighboring ORNs (Su et al., 2012). Alternatively, motor programs associated with valence might only require a subset of channel combinations to be active, with the remainder playing ancillary roles. However, reverse-engineering the exact combination mechanisms through estimated β values would require a larger set of experiments with distinct stimuli. Translating the patterns of β shifts into the specific structure of the neural network carries the risk of over-interpreting the current data set.

Technical differences between studies

It is relevant to note that several of our conclusions on olfactory valence, notably the nature of ORN-valence combination, diverge from those made in an earlier study (Bell and Wilson, 2016). Along with sampling error, it is possible that these discrepancies can be attributed to differences in experimental design and analysis, some of which are summarized in Supplementary file 2 - Sheet 3. Here, we discuss the different ways the optogenetic valence effect was controlled. As Chrimson was not available at the time, the earlier study used Channelrhodopsin-2 (ChR2), which requires intense blue light (up to 1500 µW/mm2) that adds significant heat and elicits strong responses in the fly visual system (Supplementary file 2 - Sheet 3). As these stimuli can profoundly alter behavior, especially during a lengthy illumination regime (64 min total), the researchers used genetically blind flies and an infrared laser for compensatory heat. While these technical measures were prudent, they were used largely in place of genetic controls: relative to the experimental animals (N = ~ 2512), the study used few responder UAS controls (N = 88) and no driver Gal4 controls (Supplementary file 2 - Sheet 3). The earlier study appears to have averaged technical replicates to represent behavioral variation, a procedure that under-reports variation (Bell, 2016). The present study dealt with these issues with two changes. First, it was able to make use of the Chrimson channel, which requires lower light intensity and a red wavelength to which the fly eye is less receptive. Second, all non-optogenetic effects were accounted for with balanced experimental, driver, and responder groups (N = ~ 5148 in each group). In our opinion, our approach of testing all three groups in sufficient sample sizes—and using them to calculate effect sizes—enables the exclusion of all confounding influences, including heat, visual effects, and genetic background.

Limitations of the study

This study incorporates at least four assumptions that potentially limit the generalizability of the findings. First, the use of narrow chambers effectively canalizes fly locomotion into one dimension (Bell and Wilson, 2016; Claridge-Chang et al., 2009). While this simplifies analysis, it could either alter valence, and/or increase locomotion relative to either broader chambers or behavior in the wild. Second, although the olfactory receptors have been shown to express in specific ORN types (Couto et al., 2005; Fishilevich and Vosshall, 2005), we cannot rule out that they might express elsewhere in the fly nervous system, so some of the valence results could be the result of activity in several neuron types. Third, optogenetic light is not equivalent to odor stimulation: there may be differences in firing rate, response dynamics, or other features of ORN stimulation that mean these data are not directly generalizable to odor-elicited activity. Fourth, no electrophysiological recordings were made from the optogenetically activated ORNs, so their firing rates relative to odor-elicited activity are unknown; published recordings from optogenetic and odor stimuli suggest an approximately five-fold difference in firing rates in some cases (Bell and Wilson, 2016; Hallem and Carlson, 2006).

Do single-ORN-class properties govern valence?

While larval valence has been predicted with a summation model, similar models for olfactory behavior in other systems have not been successful. Studies in various model animals have invoked complex computations to explain olfactory valence (Duchamp-Viret et al., 2003; Kuebler et al., 2012; Kundu et al., 2016; Meyer and Galizia, 2011; Riffell et al., 2009; Shen et al., 2013; Silbering and Galizia, 2007). In adult Drosophila, an analysis of the physiological and behavioral responses to 110 odors found that ORN activities and odor valence have no linear correlation (Knaden et al., 2012), suggesting that valence determination could arise from a downstream computation, for example in the antennal lobe, where incoming ORN activity patterns and outgoing projection-neuron activity patterns are dissimilar (Groschner and Miesenböck, 2019). Indeed, the physiological-activity patterns in projection neurons have been reported to be at least partially predictive of odor valence (Badel et al., 2016; Knaden et al., 2012; Parnas et al., 2013), leading to the idea that the antennal lobe extracts valence features. To examine this, we modeled published ORN activity and valence data sets, finding that linear models of ORN activity data could account for a minor fraction of olfactory-behavior variance, and this was not substantially improved with single-class ORN valence information. All the observations in the present study point to a minimal role for simple pooling rules, and support the idea that odor valence is primarily governed by complex circuit dynamics.

Materials and methods

Drosophila strains

Request a detailed protocol

Flies were raised on fly medium (Temasek Life Sciences Laboratories, 2018) at 25 °C in a 12 hr light:12 hr dark cycle. For optogenetic experiments, the flies were kept in the dark and reared on fly food supplemented with 0.5 mM all-trans-retinal (Sigma-Aldrich, USA) for 2 days prior to experimentation. For starvation experiments, the flies were reared on 2% agarose for 12–18 hr prior to the assay. Wild type flies were cantonized w1118; all the ORx-Gal4 and UAS-CsChrimson strains were obtained from the Bloomington Drosophila Stock Center (USA) (Supplementary file 2 - Sheet 4) (Couto et al., 2005; Dobritsa et al., 2003; Klapoetke et al., 2014; Vosshall et al., 2000).

Optogenetic preference assay

Request a detailed protocol

The wind- and light-induced self-administration response (WALISAR) assay was conducted as follows. Two rectangular assemblies (11.5 × 14.5 × 0.3 cm) were cut from acrylic sheets; each assembly contained 26 chambers (50 × 4 × 3 mm), herein referred to as WALISAR chambers. Airflow inlets and outlets were milled into the ends of each chamber. Optogenetic illumination was achieved using LEDs [LUXEON Rebel LEDs on a SinkPAD-II 10 mm Square Base; red (617 nm), green (530 nm), blue (470 nm), each equipped with lenses (17.7° 10 mm Circular Beam Optic)] and attached to heatsinks located above the arena on both sides at a ~ 45° angle. The LEDs were grouped by color and were powered by 700 mA BuckPuck drivers. Custom instrumentation software (CRITTA) was used to control the intensity and timing of the LEDs throughout the experiments. To achieve a half-dark/half-lit arena for the optogenetic choice experiments, two black acrylic shields were placed between the arena and the LEDs on each side and were adjusted to cast shade on either half of the chamber. By switching the LEDs on either side, the half of the arena that was lit could be alternated. The temperature difference between the dark and lit halves of the WALISAR chambers that could arise from the LED illumination was measured for the duration of a whole experiment using thermocouples, and found to be a negligible ~0.3 °C (https://doi.org/10.5281/zenodo.4545940). Compressed air was connected to the airflow inlets of the arena, with an intervening stopcock valve (Cole-Parmer) to modulate the flow rate. The air flow in all experiments using wind was 35 cm/s, as described in a previous study (Bell and Wilson, 2016). The airflow was measured before every experiment with an airflow meter (Cole-Parmer).

Flies were collected 2–3 days before experiments and cold anesthesia was administered immediately prior to their transfer into the WALISAR chambers. Unless otherwise stated, each experiment typically used sample sizes of N = ~ 104 Gal4 and UAS controls and N = ~ 52 test flies. Transfer to individual chambers took around 5 min. A single experimental cycle consisted of: acclimatization of the flies for 30 s; illumination of the left half of the arena for 45 s; no illumination for 30 s; illumination of the right half of the arena for 45 s; and no illumination for 30 s. The chambers were manually tapped against the incubator base before each cycle. Each group of flies was tested in six conditions comprising three light intensities (14, 42, 70 µW/mm2), each with and without wind (35 cm/s) (on/off), totaling 12 steps (S1–S12) (Figure 1—figure supplement 1A). The total experiment duration was 180 s × 12 = 2160s = 36 min. The light intensities were measured with a thermal power sensor (Thorlabs S310C) connected to a power and energy-meter console (Thorlabs PM100D). The flies were recorded with an AVT Guppy PRO F046B camera fitted with an IR bandpass filter, which was positioned on the top of the arena. The camera was connected to a computer running custom LabVIEW software (CRITTA), which was used to determine the flies’ head positions.

WALISAR protocol and validation

Request a detailed protocol

Given that the ordering of the experiments in repeated-measure designs can affect outcomes, the WALISAR protocol was tested with two sets of experiments performed on Orco-Gal4> UAS-CsChrimson flies, in an ascending or descending light-intensity order (Collie et al., 2003; Howitt and Cramer, 2007; McCall and Appelbaum, 1973). The responses in the two orders were similar for eight of the 12 epochs, being different only for S2, S3, S10, and S11. The difference was due to the weak valence responses in the second- and third-order epochs: S2 and S3 produced lower effects in the ascending order (Figure 2—figure supplement 1E), S10 and S11 produced lower effects in the descending order (Figure 2—figure supplement 1F). The underestimation of the second- and third-order epochs had little effect on the overall interpretation of the results because eight of the epochs generated similar results. Moreover, even a hypothetical extreme case in which the valence is overlooked in the second- and third-order epochs would result in a false negative, rather than a false positive. As such, the experimental order was concluded to not have a major effect on the WALISAR results. Additionally, the effect sizes in the first epoch tended to be smaller than the second epoch when the light was downwind of the air flow (Figure 2—figure supplement 1); to eliminate this bias, only second epochs were used for further data analysis.

One concern is that the activity rates and temporal structure of optogenetic stimulation are likely different from direct odor stimulation. To mitigate this, we conducted a study of the effect of optogenetic temporal structure, finding that—while this is a relevant concern—continuous illumination is a more conservative method (Tumkaya et al., 2019). We also benchmarked behavioral responses for the Orco neurons against results from a prior study that performed physiological recordings and used a different temporal structure (Bell and Wilson, 2016), finding that the WALISAR protocol has comparable sensitivity (Figure 1—figure supplement 3).

WALISAR data analysis: the wTSALE metric

Request a detailed protocol

Fly-position data were analyzed with custom Python scripts. The valence of each fly was measured in terms of how much time it spent in the light after first encountering one of the lit zones. Specifically, the first frame in which a fly entered the lit zone was considered the start of the test session; after this initial light encounter, the amount of time spent in the dark was subtracted from time spent in the light and finally divided by the total amount of time. This metric is designated ‘Time Spent After Light Encounter’ (TSALE). The duration between the first light discovery and the end of the light epoch varied across individual files, from never discovering the light (0 s) to being exposed to the light at the start of the epoch (60 s). As such, each fly’s TSALE score was weighted by the post-light duration, termed ‘weighted Time Spent After Light Encounter’ (wTSALE). This weighting was achieved by multiplying the TSALE with the ratio between the remaining time and the full duration of the test epoch. The wTSALE score was calculated for each fly and then averaged for the control and test genotypes. To calculate the effect sizes, responder and driver controls were pooled into a single control group. The mean difference (∆) between the pooled-control and test groups was taken as the effect size (ΔwTSALE). The ∆ distributions and 95% bootstrapped confidence intervals (CIs) were calculated using the DABEST Python package (Ho et al., 2019), and presented in the results in the following format: “∆ [95 CI lower bound, upper bound].”

Overview of the Empirical Bayes method

Request a detailed protocol

Empirical Bayes (EB) is a method for statistical inference using Bayesian hierarchical models. While relatively unknown in behavioral genetics, EB is widely used to filter omics data, having been originally developed for microarray data, and currently routinely used for, among other applications, mass spectrometry proteomics (Koh et al., 2019). This approach is intrinsically connected to the traditional analysis workflow based on hypothesis testing with multiple-testing correction, but with a major difference (Efron and Tibshirani, 2002). The hypothesis-testing approach is solely based on tail probabilities under the null hypothesis, and does not consider true signals. The key advantage of EB is that it explicitly models distributions of both the noises (null hypotheses) and the true signals (alternative hypotheses). Thus, for the analysis of a genetic screen, the key difference of this approach (compared to the conventional hypotheses testing-based analysis) is that we evaluate the significance of response levels across all ORNs simultaneously, modeling their distributions as a mixture of a random noise component and a real signal component. Empirically speaking, when the true effect sizes are modest, this typically increases the sensitivity of detecting mild effect sizes. Further, the mixture modeling produces a confidence score, that is the posterior probability of true signal from the underlying Bayesian model, which is the complement of the local false discovery rate (Efron, 2010).

Data analysis with Empirical Bayes

Request a detailed protocol

To draw conclusions from multiple experiments (performed at one of three light intensities) for a given ORN, the dimension of the data was reduced into one summary statistic (D) for each ORN, as follows Efron et al., 2001:

Di=Mexp-Mctrl1+Mctrl2*0.5

where M is the mean of the wTSALE score for experimental and control flies. Then, D was used to calculate a Z score:

Zi=Di/α0+Si

where Di is the effect size of each ORN, Si is the standard deviation, and α0 is 90th percentile of all the S values (Tusher et al., 2001).

After calculating a vector of Z scores, the EBprot software package (Koh et al., 2015) was used to apply the empirical Bayes method to differentiate true behavioral changes from noisy observations through direct estimation of the respective distributions. The empirical Bayes method directly models the valence Z scores as random variables from a two-component mixture, where the noise component is modeled as Gaussian distribution and the signal component (true behavioral changes) is modeled by a non-parametric distribution (Efron, 2010). In effect, the method borrows statistical information across the ORNs to learn the data generating distributions of the null and alternative hypotheses. As the inferential method is based on the Bayesian framework, the model fit naturally yields the posterior probability of true change of each ORN as well as the false discovery rate (FDR) associated with any posterior probability threshold. The true signal component is subsequently divided into positive and negative sub-components depending on the sign of the Z-scores. From this three-component mixture, a signed probability is derived to indicate the likelihood that the valence score came from each of the three candidate distributions (negative, neutral, and positive valence). Prior to calculating the signed probabilities, EBprot removes outliers from the valence distributions (Koh et al., 2015). Effect sizes associated with a < 25% FDR were considered for further analysis.

Modeling the valence responses of ORN combinations

Request a detailed protocol

The ORN-combo valences were modeled using three functions, each widely used in neural network research for input pooling: summation, max-pooling, and min-pooling. Summation simply sums two individual valence values, while max-pooling and min-pooling return the larger or smaller absolute value of the two components, respectively (Goodfellow et al., 2016), as follows:

Summation=i0+i1
Max-pooling=maxi0,i1,key=absolutevalue
Min-pooling=mini0,i1,key=absolutevalue

where i0 and i1 are real numbers.

For example, given i0 = –2 and i1 = + 1, the summation, max-pooling, and min-pooling functions would return -1,–2, and +1, respectively.

Correlation and agreement analyses

Request a detailed protocol

Linear regressions were performed using the SciPy library in Python (Jones et al., 2001). Bootstrapped CIs for the coefficient of determination (R2) were calculated using the scikits-bootstrap package in Python (Evans, 2019). Bland-Altman plots were generated using custom scripts using the matplotlib and seaborn libraries in Python (Bland and Altman, 1999; Hunter, 2007; Waskom et al., 2017).

ORN interaction analyses

Request a detailed protocol

Custom R scripts were used to perform statistical inference on multiple linear regression models characterizing the association between single-ORN classes and ORN-combo. Single-ORNs (ORN1 and ORN2) were used as predictors, while the ORN-combo phenotype was the dependent variable, as follows:

ORN-combo=β0+β1×ORN1+β2×ORN2

From the original data, 10,000 bootstrap samples of the single-fly data were drawn from each group (ORN1, ORN2, and ORN-combo) and the flies were ordered by their valence levels and paired into trios of ORN1, ORN2, and ORN-combo. In each bootstrap sample, a multiple linear regression model was fitted, which revealed the probabilistic distributions of the beta weights (β1 and β2).

Predictive modeling

Request a detailed protocol

We considered that the available published data has a disproportionate ORN to odor ratio (23–110), with the possibility of internal correlation. Prior to building a valence prediction model, we performed exploratory analyses on the datasets: Pearson correlation among the ORNs, hierarchical clustering, and partial least squares discriminant analysis (PLS-DA). All three analyses were conducted using the pandas and scikit-learn libraries in Python (McKinney, 2010; Pedregosa et al., 2011). The eight latent variables (LVs) calculated by the PLS-DA analysis were used to train the linear and non-linear models using the scikit-learn library in Python (Pedregosa et al., 2011). The performance of the models were evaluated by using the adjusted-R2 and the root-mean-squared error (RMSE). The RMSE values were calculated using a shuffled-split ten-fold cross-validation scheme in Python’s scikit-learn library (Pedregosa et al., 2011). To weight the prediction models, we calculated the Pearson correlation coefficients between Hallem and Carlson’s firing rates and our single-ORN wTSALE valence scores of the 23 ORNs for all 110 samples. So that the odors with higher correlation play a more important role in the model fitting, absolute value of these correlations were then used to re-weight the odor samples in the regression model.

Re-analyzing previous valence studies

Request a detailed protocol

The relevant studies were downloaded in pdf format, and the data of interest were extracted by using the measuring tool in Adobe Acrobat Pro (Adobe Systems USA). The extracted values for control and experimental groups were then used to calculate the standardized effect size, Cohen’s d (Cumming and Calin-Jageman, 2016). If technical replicates were used to calculate statistics, the effect sizes were corrected accordingly (Supplementary file 2 - Sheet 3).

Code Availability

Request a detailed protocol

All of the data generated by this study are available to download from Zenodo (https://doi.org/10.5281/zenodo.3994033). The code is available at (https://github.com/ttumkaya/WALiSuite_V2.0; Tumkaya, 2022; copy archived at swh:1:rev:35ca421e06223bc2d5f167783c3029b2a8240a85).

Data availability

Data and code availability: All of the data generated by this study are available to download from Zenodo (https://doi.org/10.5281/zenodo.3994033). The code is available at https://github.com/ttumkaya/WALiSuite_V2.0 copy archived at swh:1:rev:35ca421e06223bc2d5f167783c3029b2a8240a85.

The following data sets were generated
    1. Tumkaya T
    2. Burhanudin S
    (2020) Zenodo
    Dataset for: Majority of olfactory-receptor neurons have individually neutral effects on behavior.
    https://doi.org/10.5281/zenodo.3994033

References

  1. Book
    1. Goodfellow I
    2. Bengio Y
    3. Courville A
    (2016)
    Deep Learning
    MIT Press.
  2. Book
    1. Howitt D
    2. Cramer D.
    (2007)
    Introduction to research methods in psychology
    Pearson Education.
  3. Conference
    1. McKinney W
    (2010)
    Proceedings of the 9th Python in Science Conference
    Data structures for statistical computing in python. pp. 51–56.
    1. Pedregosa F
    2. Varoquaux G
    3. Gramfort A
    4. Michel V
    5. Thirion B
    (2011)
    Scikit-learn: Machine learning in Python
    The Journal of Machine Learning Research 12:2825–2830.

Decision letter

  1. Sonia Sen
    Reviewing Editor; Tata Institute for Genetics and Society, India
  2. Piali Sengupta
    Senior Editor; Brandeis University, United States
  3. Matthew C Smear
    Reviewer; University of Oregon, United States

Our editorial process produces two outputs: (i) public reviews designed to be posted alongside the preprint for the benefit of readers; (ii) feedback on the manuscript for the authors, including requests for revisions, shown below. We also include an acceptance summary that explains what the editors found interesting or important about the work.

Decision letter after peer review:

Thank you for submitting your article "Most primary olfactory neurons have individually neutral effects on behavior" for consideration by eLife. Your article has been reviewed by 3 peer reviewers, and the evaluation has been overseen by a Reviewing Editor and Piali Sengupta as the Senior Editor. The following individual involved in review of your submission has agreed to reveal their identity: Matthew C Smear (Reviewer #2).

The reviewers have discussed their reviews with one another, and the Reviewing Editor has drafted this to help you prepare a revised submission.

We appreciated the quality, extent, and importance of this work. While doing so, we had a few concerns that we think the authors should be able to address. They are listed below.

1. Anaesthesia: We were concerned about the short recovery period post cold anesthesia and prior to the behavioural assay. Since cold anesthesia is known to have effects on behaviour, could the authors please demonstrate that a longer duration of recovery doesn't alter their findings of neutral ORN valence?

2. wTSALE: We were concerned that this method of weighting used by the authors may be obscuring real behavioural phenomenon and therefore masking valence. Could the authors please revisit this? Providing more traces of the different response types – attraction, avoidance, weak responses, etc. – would also be helpful.

3. ORN combinations: One of the key points of this manuscript is what it tells us about the rules by which ORN combinations work. While the authors show what their study rules out, we felt that they fall short of discussing what might be occurring instead. So, could the authors please include some discussion around this point?

In this section we also recommend incorporating SupFig8 into the main figure 5. It possible that different ORN pair use different interaction rules and the grouped analysis in figure 5 would mask this. Sup Fig8 is more informative in this regard.

4. Statistics: The statistics in this manuscript are quite involved. We recognise that the authors are promoting the use of Empirical-Bayes methods for statistical inferences. Since this is not commonplace, could the authors please incorporate an intuitive explanation about Empirical Bayes, its assumptions, and why it's better suited to the analysis? We think this will greatly improve accessibility of this manuscript, and therefore its impact.

5. Comparison with the Bell and Wilson study: Could the authors please include the number of ORN pairs tested in the B and W study and their own (28 and 7)? With respect to the stimulation conditions listed in this table (Supp Table 4), we assume that the authors' count of 6 conditions is because they are including with and without airflow for their 3 light intensities. In this case, the B and W study should be listed as 16. Alternatively (indeed preferably), the two studies should be listed as 8 and 3 respectively.

Reviewer #1 (Recommendations for the authors):

Figure 1A I cannot tell the direction of airflow in the corridors. The air port is only shown on one side, does that mean airflow is unidirectional? Where is the exhaust?

I would prefer a more schematic/conceptual drawing of the arena than this quasi-realistic one where the main feature that pops out to me are the flies themselves. I would prefer that the drawing conveys the technical details needed to evaluate what the flies experience in the assay.

Fig1F The y-axis label wTSALE should be swapped for a term with some intrinsic meaning. There isn't even any basic description of what wTSALE means in the Results section, the reader has to go to the Methods. I think it would be helpful for the reader to understand the assay more clearly if the full description is in the Results. It is basically the proportion of time the flies spend outside the light, with the clock starting only after the fly has experienced the light the first time. That's pretty easy to understand and the y-axis label could be % time outside light' or even 'preference light' for positive values and 'preference dark' for negative values without criminally oversimplifying the measurement (IMO).

Also I am not convinced the weighting the authors use (i.e. wTSALE vs TSALE) is really justified. Essentially they are trying to control for shorter sampling periods of the fly's behavior in an extremely simple linear way. That implies that short sampling periods may not be representative – is it fair to simply weight those down so the score goes closer to zero since that actually indicates a lack of preference? Basically since any value on their y-axis carries meaning, it seems unfair to weight some points down simply because sampling wasn't extensive enough. Why not just require some minimal time window for flies to have experienced the light (i.e. know what they are choosing) and look at the overall proportion of time in light vs dark?

P.10 "This result indicates that wind has essentially no impact on ORN-elicited behavior in walking flies" this should read 'single-ORN-elicited behavior' since this is all the authors tested. ORN-elicited behavior could be read as ORN activation in general i.e. odor-based activation, where there is likely an effect of wind at least in some assays.

Figure 5: The authors analyze all ORN pairs together to test whether they summate/max pool/min pool but prior work (Bell and Wilson) showed that some pairs summate while others max pool, which would confound the style of grouped analysis in Figure 5G-I.

Additionally since only 7 combinations were used and only 3 intensity levels, this figure is the weakest part of the paper, which up to this point has been extremely extensive. It also makes the first entry in Table 4 (Number ORNs tested = 45 vs n=8 for Bell and Wilson) unfair since Bell and Wilson actually looked at all combinations among 8 ORNs.

Can the authors discuss more about how the dominant β value can flip as the stimulation intensity increases? How would that work in terms of neural activity in the biological network? Also, what does the diversity of β values imply about biological network, does it potentially correspond to different weights on different downstream targets?

Ending on a negative result is a little disappointing – one positive point the authors could make is that (with one exception) the ORN combinations all transition towarrds more max pooling at higher stimulation rates. This suggests an competitive interaction between channels, which is easy to imagine. Undoubtedly it is complex with different downstream targets having different rules, but this is one fairly consistent trend.

I should say that I found Bell and Wilson more convincing because they examine interactions for each ORN pair over a wider range of spike rates. Here there are just three points for comparison, and when my eyes look at Figure 5C-E it seems that there is not a lot of difference between the three interaction modes.

Finally, the authors should somehow incorporate FigS8 into the main text since I'm sure that the interaction mode depends on the pair of ORNs being examined.

Reviewer #2 (Recommendations for the authors):

– I need to understand the raw data better. What are the flies actually doing here? In 1B, the example fly seems to be walking back and forth at about 0.1 Hz. Is this representative of the population? Do the flies ever not move at all? How is this outcome dealt with? The methods mention that Empirical Bayes has a principled way of excluding outliers. What is that way? When the fly's path enters into the illuminated region, it seems to immediately stop and walk back to the opposite wall, and then on its next two cycles it stops before entering the illuminated region. Is this because the light spreads or does the fly remember where it hit the light before? What happens when the illumination occurs when the fly is already on the illuminated side? The effect of Gr66a>Chr (1F) is much larger than any of the OR effects. What does a weaker avoidance response look like? What does an attraction response look like? The mean +/- 95% CI plots of 1C-E do not answer these questions. More individual animal trajectories and population occupancy heat maps would help a lot. Exclusively compressing the data to the one wTSALE number may well be obscuring worthwhile features of the behavior. With a richer characterization of the behavior, it might be possible to reduce the sample size and simplify the statistics.

– The statistical methods are unusual and seem unnecessarily complicated (at least to me). Further, why these were used instead of something more conventional? Readers (at least this reader) would benefit greatly from clear language giving an intuition for how Empirical Bayes works, what are its assumptions, and why it is superior to more conventional, easier-to-understand methods.

– The distribution of wTSALE in Figure 2 F and G is striking. In these plots, including the controls, there is a large mode at wTSALE=1. This mode is not apparent in the distributions of 1F. Why are control flies so much more attracted to light in these experiments? How does Empirical Bayes deal with non-Gaussian distributions?

– What direction does the wind flow through the chamber? It appears to run perpendicular to the illumination axis. Could this matter? Does wind itself impact the locomotion of the flies? Since only δ-wTSALE is shown, it seems possible that wind may affect the behavior in a way that would obscure an effect. Here again it would be helpful to show more of what the flies are actually doing.

– The authors invoke "complex circuit dynamics" to explain the results of the combined-receptor experiments. I'm not sure what the authors mean here. "Dynamics" implies that time-dependent processes determine valence. If this were the case, these experiments would show no effect, since the stimuli don't recapitulate the dynamics of odor-evoked ORN activity. The discussion in a recent paper by Ron Yu's group (Qiu et al., 2021; Current Biology), deals with the non-labelled-line-ness of the mouse olfactory system in a thoughtful way. A similar discussion would benefit this paper as well.

Reviewer #3 (Recommendations for the authors):

1. The authors use a large array of GAL4 driver lines that they claim cover only the relevant ORN type. However, for most of these lines this was not examined. Although in the past such lines were used for behavior experiments, recent studies are much stricter with the use of driver lines. Many studies have demonstrated that even expression in a single neuron (other than the target neurons) either in the central brain or in the VNC can affect behavioral results. The authors therefore must show that the lines used in this study only label the target neurons either by providing adequate citations or by examining this directly with confocal stacks of both whole brain and VNC.

2. The authors do not show the relevance of their optogenetic activation of ORNs to odor activation of ORNs. Previous studies have shown that optogenetic activation of ORNs generates a firing rate of approximately 30 Hz (Bell and Wilson, 2016; Fox and Nagel, 2021). In contrast, ORNs can reach firing frequencies of up to 250 Hz in response to odors (Hallem and Carlson, 2006). In addition, ORNs show temporal dynamics, whereas I presume that the continuous illumination generates a more uniform response. The authors briefly discuss this in the methods section. They claim that "continuous illumination is a more conservative method (Tumkaya et al., 2019)". However, the same authors claim in their Tumkaya et al., 2019 manuscript that "These results suggest that neither stimulation type is necessarily superior to the other: static- or pulsed-light stimulation can capture more of the native responses than the other in inducing olfactory behavior, depending on the neuronal type". The authors also claim that "We also benchmarked behavioral responses for the Orco neurons against results from a prior study that performed physiological recordings and used a different temporal structure (Bell and Wilson, 2016), finding that the WALISAR protocol has comparable sensitivity (Figure S3)". The fact that both optogenetic activations has similar behavioral results does not imply any relevance to an olfactory cue.

My main concern is that the current optogentic stimulation probably activates ORNs relatively weakly, thus mimicking low odor concentration. As low odor concentrations elicit in many cases only weak behavioral responses it is more than possible that the lack of behavioral effect is just due to "low concentration" and not an indication to the actual role of each ORN.

Taken together, I think the authors should go the extra mileage and show some relevance to olfactory stimuli.

3. The authors own data raises potential problems with their approach. Some of the ORNs that are classified as driving aversion or attraction seem to change valence value they induce with the light intensity. For example, the authors report Or42b to drive attraction in agreement with published literature. However, at the strongest light intensity it is actually neutral. Similarly, the authors report Or85d to drive aversion. However, at the strongest light intensity it is also neutral. So, are these ORs "neutral"?

4. The authors test a number of previously suggested linear models and find that they do not predict how two-ORN odor valence emerges from single-ORN valence. However, linear models were shown to be insufficient to predict odor valence (Badel et al., 2016). It is thus not surprising that these linear models failed.

5. The authors use two databases, one of odor responses (Hallem and Carlson 2006) and one of behavioral responses (Knaden et al., 2012) along with a linear model to try and predict odor valence from ORN activity. However, as mentioned above linear models are not adequate for describing the relation between ORN activity and Odor valence. Furthermore, I think the Knaden et al., database is a wrong database to use in this context. Knaden et al., used a trap assay. In this assay, flies are captured in the trap after a single entrance to the odor source. Thus, exploratory behavior, in which flies examine the odor and then can decide to avoid it, cannot occur, and this assay is expected to be biased towards reporting odors as attractive. Indeed, this was the case in the Knaden et al., database in contrast to other published results. This database was suitable for the claim raised by Knaden et al., that looked only at the most aversive and attractive odors, but it cannot be used to try to predict any odor valence.

6. The authors used cold anesthesia just prior to loading the flies to the chambers and only 30 second acclimation following the cold anesthesia. However, cold anesthesia is known to have effects on behavior, increasing response time, reducing locomotion and reducing overall responses (just a few examples, Barron, 2000; MacMillan et al., 2017, Trannoy et al., 2015). I think most studies today try to avoid cold anesthesia just before the experiment. My concern here is that the lack of effect for most ORNs, may arise from general behavior impairment. Can the authors give a few examples from the neutral ORNs without cold anesthesia?

7. The authors conclude that: 1. "the majority of primary olfactory sensory neurons have neutral behavioral effects individually". This conclusion (as mentioned above) is definitely correct for the optogenetic activation, but its relevance to odor valence is questionable. Furthermore, Badel et al., 2016 already demonstrated with actual odor stimuli that “We find that the behavior is accurately predicted by a model summing normalized glomerular responses, in which each glomerulus contributes a specific, small amount to odor preference.” Thus, the novelty of the current study is not large.

Their second conclusion is that “olfactory sensory neurons…participate in broad, odor-elicited ensembles with potent behavioral effects arising from complex interactions”. I agree with them that olfactory coding is complex. However, they did not show any actual odor responses to support their claim, neither did they provided even one complex mechanism. I think that stating that olfaction is complex is just not enough.

8. To my understanding the order of the β coefficients can affect the interpretation of the data. However, I could not find a reference for this in the methods. Can the authors please elaborate on this?

https://doi.org/10.7554/eLife.71238.sa1

Author response

We appreciated the quality, extent, and importance of this work. While doing so, we had a few concerns that we think the authors should be able to address. They are listed below.

1. Anaesthesia: We were concerned about the short recovery period post cold anesthesia and prior to the behavioural assay. Since cold anesthesia is known to have effects on behaviour, could the authors please demonstrate that a longer duration of recovery doesn’t alter their findings of neutral ORN valence?

We appreciate this concern, however there are three lines of evidence that suggest the brief ice anesthesia is not a major cause of effect-size underestimation.

First, fly loading takes around 5 min and the full experimental duration is 36 min, meaning that, for later epochs, the effective total-recovery durations are over 30 min. Thus, even if earlier epochs were underestimating valence, the latter epochs are less likely to be affected.

Second, when compared with the literature, the positive controls are concordant. To benchmark our method, we used a quantitative literature review with standardized effect sizes of two reference lines: Gr66a-Gal4 and Orco-Gal4 (Figure 1-S3). It shows that our assay yields valence that is very similar to previous studies.

Comparing maximal scores, the comparisons of these lines include a 20% decrease (Gr66a), a 10% increase (Orco-Gal4 with wind), and a 5-fold increase (Orco-Gal4 in no wind). This shows that the overall protocol with the brief cold anesthesia does not systematically underestimate valence when compared with published protocols.

Third, the unbiased screen turned up a majority of hits that were already reported in the literature as being valent ORN types. Discussed at greater length below, this also indicates that there is nothing especially defective about the ice protocol.

Action taken: We added text to Methods to clarify the total duration. “The total experiment duration was 180 s × 12 = 2160s = 36 min.”

We added a row to the comparison in Supplementary File 2 – Sheet 3 to include that Bell and Wilson used aspiration, while we used cold anesthesia.

2. wTSALE: We were concerned that this method of weighting used by the authors may be obscuring real behavioural phenomenon and therefore masking valence. Could the authors please revisit this? Providing more traces of the different response types – attraction, avoidance, weak responses, etc. – would also be helpful.

We appreciate this concern and have addressed it by showing a case study of Or67d-Gal4 (Author response image 1) and several other drivers (Author response images 2–7) finding that the choice of metric has a modest impact on the overall valence outcome. For Or67d, the majority of the flies discover the optogenetic light within the first ~5 seconds of a WALISAR experiment (Author response image 1F). The full-epoch metric conserves more preference information than a short-epoch cutoff, but there still are a few flies whose first light encounter does not occur until much later in the epoch (Author response image 1F). For example, a fly that encounters the light in the final three seconds of the test will give less reliable preference information. To account for this difference, we linearly weighed the preference scores by the percentage of epoch that the flies actually experienced. So, wTSALE does not underestimate valence, but rather is an unbiased and robust metric that is fairly similar to other metrics.

Author response image 1
wTSALE presents Or67d-induced behaviour more accurately than commonly used metrics in the fieldA.

An estimation plot presents the individual preference (upper axes) and valence (lower) of flies with activatable Or67d+ neurons in the WALISAR assay. In the upper panel, each dot indicates a preference index (PI) for an individual fly that was calculated by using the last 3 second of the epoch: w1118; Or67d-Gal4 and w1118; UAS-CsChrimson flies are colored blue; and Or67d-Gal4 > UAS-CsChrimson flies are in orange. The mean and 95% Cis associated with each group are shown by the adjacent broken line. In the bottom panel, the black dots indicate the mean difference (ΔwTSALE) between the relevant two groups: the valence effect size. The black whiskers span the 95% Cis, and the orange curve represents the distribution of the mean difference. B. An estimation plot presents the individual preference (upper axes) and valence (lower) of flies with activatable Or67d+ neurons in the WALISAR assay that were calculated by using the last 5 seconds of the experiment. C. Preference index (PI) for each fly that used only the final 10 seconds of the experiment. The color-code, layout, and statistics are the same as Panels A and B. D. Preference index (PI) for each fly when the whole epoch (45 seconds) was included into the analysis. The color-code, layout, and statistics are the same as Panels A and B. E. The individual preference (upper axes) and valence (lower) of flies by the wTSALE metric. F. A scatter plot shows how long it takes for each fly to encounter the optogenetic light once it is switched on. The median and 95% Cis associated with each group are shown by the adjacent broken line.

Author response image 2
Or67d neuron activation is attractive to flies A-C-E.

Trace plots representing the mean location of controls (w1118; Or67d-Gal4 and w1118; UAS-CsChrimson), and test flies (Or67d-Gal4 > UAS-CsChrimson) throughout an experiment at 14, 42, and 70 μW/mm2 light intensities. The blue and orange ribbons indicate 95% CIs for the control and test flies, respectively. Second epochs were used to calculate the preference scores (shown in black rectangle). B. An estimation plot presents the individual preference (upper axes) and valence (lower) of flies with activatable Or67d+ neurons in the WALISAR assay. In the upper panel, each dot indicates a preference score (wTSALE) for an individual fly: w1118; Or67d-Gal4 and w1118; UAS-CsChrimson flies are colored blue; and Or67d-Gal4 > UAS-CsChrimson flies are in orange. The mean and 95% CIs associated with each group are shown by the adjacent broken line. In the bottom panel, the black dots indicate the mean difference (ΔwTSALE) between the relevant two groups: the valence effect size. The black whiskers span the 95% CIs, and the orange curve represents the distribution of the mean difference. D. A scatter plot shows how long it takes for each fly to encounter the optogenetic light once it is switched on. The median and 95% CIs associated with each group are shown by the adjacent broken line. F. An estimation plot presents the preference index (PI) for each fly calculated by using the locomotion data from the whole epoch (45 seconds). The color-code, layout, and statistics are the same as Panel B.

Author response image 3
Orco activation triggers attraction A-C-E.

Trace plots representing the mean location of controls (w1118; Orco-Gal4 and w1118; UAS-CsChrimson), and test flies (Orco-Gal4 > UAS-CsChrimson) throughout an experiment at 14, 42, and 70 μW/mm2 light intensities. The blue and orange ribbons indicate 95% CIs for the control and test flies, respectively. Second epochs were used to calculate the preference scores (shown in black rectangle). B. An estimation plot presents the individual preference (upper axes) and valence (lower) of flies with activatable Orco+ neurons in the WALISAR assay. In the upper panel, each dot indicates a preference score (wTSALE) for an individual fly: w1118; Orco-Gal4 and w1118; UAS-CsChrimson flies are colored blue; and Orco-Gal4 > UAS-CsChrimson flies are in orange. The mean and 95% CIs associated with each group are shown by the adjacent broken line. In the bottom panel, the black dots indicate the mean difference (ΔwTSALE) between the relevant two groups: the valence effect size. The black whiskers span the 95% CIs, and the orange curve represents the distribution of the mean difference. D. A scatter plot shows how long it takes for each fly to encounter the optogenetic light once it is switched on. The median and 95% CIs associated with each group are shown by the adjacent broken line. F. An estimation plot presents the preference index (PI) for each fly calculated by using the locomotion data from the whole epoch (45 seconds). The color-code, layout, and statistics are the same as Panel B.

Author response image 4
Or7a has neutral valence A-C-E.

Trace plots representing the mean location of controls (w1118; Or7a-Gal4 and w1118; UAS-CsChrimson), and test flies (Or7a-Gal4 > UAS-CsChrimson) throughout an experiment at 14, 42, and 70 μW/mm2 light intensities. The blue and orange ribbons indicate 95% CIs for the control and test flies, respectively. Second epochs were used to calculate the preference scores (shown in black rectangle). B. An estimation plot presents the individual preference (upper axes) and valence (lower) of flies with activatable Or7a+ neurons in the WALISAR assay. In the upper panel, each dot indicates a preference score (wTSALE) for an individual fly: w1118; Or7a-Gal4 and w1118; UAS-CsChrimson flies are colored blue; and Or7a-Gal4 > UAS-CsChrimson flies are in orange. The mean and 95% CIs associated with each group are shown by the adjacent broken line. In the bottom panel, the black dots indicate the mean difference (ΔwTSALE) between the relevant two groups: the valence effect size. The black whiskers span the 95% CIs, and the orange curve represents the distribution of the mean difference. D. A scatter plot shows how long it takes for each fly to encounter the optogenetic light once it is switched on. The median and 95% CIs associated with each group are shown by the adjacent broken line. F. An estimation plot presents the preference index (PI) for each fly calculated by using the locomotion data from the whole epoch (45 seconds). The color-code, layout, and statistics are the same as Panel B.

Author response image 5
Or69a+ neurons trigger no response A-C-E.

Trace plots representing the mean location of controls (w1118; Or69a-Gal4 and w1118; UAS-CsChrimson), and test flies (Or69a-Gal4 > UAS-CsChrimson) throughout an experiment at 14, 42, and 70 μW/mm2 light intensities. The blue and orange ribbons indicate 95% CIs for the control and test flies, respectively. Second epochs were used to calculate the preference scores (shown in black rectangle). B. An estimation plot presents the individual preference (upper axes) and valence (lower) of flies with activatable Or69a+ neurons in the WALISAR assay. In the upper panel, each dot indicates a preference score (wTSALE) for an individual fly: w1118; Or69a-Gal4 and w1118; UAS-CsChrimson flies are colored blue; and Or69a-Gal4 >UAS-CsChrimson flies are in orange. The mean and 95% CIs associated with each group are shown by the adjacent broken line. In the bottom panel, the black dots indicate the mean difference (ΔwTSALE) between the relevant two groups: the valence effect size. The black whiskers span the 95% CIs, and the orange curve represents the distribution of the mean difference. D. A scatter plot shows how long it takes for each fly to encounter the optogenetic light once it is switched on. The median and 95% CIs associated with each group are shown by the adjacent broken line. F. An estimation plot presents the preference index (PI) for each fly calculated by using the locomotion data from the whole epoch (45 seconds). The color-code, layout, and statistics are the same as Panel B.

Author response image 6
Or56a induces aversion response in flies A-C-E.

Trace plots representing the mean location of controls (w1118; Or56a-Gal4 and w1118; UAS-CsChrimson), and test flies (Or56a-Gal4 > UAS-CsChrimson) throughout an experiment at 14, 42, and 70 μW/mm2 light intensities. The blue and orange ribbons indicate 95% CIs for the control and test flies, respectively. Second epochs were used to calculate the preference scores (shown in black rectangle). B. An estimation plot presents the individual preference (upper axes) and valence (lower) of flies with activatable Or56a+ neurons in the WALISAR assay. In the upper panel, each dot indicates a preference score (wTSALE) for an individual fly: w1118; Or56a-Gal4 and w1118; UAS-CsChrimson flies are colored blue; and Or56a-Gal4 > UAS-CsChrimson flies are in orange. The mean and 95% CIs associated with each group are shown by the adjacent broken line. In the bottom panel, the black dots indicate the mean difference (ΔwTSALE) between the relevant two groups: the valence effect size. The black whiskers span the 95% CIs, and the orange curve represents the distribution of the mean difference. D. A scatter plot shows how long it takes for each fly to encounter the optogenetic light once it is switched on. The median and 95% CIs associated with each group are shown by the adjacent broken line. F. An estimation plot presents the preference index (PI) for each fly calculated by using the locomotion data from the whole epoch (45 seconds). The color-code, layout, and statistics are the same as Panel B.

Author response image 7
Flies avoid Or85d activation A-C-E.

Trace plots representing the mean location of controls (w1118; Or85d-Gal4 and w1118; UAS-CsChrimson), and test flies (Or85d-Gal4 > UAS-CsChrimson) throughout an experiment at 14, 42, and 70 μW/mm2 light intensities. The blue and orange ribbons indicate 95% CIs for the control and test flies, respectively. Second epochs were used to calculate the preference scores (shown in black rectangle). B. An estimation plot presents the individual preference (upper axes) and valence (lower) of flies with activatable Or85d+ neurons in the WALISAR assay. In the upper panel, each dot indicates a preference score (wTSALE) for an individual fly: w1118; Or85d-Gal4 and w1118; UAS-CsChrimson flies are colored blue; and Or85d-Gal4 > UAS-CsChrimson flies are in orange. The mean and 95% CIs associated with each group are shown by the adjacent broken line. In the bottom panel, the black dots indicate the mean difference (ΔwTSALE) between the relevant two groups: the valence effect size. The black whiskers span the 95% CIs, and the orange curve represents the distribution of the mean difference. D. A scatter plot shows how long it takes for each fly to encounter the optogenetic light once it is switched on. The median and 95% CIs associated with each group are shown by the adjacent broken line. F. An estimation plot presents the preference index (PI) for each fly calculated by using the locomotion data from the whole epoch (45 seconds). The color-code, layout, and statistics are the same as Panel B.

3. ORN combinations: One of the key points of this manuscript is what it tells us about the rules by which ORN combinations work. While the authors show what their study rules out, we felt that they fall short of discussing what might be occurring instead. So, could the authors please include some discussion around this point?

In this section we also recommend incorporating Sup Figure 8 into the main figure 5. It possible that different ORN pair use different interaction rules and the grouped analysis in figure 5 would mask this. Sup Fig8 is more informative in this regard.

We have incorporated the Supplementary Figure into Figure 5, and written a new Discussion section about possible mechanisms of valence computation.

4. Statistics: The statistics in this manuscript are quite involved. We recognise that the authors are promoting the use of Empirical-Bayes methods for statistical inferences. Since this is not commonplace, could the authors please incorporate an intuitive explanation about Empirical Bayes, its assumptions, and why it’s better suited to the analysis? We think this will greatly improve accessibility of this manuscript, and therefore its impact.

We thank the reviewing team for the suggestion. Since Empirical Bayes is a method with potential applicability for screens, including neurogenetic screens, we agree that it is important to explain the method in accessible terms.

We have added a version of the following explanation about Empirical Bayes, its assumptions, and why it’s better suited to the analysis.

“Empirical Bayes (EB) is a method for statistical inference using hierarchical Bayesian models. While relatively unknown in behavioral genetics, EB is widely used to filter omics data, having been originally developed for microarray data, and currently routinely used for, among other applications, mass spectrometry proteomics (Koh et al., 2019). This approach is intrinsically connected to the traditional analysis workflow based on hypothesis testing with multiple-testing correction, but with a major difference (Efron and Tibshirani, 2002). The hypothesis-testing approach is solely based on tail probabilities under the null hypothesis, and does not consider true signals. The key advantage of EB is that it explicitly models distributions of both the noises (null hypotheses), and the true signals (alternative hypotheses). Thus, for the analysis of a genetic screen, the key difference of this approach (compared to the conventional hypotheses testing-based analysis) is that we evaluate the significance of response levels across all ORNs simultaneously, abelled their distributions as a mixture of a random noise component and a real signal component. Empirically speaking, when the true effect sizes are modest, this typically increases sensitivity. Further, the mixture abelled naturally produces a confidence score, i.e. the posterior probability of true signal from the underlying Bayesian model, which is the inverse of the local false discovery rate (Efron, 2010).”

5. Comparison with the Bell and Wilson study: Could the authors please include the number of ORN pairs tested in the B and W study and their own (28 and 7)? With respect to the stimulation conditions listed in this table (Supp Table 4), we assume that the authors’ count of 6 conditions is because they are including with and without airflow for their 3 light intensities. In this case, the B and W study should be listed as 16. Alternatively (indeed preferably), the two studies should be listed as 8 and 3 respectively.

We thank the Reviewers and Editors for their guidance. The B and W study did an experiment with and without airflow only for Orco-Gal4, and then proceeded with airflow experiments from there onwards. The direct comparison is eight versus three light intensities. For the airflow, the comparison is 1 ORN (Orco) versus 45 ORNs.

We have added two rows and updated the numbers in Supplementary Table 4.

Reviewer #1 (Recommendations for the authors):

Figure 1A I cannot tell the direction of airflow in the corridors. The air port is only shown on one side, does that mean airflow is unidirectional? Where is the exhaust?

I would prefer a more schematic/conceptual drawing of the arena than this quasi-realistic one where the main feature that pops out to me are the flies themselves. I would prefer that the drawing conveys the technical details needed to evaluate what the flies experience in the assay.

An additional schematic was added to Figure 1.

Fig1F The y-axis label wTSALE should be swapped for a term with some intrinsic meaning. There isn’t even any basic description of what wTSALE means in the Results section, the reader has to go to the Methods. I think it would be helpful for the reader to understand the assay more clearly if the full description is in the Results. It is basically the proportion of time the flies spend outside the light, with the clock starting only after the fly has experienced the light the first time. That’s pretty easy to understand and the y-axis label could be % time outside light’ or even ‘preference light’ for positive values and ‘preference dark’ for negative values without criminally oversimplifying the measurement (IMO).

We thank the Reviewer for this helpful suggestion. We have improved the description of wTSALE to the Results, in a section titled “WALISAR data analysis: the wTSALE metric”.

Also I am not convinced the weighting the authors use (i.e. wTSALE vs TSALE) is really justified. Essentially they are trying to control for shorter sampling periods of the fly’s behavior in an extremely simple linear way. That implies that short sampling periods may not be representative – is it fair to simply weight those down so the score goes closer to zero since that actually indicates a lack of preference? Basically since any value on their y-axis carries meaning, it seems unfair to weight some points down simply because sampling wasn’t extensive enough. Why not just require some minimal time window for flies to have experienced the light (i.e. know what they are choosing) and look at the overall proportion of time in light vs dark?

Requiring a minimum time window and weighting the score are doing similar things, just that one is thresholded while the other is graded. The weighting was introduced as a non-thresholded way to allow all the available data to contribute in proportion to the time the animal was responding to the optogenetic light, which we consider to be the relevant metric. Regardless, as we show in the metric case study, the exact choice of metric does not dramatically alter the preference outcomes. We are aware that we made many assumptions and human decisions in the course of this project, however the use of wTSALE cannot account for inter-study differences.

Extended discussion: Extracting data from the final seconds of an experiment, and using this data to calculate a preference index is a common practice in the field. These brief arbitrary durations may range from as short as 5 s to 10 s or even just the final position (Aso et al., 2014; Dolan et al., 2019). While widely used, the conventional approach has potential issues, which encouraged us to implement wTSALE.

We investigated possible undersampling arising from using the last 3, 5, or 10 seconds of the optogenetic activation of Or67d neurons (Author response image 1A-C). The overall conclusion remains the same across the three calculations in this example: that Or67d triggers attraction. This comparison shows that wTSALE is not masking or obscuring valence. However, there is a problem with the short-duration metrics: the majority of the flies have either a -1 or +1 preference index (PI). (Indeed, there are so many data points at the extremes, the swarm plot cannot display them). This distorted, tail-heavy, high-variability distribution shows that, for short-duration epochs, P.I. scores are heavily dependent on where in the chamber the flies happened to be for those few seconds (Author response image 1A-C).

Instead of an under-sampled short epoch, one could use the entire test to calculate preference index. However this approach introduces a new pitfall: it would include all flies even if they never encountered optogenetic light—the intervention that is under investigation. As our analysis of the Or67d activation experiment demonstrates, when the whole epoch was used to measure preference, there are many flies with apparent “total aversion” (PI = –1; Author response image 1D). Given that these flies never crossed into the illuminated side of the chamber and experienced activity in the Or67d neurons, it is misleading to conclude that these flies ‘avoid’ the activation of Or67d neurons 100% of the time.

The metric used in our study—wTSALE—utilizes all the locomotion data following a fly encountering the optogenetic light. It (1) eliminates the need for selecting an arbitrary cutoff; (2) uses all of the relevant tracking data for the analysis, thereby reducing sampling error and variation, while capturing the attraction response triggered by the Or67d neurons (Author response image 1E).

We have shown the metrics case study here, which demonstrates that wTSALE is not misestimating valence when compared to more conventional metrics.

P.10 “This result indicates that wind has essentially no impact on ORN-elicited behavior in walking flies” this should read ‘single-ORN-elicited behavior’ since this is all the authors tested. ORN-elicited behavior could be read as ORN activation in general i.e. odor-based activation, where there is likely an effect of wind at least in some assays.

We thank the Reviewer for catching this. We have changed the text as suggested.

Figure 5: The authors analyze all ORN pairs together to test whether they summate/max pool/min pool but prior work (Bell and Wilson) showed that some pairs summate while others max pool, which would confound the style of grouped analysis in Figure 5G-I.

We agree with the Reviewer, which is why we also did the multiple linear regression.

Extended discussion: Of several analyses addressing this hypothesis, the multiple linear regression addresses this directly. The charts shown in Figure 5J display the axes and diagonal lines, abelled either “summation” or “max pooling”. If either/both of the simple rules governed ORN combinations, the β weights would be expected to cluster along these lines. Instead, we observe a broad scatter that shifts with increasing intensity. Both of these features (the locations of the weight markers and the fact that they shift) erode confidence in the either/or simple-rule hypothesis.

The present study was started a year before the publication of Bell and Wilson, and, while we realized early on that this study would inevitably be at least a partial replication, we never intended it to be a full-throated critique of the prior study. Bell and Wilson is an excellent, pioneering study. Nevertheless it is, like the present study, imperfect. We have not used effect sizes and other state-of-the-art analytic methods for their own sake, but because they have been established in the statistical literature to work better than ad hoc methods. To determine the combination type, Bell and Wilson used significance tests between the ORN-combination P.I. and two other metrics (the sum of the components and the larger component), combined with the calculation of P value ratios. This analysis shares all the well-documented problems associated with significant testing for confirmatory analyses, including false dichotomization. The primary B and W analysis workflow is: (1) continuous P.I. data is recast as binary outcomes (significant/not significant), and (2) this new binary data is then mapped to a dual mechanistic hypothesis (sum or max). However, this is based on the false assumption that significant differences are only compatible with the sum or max hypotheses.

It is likely that the B and W data, if it were analyzed with effect-size methods, would also show a variety of non-sum, non-max, weighted poolings. Indeed, two exemplar combinations (Bell and Wilson, Figure 5BC) have response magnitudes that exceed both models, and the overview matrix (Bell and Wilson, Figure 6A) shows that 11 of 28 tests were non-significant. We would have liked to re-analyze B and W’s data with best-practice methods, but when we requested the authors to share their data, they declined. Another barrier to a re-analysis is that their experiments included no Gal4 controls (not done) and very few UAS controls (Supplementary File 2 – Sheet 3), making it impossible to calculate effect sizes that correctly isolate the optogenetic influence.

We moved the former Figure 5GHI to the supplement, and Figure S8 to the main Figure.

Additionally since only 7 combinations were used and only 3 intensity levels, this figure is the weakest part of the paper, which up to this point has been extremely extensive. It also makes the first entry in Table 4 (Number ORNs tested = 45 vs n=8 for Bell and Wilson) unfair since Bell and Wilson actually looked at all combinations among 8 ORNs.

We have added a row to Supplementary File 2 – Sheet 3 describing the difference in combinations.

Can the authors discuss more about how the dominant β value can flip as the stimulation intensity increases? How would that work in terms of neural activity in the biological network? Also, what does the diversity of β values imply about biological network, does it potentially correspond to different weights on different downstream targets?

We thank the Reviewer for this valuable suggestion.

In the current experiment, the variability of β values across a range of stimuli indicates that we captured particular instances of ORN activation among many possible neural response mechanisms, and that many ORN combinations likely share downstream targets. Therefore we can cautiously infer that the underlying neural network is multifactorial and perhaps context-specific. However, we also believe that reverse-engineering the exact on-off combinations through estimated β values requires a wider set of experiments with distinct stimuli than we present. Attempting to translate the meaning of β shifts into the structure or variability of the neural network carries a high risk of over-interpretation.

We have added a section (“Possible models of ORN combination combination”) to mention possible mechanisms underlying the β shifts.

Ending on a negative result is a little disappointing – one positive point the authors could make is that (with one exception) the ORN combinations all transition abelle more max pooling at higher stimulation rates. This suggests an competitive interaction between channels, which is easy to imagine. Undoubtedly it is complex with different downstream targets having different rules, but this is one fairly consistent trend.

See above.

I should say that I found Bell and Wilson more convincing because they examine interactions for each ORN pair over a wider range of spike rates. Here there are just three points for comparison, and when my eyes look at Figure 5C-E it seems that there is not a lot of difference between the three interaction modes.

Although the interaction analysis in B and W does cover more light intensities, we consider the smoothness of these curves to arise from deprecated experimental and statistical practices. The curves use data from experiments that only indirectly refer to UAS controls (just Nflies = 88 total), and lack driver controls entirely. This is unusual for optogenetic experiments in Drosophila, which typically use either no-retinal controls (Badel et al., 2016) and/or driver and responder controls as was done in the present study.

The precision in the B and W response curves appears to be overrepresented in two ways. First, the optogenetic response should incorporate variance from both experimental and test animals; by omitting the controls (and therefore control variance) is expected to increase precision. Second, based on our reading of their Experimental Procedures, the optogenetic response in B and W appears to be averaged from the 16 technical replicates, potentially increasing apparent precision by √16, i.e. four-fold. With both statistical practices applied in tandem, it seems likely that the overall precision could be increased by as much as eight-fold.

Using published effects and our own preliminary control experiments as guidelines, the present study selected a sample size (N = 52, 52, 52 for test, UAS, and Gal4 control flies) that would be adequate to detect medium effect sizes. To our surprise, the effect sizes of the majority of ORNs were small and even undetectable using the sensitive EB method.

Finally, the authors should somehow incorporate FigS8 into the main text since I’m sure that the interaction mode depends on the pair of ORNs being examined.

We have incorporated Figure S8 in the main text.

Reviewer #2 (Recommendations for the authors):

– I need to understand the raw data better. What are the flies actually doing here? In 1B, the example fly seems to be walking back and forth at about 0.1 Hz. Is this representative of the population? Do the flies ever not move at all? How is this outcome dealt with?

If the flies remain in the light (either moving or non-moving) they are counted as P.I. = 1.0. If they remain in the dark, and never encounter the light, they are excluded from analysis. Overall motionless flies were excluded from the analysis. While we share the Reviewer’s interest in odor-related locomotion, a detailed analysis of trajectories is beyond the scope of the present study. We agree that the data set is rich, and will make it freely available for others to download and analyze.

We have uploaded all tracking data to Zenodo (https://doi.org/10.5281/zenodo.3994033).

The methods mention that Empirical Bayes has a principled way of excluding outliers. What is that way?

The current implementation of the Ebprot software offers an option to exclude valence Z scores that are four reference standard deviations (RSD) away from the mean of the same ORN. Reference standard deviation is computed as the median of SDs in the Z scores across all ORNs

When the fly’s path enters into the illuminated region, it seems to immediately stop and walk back to the opposite wall, and then on its next two cycles it stops before entering the illuminated region. Is this because the light spreads or does the fly remember where it hit the light before?

This is an interesting question, and is beyond the scope of the present study. We are preparing several other studies that address this kind of question.

What happens when the illumination occurs when the fly is already on the illuminated side?

The fly is considered to have encountered the light, and the data for wTSALE is calculated from this.

The effect of Gr66a>Chr (1F) is much larger than any of the OR effects. What does a weaker avoidance response look like?

We present the Response plots here.

What does an attraction response look like? The mean +/- 95% CI plots of 1C-E do not answer these questions. More individual animal trajectories and population occupancy heat maps would help a lot. Exclusively compressing the data to the one wTSALE number may well be obscuring worthwhile features of the behavior. With a richer characterization of the behavior, it might be possible to reduce the sample size and simplify the statistics.

Compressing complex behavior to a P.I. metric has a long history across behavioral neuroscience. For example, the widely used Drosophila olfactory T-maze uses endpoint counts, without any tracking data ever collected.

– The statistical methods are unusual and seem unnecessarily complicated (at least to me). Further, why these were used instead of something more conventional? Readers (at least this reader) would benefit greatly from clear language giving an intuition for how Empirical Bayes works, what are its assumptions, and why it is superior to more conventional, easier-to-understand methods.

We thank the Reviewer for this helpful suggestion. We have clarified the EB method above, and added it to the Methods.

– The distribution of wTSALE in Figure 2 F and G is striking. In these plots, including the controls, there is a large mode at wTSALE=1. This mode is not apparent in the distributions of 1F. Why are control flies so much more attracted to light in these experiments?

It is true, these kinds of preference experiments often have heavy-tailed distributions, sometimes much more extreme than the examples in Figure 1F and as presented here. For this reason, we analyze the data with robust bootstrap methods. The δ curve shows, as per the central limit theorem, that the difference distributions are often approximately Gaussian despite these heavy tails.

How does Empirical Bayes deal with non-Gaussian distributions?

The Empirical Bayes inference method (implemented in Ebprot) estimates a mixture distribution with two components: the Gaussian null distribution and the non-Gaussian alternative distribution(s). In Figure 4C, for example, the middle part of the overall distribution corresponds to the Gaussian null distribution, whereas the red and green distributions are non-parametric alternative distributions. In sum, the overall distribution of valence Z-scores is therefore abelled as a flexible non-parametric distribution, not bound by Gaussian distribution assumption.

– What direction does the wind flow through the chamber? It appears to run perpendicular to the illumination axis. Could this matter? Does wind itself impact the locomotion of the flies? Since only δ-wTSALE is shown, it seems possible that wind may affect the behavior in a way that would obscure an effect. Here again it would be helpful to show more of what the flies are actually doing.

Wind runs through the narrow chamber from one end to the other, perpendicular to the light bands. We have not observed wind directly affecting locomotion. We are using the same wind speed as Bell and Wilson: 35 cm/s, around 0.7 knots.

– The authors invoke “complex circuit dynamics” to explain the results of the combined-receptor experiments. I’m not sure what the authors mean here. “Dynamics” implies that time-dependent processes determine valence. If this were the case, these experiments would show no effect, since the stimuli don’t recapitulate the dynamics of odor-evoked ORN activity. The discussion in a recent paper by Ron Yu’s group (Qiu et al., 2021; Current Biology), deals with the non-labelled-line-ness of the mouse olfactory system in a thoughtful way. A similar discussion would benefit this paper as well.

We thank the Reviewer for the suggestion. We added a Discussion section about possible circuit dynamics that support the observed results.

Reviewer #3 (Recommendations for the authors):

1. The authors use a large array of GAL4 driver lines that they claim cover only the relevant ORN type. However, for most of these lines this was not examined. Although in the past such lines were used for behavior experiments, recent studies are much stricter with the use of driver lines. Many studies have demonstrated that even expression in a single neuron (other than the target neurons) either in the central brain or in the VNC can affect behavioral results. The authors therefore must show that the lines used in this study only label the target neurons either by providing adequate citations or by examining this directly with confocal stacks of both whole brain and VNC.

Ultimately, in every project, decisions must be made on how to allocate limited resources, and early on we decided to pursue a behavioral study, not an anatomical survey. Anatomical surveys of this scale are major undertakings (Couto et al., 2005; Fishilevich and Vosshall, 2005). While we appreciate that this is a potential concern, this is outside the scope of this project, and respectfully request that this is not made a prerequisite for publication.

We have added a description to the new Limitations section in the Discussion that highlights that non-antennal cells could be responsible for some of the effects observed.

2. The authors do not show the relevance of their optogenetic activation of ORNs to odor activation of ORNs. Previous studies have shown that optogenetic activation of ORNs generates a firing rate of approximately 30 Hz (Bell and Wilson, 2016; Fox and Nagel, 2021). In contrast, ORNs can reach firing frequencies of up to 250 Hz in response to odors (Hallem and Carlson, 2006). In addition, ORNs show temporal dynamics, whereas I presume that the continuous illumination generates a more uniform response. The authors briefly discuss this in the methods section. They claim that “continuous illumination is a more conservative method (Tumkaya et al., 2019)”. However, the same authors claim in their Tumkaya et al., 2019 manuscript that “These results suggest that neither stimulation type is necessarily superior to the other: static- or pulsed-light stimulation can capture more of the native responses than the other in inducing olfactory behavior, depending on the neuronal type”. The authors also claim that “We also benchmarked behavioral responses for the Orco neurons against results from a prior study that performed physiological recordings and used a different temporal structure (Bell and Wilson, 2016), finding that the WALISAR protocol has comparable sensitivity (Figure S3)”. The fact that both optogenetic activations has similar behavioral results does not imply any relevance to an olfactory cue.

My main concern is that the current optogentic stimulation probably activates ORNs relatively weakly, thus mimicking low odor concentration. As low odor concentrations elicit in many cases only weak behavioral responses it is more than possible that the lack of behavioral effect is just due to “low concentration” and not an indication to the actual role of each ORN.

Taken together, I think the authors should go the extra mileage and show some relevance to olfactory stimuli.

The reviewer is correct: neuronal photoactivation is not a naturalistic stimulus, but a physiological intervention. In this case, we are using a behavioral readout of the depolarizing input into the system. The use of simplified, well-controlled non-natural experimental preparations has a long tradition in the history of neuroscience. All such reductionist experiments have their limitations and serve as a counterpoint to more naturalistic experiments, which typically have the converse limitation of complexity.

Extended discussion: The relevance of the optogenetic screen to olfactory valence is supported by at least two lines of evidence.

First, of the ten ORN types identified by the screen, six ORN types have been previously identified by other groups (using odorants) as ORNs that drive attraction or aversion. The re-discovery of these ORN types lends further credibility to the idea that the optogenetic screen is relevant to olfactory preference. Considering the noise and poor reproducibility in behavioral experiments and neuroscience at large (Button et al., 2013), it is reasonable to characterize these independent replications as displaying remarkable fidelity.

Second, there is a point of reference with an excellent odor-based study that used an odorant-triggered, olfactory-receptor method of ORN activation: geosmin–Or56a heterologous olfactory stimulation (Chin et al., 2018). In comparison with this olfactogenetic assessment of ORN-evoked behaviors, there are five receptor types (35a, 42b, 47a, 67d, Gr21a/Gr63a) that are neutral in with the odorant-evoked intervention, but show valence in our assay (Supplementary File 1, Author Response Tables 1, 2). Of these, for the two receptor types covered by multiple studies (42b, Gr21/Gr63a), the olfactogenetic result is the outlier while our screen is compatible with the broad consensus (Supplementary File 1). These observations indicate that—when compared with a naturalistic mode of odorant activation—the Chrimson activation protocol used in the present study does not consistently underestimate ORN valences.

As experimental interventions, the complexity of naturalistic stimuli poses a challenge. For many odorants, receptor deconvolution is a complex problem that makes it challenging to unambiguously determine ORN→behavior causality. Many ORNs bind a multitude of odorants promiscuously. Even for particular odorant–ORN pairs with high affinity and specificity (abelled lines), it is difficult to completely rule out weak but nevertheless potentially widespread binding to ‘off-target’ ORNs.

We consider that the study already has relevance to olfactory processing.

We have added a passage to the new Limitations paragraph about the differences in firing rates between optogenetic and olfactory stimuli, and the important concern about relevance to olfactory stimuli. We request that odorant experiments are not made a requirement of publication.

Author response table 1
Receptor Valence AssayStimulusSex StageReference
Or7aovipositionolfactogeneticsFadult(Chin et al., 2018)
otwo-choiceoptogeneticMadultthis study
Or19aoovipositionolfactogeneticsFadult(Chin et al., 2018)
otwo-choiceoptogeneticMadultthis study
Or22aoovipositionolfactogeneticsFadult(Chin et al., 2018)
otwo-choiceoptogeneticMadultthis study
Or23aoovipositionolfactogeneticsFadult(Chin et al., 2018)
otwo-choiceoptogeneticMadultthis study
Or35aoovipositionolfactogeneticsFadult(Chin et al., 2018)
+two-choiceoptogeneticMadultthis study
oovipositionolfactogeneticsFadult(Chin et al., 2018)
Or42aotwo-choiceolfactogeneticsFadult(Chin et al., 2018)
otwo-choiceoptogeneticMadultthis study
oovipositionolfactogeneticsFadult(Chin et al., 2018)
Or42b+two-choiceoptogeneticMadultthis study
Or43aoovipositionolfactogeneticsFadult(Chin et al., 2018)
otwo-choiceoptogeneticMadultthis study
Or47aoovipositionolfactogeneticsFadult(Chin et al., 2018)
+two-choiceoptogeneticMadultthis study
Or47bovipositionolfactogeneticsFadult(Chin et al., 2018)
+two-choiceoptogeneticMadultthis study
Or49aovipositionolfactogeneticsFadult(Chin et al., 2018)c
otwo-choiceoptogeneticMadultthis study
Or59covipositionolfactogeneticsFadult(Chin et al., 2018)
two-choiceoptogeneticMadultthis study
Or65aoovipositionolfactogeneticsFadult(Chin et al., 2018)
otwo-choiceoptogeneticMadultthis study
Or67bovipositionolfactogeneticsFadult(Chin et al., 2018)
otwo-choiceoptogeneticMadultthis study
Or67doovipositionolfactogeneticsFadult(Chin et al., 2018)
+two-choiceoptogeneticMadultthis study
ovipositionolfactogeneticsFadult(Chin et al., 2018)
Or71aotwo-choiceolfactogeneticsFadult(Chin et al., 2018)
otwo-choiceoptogeneticMadultthis study
Or82aovipositionolfactogeneticsFadult(Chin et al., 2018)
otwo-choiceoptogeneticMadultthis study
Or83covipositionolfactogeneticsFadult(Chin et al., 2018)
+two-choiceoptogeneticMadultthis study
Or85aovipositionolfactogeneticsFadult(Chin et al., 2018)
otwo-choiceolfactogeneticsFadult(Chin et al., 2018)
otwo-choiceoptogeneticMadultthis study
Or85dovipositionolfactogeneticsFadult(Chin et al., 2018)
two-choiceoptogeneticMadultthis study
Or88aoovipositionolfactogeneticsFadult(Chin et al., 2018)
otwo-choiceoptogeneticMadultthis study
Or92aoovipositionolfactogeneticsFadult(Chin et al., 2018)
otwo-choiceoptogeneticMadultthis study
Gr21aoovipositionolfactogeneticsFadult(Chin et al., 2018)
/Gr63aotwo-choiceolfactogeneticsFadult(Chin et al., 2018)
two-choiceoptogeneticMadultthis study
Author response table 2
Valence results in olfactogenetics and current studies.
Olfacto ValentOlfacto Neutral
Opto Valent4 (2)5
Opto Neutral68

3. The authors own data raises potential problems with their approach. Some of the ORNs that are classified as driving aversion or attraction seem to change valence value they induce with the light intensity. For example, the authors report Or42b to drive attraction in agreement with published literature. However, at the strongest light intensity it is actually neutral. Similarly, the authors report Or85d to drive aversion. However, at the strongest light intensity it is also neutral. So, are these Ors “neutral”?

The data do not support these ORNs as neutral. The optogenetic data give robust signals in the EB screen, showing signed probabilities of +1.0 (42b) or –1.0 (85d) for some intensities. Both of these ORN-type results corroborate either the majority of published reports (42b) or the sole available published result (85d). Thus, the available data strongly support these two ORNs as being attractive and aversive, respectively.

4. The authors test a number of previously suggested linear models and find that they do not predict how two-ORN odor valence emerges from single-ORN valence. However, linear models were shown to be insufficient to predict odor valence (Badel et al., 2016). It is thus not surprising that these linear models failed.

We thank the Reviewer for their comment.

5. The authors use two databases, one of odor responses (Hallem and Carlson 2006) and one of behavioral responses (Knaden et al., 2012) along with a linear model to try and predict odor valence from ORN activity. However, as mentioned above linear models are not adequate for describing the relation between ORN activity and Odor valence. Furthermore, I think the Knaden et al., database is a wrong database to use in this context. Knaden et al., used a trap assay. In this assay, flies are captured in the trap after a single entrance to the odor source. Thus, exploratory behavior, in which flies examine the odor and then can decide to avoid it, cannot occur, and this assay is expected to be biased towards reporting odors as attractive. Indeed, this was the case in the Knaden et al., database in contrast to other published results. This database was suitable for the claim raised by Knaden et al., that looked only at the most aversive and attractive odors, but it cannot be used to try to predict any odor valence.

We respectfully disagree: we consider the title of Knaden et al., (“Spatial Representation of Odorant Valence in an Insect Brain”) to be justified. A trap assay can be viewed as a valence assay that depends on movements between the two trap entry points, around which there are odorant gradients and/or intermingling plumes. Each fly interacts with and moves in these plumes. While each fly is associated with a single outcome, the proportion of outcomes over many flies is used to estimate a preference. Using count-type data is fairly common in fly behavior, such as the olfactory T-maze which uses the endpoint locations of flies as the end of a two minute epoch. As detailed above in the discussion around wTSALE, we prefer video analysis, but nevertheless consider that endpoint and count data are valid ways to measure valence.

6. The authors used cold anesthesia just prior to loading the flies to the chambers and only 30 second acclimation following the cold anesthesia. However, cold anesthesia is known to have effects on behavior, increasing response time, reducing locomotion and reducing overall responses (just a few examples, Barron, 2000; MacMillan et al., 2017, Trannoy et al., 2015). I think most studies today try to avoid cold anesthesia just before the experiment. My concern here is that the lack of effect for most ORNs, may arise from general behavior impairment. Can the authors give a few examples from the neutral ORNs without cold anesthesia?

Please see the above discussion about cold anesthesia.

7. The authors conclude that: 1. "the majority of primary olfactory sensory neurons have neutral behavioral effects individually". This conclusion (as mentioned above) is definitely correct for the optogenetic activation, but its relevance to odor valence is questionable. Furthermore, Badel et al., 2016 already demonstrated with actual odor stimuli that “We find that the behavior is accurately predicted by a model summing normalized glomerular responses, in which each glomerulus contributes a specific, small amount to odor preference.” Thus, the novelty of the current study is not large.

We thank the Reviewer for the summary of and their viewpoint on the excellent Badel et al., study.

Their second conclusion is that "olfactory sensory neurons…participate in broad, odor-elicited ensembles with potent behavioral effects arising from complex interactions". I agree with them that olfactory coding is complex. However, they did not show any actual odor responses to support their claim, neither did they provided even one complex mechanism. I think that stating that olfaction is complex is just not enough.

Many possible neuronal mechanisms could generate the weighting shifts observed. We are reluctant to make over-reaching specific conclusions. We note that even in deep-learning neural nets that can perform impressive classification tasks—where there is complete knowledge of the ‘synaptic’ weights—there is typically little understanding how they do so (Knight, 2017; Zeiler and Fergus, 2014).

We have added a section to the Discussion that discusses possible mechanisms.

8. To my understanding the order of the β coefficients can affect the interpretation of the data. However, I could not find a reference for this in the methods. Can the authors please elaborate on this?

The question is ambiguous, answers to two possible interpretations:

– The ordering of the β weights in the regression does not affect the outcome.

– The ordering of the β weights in Figures 5 and S7 are the same.

References

Aso Y, Sitaraman D, Ichinose T, Kaun KR, Vogt K, Belliart-Guérin G, Plaçais P-Y, Robie AA, Yamagata N, Schnaitmann C, Rowell WJ, Johnston RM, Ngo T-TB, Chen N, Korff W, Nitabach MN, Heberlein U, Preat T, Branson KM, Tanimoto H, Rubin GM. 2014. Mushroom body output neurons encode valence and guide memory-based action selection in Drosophila. eLife 3:e04580.

Badel L, Ohta K, Tsuchimoto Y, Kazama H. 2016. Decoding of Context-Dependent Olfactory Behavior in Drosophila. Neuron 91:155–167.

Bernard C. 2019. Changing the Way We Report, Interpret, and Discuss Our Results to Rebuild Trust in Our Research. eNeuro. doi:10.1523/ENEURO.0259-19.2019

Button KS, Ioannidis JPA, Mokrysz C, Nosek BA, Flint J, Robinson ESJ, Munafò MR. 2013. Power failure: why small sample size undermines the reliability of neuroscience. Nat Rev Neurosci 14:365–376.

Chin SG, Maguire SE, Huoviala P, Jefferis GSXE, Potter CJ. 2018. Olfactory Neurons and Brain Centers Directing Oviposition Decisions in Drosophila. Cell Rep 24:1667–1678.

Claridge-Chang A, Assam PN. 2016. Estimation statistics should replace significance testing. Nat Methods 13:108–109.

Cleveland WS. 1994. The Elements of Graphing Data. AT&T Bell Laboratories.

Couto A, Alenius M, Dickson BJ. 2005. Molecular, anatomical, and functional organization of the Drosophila olfactory system. Curr Biol 15:1535–1547.

Cumming G, Calin-Jageman R. 2016. Introduction to the New Statistics: Estimation, Open Science, and Beyond. Routledge.

Dolan M-J, Frechter S, Bates AS, Dan C, Huoviala P, Roberts RJ, Schlegel P, Dhawan S, Tabano R, Dionne H, Christoforou C, Close K, Sutcliffe B, Giuliani B, Li F, Costa M, Ihrke G, Meissner GW, Bock DD, Aso Y, Rubin GM, Jefferis GS. 2019. Neurogenetic dissection of the Drosophila lateral horn reveals major outputs, diverse behavioural functions, and interactions with the mushroom body. eLife 8. doi:10.7554/eLife.43079

Editor. 2017. Show the dots in plots. Nature Biomedical Engineering 1:79.

Efron B. 2010. Large-Scale Inference: Empirical Bayes Methods for Estimation, Testing, and Prediction. Cambridge University Press.

Efron B, Tibshirani R. 2002. Empirical bayes methods and false discovery rates for microarrays. Genet Epidemiol 23:70–86.

Fishilevich E, Vosshall LB. 2005. Genetic and functional subdivision of the Drosophila antennal lobe. Curr Biol 15:1548–1553.

Gardner MJ, Altman DG. 1986. Confidence intervals rather than P values: estimation rather than hypothesis testing. Br Med J 292:746–750.

Ho J, Tumkaya T, Aryal S, Choi H, Claridge-Chang A. 2019. Moving beyond P values: data analysis with estimation graphics. Nat Methods. doi:10.1038/s41592-019-0470-3

Knight W. 2017. The Dark Secret at the Heart of AI. MIT Technology Review.

Koh HWL, Zhang Y, Vogel C, Choi H. 2019. EBprotV2: A Perseus Plugin for Differential Protein Abundance Analysis of Labeling-Based Quantitative Proteomics Data. J Proteome Res 18:748–752.

Root CM, Ko KI, Jafari A, Wang JW. 2011. Presynaptic facilitation by neuropeptide signaling mediates odor-driven food search. Cell 145:133–144.

Tukey JW. 1977. Exploratory Data Analysis. Addison-Wesley Publishing Company.

Zeiler MD, Fergus R. 2014. Visualizing and understanding convolutional networksComputer Vision – ECCV 2014, Lecture Notes in Computer Science. Cham: Springer International Publishing. pp. 818–833.

https://doi.org/10.7554/eLife.71238.sa2

Article and author information

Author details

  1. Tayfun Tumkaya

    1. Institute for Molecular and Cell Biology, A*STAR, Singapore, Singapore
    2. Program in Neuroscience and Behavioral Disorders, Duke NUS Graduate Medical School, Singapore, Singapore
    Contribution
    Conceptualization, Data curation, Investigation, Methodology, Resources, Software, Visualization, Writing - original draft, Writing - review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-8425-3360
  2. Safwan Burhanudin

    Institute for Molecular and Cell Biology, A*STAR, Singapore, Singapore
    Contribution
    Investigation
    Competing interests
    No competing interests declared
  3. Asghar Khalilnezhad

    Institute for Molecular and Cell Biology, A*STAR, Singapore, Singapore
    Contribution
    Investigation
    Competing interests
    No competing interests declared
  4. James Stewart

    Institute for Molecular and Cell Biology, A*STAR, Singapore, Singapore
    Contribution
    Formal analysis, Methodology, Software, Software: LabView; Methodology: Instrumentation
    Competing interests
    No competing interests declared
  5. Hyungwon Choi

    1. Institute for Molecular and Cell Biology, A*STAR, Singapore, Singapore
    2. Department of Medicine, National University of Singapore, Singapore, Singapore
    Contribution
    Formal analysis, Methodology, Software, Software: R, Supervision, Writing - review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-6687-3088
  6. Adam Claridge-Chang

    1. Institute for Molecular and Cell Biology, A*STAR, Singapore, Singapore
    2. Program in Neuroscience and Behavioral Disorders, Duke NUS Graduate Medical School, Singapore, Singapore
    3. Department of Physiology, National University of Singapore, Singapore, Singapore
    Contribution
    Conceptualization, Funding acquisition, Methodology, Project administration, Supervision, Writing - review and editing
    For correspondence
    claridge-chang.adam@duke-nus.edu.sg
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-4583-3650

Funding

Agency for Science, Technology and Research (AGA-SINGA)

  • Tayfun Tumkaya

Agency for Science, Technology and Research (Block grant)

  • Tayfun Tumkaya
  • James Stewart
  • Hyungwon Choi
  • Adam Claridge-Chang

Ministry of Education - Singapore (MOE2013-T2-2-054)

  • Tayfun Tumkaya
  • James Stewart
  • Adam Claridge-Chang

Ministry of Education - Singapore (MOE2017-T2-1-089)

  • Tayfun Tumkaya
  • James Stewart
  • Adam Claridge-Chang

Ministry of Education - Singapore (MOE-2016-T2-1-001)

  • Hyungwon Choi

National Medical Research Council (NMRC-CG-2017-M009)

  • Hyungwon Choi

Duke-NUS Medical School (Block grant)

  • Adam Claridge-Chang

Agency for Science, Technology and Research (JCO-1231AFG030)

  • James Stewart
  • Adam Claridge-Chang

Agency for Science, Technology and Research (JCO-1431AFG120)

  • James Stewart
  • Adam Claridge-Chang

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Acknowledgements

We thank Joses Ho, other members of the lab and TT’s thesis committee for helpful feedback. We thank Jessica Tamanini of Insight Editing London for a critical reading of the manuscript before submission. We thank Franz Anthony for the assay illustration. Funding: The authors were supported by grants from the Ministry of Education (grant numbers MOE2013-T2-2-054 and MOE2017-T2-1-089) awarded to ACC. TT was supported by a Singapore International Graduate Award from the A*STAR Graduate Academy. HC was supported by grant MOE-2016-T2-1-001 from the Singapore Ministry of Education and NMRC-CG-2017-M009 from the National Medical Research Council. The authors received additional support from Duke–NUS Medical School, a Biomedical Research Council block grant to the Institute of Molecular and Cell Biology, and grants from the A*STAR Joint Council Office (grant numbers 1,231AFG030 and 1,431AFG120) awarded to ACC.

Senior Editor

  1. Piali Sengupta, Brandeis University, United States

Reviewing Editor

  1. Sonia Sen, Tata Institute for Genetics and Society, India

Reviewer

  1. Matthew C Smear, University of Oregon, United States

Publication history

  1. Preprint posted: June 10, 2021 (view preprint)
  2. Received: June 13, 2021
  3. Accepted: January 17, 2022
  4. Accepted Manuscript published: January 19, 2022 (version 1)
  5. Accepted Manuscript updated: January 21, 2022 (version 2)
  6. Version of Record published: February 1, 2022 (version 3)
  7. Version of Record updated: May 16, 2022 (version 4)

Copyright

© 2022, Tumkaya et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 1,416
    Page views
  • 218
    Downloads
  • 1
    Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Open citations (links to open the citations from this article in various online reference manager services)

Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)

  1. Tayfun Tumkaya
  2. Safwan Burhanudin
  3. Asghar Khalilnezhad
  4. James Stewart
  5. Hyungwon Choi
  6. Adam Claridge-Chang
(2022)
Most primary olfactory neurons have individually neutral effects on behavior
eLife 11:e71238.
https://doi.org/10.7554/eLife.71238
  1. Further reading

Further reading

    1. Neuroscience
    Maria Cecilia Martinez, Camila Lidia Zold ... Mariano Andrés Belluscio
    Research Article

    The automatic initiation of actions can be highly functional. But occasionally these actions cannot be withheld and are released at inappropriate times, impulsively. Striatal activity has been shown to participate in the timing of action sequence initiation and it has been linked to impulsivity. Using a self-initiated task, we trained adult male rats to withhold a rewarded action sequence until a waiting time interval has elapsed. By analyzing neuronal activity we show that the striatal response preceding the initiation of the learned sequence is strongly modulated by the time subjects wait before eliciting the sequence. Interestingly, the modulation is steeper in adolescent rats, which show a strong prevalence of impulsive responses compared to adults. We hypothesize this anticipatory striatal activity reflects the animals’ subjective reward expectation, based on the elapsed waiting time, while the steeper waiting modulation in adolescence reflects age-related differences in temporal discounting, internal urgency states, or explore–exploit balance.

    1. Computational and Systems Biology
    2. Neuroscience
    Sergio Oscar Verduzco-Flores, Erik De Schutter
    Research Article Updated

    How dynamic interactions between nervous system regions in mammals performs online motor control remains an unsolved problem. In this paper, we show that feedback control is a simple, yet powerful way to understand the neural dynamics of sensorimotor control. We make our case using a minimal model comprising spinal cord, sensory and motor cortex, coupled by long connections that are plastic. It succeeds in learning how to perform reaching movements of a planar arm with 6 muscles in several directions from scratch. The model satisfies biological plausibility constraints, like neural implementation, transmission delays, local synaptic learning and continuous online learning. Using differential Hebbian plasticity the model can go from motor babbling to reaching arbitrary targets in less than 10 min of in silico time. Moreover, independently of the learning mechanism, properly configured feedback control has many emergent properties: neural populations in motor cortex show directional tuning and oscillatory dynamics, the spinal cord creates convergent force fields that add linearly, and movements are ataxic (as in a motor system without a cerebellum).