Olfactory receptor neurons generate multiple response motifs, increasing coding space dimensionality

  1. Brian Kim
  2. Seth Haney
  3. Ana P Milan
  4. Shruti Joshi
  5. Zane Aldworth
  6. Nikolai Rulkov
  7. Alexander T Kim
  8. Maxim Bazhenov  Is a corresponding author
  9. Mark A Stopfer  Is a corresponding author
  1. Eunice Kennedy Shriver National Institute of Child Health and Human Development (NICHD), National Institutes of Health (NIH), United States
  2. Brown University - National Institutes of Health Graduate Partnership Program, United States
  3. Department of Medicine, University of California, San Diego, United States
  4. Department of Clinical Neurophysiology and MEG Center, Amsterdam Neuroscience, Vrije Universiteit Amsterdam, Netherlands
  5. Biocircuits Institute, University of California, San Diego, United States

Abstract

Odorants binding to olfactory receptor neurons (ORNs) trigger bursts of action potentials, providing the brain with its only experience of the olfactory environment. Our recordings made in vivo from locust ORNs showed that odor-elicited firing patterns comprise four distinct response motifs, each defined by a reliable temporal profile. Different odorants could elicit different response motifs from a given ORN, a property we term motif switching. Further, each motif undergoes its own form of sensory adaptation when activated by repeated plume-like odor pulses. A computational model constrained by our recordings revealed that organizing responses into multiple motifs provides substantial benefits for classifying odors and processing complex odor plumes: each motif contributes uniquely to encode the plume’s composition and structure. Multiple motifs and motif switching further improve odor classification by expanding coding dimensionality. Our model demonstrated that these response features could provide benefits for olfactory navigation, including determining the distance to an odor source.

Editor's evaluation

This important work describes the temporal mechanisms of odor coding in the olfactory neurons of the locust. The supporting evidence is compelling and based on extensive experimental and computational analyses. This work will be of interest to sensory neuroscientists.

https://doi.org/10.7554/eLife.79152.sa0

Introduction

Odors provide many types of important information about the environment, and are characterized by their chemical compositions and concentrations. Unlike vision or audition, which, in most basic form, can be described with the two dimensions of intensity and frequency, the tens or hundreds of thousands of detectable odorant molecules come in many different shapes, sizes, and charge distributions, an assortment of attributes requiring a high-dimensional description. Also, odorants often reach detectors as chaotic and turbulent plumes comprised of odorized pulses separated by expanses of clean media, air or water. The structures of odor pulses within a plume are shaped by many factors, including the medium’s speed, environmental features such as hills, trees, or buildings, and the detector’s distance from the source. Therefore, the timing of odorants reaching a sensor can convey important information about the surroundings. To make use of this information, olfactory systems must generate high-dimensional representations of both chemical and timing information in a format usable by downstream circuits. This task is challenging, yet animals rely on the results for survival. Notably, olfaction is achieved by relatively few layers of neurons. Further, similar anatomical structures and physiological mechanisms for processing odors appear in widely divergent species, suggesting evolution has converged upon an optimized set of solutions to olfactory challenges.

In animals, the frontline encoders of odors are arrays of olfactory receptor neurons (ORNs), each expressing one of many types of olfactory receptor protein. The number of receptor types varies with species, from 119 in the desert locust (Pregitzer et al., 2017) to ~400 in humans (Malnic et al., 2004; Trimmer et al., 2019), and over 1000 in mice (Zhang and Firestein, 2002). Individual ORNs may be tuned narrowly or broadly , and typically, many types of ORNs respond to any given odor (Hallem and Carlson, 2006). Together, their responses give rise to high-dimensional, combinatorial representations that encode information about the environment. Because ORNs provide all the information available to the animal about the olfactory environment, it is important to understand the diversity and complexity of their responses.

To better characterize ORN response dynamics, we made extracellular recordings from the locust antenna while presenting odor pulses individually, or in regular trains, or as realistic, chaotic odor plumes. We found that ORNs can generate four distinct types of firing pattern motifs. While ORNs respond reliably with a given motif to a given odor across multiple trials, we also found that a single ORN can respond to different odors with different motifs, a phenomenon we term ‘motif switching.’ Further, when elicited by repeated odor pulses, the four response motifs undergo distinct forms of adaptation. With computational modeling, we found these novel aspects of stimulus encoding contribute substantially to classifying odors and to processing complex natural odor plumes, including extracting odor-invariant information about distance to the odor’s source. Together, these results reveal new ways ORNs utilize the temporal domain to expand their coding space dimensionality.

Results

ORNs can respond to odors with four distinct firing patterns

To characterize the response properties of ORNs, we made extracellular recordings from antennae of intact locusts by placing electrodes against the base of sensilla while delivering pulses of odors well-separated from each other in time (40 animals, 62 sensilla, 198 odor-ORN pairs) (Figure 1A). Because sensilla contain multiple ORNs (Ochieng et al., 1998), we then analyzed the recorded waveforms with a spike sorting algorithm to assign odor-elicited spikes to individual ORNs (Figure 1—figure supplement 1). We recorded mainly from trichoid sensilla because they contain small numbers of ORNs (3–5, Ochieng et al., 1998) making spike sorting tractable, but we found unsorted, population activity recorded from other types of sensilla yielded results consistent with the responses of sorted ORNs, including prominent onset and offset activity (Figure 1—figure supplement 2). All results other than those shown in Figure 1—figure supplement 2 are based on recordings from trichoid sensilla.

Figure 1 with 4 supplements see all
Olfactory receptor neuron (ORN) responses cluster into four distinct motifs.

(A) Extracellular recordings were made with a blunt quartz electrode placed on the base of the sensillum. (B) Raster plot shows ORN spikes elicited by 1 s odor pulse (black bar at top). Horizontal lines separate ORN-odor pairs (five trials each). A hierarchical clustering algorithm (see ‘Materials and methods’) grouped these responses into four distinct motifs. (C) Histograms show the four motifs (excitatory, delayed, offset, and inhibitory) reliably elicited by 1 s and 4 s odor pulses. Five trials each; bold lines: means; shading: SEM.

We used an unsupervised hierarchical clustering method to group odor-elicited responses, pooled across odor-ORN combinations, into categories (Figure 1B). Notably, we found the responses, particularly to longer stimuli, did not fall into a continuum of patterns, but rather clustered into distinct motifs; a boot-strap analysis indicated that four motifs provided the best clustering of responses (see ‘Materials and methods,’ Figure 1—figure supplement 3). We termed these four response motifs excitatory, delayed, offset, and inhibitory (Figure 1C, Figure 1—figure supplement 4; see ‘Materials and methods’). Excitatory responses featured a sharp increase in firing rate at the immediate onset of the odor and decayed rapidly back to baseline even as the stimulus persisted. Delayed responses featured a slower increase in firing rate and a gradual decay back to baseline throughout an odor pulse. Inhibitory responses featured a sharp decrease in firing rate during an odor presentation that returned to the baseline level after the odor was removed. Offset responses were also inhibited during odor presentation but, upon the removal of odor, immediately showed an increase in firing rate exceeding the baseline level. Also, 1 s and 4 s odor pulses elicited similar results (Figure 1C).

A given ORN can generate more than one response motif

Notably, a given ORN could respond with different patterns of spiking when presented with different odors. For example, from the same ORN, a brief pulse of pentyl acetate reliably elicited a delayed response motif, but an identical pulse of cyclohexanol reliably elicited an excitatory response motif (Figure 2A). This phenomenon, termed response motif switching, was not rare, occurring with probabilities ranging from 0.31 to 0.66 (mean = 0.38) for different pairs of odors (see Figure 2B–D). By contrast, we found very little switching across repeated trials of the same odor (0.096), and an intermediate amount of switching across concentrations of the same odor (0.25, see Figure 2—figure supplement 1). Thus, motif switching appears to convey information about odorant identity and concentration in a way that is robust to trial-to-trial variation.

Figure 2 with 1 supplement see all
Different odors can elicit different response motifs from a given olfactory receptor neuron (ORN).

(A) A single example ORN responds to two different odorants with different motifs. Five trials; bold lines: mean firing rate; shading: SEM. (B) Raster plots show spiking responses of 41 ORNs to 1 s (black bar) pulses of hexanol and cyclohexanol. Responses to both odors are sorted by motifs elicited by hexanol. (C) Conditional probability of an ORN producing a specific motif in response to hexanol given its response motif to octanol. (D) Switching probabilities varied with odor pairs (numbers at bottoms of bars: numbers of tested ORNs).

Different ORN response motifs have different adaptation profiles

ORN responses adapt when odor pulses are lengthy or repeated rapidly (Barrozo and Kaissling, 2002; Bau et al., 2002; Lemon and Getz, 1997; Marion-Poll and Tobin, 1992). To characterize response adaptation properties of each ORN response motif, we delivered repeated 200 ms odor pulses at a range of inter-pulse intervals (IPIs) (see ‘Materials and methods’; Figure 3A, Figure 3—figure supplement 1, Table 1). Notably, each response motif showed a distinct adaptation profile (Figure 3B; see ‘Materials and methods’ for details of analysis). The peak of excitatory motif responses decreased significantly over a train of odor pulses presented at brief, 0.5 s IPI. However, under the same stimulus conditions, the peaks of offset motif responses increased significantly, even with IPIs as long as 1.0 s; the average final response to an IPI of 0.5 s was threefold larger than the first. The delayed motif tended to increase after the first odor pulse, but this trend fell just short of statistical significance (Table 1). The inhibitory motif appeared to show no adaptation.

Figure 3 with 1 supplement see all
Each olfactory receptor neuron (ORN) response motif has a distinct adaptation profile.

(A) Responses of ORNs, grouped by motif, to odors pulsed at different inter-pulse intervals (IPIs). Peaks for different motifs and pulses were measured as maximum absolute change from baseline within detection windows (shaded areas). (B) Adaptation characteristics of excitatory, delayed, offset, and inhibitory response motif to 10 pulses for each IPI. Left: response motifs; Right: normalized response change from baseline. *Statistically significant changes elicited by odor pulses delivered at 0.50 s IPI. Excitatory motif responses significantly decreased; offset motif responses significantly increased; delayed motif responses modestly increased; and inhibitory motif responses did not change. See Table 1 for statistical tests.

Table 1
Results of ANOVA tests for sensory adaptation experiments shown in Figure 3B and Figure 3—figure supplement 1.
Excitatory motifDelayed motif
IPIs (s)DfF-valuep-valueIPIs (s)DfF-valuep-value
0.501, 93.110.0020.501, 91.980.052
0.751, 90.510.8650.751, 90.880.549
1.001, 90.800.6141.001, 90.380.938
1.251, 90.840.5761.251, 90.730.675
1.751, 90.380.9411.751, 90.320.966
5.001, 90.670.7345.001, 91.410.210
Offset motifInhibitory motif
IPIs (s)DfF-valuep-valueIPIs (s)DfF-valuep-value
0.501, 933.061.484E-370.501, 90.780.634
0.751, 96.594.663E-080.751, 91.050.406
1.001, 94.294.650E-051.001, 91.460.162
1.251, 92.940.0031.251, 91.210.293
1.751, 91.320.2291.751, 90.660.747
5.001, 91.760.0775.001, 91.570.127
  1. IPI: inter-pulse interval.

Computational modeling of ORN response motifs

The response motifs we observed in vivo seemed likely to contribute to the processing of olfactory information. To test this idea, we designed a computational model based on observations made in vivo. Using the responses of ORNs to 4 s pulses of odor as templates, we constructed models of individual neurons to represent each response motif (see ‘Materials and methods’; Bazhenov et al., 2008; Rulkov, 2002; Rulkov et al., 2004); this approach allowed us to model a biologically realistic population of 10,000 ORNs. The models included realistic levels of input noise and variations in baseline activity. Individual trials were generated by creating different random instantiations for the input noise (see ‘Materials and methods’). The magnitude of each odor-specific ORN response was determined by the angle between two high-dimensional vectors: an ORN-specific chemical selectivity vector, VORN , and an odor-specific characteristic vector, Vodor (see ‘Materials and methods’), so similar odors elicited similar responses. As desired, model ORN responses to a single odor pulse (Figure 4A and B) and trains of pulses (Figure 4C) matched those found in vivo (see Figures 1 and 3). The model’s responses to pulse trains were set to adapt in motif-specific fashions as found in vivo, with the excitatory motif responses decreasing and the offset motif responses increasing with each pulse (Figure 4C). As desired, our model provided an accurate simulation of ORN responses observed in vivo (Figure 4—figure supplement 1 provides a quantitative comparison of response peaks and latencies in vivo and in the model).

Figure 4 with 1 supplement see all
Computational model simulates response motifs.

(A) Raster plots of simulated olfactory receptor neuron (ORN) responses illustrate each motif. Rows: different trials; horizontal lines separate responses of different ORNs. Black bar: odor pulse. (B) Firing rates averaged across trials and ORNs by motif. Bold line: mean; shading: SEM; compare to Figure 1C. (C) Model driven at 0.50 s inter-pulse interval (IPI) simulates adaptation profiles of each motif; firing rate averaged across ORNs within a motif; compare to Figure 3A. (D) Principal component analysis of excitatory only ORN responses (blue) or all response motifs (red) shows the inclusion of multiple motifs adds to the complexity of the responses. (E) The model can vary response magnitude and motif independently; responses of a single model ORN to three different odors are shown.

We used this model to determine how the diversity of ORN motifs contributes to the complexity of the odor response. In principle, more complex odor responses allow for higher coding dimensionality. We quantified response complexity by performing principal component analysis (PCA) and evaluating the number of components needed to explain the variance across the entire population; a more complex odor response would require more principal components to explain the same amount of variance. We compared responses encoded only by excitatory motifs to responses including all motifs using numbers of ORNs matched between these two cases (N = 1374 each). When only excitatory motifs were included in the analysis, the first component alone explained nearly 30% of the variance and the second component added only ~1%. In contrast, when all motifs were included, the first component only explained ~15% of variance and 27 principal components were required to explain 30% of the variance (Figure 4D). Thus, the existence of four response motifs dramatically expanded the dimensionality of the odor representation.

Response motif switching contributes substantially to odor classification

With our model, we could independently vary odor-elicited response motifs and response magnitudes (Figure 4E), allowing us to evaluate the extent to which motif switching benefited odor classification in a way that cannot be tested in vivo. Thus, we simulated a realistically large number of ORNs (10,000) and compared the relative success of classifying two different odors (odor 1 and odor 2) with three different versions of our model in which we systematically varied the probability of motif switching. Model version 1: the probability of switching response motif when switching from odor 1 to odor 2 was 0%; version 2: 10%; version 3: 50%. We found that the model versions that simulated higher motif switching probability made it easier to distinguish these two similar odors. Figure 5A shows the responses of 41 example ORNs. Trajectories of the responses of the 10,000 ORN population over time plotted in PCA-reduced space (see ‘Materials and methods’) increasingly separated as the probability of motif switching increased (Figure 5B), demonstrating that motif switching made the ORN population responses to the two odors more different from each other. When we independently varied the similarity of the odors and the probability of motif switching (see Figure 4E), support vector machine (SVM)-based classification of the ORN responses showed classification accuracy improved as motif switching probability increased for all degrees of odor similarity (Figure 5C). With 4 s odor pulses, even the lowest tested probability of motif switching (10%) substantially improved odor classification to the point that it made a difficult classification task (when test odors were very similar) as successful as an easy classification task (when odors were very different). Together, these results demonstrate that motif switching, even if infrequent, can contribute substantially to successful odor classification.

Computational model shows response motif switching substantially improves odor classification.

(A) Simulated olfactory receptor neuron (ORN) spiking illustrates different motif switching probabilities. Odors 1 and 2 are similar (see ‘Materials and methods’). Each ORN response is sorted by motifs elicited by odor 1. Raster plots show the responses to odor 2 become increasingly different from responses to odor 1 as motif switching probability increases. (B) ORN odor-elicited response trajectories in reduced principal component analysis (PCA) space show motif switching increases the separation between responses to similar Odors 1 and 2; response to odor 1 (blue) is the same in each panel; response to odor 2 (red) changes with switching probability. (C) Odor classification success as a function of odor similarity and motif switching probability for 1 s (top) and 4 s (bottom) stimulus pulses; even low switching probabilities improve classification performance; darker shading indicates lower classification accuracy. Odor similarity is quantified by angles (degrees) between odor vectors (see ‘Materials and methods’).

Response motifs represent complex odor plume features

The four response motifs of ORNs differed in their timing and adaptation properties, suggesting that they could contribute to encoding temporal features of olfactory stimuli. In natural environments, odorant molecules are usually arranged by a turbulent medium, air or water, into temporally complex plumes. We used the statistics of open-field measurements of real plumes (Murlis et al., 2000) to generate realistic, repeatable artificial plumes that included a distribution of ‘burst’ lengths (see ‘Materials and methods’), and delivered them to the antennae of intact locusts while recording responses from ORNs in vivo (Figure 6A, top and middle). We also simulated the same plumes in our model (Figure 6A, bottom). The motif-dependent adaptation we had observed in vivo (Figure 3A) and reproduced in silico (Figure 4C) introduced motif-specific sensitivities to stimulus history. Notably, long bursts within a plume led to decreased excitatory motif responses but increased delayed motif responses (e.g., Figure 6A, ~9–12 s). On the other hand, short bursts led to increased excitatory motif responses but did not change the delayed motif responses (e.g., Figure 6A, ~5 s). This motif-specific history dependence appeared likely to contribute to encoding temporally complex stimuli.

Figure 6 with 1 supplement see all
Responses to realistic odor plume stimulus.

(A) Top: odor plume generated by opening and closing an olfactometer valve, measured by a photoionization detector (PID), and delivered to locust antennae in vivo. Plume represents sampling 2.5 m from odor source. Middle: responses of olfactory receptor neurons (ORNs) recorded in vivo; bottom: responses of model ORNs. (B) Linear–nonlinear prediction of excitatory-type ORN activity in vivo. The first 16 s of 40 s odor plumes were used for training and the next 16 s were used for testing. To assess the agreement, the test stimulus (left) was convolved with the learned filter and the result was passed to the learned threshold function (middle) to generate the predicted firing rate (right). Shading denotes a 95% confidence interval of prediction (see ‘Materials and methods’). Stimulus correlations in naturalistic stimuli were accounted for (see ‘Materials and methods’). (C) Averages of learned filters across all ORNs within a response motif in vivo (top) and in the model (bottom). Delay indicates time since the stimulus arrived at the antenna. (D) Principal component analysis of filters in the delay dimension. Filters of the same response motif cluster together. Top: filters colored by motif type. Bottom: two principal components in the analysis.

To quantify how each motif contributes to processing time-varying odor plumes, we applied a commonly used linear–nonlinear analysis scheme (Butts et al., 2007; Geffen et al., 2009). In this approach, the true firing rate of the neuron, r(t), is predicted by convolving a linear filter, f(t), with the time-varying stimulus, s(t), and then the convolved product is thresholded with a nonlinear function, g(x). This filter is useful because the waveform of the linear filter, f(t), derived from this method precisely describes the sensitivity of the neuron to the history of the stimulus. To compute these filters, briefly, we deconvolved the stimulus, s(t), from the firing rate data, r(t), collected from ORNs. This approach allowed us to calculate filters for each ORN response in vivo and in the model (Figure 6B illustrates this approach; see ‘Materials and methods’). As desired, these reconstructions accurately reproduced the firing rates of the ORNs from their trained filters on data not used for training.

As desired, filters generated from the model and from responses recorded in vivo matched closely (Figure 6C, Figure 6—figure supplement 1). We observed that different response motifs generated distinct filter waveforms, describing the different sensitivities of each motif to the temporal features of the odor plume. For example, filters for excitatory motifs showed that the ORNs generating them were most responsive ~0.4 s after odor filaments arrived (Figure 6C). To examine the stimulus-history dependence of the ORNs as a population, we embedded the filter waveforms of all the individual ORNs into an N-dimensional space (where N was the number of sample points for each filter) and used PCA to reduce each filter waveform into a single point in a two-dimensional space. We found that ORN filters clustered by motif in this space. Inhibition motifs (offset and inhibitory) and excitation motifs formed two ends of the principal filter axis, with the second dimension separating the delay motif from the other three (Figure 6C and D). This analysis revealed that each ORN response motif contributes uniquely to the olfactory system’s representation of the odor plume’s complex temporal structure.

Response motif diversity increases sensitivity to the distance to odor source

Our finding that different response motifs are sensitive to the different temporal features of a complex natural stimulus raised the possibility that this information could be used for ecologically relevant tasks that rely on assessing the timing of a stimulus. Published reports indicate a reliable relationship between the frequency of filaments within an odor plume and the distance to its source (Murlis et al., 2000). Further, evidence from walking fruit flies suggests that they navigate toward an odor source in response to both the overall frequency and inter-pulse intervals of their encounters with the odor plume (Demir et al., 2020; Álvarez-Salvado et al., 2018; Jayaram et al., 2022). To test whether, in principle, the diversity of ORN response motifs could help an animal determine the distance to an odor source (a key aspect of navigation), we simulated odor plumes characteristic of an odorant located 2.5 m, 5 m, 10 m, or 20 m away (Murlis et al., 2000). We used our model to generate responses of ORNs to these distance-specific plumes (Figure 7A; see ‘Materials and methods’).

Response motifs encode information needed to determine distance to odor plume source.

(A) Model olfactory receptor neuron (ORN) responses, averaged by motif, to four distance-dependent plumes (see ‘Materials and methods’). Green bars: simulated olfactometer valve open times. (B) Depiction of stimulus classification task. Ten realizations of plume stimuli (trials) are shown for each distance. Plumes are colored by distance to source. The goal of the classification task is to correctly identify all plume stimuli of the same distance as belonging to the same group. (C) The success of distance-dependent plume classification. Each group of points represents 25 differently trained classifiers only including a subset of indicated response motifs. Vertical bars separate groups into statistically distinct sets, where any group within a set is statistically different from all groups outside this set (one-way ANOVA F(14,374) = 1830.2, p=0; see Figure 3—figure supplement 1). The x-axis labels denote which motifs are included in each analysis group: D, delayed; E, excitatory; I, inhibitory; O, offset. (D) Scatter plots show firing activity averaged over the entire 15 s window by motif. (E) Correlation between each motif and the excitatory motif calculated using different sample durations (top). Classification success for different sample durations using responses from all motifs (bottom).

We trained SVM classifiers to discriminate among the distances based on responses of the model ORNs (Figure 7B). We constructed different classifiers using only subsets of ORNs that respond with specific motifs, testing all possible combinations to assess their separate and combined contributions to classification success. We found that classifiers that included the excitatory and delayed motifs were most successful in determining distance to the source (Figure 7C). Excitatory and delayed motif responses were less correlated with each other than the other two possible combinations (Figure 7D), indicating that the excitatory and delayed motif responses encoded different information about stimulus timing. This result was robust to the duration of the response used to calculate both correlation and classification (Figure 7E) with longer durations increasing classification success. Together, these results show the different response motifs each extract distinct information that could be used to navigate toward odor sources.

Discussion

Spiking patterns of olfactory neurons have been shown to convey information about the chemical composition of odors. In insects, projection neurons (PNs) generate spiking patterns that change with the identity and concentration of the odor. These complex dynamics arise in part from network interactions among PNs and local inhibitory neurons within the antennal lobe driven by heterogeneities in the temporal structures of odor-elicited ORN responses (Martelli and Fiala, 2019; Raman et al., 2010). Deeper in the brain, Kenyon cells (KCs), driven by PNs, change the timing of their spiking responses when the eliciting odor changes (Gupta and Stopfer, 2014), and even subtle changes in the timing of KC spiking can alter the responses of their follower neurons (Gupta and Stopfer, 2014). Because these information-bearing spiking patterns throughout the brain originate with the combinatorial spiking of populations of ORNs, it is important to understand the properties of these peripheral responses.

In natural settings, ORNs typically encounter odor plumes characterized by complex temporal structure. Turbulent flowing media churn the headspaces above odor sources into chaotic plumes (Jacob et al., 2017; Levakova et al., 2018; Murlis et al., 1992), and active sampling movements such as sweeping antennae or sniffing generate additional temporal structure (Huston et al., 2015). Animals can make use of timing information within odor plumes. Mice, for example, can use the separate timings of multiple odors comingled in plumes to isolate their separate sources (Ackels et al., 2021), and moths and mosquitoes will fly upwind only when attractive odors are appropriately pulsed (Baker et al., 1985; Geier et al., 1999). Although patterned activity originating in the olfactory environment may appear to conflict with patterned activity generated by olfactory neural circuits, combinatorial codes within the locust antennal lobes and mushroom bodies have been shown to disambiguate the two, representing not only the chemical identity and concentration of odors, but also their delivery timing (Brown et al., 2005).

We found that ORNs respond to odors with a distinct set of four motifs that can include immediate and delayed periods of spiking and periods of inhibition. Our computational analysis further revealed how each response motif contributes to processing complex temporal stimuli such as odor plumes. Excitatory and offset motifs formed two ends of a single encoding dimension, and the delayed motif formed another dimension. Thus, each motif, with its own sensitivity to specific temporal features of a stimulus, contributes in different ways to extracting timing information, and together provide a more thorough description of the stimulus to downstream neurons than that of any single motif.

Our experiments did not allow us to determine the molecular or neural mechanisms that generate each of the four response motifs. Earlier work established that the intrinsic dynamic properties of odorants, described as ‘fast’ or ‘slow,’ can contribute to variations in the timing of ORN responses (Martelli et al., 2013; Su et al., 2011). However, our experiments ruled out the possibility that intrinsic odorant dynamics underlie the response motifs we describe here. First, across our extensive dataset, all odors could elicit all four response motifs; second, photoionization detector recordings of our odor presentations all revealed ‘fast’ dynamics (not shown). It seems likely that ‘slow’ odors would elicit concentration-dependent elaborations in the response motifs we observed. In future work, it will be interesting to investigate the ways intrinsic odor dynamics interact with ORN response motifs. We predict such interactions would further increase ORN response dimensionality. Overall, our results suggest that these responses, together and in combination, play fundamental roles in olfactory processing.

We were surprised to find that a given ORN could generate different response motifs when stimulated by different odors. ORNs responded reliably and robustly with the same motif across nearly all trials of the same odor, but when the odor changed, the response motif sometimes changed. Motif switching was not rare; under some conditions in our sample, 66% of tested ORNs switched motifs when odors were switched. Even relatively subtle changes in odorant chemistry (e.g., hexanol to cyclohexanol), or changes in the concentration of an odor, could evoke changes in the response motifs of ORNs (e.g., immediate to delayed spiking), leading to dramatic differences in the timing of the odor response. Our computational model revealed that even infrequent motif switching improves odor classification by increasing the distance between the spatiotemporal representations of odors. The improvement was substantial, comparable to that obtained by changing test odor pairs from the most similar to the most different. Our study demonstrates a new property of ORNs: they can signal changes in an odor stimulus through discrete step changes in response motif as well as through continuous shifts in response magnitude. Motif switching thus appears to increase contrast between odors. Our experiments did not allow us to test whether all ORNs are capable of motif switching, or what, if any, organizational principle determines when a given ORN will switch motifs. But because some odor pairs were more likely than others to elicit motif switching, we speculate that motif switching may follow a logic that evolved to help discriminate ecologically relevant but chemically similar odors (Su et al., 2012). For these studies, we used odorants known to be ecologically relevant to locusts, including several found in the head space of wheat grass. Future experiments with larger sets of odorants, including blends or locust pheromones like 4-vinylanisole (4VA) and phenylacetonitrile (PAN), may help clarify the logic of motif switching.

Our work did not address the neural mechanism underlying motif switching. Odorant receptor proteins in ORNs may, for example, couple with multiple transduction pathways. Ephaptic coupling between ORNs cohabiting a sensillum could, in principle, allow one spiking ORN to change the firing pattern of its neighbor (Su et al., 2012; Zhang et al., 2019) in an odor-specific fashion, but our sensillar recordings did not reveal the anticorrelated spiking in pairs of ORNs expected of such an interaction (see Figure 1—figure supplement 1). Synaptic connections between ORNs have been identified in a connectomics analysis of the fly brain (Schlegel et al., 2021), and similar connectivity has been suggested to exist among gustatory receptor neurons (Chu et al., 2014). Such connectivity could potentially enable network interactions to generate multiple and switching response motifs, even at the sensory periphery.

Repeated stimuli elicit sensory adaptation, a form of short-term plasticity that encodes recent stimulus history. In agreement with prior work (Barrozo and Kaissling, 2002; Bau et al., 2002; Lemon and Getz, 1997; Marion-Poll and Tobin, 1992), we found ORNs adapted when they were activated by repeated pulses of an odor. However, we also found, surprisingly, that different response motifs adapted in distinct ways, including decreasing or increasing response intensity across a series of odor presentations. This response diversity was undetectable by previous studies employing electroantennogram (EAG) recordings, which sum the activity of populations of ORNs (Huston et al., 2015). The individual odor filaments comprising natural odor plumes vary from brief to long, each evoking, respectively, lesser or greater amounts of sensory adaptation in ORNs. Thus, adaptation adds contrast to the olfactory system’s description of the plume’s temporal structure. Additionally, each response motif adapts differently, enriching the description by highlighting different time-varying aspects of the plume. To analyze these responses, we used our recordings made in vivo to generate filters that provide optimal descriptions of each ORN’s sensitivity to changes in the stimulus with respect to time, including motif-specific responses and adaptation. These results show that different response motifs highlight different temporal features of a stimulus and, together, the combination of distinct response motifs with different forms of adaptation improves representations of temporal features in complex stimuli, including plume onset, offset, and the spacing between odor filaments. While our analysis did not investigate the mechanisms underlying the different forms of adaptation we observed, recent studies of ORNs in Drosophila reveal calcium dynamics play a role in adaptation (Martelli et al., 2013; Martelli and Fiala, 2019).

We examined whether timing information available in the multiple and switching response motifs in ORNs could provide practical benefits to animals. It has been shown that the structure of natural odor plumes contains information about the environment. For example, the length and spacing of odor filaments within plumes vary reliably with the distance to the odor source (Murlis et al., 2000). Further, behavioral (Demir et al., 2020; Álvarez-Salvado et al., 2018) and modeling studies (Jayaram et al., 2022) have shown that the first layer of the insect olfactory system extracts information sufficient for navigation from the temporal statistics of plumes. Using our model, we evaluated how the multiple response motifs generated by the ORNs could contribute to extract this information. We classified responses of the ORN population to odor plumes designed to simulate different distances to the odor source. We found that the distance to the odor source could be determined readily from information contained within the ORN responses. Notably, providing multiple ORN response motifs significantly improved the ability to classify source distance. Our model predicted that combining excitatory and delayed motifs provided the most successful way to determine distance to the source, suggesting that these motifs extract complimentary information about the odor statistics most needed to characterize distance. Other motifs likely contribute to encoding different odor features. Earlier work performed in vivo and with computational models has investigated the ways insect olfactory systems encode distance-dependent plume statistics (e.g., Jacob et al., 2017; Levakova et al., 2018). A recent modeling study found that adaptation of ORN responses contributes substantially to the sparsening of responses downstream, and to the sampling of the statistics of odor encounters that could aid navigation to food sources (Rapp and Nawrot, 2020). A common theme of these findings is that olfactory systems include mechanisms to extract information needed for navigation.

Earlier work has shown that odor-elicited responses of ORNs can be more complex than simple bursts of spikes that track the stimulus. For example, locust ORNs were shown to generate heterogeneous responses (Raman et al., 2010) and, in Drosophila, ORNs have been shown to produce either excitatory or inhibitory responses to odorants (de Bruyne et al., 2001; Su et al., 2012). Similar responses have also been observed in the ORNs of aquatic organisms, including lobster (Bobkov et al., 2012), where a subpopulation of lobster ORNs has been shown to rhythmically burst in a fashion essential for determining the edges of a turbulent plume during odor tracking (Michaelis et al., 2020). Recent results in Caenorhabditis elegans show that hedonic changes in olfactory context can change the responses of an ORN (Khan et al., 2021). However, to our knowledge, our results provide the first evidence that ORNs can generate a well-defined set of discrete temporally structured response motifs for different odors.

ORNs provide the olfactory system’s first of several stages of signal processing. In insects, ORNs pass information to the antennal lobe, where spiking patterns originating in the periphery drive further processing in networks of local and projection neurons, leading to richer, higher dimensional olfactory representations (Geffen et al., 2009; Raman et al., 2010). Our study did not address how multiple and switching response motifs in ORNs affect the responses of downstream neurons such as LNs and PNs. Our earlier work established that antennal lobe circuitry generates high-dimensional, complex temporally structured responses, but only when it is driven by input with heterogeneous timing from ORNs (Raman et al., 2010). The patterned ORN responses we report here likely contribute substantially to the variety of this input, and thus to the complexity and high dimensionality of PN responses. In other species, how ORN response patterns are utilized downstream may depend on species-specific variations in connectivity between ORNs and the antennal lobe and its glomeruli.

Although our model is phenomenological, it was designed to simulate properties found in real neurons that can be traced to specific ion channels (Komarov et al., 2018). Also, earlier work including biophysically realistic models has been shown to generate spiking behaviors comparable to the response motifs we identified (e.g., Betkiewicz et al., 2020; Farkhooi et al., 2013). Realistic biophysical models may point toward a more granular and specific mechanistic understanding of the responses we observed in ORNs.

In summary, we found that the responses of ORNs in the locust are organized into a discrete small set of spiking motifs, and that different odors can elicit different motifs from a given ORN. Multiple response motifs increase the dimensionality of the neural code for odors, and motif switching dramatically amplifies distinctions between similar chemical inputs, allowing for more successful classification. Each motif has its own adaptation profile, contributing to the encoding of temporal features of odor stimuli. These features provide benefits for olfactory tracking, such as determining distance to an odor source. Understanding and applying these processing features to stimulus classifying or tracking devices could lead to advancements in artificial intelligence and robotics.

Materials and methods

Electrophysiology

Request a detailed protocol

Recordings were made from 132 young adult locusts (Schistocerca americana) raised in a crowded colony. For experiments, locusts were secured with tape and wax into a Petri dish with one antenna exposed. An AgCl reference electrode was placed into the eye contralateral from the recording antenna. A saline-filled glass electrode (~10 µm, ~10 MΩ) was placed into the exposed antenna as a differential electrode. An identical electrode attached to a head stage (Axon instruments, X0.1LU) was placed at the base of a sensillum to acquire signals through an amplifier (Axon Instruments, Axoclamp 2b). The acquired signal was amplified ×3000 (Brownlee Precision, Model 440) and digitally sampled (National Instruments, USB-6212 and USB-6215 DAQ; Labview Software) at 10 kHz. Extracellular recordings from ORNs were made from different types of sensilla (sensilla chaetica and sensilla trichodea). Spike sorting was done offline with a whole wave algorithm (Pouzat et al., 2002) implemented in MATLAB (MathWorks). Sorted recordings were only included in the study if they were estimated to have <5% false positives and <5% false negatives during stimulation (Hill et al., 2011).

Odor stimulation

Request a detailed protocol

The odor delivery system has been described previously (Brown et al., 2005). Briefly, a constant stream of dried and activated carbon-filtered air (0.9 l min–1) was directed to the antenna through a plastic tube (6.5 mm inner diameter). A vacuum funnel (7 cm) was placed behind the animal to clear the odor space. Odorized air (0.5 L min–1) was delivered by injecting air with a pneumatic pump (Reliable Pneumatic PicoPump, World Precision Instruments) into the head space of a 100 ml glass bottle containing odorant solutions diluted in mineral oil (JT Baker) to various concentrations, and then into the constant stream. The odorant chemicals (Sigma) used in this study are components of the locust diet, wheat grass: 1-octanol (OCT), 1-hexanol (HEX0.1, HEX1, HEX, HEX100; 0.1, 1, 10, and 100% by volume, respectively), and cyclohexanol (CYC). Pentyl acetate (PET), a naturally occurring chemical with an apple-like scent, was used as well.

Odorants other than 1-hexanol solutions were diluted to 10% by volume. The artificial plume stimulus was based on the burst length and burst return parameters derived from real odor plumes measured outdoors; see the ‘Distance-based artificial plume generation’ section (Aldworth and Stopfer, 2015; Murlis et al., 2000).

Response motif clustering

Request a detailed protocol

We binned the spike-sorted responses of individual ORNs to 1 s odor pulses into 50 ms windows. Only responsive cells, those whose stimulus-evoked activity exceeded 2 standard deviations from the pre-stimulus mean, were included in the clustering analysis. A hierarchical cluster tree was created through unsupervised agglomerative hierarchical clustering (‘linkage’ function in MATLAB) using the Euclidean distance metric and the Ward agglomeration method. Responses were centered and normalized by the standard deviation before clustering.

We employed two tests to determine whether our clustering method separated responses patterns into statistically meaningful sets. Test 1 ensured that responses within clusters were more similar to each other than to random same-size subsets of all responses. Using a bootstrap approach, we chose 10,000 different random subsets of responses and, for each subset, calculated the mean distance between each response pair, yielding a distribution of mean distances. Each cluster was determined to be a statistically significant subset if the mean distance between response pairs within the cluster overlapped with less than 5% of the random distance distribution. Test 2 was used to determine whether each cluster was significantly distinct from all other clusters. We evaluated this by comparing intra- and inter-cluster distances with two-sample t-tests, using Bonferroni corrections to adjust for multiple comparisons. (Figure 1—figure supplement 3). It should be noted that this measure is nonsymmetric, changing with the cluster chosen for the intra-cluster distance comparison.

This analysis showed that response motifs formed significantly distinct clusters (passed Test 1) when we directed the algorithm to produce two, three, or four clusters. When five or more clusters were produced, the results failed Test 1 and/or Test 2. The choice of four clusters matched the impressions of experimenters viewing the data. Test 2, however, showed that inhibitory and offset clusters based on responses to 1 s stimuli were not always statistically distinct from one another, likely because floor effects limited the extent to which the inhibitory portions of these responses could vary, and because the 1 s odor stimulus elicited brief responses in which some onset and offset features overlapped. However, the same analysis applied to response motifs elicited by lengthier 4 s pulses of odorants always yielded statistically distinct inhibitory and offset clusters. Further, inhibitory and offset responses showed different adaptation profiles (see below), suggesting that they are driven, at least in part, by different mechanisms. Also, inhibitory and offset motifs would trigger different consequences downstream, with only the offset motif eliciting post-stimulus spiking in follower neurons. Thus, going forward, we included four distinct response motifs in our analysis.

Quantification of motif switching

Request a detailed protocol

Motif switching rates were calculated by dividing the number of observed motif switches by the number of motif switching opportunities. The number of switching opportunities varied across subjects because some odor presentations did not elicit detectable responses from some ORNs.

Adaptation analysis

Request a detailed protocol

Sensory adaptation was tracked by measuring the maximum absolute change in firing rate from baseline to a detection window (highlighted in gray, Figure 3A, Figure 3—figure supplement 1) that followed each odor pulse in the trial. The baseline was defined as the spike rate during the 2 s prior to the delivery of the first odor pulse in each trial. The maximum absolute change value following each odor pulse was then normalized within each trial by that trial’s maximum firing rate. This was done to measure the proportion of change within each trial as an index of adaptation. Results were tested for significance by one-way repeated-measures ANOVAs.

Model of ORN response

Request a detailed protocol

We designed ORN models based on a computationally efficient approach we previously proposed to simulate the neuronal dynamics of different types of neurons (Bazhenov et al., 2008; Rulkov, 2002; Rulkov et al., 2004). This phenomenological neuron model is described by a set of difference (map) equations and offers several numerical advantages: it avoids the common problem of selecting the proper integration scheme since the model is already written in the form needed for computer simulations; simulations with large time steps are stable and precise and tens to thousands of times faster than models based on differential equations (e.g., Hodgkin–Huxley models); model parameters can be adjusted to match experimental data (Bazhenov et al., 2008; Rulkov and Bazhenov, 2008; Rulkov et al., 2004). We previously used this approach to describe populations of locust Kenyon cells in Assisi et al., 2020; Assisi et al., 2007; Kee et al., 2015; Sanda et al., 2016. Here, briefly, difference equations, rather than ordinary differential equations, were used to generate a sequence of membrane potential samples at discrete time points with time step h = 0.5 ms,

(1) xn+1=fαxn,yn+βn, yn+1=yn-μ1+xn-σ-σn,

where the variable xn given by the first equation modeled the fast dynamics of a neuron where each spike is formed by a single iteration with xn=1 due to the use of the discontinuous nonlinear function

(2) fα(xn,u)={α(1xn)+u,if xn<0.5,1,if0.5 xn<1,1,if xn1.

The second equation for variable yn controlled the slower (0<μ1) transient dynamics of the spiking activity and subthreshold oscillations. Intrinsic parameters α and σ define the baseline state and the regime of spiking in each neuron, determining whether the neuron produces tonic spiking or a burst of spikes. βn and σn are input variables setting fast (βn) and slow (σn) responsivity on the map model to external influences. βr and σr are used to tune the shape of receptor response. For inhibitory cells, βr controls the level of rebound activity. It is the parameter driving the post-stimulus effect. For excitatory cells, it controls the fast response to the stimulus and the deceleration of spiking responses as follows:

(3) σn= σξ ξn+σr In,

where coefficients βr and σr in front of the stimulus In were used to tune the shape of the receptor response for features such as speed of reaction in excitatory cells and the level of rebound activity in inhibitory cells. The noise ξn in the ORN neurons was modeled as Ornstein–Uhlnebeck (OU) process:

(4) ξn+1=qξn-p wn,

p = d 1-q2, h = 0.5 ms, the iteration time step, τc is the correlation time in ms, and d is the standard deviation of white Gaussian noise wn . Parameters of the noise model were set to τc = 3 ms, d = 0.01, βξ = 0.1, σξ = 1.0. Distinct trials were generated using different random initialization for these noise processes.

The input current, In , was shaped using a first-order low-pass filter:

(5) In+1=γsIn+(1-γs)assn,

where sn is the stimulus (odor concentration), the parameter γs is the controlled relaxation time, and as is the controlled responsivity strength and type (excitation or inhibition).

To model the different types of responses to an odor we observed in ORNs in vivo, we altered the parameters βr and γs , and as defining response onset speed, and other equations to shape transient characteristics of input current In with discrete-time filters based on the neuron motif. For a given motif, each of these parameters is fixed for all time and for all neurons of that motif.

Response motifasγsβrσr
Excitatory0.0390.9950.11
Inhibited–0.020.9980.11
Offset–0.040.9980.21

For example, the inhibited and offset motif are driven below baseline firing rate by stimulus due to the negative sign of the as parameter. The delayed response motif features inhibition first, and, then excitation with a slow relaxation rate. To model the initial inhibition, we used a fast high-pass-filtered current and mixed it with the low-pass output to form a short negative pulse at the beginning.

(6) In+1d=γdInd+ad(sn-sn-1),
hn+1p=γhhnp+(sn-sn-1),
In+1p=γpInp+(1-γp)apsnhn+1p,

which are limited using the Heaviside function H for Ind and mixed with Inp

(7) InM=Inp-HIndInd

making input current for such receptors as

(8) In=H(InM+L)InM,

where parameter L in Heaviside limits the depth of inhibition at the beginning of response.

Response motifapγpadγdγhLβrσr
Delayed0.080.9990.80.990.99985–0.040.011.0

Model of odor selectivity

Request a detailed protocol

We used an angle-based method to model the interactions between ORNs and odors. Each model ORN was assigned a selectivity vector, VORN , and similarly each odor was assigned a characteristic vector, Vodor . These vectors are defined in an arbitrary n-dimensional space. The efficacy or the odor–receptor interaction is given by Raman et al., 2010

(9) RiA=σ( |Vodor|cosp(θ))

where θ is the angle between VORN and Vodor in the n-dimensional space, p is a parameter that defines the receptive field width of the receptor, and σ(x) is the sigmoid response function defined as

(10) σ(x)=[1+exp(-a1(x-a2))]-1

with a1=15, a2=0.3. Consequently, odors whose characteristic vectors are close to a given ORN’s selectivity vector will elicit a strong response in that ORN. Two odors that are close in this n-dimensional space are similar to each other and thus elicit similar responses across the population of ORNs.

Here, we defined the odor and receptor vectors in a 3D space, n = 3, with each coordinate xi[0,1] (Raman et al., 2010). ORN response vectors were chosen randomly in the positive octant. For classification experiments, we defined a fixed initial odor vector (odor 1), we then created additional odor vectors at increasing angles of separation from the first odor. Odors were numbered by decreasing similarity to odor 1, for example, odor 2 is more similar to odor 1 than odor 9 to odor 1. All odor vectors were normalized.

Principal component analysis

Request a detailed protocol

We created binned spike counts of model ORN responses using 50 ms bins. PCA was performed over these binned responses to reduce the 10,000-dimensional ORN space into two dimensions corresponding to the maximum explained variance for visualizing the odor trajectories. The trajectories were averaged over 10 trials for each odor. We used the pca function in MATLAB for this analysis.

Classification of model odors

Request a detailed protocol

To classify odors, we first counted the spikes of model ORN responses into 50 ms bins. Binned response vectors were then used to train a SVM classifier to distinguish responses to different odors at each bin. We performed out-of-sample testing: the trial used for testing was chosen from 1 to 10, and the remaining 9 trials of each odor were used for training the SVM. Odor responses were classified at each 50 ms time bin and classification accuracy was averaged over the whole response. We constructed these SVMs using the fitcecoc function in MATLAB.

Temporal filters using linear nonlinear models

Request a detailed protocol

Neural responses to temporally complex inputs have been modeled by combining a linear filter, f(t), and a nonlinear threshold function, gx (Butts et al., 2007; Geffen et al., 2009). Using this approach, the linear filter is convolved with the stimulus and then passed through a nonlinear thresholding function to generate the approximate firing rate, rt=g[st*ft]. To derive the linear filter, we applied a standard deconvolution process: we deconvolved the signal, s(t), from the true firing rates of ORNs following published techniques (Butts et al., 2007; Geffen et al., 2009). Briefly, trial-averaged cell responses were convolved with a Gaussian filter with width σ= 50 ms; deconvolution occurred by inverting the linear convolution matrix. To account for correlations occurring in naturalistic stimuli, we multiplied each filter by the inverse of the stimulus covariance matrix (Sharpee, 2013). Before inverting the stimulus covariance matrix performed regularization using singular value decomposition and keeping components that account for 70% of the variance (Sharpee et al., 2008). The resulting filters were normalized by their L2 norm and low-pass filtered with a stopband of 10 Hz. The nonlinear thresholding function, which connected the convolution product to the final estimate, was created by constructing a histogram of all potential outputs of the linear filtering. For each bin, i, in this histogram we averaged all linear products, forming xi , and all true firing rates, forming yi . This procedure created a series of inputs (averaged linear product values) and outputs (firing rate values) that defined the thresholding function, y=g(x). For this analysis, we used 100 bins. The first 16 s of 40 s odor plumes were used for training and the next 16 s were used for testing. We then altered which 50% stretch of the data was used to train the model to derive a 95% confidence estimate of our L-NL prediction.

Distance-based artificial plume generation

Request a detailed protocol

We generated artificial odor plumes based on experiments in which plumes of ionic tracers in outdoor settings were measured at various distances from the source (Murlis et al., 2000). The statistics of the plume burst lengths and inter-burst intervals in these published experiments closely matched gamma distributions with the appropriate length and shape parameters. We used the arithmetic and geometric mean statistics of the real plumes to generate gamma distributions for burst length and inter-burst intervals representing plumes measured 2.5 m, 5 m, 10 m, and 20 m from the source. This optimization was achieved by the following relationship for gamma distributions:

(11) logk-ψk=logμ-logγ

where k is the shape parameter, ψ is the polygamma function, µ is the arithmetic mean, and γ is the geometric mean. This relationship was solved numerically using the MATLAB function fsolve. The length parameter was then given from the relationship θ=μ/k.

Distance-based stimulus classification

Request a detailed protocol

We delivered 15 s trials of artificial plumes constructed to reflect different distances from the source (2.5 m, 5 m, 10 m, and 20 m) to simulated ORNs. The ORN responses were then used to classify the stimulus by distance to source. Spiking responses of each ORN were counted into 50 ms bins and were then concatenated across cells, resulting in NCells * NBins length feature vectors. SVM classifiers were then trained using these vectors. Equal numbers of trials were used for training or testing. Each SVM was trained only using cells generating a given set of response motifs. To compare results across different SVMs, we used the same number of cells (n = 192, the smallest number of responsive cells across all response motifs) regardless of the motifs used in the classifier. To ensure the specific selection of cells did not influence the results, we created 25 different SVMs based on different selected cells and randomly chosen trials for training or testing. As before, SVMs were trained using MATLABs fitcecoc function.

Data availability

All data generated or analyzed during this study have been deposited at Open Science Framework and can be accessed here: Kim B and Haney S, Joshi S, Bazhenov M, Stopfer M (2022) Open Science Framework. Olfactory receptor neurons generate multiple response motifs, increasing coding space dimensionality https://osf.io/8bs72/.

The following previously published data sets were used
    1. Kim B
    2. Haney S
    3. Joshi S
    4. Bazhenov M
    5. Stopfer M
    (2022) Open Science Framework
    ID 8bs72. Olfactory receptor neurons generate multiple response motifs, increasing coding space dimensionality.

References

    1. Marion-Poll F
    2. Tobin TR
    (1992) Temporal coding of pheromone pulses and trains in manduca sexta
    Journal of Comparative Physiology. A, Sensory, Neural, and Behavioral Physiology 171:505–512.
    https://doi.org/10.1007/BF00194583

Decision letter

  1. Piali Sengupta
    Senior and Reviewing Editor; Brandeis University, United States
  2. Martin Nawrot
    Reviewer

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 "Olfactory receptor neurons generate multiple response motifs, increasing coding space dimensionality" 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 the review of your submission has agreed to reveal their identity: Martin Nawrot (Reviewer #3).

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

Essential revisions:

As you can see from the individual reviews, all reviewers (and myself) agree that this is a study of high potential importance with expertly executed experiments that result in an impressive and highly valuable dataset. However, we also all agree that some work is needed not only to expand the discussion but in particular to link the modelling part better to the experimental/data analysis one.

1) Presentation and level of detail: Please include detailed statistical descriptions throughout, details about the recording configuration including sensillum type, as well as detailed statistics about the interaction of motif switching and odour types. Do expand figure descriptions to ensure that all figure parts are properly explained.

2) Computational modelling: Locusts have a specific and unusual arrangement in that OSNs expressing the same receptor project to different glomeruli. Moreover, they have significantly more OSNs than e.g. Drosophila. Please assess how OSN number and the specific projection pattern affect the conclusions of the model to qualify its generality.

This is indeed a key issue – with the architecture of the locust's olfactory system so different from other insects (and lots of unknowns with respect to the molecular make-up), this not only requires extensive and detailed discussion and comparison but also possibly exploring larger parameter space in the model.

3) Computational modelling: Tighten the link between experimental approach and model, for example, by comparing the distribution of response latencies and peak rates. Assess the validity of the filters used and perform explicit cross-validation. Individual model parameters need to be explicitly described and it needs to be made clear how they impact the model predictions. Figure 5 seems to assess the consequence of motif-switching for a situation where responses to one odor are held constant. Importantly, the comparison to ORN activity in Figure 6 needs to be quantified.

4) Computational modelling: Generally, the model needs to be described much better, and accessible to non-experts. Equations need to be labelled, parameters described in detail, etc.

5) The set of odours used is quite limited. Ideally, a broader set of odours would be presented including odour mixtures and ethologically relevant "specialist" odours. Please at least discuss your findings and their applicability to other odour classes such as complex odour mixtures or ecologically particularly relevant odours where possibly a specialized circuitry could be expected.

6) As you can see from the detailed reviews, several important references and discussion points are missed. Please thoroughly go through those and amend accordingly.

Reviewer #1 (Recommendations for the authors):

This study would largely improve and be less specialized if the authors would broaden their discussion and provide further insight into their modeling approach. Furthermore, by expanding the odor set with ecologically relevant and rather diverse odors (such as PAN), it would be highly interesting to see whether the separation in these four response motifs would persist and whether odor identity coding would be more prominent. In addition, I have several specific suggestions that would improve readability and the value of the MS:

– The authors used two odors at different concentrations, while the other two odors were applied in just one concentration. Please add the concentration of the other two odors. It would be interesting to see the impact of odor concentrations on the dynamic of the OSN responses. Are the observed response motif concentration-dependent?

– The authors recorded mainly from trichoid sensilla, since those contain small numbers of OSNs which simplifies spike sorting. However, the authors mention that they also included recordings of other sensillum types. Please specify which types were measured and whether the response dynamics were sensillum-specific since those should express different types of ORs or IRs.

– In general, I am missing detailed info about the statistics used. The authors mentioned e.g. in Figure 3 that offset responses increased significantly during adaptation without providing any details. This increase is hardly visible in the corresponding figure.

– The descriptions of the figures in the figure legends are often insufficient and do not allow us to understand what is actually represented. For example, what is exactly shown in the two panels in Figure 3B? Does each line correspond to the same OSN responding to either hexanol or cyclohexanol? What is the meaning of the color code in Figure 3C? What are the shaded traces in Figure 4B? What is exactly represented in Figure 5C and what is the meaning of the color code? Please also increase the font size in Figure 3 to make it readable.

– The authors state that each OSN can exhibit different response motifs. How reliable/reproducible are the recorded responses? Do individual OSNs also switch between motifs for the same odor? Furthermore, it would be very informative if the motif switches would be analyzed with regard to odor specificity. Are the motif switches odor-dependent? Does for example an OSN that is tuned to hexanol always reveal an inhibition to cyclohexanol?

– As mentioned above, the computational modeling part is rather written for specialists and should be revised.

– Please cite and discuss the paper by Martelli and Fiala (eLife, 2019), which addresses adaptation mechanisms to odor pulses in OSNs in Drosophila.

Reviewer #2 (Recommendations for the authors):

Specific suggestions to address concerns are as follows:

1) There are 119 ORNs in the locust but simulations use a much greater number of receptors (10,000). Please justify this number or reduce it to match the experimental system.

2) The set of simulations in Figure 5 shows that motif switching improves odor classification. However, the comparison is made while holding responses to one odor constant while shifting responses to the second odor in different ways. This does not reflect the experimental situation where there is no motif switching for a given odorant.

3) The method that is used here to reconstruct neural filters is not appropriate for strongly correlated and natural odorant stimuli delivered experimentally. For a review of methods, please see https://pubmed.ncbi.nlm.nih.gov/23841838/.

4) It is imperative to add quantification for how accurately the model describes the ORN activity in Figure 6B.

5) Add error bars in Figure 6B for the estimated filters and gain functions.

Reviewer #3 (Recommendations for the authors):

Based on my opinion phrased in the public review above, I believe that this manuscript deserves publication with eLife if revised appropriately. That is, I have no doubt about the quality of experimental and theoretical methods and results. My concerns below refer to minor points in the description of the theoretical methods that may be improved, to a number of relevant but missing references including the observation of ORN response patterns in the fly, and to the Discussion that should be strengthened/deepened to further increase the impact of the present MS.

Specific review comments

1. ORN response motifs and temporal stimulus pattern responses in vivo and in silico

Galili et al. (2011) have shown clear offset responses (termed "post-odor responses") on ORNs in Drosophila. Martelli et al. (2013) have performed an extensive study on in vivo ORN responses and linear filter modeling of different stimulus-response patterns that partially fit the motif observations in the present manuscript. To my interpretation of their results, excitatory vs. delayed responses may also be odor concentration-dependent, different from the conclusion in the present supplementary figure. These studies should be cited and discussed; possibly there are additional references to that point that I am not aware of.

2. Discussion: ORN motifs in PN responses

PN responses have been shown to be highly complex in their odor-dependent temporal profile, in particular, they show inhibitory responses and delayed (late) responses (Krofczik et al., 2009), and off-responses (Galili et al., 2011). Individual PNs can respond rapidly to odor onset for one odor and exhibit a strong inhibition of other odors. This inhibition can be very fast abolishing PN response efficiently (Krofczik et al., 2009). The ORN motif switching shown here could provide an explanation for this observed behavior in PNs. Also, ORNs make direct connections to both, PNs and LNs (shown in detail in Drosophila) and this may further accentuate the expression of similar motifs in PNs odor code, e.g. if individual ORNs predominantly or exclusively target PNs and inhibitory LNs. Indicating the potential effect on PN coding in the Discussion will add to the impact of the present MS.

3. Discussion: mechanistic modeling of ORN adaptation in biophysical model neurons

The authors use a phenomenological linear filter model to describe the stimulus-response current. The Discussion does not indicate how realistic biophysical models for adaptive ORNs can at least capture the excitatory motif (as in classical stimulus adaptation reviewed in Benda 2021) and post-stimulus rebound effects. E.g the conductance-based spike frequency adaptation model (Farkhooi et al., 2013) has been shown to fit the excitatory response motif; this paper also showed that these mechanisms significantly increase response reliability across trials that cannot be captured by the phenomenological model presented here. This model of adaptation also explains the inhibitory post-stimulus effect after an excitatory response and the post-inhibition rebound as offset-response (Farkhooi et al., 2013, Betkiewicsz et al., 2020) that are also present in the response motifs where the avg. excitatory firing rate drops below baseline (Figure 1C, bottom and top) and in the inhibitory rebound after the offset of the long stimulus (Figure 1C, bottom, Figure 2A). The stimulus-response filters are designed to capture both these effects in the present MS (Figure 4B).

4. Discussion: Functional interpretation for odor sensing in a complex odor environment

The results are relevant for biologically realistic integrative models of the sensory pathway and sensory-motor transformations. The temporal dynamics of ORN responses are particularly relevant when simulating odor navigation behavior in flying or walking insects. The recent study by Rapp and Nawrot (2020) has shown that ORN adaptation (classical) translates to PN firing (which are modeled as non-adaptive neurons) and is important to reproduce temporally sparse coding in the mushroom body and is thus required for the active sampling of the statistics of odor encounters that can subserve navigation.

5. Discussion: Distance-related odor plume statistics

There have been several studies on the distance-dependent odor plume statistics and their mimicking in temporally patterned stimulation during physiological recordings, in particular in the moth (Jacob et al. 2017, Levakova et al. 2018).

6. Quantitative measures to inform models

For modeling purposes, it would be valuable if the authors can provide additional quantitative measures such as a distribution of response latencies and peak rates. Also, Figure 1C shows average response rates +-SEM for the 4 motifs in the overlay. A supplemental figure that shows trial-averaged average responses per unit in overlay separately for the four motives would allow for variability across neurons.

7. Method description of the theoretical model

– The methods section appears in front of the Results section. I believe it should appear after the Discussion according to eLife instructions.

– Please number equations.

– Eqn. following l.189: I understand u=(y_n+β_n) as in the first equation, correct? I was puzzled by the spike that "can" be produced if -0.5 < x_n < 1 and expected a stochastic process. The correct interpretation is that a spike is produced when threshold -0.5 is crossed I assume. x_n+1 is set to -1 after a spike, which is a reset. Can x_n become smaller than -1 due to input and noise or is it bound?

– Eqn. following l.207: Is γ^S < = 1 and > = 0? Is it a fixed parameter?

– The description of the effect of the model parameters β, γ, and α remains somewhat vague and very short in lines 212/213; it is not entirely transparent to me which parameter drives post-stimulus effects and whether γ is fixed or, likely, follows the stimulus step response.

– The authors show PID responses in Figure 6 for random stimulus trains. What do they look like for the long stimulus pulses? I would expect a low-pass type PID response (charging curve) and could this account for some of the low-pass filter properties in the model?

– Prediction with linear non-linear cascade model. It is not clear to me how cross-validation is performed here. The filter and transfer functions are estimated from the responses to the stochastic pulse presentation. How is the cross-validation done? Training on a set of trials (how large) and prediction on a different set? Even more convincing would be to train the model on the first half of the stimulus train and test on the second half. I understand that all animals were presented with the same stimulus. Can the authors train the model on individual neurons and predict the response of the pseudo population of non-simultaneously recorded cells? How well does the model work when the filter is estimated from single or repeated pulse presentations, does this easily transfer?

References:

Benda, J. (2021). Neural adaptation. Current Biology, 31(3), R110-R116.

Betkiewicz, R., Lindner, B., and Nawrot, M. P. (2020). Circuit and cellular mechanisms facilitate the transformation from dense to sparse coding in the insect olfactory system. Eneuro, 7(2).

Farkhooi, F., Froese, A., Muller, E., Menzel, R., and Nawrot, M. P. (2013). Cellular adaptation facilitates sparse and reliable coding in sensory pathways. PLoS computational biology, 9(10), e1003251.

Galili, D. S., Lüdke, A., Galizia, C. G., Szyszka, P., and Tanimoto, H. (2011). Olfactory trace conditioning in Drosophila. Journal of Neuroscience, 31(20), 7240-7248.

Jacob, V., Monsempès, C., Rospars, J. P., Masson, J. B., and Lucas, P. (2017). Olfactory coding in the turbulent realm. PLOS Computational Biology, 13(12), e1005870.

Krofczik, S., Menzel, R., and Nawrot, M. P. (2009). Rapid odor processing in the honeybee antennal lobe network. Frontiers in computational neuroscience, 2, 9.

Levakova, M., Kostal, L., Monsempès, C., Jacob, V., and Lucas, P. (2018). Moth olfactory receptor neurons adjust their encoding efficiency to temporal statistics of pheromone fluctuations. PLoS computational biology, 14(11), e1006586.

Martelli, C., Carlson, J. R., and Emonet, T. (2013). Intensity invariant dynamics and odor-specific latencies in olfactory receptor neuron response. Journal of Neuroscience, 33(15), 6285-6297.

Rapp, H., and Nawrot, M. P. (2020). A spiking neural program for sensorimotor control during foraging in flying insects. Proceedings of the National Academy of Sciences, 117(45), 28412-28421.

https://doi.org/10.7554/eLife.79152.sa1

Author response

Essential revisions:

As you can see from the individual reviews, all reviewers (and myself) agree that this is a study of high potential importance with expertly executed experiments that result in an impressive and highly valuable dataset. However, we also all agree that some work is needed not only to expand the discussion but in particular to link the modelling part better to the experimental/data analysis one.

Many thanks for this very positive and helpful assessment of our work! As we describe below, we’ve now made many revisions to improve our manuscript. Please note that line numbers given in this letter refer to the clean copy provided with this resubmission.

1) Presentation and level of detail: Please include detailed statistical descriptions throughout, details about the recording configuration including sensillum type, as well as detailed statistics about the interaction of motif switching and odour types. Do expand figure descriptions to ensure that all figure parts are properly explained.

We thank the editor and reviewers for their attention to our manuscript and for their helpful suggestions. As requested, we have now expanded these descriptions.

2) Computational modelling: Locusts have a specific and unusual arrangement in that OSNs expressing the same receptor project to different glomeruli. Moreover, they have significantly more OSNs than e.g. Drosophila. Please assess how OSN number and the specific projection pattern affect the conclusions of the model to qualify its generality.

This is indeed a key issue – with the architecture of the locust's olfactory system so different from other insects (and lots of unknowns with respect to the molecular make-up), this not only requires extensive and detailed discussion and comparison but also possibly exploring larger parameter space in the model.

As the reviewer notes, it is clearly true that species-specific differences exist in olfactory circuitry downstream from OSNs. The responses of OSNs in other species, including Drosophila, have been shown to include more than one spiking pattern, as we discuss (lines 471-481). But, briefly, we feel strongly that a substantial expansion of our treatment of species-specific differences falls outside of the subject of our manuscript which addresses only responses of OSNs, not their followers.

For this project we did not record from follower neurons. Our computational analysis of OSN activity was designed to reveal the information content of OSN responses without making any assumptions about, or drawing upon any features of, follower neurons or their projection patterns. Our analysis, for example, does not in any sense include a model of glomeruli or the antennal lobe or any other aspect of the architecture of the locust brain. We agree that such questions are interesting. In fact, earlier work from our group (Raman et al., 2010) investigated downstream consequences of patterned activity in ORNs. In our revised manuscript we discuss these ideas in lines 337-346 and 480-490. But revising our model to include downstream olfactory structures, such as the antennal lobe and glomeruli, would essentially constitute a major new study, one likely to raise many new questions about the assumptions we would have to make. We do plan to begin such a study in the near future.

The reviewer is also correct that different species have different numbers of OSNs, and different numbers of types of olfactory receptor proteins (we acknowledged this in our manuscript (lines 57-61)). However, these differences also do not affect our conclusions. Only one calculation in our manuscript, odor classification accuracy, would be affected by numbers of OSNs and OSN types; increasing numbers of OSNs and OSN types would, from first principles, be expected to allow greater absolute classification success; we think readers will correctly assume this is so. However, our conclusions do not rely on absolute levels of classification accuracy. Rather, we focus on the improvement in accuracy provided by including multiple response motifs in the analysis. The fact of this improvement would not be affected by varying the numbers of OSNs or OSN types.

Because our conclusions do not depend to any extent upon the architecture of the locust's olfactory system or numbers of OSNs or OSN types, we would prefer to omit detailed and necessarily speculative discussions and analyses of these factors from our manuscript.

3) Computational modelling: Tighten the link between experimental approach and model, for example, by comparing the distribution of response latencies and peak rates. Assess the validity of the filters used and perform explicit cross-validation. Individual model parameters need to be explicitly described and it needs to be made clear how they impact the model predictions. Figure 5 seems to assess the consequence of motif-switching for a situation where responses to one odor are held constant. Importantly, the comparison to ORN activity in Figure 6 needs to be quantified.

These are excellent suggestions. We have now addressed all of them with new analyses, new figures, and extensive revisions to the text, as described in detail below.

4) Computational modelling: Generally, the model needs to be described much better, and accessible to non-experts. Equations need to be labelled, parameters described in detail, etc.

We agree and have revised descriptions of our model, throughout our manuscript, for greater accessibility.

5) The set of odours used is quite limited. Ideally, a broader set of odours would be presented including odour mixtures and ethologically relevant "specialist" odours. Please at least discuss your findings and their applicability to other odour classes such as complex odour mixtures or ecologically particularly relevant odours where possibly a specialized circuitry could be expected.

We thank the reviewers for raising this concern which often arises in olfactory studies. We have now expanded our discussion to address the applicability of our findings to complex odor mixtures and “specialist” odorants such as PAN in lines 401-405: “For these studies we used odorants known to be ecologically relevant to locusts, including several found in the head space of wheat grass. Future experiments with larger sets of odorants, including blends or locust pheromones like 4-vinylanisole (4VA) and phenylacetonitrile (PAN), may help clarify the logic of motif switching.”

We also note, as above, that our conclusions reflect only OSN responses, not possibly specialized circuitry that may follow.

6) As you can see from the detailed reviews, several important references and discussion points are missed. Please thoroughly go through those and amend accordingly.

We agree and have now extensively revised the text to add suggested references and discussion points.

Reviewer #1 (Recommendations for the authors):

This study would largely improve and be less specialized if the authors would broaden their discussion and provide further insight into their modeling approach. Furthermore, by expanding the odor set with ecologically relevant and rather diverse odors (such as PAN), it would be highly interesting to see whether the separation in these four response motifs would persist and whether odor identity coding would be more prominent. In addition, I have several specific suggestions that would improve readability and the value of the MS:

– The authors used two odors at different concentrations, while the other two odors were applied in just one concentration. Please add the concentration of the other two odors. It would be interesting to see the impact of odor concentrations on the dynamic of the OSN responses. Are the observed response motif concentration-dependent?

We thank the reviewer for these suggestions. We have extensively revised the manuscript to broaden our discussion and provide further insight into our modeling approach.

The reviewer’s suggestion to test additional odorants often arises in olfaction studies – there are always more odors to test and good reasons for testing them. We note, along with the other reviewers, that our dataset is already extensive, and we are reluctant to extend it further at present. However, to address the reviewer’s concerns we have now added the following text to our discussion in lines 401-405: “For these studies we used odorants known to be ecologically relevant to locusts, including several found in the head space of wheat grass. Future experiments with larger sets of odorants, including blends or locust pheromones like 4-vinylanisole (4VA) and phenylacetonitrile (PAN), may help clarify the logic of motif switching.”

We share the reviewer’s interest in ORN responses elicited by different concentrations of odors. Our revised manuscript provides an analysis of the impact of odor concentrations on the dynamic of the OSN responses in Figure 2 —figure supplement 1, and a description in the text in lines 122-126: “…we found very little switching across repeated trials of the same odor (p=0.096), and an intermediate amount of switching across concentrations of the same odor (p=0.25, see Figure 2 —figure supplement 1). Thus, motif switching appears to provide a mechanism to convey information about odorant identity and concentration that is robust to trial-to-trial variation.”

As this analysis shows, response motifs do vary with odor concentration, but only slightly, to a much lesser extent than variance due to changes in odor identity. Our discussion now expands on the idea that response motifs contribute to information about odor concentration in lines 386-390: “Even relatively subtle changes in odorant chemistry (hexanol vs cyclohexanol, for example), or changes in the concentration of an odor, could evoke changes in the response motifs of ORNs (immediate vs delayed spiking, for example), leading to dramatic differences in the timing of the odor response.”

– The authors recorded mainly from trichoid sensilla, since those contain small numbers of OSNs which simplifies spike sorting. However, the authors mention that they also included recordings of other sensillum types. Please specify which types were measured and whether the response dynamics were sensillum-specific since those should express different types of ORs or IRs.

We agree -- it is important to specify which sensillum types were measured and whether the response dynamics were sensillum-specific. Some of this information is already present in our manuscript, but to make this clear, our revision now states (in lines 92-94) that “All results other than those shown in Figure 1 —figure supplement 2 are based on recordings from trichoid sensilla.” Figure 1 —figure supplement 2 shows that “unsorted, population activity recorded from other types of sensilla yielded results consistent with the responses of sorted ORNs, including prominent onset and offset activity” (lines 90-92). Thus, as our manuscript now makes clear, the response dynamics we observed do not appear to be sensillum-specific.

– In general, I am missing detailed info about the statistics used. The authors mentioned e.g. in Figure 3 that offset responses increased significantly during adaptation without providing any details. This increase is hardly visible in the corresponding figure.

We agree. To address the reviewer’s concern, we have now revised throughout our text and figures to provide more information. Regarding the specific information requested for Figure 3, we found normalized offset responses increased from -1 to +1, a substantial and statistically significant change that can be seen in the histograms in panel A, and in the graph in panel B. To clarify this issue we revised the caption for Figure 3 to state: “Each ORN response motif has a distinct adaptation profile. (A) Responses of ORNs, grouped by motif, to odors pulsed at different inter-pulse intervals (IPIs). Peaks for different motifs and pulses were measured as maximum absolute change from baseline within detection windows (shaded areas). (B) Adaptation characteristics of excitatory, delayed, offset, and inhibitory response motif to 10 pulses for each IPI. Left: response motifs; Right: Normalized response change from baseline. *=statistically significant changes elicited by odor pulses delivered at 0.50s IPI. Excitatory motif responses significantly decreased; offset motif responses significantly increased; delayed motif responses modestly increased; and inhibitory motif responses did not change. See Table 1 for statistical tests” (lines 169-175).

– The descriptions of the figures in the figure legends are often insufficient and do not allow us to understand what is actually represented. For example, what is exactly shown in the two panels in Figure 3B? Does each line correspond to the same OSN responding to either hexanol or cyclohexanol? What is the meaning of the color code in Figure 3C? What are the shaded traces in Figure 4B? What is exactly represented in Figure 5C and what is the meaning of the color code? Please also increase the font size in Figure 3 to make it readable.

We thank the reviewer for this suggestion and have now revised our figure captions to include more information. We have also completely remade Figure 3 to make it more readable with an enlarged font. Briefly, the left panel in Figure 3B illustrates the response motif indicated by the label above it, and the right panel shows the normalized response in this motif’s change from baseline. The revised figure caption is given above. The shaded traces in Figure 4B indicate standard error of the mean; the caption now states this. In Figure 5C, darker shading indicates lower classification accuracy; the caption now states this.

– The authors state that each OSN can exhibit different response motifs. How reliable/reproducible are the recorded responses? Do individual OSNs also switch between motifs for the same odor?

Thank you – we have now revised our text to clarify these results. Briefly, the recorded responses are very reliable and reproducible, as can be seen in Figure 2. The revised text now states (lines 120-126): “Response motif switching was not rare, occurring with probabilities ranging from 0.31 to 0.66 (mean = 0.38) for different pairs of odors (see Figure 2B-D). By contrast, we found very little switching across repeated trials of the same odor (0.096), and an intermediate amount of switching across concentrations of the same odor (0.25, see Figure 2 —figure supplement 1). Thus, motif switching appears to provide a mechanism to convey information about odorant identity and concentration that is robust to trial-to-trial variation.”

Furthermore, it would be very informative if the motif switches would be analyzed with regard to odor specificity. Are the motif switches odor-dependent? Does for example an OSN that is tuned to hexanol always reveal an inhibition to cyclohexanol?

We thank the reviewer for these questions and have now revised the manuscript to make these results clearer and to provide appropriate context for them. Briefly, motif switches are to some extent odor specific. As shown in Figure 2C, an OSN responding with excitation to hexanol with a probability of 0.880 shows a 0.000 probability of responding to cyclohexanol with inhibition. Our results suggest an underlying logic to motif switching and odor specificity. For example, a chi-square analysis showed the distribution of odor motifs switches within odor pairs is not what one would expect by chance, indicating some underlying structure, but a rigorous analysis of this phenomenon would require a much larger dataset with many more odorants. We plan to address this phenomenon in a future study. We have now expanded our discussion to address these important points.

Our revised discussion now states: “Our experiments did not allow us to test whether all ORNs are capable of motif switching, or what, if any, organizational principle determines when a given ORN will switch motifs. But because some odor pairs were more likely than others to elicit motif switching, we speculate that motif switching may follow a logic that evolved to help discriminate ecologically relevant but chemically similar odors” (lines 397-401).

– As mentioned above, the computational modeling part is rather written for specialists and should be revised.

We thank the reviewer for this suggestion. We would like our work to be accessible to many readers and have revised the computational modeling part in several places for clarity.

– Please cite and discuss the paper by Martelli and Fiala (eLife, 2019), which addresses adaptation mechanisms to odor pulses in OSNs in Drosophila.

Thank you. We now cite and describe this paper (see line 401 and lines 438-401): “Our analysis did not investigate mechanisms underlying the different forms of adaptation we observed; a recent study of ORNs in Drosophila has revealed calcium dynamics play a role in adaption (Martelli et al., 2019).”

Reviewer #2 (Recommendations for the authors):

Specific suggestions to address concerns are as follows:

1) There are 119 ORNs in the locust but simulations use a much greater number of receptors (10,000). Please justify this number or reduce it to match the experimental system.

The exact number of ORNs in the locust is not known, but estimates range from 45,000 to 113,000 per antenna (Leitch and Laurent 1996; Perez-Orive et al. 2002; Galizia and Sachse 2010). We believe our choice to model 10,000 ORNs is a reasonable compromise between the ideal size (which would be true number of ORNs in locust) and practical limitations needed to achieve computational efficiency.

The referee may be thinking about numbers of types of ORs rather than numbers of ORNs. Our simulations did not explicitly organize ORs into types. However, we are confident that doing so would not affect our conclusions. Our model is phenomenological, designed to simulate the four spiking motifs we observed in vivo. This model does not rely to any extend on the receptor proteins and cellular mechanisms determined by OR types and that generate these spiking motifs in vivo. Several figures in our manuscript (Figures 4, Figure 4 —figure supplement 1, Figure 6) demonstrate our model does a good job of simulating these four motifs. It is only these spiking patterns, not the cellular mechanisms underlying them, that matter to our conclusions. Overall, we believe our approach is appropriate for our questions and conclusions.

2) The set of simulations in Figure 5 shows that motif switching improves odor classification. However, the comparison is made while holding responses to one odor constant while shifting responses to the second odor in different ways. This does not reflect the experimental situation where there is no motif switching for a given odorant.

We regret that our description of this analysis was unclear, as mentioned above (public review, comment #2). As described in detail above, we have now revised our text to clarify our goals and procedures (lines 191-195, 206-209).

3) The method that is used here to reconstruct neural filters is not appropriate for strongly correlated and natural odorant stimuli delivered experimentally. For a review of methods, please see https://pubmed.ncbi.nlm.nih.gov/23841838/.

We agree and thank the reviewer for raising this important point and recommending an appropriate analysis.

To account for natural correlations present in the stimuli we used in experiments, we have now completely redone our analysis, using the method suggested by the reviewer. Briefly, filters calculated through deconvolution were multiplied by the inverse stimulus covariance matrix as described in the paper cited (Sharpee, 2013). The results of this new analysis, shown in revised Figure 6, are consistent with results in the previous version of our manuscript.

This new analysis is described in the methods section of the revised manuscript, as follows: “To account for correlations occurring in naturalistic stimuli, we multiplied each filter by the inverse of the stimulus covariance matrix (Sharpee, 2013). Before inverting the stimulus covariance matrix performed regularization using singular value decomposition and keeping components that account for 70% of the variance (Sharpee et al., 2008). The resulting filters were normalized by their L2 norm and low-pass filtered with a stopband of 10Hz” (lines 713-718).

4) It is imperative to add quantification for how accurately the model describes the ORN activity in Figure 6B.

We agree and thank the reviewer for raising this important point. To qualitatively compare ORN activity in vivo and in the model we have now prepared new supplementary figure; Figure 4 —figure supplement 1. As shown, results obtained in vivo and in the model are highly correlated, indicating the model accurately describes ORN activity. We now state this directly in the text, as follows: “As desired, our model provided an accurate simulation of ORN responses observed in vivo (Figure 4 —figure supplement 1 provides a quantitative comparison of response latencies and peaks in vivo and in the model)” (lines 164-167).

5) Add error bars in Figure 6B for the estimated filters and gain functions.

Thank you. We have now revised the figure to include shading that indicates 95% confidence intervals.

Reviewer #3 (Recommendations for the authors):

Based on my opinion phrased in the public review above, I believe that this manuscript deserves publication with eLife if revised appropriately. That is, I have no doubt about the quality of experimental and theoretical methods and results. My concerns below refer to minor points in the description of the theoretical methods that may be improved, to a number of relevant but missing references including the observation of ORN response patterns in the fly, and to the Discussion that should be strengthened/deepened to further increase the impact of the present MS.

Specific review comments

1. ORN response motifs and temporal stimulus pattern responses in vivo and in silico

Galili et al. (2011) have shown clear offset responses (termed "post-odor responses") on ORNs in Drosophila. Martelli et al. (2013) have performed an extensive study on in vivo ORN responses and linear filter modeling of different stimulus-response patterns that partially fit the motif observations in the present manuscript. To my interpretation of their results, excitatory vs. delayed responses may also be odor concentration-dependent, different from the conclusion in the present supplementary figure. These studies should be cited and discussed; possibly there are additional references to that point that I am not aware of.

We thank the reviewer for raising this interesting point. We now cite and discuss these studies. Briefly, we think our results are readily reconciled with those of Galili et al. and Martelli et al. by considering intrinsic odor dynamics. These earlier groups showed that odorants can have intrinsic dynamic properties and can be “fast” or “slow.” Our PID recordings showed none of the odors we used were “slow” – that is, none elicited PID signals that increased gradually throughout a stimulus presentation. Rather, all our odors elicited “fast” PID responses that approximated square pulses at all tested concentrations (please see the example below in response to point #7). Our observation rules out the possibility that differences in intrinsic odor dynamics could explain the origin of the response motifs we observed, and likely also explains why we did not observe the concentration-driven changes in ORN responses reported earlier.

2. Discussion: ORN motifs in PN responses

PN responses have been shown to be highly complex in their odor-dependent temporal profile, in particular, they show inhibitory responses and delayed (late) responses (Krofczik et al., 2009), and off-responses (Galili et al., 2011). Individual PNs can respond rapidly to odor onset for one odor and exhibit a strong inhibition of other odors. This inhibition can be very fast abolishing PN response efficiently (Krofczik et al., 2009). The ORN motif switching shown here could provide an explanation for this observed behavior in PNs. Also, ORNs make direct connections to both, PNs and LNs (shown in detail in Drosophila) and this may further accentuate the expression of similar motifs in PNs odor code, e.g. if individual ORNs predominantly or exclusively target PNs and inhibitory LNs. Indicating the potential effect on PN coding in the Discussion will add to the impact of the present MS.

Thank you -- we agree that it is important to consider the consequences of ORN motif switching on their downstream followers such as PNs. Our manuscript already included some of this discussion. In lines 337-344 we described how the responses of several layers of follower neurons are shaped by heterogeneities in the temporal structures of odor-elicited ORN responses, and cited our earlier work Raman et al., 2010 and Gupta and Stopfer, 2014, in support. Also, in Discussion (lines 480-483) we wrote: “ORNs provide the olfactory system’s first of several stages of signal processing. In insects, ORNs pass information to the antennal lobe, where spiking patterns originating in the periphery drive further processing in networks of local and projection neurons, leading to richer, higher dimensional olfactory representations.” As noted above, we predict interactions of ORN response motifs with intrinsic odor dynamics would likely increase ORN response dimensionality, and that greater response heterogeneity would further increase the dimensionality of PN responses.

To more directly address this point we have added the following to our discussion: “Our study did not address how multiple and switching response motifs in ORNs affect the responses of downstream neurons such as LNs and PNs. Our earlier work established that antennal lobe circuitry generates high-dimensional, complex temporally structured responses, but only when it is driven by input with heterogeneous timing from ORNs (Raman et al., 2010). The patterned ORN responses we report here likely contribute substantially to the variety of this input, and thus to the complexity and high dimensionality of PN responses” (lines 484-490).

3. Discussion: mechanistic modeling of ORN adaptation in biophysical model neurons

The authors use a phenomenological linear filter model to describe the stimulus-response current. The Discussion does not indicate how realistic biophysical models for adaptive ORNs can at least capture the excitatory motif (as in classical stimulus adaptation reviewed in Benda 2021) and post-stimulus rebound effects. E.g the conductance-based spike frequency adaptation model (Farkhooi et al., 2013) has been shown to fit the excitatory response motif; this paper also showed that these mechanisms significantly increase response reliability across trials that cannot be captured by the phenomenological model presented here. This model of adaptation also explains the inhibitory post-stimulus effect after an excitatory response and the post-inhibition rebound as offset-response (Farkhooi et al., 2013, Betkiewicsz et al., 2020) that are also present in the response motifs where the avg. excitatory firing rate drops below baseline (Figure 1C, bottom and top) and in the inhibitory rebound after the offset of the long stimulus (Figure 1C, bottom, Figure 2A). The stimulus-response filters are designed to capture both these effects in the present MS (Figure 4B).

This is an excellent point and we thank the reviewer for raising it. Although our model is basically descriptive and phenomenological, earlier work including biophysically realistic models can generate spiking behaviors comparable to the response motifs we identified. More realistic models may point toward a more granular and specific mechanistic understanding of the phenomena we describe.

It is also worth mentioning that, while our model is phenomenological, it was designed to simulate properties found in real neurons that can be traced to specific ion channels. In essence, the first equation (x variable) describes fast spiking dynamics and thus represent effects of fast Na+, K+ currents (INa, IK). The second equation describes slow membrane adaptation (variable y) that captures responses commonly represented by ca2+ dependent potassium currents (IK(Ca)), as well as rebound burst effects associated with low-threshold ca2+ currents (IT) found in some classes of neurons. Thus, in principle, the model suggests that a minimal set of currents needed to reproduce the excitatory and offset ORN responses would include INa, IK, IK(Ca), IT. Some additional voltage-gated K+ currents may be needed to explain delayed responses. These and other properties of the model are discussed in more detail here: Maxim Komarov, Giri Krishnan, Sylvain Chauvette, Nikolai Rulkov, Igor Timofeev and Maxim Bazhenov (2018) New class of reduced computationally efficient neuronal models for large-scale simulations of brain dynamics, Journal of Computational Neuroscience, 44:1–24.

To address the reviewer’s point we have expanded our discussion as follows: “Although our model is phenomenological, it was designed to simulate properties found in real neurons that can be traced to specific ion channels (Komarov et al., 2018). Also, earlier work including biophysically realistic models have been shown to generate spiking behaviors comparable to the response motifs we identified (e.g., Farkhooi et al., 2013; Betkiewicsz et al., 2020). Realistic biophysical models may point toward a more granular and specific mechanistic understanding of the responses we observed in ORNs” (lines 494-500).

4. Discussion: Functional interpretation for odor sensing in a complex odor environment

The results are relevant for biologically realistic integrative models of the sensory pathway and sensory-motor transformations. The temporal dynamics of ORN responses are particularly relevant when simulating odor navigation behavior in flying or walking insects. The recent study by Rapp and Nawrot (2020) has shown that ORN adaptation (classical) translates to PN firing (which are modeled as non-adaptive neurons) and is important to reproduce temporally sparse coding in the mushroom body and is thus required for the active sampling of the statistics of odor encounters that can subserve navigation.

This paper is indeed relevant to our work and we now discuss it in our manuscript, as follows: “A recent modeling study found that adaptation of ORN responses contributes substantially to the sparsening of responses downstream, and to the sampling of the statistics of odor encounters that could aid navigation to food sources (Rapp and Nawrot 2020)” (lines 462-465).

5. Discussion: Distance-related odor plume statistics

There have been several studies on the distance-dependent odor plume statistics and their mimicking in temporally patterned stimulation during physiological recordings, in particular in the moth (Jacob et al. 2017, Levakova et al. 2018).

We thank the reviewer for pointing us to these earlier, relevant studies. We now mention and cite these additional papers in two places as follows: “Earlier work performed in vivo and with computational models have investigated ways insect olfactory systems encode distance-dependent plume statistics (e.g., Jacob et al., 2017; Levakova et al. 2018)” (lines 460-462); and “A common theme of these findings is that olfactory systems include mechanisms to extract information needed for navigation” (lines 465-466).

6. Quantitative measures to inform models

For modeling purposes, it would be valuable if the authors can provide additional quantitative measures such as a distribution of response latencies and peak rates. Also, Figure 1C shows average response rates +-SEM for the 4 motifs in the overlay. A supplemental figure that shows trial-averaged average responses per unit in overlay separately for the four motives would allow for variability across neurons.

We agree and have now added two supplementary figures to provide this information. Figure 4 —figure supplement 1, as requested, compares the distributions of response latencies and peak rates observed for each motif in vivo and in our model. The caption: “Distribution of peak firing rates and response latencies for different motifs in model and in vivo experiments for a stimulus time period of 1000ms. (A) The peak responses were calculated as the maximum firing rate for the Excitatory, Delayed and Offset motifs, and as the minimum firing rate for the Inhibitory motif. (B) The response latencies were calculated as the time required to reach the peak response.”

A second new supplementary figure, Figure 1 —figure supplement 4, as requested, shows trial-averaged average responses per unit in overlay separately for the four motifs to illustrate variability across neurons. The caption: “1 second odor-olfactory receptor neuron responses. A total of 198 O-ORN responses are plotted in thin grey lines. The average response for each motif is overlaid in a thicker color line. Black bar denotes stimulus delivery.”

7. Method description of the theoretical model

– The methods section appears in front of the Results section. I believe it should appear after the Discussion according to eLife instructions.

Fixed – thank you!

– Please number equations.

Done.

– Eqn. following l.189: I understand u=(y_n+β_n) as in the first equation, correct? I was puzzled by the spike that "can" be produced if -0.5 < x_n < 1 and expected a stochastic process. The correct interpretation is that a spike is produced when threshold -0.5 is crossed I assume. x_n+1 is set to -1 after a spike, which is a reset. Can x_n become smaller than -1 due to input and noise or is it bound?

Yes, u=yn+βn is the first equation. The interpretation that a spike is produced when 0.5<x< 1 is correct – this is not a stochastic process. xn is not bound by -1 on the lower side as it can become smaller due to noise and other inputs. This equation is meant to ensure correct spike generation in the discrete-time model. Setting xn+1 to -1 ends the spike event but does not reset the system -- that is achieved by two-dimensional dynamics of the model determined by variables yn and βn. More information about the dynamics of the model can be found in cited references. To clarify this in the revised manuscript we no longer use u in the equation for nonlinear functions, and we have revised our explanation as follows, in lines 616-621:

“Here, briefly, difference equations, rather than ordinary differential equations, were used to generate a sequence of membrane potential samples at discrete time points with time step h= 0.5ms,

 xn+1=fα(xn,yn+βn),          </p><p>yn+1=ynμ(1+xnσσn), </p><p>

where the variable xn given by the first equation modeled the fast dynamics of a neuron where each spike is formed by a single iteration with xn=1 due to the use of the discontinuous nonlinear function fα(xn,u)={α(1xn)+u,if xn<0.5,</p><p>1,if0.5 xn<1,</p><p>1,if xn1.</p><p>

– Eqn. following l.207: Is γ^S < = 1 and > = 0? Is it a fixed parameter?

Yes, γs is a fixed parameter between 0 and 1 to define the model’s relaxation rate. It is fixed for a particular motif and does not change with the input. The actual values of this parameter are given in the table beginning on line 650. We now mention this in the revised text as follows: “For a given motif, each of these parameters is fixed for all time and for all neurons of that motif” (lines 646-647).

– The description of the effect of the model parameters β, γ, and α remains somewhat vague and very short in lines 212/213; it is not entirely transparent to me which parameter drives post-stimulus effects and whether γ is fixed or, likely, follows the stimulus step response.

We agree it is important to clarify these points, and now include, in many small changes throughout the Methods, the following information:

– Parameters α and σ define the baseline state and the regime of spiking in each neuron, determining whether the neuron produces tonic spiking or a burst of spikes.

βn and σn are input variables setting fast (βn) and slow (σn) responsivity on the map model to external influences. βr and σr are used to tune the shape of receptor response. For inhibitory cells, βr controls the level of rebound activity. It is the parameter driving the post-stimulus effect. For excitatory cells it controls the fast response to the stimulus and the deceleration of spiking responses.

γs is a fixed parameter for a particular motif and controls the relaxation time of the response. It does not follow the stimulus step response.

as is also a fixed parameter for a particular motif and controls the responsivity strength and the type (excitation or inhibition).

We have added to the text to explain which parameters drive peri-stimulus effects in inhibitory and offset motifs. For example, both motifs are driven below baseline firing rates by setting the as parameter negative.

– The authors show PID responses in Figure 6 for random stimulus trains. What do they look like for the long stimulus pulses? I would expect a low-pass type PID response (charging curve) and could this account for some of the low-pass filter properties in the model?

As we describe above, the PID responses we recorded to our odors had fast temporal profiles with rapidly rising and falling edges, as responses to long stimuli in Author response image 1 illustrate (4 sec stimulus pulse, green: individual traces; black: average). Because we observed throughout our dataset that every odor could elicit each of the four response motifs, we can conclude that the temporal properties of the odorants do not account for the model’s filter properties.

Author response image 1

– Prediction with linear non-linear cascade model. It is not clear to me how cross-validation is performed here. The filter and transfer functions are estimated from the responses to the stochastic pulse presentation. How is the cross-validation done? Training on a set of trials (how large) and prediction on a different set? Even more convincing would be to train the model on the first half of the stimulus train and test on the second half. I understand that all animals were presented with the same stimulus. Can the authors train the model on individual neurons and predict the response of the pseudo population of non-simultaneously recorded cells? How well does the model work when the filter is estimated from single or repeated pulse presentations, does this easily transfer?

Cross-validation was performed just as suggested by the reviewer: we used the first 16sec of 40sec odor plumes for training and the next 16sec for testing. We now describe this in the text (lines 723-725).

We don’t have any evidence for a mechanism to correlate firing among ORNs in different sensilla, but we note that in other sensory systems (such as vertebrate retina) correlated firing between afferent neurons accounts for a small but significant amount of stimulus encoding (Latham and Nirenberg, 2005; Paninski et al., 2008). Because of this we treat our model, one trained on individual neurons recorded separately, as providing a lower bound estimate of the stimulus information contained in the population response.

https://doi.org/10.7554/eLife.79152.sa2

Article and author information

Author details

  1. Brian Kim

    1. Eunice Kennedy Shriver National Institute of Child Health and Human Development (NICHD), National Institutes of Health (NIH), Bethesda, United States
    2. Brown University - National Institutes of Health Graduate Partnership Program, Providence, United States
    Contribution
    Conceptualization, Resources, Data curation, Software, Formal analysis, Validation, Investigation, Visualization, Methodology, Writing - original draft, Writing - review and editing
    Contributed equally with
    Seth Haney
    Competing interests
    No competing interests declared
  2. Seth Haney

    Department of Medicine, University of California, San Diego, San Diego, United States
    Contribution
    Conceptualization, Resources, Data curation, Software, Formal analysis, Validation, Investigation, Visualization, Methodology, Writing - original draft, Writing - review and editing
    Contributed equally with
    Brian Kim
    Competing interests
    No competing interests declared
  3. Ana P Milan

    Department of Clinical Neurophysiology and MEG Center, Amsterdam Neuroscience, Vrije Universiteit Amsterdam, Amsterdam, Netherlands
    Contribution
    Formal analysis, Investigation, Visualization, Methodology
    Competing interests
    No competing interests declared
  4. Shruti Joshi

    Department of Medicine, University of California, San Diego, San Diego, United States
    Contribution
    Data curation, Formal analysis, Validation, Visualization, Methodology, Writing - review and editing
    Competing interests
    No competing interests declared
  5. Zane Aldworth

    Eunice Kennedy Shriver National Institute of Child Health and Human Development (NICHD), National Institutes of Health (NIH), Bethesda, United States
    Contribution
    Conceptualization, Resources, Supervision, Validation, Methodology, Writing - review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-0647-8465
  6. Nikolai Rulkov

    Biocircuits Institute, University of California, San Diego, La Jolla, United States
    Contribution
    Data curation, Methodology
    Competing interests
    No competing interests declared
  7. Alexander T Kim

    Eunice Kennedy Shriver National Institute of Child Health and Human Development (NICHD), National Institutes of Health (NIH), Bethesda, United States
    Contribution
    Data curation, Methodology
    Competing interests
    No competing interests declared
  8. Maxim Bazhenov

    Department of Medicine, University of California, San Diego, San Diego, United States
    Contribution
    Conceptualization, Resources, Software, Supervision, Funding acquisition, Project administration, Writing - review and editing
    For correspondence
    mbazhenov@health.ucsd.edu
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-1936-0570
  9. Mark A Stopfer

    Eunice Kennedy Shriver National Institute of Child Health and Human Development (NICHD), National Institutes of Health (NIH), Bethesda, United States
    Contribution
    Conceptualization, Resources, Formal analysis, Supervision, Funding acquisition, Validation, Investigation, Writing - original draft, Project administration, Writing - review and editing
    For correspondence
    stopferm@mail.nih.gov
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-9200-1884

Funding

Office of Naval Research (N00014-16-1-2829)

  • Maxim Bazhenov

National Institutes of Health (RF1MH117155)

  • Maxim Bazhenov

National Institutes of Health (R01NS109553)

  • Maxim Bazhenov

National Science Foundation (IIS-1724405)

  • Maxim Bazhenov

Obra Social La Caixa (ID 100010434 with code LCF/BQ/ES15/10360004)

  • Ana P Milan

Eunice Kennedy Shriver National Institute of Child Health and Human Development (Intramural)

  • Mark A Stopfer

Intel Corp (CG42647565 FE2018)

  • Maxim Bazhenov

NIH (1R01DC020892)

  • Maxim Bazhenov

NSF (EFRI BRAID 2223839)

  • Maxim Bazhenov

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

Acknowledgements

Supported by ONR (N00014-16-1-2829), NIH (RF1MH117155 and R01NS109553), Intel Corp (CG42647565 FE2018), NSF (IIS-1724405), NIH (1R01DC020892), and NSF (EFRI BRAID 2223839) awarded to MB, a doctoral thesis grant from Obra Social La Caixa (ID 100010434 with code LCF/BQ/ES15/10360004) awarded to APM, and an intramural grant from the National Institute of Child Health and Human Development, National Institutes of Health (MS).

Senior and Reviewing Editor

  1. Piali Sengupta, Brandeis University, United States

Reviewer

  1. Martin Nawrot

Version history

  1. Received: April 1, 2022
  2. Preprint posted: April 12, 2022 (view preprint)
  3. Accepted: January 31, 2023
  4. Accepted Manuscript published: January 31, 2023 (version 1)
  5. Version of Record published: February 13, 2023 (version 2)

Copyright

This is an open-access article, free of all copyright, and may be freely reproduced, distributed, transmitted, modified, built upon, or otherwise used by anyone for any lawful purpose. The work is made available under the Creative Commons CC0 public domain dedication.

Metrics

  • 1,275
    Page views
  • 218
    Downloads
  • 2
    Citations

Article citation count generated by polling the highest count across the following sources: PubMed Central, Crossref, 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. Brian Kim
  2. Seth Haney
  3. Ana P Milan
  4. Shruti Joshi
  5. Zane Aldworth
  6. Nikolai Rulkov
  7. Alexander T Kim
  8. Maxim Bazhenov
  9. Mark A Stopfer
(2023)
Olfactory receptor neurons generate multiple response motifs, increasing coding space dimensionality
eLife 12:e79152.
https://doi.org/10.7554/eLife.79152

Share this article

https://doi.org/10.7554/eLife.79152

Further reading

    1. Neuroscience
    Maureen van der Grinten, Jaap de Ruyter van Steveninck ... Yağmur Güçlütürk
    Tools and Resources

    Blindness affects millions of people around the world. A promising solution to restoring a form of vision for some individuals are cortical visual prostheses, which bypass part of the impaired visual pathway by converting camera input to electrical stimulation of the visual system. The artificially induced visual percept (a pattern of localized light flashes, or ‘phosphenes’) has limited resolution, and a great portion of the field’s research is devoted to optimizing the efficacy, efficiency, and practical usefulness of the encoding of visual information. A commonly exploited method is non-invasive functional evaluation in sighted subjects or with computational models by using simulated prosthetic vision (SPV) pipelines. An important challenge in this approach is to balance enhanced perceptual realism, biologically plausibility, and real-time performance in the simulation of cortical prosthetic vision. We present a biologically plausible, PyTorch-based phosphene simulator that can run in real-time and uses differentiable operations to allow for gradient-based computational optimization of phosphene encoding models. The simulator integrates a wide range of clinical results with neurophysiological evidence in humans and non-human primates. The pipeline includes a model of the retinotopic organization and cortical magnification of the visual cortex. Moreover, the quantitative effects of stimulation parameters and temporal dynamics on phosphene characteristics are incorporated. Our results demonstrate the simulator’s suitability for both computational applications such as end-to-end deep learning-based prosthetic vision optimization as well as behavioral experiments. The modular and open-source software provides a flexible simulation framework for computational, clinical, and behavioral neuroscientists working on visual neuroprosthetics.

    1. Neuroscience
    Simon Lui, Ashleigh K Brink, Laura H Corbit
    Research Article

    Extinction is a specific example of learning where a previously reinforced stimulus or response is no longer reinforced, and the previously learned behaviour is no longer necessary and must be modified. Current theories suggest extinction is not the erasure of the original learning but involves new learning that acts to suppress the original behaviour. Evidence for this can be found when the original behaviour recovers following the passage of time (spontaneous recovery) or reintroduction of the reinforcement (i.e. reinstatement). Recent studies have shown that pharmacological manipulation of noradrenaline (NA) or its receptors can influence appetitive extinction; however, the role and source of endogenous NA in these effects are unknown. Here, we examined the role of the locus coeruleus (LC) in appetitive extinction. Specifically, we tested whether optogenetic stimulation of LC neurons during extinction of a food-seeking behaviour would enhance extinction evidenced by reduced spontaneous recovery in future tests. LC stimulation during extinction trials did not change the rate of extinction but did serve to reduce subsequent spontaneous recovery, suggesting that stimulation of the LC can augment reward-related extinction. Optogenetic inhibition of the LC during extinction trials reduced responding during the trials where it was applied, but no long-lasting changes in the retention of extinction were observed. Since not all LC cells expressed halorhodopsin, it is possible that more complete LC inhibition or pathway-specific targeting would be more effective at suppressing extinction learning. These results provide further insight into the neural basis of appetitive extinction, and in particular the role of the LC. A deeper understanding of the physiological bases of extinction can aid development of more effective extinction-based therapies.