Article activity feed

  1. Author Response:

    Reviewer #3 (Public Review):

    1. The two algorithms presented are essentially a low-pass and high-pass filter on binarized odor. As such, it may not be so surprising that there is a tradeoff between which algorithm works better depending on the frequency content of different environments. The low-pass filter (algorithm 1) works better in environments with mostly low-frequency fluctuations (boundary layer plume, low wind-speed, high diffusivity) while the high-pass filter (algorithm 2) works better in environments with mostly high-frequency fluctuations (high windspeed, low diffusivity). To understand what is essential in these algorithms I think it would be useful to (1) compare the two algorithms to a "null" algorithm that drives upwind orientation whenever odor is present (i.e. include thresholding and binarization but no filtering), (2) compare navigation success metrics directly to the frequency content of different environments, (3) examine how navigation success depends on the filtering cutoff of the two algorithms (tau_on and tau_w). Comparing to the null algorithm with no filtering I think is important to determine whether there is actually a tradeoff to be made, or whether a system that can approximate a flat transfer function (or at least capture all relevant frequencies in the environment) is ideal and must be approximated with biological parts.

    For (1) and (3), we have now added simulations of the models for a range of different timescales, including an integrator with an infinitely fast timescale corresponding to the “null” model the reviewer describes (Results lines 376-380, Figure 4—figure supplement 2 and Materials and methods lines 1008-1025). We find that changing the timescale of the intermittency filter largely leaves performance unchanged whereas changing the timescale of the frequency filter is akin to changing the gain on the frequency filter, as predicted by Equations 24 and 29. Since we do find a local maximum in the frequency filter timescale, we conclude that there are benefits to filtering in time. For (2), many plumes we simulate in Fig. 5 span a wide range of frequencies and intermittencies; we chose to plot performance as a function of diffusivity / windspeed to emphasize how performance depends on environment parameters that shape the statistics of the plume (flow and odor dynamics). Note that we renamed 𝜏! to 𝜏".

    1. While the two algorithms presented here present a nice conceptual division, biological filtering algorithms are likely to incorporate elements of both. For example, the adaptive compression algorithm of Alvarez-Salvado (which is eliminated in the simplification used here) provides some sensitivity to odor onsets and is based on well-described adaptation at the olfactory periphery. Synaptic depression algorithms likewise provide sensitivity to derivatives as well as integration over time, and synaptic depression with multiple timescales has been described in detail at various stages of the olfactory system. A productive extension of the work done here would be to explore the utility of biophysically-motivated filtering algorithms for navigation in different environments.

    Thank you for this suggestion, which led us to extend our work in that interesting direction. We have now generalized our model to respond to odor intensity (rather than its binarized version) by implementing an adaptive compression taken from prior modeling efforts (Alvarez-Salvado et al, eLife 2018) (added to Fig. 3; also see additional Fig. 3 Supplement 1). Moreover, we now also consider navigators that respond to odor signals using a biophysical model of odor transduction, ORN firing, and PN firing, in addition to synaptic depression within the ORN-PN synapse, which combines modeling efforts from prior works (Gorur-Shandilya, Demir, et al, eLife 2017; Nagel & Wilson, Nat. Neurosci. 2015; Fox & Nagel, “Synaptic control of temporal processing in the Drosophila olfactory system” arXiv 2021). This realistic circuit model produced exciting results that indicate that the natural ORN-PN circuitry can, to some degree, satisfy the dual demands of intermittency and frequency sensing. These results are shown in the new Fig. 6.

    1. It would be helpful in the Discussion to present a clearer picture of what the frequency content of natural environments is likely to be. For example, flies stop walking at windspeeds above ~70cm/s (Yorozu 2009). In contrast, flies in flight are likely to encounter much sparser and high frequency plume encounters, as they are moving through the air at much faster speeds and because odors encountered here would be away from the boundary layer. Therefore the best test of the tradeoff hypothesis would likely be to compare temporal filtering of odor plumes by neural circuitry in flying vs walking flies. This would connect to the literature in motion detection as well, where octopamine release during flight causes a speeding of the motion detection algorithm.

    We have added lines 47-48 to the introduction describing the natural frequency content of plumes and lines 574-578 discussing how one might see evidence of this tradeoff when comparing between walking and flying flies.

    Was this evaluation helpful?
  2. Evaluation Summary:

    This manuscript by Jayaram and colleagues uses computational modeling approaches to examine how temporal filtering of an odor signal contributes to navigation success in different odor environments. The manuscript advances the literature in considering how different algorithms may be optimal for different environments. Further evidence is required to more convincingly prove the intriguing trade-off between frequency and "intermittency" sensing described here.

    (This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. Reviewer #1 agreed to share their name with the authors.)

    Was this evaluation helpful?
  3. Reviewer #1 (Public Review):

    In the manuscript by Jayaram and coworkers, the authors model how temporal features of the olfactory environment impact the navigation of walking fruit flies. The authors find that under certain stimulus conditions, utilization of both intermittency and odor encounter rate increases navigational success in the agent-based model.

    The strengths of the study lie in the examination of distinct aspects of the features of the plume as the animal navigates. The authors build on their previous and important work in this domain that used spatiotemporal complex plumes and navigating flies.

    The weaknesses of the study are concerns about the oversimplification of the spatiotemporal dynamics of the plume and its encounters with the navigating fly and the other features of the plume (intensity). The authors describe other processes that were excluded from the model (bilateral sensing, learning, adaptation) - the model results that include these features would more realistically define the model results. Testing only two different plumes (environments; high frequency, high intermittency) may also over-simply the processes at play and results.

    Of course, the authors are the experts, but in atmospheric sciences and past odor plume studies, the term intermittency is related to the stimulus (conditional statistics on the presence of the volatile chemical signal), rather than the reference frame of the navigating organism. Similarly, "rate" has been used in the more powerful characterization of "flux" (C/time) as the animal navigates to the plume. Defining these terms at the onset, and incorporating intensity in the model, would be helpful.

    Was this evaluation helpful?
  4. Reviewer #3 (Public Review):

    1. The two algorithms presented are essentially a low-pass and high-pass filter on binarized odor. As such, it may not be so surprising that there is a tradeoff between which algorithm works better depending on the frequency content of different environments. The low-pass filter (algorithm 1) works better in environments with mostly low-frequency fluctuations (boundary layer plume, low wind-speed, high diffusivity) while the high-pass filter (algorithm 2) works better in environments with mostly high-frequency fluctuations (high windspeed, low diffusivity). To understand what is essential in these algorithms I think it would be useful to (1) compare the two algorithms to a "null" algorithm that drives upwind orientation whenever odor is present (i.e. include thresholding and binarization but no filtering), (2) compare navigation success metrics directly to the frequency content of different environments, (3) examine how navigation success depends on the filtering cutoff of the two algorithms (tau_on and tau_w). Comparing to the null algorithm with no filtering I think is important to determine whether there is actually a tradeoff to be made, or whether a system that can approximate a flat transfer function (or at least capture all relevant frequencies in the environment) is ideal and must be approximated with biological parts.

    2. While the two algorithms presented here present a nice conceptual division, biological filtering algorithms are likely to incorporate elements of both. For example, the adaptive compression algorithm of Alvarez-Salvado (which is eliminated in the simplification used here) provides some sensitivity to odor onsets and is based on well-described adaptation at the olfactory periphery. Synaptic depression algorithms likewise provide sensitivity to derivatives as well as integration over time, and synaptic depression with multiple timescales has been described in detail at various stages of the olfactory system. A productive extension of the work done here would be to explore the utility of biophysically-motivated filtering algorithms for navigation in different environments.

    3. It would be helpful in the Discussion to present a clearer picture of what the frequency content of natural environments is likely to be. For example, flies stop walking at windspeeds above ~70cm/s (Yorozu 2009). In contrast, flies in flight are likely to encounter much sparser and high frequency plume encounters, as they are moving through the air at much faster speeds and because odors encountered here would be away from the boundary layer. Therefore the best test of the tradeoff hypothesis would likely be to compare temporal filtering of odor plumes by neural circuitry in flying vs walking flies. This would connect to the literature in motion detection as well, where octopamine release during flight causes a speeding of the motion detection algorithm.

    Was this evaluation helpful?