Continuous, long-term crawling behavior characterized by a robotic transport system
Curation statements for this article:-
Curated by eLife
eLife assessment
This study describes a useful method to monitor the behavior of Drosophila larvae in a uniform environment over much longer time scales than was possible with previous methods. The authors provide a solid characterization of aspects of the method and show that the behavior of single larvae can be quantified over several hours. The experiments offer a proof-of-concept for a robotic device that will enable the investigation of behavior in long-term experiments in ways that were previously unimaginable.
This article has been Reviewed by the following groups
Listed in
- Evaluated articles (eLife)
Abstract
Detailed descriptions of behavior provide critical insight into the structure and function of nervous systems. In Drosophila larvae and many other systems, short behavioral experiments have been successful in characterizing rapid responses to a range of stimuli at the population level. However, the lack of long-term continuous observation makes it difficult to dissect comprehensive behavioral dynamics of individual animals and how behavior (and therefore the nervous system) develops over time. To allow for long-term continuous observations in individual fly larvae, we have engineered a robotic instrument that automatically tracks and transports larvae throughout an arena. The flexibility and reliability of its design enables controlled stimulus delivery and continuous measurement over developmental time scales, yielding an unprecedented level of detailed locomotion data. We utilize the new system’s capabilities to perform continuous observation of exploratory search behavior over a duration of 6 hr with and without a thermal gradient present, and in a single larva for over 30 hr. Long-term free-roaming behavior and analogous short-term experiments show similar dynamics that take place at the beginning of each experiment. Finally, characterization of larval thermotaxis in individuals reveals a bimodal distribution in navigation efficiency, identifying distinct phenotypes that are obfuscated when only analyzing population averages.
Article activity feed
-
-
Author Response
Reviewer #1 (Public Review):
The paper describes a robotic system that can be used for prolonged recording of forced activity in crawling Drosophila larvae. This is mostly intended to be a proof of principle description of a tool potentially useful for the community. The system - whose value lies completely in its reproducibility and adoption - is only superficially described in the paper, but a more detailed description is made available through Github, along with the software used for the collection and analysis of data.
There is good, convincing evidence this can work as some sort of "larval conveyor belt", used to artificially prolong food crawling behaviour in the animals. More could be said about the ecological implications of the assay (for instance: how relevant is it to an animal's natural behaviour? Does …
Author Response
Reviewer #1 (Public Review):
The paper describes a robotic system that can be used for prolonged recording of forced activity in crawling Drosophila larvae. This is mostly intended to be a proof of principle description of a tool potentially useful for the community. The system - whose value lies completely in its reproducibility and adoption - is only superficially described in the paper, but a more detailed description is made available through Github, along with the software used for the collection and analysis of data.
There is good, convincing evidence this can work as some sort of "larval conveyor belt", used to artificially prolong food crawling behaviour in the animals. More could be said about the ecological implications of the assay (for instance: how relevant is it to an animal's natural behaviour? Does the system introduce artifactual distortions in the analysis, driven by the fact that animals crawl greater distances than they would normally crawl in nature? Will this extensive activity affect their development to pupation or adulthood?).
In addition all our code being available on GitHub, we have added substantially to Materials and Methods in the manuscript (1-1.5 pages) detailing the analysis pipeline more thoroughly.
We agree that a more thorough comparison of ecological vs. laboratory conditions was warranted here, and have addressed this in new Discussion section material (6th paragraph especially). The developmental effect due to prolonged locomotion is a very good point – with only a single animal measured for more than 24 hours, we do not yet know whether instar molting or pupation is delayed, but this could certainly be a concern in longer experiments moving forward.
Reviewer #3 (Public Review):
"Continuous, long-term crawling behavior characterized by a robotic transport system" by Yu et al. presents their new robotic device to track, reposition, and feed Drosophila larvae as they crawl on an arena. By using a water droplet (or if necessary, suction) to transport larvae from the edge of the arena to the middle, long behavior trajectories can be recorded without losing larvae from the arena or camera field of view. The picker robot is also able to dispense small amounts of apple juice at precise locations to keep larvae alive for extended periods although the food was not sufficient to trigger molting and the development to the next instar stage.
The approach is interesting, but the authors could provide more details on why the approach is necessary for non-expert readers. For example, what are the advantages of using the robot picker compared to simply confining larvae in a closed arena? It's not obvious (to me) that being picked back to the center of the arena is a smaller perturbation compared to running into a chamber wall and changing direction.
Thank you for this suggestion, it’s a very good point. We have expanded our Introduction considerably, and directly address this issue (4th paragraph in particular). We do quantify the perturbation due to robot pick-ups and drop-offs (Fig. 3D), but that only addresses the short term. We prefer not to use a closed arena for three reasons: (1) in a gradient navigation experiment, reaching the edge would effectively end “navigation” and we would be unable to study that behavior over longer times, (2) larvae can crawl up the sides of walls and will be lost to the tracker (they do this all the time in the Petri dishes they are raised in), and (3) larvae often do not bounce off walls and resume crawling, they tend to dwell near edges they find. To this last point, we have added a new Supplemental figure (Figure 1 – supplement 1) illustrating this effect with a representative example.
The first paragraph of the introduction emphasizes the multiple time scales that are relevant for behavior from rapid stimulus response up to developmental times. This is to set the context of the authors' contribution but I'm not sure it's a fair representation of the state of the art. For example, the authors state that high-bandwidth measurement over long times is prohibitive and cite three Drosophila papers, but there are home-cage monitoring systems that allow continuous recording of mouse behavior over long times with high resolution. At the other end of the spectrum, there have been some long-term behaviour experiments done on worm behaviour with reasonably high time resolution (e.g Stern et al. 10.1016/j.cell.2017.10.041).
This is absolutely correct, the context needed to be much broader than our own prior larva results. We have overhauled that section and written a wider introduction that includes the C. elegans paper you mentioned, and also brings in other model systems like adult flies, mice, and rats. We frame our own work as (1) in a new animal, for long term measurements; (2) investigating non-confined free locomotion over a long time scale.
The authors train a neural network to segment and track the larvae, however, little information is given on the training process and I don't think it would be possible to reproduce the model based on the description. More details of the network, hyperparameters, and training data would be required to evaluate it.
Definitely! We have added a new section to Materials and Methods (1-1.5 pages in length), detailing our analysis pipeline, with sections for position tracking, postural analysis, and behavioral classification.
The authors also state several times that larval identity is maintained throughout the recording, but this isn't quantified. It's not clear whether identity is maintained across collisions of two or more animals by the tracking algorithm or whether these collisions simply don't happen in their data because density is low.
This has also been addressed and clarified in the same new part of the Materials and Methods section. We quantify collision rates and give the accuracy maintaining identity after collisions.
The environment is nominally isotropic, but once larvae have been crawling on the surface for hours, including periodic feeding, there will likely be multiple gradients the larvae may sense. This may not be observable in the data, but should perhaps be mentioned in the text.
This is certainly true. Other than the single animal 30-hour experiment described in the manuscript, there is no food introduced to the larvae during our 6-hour experiments. Looking ahead, the presence of food remnants in the arena could become a serious confounding factor in nominally isotropic experiments, as the reviewer points out. We have added substantially to the Discussion section to discuss various limitations of the design and experiments, and directly talk about the odor/taste stimuli being introduced by food (second to last paragraph in Discussion).
The authors show that the picking action results in a small but detectable increase in speed. The degree of perturbation overall depends on the picking frequency so some quantification of the inter-pick time interval would help to interpret whether this perturbation is relevant for a particular experiment. Is there a difference in excitation when larvae are picked successfully on the first try compared to when multiple tries or suction are required?
We have now quantified the amount of time between pickups and added that in the Materials and Methods section directly (it’s 0.87 pick-ups per hour per animal). We do not have a sufficient amount of data to determine whether there is a statistically significant difference in behavior for multiple pickup attempts – this can also be confounded because sometimes an unsuccessful pickup is one that does not touch the larva at all (so would presumably not introduce additional perturbations).
From the reconstructed trajectory in Figure 4, this interval looks very long compared to speed increase after picking. When reconstructing the trajectory, how are the segments joined? Is it simply by resetting the xy position or also updating rotating to match the previous direction of travel? (I'm guessing the larva can rotate during transport?)
We have updated the Figure 4 caption to make it clear that the segments are only joined translationally, by resetting the xy position.
The authors present a simple model in Figure 6 to illustrate the differences between individuals that can be hidden when looking at population distributions. However, the differences they show in the simulation don't seem relevant to the differences they observe in the experiments. Specifically, Fig. 6A and B show a contrast between individuals with similar mean speeds compared to individuals with different (but still unimodal) mean speeds. In contrast, the experimental data in Fig. D shows individual distributions that are quite similar but that are bimodal. So, there is indeed a difference between the individual distributions that is obscured in the population distribution, but is there evidence of larval personality types (line 444)? Similarly, the sentence beginning line 381 doesn't seem right either.
We are really glad this was brought up so that we could clarify better in the text, as it’s an important point. We have edited the text in the Results subsection related to Figure 6 and the Figure 6 caption to clear things up. The individual distributions in 6D are not bimodal, there are 38 traces shown that are all essentially unimodal. In addition to stating this directly in the text, we have quantified this by adding the average BC for individuals in both isotropic and thermal gradient contexts (they are essentially the same, i.e. equally unimodal in both cases).
-
eLife assessment
This study describes a useful method to monitor the behavior of Drosophila larvae in a uniform environment over much longer time scales than was possible with previous methods. The authors provide a solid characterization of aspects of the method and show that the behavior of single larvae can be quantified over several hours. The experiments offer a proof-of-concept for a robotic device that will enable the investigation of behavior in long-term experiments in ways that were previously unimaginable.
-
Reviewer #1 (Public Review):
The paper describes a robotic system that can be used for prolonged recording of forced activity in crawling Drosophila larvae. This is mostly intended to be a proof of principle description of a tool potentially useful for the community. The system - whose value lies completely in its reproducibility and adoption - is only superficially described in the paper, but a more detailed description is made available through Github, along with the software used for the collection and analysis of data.
There is good, convincing evidence this can work as some sort of "larval conveyor belt", used to artificially prolong food crawling behaviour in the animals. More could be said about the ecological implications of the assay (for instance: how relevant is it to an animal's natural behaviour? Does the system introduce …
Reviewer #1 (Public Review):
The paper describes a robotic system that can be used for prolonged recording of forced activity in crawling Drosophila larvae. This is mostly intended to be a proof of principle description of a tool potentially useful for the community. The system - whose value lies completely in its reproducibility and adoption - is only superficially described in the paper, but a more detailed description is made available through Github, along with the software used for the collection and analysis of data.
There is good, convincing evidence this can work as some sort of "larval conveyor belt", used to artificially prolong food crawling behaviour in the animals. More could be said about the ecological implications of the assay (for instance: how relevant is it to an animal's natural behaviour? Does the system introduce artifactual distortions in the analysis, driven by the fact that animals crawl greater distances than they would normally crawl in nature? Will this extensive activity affect their development to pupation or adulthood?).
-
Reviewer #2 (Public Review):
This study describes the development of a robotic system that allows investigators to track the movements of Drosophila larvae for extremely long time durations. Prior studies were limited by the fact that tracking of larval movements needed to be stopped whenever the animal reached the edge of a behavioral arena. This new study overcomes this limitation with a robot arm that gently picks up the larvae when they reach the edge of the arena and then gently releases them again so that tracking can be resumed. The very long periods of data acquisition are performed with a video camera that provides a low-resolution 64x64 pixel representation of the larvae. Nevertheless, the authors are able to extract postural information from the animals using a sophisticated machine vision based neural network. The authors …
Reviewer #2 (Public Review):
This study describes the development of a robotic system that allows investigators to track the movements of Drosophila larvae for extremely long time durations. Prior studies were limited by the fact that tracking of larval movements needed to be stopped whenever the animal reached the edge of a behavioral arena. This new study overcomes this limitation with a robot arm that gently picks up the larvae when they reach the edge of the arena and then gently releases them again so that tracking can be resumed. The very long periods of data acquisition are performed with a video camera that provides a low-resolution 64x64 pixel representation of the larvae. Nevertheless, the authors are able to extract postural information from the animals using a sophisticated machine vision based neural network. The authors use this system to continuously track the behaviors of individual larvae for six hours in the presence or absence of a thermal gradient. They argue that high inter-animal variability in a navigation index occurs in the presence of a thermal gradient but not in its absence. The intra-animal mean navigation also appears to be bimodal, apparently switching between "non-navigating" and "strongly navigating" states (not the authors' words). Interestingly, when only the population means are investigated a single mode is indicated with an overall weak navigation index. This comparison very nicely illustrates the power of this method to reveal richness in the data that leads to insights that cannot be observed with short-term measurements. Another impressive feature of the robotic system design is that it is capable of delivering small droplets of food to individual larvae. This allowed the authors to track a single larva for a remarkable 30 hours in which it is seen to crawl for more than 48 meters. Overall, the robotic system presented here will allow the researchers to investigate behaviors of larvae in long-term experiments in ways that were previously unimaginable.
-
Reviewer #3 (Public Review):
"Continuous, long-term crawling behavior characterized by a robotic transport system" by Yu et al. presents their new robotic device to track, reposition, and feed Drosophila larvae as they crawl on an arena. By using a water droplet (or if necessary, suction) to transport larvae from the edge of the arena to the middle, long behavior trajectories can be recorded without losing larvae from the arena or camera field of view. The picker robot is also able to dispense small amounts of apple juice at precise locations to keep larvae alive for extended periods although the food was not sufficient to trigger molting and the development to the next instar stage.
The approach is interesting, but the authors could provide more details on why the approach is necessary for non-expert readers. For example, what are the …
Reviewer #3 (Public Review):
"Continuous, long-term crawling behavior characterized by a robotic transport system" by Yu et al. presents their new robotic device to track, reposition, and feed Drosophila larvae as they crawl on an arena. By using a water droplet (or if necessary, suction) to transport larvae from the edge of the arena to the middle, long behavior trajectories can be recorded without losing larvae from the arena or camera field of view. The picker robot is also able to dispense small amounts of apple juice at precise locations to keep larvae alive for extended periods although the food was not sufficient to trigger molting and the development to the next instar stage.
The approach is interesting, but the authors could provide more details on why the approach is necessary for non-expert readers. For example, what are the advantages of using the robot picker compared to simply confining larvae in a closed arena? It's not obvious (to me) that being picked back to the center of the arena is a smaller perturbation compared to running into a chamber wall and changing direction.
The first paragraph of the introduction emphasizes the multiple time scales that are relevant for behavior from rapid stimulus response up to developmental times. This is to set the context of the authors' contribution but I'm not sure it's a fair representation of the state of the art. For example, the authors state that high-bandwidth measurement over long times is prohibitive and cite three Drosophila papers, but there are home-cage monitoring systems that allow continuous recording of mouse behavior over long times with high resolution. At the other end of the spectrum, there have been some long-term behaviour experiments done on worm behaviour with reasonably high time resolution (e.g Stern et al. 10.1016/j.cell.2017.10.041).
The authors train a neural network to segment and track the larvae, however, little information is given on the training process and I don't think it would be possible to reproduce the model based on the description. More details of the network, hyperparameters, and training data would be required to evaluate it.
The authors also state several times that larval identity is maintained throughout the recording, but this isn't quantified. It's not clear whether identity is maintained across collisions of two or more animals by the tracking algorithm or whether these collisions simply don't happen in their data because density is low.
The environment is nominally isotropic, but once larvae have been crawling on the surface for hours, including periodic feeding, there will likely be multiple gradients the larvae may sense. This may not be observable in the data, but should perhaps be mentioned in the text.
The authors show that the picking action results in a small but detectable increase in speed. The degree of perturbation overall depends on the picking frequency so some quantification of the inter-pick time interval would help to interpret whether this perturbation is relevant for a particular experiment. Is there a difference in excitation when larvae are picked successfully on the first try compared to when multiple tries or suction are required?
From the reconstructed trajectory in Figure 4, this interval looks very long compared to speed increase after picking. When reconstructing the trajectory, how are the segments joined? Is it simply by resetting the xy position or also updating rotating to match the previous direction of travel? (I'm guessing the larva can rotate during transport?)
The authors present a simple model in Figure 6 to illustrate the differences between individuals that can be hidden when looking at population distributions. However, the differences they show in the simulation don't seem relevant to the differences they observe in the experiments. Specifically, Fig. 6A and B show a contrast between individuals with similar mean speeds compared to individuals with different (but still unimodal) mean speeds. In contrast, the experimental data in Fig. D shows individual distributions that are quite similar but that are bimodal. So, there is indeed a difference between the individual distributions that is obscured in the population distribution, but is there evidence of larval personality types (line 444)? Similarly, the sentence beginning line 381 doesn't seem right either.
-
