Author response:
The following is the authors’ response to the original reviews
Public Reviews:
Reviewer #1 (Public review):
(1) The authors repeatedly assert that an individual's behavior in the foraging assay depends on its prior history (particularly cultivation conditions). While this seems like a reasonable expectation, it is not fully fleshed out. The work would benefit from studies in which animals are raised on more or less abundant food before the behavioral task.
Cultivation density: While we agree with the reviewer that testing the effects of varying bacterial density during animal development (cultivation) is an interesting experiment, it is not feasible at this time. We previously attempted this experiment but found it nontrivial to maintain stable bacterial density conditions over long timescales as this requires matching the rate of bacterial growth with the rate of bacterial consumption. Despite our best efforts, we have not been able to identify conditions that satisfy these requirements. Thus, we focused our revised manuscript to include only assertions about the effects of recent experiences and added this inquiry as a future direction (lines 618-624).
(2) The authors convincingly show that the probability of particular behavioral outcomes occurring upon patch encounter depends on time-associated parameters (time since last patch encounter, time since last patch exploitation). There are two concerns here. First, it is not clear how these values are initialized - i.e., what values are used for the first occurrence of each behavioral state? More importantly, the authors don't seem to consider the simplest time parameter, the time since the start of the assay (or time since worm transfer). Transferring animals to a new environment can be associated with significant mechanical stimulus, and it seems quite possible that transferring animals causes them to enter a state of arousal. This arousal, which certainly could alter sensory function or decision-making, would likely decay with time. It would be interesting to know how well the model performs using time since assay starts as the only time-dependent parameter.
Parameter Initialization: We thank the reviewer for pointing out an oversight in our methods section regarding the model parameter values used for the first encounter. We clarified the initialization of parameters in the manuscript (lines 1162-1179). In short, for the first patch encounter where k = 1:
● ρk is the relative density of the first patch.
● τs is the duration of time spent off food since the beginning of the recorded experiment. For the first patch, this is equivalent to the total time elapsed.
● ρh is the approximated relative density of the bacterial patch on the acclimation plates (see Assay preparation and recording in Methods). Acclimation plates contained one large 200 µL patch seeded with OD600 = 1 and grown for a total of ~48 hours. As with all patches, the relative density was estimated from experiments using fluorescent bacteria OP50-GFP as described in Bacterial patch density estimation in Methods.
● ρe is equivalent to ρh.
Transfer Method: We thank the reviewer for their thoughtful comment on how the stress of transferring animals to a new plate may have resulted in an increased arousal state and thus a greater probability of rejecting patches. We anticipated this possibility and, in order to mitigate the stress of moving, we used an agar plug method where animals were transferred using the flat surface of small cylinders of agar. Importantly, the use of agar as a medium to transfer animals provides minimal disruption to their environment as all physical properties (e.g. temperature, humidity, surface tension) are maintained. Qualitatively, we observed no marked change in behavior from before to after transfer with the agar plug method, especially as compared to the often drastic changes observed when using a metal or eyelash pick. We added these additional methodological details to the methods (lines 791-796).
Time Parameter: However, the reviewer’s concern that the simplest time parameter (time since start of the assay) might better predict animal behavior is valid. We thank the reviewer for pointing out the need to specifically test whether the time-dependent change in explore-exploit decision-making corresponds better with satiety (time off patch) or arousal (time since transfer/start of assay) state. To test this hypothesis, we ran our model with varying combinations of the satiety term τs and a transfer term τt. We found that when both terms were included in the model, the coefficient of the transfer term was non-significant. This result suggests that the relevant time-dependent term is more likely related to satiety than transfer-induced stress (lines 343-358; Figure 4 - supplement 4D).
(3) Similarly, Figures 2L and M clearly show that the probability of a search event occurring upon a patch encounter decreases markedly with time. Because search events are interpreted as a failure to detect a patch, this implies that the detection of (dilute) patches becomes more efficient with time. It would be useful for the authors to consider this possibility as well as potential explanations, which might be related to the point above.
Time-dependent changes in sensing: We agree with the reviewer that we observe increased responsiveness to dilute patches with time. Although this is interesting, our primary focus was on what decision an animal made given that they clearly sensed the presence of the bacterial patch. Nonetheless, we added this observation to the discussion as an area of future work to investigate the sensory mechanisms behind this effect (lines 563-568).
(4) Based on their results with mec-4 and osm-6 mutants, the authors assert that chemosensation, rather than mechanosensation, likely accounts for animals' ability to measure patch density. This argument is not well-supported: mec-4 is required only for the function of the six non-ciliated light-touch neurons (AVM, PVM, ALML/R, PLML/R). In contrast, osm-6 is expected to disrupt the function of the ciliated dopaminergic mechanosensory neurons CEP, ADE, and PDE, which have previously been shown to detect the presence of bacteria (Sawin et al 2000). Thus, the paper's results are entirely consistent with an important role of mechanosensation in detecting bacterial abundance. Along these lines, it would be useful for the authors to speculate on why osm-6 mutants are more, rather than less, likely to "accept" when encountering a patch.
Sensory mutant behavior: We thank the reviewer for pointing out the error in our interpretation of the behavior of osm-6 and mec-4 animals. We further elaborated on our findings and edited the text to better reflect that osm-6 mutants lack both chemosensory and mechanosensory ciliated sensory neurons (lines 406-448; lines 567-577). Specifically, we provided some commentary on the finding that osm-6 mutants show an augmented ability to detect the presence of bacterial patches but a reduced ability to assess their bacterial density. While this finding seems contradictory, it suggests that in the absence of the ability to assess bacterial density, animals must prioritize exploiting food resources when available.
(5) While the evidence for the accept-reject framework is strong, it would be useful for the authors to provide a bit more discussion about the null hypothesis and associated expectations. In other words, what would worm behavior in this assay look like if animals were not able to make accept-reject decisions, relying only on exploit-explore decisions that depend on modulation of food-leaving probability?
Accept-reject vs. stay-switch: We thank the reviewer for alerting us to this gap in our discussion. We have revised the text to further extrapolate upon our point of view on this somewhat philosophical distinction and what it predicts about C. elegans behavior (lines 507-533).
Reviewer #3 (Public review):
(1) Sensing vs. non-sensing
The authors claim that when animals encounter dilute food patches, they do not sense them, as evidenced by the shallow deceleration that occurs when animals encounter these patches. This seems ethologically inaccurate. There is a critical difference between not sensing a stimulus, and not reacting to it. Animals sense numerous stimuli from their environment, but often only behaviorally respond to a fraction of them, depending on their attention and arousal state. With regard to C. elegans, it is well-established that their amphid chemosensory neurons are capable of detecting very dilute concentrations of odors. In addition, the authors provide evidence that osm-6 animals have altered exploit behaviors, further supporting the importance of amphid chemosensory neurons in this behavior.
Interpretation of “non-sensing” encounters: We thank the reviewer for their comment and agree that we do not know for certain whether the animals sensed these patches or were merely non-responsive to them. We are, however, confident that these encounters lack evidence of sensing. Specifically, we note that our analyses used to classify events as sensing or non-sensing examined whether an animal’s slow-down upon patch entry could be distinguished from either that of events where animals exploited or that of encounters with patches lacking bacteria. We found that “non-sensing” encounters are indeed indistinguishable from encounters with bacteria-free patches where there are no bacteria to be sensed (see Figure 2 - Supplement 8A-C and Patch encounter classification as sensing or non-responding in Methods). Regardless, we agree with the reviewer that all that can be asserted about these events is that animals do not appear to respond to the bacterial patch in any way that we measured. Therefore, we have replaced the term “non-sensing” with “non-responding” to better indicate the ethological interpretation of these events and clarified the text to reflect this change (lines 193-200; lines 211-212).
(2) Search vs. sample & sensing vs. non-sensing
In Figures 2H and 2I, the authors claim that there are three behavioral states based on quantifying average velocity, encounter duration, and acceleration, but I only see three. Based on density distributions alone, there really only seem to be 2 distributions, not 3. The authors claim there are three, but to come to this conclusion, they used a QDA, which inherently is based on the authors training the model to detect three states based on prior annotations. Did the authors perform a model test, such as the Bayesian Information Criterion, to confirm whether 2 vs. 3 Gaussians is statistically significant? It seems like the authors are trying to impose two states on a phenomenon with a broad distribution. This seems very similar to the results observed for roaming vs. dwelling experiments, which again, are essentially two behavioral states.
Validation of sensing clusters: We are grateful to the reviewer for pointing out the difficulty in visualizing the clusters and the need for additional clarity in explaining the semi-supervised QDA approach. We added additional visualizations and methods to validate the clusters we have discovered. Specifically, we used Silverman’s test to show that the sensing vs. non-responding data were bi-modal (i.e. a two-cluster classification method fits best) and accompanied this statistical test with heat maps which better illustrate the clusters (lines 171-173; lines 190-191; lines 948-972; lines 1003-1005; Figure 2 - supplement 6A-C; Figure 2 - supplement 7C-F).
Further, it seems that there may be some confusion as to how we arrived at 3 encounter types (i.e. search, sample, exploit). It’s important to note that two methods were used on two different (albeit related) sets of parameters. We first used a two-cluster GMM to classify encounters as explore or exploit. We then used a two-cluster semi-supervised QDA to classify encounters as sensing or non-sensing (now changed to “non-responding”, see above response) using a different set of parameters. We thus separated the explore cluster into two (sensing and non-responding exploratory events) resulting in three total encounter types: exploit, sample (explore/sensing), and search (explore/non-sensing).
(4) History-dependence of the GLM
The logistic GLM seems like a logical way to model a binary choice, and I think the parameters you chose are certainly important. However, the framing of them seems odd to me. I do not doubt the animals are assessing the current state of the patch with an assessment of past experience; that makes perfect logical sense. However, it seems odd to reduce past experience to the categories of recently exploited patch, recently encountered patch, and time since last exploitation. This implies the animals have some way of discriminating these past patch experiences and committing them to memory. Also, it seems logical that the time on these patches, not just their density, should also matter, just as the time without food matters. Time is inherent to memory. This model also imposes a prior categorization in trying to distinguish between sensed vs. not-sensed patches, which I criticized earlier. Only "sensed" patches are used in the model, but it is questionable whether worms genuinely do not "sense" these patches.
Model design: We thank the reviewer for their thoughtful comments on the model. We completed a number of analyses involving model selection including model selection criteria (AIC, BIC) and optimization with regularization techniques (LASSO and elastic nets) and found that the problem of model selection was compounded by the enormous array of highly-correlated variables we had to choose from. Additionally, we found that both interaction terms and non-linear terms of our task variables could be predictive of accept-reject decisions but that the precise set of terms selected depended sensitively on which model selection technique was used and generally made rather small contributions to prediction. The diverse array of results and combinatorial number of predictors to possibly include failed to add anything of interpretable value. We therefore chose to take a different approach to this problem. Rather than trying to determine what the “best” model was we instead asked whether a minimal model could be used to answer a set of core questions. Indeed, our goal was not maximal predictive performance but rather to distinguish between the effects of different influences enough to determine if encounter history had a significant, independent effect on decision making. We thus chose to only include task variables that spanned the most basic components of behavioral mechanisms to ask very specific questions. For example, we selected a time variable that we thought best encapsulated satiety. While we could have included many additional terms, or made different choices about which terms to include, based on our analyses these choices would not have qualitatively changed our results. Further, we sought to validate the parameters we chose with additional studies (i.e. food-deprived and sensory mutant animals). We regard our study as an initial foray into demonstrating accept-reject decision-making in nematodes. The exact mechanisms and, consequently, the best model design are therefore beyond the scope of this study.
Lastly, in regards to the use of only sensed patches in the model; while we acknowledge that we are not certain as to whether the “non-responding” encounters are truly not sensed, we find qualitatively similar results when including all exploratory patches in our analyses. However, we take the position that sensation is necessary for decision-making and thus believe that while our model’s predictive performance may be better using all encounters, the interpretation of our findings is stronger when we only include sensing events. We have added additional commentary about our model to the discussion section (lines 667-695).
(5) osm-6
The osm-6 results are interesting. This seems to indicate that the worms are still sensing the food, but are unable to assess quality, therefore the default response is to exploit. How do you think the worms are sensing the food? Clearly, they sense it, but without the amphid sensory neurons, and not mechanosensation. Perhaps feeding is important? Could you speculate on this?
We thank the reviewer for their thoughtful remarks. We have added additional commentary about the result of our sensory mutant experiments as described above in response to Reviewer #1 under Sensory mutant behavior.
(7) Impact:
I think this work will have a solid impact on the field, as it provides tangible variables to test how animals assess their environment and decide to exploit resources. I think the strength of this research could be strengthened by a reassessment of their model that would both simplify it and provide testable timescales of satiety/starvation memory.
Recommendations for the authors:
Reviewer #2 (Recommendations for the authors):
The authors title the work as an "ethological study" and emphasize the theme of "foraging in naturalistic environments" in contrast to typical laboratory conditions. The only difference in this study relative to typical laboratory conditions is that the food bacteria is distributed in many small patches as compared to one large patch. First, it is not clear to the reviewer that the size of the food patches in these experiments is more relevant to C. elegans in its natural context than the standard sizes of food patches. Furthermore, all the other highly unnatural conditions typical of laboratory cultivation still apply: the use of a 2D agar substrate, a single food bacteria that is not a component of a naturalistic diet, and the use of a laboratory-adapted strain of C. elegans with behavior quite distinct from that of natural isolates. The reviewer is not suggesting that the authors need to make their experiments more naturalistic, only that the experiments as described here should not be described as naturalistic or ethological as there is no support for such claims.
Ethological interpretation: We thank the reviewer for their comments about the use of the term ethological to describe this study. We chose to develop a patchy bacterial assay to mimic the naturalistic “boom-or-bust” environment. While we agree with the reviewer that we do not know if the size and distribution of the food patches in these experiments is more relevant to C. elegans, we maintain that these experiments were ecologically-inspired and revealed behavior that is difficult to observe in environments with large, densely-seeded bacterial patches. We have updated our text to better reflect that this study was “ecologically-inspired” rather than truly “ethological” in nature (lines 94, 693).
The main finding of the paper is that worms explore and then exploit, i.e. they frequently reject several bacterial patches before accepting one. This result requires additional scrutiny to reject other possible interpretations. In particular, when worms are transferred to a new plate we would expect some period of increased arousal due to the stressful handling process. A high arousal state might cause rejection of food patches. Could the measured accept/reject decisions be influenced by this effect? One approach to addressing this concern would be to allow the animals to acclimate to the new plate on a bare region before encountering the new food patches.
We thank the reviewer for their comment on how the stress of transferring animals to a new plate may have resulted in an increased arousal state and thus a greater probability of rejecting patches. We addressed this above in response to Reviewer #1 under Transfer Method and Time Parameter. In brief, we used a worm picking method that mitigated stress and added additional analyses showing that a transfer-related term was less predictive than a satiety-related term.
Related to the above, in what circumstances exactly are the authors claiming that worms first explore and then exploit? After being briefly deprived of food? After being handled?
Explore-then-exploit: All animals were well-fed and handled gently as described above under Transfer Method (lines 787-795). Our results suggest that the appearance of an explore-then-exploit strategy is a byproduct of being transferred from an environment with high bacterial density to an environment with low bacterial density as described in the manuscript (lines 461-466).
The authors emphasize their analysis of the accept/reject decision as a critical innovation. However, the accept/reject decision does not strike me as substantially different from the previously described stay/switch decision. When a worm encounters a new patch of bacteria, accepting this bacteria is equivalent to staying on it and rejecting (leaving) it is equivalent to switching away from it. The authors should explain how these concepts are significantly distinct.
Accept-reject vs. stay-switch: We thank the reviewer for alerting us to this gap in our discussion. We have revised the text to further extrapolate upon our point of view on this somewhat philosophical distinction and what it predicts about C. elegans behavior (lines 507-533).
During patch encounter classification, the authors computed three of the animals' behavioral metrics (Line 801-804) and claimed that the combination of these three metrics reveals two non-Gaussian clusters representing encounters where animals sensed the patch or did not appear to sense the patch. The authors also refer to a video to demonstrate the two clusters by rotating the 3-dimension scatter plot. However, the supposed clusters, if any, are difficult to see in a 3D (Video 5) or in a 2D scatter plot (Figure 3I). The authors need to clearly demonstrate the distinct clustering as claimed in the paper as this feature is fundamental and necessary for the model implementation and interpretation of results.
We are grateful to the reviewer for pointing out the difficulty in visualizing the clusters. We added additional visualizations and methods to validate the clusters we have discovered as described in our above response to Reviewer #3 under Validation of sensing clusters.
When selecting parameters (covariates) for their model, it is critical to avoid overfitting. Therefore, the authors used AIC and BIC (Figure 4- supplement 1) to demonstrate that the full GLM model has a better model performance than the other models which contain only a subset of the full covariates (in a total of 5). However, the authors compare the full set with only 4 other models whereas the total number of models that need to be compared with is 2^5-2. The authors at least need to include the AIC and BIC scores of all possible models in order to draw the conclusion about the performance of the full model.
Model selection criterion: We thank the reviewer for pointing out this gap in our methodology. We have now run the model with all combinations of subsets of model parameters and have confirmed that the model with all 5 covariates outperforms all other models even when using BIC, the strictest criterion for overfitting (Figure 1 - supplement 1A). The only other model that performs well (though not as often as the 5-term model) is the 4-term model lacking ρh. This result is not surprising as ρh only changes substantially once in an animal’s encounter history for the single-density, multi-patch data that this model was fit to. For example, for an animal foraging on patches of density 10, on the first encounter ρh = ~200 (see Parameter initialization above), but on every subsequent encounter ρh = ~10. Resultantly, the effect of ρh on the probability of exploiting is somewhat binary on the single-density, multi-patch data set. Nevertheless, we see significantly improved prediction of behavior in the novel multi-density, multi-patch data (Figure 4F) as we observe an effect of the most recently encountered patch. Additionally, we observe a similar impact (i.e., significant coefficient of negative sign) of the ρh term when the model is fit to the multi-density, multi-patch data set (Figure 4 - supplement 4D).
In any bacterial patch, the edges have a higher density of bacteria than the patch center. Thus, it is possible that a worm scans the patch edge density, on the basis of which it decides to accept or reject the patch whose average density is smaller. This could potentially cause an underestimate of the bacteria density used in the model. Furthermore, the potential inhomogeneity of the patch may further complicate the worm's decision-making, and the discrepancy between the reality and the model assumption will reduce the validity of the model. The authors need to estimate the inhomogeneity of the bacterial patches used in their assays and discuss how the edge effects may affect their results and conclusions.
Bacterial patch inhomogeneity: We extensively tested the landscape of the bacterial patches by imaging fluorescently-labeled bacteria OP50-GFP (Bacterial Patch Density in Methods; Figure 2 - supplement 1-3). As the reviewer mentions, we observe significantly greater bacterial density at the patch edge. This within-patch spatial inhomogeneity results from areas of active proliferation of bacteria and likely complicates an animal’s ability to accurately assess the quantity of bacteria within a patch and, consequently, our ability to accurately compute a metric related to our assumptions of what the animal is sensing. In our study, we used the relative density of the patch edge where bacterial density is highest as a proxy for an animal’s assessment of bacterial patch density (Figure 2 – supplement 1). This decision was based on a previous finding that the time spent on the edge of a bacterial patch affected the dynamics of subsequent area-restricted search. While within-patch spatial inhomogeneity likely affects an animal’s ability to assess patch density, we do not believe that this qualitatively affects the results of our study. Both the patch densities tested (Figure 2 – supplement 3A) as well as our observations of time-dependent changes in exploitation (Figure 2E,N-O; Figure 3H-I) maintained a monotonic relationship. Therefore, alternative methods of patch density estimation should yield similar results. We have added additional discussion on this topic to our manuscript (lines 578-593).
The authors claim that their methods (GMM and semi-supervised QDA) are unbiased. This seems unlikely as the QDA involves supervision. The authors need to provide additional explanation on this point.
Semi-supervised QDA labelling: We have removed the term “unbiased” to avoid any misinterpretation of the methodology and clarified our method of labelling used for “supervising” QDA. Specifically, we made two simple assumptions: 1) animals must have sensed the patch if they exploited it and 2) animals must not have sensed the patch if there was no bacteria to sense. Thus, we labeled encounters as sensing if they were found to be exploitatory as we assume that sensation is prerequisite to exploitation; and we labeled encounters as non-sensing for events where animals encountered patches lacking bacteria (OD600 = 0). All other points were non-labeled prior to learning the model. In this way, our labels were based on the experimental design and results of the GMM, an unsupervised method; rather than any expectations we had about what sensing should look like. The semi-supervised QDA method then used these initial labels to iteratively fit a paraboloid that best separated these clusters, by minimizing the posterior variance of classification (lines 1012-1021). See Figure 2 - supplement 8A-B for a visualization showing the labelled data.
Based on the authors' result, worms behaviorally exhibit their preferences toward food abundance (density), which results in a preference scale for a range of densities. Does this scale vary with the worms' initial cultivation states? The author partially verified that by observing starved worms. This hypothesis could be better tested if the authors could analyze the decision-making of the worms that were initially cultivated with different densities of bacterial food.
While we agree with the reviewer that testing the effects of varying bacterial density during animal development (cultivation) is a very interesting experiment, it is not feasible at this time. We focused our revised manuscript to include only assertions about the effects of recent experiences and added this inquiry as a future direction as described above in our response to Reviewer #1 under Cultivation density.
It would be helpful to elaborate more on how the framework developed in this paper can be applied more broadly to other behaviors and/or organisms and how it may influence our understanding of decision-making across species.
We thank the reviewer for alerting us to this gap in our discussion. We have added additional commentary about our model and its utility to the discussion section (lines 667-695).
Reviewer #3 (Recommendations for the authors):
Sensing vs. non-sensing
Perhaps a more ethologically accurate term to describe this behavior would be "ignoring" rather than "not sensing". If the authors feel strongly about using the term "not sensing", then they should provide experimental evidence supporting this claim. However, I think simply changing the terminology negates these experiments.
We thank the reviewer for their thoughtful comments. While we agree with the reviewer that the term “non-sensing” may not be ethologically accurate (see response to Public Review above under Interpretation of “non-sensing” encounters), we interpret the term “ignoring” to mean that the animal sensed the patches but decided not to react. We have chosen to replace the term “non-sensing” with “non-responding” to best indicate the ethological interpretation of our observation. Nonetheless, we believe that it remains possible that animals are truly not sensing the bacterial patches as our method of classification compared the behavior against encounters with patches lacking bacteria (as described above in response to Reviewer #2 under Semi-supervised QDA labelling).
History-dependence of the GLM
Perhaps a simpler approach would be to say the worm senses everything, and this accumulative memory affects the decision to exploit. For example, the animal essentially experiences two feeding states: feeding on patches, and starvation off of patches.
The level of satiety could be modeled linearly:
Satiety(t_enter:t_leave) = k_feed*patch_density*delta_t
Where k_feed is some model parameter for rate of satiety signal accumulation, t_enter is the time the animal entered the patch, t_leave is the time the animal left the patch, and delta_t is the difference between the two. Perhaps you could add a saturation limit to this, but given your data, I doubt that is the case.
Starvation could be modeled as simply a decay from the last satiety signal:
Starvation(t_leave:t_enter) = Satiety(t_leave)*exp(-k_starve*delta_t).
Where starvation is the rate constant for the decay of the satiety signal.
For the logistic model, the logistic parameter is simply the difference between the current patch density and the current satiety signal.
A nice thing about this approach is that it negates the need to categorize your patches. All patch encounters matter. Brief patch encounters (categorized as non-sensing and not used in the prior GLM) naturally produce a very small satiety signal and contribute very little to the exploit decision. Another nice thing about this approach is that it gives you memory timescales, that are testable. There is a rate of satiety accumulation and a rate of satiety loss. You should be able to predict behavior with lower patch density, assuming the rate constants hold. (I am not advocating you do more experiments here, just pointing out a nice feature of this approach).
You could possibly apply this to a GLM for velocity on a non-exploited patch as well, though I assume this would be a linear GLM, given the velocity distributions you provided.
We thank the reviewer for their time and thoughtfulness in thinking about our model. The reviewer’s proposed model seems entirely reasonable and could aid in elucidating the time component of how prior experience affects decision-making. However, we decided to keep our paper focused on using a minimal model to answer a set of core questions (e.g., Does encounter history or satiety influence decision-making?) (see above under Model design for a more detailed response). Future studies investigating the mechanisms of these foraging decisions should open the door for more mechanistically accurate models. We have expanded our discussion of the model to include this assertion (lines 667-695).
