Environmental statistics and sensory experience shape patch foraging strategies in Drosophila larvae

Reviewer #1 (Public review):

Summary:

Mudunuri et al. investigate the foraging response of Drosophila larvae in response to patchy resources of distinct value (concentration of nutrient or valence). They show that larvae adjust their behavior according to both the quality and valence of available resources. Interestingly, previous exposure to resources of lower value increases the permanence time in resources of greater value. This suggests that larvae can value, remember and adapt their behaviour in response to previous foraging experience.

They perform a simple integration model that recapitulates the larval behaviour.

Strengths:

This paper uses a very well-controlled foraging set-up where larvae are tested individually and for 3 hours, allowing for a good statistical analysis of their behaviour.

They investigate for the first time the ability of Drosophila larvae to perceive, remember and compare the quality and valence of distinct resources. It is very exciting, as it will open up the field of foraging decision studies using the fruitfly larvae.

Weaknesses:

(1) Most of the analysis depends on the thresholding, but it is not clear what increasing the radius of analysis means in terms of foraging. There are two issues here:

a) What is the behaviour of the larvae on the edges of the patch? It is obvious that the fructose or the NaCl will diffuse at the edge, so are they remaining in the proximity because they are actively feeding (exploiting) on this decaying concentration, or are they sensing the lower gradient and they are actually looking (chemosensing) for the higher concentration? The behaviour at the edge is really different (check sucrose in Wosniack et al. 2022), and there might be a way of avoiding the diffusion by actually adding a plastic ring and pouring the agar + resource in there. The effect of the ring, per se, would still have to be tested.

b) How was the threshold selected? It is very likely that the concentration at the patch boundary will be very different for 1M and 0.1 M. Could the authors explain why they chose such a distance? What does majority of larvae mean? Is the "majority" the same for 0.1M and 1M? Is there a relationship between the threshold chosen and the diffusion of fructose and NaCl?

(2) The word exploitation is used in the paper, but there are many instances where it is unclear whether that is the case. This should be clarified since there are no controls for exploitation.

(3) In the experiments analysing the adaptation of foraging behaviour, it is not clear if the first and second patch means that only 2 patches were analysed per larva or the first and second in a sequence of patches visited. I think it is the second option (because of Figure S3D), but the authors should clarify this. Also, we do not know how many animals were tested. The number of data points in 4C (4G) compared to 4D (4H) seems very different.
Regarding the results, which are very interesting, why aren't the larvae spending less time in the 0.1M sucrose patch after having fed on a 1M patch, while they spend more time in a 1M after a 0.1M? Could it be that the difference in residence time is correlated with their hunger rather than the comparison between conditions?

(4) I am not an expert in this type of model, and I would appreciate it if the authors could explain how the values of the drift and leak have been fitted in Figure 5H. If possible, I would recommend adding a graph showing the parameter exploration of distinct possible combinations of values.

Reviewer #2 (Public review):

Summary:

This manuscript investigates how Drosophila larvae make foraging decisions in patchy environments with controlled resource density and valence. Using movement tracking in bounded arenas, the authors show that larvae's patch residence time (PRT) differs depending on resource type, environmental context, and prior experience.

The authors vary whether the environment is homogenous (all patches are equal) or heterogenous (mixed patches) and whether a higher density of the resource is appetitive (food) or aversive (salt). The most salient results are that in heterogeneous environments, larvae remain longer on higher-density patches of fructose, while they stay shorter in higher-density salt patches. The study further demonstrates that prior foraging experience influences subsequent patch residence time (PRT).

A drift-diffusion model is used to describe patch-leaving behavior, suggesting that an integration process may underlie stay-leave decisions during foraging. Overall, the work provides a useful behavioral system for studying foraging behaviour and highlights the role of context and experience in shaping larval foraging strategies.

Strengths:

A major strength of the manuscript is the behavioral system. The assay is simple, well-controlled, and suitable for realistic spatial and temporal scale tracking of individual larvae. The use of non-volatile resources and embedded patches minimizes confounds from olfactory navigation and allows the authors to focus on local patch exploitation, return behavior, and experience-dependent decisions.

The results regarding patch resident time (how long larvae stay in patches of different resource density) are convincing. In homogeneous environments, larvae spend more time on patches with a higher density of food (0.1M > 0.01M) and less time in patches with a lower density of salt (0.01M > 0.1M), indicating that their behaviour is sensitive to the valence of the resource. Further, larvae do not simply respond to current circumstances, since PRT in a given patch is sensitive to the quality of the preceding one encountered, showing some kind of memory.

Weaknesses:

(1) The theoretical background of the experiment, as exposed in the Introduction, is somewhat misleading. The experiment is based on patches of sufficient size for the individual larvae not to deplete them through their activity, so that the intake rate is constant while exploiting a given patch. In those circumstances, the theoretical rate-maximizing strategy would be to either reject a patch on encounter or stay in it indefinitely (until pupation). The threshold for rejection or acceptance will depend on travel time, but patch residence time would be either zero (or minimal identification time) or lifelong. In the introduction, it appears as if the system follows the classical Marginal Value Theorem assumptions as used in classical foraging theory. In that case, patch residence time is fundamentally sensitive to a decline in intake rate while in a patch. This raises questions about what factors drive patch-leaving in the present protocol. A better theoretical framework would focus on behavioural variables that can be expected to depend on the circumstances of the experiment, as discussed below.

(2) Rather than make predictions about time in the patch, which as explained above do not reflect the present system, larval behaviour could be modelled and described as a function of observable properties such as: (a) speed of locomotion; (b) tendency to deviate from straight progress (area restricted searching); (c) probability of return after leaving a patch, possibly controlled through rea restricted searching; (d) a response to concentration gradient, since patch boundaries are probably gradual through diffusion. There is a useful literature in this regard in studies of parasitic wasps such as Venturia canescens (formerly Nemeritis canescens, see Waage 1979). Larva may respond directly to local resource concentration (see van Alphen, J. J., Bernstein, C., & Driessen, G., 2003), where higher concentration leads to increased feeding rate, reduced locomotion, and consequently results in longer time in each patch. This could still be a normative model, but based on realistic driving inputs. The dimensions of the system make it unlikely that larvae have the opportunity to adjust to travel time, or patch composition, on which classical foraging models are based. The original versions of the marginal value theorem were thought for cases where birds exploited pine cones, so that each bird had multiple encounters, and also on dung flies that mated in dung patches, which also dried out. A system with heritable optimised parameters could work for other natural systems where the parameters can be heritable, but not here.

(3) The previous argument indicates that patch time, while it is a real quantitative consequence, is not ideal as the major dependent variable for this system. Given that the authors have the full trajectories, they could treat movement in discrete time bins and ask if the tendency to depart from linear progression (i.e. from moving straight ahead) is a function of the density of the resource. It would appear as if all the results, including return to patches (but not memory), could be explained by area-restricted searching (see Dorfman, A., Hills, T. T., & Scharf, I. (2022). A guide to area‐restricted search: a foundational foraging behaviour. Biological Reviews, 97(6), 2076-2089.). Slower movement (perhaps directly caused by eating) and more twisted progress could generate longer times in higher food densities.

(4) The evidence for an effect of prior experience is interesting but could be strengthened. The authors state that PRT on the second patch depends on the concentration in the first patch. However, statistically significant modulation of prior experience was only found when the second food patch was richer, namely 1M fructose (Figure 4C). If the change in patch time is due to a form of learning and contrast, one might expect significantly shorter times in any second patch if the first one was richer, which is not the case. One difficulty is that the 'patchy' nature of the environment may not be evident to the larvae, because they are much smaller than the patches. From a larva's perspective, a patch is an environment, potentially suitable to remain in until pupation (which is what they ought to do in richer food patches).

(5) The modelling section is promising but currently somewhat underdeveloped relative to the strength of the claims. The authors fit a drift-diffusion model to data and report that a drift-only model captures homogeneous environments, whereas adding a leak term improves the fit in heterogeneous environments. This provides a useful quantitative summary of behavior but the biological interpretation of the leak parameter is not clear. In addition, the valence condition was not modelled.

Reviewer #3 (Public review):

Summary:

The work investigates how the foraging behaviour of Drosophila larvae depends on resource quality, valence, and heterogeneity in the foraging environment. A specific focus of the work was to study how foraging decisions depend on the prior experience of alternative resource patches in the same environment. Moreover, the work presents computational models (drift diffusion models) that recapitulate foraging decisions, and whose parameters appear to depend on resource quality and environment statistics, providing potential insights into the dynamics of the decision-making process.

I am not familiar with previous literature on foraging decisions in Drosophila, but I was specifically consulted to comment on the computational modelling. Therefore, my comments will mostly focus on the modelling aspects.

Strengths:

In my understanding, the two strengths of the current study are that:
(1) it uses non-volatile resources, providing better control of the available cues that could guide foraging decisions, and
(2) it tracks foraging behaviour over an extended period of time (3h), generating a rich dataset of foraging behaviour in the same environment.

Overall, the study appears to have been carefully conducted.

Weaknesses:

The computational modelling currently provides limited additional value beyond the empirical results. There are no prior hypotheses that are addressed by the computational models. Given the flexibility of DDMs, fitting foraging times is expected to be feasible. The question is whether the fits provide mechanistic insight. The main insight appears to be that describing foraging times in a homogeneous environment requires a single free parameter (drift rate), while the heterogenous environment requires a second parameter (leak). However, the effective complexity of the model is higher than the stated parameter count suggests, as each patch quality is fit with a different drift rate, which does not generalise across environments: in the heterogeneous environment, the drift rate differs substantially across fructose concentrations, whereas in the homogeneous environment, the same concentrations yield nearly identical drift rates. Counter their claims, the authors also do not systematically explore the effect of specific prior foraging experience on computational parameters, but only contrast model fits to environments with different statistics, in which prior experiences will be generally different. Overall, at the moment these modelling results have a rather descriptive character, and provide very little insight into the underlying computational principles that drive foraging decisions.

A second weakness is that the study does not report the detailed results of the statistical tests, and it seems that the authors interpret several differences that are not marked as statistically significant in the figures. Furthermore, the model comparisons do not account for different degrees of freedom of the models, and the goodness of fit values alone are insufficient to conclude that one model is better than the other (rather than overfitting).