Stochastic characterization of navigation strategies in an automated variant of the Barnes maze

  1. Ju-Young Lee
  2. Dahee Jung
  3. Sebastien Royer

  • Curated by eLife

    eLife logo

    eLife assessment

    This study presents a valuable new behavioral apparatus aimed at differentiating the strategies animals use to orient themselves in an environment. The evidence supporting the claims is solid, with statistical modeling of animal behavior. Overall, this study will attract the interest of researchers exploring spatial learning and memory.

Reviewed by

Read the full article

Listed in

Log in to save this article

Abstract

Animals can use a repertoire of strategies to navigate in an environment, and it remains an intriguing question how these strategies are selected based on the nature and familiarity of environments. To investigate this question, we developed a fully automated variant of the Barnes maze, characterized by 24 vestibules distributed along the periphery of a circular arena, and monitored the trajectories of mice over 15 days as they learned to navigate towards a goal vestibule from a random start vestibule. We show that the patterns of vestibule visits can be reproduced by the combination of three stochastic processes reminiscent of random, serial, and spatial strategies. The processes randomly selected vestibules based on either uniform (random) or biased (serial and spatial) probability distributions. They closely matched experimental data across a range of statistical distributions characterizing the length, distribution, step size, direction, and stereotypy of vestibule sequences, revealing a shift from random to spatial and serial strategies over time, with a strategy switch occurring approximately every six vestibule visits. Our study provides a novel apparatus and analysis toolset for tracking the repertoire of navigation strategies and demonstrates that a set of stochastic processes can largely account for exploration patterns in the Barnes maze.

Article activity feed

  1. Version published to 10.7554/elife.88648.4 on eLife
  2. Version published to 10.7554/elife.88648 on eLife
  3. Version published to 10.7554/elife.88648.3 on eLife
  4. eLife

    Author response:

    The following is the authors’ response to the previous reviews.

    Public Reviews:

    Reviewer #1 (Public Review):

    The authors design an automated 24-well Barnes maze with 2 orienting cues inside the maze, then model what strategies the mice use to reach the goal location across multiple days of learning. They consider a set of models and conclude that the animals begin with a large proportion of random choices (choices irrespective of the goal location), which over days of experience becomes a combination of spatial choices (choices targeted around the goal location) and serial choices (successive stepwise choices in a given direction). Moreover, the authors show that after the animal has many days of experience in the maze, they still often began each trial with a random choice, followed by spatial or serial choices.

    This study is written concisely and the results are presented concisely. The best fit model provides valuable insight into how the animals solve this task, and therefore offers a quantitative foundation upon which tests of neural mechanisms of the components of the behavioral strategy can be performed. These tests will also benefit from the automated nature of the task.

    Reviewer #2 (Public Review):

    This paper uses a novel maze design to explore mouse navigation behaviour in an automated analogue of the Barnes maze. A major strength is the novel and clever experimental design which rotates the floor and intramaze cues before the start of each new trial, allowing the previous goal location to become the next starting position. The modelling sampling a Markov chain of navigation strategies is elegant, appropriate and solid, appearing to capture the behavioural data well. This work provides a valuable contribution and I'm excited to see further developments, such as neural correlates of the different strategies and switches between them.

    Reviewer #3 (Public Review):

    Strength:

    The development of an automated Barnes maze allows for more naturalistic and uninterrupted behavior, facilitating the study of spatial learning and memory, as well as the analysis of the brain's neural networks during behavior when combined with neurophysiological techniques. The system's design has been thoughtfully considered, encompassing numerous intricate details. These details include the incorporation of flexible options for selecting start, goal, and proximal landmark positions, the inclusion of a rotating platform to prevent the accumulation of olfactory cues, and careful attention given to atomization, taking into account specific considerations such as the rotation of the maze without causing wire shortage or breakage. When combined with neurophysiological manipulations or recordings, the system provides a powerful tool for studying spatial navigation system.

    The behavioral experiment protocols, along with the analysis of animal behavior, are conducted with care, and the development of behavioral modeling to capture the animal's search strategy is thoughtfully executed. It is intriguing to observe how the integration of these innovative stochastic models can elucidate the evolution of mice's search strategy within a variant of the Barnes maze.

    Comments on revised version:

    The authors have addressed all the points I outlined in the previous round of review, resulting in significant improvements to the manuscript. However, I have one remaining comment. Given the updated inter-animal analysis (Supplementary Figure 8), it appears that male and female mice develop strategies differently across days. Male mice seem to progressively increase their employment of spatial strategy across days, at the expense of the random strategy. Conversely, female mice exhibit both spatial and serial strategies at their highest levels on day 2, with minimal changes observed on the subsequent days.

    These findings could alter the interpretation of Figure 5 and the corresponding text in the section "Evolution of search strategy across days".

    For instance, this statement on page 6 doesn't hold for female mice: "The spatial strategy was increased across days, ... largely at the expense of the random strategy."

    We agree with the reviewer. While the text on page 6 is still valid for the male-female pooled data, we have clarified in the next section describing male-female differences that this trend is not observed in female. Furthermore, we adjusted the relevant part of the discussion the following manner:

    “A shift in the proportion of random, spatial and serial strategies was observed across days. Several factors might contribute to this shift, including learning of the environment and goal location, changes in motivation for exploration versus goal-directed navigation, and the evaluation of each strategy’s benefit via reinforcement learning. The spatial strategy progressively increased, mostly at the expense of the random strategy. This trend suggests a diminishing interest in exploration and an increasing benefit from employing the spatial strategy as the mice became more familiar with the environment and goal location. Consistent with this hypothesis, the development of the spatial strategy approximately matched the development of spatial maps in the hippocampus37 and the growth pattern of hippocampal feedforward inhibitory connectivity62, both showing progressive increases that reached plateaus after a week. In contrast, the serial strategy showed a sudden increase from day 1 to day 2, indicating that this goal-directed strategy is associated with rapid learning and could already be reinforced on day 2. However, the strategy shift was not uniform across the mouse population, as male and female mice showed distinct trends. Female mice showed no progressive increase in spatial strategy and initially relied more on the spatial strategy while using the random strategy less compared to male mice. This difference might be explained by faster learning of goal location and/or a stronger inclination towards goal-directed navigation over exploration in female mice.”

    Recommendations for the authors:

    Reviewer #1 (Recommendations For The Authors):

    Minor points:

    (1) The following sentence in the abstract is not grammatical: "The processes randomly selected vestibules based on either uniform (random) or biased (serial and spatial) probability distributions; closely matched experimental data across a range of statistical distributions characterizing the length, distribution, step size, direction, and stereotypy of vestibule sequences; and revealed a shift from random to spatial and serial strategies over time, with a strategy switch occurring approximately every 6 vestibule visits."

    One possible revision is: "The processes randomly selected vestibules based on either uniform (random) or biased (serial and spatial) probability distributions; [they] closely matched experimental data across a range of statistical distributions characterizing the length, distribution, step size, direction, and stereotypy of vestibule sequences, [revealing] a shift from random to spatial and serial strategies over time, with a strategy switch occurring approximately every 6 vestibule visits."

    We followed the reviewer’s suggestion.

    (2) There is a missing word in the following sentence in the last paragraph of the discussion: "Our tools might be combined in the future with optogenetic and/or pharmacogenetic [missing word here] to investigate the neural mechanisms underlying strategy selection"

    We added the word ‘manipulations’: ‘… optogenetic, pharmacogenetic manipulations …’

    Reviewer #2 (Recommendations For The Authors):

    I have two minor suggestions:

    (1) Results - Automated Maze section: It would be beneficial to clarify here that the floor and cues rotate allowing automation by chining start/end positions together. This information is key to the reader understanding the task and currently they would only know this by studying fig1 or delving into the methods

    As suggested by the reviewer, we have added the following text in the Results - Automated Maze section:

    “The maze consist of an enclosed arena with an array of 24 doors evenly spaced along the periphery, and two home boxes moving around the arena perimeter. Start positions are changed by rotating the arena and the home boxes (Fig. 1b). Furthermore, the arena has a tinted cover that prevents mice from seeing room cues while still allowing for infrared tracking of mouse trajectories.”

    (2) I still find the author's decision to exclude days from some of the line plots, e.g. days 3,4,5 from Fig2 etc, a little odd as this makes the reader wary. I appreciate their argument about clarity, but this can still be achieved while partitioning all of the data rather than excluding certain days. NB I do not find the heat map distributions in the far panel a particularly good substitute for this as pixel intensities are far less interpretable

    We appreciate the reviewer’s comment. We want to point out that line plots for all individual days are actually displayed in Supplementary Figure 7a.

    Reviewer #3 (Recommendations For The Authors):

    Although the difference between females and males is clear in Figure S8b, please note that the statistics in panels C and D might not be appropriate, as many of them may become insignificant if adjusted for multiple comparisons.

    If we understand correctly, a Bonferroni correction would need to consider the 3 day intervals in Figure S8c and the 2 day groups in Figure S8d. This would mean a significance threshold of 0.05/3 = 0.016667 for Figure S8c and 0.05/2 = 0.025 for Figure S8d, after Bonferroni correction. As it stands, all comparisons that are not labelled ’ns’ in Figure S8c-d remain significant even after applying the Bonferroni correction.

  5. eLife

    eLife assessment

    This study presents a valuable new behavioral apparatus aimed at differentiating the strategies animals use to orient themselves in an environment. The evidence supporting the claims is solid, with statistical modeling of animal behavior. Overall, this study will attract the interest of researchers exploring spatial learning and memory.

  6. eLife

    Reviewer #1 (Public Review):

    The authors design an automated 24-well Barnes maze with 2 orienting cues inside the maze, then model what strategies the mice use to reach the goal location across multiple days of learning. They consider a set of models and conclude that the animals begin with a large proportion of random choices (choices irrespective of the goal location), which over days of experience becomes a combination of spatial choices (choices targeted around the goal location) and serial choices (successive stepwise choices in a given direction). Moreover, the authors show that after the animal has many days of experience in the maze, they still often began each trial with a random choice, followed by spatial or serial choices.

    This study is written concisely and the results are presented concisely. The best fit model provides valuable insight into how the animals solve this task, and therefore offers a quantitative foundation upon which tests of neural mechanisms of the components of the behavioral strategy can be performed. These tests will also benefit from the automated nature of the task.

  7. eLife

    Reviewer #2 (Public Review):

    This paper uses a novel maze design to explore mouse navigation behaviour in an automated analogue of the Barnes maze. A major strength is the novel and clever experimental design which rotates the floor and intramaze cues before the start of each new trial, allowing the previous goal location to become the next starting position. The modelling sampling a Markov chain of navigation strategies is elegant, appropriate and solid, appearing to capture the behavioural data well. This work provides a valuable contribution and I'm excited to see further developments, such as neural correlates of the different strategies and switches between them.

  8. eLife

    Reviewer #3 (Public Review):

    The development of an automated Barnes maze allows for more naturalistic and uninterrupted behavior, facilitating the study of spatial learning and memory, as well as the analysis of the brain's neural networks during behavior when combined with neurophysiological techniques. The system's design has been thoughtfully considered, encompassing numerous intricate details. These details include the incorporation of flexible options for selecting start, goal, and proximal landmark positions, the inclusion of a rotating platform to prevent the accumulation of olfactory cues, and careful attention given to atomization, taking into account specific considerations such as the rotation of the maze without causing wire shortage or breakage. When combined with neurophysiological manipulations or recordings, the system provides a powerful tool for studying spatial navigation system.

    The behavioral experiment protocols, along with the analysis of animal behavior, are conducted with care, and the development of behavioral modeling to capture the animal's search strategy is thoughtfully executed. It is intriguing to observe how the integration of these innovative stochastic models can elucidate the evolution of mice's search strategy within a variant of the Barnes maze.

  9. Version published to 10.1101/2023.04.14.536859v4 on bioRxiv
  10. Version published to 10.7554/elife.88648.2 on eLife
  11. eLife

    Author response:

    The following is the authors’ response to the original reviews.

    We are very grateful to the reviewers for their constructive comments. Here is a summary of the main changes we made from the previous manuscript version, based on the reviewers’ comments:

    (1) Introduction of a new model, based on a Markov chain, capturing within-trial evolution in search strategy .

    (2) Addition of a new figure investigating inter-animal variations in search strategy.

    (3) Measurement of model fit consistency across 10 simulation repetitions, to prevent the risk of model overfitting.

    (4) Several clarifications have been made in the main text (Results, Discussion, Methods) and figure legends.

    (5) We now provide processed data and codes for analyses and models at GitHub repository

    (6) Simplification of the previous modeling. We realized that the two first models in the previous manuscript version were simply special cases of the third model. Therefore, we retained only the third model, which has been renamed as the ‘mixture model’.

    (7) Modification of Figure 4-6 and Supplementary Figure 7-8 (or their creation) to reflect the aforementioned changes

    Public Reviews:

    Reviewer #1 (Public Review):

    The authors design an automated 24-well Barnes maze with 2 orienting cues inside the maze, then model what strategies the mice use to reach the goal location across multiple days of learning. They consider a set of models and conclude that one of these models, a combined strategy model, best explains the experimental data.

    This study is written concisely and the results presented concisely. The best fit model is reasonably simple and fits the experimental data well (at least the summary measures of the data that were presented).

    Major points:

    (1) One combined strategy (once the goal location is learned) that might seem to be reasonable would be that the animal knows roughly where the goal is, but not exactly where, so it first uses a spatial strategy just to get to the first vestibule, then switches to a serial strategy until it reaches the correct vestibule. How well would such a strategy explain the data for the later sessions? The best combined model presented in the manuscript is one in which the animal starts with a roughly 50-50 chance of a serial (or spatial strategy) from the start vestibule (i.e. by the last session before the reversal the serial and spatial strategies are at ~50-50m in Fig. 5d). Is it the case that even after 15 days of training the animal starts with a serial strategy from its starting point approximately half of the time? The broader point is whether additional examination of the choices made by the animal, combined with consideration of a larger range of possible models, would be able to provide additional insight into the learning and strategies the animal uses.

    Our analysis focused on the evolution of navigation strategies across days and trials. The reviewer raises the interesting possibility that navigation strategy might evolve in a specific manner within each trial, especially on the later days once the environment is learned. To address this possibility, we first examined how some of the statistical distributions, previously analyzed across days, evolved within trials. Consistent with the reviewer’s intuition, the statistical distributions changed within trials, suggesting a specific strategy evolution within trials. Second, we developed a new model, where strategies are represented as nodes of a Markov chain. This model allows potential strategy changes after each vestibule visit, according to a specific set of transition probabilities. Vestibules are chosen based on the same stochastic processes as in the previous model. This new model could be fitted to the experimental distributions and captured both the within-trial evolution and the global distributions. Interestingly, the trials were mostly initiated in the random strategy (~67% chance) and to a lesser extent in the spatial strategy (~25% chance), but rarely in the serial strategy (~8% chance). This new model is presented in Figure 6.

    (2) To clarify, in the Fig. 4 simulations, is the "last" vestibule visit of each trial, which is by definition 0, not counted in the plots of Fig. 4b? Otherwise, I would expect that vestibule 0 is overrepresented because a trial always ends with Vi = 0.

    The last vestibule visit (vestibule 0 by definition) is counted in the plots of Fig.4b. We initially shared the same concern as the reviewer. However, upon further consideration, we arrived at the following explanation: A factor that might lead to an overrepresentation of vestibule 0 is the fact that, unlike other vestibules, it has to be contained in each trial, as trials terminated upon the selection of vestibule 0. Conversely, a factor that might contribute to an underrepresentation of vestibule 0 is that, unlike other vestibules, it cannot be counted more than once per trial. Somehow these two factors seem to counterbalance each other, resulting in no discernible overrepresentation or underrepresentation of vestibule 0 in the random process.

    Reviewer #2 (Public Review):

    This paper uses a novel maze design to explore mouse navigation behaviour in an automated analogue of the Barnes maze. Overall I find the work to be solid, with the cleverly designed maze/protocol to be its major strength - however there are some issues that I believe should be addressed and clarified.

    (1) Whilst I'm generally a fan of the experimental protocol, the design means that internal odor cues on the maze change from trial to trial, along with cues external to the maze such as the sounds and visual features of the recording room, ultimately making it hard for the mice to use a completely allocentric spatial 'place' strategy to navigate. I do not think there is a way to control for these conflicts between reference frames in the statistical modelling, but I do think these issues should be addressed in the discussion.

    It should be pointed out that all cues on the maze (visual, tactile, odorant) remained unchanged across trials, since the maze was rotated together with goal and guiding cues. Furthermore, the maze was equipped with an opaque cover to prevent mice from seeing the surrounding room (the imaging of mouse trajectories was achieved using infrared light and camera). It is however possible that some other cues such as room sounds and odors could be perceived and somewhat interfered with the sensory cues provided inside the maze. We have now mentioned this possibility in the discussion.

    (2) Somewhat related - I could not find how the internal maze cues are moved for each trial to demarcate the new goal (i.e. the luminous cues) ? This should be clarified in the methods.

    The luminous cues were fixed to the floor of the arena. Consequently, they rotated along with the arena as a unified unit, depicted in figure 1. We have added some clarifications in Figure 1 legend and methods.

    (3) It appears some data is being withheld from Figures 2&3? E.g. Days 3/4 from Fig 2b-f and Days 1-5 on for Fig 3. Similarly, Trials 2-7 are excluded from Fig 3. If this is the case, why? It should be clarified in the main text and Figure captions, preferably with equivalent plots presenting all the data in the supplement.

    The statistical distributions for all single days/trials are shown in the color-coded panels of Figure2&3. In the line plots of Figure2&3, we show only the overlay of 2-3 lines for the sake of clarity. The days/trials represented were chosen to capture the dynamic range of variability within the distributions. We have added this information in the figure legends.

    (4) I strongly believe the data and code should be made freely available rather than "upon reasonable request".

    Matrices of processed data and various codes for simulations and analyses are now available at https://github.com/ sebiroyerlab/Vestibule_sequences.

    Reviewer #3 (Public Review):

    Royer et al. present a fully automated variant of the Barnes maze to reduce experimenter interference and ensure consistency across trials and subjects. They train mice in this maze over several days and analyze the progression of mouse search strategies during the course of the training. By fitting models involving stochastic processes, they demonstrate that a model combined of the random, spatial, and serial processes can best account for the observed changes in mice's search patterns. Their findings suggest that across training days the spatial strategy (using local landmarks) was progressively employed, mostly at the expense of the random strategy, while the serial strategy (consecutive nearby vestibule check) is reinforced from the early stages of training. Finally, they discuss potential mechanistic underpinnings within brain systems that could explain such behavioral adaptation and flexibility.

    Strength:

    The development of an automated Barnes maze allows for more naturalistic and uninterrupted behavior, facilitating the study of spatial learning and memory, as well as the analysis of the brain's neural networks during behavior when combined with neurophysiological techniques. The system's design has been thoughtfully considered, encompassing numerous intricate details. These details include the incorporation of flexible options for selecting start, goal, and proximal landmark positions, the inclusion of a rotating platform to prevent the accumulation of olfactory cues, and careful attention given to atomization, taking into account specific considerations such as the rotation of the maze without causing wire shortage or breakage. When combined with neurophysiological manipulations or recordings, the system provides a powerful tool for studying spatial navigation system.

    The behavioral experiment protocols, along with the analysis of animal behavior, are conducted with care, and the development of behavioral modeling to capture the animal's search strategy is thoughtfully executed. It is intriguing to observe how the integration of these innovative stochastic models can elucidate the evolution of mice's search strategy within a variant of the Barnes maze.

    Weakness:

    (1) The development of the well-thought-out automated Barnes maze may attract the interest of researchers exploring spatial learning and memory. However, this aspect of the paper lacks significance due to insufficient coverage of the materials and methods required for readers to replicate the behavioral methodology for their own research inquiries.

    Moreover, as discussed by the authors, the methodology favors specialists who utilize wired recordings or manipulations (e.g. optogenetics) in awake, behaving rodents. However, it remains unclear how the current maze design, which involves trapping mice in start and goal positions and incorporating angled vestibules resulting in the addition of numerous corners, can be effectively adapted for animals with wired implants.

    The reviewer is correct in pointing out that the current maze design is not suitable for performing experiments with wired implant, particularly due to the maze’s enclosed structure and the access to the start/goal boxes through side holes. Instead, pharmacogenetics and wireless approaches for optogenetic and electrophysiology would need to be used. We have now mentioned this limitation in the discussion.

    (2) Novelty: In its current format, the main axis of the paper falls on the analysis of animal behavior and the development of behavioral modeling. In this respect, while it is interesting to see how thoughtfully designed models can explain the evolution of mice search strategy in a maze, the conclusions offer limited novel findings that align with the existing body of research and prior predictions.

    We agree with the reviewer that our study is weakly connected to previous researches on hippocampus and spatial navigation, as it consists mainly of animal behavior analysis and modeling and addresses a relatively unexplored topic. We hope that the combination of our behavioral approach with optogenetic and electrophysiology will allow in the future new insights that are in line with the existing body of research.

    (3) Scalability and accessibility: While the approach may be intriguing to experts who have an interest in or are familiar with the Barnes maze, its presentation seems to primarily target this specific audience. Therefore, there is a lack of clarity and discussion regarding the scalability of behavioral modeling to experiments involving other search strategies (such as sequence or episodic learning), other animal models, or the potential for translational applications. The scalability of the method would greatly benefit a broader scientific community. In line with this view, the paper's conclusions heavily rely on the development of new models using custom-made codes. Therefore, it would be advantageous to make these codes readily available, and if possible, provide access to the processed data as well. This could enhance comprehension and enable a larger audience to benefit from the methodology.

    The current approach might indeed extend to other species in equivalent environments and might also constitute a general proof of principle regarding the characterization of animal behaviors by the mixing of stochastic processes. We have now mentioned these points in the discussion.

    As suggest by the reviewer, we have now provided model/simulation codes and processed data to replicate the figures, at https://github.com/sebiroyerlab/Vestibule_sequences

    (4) Cross-validation of models: The authors have not implemented any measures to mitigate the risk of overfitting in their modeling. It would have been beneficial to include at least some form of cross-validation with stochastic models to address this concern. Additionally, the paper lacks the presence of analytics or measures that assess and compare the performance of the models.

    To avoid the risk of model overfitting, the most appropriate solution appeared to be repeating the simulations several times and examining the consistency of the obtained parameters across repetitions. For the mixture model, we now show in Supplementary figure 7 the probabilities obtained from 10 repetitions of the simulation. Similarly, for the Markov chain model, the probabilities obtained from 10 repetitions of the simulation are shown in Figure 6.

    Regarding model comparison, we have simplified our mixture model into only one model, as we realized the 2 other models in the previous manuscript version were simply special cases of the 3rd model. Nevertheless, comparison was still needed for the estimation for the best value of N (the number of consecutive segments that a strategy lasts) in the mixture model. We now show the comparison of mean square errors obtained for different values of N, using t-test across 10 repetitions of the simulations (Figure 5c).

    (5) Quantification of inter-animal variations in strategy development: It is important to investigate, and address the argument concerning the possibility that not all animals recruit and develop the three processes (random, spatial, and serial) in a similar manner over days of training. It would be valuable to quantify the transition in strategy across days for each individual mouse and analyze how the population average, reflecting data from individual mice, corresponds to these findings. Currently, there is a lack of such quantification and analysis in the paper.

    We have added a figure (Supplementary figure 8) showing the mixture model matching analyses for individual animals. A lot of variability is indeed observed across animals, with some animals displaying strong preferences for certain strategies compare to others. The average across mouse population showed a similar trend as the result obtained with the pooled data.

    Recommendations for the authors:

    Summary of Reviewer Comments:

    (1) In its present form, the manuscript lacks sufficient coverage of the materials and methods necessary for readers to replicate the behavioral methodology in their own research inquiries. For instance, it would be beneficial to clarify how the cues are rotated relative to the goal.

    (2) The models may be over-fitted, leading to spurious conclusions, and cross-validation is necessary to rule out this possibility.

    (3) The specific choice of the three strategies used to fit behavior in this model should be better justified, as other strategies may account for the observed behavior.

    (4) The study would benefit from an analysis of behavior on an animal-by-animal basis, potentially revealing individual differences in strategies.

    (5) Spatial behavior is not necessarily fully allocentric in this task, as only the two cues in the arena can be used for spatial orientation, unlike odor cues on the floor and sound cues in the room. This should be discussed.

    (6) Making the data and code fully open source would greatly strengthen the impact of this study.

    In addition, each reviewer has raised both major and minor concerns which should be addressed if possible.

    Reviewer #1 (Recommendations For The Authors):

    Minor points:

    (1) Change "tainted" to "tinted" in Fig. 1a

    (2) Should note explicitly in Fig. 2d that the goal is at vestibule 0, and also in the legend

    (3) Fig. 3 legend should say "c-e)", not "c-f)"

    (4) Supplementary Fig. 8 legend repeats "d)" twice

    Reviewer #2 (Recommendations For The Authors):

    Packard & McGaugh 1996 is cited twice as refs 5 and 14

    Reviewer #3 (Recommendations For The Authors):

    - Figure 3: Please correct the labels referenced as "c-f)" in the figure's legend.

    - Rounding numbers issue on page 4: 82.62% + 17.37% equals 99.99%, not 100%.

    We fixed all minor points. We are very thankful to the reviewers for their constructive comments.

  12. eLife

    eLife assessment

    This study presents a valuable new behavioral apparatus aimed at differentiating the strategies animals use to orient themselves in an environment. The evidence supporting the claims is solid, with statistical modeling of animal behavior. Overall, this study will attract the interest of researchers exploring spatial learning and memory.

  13. eLife

    Reviewer #1 (Public Review):

    The authors design an automated 24-well Barnes maze with 2 orienting cues inside the maze, then model what strategies the mice use to reach the goal location across multiple days of learning. They consider a set of models and conclude that the animals begin with a large proportion of random choices (choices irrespective of the goal location), which over days of experience becomes a combination of spatial choices (choices targeted around the goal location) and serial choices (successive stepwise choices in a given direction). Moreover, the authors show that after the animal has many days of experience in the maze, they still often began each trial with a random choice, followed by spatial or serial choices.

    This study is written concisely and the results are presented concisely. The best fit model provides valuable insight into how the animals solve this task, and therefore offers a quantitative foundation upon which tests of neural mechanisms of the components of the behavioral strategy can be performed. These tests will also benefit from the automated nature of the task.

  14. eLife

    Reviewer #2 (Public Review):

    This paper uses a novel maze design to explore mouse navigation behaviour in an automated analogue of the Barnes maze. A major strength is the novel and clever experimental design which rotates the floor and intramaze cues before the start of each new trial, allowing the previous goal location to become the next starting position. The modelling sampling a Markov chain of navigation strategies is elegant, appropriate and solid, appearing to capture the behavioural data well. This work provides a valuable contribution and I'm excited to see further developments, such as neural correlates of the different strategies and switches between them.

  15. eLife

    Reviewer #3 (Public Review):

    Strength:

    The development of an automated Barnes maze allows for more naturalistic and uninterrupted behavior, facilitating the study of spatial learning and memory, as well as the analysis of the brain's neural networks during behavior when combined with neurophysiological techniques. The system's design has been thoughtfully considered, encompassing numerous intricate details. These details include the incorporation of flexible options for selecting start, goal, and proximal landmark positions, the inclusion of a rotating platform to prevent the accumulation of olfactory cues, and careful attention given to atomization, taking into account specific considerations such as the rotation of the maze without causing wire shortage or breakage. When combined with neurophysiological manipulations or recordings, the system provides a powerful tool for studying spatial navigation system.
    The behavioral experiment protocols, along with the analysis of animal behavior, are conducted with care, and the development of behavioral modeling to capture the animal's search strategy is thoughtfully executed. It is intriguing to observe how the integration of these innovative stochastic models can elucidate the evolution of mice's search strategy within a variant of the Barnes maze.

    Comments on revised version:

    The authors have addressed all the points I outlined in the previous round of review, resulting in significant improvements to the manuscript. However, I have one remaining comment. Given the updated inter-animal analysis (Supplementary Figure 8), it appears that male and female mice develop strategies differently across days. Male mice seem to progressively increase their employment of spatial strategy across days, at the expense of the random strategy. Conversely, female mice exhibit both spatial and serial strategies at their highest levels on day 2, with minimal changes observed on the subsequent days.
    These findings could alter the interpretation of Figure 5 and the corresponding text in the section "Evolution of search strategy across days".
    For instance, this statement on page 6 doesn't hold for female mice: "The spatial strategy was increased across days, ... largely at the expense of the random strategy."

  16. Version published to 10.1101/2023.04.14.536859v3 on bioRxiv
  17. Version published to 10.7554/elife.88648.1 on eLife
  18. eLife

    eLife assessment

    This study presents a valuable new behavioral apparatus aimed at differentiating the strategies animals use to orient themselves in an environment. The evidence supporting the claims is solid, with statistical modeling of animal behavior. Overall, this study will attract the interest of researchers exploring spatial learning and memory.

  19. eLife

    Reviewer #1 (Public Review):

    The authors design an automated 24-well Barnes maze with 2 orienting cues inside the maze, then model what strategies the mice use to reach the goal location across multiple days of learning. They consider a set of models and conclude that one of these models, a combined strategy model, best explains the experimental data.

    This study is written concisely and the results presented concisely. The best fit model is reasonably simple and fits the experimental data well (at least the summary measures of the data that were presented).

    Major points:

    1. One combined strategy (once the goal location is learned) that might seem to be reasonable would be that the animal knows roughly where the goal is, but not exactly where, so it first uses a spatial strategy just to get to the first vestibule, then switches to a serial strategy until it reaches the correct vestibule. How well would such a strategy explain the data for the later sessions? The best combined model presented in the manuscript is one in which the animal starts with a roughly 50-50 chance of a serial (or spatial strategy) from the start vestibule (i.e. by the last session before the reversal the serial and spatial strategies are at ~50-50m in Fig. 5d). Is it the case that even after 15 days of training the animal starts with a serial strategy from its starting point approximately half of the time? The broader point is whether additional examination of the choices made by the animal, combined with consideration of a larger range of possible models, would be able to provide additional insight into the learning and strategies the animal uses.

    2. To clarify, in the Fig. 4 simulations, is the "last" vestibule visit of each trial, which is by definition 0, not counted in the plots of Fig. 4b? Otherwise, I would expect that vestibule 0 is overrepresented because a trial always ends with Vi = 0.

  20. eLife

    Reviewer #2 (Public Review):

    This paper uses a novel maze design to explore mouse navigation behaviour in an automated analogue of the Barnes maze. Overall I find the work to be solid, with the cleverly designed maze/protocol to be its major strength - however there are some issues that I believe should be addressed and clarified.

    1. Whilst I'm generally a fan of the experimental protocol, the design means that internal odor cues on the maze change from trial to trial, along with cues external to the maze such as the sounds and visual features of the recording room, ultimately making it hard for the mice to use a completely allocentric spatial 'place' strategy to navigate. I do not think there is a way to control for these conflicts between reference frames in the statistical modelling, but I do think these issues should be addressed in the discussion.

    2. Somewhat related - I could not find how the internal maze cues are moved for each trial to demarcate the new goal (i.e. the luminous cues) ? This should be clarified in the methods.

    3. It appears some data is being withheld from Figures 2&3? E.g. Days 3/4 from Fig 2b-f and Days 1-5 on for Fig 3. Similarly, Trials 2-7 are excluded from Fig 3. If this is the case, why? It should be clarified in the main text and Figure captions, preferably with equivalent plots presenting all the data in the supplement.

    4. I strongly believe the data and code should be made freely available rather than "upon reasonable request".

  21. eLife

    Reviewer #3 (Public Review):

    Royer et al. present a fully automated variant of the Barnes maze to reduce experimenter interference and ensure consistency across trials and subjects. They train mice in this maze over several days and analyze the progression of mouse search strategies during the course of the training. By fitting models involving stochastic processes, they demonstrate that a model combined of the random, spatial, and serial processes can best account for the observed changes in mice's search patterns. Their findings suggest that across training days the spatial strategy (using local landmarks) was progressively employed, mostly at the expense of the random strategy, while the serial strategy (consecutive nearby vestibule check) is reinforced from the early stages of training. Finally, they discuss potential mechanistic underpinnings within brain systems that could explain such behavioral adaptation and flexibility.

    Strength:
    The development of an automated Barnes maze allows for more naturalistic and uninterrupted behavior, facilitating the study of spatial learning and memory, as well as the analysis of the brain's neural networks during behavior when combined with neurophysiological techniques. The system's design has been thoughtfully considered, encompassing numerous intricate details. These details include the incorporation of flexible options for selecting start, goal, and proximal landmark positions, the inclusion of a rotating platform to prevent the accumulation of olfactory cues, and careful attention given to atomization, taking into account specific considerations such as the rotation of the maze without causing wire shortage or breakage. When combined with neurophysiological manipulations or recordings, the system provides a powerful tool for studying spatial navigation system.
    The behavioral experiment protocols, along with the analysis of animal behavior, are conducted with care, and the development of behavioral modeling to capture the animal's search strategy is thoughtfully executed. It is intriguing to observe how the integration of these innovative stochastic models can elucidate the evolution of mice's search strategy within a variant of the Barnes maze.

    Weakness:
    1. The development of the well-thought-out automated Barnes maze may attract the interest of researchers exploring spatial learning and memory. However, this aspect of the paper lacks significance due to insufficient coverage of the materials and methods required for readers to replicate the behavioral methodology for their own research inquiries.
    Moreover, as discussed by the authors, the methodology favors specialists who utilize wired recordings or manipulations (e.g. optogenetics) in awake, behaving rodents. However, it remains unclear how the current maze design, which involves trapping mice in start and goal positions and incorporating angled vestibules resulting in the addition of numerous corners, can be effectively adapted for animals with wired implants.

    2. Novelty: In its current format, the main axis of the paper falls on the analysis of animal behavior and the development of behavioral modeling. In this respect, while it is interesting to see how thoughtfully designed models can explain the evolution of mice search strategy in a maze, the conclusions offer limited novel findings that align with the existing body of research and prior predictions.

    3. Scalability and accessibility: While the approach may be intriguing to experts who have an interest in or are familiar with the Barnes maze, its presentation seems to primarily target this specific audience. Therefore, there is a lack of clarity and discussion regarding the scalability of behavioral modeling to experiments involving other search strategies (such as sequence or episodic learning), other animal models, or the potential for translational applications. The scalability of the method would greatly benefit a broader scientific community. In line with this view, the paper's conclusions heavily rely on the development of new models using custom-made codes. Therefore, it would be advantageous to make these codes readily available, and if possible, provide access to the processed data as well. This could enhance comprehension and enable a larger audience to benefit from the methodology.

    4. Cross-validation of models: The authors have not implemented any measures to mitigate the risk of overfitting in their modeling. It would have been beneficial to include at least some form of cross-validation with stochastic models to address this concern. Additionally, the paper lacks the presence of analytics or measures that assess and compare the performance of the models.

    5. Quantification of inter-animal variations in strategy development: It is important to investigate, and address the argument concerning the possibility that not all animals recruit and develop the three processes (random, spatial, and serial) in a similar manner over days of training. It would be valuable to quantify the transition in strategy across days for each individual mouse and analyze how the population average, reflecting data from individual mice, corresponds to these findings. Currently, there is a lack of such quantification and analysis in the paper.

  22. Version published to 10.1101/2023.04.14.536859v2 on bioRxiv
  23. Version published to 10.1101/2023.04.14.536859v1 on bioRxiv