Fully autonomous mouse behavioral and optogenetic experiments in home-cage

  1. Yaoyao Hao
  2. Alyse Marian Thomas
  3. Nuo Li

  • Curated by eLife

    eLife logo

    Evaluation Summary:

    This manuscript describes extensively a fully automated procedure to train mice to perform voluntary head-fixation, and a whisker-based tactile discrimantion task. In addition the authors demonstrate that with this procedure, light illumination of red-shifted opsins expressed in inhibitory neurons can be used to selectively silence targeted brain regions during the task in a non-invasive manner. Together, although volontary head-fixation training and automated behavior has been readily implemented in different contexts, this study elegantly delineates important steps to boost the acceptancy and duration of head-fixations and thereby train more complex tasks. The demonstration of transcranial optogenetics in this context also opens the possibility to perform precise brain inactivations during well-controlled sensory stimulations, in self-initiated behavior.

    (This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. Reviewer #1 and Reviewer #3 agreed to share their names with the authors.)

Reviewed by

Read the full article See related articles

Listed in

Log in to save this article

Abstract

Goal-directed behaviors involve distributed brain networks. The small size of the mouse brain makes it amenable to manipulations of neural activity dispersed across brain areas, but existing optogenetic methods serially test a few brain regions at a time, which slows comprehensive mapping of distributed networks. Laborious operant conditioning training required for most experimental paradigms exacerbates this bottleneck. We present an autonomous workflow to survey the involvement of brain regions at scale during operant behaviors in mice. Naive mice living in a home-cage system learned voluntary head-fixation (>1 hr/day) and performed difficult decision-making tasks, including contingency reversals, for 2 months without human supervision. We incorporated an optogenetic approach to manipulate activity in deep brain regions through intact skull during home-cage behavior. To demonstrate the utility of this approach, we tested dozens of mice in parallel unsupervised optogenetic experiments, revealing multiple regions in cortex, striatum, and superior colliculus involved in tactile decision-making.

Article activity feed

  1. Version published to 10.7554/elife.66112 on eLife
  2. eLife

    Author Response:

    Reviewer #1:

    The authors demonstrate in this study that it is possible to train mice to perform a challenging tactile discrimination task, in a highly controlled manner, in a fully automated setup in which the animals learn to head-fix voluntarily. A number of well described tricks are used to prolong the self-fixation time and thereby obtain enough training time to reach good performance when the decision perceptual decision is difficult. In addition the study establish that this experimental design allows targeted silencing of relatively deep brain areas through a clear skull preparation.

    It has already been demonstrated that mice can perform voluntary head-fixation and can do behavioral tasks in this context. However, this is the first time this methodology is applied to first to a tactile task and second to a task that mice learn is thousands of trials. Another advantage of the present technique is that it is fully automated and allows training without virtually any human intervention.

    The demonstration that optogenetic silencing can be performed in this context is nice but not very surprising as already done in other contexts. Nevertheless it is an interesting application of self head-fixation. The authors should make sure that a maximum of information is available relative to the efficiency of the silencing (fraction of cells silenced) and about its impact on the behavior (does it result or not in a complete impairment?).

    We have improved presentation in various places of the paper to provide more information about the optogenetic manipulation. We added new analysis of the fraction of neurons affected by photostimulation (Figure 8E). We also analyzed the impact on behavioral performance relative to chance performance (Figure S4A and S6). We compared the effect size to prior studies (Figure S4) and we discuss the interpretation of effect size (Discussion, page 22).

    In the power range tested in this study, photostimulation did not reduce performance to chance level (Figure S6). One limitation of the optogenetic workflow is the interpretation of behavioral deficit effect size. We examined this issue in ALM, a brain region from which we have the most extensive data. In previous studies, we have shown that bilateral photoinhibition of ALM results in chance level performance (Li et al 2016, Fig 2b; Gao et al, 2018, Extended Data Fig 6b). Here, mice performance was above chance during photoinhibition of ALM (Figure S4). This difference in effect size likely resulted from incomplete silencing of ALM. The photostimulus intensity used here was much less than those used in previous studies (0.3 vs. 11.9 mW/mm2). In addition, a single virus injection was not sufficient to cover the entire ALM. Thus a partial behavioral effect could be due to incomplete silencing of a brain region, or partial involvement of the brain region in the task. Given this limitation, we caution that the function of a brain region could only be fully deduced in more detailed analysis and together with neurophysiology. The workflow presented here can be used as a discovery platform to quickly identify regions of interest for more detailed neurophysiology analysis. We now better highlight these points in the Discussion.

    Reviewer #2:

    Hao and colleagues developed an automatic system for high-throughput behavioral and optogenetic experiments for mice in home cage settings. The system includes a voluntary head-fixation apparatus and integrated fiber-free optogenetic capabilities. The authors describe in detail the design of the system and the stages for successful automatic training. They perform proof-of-concept experiments to validate their system. The experiments are technically solid and I am convinced that their system will be of interest to some laboratories that perform similar experiments. Despite the large variety of similar automated systems out there, this one may prove to become a popular design.

    The weak side of the work is that it is not particularly novel scientifically. The system is complex but there it is not an innovative technology. The body of the study has too many technical details as if it is a Methodological section of a regular manuscript. There are bits of interesting information scattered around the paper (like the insights about the strategy mice use, which stem from the regression analysis), but these are not developed into any coherent direction that answers outstanding questions. The potential advantages of this system compared to other systems is marginal. In my eyes, the fact that manual training is so similar to the automatic one is not only a positive point. Rather, it signifies that the differences are mainly quantitative (e.g. # of mice a lab can train per day, etc). Thus, even as a methods paper, the lack of qualitative difference between this and other methods weakens it as a potential substrate for novel findings.

    The automated workflow presented here significantly boosts the yield and duration of training to rival and slightly surpass that of manual training for the first time (new Supplemental Table 1). We think this degree of automation is an important technical advance. We show that the workflow can significantly scale up the throughput of optogenetic experiments probing behaviors that require thousands of trials to learn. This enables efficient and systematic mapping of large subcortical structures that are previously difficult to achieve. We better highlight comparisons to previous methods in several key areas in the Supplemental Table 1. We have also strengthened the Discussion (page 20).

    We highlight one line of inquiry enabled by our workflow, a systematic mapping of the cortico-basal- ganglia loops during perceptual decision-making. The striatum is topographically organized. Previous studies examined different subregions of the striatum in different perceptual decision behaviors, making comparisons across studies difficult. The striatum in the mouse brain is ~21.5 mm3 in size (Allen reference brain, (Wang, et al, Cell 2020)). Optogenetic experiments using optical fibers manipulate activity near the fiber tip (approximately 1 mm3). A systematic survey of different striatal domains’ involvement in specific behaviors is currently difficult. In our workflow, individual striatal subregions (~1 mm3, Figure 8) could be rapidly screened through parallel testing. At moderate throughput (15 mice / 2 months), a screen that tiles the entire striatum could be completed in under 12 months with little human effort. To illustrate its feasibility, we tested 3 subregions in the striatum previously implicated in different types of perceptual decision behaviors (Yartsev et al, eLife 2018; Sippy et al, Neuron 2015; Znamenskiy & Zador, Nature 2013), including an additional region in the posterior striatum that do not receive ALM and S1 inputs. The results revealed a hotspot in the dorsolateral striatum that biased tactile-guided decision-making (Figure 8). Our approach thus opens the door to rapid screening of the striatal domains during complex operant behaviors.

    Moreover, by eliminating human intervention, automated training allows quantitative assaying of task learning (Figure 4). Home-cage testing also exposes behavioral signatures of motivation in self-initiated behavior (Figure 6). These observations suggest additional opportunities for inquires of goal-directed behaviors in the context of home-cage testing.

    Reviewer #3:

    In this study, Hao et al. developed an automatized operant box to perform decision-making tasks and optogenetic perturbations without requiring the experimenter's manipulation. For this aim, mice learn to head-fix and to perform a task by themselves. The optogenetic experiment using red-shifted opsins allows manipulation of circuits without the need of an implanted optical fiber. The automation of behavioral tasks in home cages (isolated rodents or in groups) is an intense area of research in neuroscience. The possibility of coupling home cage behavioral analysis with optogenetic manipulation and with complex tasks that require precise positioning of the animal for controlled stimulations (vibrating stimulation, visual …..) is thus of great interest and I commend the authors for their comprehensive dissection of the automated behavioral training setup. Some clarification, reporting of additional behavioral measures and refinement of analyses could improve the impact of this work.

    1. The first part of the paper nicely describes the experimental procedure to automate such a complex task. The procedure is very well described, the important points (e.g. the possibility for the animal to disengage…) are properly highlighted, and the online site allows to download the plans and 3D descriptions of the tools and the procedures. The authors compare task learning in automated versus manual training and show that there are overall very few differences. Whisker trimming reduces performance, indicating that animal used information to make the choice. This part of the work is already impressive. Apart from that, the authors do not consider in their description what could be an essential aspect of experiments in a home-cage, i.e the control of the motivation to perform the task. Mice perform the task (here, engage in the head fixation to obtained reward) when they wish and thus, compared with the manual training, there is no explicit control of the animal motivation. This could have consequence on i) the inter-fixation intervals that become an element of the decision and ii) questioned whether the commitment to the task is always motivated by drinking, or whether there is also a commitment to explore, or to check… This could impact the success in the task (e.g. if the animal is not motivated by water, it can explore…). Adding data analyses (information about the daily water consumption, are the inter-fixation intervals correlated with the success or failure in the last trial …) and even short discussion or introduction of these aspects (see for example Timberlake et al, JEAB 1987 or Rowland et al 2008, Physiol behavior for distinction between close and open economies paradigm) could strengthened the behavioral description.

    We thank the reviewer for these suggestions. We performed additional analyses to examine these issues which led us to include a new section of Results in the revised manuscript (page 13-14 and Figure 6).

    We have added a new Figure 6 showing water consumption and body weight information in home-cage testing. At steady state, a mouse typically consumed ~1mL of water daily (~400 rewarded trials) while maintaining stable body weight. This amount of water consumption was similar to mice engaged in daily manual experiments (Guo et al, Plos ONE 2014). The number of head-fixations per day was correlated with body weight (Figure 6). Since body weight reflects prior water consumption, this indicates different levels of motivation due to thirst, which drives engagement in the task.

    We also examined the inter-fixation-interval. Interestingly, the inter-fixation-interval after an error (which led to no reward) was significantly longer than following a correct trial (Figure 6E). This is inconsistent with error from exploration. Rather it likely reflects a loss of motivation after an error, perhaps due to the loss of an expected reward. We suspect that error trials violated the mice’s expectation of reward, and therefore discouraged the mice, leading to a loss in motivation. Consistent with this interpretation, we also found a significant increase in inter-fixation-intervals shortly after a sensorimotor contingency reversal (Figure 6F), coinciding with an increase in error rate due to the rule change.

    Despite these changes in motivation to engage in the task, the choice behavior in the task was similar. In highly trained mice, task performance was stable despite the body weight change (Figure 6D). Logistic regression analysis of the choice behavior shows that mice maintained the same strategy in their choice behavior (Figure 6G).

    1. In the second part of the work, the authors focus on the description of choice behavior. To characterize it, the authors used a logistic model to predict choices. They suggest that at the beginning of the task the animals biased their current choice by their last choice (parameter A1) and that once the task is learned they alternate according to the current stimulation (parameters S0). The model was a logistic function of the weighted sum of several behavioral and task variables and has 19 parameters (the ß parameters). If the animal only used these two informations, can a model that only takes into account A1 and S0 reproduce the data? If not, this certainly indicates that other informations (even distributed) are necessary; and also indicates individual strategies. Finally, analyses are made by considering trials as a discrete chain (trial n, n+1…). However, the self-head-fixed methodology causes the trials to be organized with more or less time between successive trials depending on motivation (see above). Again, do the authors note differences in performance according to the timing between trials? Could it be a variable in the model?

    We thank the reviewer for these great suggestions. We tested a model that included only choice history A1, tactile stimulus S0, and a constant bias term (β0). This 3-parameter model performed as well as the full model in predicting choice. This indicates that other factors do not contribute significantly to the choice behavior. We have included this result in the revised Figure 4C.

    We next examined whether inter-fixation-interval (i.e. the time elapsed between head-fixations and presumably the motivation to engage in the task) could impact mice’s choice behavior. There are multiple ways inter-fixation-interval could be incorporated into the logistic regression model. For example, it could be modeled as an explicit variable that biases left/right choice, or modulations on existing regressors (i.e. a gain variable that modulates the contribution of specific regressors). Each approach requires assumptions about how motivation affects the behavioral strategy of the mice. Instead, as a first order analysis, we examined whether the logistic regression model could predict choice equally well in trials following short vs. long inter-fixation-intervals. Our logic is that if mice adapted different strategies in different motivational states (reflected in short vs. long inter-fixation- intervals), the predictive power of the model would differ between these conditions. We fit the logistic regression model using trials in their natural sequential order (regardless of the inter-fixation-intervals). The model was then used to predict choice on independent trials. Trials were then sorted by the preceding inter-fixation-intervals. Prediction performance was calculated separately for trials following short vs. long inter-fixation-intervals. We did not find a significant difference in the model prediction performance. The result was similar in early and late stages of task learning (Figure 6G), even though mice used distinct strategies during these periods (Figure 4). These results suggest consistent strategies in the choice behavior. We have included this analysis in the new Figure 6.

    1. The third part described optogenetic manipulations. It is clear that group sizes are small. Nevertheless, if the objective was to show that the method works, the results are convincing. Some experimental details and in particular the choice of the statistical procedure need clarification.

    We have improved the presentation and clarified experimental details of the task, hypotheses for targeting specific brain regions, and statistical procedures.

  3. eLife

    Reviewer #3 (Public Review):

    In this study, Hao et al. developed an automatized operant box to perform decision-making tasks and optogenetic perturbations without requiring the experimenter's manipulation. For this aim, mice learn to head-fix and to perform a task by themselves. The optogenetic experiment using red-shifted opsins allows manipulation of circuits without the need of an implanted optical fiber. The automation of behavioral tasks in home cages (isolated rodents or in groups) is an intense area of research in neuroscience. The possibility of coupling home cage behavioral analysis with optogenetic manipulation and with complex tasks that require precise positioning of the animal for controlled stimulations (vibrating stimulation, visual .....) is thus of great interest and I commend the authors for their comprehensive dissection of the automated behavioral training setup. Some clarification, reporting of additional behavioral measures and refinement of analyses could improve the impact of this work.

    1. The first part of the paper nicely describes the experimental procedure to automate such a complex task. The procedure is very well described, the important points (e.g. the possibility for the animal to disengage...) are properly highlighted, and the online site allows to download the plans and 3D descriptions of the tools and the procedures. The authors compare task learning in automated versus manual training and show that there are overall very few differences. Whisker trimming reduces performance, indicating that animal used information to make the choice. This part of the work is already impressive. Apart from that, the authors do not consider in their description what could be an essential aspect of experiments in a home-cage, i.e the control of the motivation to perform the task. Mice perform the task (here, engage in the head fixation to obtained reward) when they wish and thus, compared with the manual training, there is no explicit control of the animal motivation. This could have consequence on i) the inter-fixation intervals that become an element of the decision and ii) questioned whether the commitment to the task is always motivated by drinking, or whether there is also a commitment to explore, or to check... This could impact the success in the task (e.g. if the animal is not motivated by water, it can explore...). Adding data analyses (information about the daily water consumption, are the inter-fixation intervals correlated with the success or failure in the last trial ...) and even short discussion or introduction of these aspects (see for example Timberlake et al, JEAB 1987 or Rowland et al 2008, Physiol behavior for distinction between close and open economies paradigm) could strengthened the behavioral description.

    2. In the second part of the work, the authors focus on the description of choice behavior. To characterize it, the authors used a logistic model to predict choices. They suggest that at the beginning of the task the animals biased their current choice by their last choice (parameter A1) and that once the task is learned they alternate according to the current stimulation (parameters S0). The model was a logistic function of the weighted sum of several behavioral and task variables and has 19 parameters (the ß parameters). If the animal only used these two informations, can a model that only takes into account A1 and S0 reproduce the data? If not, this certainly indicates that other informations (even distributed) are necessary; and also indicates individual strategies. Finally, analyses are made by considering trials as a discrete chain (trial n, n+1...). However, the self-head-fixed methodology causes the trials to be organized with more or less time between successive trials depending on motivation (see above). Again, do the authors note differences in performance according to the timing between trials? Could it be a variable in the model?

    3. The third part described optogenetic manipulations. It is clear that group sizes are small. Nevertheless, if the objective was to show that the method works, the results are convincing. Some experimental details and in particular the choice of the statistical procedure need clarification.

  4. eLife

    Reviewer #2 (Public Review):

    Hao and colleagues developed an automatic system for high-throughput behavioral and optogenetic experiments for mice in home cage settings. The system includes a voluntary head-fixation apparatus and integrated fiber-free optogenetic capabilities. The authors describe in detail the design of the system and the stages for successful automatic training. They perform proof-of-concept experiments to validate their system. The experiments are technically solid and I am convinced that their system will be of interest to some laboratories that perform similar experiments. Despite the large variety of similar automated systems out there, this one may prove to become a popular design.

    The weak side of the work is that it is not particularly novel scientifically. The system is complex but there it is not an innovative technology. The body of the study has too many technical details as if it is a Methodological section of a regular manuscript. There are bits of interesting information scattered around the paper (like the insights about the strategy mice use, which stem from the regression analysis), but these are not developed into any coherent direction that answers outstanding questions. The potential advantages of this system compared to other systems is marginal. In my eyes, the fact that manual training is so similar to the automatic one is not only a positive point. Rather, it signifies that the differences are mainly quantitative (e.g. # of mice a lab can train per day, etc). Thus, even as a methods paper, the lack of qualitative difference between this and other methods weakens it as a potential substrate for novel findings.

  5. eLife

    Reviewer #1 (Public Review):

    The authors demonstrate in this study that it is possible to train mice to perform a challenging tactile discrimination task, in a highly controlled manner, in a fully automated setup in which the animals learn to head-fix voluntarily. A number of well described tricks are used to prolong the self-fixation time and thereby obtain enough training time to reach good performance when the decision perceptual decision is difficult. In addition the study establish that this experimental design allows targeted silencing of relatively deep brain areas through a clear skull preparation.

    It has already been demonstrated that mice can perform voluntary head-fixation and can do behavioral tasks in this context. However, this is the first time this methodology is applied to first to a tactile task and second to a task that mice learn is thousands of trials. Another advantage of the present technique is that it is fully automated and allows training without virtually any human intervention.

    The demonstration that optogenetic silencing can be performed in this context is nice but not very surprising as already done in other contexts. Nevertheless it is an interesting application of self head-fixation. The authors should make sure that a maximum of information is available relative to the efficiency of the silencing (fraction of cells silenced) and about its impact on the behavior (does it result or not in a complete impairment?).

  6. eLife

    Evaluation Summary:

    This manuscript describes extensively a fully automated procedure to train mice to perform voluntary head-fixation, and a whisker-based tactile discrimantion task. In addition the authors demonstrate that with this procedure, light illumination of red-shifted opsins expressed in inhibitory neurons can be used to selectively silence targeted brain regions during the task in a non-invasive manner. Together, although volontary head-fixation training and automated behavior has been readily implemented in different contexts, this study elegantly delineates important steps to boost the acceptancy and duration of head-fixations and thereby train more complex tasks. The demonstration of transcranial optogenetics in this context also opens the possibility to perform precise brain inactivations during well-controlled sensory stimulations, in self-initiated behavior.

    (This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. Reviewer #1 and Reviewer #3 agreed to share their names with the authors.)

  7. Version published to 10.1101/2020.12.27.424480v2 on bioRxiv
  8. Version published to 10.1101/2020.12.27.424480v1 on bioRxiv