Neural spiking signatures predict behavioral phenotypes of cerebellar movement disorders

Reviewer #1 (Public Review):

Summary:
Van der Heijden et al perform an ambitious analysis of single-unit activity in the interposed nuclei of multiple mouse models of cerebellar dysfunction. Based on these recordings, they develop a classifier to predict the behavioral phenotype (ataxic, dystonic, or tremor) of each model, suggesting that highly regular spiking is associated with ataxia, irregular spiking is associated with dystonia and rhythmic spiking is associated with tremor. After developing this classifier, they show that activating Purkinje neurons in different patterns that evoke interposed nuclear activity similar to their "ataxic", "dystonic", and "tremor" firing patterns induce similar behaviors in healthy mice. These results show convincingly that specific patterns of cerebellar output are sufficient to cause specific movement abnormalities. The extent to which cerebellar nuclear firing patterns are solely responsible for phenotypes in human disease remains to be established, however.

Strengths:
Major strengths are the recordings across multiple phenotypic models including genetic and pharmacologic manipulations, and the robust phenotypes elicited by Purkinje neuron stimulation.

Weaknesses:
The number of units recorded was small for each model (on the order of 20), limiting the conclusions that can be drawn from the recording/classifier experiments.

Reviewer #2 (Public Review):

Cerebellar diseases can manifest as various behavioral phenotypes, such as ataxia, dystonia, and tremor. In this study, Heijden and colleagues aim to understand whether these differing behavioral phenotypes are associated with disease-specific changes in the firing patterns of cerebellar output neurons in the cerebellar nuclei (CN). The authors effectively demonstrate that across different mouse models of cerebellar disease, there are distinct changes in the firing properties of CN neurons. They take a crucial step further by attempting to replicate disease-specific firing patterns in the cerebellar output neurons of healthy (control) mice using optogenetics. When Purkinje cells are stimulated in a manner that results in similar firing properties in CN neurons, the authors observe a variety of atypical behavioral responses, many of which align with the behavioral phenotypes observed in mouse models of the respective diseases.

Overall, the primary results are quite convincing. Specifically, they show that (1) different mouse models of cerebellar disease exhibit different statistics of firing in CN neurons, and (2) driving CN neurons in a time-varying manner that mimics the statistics measured in disease models results in behavioral phenomena reminiscent of the disease states. These findings suggest that aberrant activity in the CN can originate from various sources (e.g., developmental circuit deficits, abnormal plasticity, insult), but ultimately, these changes are funneled through the CN neurons, whose firing rates are affected, and this, in turn, drives aberrant behavior. This is a noteworthy observation that underscores the potential of targeting these output neurons in the treatment of cerebellar disease. Moreover, this manuscript provides valuable insights into the firing patterns associated with the most common cerebellar-dependent disease phenotypes.

However, the paper falls short in terms of the classifier model itself. The current implementation of this classifier appears to be rather weak. For instance, the cross-validated performance on the same disease line mouse model for tremor is only 56%. While I understand that the classifier aims to simplify a high-dimensional dataset into a more manageable decision tree, its rather poor performance undermines the authors' main objectives. In a similar vein, although focusing on three primary features of spiking statistics identified by the decision tree model (CV, CV2, and median ISI) is useful for understanding the primary differences between the firing statistics of different mouse models, it results in an overly simplistic view of this complex data. The classifier and its reliance on the reduced feature set are the weakest points of the paper and could benefit from further analysis and a different classification architecture. Nevertheless, it is commendable that the authors have collected high-quality data to validate their classifier. Particularly impressive is their inclusion of data from multiple mouse models of ataxia, dystonia, and tremor, enabling a true test of the classifier's generalizability.

Reviewer #3 (Public Review):

Summary:
This manuscript looks at the single-cell spike signatures taken from in vivo cerebellar nuclear neurons from awake mice suffering from 3 distinct diseases and uses a sophisticated classifier model to predict disease based on a number of different parameters about the spiking patterns, rather than just one or two. Single read-outs of spike firing patterns did not show significant differences between all 4 groups meaning that you need to analyze multiple parameters of the spike trains to get this information. The results are really satisfying and intriguing, with some diseases separating very well, and others having more overlap. It also represents a significant advancement for the rigor and creativity used for analyzing cerebellar output spike patterns. I really like this paper, it's a clever idea and has been done very well.

The authors examine multiple distinct forms of different diseases, including different types of ataxia, dystonia, and tremor. While some of the interpretation of this work remains unclear to this reviewer (in particular Figure 2, with ataxia models), I applaud the rigor and sharing of complex data that is not always straightforward to understand.

Strengths:
The work is technically impressive and the analysis pushes the envelope of how cerebellar dysfunction is classified, which makes it an important paper for the field. It's well written. The approach it is taking is clever. The analysis is thorough, and the authors examine a wide array of different disease models, which is time-consuming, costly, and very challenging to do. It's a very strong manuscript.

Weaknesses:
Weaknesses are few and quite minor. Some rewriting could be done to make certain sections clearer.

Author Response:

Update, January 11, 2024:

During the course of our careful revising of the paper, we discovered an inconsistency in the way we presented data for figures 5 and 6. Specifically, we used optogenetics to induce ataxia in mice. However, "ataxia", as a phenotype, can be initiated by a spectrum of cell dysfunctions as revealed by previous studies. We systematically explored this with optogenetics in this current work. Our error is that we presented one stimulation paradigm to show ataxic cell firing (2 ms on / 11 ms off square wave) and then presented a slightly different paradigm to show ataxic animal behavior (10 ms on / 10 ms off square wave). We note that our ataxia paradigms do not affect the outcomes of the dystonia and tremor stimulations. Importantly, the choice of ataxia paradigm does not change the conclusions of the paper. Regardless, for clarity we are actively working to make the stimulation parameters that we present consistent between figures 5 and 6.

October 10, 2023:

We would like to thank all three reviewers for providing excellent suggestions that will enable us to strengthen our manuscript and enhance the impact of our findings. We plan on addressing the comments by altering the text, providing additional data, revising the figures as requested, and most importantly by providing an improved classifier model. Where relevant, we will also provide the reviewers with a response to specific questions that they raised. We will respond to the reviewer’s comments in a point-by-point manner when we submit a revised manuscript. Below, we include an outline of the main points that we intend to address.

Although we will respond in full to all comments and suggestions in the revised documents, here we outline only the major areas in order provide context for our revisions. 1) The major point of concern raised by the reviewers is the strength of the classifier model. We agree with the reviewers that we should put forward the strongest model possible as this forms a core component of our paper. We are planning on retraining our model using the suggestions put forward by the reviewers in the public and author-directed comments. Importantly, given the healthy discussion about our model, our revised manuscript will now also include additional clarification about the choice of the model architecture and limitations of our data structure. Based on the reviewers’ comments, we will include a brief discussion about possible future ways of improving the model. 2) We will provide additional figures and updated figure panels to reflect the new data analyses. Ultimately, we agree that the major strength of our manuscript lies within the many mouse models tested and validation of the classification in different genetic, pharmacological, and optogenetic mouse models, a point raised by all three reviewers. We are confident that the revised images will reflect these strengths. 3) In addition to improving our classifier model, we are planning on making textual changes to clarify several parts of the text and propose a new title that better reflects the data put forth in our manuscript. 4) There are several minor but important comments that were raised by all three reviewers. We will also incorporate these changes as suggested.