Two functionally distinct Purkinje cell populations implement an internal model within a single olivo-cerebellar loop
Curation statements for this article:-
Curated by eLife
Evaluation Summary:
This manuscript features high quality experimental data with a detailed and clear analysis, combined with a neural network model to address the concept of differentiation in cerebellar functioning. This is an intensively debated topic currently and this work has an important, clear message to add to that debate. The data is very exciting, and the analyses and computational modeling very revealing and insightful. This stands on its own as a major contribution. The authors also raise an extremely interesting mechanistic interpretation of these data, which is tantalising but requires further support.
(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 #2 agreed to share their name with the authors.)
This article has been Reviewed by the following groups
Listed in
- Evaluated articles (eLife)
Abstract
Olivo-cerebellar loops, where anatomical patches of the cerebellar cortex and inferior olive project one onto the other, form an anatomical unit of cerebellar computation. Here, we investigated how successive computational steps map onto olivo-cerebellar loops. Lobules IX-X of the cerebellar vermis, i.e. the nodulus and uvula, implement an internal model of the inner ear’s graviceptor, the otolith organs. We have previously identified two populations of Purkinje cells that participate in this computation: Tilt-selective cells transform egocentric rotation signals into allocentric tilt velocity signals, to track head motion relative to gravity, and translation-selective cells encode otolith prediction error. Here we show that, despite very distinct simple spike response properties, both types of Purkinje cells emit complex spikes that are proportional to sensory prediction error. This indicates that both cell populations comprise a single olivo-cerebellar loop, in which only translation-selective cells project to the inferior olive. We propose a neural network model where sensory prediction errors computed by translation-selective cells are used as a teaching signal for both populations, and demonstrate that this network can learn to implement an internal model of the otoliths.
Article activity feed
-
Evaluation Summary:
This manuscript features high quality experimental data with a detailed and clear analysis, combined with a neural network model to address the concept of differentiation in cerebellar functioning. This is an intensively debated topic currently and this work has an important, clear message to add to that debate. The data is very exciting, and the analyses and computational modeling very revealing and insightful. This stands on its own as a major contribution. The authors also raise an extremely interesting mechanistic interpretation of these data, which is tantalising but requires further support.
(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 #2 agreed to share their name …
Evaluation Summary:
This manuscript features high quality experimental data with a detailed and clear analysis, combined with a neural network model to address the concept of differentiation in cerebellar functioning. This is an intensively debated topic currently and this work has an important, clear message to add to that debate. The data is very exciting, and the analyses and computational modeling very revealing and insightful. This stands on its own as a major contribution. The authors also raise an extremely interesting mechanistic interpretation of these data, which is tantalising but requires further support.
(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 #2 agreed to share their name with the authors.)
-
Reviewer #1 (Public Review):
In this very interesting manuscript, Angelaki and Laurens suggest that the same climbing fiber feedback from the inferior olive may innervate functionally distinct classes of Purkinje cells. That is, the same inferior olivary signals, as assessed via the relationship between vestibular stimuli and complex spike frequency, appears to exist in multiple classes of vestibular Purkinje cells. This is an interesting and, in some ways, unexpected observation.
Strengths:
Overall, this is a well written and thorough manuscript. The neurophysiology data are clear, and the data analyses are presented clearly and performed rigorously. To that end, the authors' primary claim is well-supported. It is clear that "tilt" and "translation" selective Purkinje cells are innervated by the same or very similar climbing fibers …
Reviewer #1 (Public Review):
In this very interesting manuscript, Angelaki and Laurens suggest that the same climbing fiber feedback from the inferior olive may innervate functionally distinct classes of Purkinje cells. That is, the same inferior olivary signals, as assessed via the relationship between vestibular stimuli and complex spike frequency, appears to exist in multiple classes of vestibular Purkinje cells. This is an interesting and, in some ways, unexpected observation.
Strengths:
Overall, this is a well written and thorough manuscript. The neurophysiology data are clear, and the data analyses are presented clearly and performed rigorously. To that end, the authors' primary claim is well-supported. It is clear that "tilt" and "translation" selective Purkinje cells are innervated by the same or very similar climbing fibers (e.g., Figures 3 and 5). This observation expands the idea of functional organization in the cerebellum by climbing fibers.
Weaknesses:
While the primary conclusion of this study is well-supported, the model and theoretical aspects should be enhanced. It is difficult to assess whether the authors' secondary claim about the ability for a shared climbing fiber input to suitably "teach" both tilt and translation responsive Purkinje cells is justified. Much of the issue with the model likely stem from a lack of sufficient information in the methods/results sections to fully appreciate how the model functions.
-
Reviewer #2 (Public Review):
Summary:
This manuscript by Angelaki and Laurens investigates the complex spike response properties of Purkinje cells in the nodulus and uvula during translation and/or tilt. The authors find that two previously identified populations of Purkinje cells, one that has tilt-selective simple spike responses and one with translation-selective simple spikes, have very comparable complex spike response properties. Using additional experimental and modeling data they propose a configuration in which the computations in translation-selective cells are used as teaching signals for both populations. These data are interesting and have an important message to add to a growing field of data on (functional) differentiation within the cerebellum.
Review:
The manuscript is based on a subset of recordings from a dataset that …
Reviewer #2 (Public Review):
Summary:
This manuscript by Angelaki and Laurens investigates the complex spike response properties of Purkinje cells in the nodulus and uvula during translation and/or tilt. The authors find that two previously identified populations of Purkinje cells, one that has tilt-selective simple spike responses and one with translation-selective simple spikes, have very comparable complex spike response properties. Using additional experimental and modeling data they propose a configuration in which the computations in translation-selective cells are used as teaching signals for both populations. These data are interesting and have an important message to add to a growing field of data on (functional) differentiation within the cerebellum.
Review:
The manuscript is based on a subset of recordings from a dataset that formed the basis for previous work of the authors. The complex spike data are a very relevant part of Purkinje cell functioning, so it is good to see an in-depth analysis here. The recordings are of high quality and the analysis is detailed and clear. The only concern I have is in the interpretation, or perhaps the semantics of the interpretation, of the results.
This concerns centers around the use of what, by visual inspection, indeed appear to be similar CS histograms for translation, tilt and tilt-trans, as conclusive evidence that the Purkinje cells in both populations receive the same IO input. Although I agree this is an understandable conclusion when taking anatomical data of others into account, I think the evidence, consisting of comparably shaped CS histograms, is not conclusive. For instance, it is possible that a simultaneous CS activation in groups of PCs could be provided without an identical IO projection. However, the evidence is strong and the proposed concept is very interesting.
-
Reviewer #3 (Public Review):
Angelaki and Laurens examine the complex spike responses of two types of Purkinje cells (translation-PC and tilt-PC) whose simple spike responses they had previously characterized. They find that even though the simple spike responses of translation-PCs and tilt-PCs are very different from each other during 3D head movements, the two types of Purkinje cells have similar complex spike responses. These results are interpreted within the context of an artificial neural network model, leading the authors to suggest that Purkinje cells within the same cerebellar module (i.e. Purkinje cells with similar complex spike responses) can learn to generate different simple spike responses.
The data is very exciting, and the analyses and computational modeling very revealing and insightful. The authors clearly demonstrate …
Reviewer #3 (Public Review):
Angelaki and Laurens examine the complex spike responses of two types of Purkinje cells (translation-PC and tilt-PC) whose simple spike responses they had previously characterized. They find that even though the simple spike responses of translation-PCs and tilt-PCs are very different from each other during 3D head movements, the two types of Purkinje cells have similar complex spike responses. These results are interpreted within the context of an artificial neural network model, leading the authors to suggest that Purkinje cells within the same cerebellar module (i.e. Purkinje cells with similar complex spike responses) can learn to generate different simple spike responses.
The data is very exciting, and the analyses and computational modeling very revealing and insightful. The authors clearly demonstrate that Purkinje cells with (apparently) similar climbing fiber inputs can generate very different simple spike responses during 3D head movements. This is a novel finding that will stimulate new discussion about cerebellar function among researchers in the field, and it stands on its own.
However, the authors make 2 'big' claims that are not fully supported by their analyses:
- Are translation-PCs and tilt-PCs in the same olivo-cerebellar loop? The authors state that both Purkinje cell populations are part of a single olivo-cerebellar loop because their complex spikes are driven by the same sensory prediction error.
- While both translation-PCs and tilt-PCs appear to have similar complex spike responses during the specific conditions and 3D head movements tested, examining complex spike responses in other conditions/other stimuli may reveal differences that would indicate that the Purkinje cells belong to different olivo-cerebellar loops (for example, all experiments were performed in the dark and consisted of passive head movements; is it possible that some Purkinje cells, but not others, could emit visually-driven complex spikes in addition to the translation-related complex spikes observed in the dark in this study? Or could the complex spike responses of tilt-PCs and translation PCs become different from each other during active movements?).
- Ultimately, whether two Purkinje cells belong to the same olivo-cerebellar loop is an anatomical question; it can only be demonstrated by confirming not only that (1) the Purkinje cells receive the same climbing fiber input from the inferior olive, but also that (2) the output of the Purkinje cells is sent back to the same neurons in the inferior olive. The output pathways of translation-PCs and tilt-PCs are unknown, but the correlational data in Supplementary Fig. 5S1 seems to suggest that they are different (at least with regards to their projections to the inferior olive). In the neural network model, it is assumed that only the output of translational-PCs (but not tilt-PCs) is sent to the part of the olive that projects back to both translational-PCs and tilt-PCs, whereas the output of tilt-PCs is sent to translational-PCs but not to the inferior olive. There is no anatomical data available to support this unconventional architecture.
- Can translation-PCs and tilt-PCs use the same complex spikes to learn their respective simple spike responses during 3D head movements? The authors mention that their computational model supports the hypothesis that a single olivocerebellar loop with a shared error signal (in the form of a complex spike) can learn the diverse Purkinje cell responses encountered in the cerebellum (page 18, lines 471-474). However, the neural network model only implements plasticity in the inputs to the tilt-PCs. There is no plasticity in the local synapses formed by the inputs to the translation-PCs. In addition, the model assumes that translation-PCs in different olivo-cerebellar loops will receive selective otolith inputs encoding a specific component of the GIA (e.g. GIAy). In the model, the weights of these highly selective otolith inputs are not learned; they are set at the beginning and cannot be changed. Thus, at the beginning of the simulation, it is these 'fixed' GIA inputs that are entirely responsible for the responses of translation-PCs (these responses are not learned), and for the responses of the inferior olive that drive plasticity in tilt-PCs. In effect, the specialized GIA inputs to the translation-PCs serve as the teaching signal for the tilt-PCs, at least at the beginning of the simulation period. There is no empirical evidence to support the idea that Purkinje cell inputs come in different flavors, including one set of inputs that is plastic and one set of specialized and highly selective inputs that is hardwired and unchangeable. It is unclear whether the network model would be able to learn the correct responses of tilt-PCs and translation-PCs if the weights of *all* the local inputs to both tilt-PCs and translation-PCs were randomly distributed at the beginning and allowed to change according to the same decorrelation plasticity rule.
As a final comment, although Purkinje cells were classified as tilt-PCs or translation-PCs, the data makes it clear that the responses of both cell populations, both with regards to their simple spike and complex spike responses, are not 100% selective and contain components related to both tilt and translation. The gain of the response for one component (e.g. tilt) is frequently just 2 or 4 times higher than the gain of the response for the other component (e.g. translation). How this diversity of responses could emerge, or what the functional significance could be for the operation of the circuit, was not addressed in the neural network model.
-
-