Dynamic organization of cerebellar climbing fiber response and synchrony in multiple functional components reduces dimensions for reinforcement learning

Curation statements for this article:
  • Curated by eLife

    eLife logo

    eLife assessment

    This is an important study of the dimensionality and synchrony of calcium responses in Purkinje cells measured across a large region of the cerebellar cortex over the course of learning. This work has the potential to inform our understanding of the functional organization of the cerebellum and longstanding hypotheses about the role of cerebellar climbing fibers in the induction of learning and in the timing of movement, but the evidence provided for the many sweeping claims is incomplete. The paper would benefit from additional statistical analyses to more rigorously evaluate the central claims, with consideration of appropriate comparison groups and potential confounds.

This article has been Reviewed by the following groups

Read the full article

Abstract

Cerebellar climbing fibers convey diverse signals, but how they are organized in the compartmental structure of the cerebellar cortex during learning remains largely unclear. We analyzed a large amount of coordinate-localized two-photon imaging data from cerebellar Crus II in mice undergoing ‘Go/No-go’ reinforcement learning. Tensor component analysis revealed that a majority of climbing fiber inputs to Purkinje cells were reduced to only four functional components, corresponding to accurate timing control of motor initiation related to a Go cue, cognitive error-based learning, reward processing, and inhibition of erroneous behaviors after a No-go cue. Changes in neural activities during learning of the first two components were correlated with corresponding changes in timing control and error learning across animals, indirectly suggesting causal relationships. Spatial distribution of these components coincided well with boundaries of Aldolase-C/zebrin II expression in Purkinje cells, whereas several components are mixed in single neurons. Synchronization within individual components was bidirectionally regulated according to specific task contexts and learning stages. These findings suggest that, in close collaborations with other brain regions including the inferior olive nucleus, the cerebellum, based on anatomical compartments, reduces dimensions of the learning space by dynamically organizing multiple functional components, a feature that may inspire new-generation AI designs.

Article activity feed

  1. Author Response

    Reviewer #1 (Public Review):

    Hoang, Tsutsumi and colleagues use 2-photon calcium imaging to study the activity of Purkinje cells during a Go/No-go task and related this activity to their location in Aldolase-C bands. Tensor component analysis revealed that a substantial part of the calcium responses can be linked to four functional components. The manuscript addresses an important question with an elegant technical approach and careful analysis. There are a few points that I think could be addressed to further improve the quality of the manuscript.

    1. The authors should be careful not to overstate the goal and results. For instance, in the abstract it is stated that dynamical functional organization is necessary for dimension reduction. However, the statement that the 4 TCs together account for about half of the variance (line 220) indicates that dimensionality may not be reduced that much. I would suggest revising the first and last sentence of the abstract accordingly.

    Dynamic functional organization of TC1 and TC2 by synchronization is the major finding of this study and we believe that it is one of the most efficient mechanisms of dimension reduction, given the unique anatomy of the cerebellum. In the revised manuscript, we added a supplemental result showing that the dimensionality of TC1 and TC2 neurons decreased and increased, respectively, in accordance with bi-directional changes in their synchronization (Figure 3 – figure supplement 1DE). Dimension reduction was further confirmed by conventional PCA (Figure 6 – figure supplement 1). However, we agree that the statement that the cerebellum reduces dimensions by self-organization of components is speculative, and we revised the abstract accordingly.

    At the end of the introduction, the authors refer to "the first evidence supporting the two major theories of cerebellar function" but which two theories is referred to and how this manuscript support them is not very obvious. Similarly, they state that "This study unveiled the secret of cerebellar functional architecture", which I would consider to be an unnecessary overstatement of the impact of the work described.

    In the revised Introduction, we explicitly stated that TC1 and TC2 are related to timing control and cognitive error learning, respectively, with some indirect causal evidence. We also revised the last paragraph of the Introduction to emphasize that this study provides the first evidence to support the view that distinct cerebellar components may serve divergent cerebellar functions in a single task. The statement "This study unveiled the secret of cerebellar functional architecture" was removed.

    In the title, the authors use the word modular. In the consensus paper on cerebellar modules (Apps et al., 2018) an attempt is made to unify the terms used to describe cerebellar anatomical structures. Here "module" is used for the longitudinal zone of interconnected PCs, CN neurons and olivary neurons. As the authors only studied PC activity (and indirectly the IO), I would suggest using band, stripe or subpopulation instead.

    Because we used TCA to identify functional components underlying the Go/No-go data, we changed the word “module” to “component” in the title.

    Finally, the term "CF firing" or "CF activity" is used when referring to the recorded signals. However, the authors measure postsynaptic calcium responses that are indeed likely driven by CF inputs, but could also be influenced by PF inputs. At the very least, because Purkinje cells and not climbing fibers are being imaged, "complex spike" should be used instead. It would be more accurate still to use the more general "calcium response" and make less of an assumption about the origin of the calcium response.

    In this study, CF-dependent dendritic Ca2+ signals in adjacent AldC compartments were recorded by the two-photon imaging. The HA_time algorithm (Hoang et al. 2020) was then applied to extract spike timings from the recorded signals. In the revised manuscript, we used the terms “calcium responses” and “complex spikes” when referring to the recorded Ca2+ signals and the estimated spikes, respectively.

    1. For some figure panels and statements in the manuscript error bars or confidence intervals and statistics are missing. This is the case for, for example, the changes in fraction correct, lick latency, fraction incorrect, etc. (Fig 1B, 2E-F, TC levels in 3, 4D-E and 5A-C). Including these is particularly relevant in Fig 4E as this is a key result, mentioned also in the abstract. Please indicate clearly if these plots are cumulative for all mice or per mouse and averaged. I advise the authors to statistically support the claim that the changes are significant and in opposite direction as this element of the study is referred to in the abstract and discussion (summary).

    We added the error bars / confidence intervals to the related figures. Most importantly, we added histograms of synchrony strength for TC1/TC2 neurons (Figure 4E) and conducted statistical tests to strengthen the claim of bi-directional changes in synchronization of TC1/TC2.

    1. Data presentation sometimes does not do the work justice. For example, the data in Figure 6 are very interesting, but hard to read because of the design of the figure. It is clear how the components are mostly confined to Aldolase-C domains, but within the domains the distribution is not clear. I would advise to also more clearly indicate what the locations of the colors within the bands refers to. The spatial distribution of the selected top 300 cells for each TC could be added.

    We added pie-chart plots for the fraction of TC1-4 neurons in each Ald-C zone and learning stage. We also indicated in the figure legend that the location of a single-color bar referred to the geographic distance of the corresponding neuron relative to Ald-C boundaries. We included spatial distribution of the selected neurons in Figure 4 – figure supplement 1D.

  2. eLife assessment

    This is an important study of the dimensionality and synchrony of calcium responses in Purkinje cells measured across a large region of the cerebellar cortex over the course of learning. This work has the potential to inform our understanding of the functional organization of the cerebellum and longstanding hypotheses about the role of cerebellar climbing fibers in the induction of learning and in the timing of movement, but the evidence provided for the many sweeping claims is incomplete. The paper would benefit from additional statistical analyses to more rigorously evaluate the central claims, with consideration of appropriate comparison groups and potential confounds.

  3. Reviewer #1 (Public Review):

    Hoang, Tsutsumi and colleagues use 2-photon calcium imaging to study the activity of Purkinje cells during a Go/No-go task and related this activity to their location in Aldolase-C bands. Tensor component analysis revealed that a substantial part of the calcium responses can be linked to four functional components. The manuscript addresses an important question with an elegant technical approach and careful analysis. There are a few points that I think could be addressed to further improve the quality of the manuscript.

    1. The authors should be careful not to overstate the goal and results. For instance, in the abstract it is stated that dynamical functional organization is necessary for dimension reduction. However, the statement that the 4 TCs together account for about half of the variance (line 220) indicates that dimensionality may not be reduced that much. I would suggest revising the first and last sentence of the abstract accordingly.
    At the end of the introduction, the authors refer to "the first evidence supporting the two major theories of cerebellar function" but which two theories is referred to and how this manuscript support them is not very obvious. Similarly, they state that "This study unveiled the secret of cerebellar functional architecture", which I would consider to be an unnecessary overstatement of the impact of the work described.
    In the title, the authors use the word modular. In the consensus paper on cerebellar modules (Apps et al., 2018) an attempt is made to unify the terms used to describe cerebellar anatomical structures. Here "module" is used for the longitudinal zone of interconnected PCs, CN neurons and olivary neurons. As the authors only studied PC activity (and indirectly the IO), I would suggest using band, stripe or subpopulation instead.
    Finally, the term "CF firing" or "CF activity" is used when referring to the recorded signals. However, the authors measure postsynaptic calcium responses that are indeed likely driven by CF inputs, but could also be influenced by PF inputs. At the very least, because Purkinje cells and not climbing fibers are being imaged, "complex spike" should be used instead. It would be more accurate still to use the more general "calcium response" and make less of an assumption about the origin of the calcium response.

    2. For some figure panels and statements in the manuscript error bars or confidence intervals and statistics are missing. This is the case for, for example, the changes in fraction correct, lick latency, fraction incorrect, etc. (Fig 1B, 2E-F, TC levels in 3, 4D-E and 5A-C). Including these is particularly relevant in Fig 4E as this is a key result, mentioned also in the abstract. Please indicate clearly if these plots are cumulative for all mice or per mouse and averaged. I advise the authors to statistically support the claim that the changes are significant and in opposite direction as this element of the study is referred to in the abstract and discussion (summary).

    3. Data presentation sometimes does not do the work justice. For example, the data in Figure 6 are very interesting, but hard to read because of the design of the figure. It is clear how the components are mostly confined to Aldolase-C domains, but within the domains the distribution is not clear. I would advise to also more clearly indicate what the locations of the colors within the bands refers to. The spatial distribution of the selected top 300 cells for each TC could be added.

  4. Reviewer #2 (Public Review):

    Hoang, Tsutsumi et al provide a comprehensive functional mapping of cerebellar climbing fiber responses in Lobule Crus II. The study derives from analysis of a dataset originally published in Tsutsumi et al eLife 2019, using two photon Ca2+ imaging throughout the learning of a Go/No-go reward-driven licking behavior. Each recording session yielded data from a ~two-hundred micron patch of tissue, with neurons spatially localized relative the "zebrin" banding pattern of the cerebellar cortex as reported by an aldolaceC-tdTomato transgenic line. In the present work, complex spike times were extracted at higher temporal resolution using subframe raster line-scan timing information, and then decomposed at the trial-averaged population level using tensor component analysis.

    The central conclusion is that the entirety of crus II climbing fiber responses decomposes into just a few patterns that capture key features of the behavior. Some of these patterns strengthen with learning, i.e., feature climbing fiber spiking that increases in frequency, while others decay with learning, i.e., feature climbing fiber responses that are prominent only in novice animals. These different climbing fiber activity components are in some cases associated with either positive or negative aldolace-C compartments of crus II. Finally, synchronization is concentrated among cells contributing to the same tensor components, and synchrony levels increase or decrease for different components over learning.

    The analysis therefore suggests that distinct principles of climbing fiber function can be present simultaneously in distinct cerebellar modules (and, according to the TCA cell weightings, potentially simultaneously in individual climbing fibers). This conclusion is contrary to the implied dichotomy in the literature that climbing fibers either function as "error signals" or as "timing signals" in a particular behavioral context or cerebellar region. The authors speculate that resolution of this dichotomy could result from the biophysics of the inferior olive, in which flexibly coupled oscillators might self-organize into a low dimensional decomposition of task dynamics. Relatedly, the authors speculate that changes in synchronization that contrast between different components could serve to either regulate instructive signal dimensionality or climbing fiber timing functions, depending on each component's functional contribution. From a theoretical standpoint, this is a helpful new direction. The framework is more agnostic to the details of the activity profiles of any specific group of climbing fibers, but more attuned to the systems-level distribution of activity profiles and how these might collectively serve a behavior.

    A valuable feature of the study is the simultaneous analysis of many imaging fields spanning 17 subjects and the entire dorsal surface of crus II. This bypasses some of the recurring interpretational issues with climbing fiber recordings that stem from their spatial organization across the cerebellar surface with often abrupt transitions at compartmental boundaries. By decomposing responses across many compartments simultaneously (at the trial-averaged level), the authors provide a quantitative estimate of the diversity of response patterns and their distribution across space and cells. It's worth noting that this approach is also a double-edged sword, as the trial-averaged decomposition does not depend on single-trial correlations between neurons, thus strictly speaking leaving it an open question whether apparently similar climbing fiber patterns present in distant imaging fields exhibit correlated variability either across trials or across learning.

    The data convincingly show that several dominant tensor components explain a large amount of climbing fiber variance across crus II. The authors speculate that this reflects an olivary decomposition of task dynamics. Due to the nature of the analysis - TCA applied over an entire dataset - there is not a clear test of this hypothesis in the present manuscript.

    The authors also present the interesting and compelling result that different CF response patterns undergo opposite learned changes in synchronization. They speculate that different trajectories of synchronization, specifically, increases for TC1 (hit) and decreases for TC2 (false alarm), could reflect different functional uses of TC1 and TC2, although it is difficult to assess the likelihood of this being true based on the data and analyses presented.