Consensus Paper: Towards a Systems-Level View of Cerebellar Function: the Interplay Between Cerebellum, Basal Ganglia, and Cortex

The cerebellum works in concert with cortex and the basal ganglia as a fundamental building block in motor and cognitive tasks of various complexity, from sensorimotor mapping to reasoning [1, 2, 3, 4]. This systems-level view is increasingly supported by evidence demonstrating that the cerebellum and the basal ganglia receive input from, and send output to, different cortical areas through multisynaptic loops that have been assumed to be anatomically segregated and to perform distinct functional operations [5, 6, 7, 8, 9, 10]. Moreover, recent findings reveal the existence of an anatomical substrate for the bidirectional communication between cerebellum and basal ganglia. In particular, studies on rats [11] and monkeys [12] have demonstrated that the cerebellum sends a strong disynaptic projection to the striatum through the thalamus. Furthermore, recent studies in monkeys have shown that the subthalamic nucleus sends a disynaptic projection to the cerebellar cortex by way of the pontine nuclei [13]. Similar evidence has been recently reported in the human brain [14].

These data have stimulated new research to investigate the reciprocal influence between these brain areas and the different forms of learning typically associated with them: supervised learning in the cerebellum based on plasticity of parallel fiber-Purkinje cell synapses [15, 16, 17]; unsupervised learning in the cortex based on associative Hebbian processes [18, 19, 20]; and trial-and-error (reinforcement learning) in the basal ganglia based on reward prediction errors computed at level of the striatum and depending on dopamine signals [21, 22]. Several recent studies have tried to integrate these apparently disconnected learning processes, focusing for example on the role of the cerebellum in regulating the plasticity of premotor and motor networks during sensorimotor learning [23, 24, 25, 26, 27]; the involvement of the cerebellum in the reinforcement learning processes underlying the computation of sensory and reward prediction errors during motor adaptation [28, 29, 30, 31, 32, 33]; the role of the cerebellum and basal ganglia in the cortical acquisition of high-level cognitive (non-motor) functions [1, 3, 4]; and the influence of cerebellum and basal ganglia on motor cortical plasticity in health and in Parkinson’s disease [34, 35]. Finally, data supporting a close interaction between cerebellar, basal ganglia, and cortical areas have recently stimulated new research on the role of cerebellum and basal ganglia in functions typically associated with cortical areas (e.g., action understanding [4]) and the role of cerebellar and cortical areas in impairments typically associated with basal ganglia, such as Tourette’s syndrome, dystonia, and Parkinson’s disease [36, 37, 38, 39, 40].

Despite this recent progress, we still lack a comprehensive framework that defines cerebellar function from a wider, systems-level perspective [4, 41, 42]. Viewing cerebellum, cortex, and basal ganglia as an integrated system could lead to understand their functional and learning processes in radically different ways. The goal of this consensus paper is to feed the discussion on the fundamental interactions between the functional and plasticity processes of cerebellum, cortex, and basal ganglia, conceived as forming an integrated system underlying several motor and cognitive functions.

Evidence of the interactions between the functional and learning processes in the cerebellum, cortex, and basal ganglia comes from different scientific disciplines and methodologies: from neurophysiology to computational modelling and from behavioral methods to neuropsychological and brain imaging approaches. Unfortunately, discussion between experts using these different approaches is often limited. We believe that a multi-methodological and multidisciplinary discussion is necessary to fully understand the integrated dynamics of the cerebellar-basal-ganglia-cortical system. This consensus paper offers an up-to-date overview on this discussion as it collects contributions from an interdisciplinary community of scientists actively involved in the investigation of cerebellum that provide a range of different, sometimes controversial, viewpoints.

The issue begins with the paper of Bostan and Strick that introduces and reviews the recent neurophysiological and neuroanatomical evidence on the cerebellar connections with the cerebral cortex and the basal ganglia.

The next two contributions, by Verschure et al. and by Jörntell, further expand the systems-level view of cerebellar interconnections supported by the analysis of the previous piece (by Bostan and Strick). Verschure et al. raise the fascinating possibility that the nucleo-olivary pathway balances the contribution of feedback (mediated by brain stem nuclei) and feed-forward (cerebellar) modes of motor control. Jörntell provides an overview of the spinal inputs into the cerebellum, discussing how these inputs allow the cerebellum to keep informed about the excitatory drive on low level motor functions. He thus proposes that the major role of the cerebellum is to allow many spinal inputs to be associated with one another in specific environmental contexts.

The contribution of Houk provides a theoretical, systems-level perspective of how the cerebellar-cortical and basal ganglia-cortical loops may interact to affect learning and control processes.

Along similar lines, the contribution of Doya elaborates his previous theory of motor control and learning, based on the idea that different brain areas implement different learning algorithms, to account for recent findings discussed by Bostan and Strick on the circuitry between the cerebellum and the basal ganglia.

Pivoting on recent empirical evidence, Miall suggests that the cerebellum may support a single operation, acting as a short-term predictor within the cerebellar-basal-ganglia-cortical system. He also discusses data on medium-term, post-learning, memory consolidation effects in this system.

Lago-Rodriguez and Galea summarize the importance of using non-invasive techniques (such as transcranial magnetic stimulation (TMS) or transcranial direct current stimulation (tDCS)) to investigate the causal role of the cerebellum in motor function and its interactions with other brain regions.

Finally, the last two papers provide some examples of the consequences of abnormal interactions between cerebellum, basal ganglia, and cortical systems. Popa and Kishore discuss data supporting how impaired cerebellar plasticity mechanisms could contribute to Parkinson’s disease and dyskinesia. Caligiore et al. build on recent data that suggest an involvement of cerebellum in Parkinson’s disease and discuss some cortical-subcortical circuits within the cerebellar-basal-ganglia-cortical network that could be involved in Parkinson’s resting tremor.

Taken together, these empirical and theoretical contributions offer readers a broad and up-to-date framework to describe the dynamic interplay between functioning and learning mechanisms in the cerebellum, basal ganglia, and cortex, pointing out the key open issues for future research. The final part of the article summarizes the main topics and analyzes the main elements of consensus—or lack of consensus among the authors. As discussed in the “Summary and conclusions” section, consensus exists on several key aspects of cerebellar-basal-ganglia-cortical function; at the same time, there are various topics that are still debated and where consensus cannot be reached yet, reflecting the relative novelty of the topic. These open points represent interesting and promising directions for future research.

The classical view of cerebellar function has been that it is exclusively concerned with the control of movement through descending control and influence on the primary motor cortex (M1). The cerebellum was thought to receive information from multiple neocortical areas and funnel it back to M1 [43]. More recent analyses of cerebellar outputs have resulted in a dramatic shift in the conceptualization of cerebellar function.

It is now clear that efferents from the cerebellar nuclei target multiple subdivisions of the thalamus [44] that reach widespread neocortical areas. Experiments using neurotropic viruses as transneuronal tracers have been crucial for the identification of neocortical areas that are the targets of cerebellar output (see [1] for a review). These experiments demonstrate that cerebellar output influences not only M1 but also premotor, prefrontal, and parietal areas (Fig. 1a, b). Output channels to M1 and the premotor areas are clustered in a dorsal region of the cerebellar dentate nucleus, identifying a motor domain within this nucleus. Output channels to prefrontal cortex are clustered together in a ventral (non-motor) region of the nucleus, entirely outside the motor domain [45] (Fig. 1b).

Open image in new window

To date, all of the areas in the cerebral cortex that are targets of cerebellar output also have prominent projections to the cerebellar cortex via the pontine nuclei (Fig. 1a). Thin axon collaterals from approximately 20 % of ponto-cerebellar neurons reach the dentate nucleus [46] and may allow excitatory inputs from cortical areas, such as M1, to reach the dentate [47]. Interestingly, areas of the cerebral cortex that lack substantial projections to the cerebellum (e.g., areas 46v, 12, and TE) do not appear to be major targets of cerebellar output. If these principles apply to all cerebro-cerebellar networks, then all of the areas of cerebral cortex that project to the cerebellum may be the targets of cerebellar output. This would include such diverse areas as regions of extrastriate cortex, cingulate cortex, and the parahippocampal gyrus [48, 49].

The distinct motor and non-motor domains observed in the dentate nucleus have their counterparts in cerebellar cortex (Fig. 1c, d). The motor domain includes a region primarily in the anterior lobe (lobules III–VI) and another in the paramedian lobule and adjacent cerebellar hemisphere (HVIIB and HVIII). These regions have been shown to both receive inputs from and send outputs to M1, forming a closed-loop circuit [50] (Fig. 1c). The non-motor domain is extensive and involves the portions of the posterior vermis and hemisphere that lie between the two regions of motor representation. Regions within the non-motor domain of cerebellar cortex also participate in distinct closed-loop circuits with regions of prefrontal cortex [50] (Fig. 1d). Thus, multiple closed-loop circuits represent a fundamental feature of interactions between the cerebellum and the cerebral cortex.

The evidence that cerebellar output influences multiple areas of prefrontal and posterior parietal cortex provides the anatomical substrate for a distinct cerebellar contribution to non-motor aspects of behavior [51]. To date, most theories of cerebellar function are based on the understanding of cerebellar contributions to motor control [41, 51]. Further research is needed to determine the nature and full extent of cerebellar contributions to non-motor function.

The circuits that link the cerebellum with the cerebral cortex have traditionally been considered to be anatomically and functionally distinct from those that link the basal ganglia with the cerebral cortex [12, 44]. Any interactions between cerebro-cerebellar and cerebro-basal ganglia loops were thought to occur primarily at the neocortical level. Results from recent anatomical experiments challenge this perspective and provide evidence for disynaptic pathways that link the cerebellum with the basal ganglia (Fig. 2). Retrograde transneuronal transport of rabies virus demonstrated that both motor and non-motor portions of the dentate nucleus project disynaptically to the striatum (putamen and caudate) [12]. Similarly, both motor and non-motor portions of the subthalamic nucleus (STN) of the basal ganglia project disynaptically to motor and non-motor regions of cerebellar cortex [13]. The interconnections between the cerebellum and the basal ganglia provide the neural basis for cerebellar involvement in disorders typically associated with the basal ganglia (e.g., Parkinson’s disease, dystonia, Tourette syndrome and addiction), as well as in normal basal ganglia functions, such as reward-related learning [3, 37]. Furthermore, there is evidence that cerebellar output can alter striatal plasticity [52]. These new observations indicate that exploring the physiology and functions of the interconnections between the cerebellum and the basal ganglia represents an important new direction for future research.

Open image in new window

Overall, neuroanatomical studies using virus transneuronal tracers have demonstrated that the output from the cerebellum reaches vast areas of the cerebral cortex, including regions of prefrontal and posterior parietal cortex. Furthermore, the cerebellum is reciprocally connected with the basal ganglia, indicating that the two subcortical structures are part of a densely interconnected network. These new results challenge us to discover both the entire range of behaviors that are influenced by this network and the neural computations it implements.

The Distributed Adaptive Control theory of mind and brain describes the brain as a multi-layered control system including the body, predefined reactive control, adaptive control for state space acquisition and action shaping, and lastly the contextual control for the generation of goal-oriented plans for action [53, 54]. This raises the specific question of how these different control layers interact. The study of the cerebellum allows us to investigate in detail how reactive and adaptive control systems are interfaced. In particular, the eyeblink conditioning paradigm allows the study of the integration of cerebellar predictive commands with brainstem level reactive feedback commands. Here, we discuss this interaction between cerebellum and these lower level reflexes with the idea that it can inform hypotheses about the interaction between the cerebellum and high-level structures of the central nervous system, such as cortex and basal ganglia. In eyeblink conditioning, a subject learns to respond with an anticipatory action—the so-called conditioned response (CR)—to a naturally neutral stimulus—the conditioned stimulus (CS)—that has been coupled through the experimental training with a blink-eliciting noxious stimulus—the unconditioned stimulus (US). Experimental evidence suggests that the cerebellum provides the substrate for CR acquisition [55, 56, 57, 58, 59, 60, 61, 62, 63] (but see [66] and Delgado-García in [58] for alternative views). In line with that hypothesis, it has been proposed that in eyeblink conditioning, the cerebellum substitutes reactive reflexes by anticipatory actions [64]. However, a closer look at the resultant behavior reveals that, rather than totally replacing reactive feedback commands, feed-forward cerebellar commands are merged with them [59, 65, 68].

The difference between total or gradual replacement of reflexes by anticipatory actions becomes relevant when considering an instrumental version of the eyeblink conditioning paradigm where the US is provided by an air puff whose noxious effect can partially be prevented by the anticipatory blink itself [59, 61, 66, 68]. In contrast, in classical—or Pavlovian—eyeblink conditioning electrical shocks to the periorbital area are usually employed as US, in which case neither the UR nor the CR have operational value, i.e., the responses do not ameliorate the effect of the US [60, 67]. In the instrumental contingency that uses air puffs as US, the anticipatory response establishes a behavioral feedback that, as learning progresses, diminishes the sensed intensity of the aversive and learning-inducing US [65, 66, 67]. One may expect that the subject will completely avoid any aversive effect of the unconditioned stimulus by blinking preemptively, namely, to expect a complete substitution of reactive feedback by predictive feed-forward control after learning. However, behavior indicates that at learning asymptote subjects only display a partial anticipatory closure of the eyelid to the CS and will only completely close their eyes after perception of the US [59, 68, 69, 70]. This implies that a significant part of the protective action of the putative CR still takes place reactively and is thus a UR.

Computational models of cerebellar learning coupled to a reactive controller have explained the gradual replacement of the UR by the CR in terms of internal negative feedback provided by the nucleo-olivary inhibition (NOI) [65, 71]. Via this negative feedback, the CS-dependent acquired pause of Purkinje cells firing (Purkinje cell CR) [57], leads to dis-inhibition of nucleo-olivary cells that in turn increase inhibition of the cells in the inferior olive (IO) [72] (Fig. 3). Thus, at the level of the cerebellum, this nucleo-olivary projection achieves, internally, what the behavioral negative feedback achieves externally: it reduces the intensity of the learning-inducing error signal, specifically, the driving of the activity of the IO cells above their baseline firing rate, which is in the range of 1 Hz. As a result, an air puff US can fail to initiate plasticity in the cerebellar cortex because the excitation it provides to the IO has been canceled by the NOI recruited by the cerebellar CR [73, 74]. In consequence, under the assumptions of bidirectional plasticity in the cerebellar cortex and spontaneous activity in the IO, the NOI prevents the complete substitution of a feedback by a feed-forward mode of control [65] and thus implements a mixed feedback/feed-forward controller.

Open image in new window

We can explain this hybrid feedback/feed-forward control model from the perspective of optimality [75]. Namely, if one considers that both the protective action and the failure to avoid the US carry a cost, optimal behavior will depend on an effort/error tradeoff. In the case of eyeblink conditioning, the tradeoff may comprise of keeping the eyes open to sample visual information, versus protecting the cornea from potentially harming stimuli. NOI, by adjusting the relative weight of the internal feedback, can therefore set a balance between anticipation and reaction such that the overall cost is minimized [76]. In other words, cerebellar learning is not minimizing error but rather minimizing cost, including the cost of not sampling the external environment.

Mixed feed-forward and feedback control is considered the most robust control strategy, combining the efficiency of anticipatory action with the inherent robustness of feedback control. As we just observed, in nature, a balance between both modes of control can be found in the conditioning of the eye blink response and its realization in the cerebellum. Understanding this phenomenon requires complete consideration of the systems-level interactions between the cerebellar adaptive layer, the reactive brain stem motor nuclei, and the controlled plant (eyelid/nictitating membrane). Here, the use of a systems-level modelling approach has generated the testable prediction that the NOI is the critical interface regulating the coupling of cerebellum enabled predictive feed-forward control and reactive control achieved by brainstem feedback controllers. At this point, the question arises of how the NOI is itself regulated to achieve an optimal balance between the reactive and adaptive modes of control. The substantia nigra has been shown to send dopaminergic projections to the nucleo-olivary pathway [77], whose functional relevance is not yet understood. We propose that through this dopaminergic neuromodulatory output, the basal ganglia may regulate activity in the nucleo-olivary pathway and consequently control contextually the recruitment of the cerebellum. In the case of eyeblink conditioning, such control may result in a modulation of the learning rate and of the amplitude of the CR at learning asymptote. Indeed, IO cells have been shown to reproduce, in eyeblink conditioning, the encoding of temporal-difference prediction errors already found in dopaminergic cells of the basal ganglia [78], indicating an at least correlational link between dopaminergic and inferior olivary activity. In the broader context of this consensus paper, we propose that the NOI realizes a mechanism that allows to contextually adjust the contribution of distinct cortico-cerebellar loops to cortical computations and that such regulation might depend on neuromodulatory output from the basal ganglia.

The NOI pathway is a basic element of the cerebellar circuit with pervasive projections of climbing fibers throughout the cerebellar cortex. At this point, it is attractive to speculate that the possibility of regulating the degree of recruitment of the cerebellum by contextually adjusting the level of NOI would augment the computational power and versatility of brain networks that include the cerebellum, allowing them to shift contextually between more certain (prediction-based) and/or more uncertain (reactive-based) modes of control. It appears likely that this would be a more pressing control problem when we consider higher brain structures that are interfaced to the cerebellum. Indeed, given that the balancing of feedback and feed-forward control is essential for adaptive behavior, an imbalance of the two might account for pathologies such as Parkinson’s disease dyskinesia. More concretely, we can hypothesize that the over-activation of the cerebellar cortex discussed by Popa and Kishore below could result from a too strong NOI that by silencing the inferior olive increases the tonic level of activity in Purkinje cells [79]. This points to the NOI pathway as a possible target for studies with animal models of Parkinson’s disease.

Cerebro-cerebellar communication is often implicitly assumed to refer to the cortico-pontine system and the projections back from the cerebellum to neocortex via the thalamus. However, a substantial part of the interplay between the neocortex and the cerebellum is likely to occur via the spinocerebellar systems. If the basal ganglia initiate the release of motor programs or motor command signals via the neocortex, corticospinal axons will inevitably activate large parts of the circuitry in the spinal cord. As many neuronal components in the spinal circuitry project to the cerebellum, in addition to alpha-motor neurons or spinal motor pools [87], the spinocerebellar pathways are an important source of information about the neocortical activity [80, 87].

The multitude of spinocerebellar pathways [80, 81] is of central importance to the coordination of limb movements [82]. The spinocerebellar and spino-reticulo-cerebellar pathways represent information from spinal motor circuits that are composed of various spinal interneurons. Almost all corticospinal tract axons terminate within the pool of spinal interneurons rather than the alpha-motor neurons directly [83, 84], which puts the interneurons in a key position for the majority of the motor control functions exerted by the brain. This pool of interneurons integrates the motor command signal from the neocortex with local sensory feedback, such as cutaneous information, tendon organ afferents, and muscle spindle afferents both of type I and type II. Spinal interneurons can either have direct recurrent connections that ascend all the way to the cerebellum [85] or the lateral reticular nucleus [86, 87], or they can utilize specialized ascending neurons such as the neurons of Clarke’s column [80]. These systems display differences regarding their detailed connectivity and the information they sample within the spinal cord. Consequently, the ascending projections from these systems will serve the overall function of providing the cerebellum with a wide monitoring of the activity in spinal motor circuits. Spinal interneurons are involved in the muscle synergy selection when the spinal motor circuitry is driven by descending motor control signals from the neocortex and/or subcortical motor systems [88]. Hence, the cerebellum can use the spinocerebellar systems to be informed about the relative excitatory drive on specific synergical components and on the low level motor functions resident in the spinal cord [89, 90].

In many respects, the intrinsic processing within the cerebellar cortex can to a large extent considered to be solved [91]. Mossy fiber information is transmitted through the granule cell layer and reaches the molecular layer, where the signal is given a specific synaptic weight in the Purkinje cell through learning. Due to the presence of local inhibitory interneurons, this synaptic weight can also have a negative value on the Purkinje cell. Hence, observed excitatory and inhibitory inputs to these cells in vivo can to a large extent be explained by learning [92], which is related to the climbing fiber signal the Purkinje cell receives. In light of this relatively simple circuitry function, a natural consequence is that the key function of the cerebellar cortex becomes that of a major associative element. The individual Purkinje cell has in itself massive associative power, due to the very high number of parallel fiber synapses and interneuron synapses it receives. An array of Purkinje cells that compose a functional microzone, and which targets the same group of cells in the efferent cerebellar nucleus and therefore can be considered a “super”-Purkinje cell, has a tremendous associative power [91]. One important possible consequence of this arrangement is that the cerebellum can associate many specific basal (spinal) synergy components in the right temporal order to help synthesize compound movements [89]. Although there is a debate on the pattern of convergence of mossy fiber information in individual granule cells, at least for limb controlling areas of the cerebellar cortex mossy fiber information from individual functional pathways is directly transmitted through granule cells and reflected in the output of the Purkinje cells [92, 93]. In this case, there is an unconditional transmission of information through the granule layer from individual functional pathways, which means that the information carried by these pathways can be integrated by the Purkinje cell without being dependent on concomitant input from other mossy fiber systems in contrast to what was assumed in the original Marr-Albus theory of cerebellar information processing [94].

However, there are naturally many details left to solve for the internal processing in the cerebellar cortex. One example is how the global balance of activity in the perpetually active cerebellar circuitry is maintained, where the feedback loops formed by the inhibitory Golgi cells, the nucleo-olivary inhibitory pathway, and the control of the overall Purkinje cell firing level from climbing fibers are likely to be important [95]. Another example is the role of the various neuromodulators, of which there is ample representation in the cerebellum. But with respect to the direct fast processing, which at least under limb movement control is the main functional contribution of the cerebellum, the structure and physiology of the cerebellar-cortical circuitry do not seem to provide a substrate for adding substantial computations of highly advanced functions on the information it receives. But it does allow the coupling/association of a very large amount of mossy fiber information, which in itself can represent advanced functions, as described above. In the case of spinocerebellar systems, such functions can already be represented in the information that reaches the cerebellum, and the task of the cerebellum becomes that of finding which of these many functions should be associated with each other, and under which context they should be associated.

Based on the anatomical and physiological evidence available by the end of the last century, I proposed a theoretical framework that the cerebellum, the basal ganglia, and the cerebral cortex are specialized for supervised, reinforcement, and unsupervised learning, respectively [41, 111]. This view has received some support and raised some debate among theoreticians and experimentalists.

An important combination of supervised learning and reinforcement learning is model-based action selection and learning. Supervised learning in the cerebellum, including the cerebellar cortex and nuclei, can create a “forward model” that predicts the results of executed actions [64]. Reinforcement learning in the basal ganglia can provide a “state value function” that evaluates the goodness of states. Because the dominant view at that time was that the outputs of the cerebellum and the basal ganglia form separate channels in the thalamus [6], I hypothesized that model-based action selection can be implemented by a cortico-cerebellar loop predicting the result of an imaginary action, the cortico-basal ganglia loop for evaluating its goodness, and an activation of dopamine neurons for triggering execution of the imagined action (Fig. 7 of [111], reproduced here as Fig. 6). A more specific proposal was that the striosome (or patch) compartment, which projects to dopamine neurons, learns state value function while the matrix compartment, which projects to the pallidum, learns state-action value functions [41, 99].

Open image in new window

In the last decade, there have been significant new observations regarding the interaction between the cerebellum and the basal ganglia. The cerebellar output is disynaptically connected to the basal ganglia through dentate nucleus–centrolateral thalamus–striatum pathway [12]. It was recently demonstrated that stimulation of dentate neurons evokes short-latency (about 10 ms) responses in about half of the striatal neurons [52]. It was further revealed that the basal ganglia output is also disynaptically connected to the cerebellum through subthalamic nucleus–pontine nucleus–cerebellar cortex pathway [13]. Thus, the existence of the short-cut connections between the cerebellum and the basal ganglia is becoming a new anatomical consensus [3], but what are their computational roles?

A possible role of the dentato-thalamo-striatal pathway is model-based evaluation of action candidates. The forward model learned in the lateral cerebellum predicts the resulting sensory state for a candidate action and its output is sent through the thalamus to the striatum, which estimates the value of the predicted state. An interesting observation [52] is that optogenetic stimulation of the dentate-thalamo-striatal pathway reverts cortico-striatal long-term depression (LTD) to LTP, which had been reported to happen with dopamine stimulation [114]. Previous self-stimulation experiments showed that the striosome compartment is more effective in inducing self-stimulation than the matrix compartment [115]. It is not known whether the dentato-thalamo-striatal pathway has any preference projections, but a possible scenario is that the cerebellar input activates the striosome neurons that activate dopamine neurons.

The function of the subthalamo-ponto-cerebellar pathway is harder to interpret. The subthalamic nucleus is a part of the “indirect pathway” of the basal ganglia that is implicated in action inhibition and aversive learning [116, 117]. Thus, a possible function of the pathway is to provide a “Stop!” signal to the cerebellum for withholding ongoing movements. Another possibility, though highly speculative, is to signal “off-line” status to the cerebellum, notifying that currently the basal ganglia have withheld execution of motor programs (though inhibition of the thalamus and midbrain motor nuclei via globus pallidus and substantia nigra reticulata, respectively) so that the cerebellar internal models can be safely used for off-line mental simulation.

Here, I presented some views based on the hypothesis that the cerebellum provides internal models of the body and the environmental dynamics for model-based motor control and decisions [41, 111]. There are, however, proposals and accumulating evidence suggesting that the frontal and parietal cortices as well as the hippocampus play important roles in model-based planning and decisions [118, 119, 120]. A possible difference of the internal models provided by the cerebellum and the cortex may be that the cerebellum learns deterministic input–output mapping with its mostly feed-forward circuit, while the cortex learns joint or conditional probability models through iterative dynamics using its heavily recurrent circuit. Probabilistic models are more general, as it includes deterministic models as extreme cases, but deterministic models have the virtue of simplicity. The hippocampus may enable reuse and editing of episodic memories for planning about the future. These ideas are all speculative, but recent optogenetic tools for pathway-specific circuit manipulations can make it possible to test these hypotheses. The big challenge for us theoreticians is to provide hypotheses worthy of laborious testing.

Over several decades, a substantial body of work has investigated the role of the cerebellum and its connections with other brain areas in human motor control and learning [34]. Non-invasive brain stimulation techniques such as TMS and tDCS have been used to probe the functional importance of cerebellar connectivity in motor control/learning. In general, tDCS has been used to locally modulate cerebellar excitability [156], whereas TMS has been applied to either locally disrupt cerebellar processes [127] or to evaluate cerebellar-primary motor cortex (M1) inhibitory connectivity (cerebellar brain inhibition (CBI)) [157]. CBI takes advantage of the cerebellum’s inhibitory connections with M1, applying a conditioning pulse of TMS to the cerebellum prior to stimulating the M1. The following section will provide an overview of what we have learnt using non-invasive brain stimulation regarding the role of the cerebellum and its connections with other brain regions in motor control/learning. We will focus on studies of upper limb reaching movements and highlight the issues still to be resolved.

Cerebellum, basal ganglia, and motor areas in the frontal cortex are the major nodes in the motor network, and they have unique cytoarchitectural properties, reciprocal connectivity, and synaptic plasticity mechanisms that allow efficient communication among them (Fig. 8). The discovery of reciprocal, topography-specific, subcortical connections between cerebellum and basal ganglia in primates [12] indicates that the basal ganglia and cerebellar motor networks are not independent information processing systems but can exchange information between them in real-time to provide appropriate inputs to the frontal cortices. In rats, under physiological conditions, the cerebellum modulates cortico-striatal plasticity through these connections [52], and the same pathway can also transmit aberrant cerebellar activity to basal ganglia, generating dystonia and dyskinesias in animals. Cerebellum also modulates cortical sensorimotor integration [167, 168] and is important for acquiring and maintaining [169] new cortico-spinal integrations. In young healthy humans, non-invasive stimulation studies demonstrated that cerebellum exerts a bidirectional modulation only on heterosynaptic motor cortex (M1) plasticity [170], which is dependent on peripheral sensory information reaching the M1 through complex polysynaptic pathways [171]. The cortical heterosynaptic plasticity can be suppressed by excitation of the cerebellar cortex or amplified, prolonged, and rendered less topography-specific by inhibition of the cerebellar cortex [170]. These findings could be an indirect evidence of the fine-tuning mechanism in place in the cerebello-cortical network to rapidly adapt the motor commands to new environmental challenges during movement planning. If so, excitation of the cerebellar cortex would inhibit the ipsilateral dentate output and thereby the sensory relay to M1, thus leaving M1 less receptive to any new motor programs. In contrast, the inhibition of cerebellar cortex would disinhibit the dentate output, which would facilitate the sensory relay to M1, permitting the selection of new motor programs [170]. From this perspective, such preparations of continuously anticipating and pre-planning motor programs for simple movements through cerebellar modulation would assist the efficient execution of complex movements. Pathologies, in which cerebellar bidirectional adjustment of M1 plasticity is impaired, can therefore be expected to affect motor performance and motor learning that need fine online adjustments.

Open image in new window

The abnormalities in M1 plasticity in patients with disorders such as primary dystonia and Parkinson’s disease (PD) are generally attributed to abnormal basal ganglia input to the M1. Non-invasive studies testing the efficiency of cerebellar modulation of cortical excitability were recently reported in patients with writer’s cramp [172, 173]. Though the ongoing M1 plasticity was not different from age-matched healthy volunteers, both excitation and inhibition of the cerebellar cortex were ineffective in modulating the heterosynaptic cortical plasticity in patients with writer’s cramp. Also, the extent of modulation of motor cortex plasticity by cerebellar inhibition had an impact on the efficiency of online adaptation in a sensorimotor adaptation task [172]. This deficiency might lead to degradation, with time, of the specific, vulnerable, motor programs for tasks that require reinforcement by online sensorimotor adaptation such as writing. However, clinically relevant improvement in dystonia was not observed following a single session of artificial cerebellar excitation [173] or inhibition in patients with writer’s cramp [174]. This is not surprising since abnormal motor programs for writing must have been reinforced over years, before becoming manifest and it might require many cerebellar stimulation sessions to reverse this process [175].

Studies of plasticity changes in M1 in early PD report variable results. The results are more consistent regarding the loss of bidirectional plasticity of M1 in advanced PD patients with motor complications of levodopa treatment [176, 177, 178]. There is indirect evidence that cerebellar modulation of M1 plasticity is impaired in advanced PD patients with levodopa-induced dyskinesias. Repeated sessions of cerebellar inhibition could enhance the M1 heterosynaptic plasticity [179] and concurrently reduce the severity of dyskinesias [179, 180]: the greater the enhancement of M1 plasticity following cerebellar inhibition, the greater was dyskinesias reduction. This raises an interesting question whether there is a state of deficient cerebellar inhibitory modulation or excessive cerebellar excitation in advanced PD.

Whether altered balance between the activity in the basal ganglia and cerebellum in advanced Parkinson’s disease is caused by the primary pathology (i.e., dopamine depletion in the basal ganglia) or by a maladaptive compensation (e.g., increased activity in the cerebellum) remains an important point to be clarified in the future, which might lead to new therapeutic approaches, including new surgical targets. At this moment, both hypotheses are equally plausible and not mutually exclusive: such a state of imbalance in the cerebellum could affect basal ganglia functions and generate dyskinesias just as in the rat model [167], or the aberrant basal ganglia activity in advanced PD could affect the bidirectional cerebellar modulation of M1 plasticity [179]. Based on the indirect evidence of cerebellar involvement in advanced PD and the state of cerebellar hyperexcitation in the levodopa-naïve, MPTP-treated primates [181], a model was recently proposed [182] that considers cerebellar, basal ganglia, and motor cortical networks as interlinked and their abnormal interactions as central to the generation of abnormal movements in both parkinsonism and levodopa-induced dyskinesias (Fig. 9). This model needs further validation in animal models of PD and levodopa-induced dyskinesias.

Open image in new window

Another important detail concerning this heavily interdependent circuit is the increased activity in the cerebellar cortex that inhibits the cerebellar output from the nuclei, which, in turn, project not only to the thalamic relays but also to key hubs in the brainstem, like the inferior olive. However, most of the spontaneous activity in the cerebellar cortex is driven by intrinsic (non-synaptic) mechanisms and the level of spontaneous activity in Purkinje cells is reciprocally regulated by the activity of the inferior olive (through the nucleo-olivary inhibition). Any regulation of the strength of this negative link remotely from other brain structures or artificially (pharmacologically or with invasive-non-invasive stimulation) could be a means to adjust the level of spontaneous activity in the cerebellar cortex and subsequently in the whole network.

It is commonly acknowledged that resting tremor in PD has a central origin, but the localization of the central oscillator (or oscillators) is still debated. There is converging evidence from metabolic imaging, electrophysiology, and deep brain stimulation studies that both the basal ganglia and the cerebello-thalamo-cortical circuit are causally involved in PD resting tremor [36]. Recently, Helmich and colleagues have proposed the “dimmer-switch” hypothesis to explain the different roles played by these two circuits in PD resting tremor [36, 183]. According to this systems-level perspective, the basal ganglia may work as a light switch triggering tremor-related responses in the cerebello-thalamo-cortical circuit that, in turn, generates the tremor modulating its amplitude like a light dimmer. This view is in line with previous theoretical proposals highlighting the roles of basal ganglia for triggering movement initiation and of cerebellum for movement amplification of motor patterns [42]. These physiological roles arise from the unique architecture of each circuit. Specifically, the balance between activity in the direct and indirect pathways in the basal ganglia enables the GPi to select one action while inhibiting others [184, 185]. Furthermore, the strong rhythmic (6–10 Hz) properties of the cerebello-thalamo-cortical circuit—which have been related to motor timing [186, 187]—may enable this circuit to estimate limb position in relation to the movement goal [41]. This suggests that existing functional modules—involved in voluntary action selection in healthy subjects—are driven into producing involuntary actions (resting tremor) in Parkinson’s disease, but it is not clear why this happens. Functional changes in the neurotransmitter systems projecting to these circuits may play a role: alterations in the dopaminergic, serotonergic, and noradrenergic systems have all been associated with PD resting tremor [183, 188, 189]. The dimmer-switch hypothesis also does not specify “how” the basal ganglia interact with the cerebello-thalamo-cortical network. Here we discuss four possible pathways that may play a role in this (Fig. 10), and we discuss research strategies to distinguish between these possibilities. Finally, we highlight how a systems-level view on PD may lead to the development of new treatments.

Open image in new window

Studies in non-human primates have shown the presence of several anatomical pathways between the basal ganglia and the cerebello-thalamo-cortical circuit. First, the STN has both afferent and efferent connections with the cerebello-thalamo-cortical circuit: it receives direct anatomical projections from the primary motor cortex and sends disynaptic projections to the cerebellar cortex through the pontine nuclei [4, 13]. In PD, the activity of excitatory/glutamatergic output neurons of the STN is increased [190], and this may trigger the cerebellar cortex via the pontine nuclei, as these projections are largely glutamergic [13]. In this way, the STN may work as a switch that triggers, through the subthalamo-ponto-cerebellar pathway, abnormal activity in the cerebello-thalamo-cortical circuit. Although this hypothesis explains why DBS of the STN and ventral intermediate nucleus (VIM) are very effective in treating resting tremor [191, 192], it diminishes the role of the internal globus pallidus (GPi) in the generation of tremor. This does not explain the presence of tremor-related activity in the GPi [183, 193] and the fact that DBS targeting the GPi effectively reduces tremor [194]. Second, the basal ganglia and the cerebello-thalamo-cortical circuit may interact at the level of the motor cortex, where both circuits anatomically converge [195]. Specifically, the GPi sends GABA-ergic projections to the anterior part of the ventrolateral thalamus (VLa) which in turn sends excitatory efferents to the motor cortex that projects to the posterior part of the ventrolateral thalamus (VLp, also termed VIM), which receives cerebellar projections [196]. This hypothesis would explain why DBS of the STN, GPi, and VIM are all effective in treating tremor [191, 192, 194]. Third, the basal ganglia may engage the cerebello-thalamo-cortical circuit also through the zona incerta, which receives projections from the GPi and sends GABA-ergic projections to the VLp [197]. Finally, the basal ganglia may interact with the cerebello-thalamo-cortical circuit through the peripheral nervous system. That is, after tremor initiation by the basal ganglia, tremor-related somatosensory afferents transmitted to the cerebellum may engage the cerebello-thalamo-cortical circuit [198]. However, this hypothesis does not fit with data showing that peripheral deafferentation, peripheral anesthesia of tremulous muscles, and mechanical perturbations have little effect on PD tremor (for a review, see [36]). To conclude, there is currently no definite evidence which of the pathways connecting the basal ganglia with the cerebello-thalamo-cortical circuit contribute to PD tremor.

Given the complex involvement of different neural circuits, a systems-level perspective on PD tremor is required to answer this question. Systems-level questions require systems-level outcome measures, and the development of new analysis tools has made this possible. For example, rather than testing for altered activity in a set of brain regions, or altered functional connectivity between two different brain regions, dynamic causal modelling (DCM) allows one to statistically compare different network configurations based on fMRI data. A second approach could rely on computational modelling [39, 199] that has the advantage to allow one to directly measure the effects of adding or removing specific pathways on the oscillatory output of the system. Finally, animal models of tremulous PD may be helpful to test whether the resection of specific pathways interferes with tremor. For example, this approach has been applied to a mouse model of dystonia where resection of the anatomical pathway between cerebellum and basal ganglia was shown to reduce dystonic symptoms [200].

A systems-level view of tremor may improve existing treatments, for example by tailoring GPi-DBS to ongoing activity in the motor cortex [201]. It may also facilitate the development of new treatments. For example, computational models informed by functional imaging may allow us to test new anti-tremor treatments “in silico” before transferring them to animals or patients [199].