Cerebellar physiology: links between microcircuitry properties and sensorimotor functions

Abstract

Existing knowledge of the cerebellar microcircuitry structure and physiology allows a rather detailed description of what it in itself can and cannot do. Combined with a known mapping of different cerebellar regions to afferent systems and motor output target structures, there are several constraints that can be used to describe how specific components of the cerebellar microcircuitry may work during sensorimotor control. In fact, as described in this review, the major factor that hampers further progress in understanding cerebellar function is the limited insights into the circuitry‐level function of the targeted motor output systems and the nature of the information in the mossy fiber afferents. The cerebellar circuitry in itself is here summarized as a gigantic associative memory element, primarily consisting of the parallel fiber synapses, whereas most other circuitry components, including the climbing fiber system, primarily has the role of maintaining activity balance in the intracerebellar and extracerebellar circuitry. The review explores the consistency of this novel interpretational framework with multiple diverse observations at the synaptic and microcircuitry level within the cerebellum.

Abbreviations

DCN

deep cerebellar nuclei

DSCT

dorsal spinocerebellar tract

LTD

long‐term depression

LTP

long‐term potentiation

MLI

molecular layer interneuron

PC

Purkinje cell

SCT

spinocerebellar tract

SRCT

spino‐reticulocerebellar tract

Introduction

Over the last 10–15 years, several foundations for how we understand the function of the cerebellum have changed. Apart from an overall scientific progress, in particular the capacity in several labs to record from the smaller neuron types in the cerebellar cortex in vivo, the birth of new neuroscience methods to explore the cerebellum and the development of theories of global network function have contributed to these advances. In the cerebellum, as in other brain structures, function is determined both by the physiology of the cellular and synaptic elements and the structure of the neuronal network, where the latter is being shaped as a result of another neurophysiological process, synaptic plasticity or learning. Therefore, these two aspects will be the focus of this review. The review will also have a bias towards the somatic sensorimotor functions and less focus on eye and neck motor functions and on higher order neocortical (‘mental’) functions. It is assumed, however, that many principles described here will apply also to the latter. The review will start with a brief résumé of the connectivity structure at the macro‐ and microcircuitry level of the cerebellum, present some recent insights in the details of the connectivity structure and physiology of the microcircuitry and then mainly focus on cerebellar physiology in a wider functional sense, including the constraints for cerebellar function imposed by the systems that the cerebellum is connected with. The motor control of the brain is assumed to be deeply intertwined with the massive sensory feedback with which it is provided, and neural representations of sensorimotor functions at different levels are the main platforms from which this discussion takes its origin.

Before considering the physiology of the cerebellum at a more detailed level it is important to be clear about the mode of coding that is assumed. In contrast to the original formulations in the classical ideas of Marr–Albus, and many alternative theories, it is assumed here that sparse coding, and other forms of coding that depend on low firing rates and isolated spikes generated with millisecond precision, do not apply. A long range of theoretical arguments and experimental observations support the idea that the cerebellum and its afferent mossy fibre pathways instead operate using a coding mode that is more reminiscent of dense code or high‐resolution, time‐modulated rate code. These issues have been extensively covered in a recent theoretical paper (Spanne & Jörntell, 2015), but some of the main arguments will be restated later in this review. Changing the assumption with respect to the coding away from sparse and spike‐time dependent coding has fundamental consequences for the functionality in the cerebellar neuronal circuitry, not least with respect to the crucial interplay between the Purkinje cells and the neurons of the cerebellar nuclei. Similarly, cerebellar contributions to ongoing movements can be explained by simply considering high‐resolution, time‐modulated rate code outputs in a population of Purkinje cells making functionally specific connections to the cerebellar nuclear neurons (Herzfeld et al. 2015). The behavioural control by the cerebellar cortex can hence be explained without considering synchrony in the activation of the Purkinje cells or climbing fibres, and potential effects of neural synchronization are therefore not considered here.

Anatomical macro‐organization helps understanding of regional function

Beneath the largely regularly foliated surface of the cerebellum hides an intricate but repeatable neural connectivity pattern. The most useful feature with which to understand the macro‐organization of the cerebellar neuronal network has proven to be the external connections of the cortex, especially the climbing fibre projections from the inferior olive to the cerebellar cortex and the Purkinje cell projections from the cerebellar cortex to the cerebellar and vestibular nuclei. The sagittal zonal divisions of the cortex that are defined by the anatomical organization in these two connections, often referred to as the olivocerebellar and the corticonuclear connectivity, have been reviewed many times, for many species (Oscarsson, 1973, 1979; Armstrong, 1974; Voogd & Glickstein, 1998; Apps & Garwicz, 2005), and will not be reviewed in detail here. But from a physiological point of view it is interesting to note that the olivocerebellar and the corticonuclear connections may be examples of genetically pre‐programmed connectivity (Altman & Bayer, 1997; Leto et al. 2015), which draws up the main line of circuitry organization before other connections grow in and distribute themselves according to the framework established. For example, the distribution pattern of mossy fibres is closely related to that of the climbing fibres (Pijpers et al. 2006; Ruigrok, 2011; Reeber et al. 2013), and experimental scrambling of the topographical climbing fibre organization in the cortex is paralleled by a scrambling in the topography of the ingrowing mossy fibres (Reeber et al. 2013). The zonal organization is likely to be a universal anatomical‐organizational feature of the cerebellum across many species (Voogd & Glickstein, 1998).

The zonal connectivity structure defines a functional structure. This was first described for the vermis and intermediate part of the cerebellar anterior lobe, where a number of sagittal zones could be defined solely on the basis of by which spino‐olivocerebellar pathways the climbing fibres received their peripheral input (Oscarsson, 1973; Andersson & Eriksson, 1981). The specificity in the corticonuclear connectivity (Voogd & Glickstein, 1998; Apps & Garwicz, 2005), where individual zones target specific cerebellar or vestibular nuclei (Fig. 1), also provides for a functional specificity of the output of the zones. This is because specific nuclei target specific motor systems in the brainstem (Ito, 1984), and to some extent specific regions of the cerebral cortex via the thalamus (Jörntell & Ekerot, 1999; Bostan et al. 2013). Hence, although the general network structure within the cortex follows the same basic plan across the cerebellum, the zonal regions are likely to be primarily targeting specific motor output systems in the brainstem. Regional localization then becomes important as it can be expected that the details of the connectivity within the cortex are locally adapted to the specific control function of the motor output system to which the zone is connected.

Figure 1. Schematic illustration of the cerebellar sagittal zonation and efferent connections.

Upper panel, 3‐D image of the anterior lobe of the cat cerebellum, with zones in the vermis and the pars intermedia on one side of the midline indicated. For clarity, the X and Y zones (Ekerot & Larson, 1982) are omitted from the display. DN indicates that there may be multiple D zones not yet identified in the hemispheres of higher animals such as primates. Lower panel, main output connections of the cerebellar sagittal zones. The connections from the cortex to the nuclei are well organized (Voogd & Bigaré, 1980; Trott & Armstrong, 1987 a,b). For illustration purposes, only the main connections of the output nuclei are indicated although some overlap in the targeted motor structures may occur between them. All cerebellar nuclei also target the neocortex via thalamic innervation. Connectivity based on Hirai et al. (1982), Ito (1984), Gibson et al. (1987), Zwergal et al. (2009) and Bostan et al. (2013). Targeted motor structures in turn overlap in their innervation of the spinal interneuronal motor circuitry (Jankowska, 1992; Bortoff & Strick, 1993; Hultborn, 2001; Hammar et al. 2011; Jankowska et al. 2011 a; Shrestha et al. 2012; Spanne & Jörntell, 2013).

Whereas the macro‐organization described above mostly originates from work made decades ago, the discovery that the expression of different molecular markers such as aldolase C and excitatory amino acid transporter 4 (EAAT₄) are topographically congruent with the sagittal zonal structure (Sugihara & Shinoda, 2004; Apps & Hawkes, 2009; Ruigrok, 2011; Reeber et al. 2013) suggest that a very fine molecular map parallels the zonal functional structure. Under different names, such as zebrins, these molecular markers are currently emerging as a widely used tool for reference to localization. As localization within the topographical structure has a functional relevance (Fig. 1), it is a useful tool for the physiologist.

The cerebellar cortical network as an adaptive association layer

Most of the cerebellar information processing is carried out by the main pathways of the cortical circuitry, which were established almost 50 years ago (Eccles et al. 1967) and mostly modified in the details over the years that have passed since. They have been described numerous times (Ito, 1984, 2006; Dean et al. 2010), and will only be briefly summarized here. The main flow of information in the cerebellar cortex is between the mossy fibres and the Purkinje cells, where the mossy fibres make excitatory synapses with granule cells, which in turn issue parallel fibres that make excitatory synapses on the Purkinje cells. The mossy fibre–granule cell information can also reach the Purkinje cells as an inhibitory signal by means of the parallel fibre synapses on the molecular layer interneurons (MLIs; stellate and basket cells; Jörntell et al. 2010), which in turn make inhibitory synapses on the Purkinje cells. At least in regions involved in somatic sensorimotor processing, these four types of neural elements can be expected to carry out almost all of the information processing that the cerebellar cortex is responsible for during ongoing motor control, whereas in other regions the unipolar brush cells may also be expected to contribute (Dino et al. 1999).

Some recent studies have examined the properties of the cortical intrinsic connections at a higher level of detail than previously possible. A long‐standing enigma has been the physiology of the basket formation around the Purkinje cell bodies, formed by the axon terminals of the basket cells but also the stellate cells (Jörntell et al. 2010). It seems clear that this formation does not form a regular inhibitory synapse. Instead, it may rely on ‘ephaptic’ transmission (Blot & Barbour, 2014), i.e. transmission of the electrical field generated by the spike between closely apposed neuronal membranes. How this non‐regular form of neuronal communication influences the mode of MLI inhibition of Purkinje cells as a whole is not clear. However, since the transmission of inhibitory inputs to Purkinje cells in vivo in general appears to be smooth and without rapid transients (Blazquez & Yakusheva, 2015), it seems at the moment reasonable to assume that it does not confer any functionally different effect compared to the overall inhibitory signalling from MLIs. Other advances include deepened insights into the rebound potentiation of the MLI synapses on the Purkinje cells (Tanaka et al. 2013) and a more detailed decomposition of the molecular components of this inhibitory connection (He et al. 2015).

A factor that has redefined how we understand the intrinsic connectivity of the cerebellar cortex are the findings of several forms of synaptic plasticity in addition to the classical long‐term depression (LTD) of the parallel fibre synapses on Purkinje cells. In particular the long‐term potentiation (LTP) of parallel fibre synapses on Purkinje cells and the LTP and LTD of parallel fibre synapses on the molecular layer interneurons have important consequences for the establishment of the intrinsic connectivity of the cerebellum (Dean et al. 2010). Whereas the firing behaviour of the individual neuron types follow certain patterns, these plasticity mechanisms can completely change the physiological ‘connectome’, i.e. the patterns of synaptic weights between neurons. Also other forms of plasticity in the cerebellar cortex were described in the past decade. As discussed elsewhere (Jörntell, 2016), the network position of these plasticity effects do not have the same potential for functional change as the parallel fibre synapses. They do, however, have a natural position for maintaining the balance of neuronal activity across the cerebellar network, as discussed later.

Cerebellar nuclear neurons: inhibitory cortical output integrated with excitatory drive

For all zones and functional regions of the cerebellum, the relationship between the mossy fibre–granule cell‐driven Purkinje cell inhibition of the neurons of the deep cerebellar nuclei (and relevant vestibular nuclei) and the mossy fibre‐driven excitation of the same neurons is a critical step towards understanding cerebellar function (Bengtsson & Jörntell, 2014). The Purkinje cells output a time‐varying inhibitory signal, defined by the cortical processing, to the deep cerebellar nuclei (DCN). The balance between this inhibition, the intrinsic spontaneous firing of the DCN neurons, and the time‐varying excitatory mossy fibre inputs to the DCN neurons determines the cerebellar output. As there are many Purkinje cells and many DCN neurons active under any movement, there will be a spatiotemporal pattern of Purkinje cell activation and a resulting spatiotemporal pattern of DCN neuron output. Whereas the temporal aspect of the Purkinje cell output pattern depends on the processing in the cortex, its spatial distribution follows not only the zonally specific topographical rule shown in Fig. 1, but also the finer microzonal organization within a zone is to some extent preserved in the connectivity from the Purkinje cells to the cerebellar nuclear neurons (Andersson & Oscarsson, 1978; Garwicz & Ekerot, 1994). By using artificial inputs that synchronize the climbing fibre activation of the Purkinje cells, this fine corticonuclear organization can, in the experimental situation, be used to identify the receptive field of the climbing fibre input to the Purkinje cells that inhibits a particular DCN cell (Andersson & Oscarsson, 1978; Bengtsson et al. 2011). This inhibitory climbing fibre receptive field, in turn, is a functional tag for the DCN cell as it is indicative of the type of movement it contributes to (Bengtsson & Jörntell, 2014). In this experimental setting, the external sources that drive the mossy fibre input to DCN cells show some degree of relationship to those driving the climbing fibre input (Bengtsson & Jörntell, 2014). Although a comparatively weak relationship, its presence suggests the involvement of synaptic plasticity also in shaping the mossy fibre to DCN cell connection, possibly during development, or also in adult life.

Whereas the above studies focused on the projection neurons of the DCN, the cerebellar nuclei also contain neurons that provide inhibitory feedback to the inferior olive (Andersson et al. 1988; Chaumont et al. 2013) as well as local inhibitory interneurons (Husson et al. 2014). The feedback control of the inferior olive by the nucleo‐olivary projection neurons provides a good substrate for maintaining the overall activity balance in the cerebellar cortex, as discussed below. The role of the local, glycinergic inhibitory interneurons, and the nucleo‐cortical glycinergic neurons that innervate primarily Golgi cells (Ankri et al. 2015), are less clear, although a role in regulating overall activity balance lies close at hand also for these neurons.

Mossy fibres carry most of the information used in cerebellar processing

Among the two afferent systems to the cerebellar cortex and the DCN, the mossy fibres exist in high numbers, typically have a wide dynamic range of firing (Van Kan et al. 1993; Garwicz et al. 1998; Jörntell & Ekerot, 2006; Arenz et al. 2008; Prsa et al. 2009) and, due to their extensive branching (Wu et al. 1999; Quy et al. 2011), each mossy fibre can potentially provide information to a very high number of Purkinje cells. Although the climbing fibres have a powerful effect on the Purkinje cell output, their firing during behaviour typically does not reach above 2 Hz, and since, for example, during a reaching movement they fire less than once per movement (Horn et al. 2004), they cannot carry a substantial amount of information for the guidance of such movements. Hence, it must be assumed that the mossy fibres are the main carriers of information about ongoing movements and movement commands to the cerebellum. Therefore to understand the cerebellum a fundamental step is to understand the nature of the information that is carried by the multitude of mossy fibre systems.

Mossy fibre information defines cerebellar microcircuitry function

What is the type of information that the cerebellum would need, and what type of information can the central nervous system provide? The answers to these questions require detailed insights into how the brain operates during movement, i.e. which sensorimotor circuitries are engaged and what particular functions these circuitries may have. In particular, one of the most difficult problems in the analysis of any brain area is to understand how information is represented in the system under study. It is often implicitly assumed that the mode of representation directly correlates with a particular set of arbitrary physical parameters such as force, acceleration or skin indentation, and that such parameters are coded for at the level of each individual neuron. However, brain operation may be organized using much more abstract representations of the physical world. Whereas physical parameters may or may not be explicitly represented in the peripheral sensors, any central neuron will integrate this information with internal brain representations. Hence, explicit neuronal representation of any arbitrary physical parameter cannot in general be expected to occur in central neurons, even though correlations with such parameters may always be possible to find. Instead, it is likely that any piece of information represented in the brain is distributed across a large set of neurons, and that each individual neuron carries only bits and pieces of that information (Herzfeld et al. 2015) at an abstracted level. Hence, the link between the physical world and the neural abstractions thereof needs to be understood.

Primarily since all subcortical sensorimotor systems and all parts of the neocortex seem to be engaged under movement, and therefore probably participate in the representation of relevant sensorimotor information, at least a number of basic constraints can be drawn. A first essential principle is that because of the structure of the nervous system, it is unlikely that a ‘pure’ motor command signal exists in the brain. This is because any motor command that results in muscle activation will inevitably also lead to sensory feedback (Fig. 2 A) that affects the activity of the motor output neurons at all levels of the brain. Our bodies are equipped with a very high‐density sheet of sensors distributed across skin, muscles, tendons and other connective tissue. These sensors are the filter through which the brain ‘sees’ the physical world. As even activation of a single muscle will result in changes in tendon forces, muscle lengths and skin strain patterns, it can be assumed that any kind of movement will be accompanied by a very rich set of sensor feedback patterns (Spanne & Jörntell, 2013). Importantly, these sensor feedback patterns impinge on spinal interneurons (Jankowska, 1992), which are also the main final pathway of motor commands generated by other structures of the brain (Santello et al. 2013). Hence, the final motor command, i.e. the final spatiotemporal structure of the activation of the α‐motoneurons and thereby the muscles, is a sum or a product of all the motor command signals issued and the pattern of sensory feedback that they are associated with. For the cerebellum to regulate ongoing movements, tapping off what is going on in the population of spinal interneurons is a very important source of information. This is precisely what the numerous spinocerebellar pathways focus on.

Figure 2. The sensorimotor circuitry of the brain and the spinocerebellar neurons.

A, overview of peripheral and spinal circuitry being engaged during voluntary movement commands. B, some main mossy fibre systems for somatic sensorimotor control. The connections from the spinal grey matter (outlined shape) and brainstem nuclei of the spinocerebellar and spino‐reticulocerebellar tracts/systems (SCT and SRCT, respectively) as well as the connections of the cuneocerebellar and pontocerebellar systems are illustrated. Connectivity based on Alstermark et al. (1981), Cheema et al. (1983), Kosinski et al. (1988), Ekerot (1990), Matsuyama & Drew (1997), Jankowska et al. (2011 a), Shrestha et al. (2012) and Geborek et al. (2013 a). VSCT, ventral spinocerebellar tract; DSCT, dorsal spinocerebellar tract; RSCT, rostral spinocerebellar tract; SRCT, spino‐reticulocerebellar tract; PCT, pontocerebellar tract; CCT, cuneocerebellar tract; Cun, main cuneate and external (accessory) cuneate nuclei; LRN, lateral reticular nucleus; Pons, pontine nuclei. ^* indicates that the VSCT pathway crosses the midline at the spinal segmental level and then recrosses in the brainstem to enter the cerebellum ipsilaterally, which is omitted in the illustration for clarity.

Spinocerebellar pathways have their origin in the spinal interneuron pool, and mediate combined sensory and motor information (Fig. 2 B). Some of the spinocerebellar tract (SCT) neurons appear to be specialized neurons that do not form part of the spinal interneuron circuitry but sample information from the circuitry and project directly to the cerebellum. An example of such an organization is the Clarke's column, containing neurons that form part of the dorsal spinocerebellar tract (DSCT). However, a more common principle appears to be that the neurons that are part of a spinocerebellar system are also directly participating in the spinal interneuronal circuitry, i.e. some of the spinal interneurons not only form part of a sensorimotor processing circuitry, they also send off axon branches that travel all the way up to the cerebellum. Many of the ventral spinocerebellar tract (VSCT) and rostral spinocerebellar tract (RSCT) neurons seem to be of this type. In fact, it is possible that even many DSCT neurons fall into this category – a recently described pool of DSCT neurons are located outside Clarke's column and have a location in the spinal grey matter that is no different from that of regular spinal interneurons (Shrestha et al. 2012). A final category in this group of carriers of sensorimotor information is the spino‐reticulocerebellar tract (SRCT). Similar to the spinocerebellar pathways, the SRCT is made up of spinal interneurons that project towards the cerebellum. However, rather than entering the cerebellum directly, this pathway has an additional synaptic relay in the lateral reticular nuclear neurons (Alstermark & Isa, 2012).

SCT and SRCT neurons show a range of temporal dynamics in their spike output pattern relative to an ongoing movement (Arshavsky et al. 1978 a,b; Fedirchuk et al. 2013). That is, as a population they can temporally signal the transitions between different phases of movements, where each individual neuron can signal one or several transitions. As they are driven in part by inputs from specific sensors, the relationship between their activation across different contexts or states (i.e. different types of movements/phases of movements) will follow certain patterns, as the activation of the sensors are determined by the local biomechanics. Hence, each SCT/SRCT neuron in principle describes a function between a component of the motor command (sampled from a variety of motor control structures; Oscarsson, 1973; Jankowska et al. 2011 a; Shrestha et al. 2012) and the resulting sensory feedback, and it can be expected to do so across all possible motor contexts (Spanne & Jörntell, 2013). But given the number of sensors in the body and the number of ‘upper’ motoneurons, there are so many possible functions that it is not reasonable to assume that they are all represented in the brain (Spanne & Jörntell, 2013). Instead, useful sensorimotor functions could be laid out in the spinal sensorimotor circuitry from early development, as an effect of early motor learning (Spanne & Jörntell, 2013). Hence, the cerebellum is, via the SCT/SRCT systems, provided with information about ongoing activity in specific sensorimotor functions and can, by specifically routing them through its internal network, primarily via plasticity in the molecular layer, link specific functions to each other in specific phase transitions and in specific states/motor contexts. In other words, if the incoming mossy fibre information is viewed as reflections of functions of sensorimotor relationships approximated in extracerebellar circuitry, then the pattern of connectivity within the cerebellar cortex and with the DCN cells determines which of these functions are associated with each other and in which patterns these functions will influence the neuronal networks of the extracerebellar structures from which the functions originate (Bengtsson & Jörntell, 2014).

Although the description above is focused on somatic motor control, and the specific support it has from the spinocerebellar systems, similar circuitry constructs for tapping off equivalent information also seem to be present for ocular motor control systems (Porrill et al. 2004). It seems likely that such ‘recurrent’ circuitry architecture exists also in cerebro‐cerebellar systems of the cerebellar hemispheres and the dentate nucleus (Kelly & Strick, 2003; Bostan et al. 2013) where they could be used for linking mental representations into learnt ‘spatio’‐temporal patterns of neural activation. Such mental representations could, for example, be internal models of the relationship between the self and the external world (Kawato et al. 2003; Ito, 2008), which would be useful for motor planning. Conceivably, this could form a substrate for linking thought processes and mental associations to each other, although there is a long way to go before we have a chance to understand the spatiotemporal structure of the neural representation of a thought. In this context, recent evidence indicating that the dentate nucleus also projects, via the thalamus, to the basal ganglia (Bostan et al. 2013; Chen et al. 2014), is also likely to play an important role although that role is at present unclear.

Climbing fibres regulate overall neuronal activity and synaptic plasticity

In relation to the description of the circuitry structure of the SCT/SRCT mossy fibre systems, it is interesting to first note that the spinal contribution to the input to the various parts of the inferior olive originates from the same spinal circuitry components (Fig. 3). In fact, it is even conceivable that they originate from the same spinal neurons. Hence, for the parts of the cerebellum concerned with somatic movement control, the mossy fibre input and the input to the inferior olive originate from functionally similar sources of information (Figs 2 B and 3), which are expected to be extensively engaged by any type of somatic movement (Prut & Fetz, 1999). But what does the climbing fibre system do for the function of the cerebellum?

Figure 3. The spino‐olivocerebellar pathways for activation of climbing fibres.

The connections from spino‐olivocerebellar pathways (SOCPs) to the climbing fibres of the cerebellar cortex, via specific parts of the inferior olive, are illustrated. VF, ventral funiculus (for the VF‐SOCP); LF, lateral funiculus; DLF, dorsolateral funiculus; DF, dorsal funiculus; Cun, main cuneate nucleus; cMAO, caudal medial accessory olive; cDAO, caudal dorsal accessory olive; rMAO, rostral medial accessory olive; rDAO, rostral dorsal accessory olive; PO, principal olive. Connectivity based on Oscarsson (1969), Armstrong et al. (1973), Oscarsson & Sjölund (1977), Ekerot & Larson (1979), Andersson & Eriksson (1981), Voogd & Glickstein (1998), Shrestha et al. (2012), Flavell et al. (2014) and Koutsikou et al. (2015). Cerebellar cortical zones are indicated as in Fig. 1, with D_N indicating that there may be multiple D zones not yet identified in the hemispheres of higher animals such as primates, which have different specific functions/targets in regulating the neocortex (Bostan et al. 2013). Note that, similar to the brainstem mossy fibre pathways in Fig. 2 B, the inferior olive also receives direct motor command signals (McCurdy et al. 1992), which are omitted for clarity. Question marks indicate that some of the SOCPs pass through synaptic relays in the brainstem that are yet to be identified.

The climbing fibre synapse on the Purkinje cell had already been identified as a uniquely powerful synapse of the central nervous system 50 years ago (Eccles et al. 1967). It was assumed that the activation of such a powerful synapse was a teaching or an error signal informing the Purkinje cells when they needed to re‐learn their responses to other (parallel fibre) input occurring at the same time (Marr, 1969; Albus, 1971). Activation of the climbing fibre synapse was originally thought to lead to long‐term potentiation (LTP) of the simultaneously activated parallel fibre synapses, but it was soon discovered that it instead was long‐term depression (LTD) that was induced (Ito et al. 1982; Ito, 1984). Nevertheless, the overall strong agreement between the theoretical considerations and the novel experimental data led to this form of cerebellar LTD being considered synonymous with cerebellar learning since then. Therefore, it has had profound influences on our concepts of cerebellar function.

Cerebellar LTD is a powerful phenomenon, easily demonstrable in experimental preparations using localized activation of parallel fibres or local chemical activation of AMPA receptors (Ito et al. 1982; Launey et al. 2004). Indeed, in the adult state, most parallel fibre synapses are silent, i.e. they have close to zero synaptic efficacy (Ekerot & Jörntell, 2001; Isope & Barbour, 2002; Jörntell & Ekerot, 2002, 2003). However, in the long run, any system using synaptic plasticity as the learning mechanism needs to have the means to change synaptic weights in both directions. By gradually modifying the synaptic weights, the connectivity of the system changes and the response in a given context can be gradually improved (Spanne & Jörntell, 2015). In this process, however, the synapses that were initially given a low weight may at a later stage of more mature synaptic weight distributions gradually become useful to the system again. Therefore, it is necessary that the parallel fibre synapses on the Purkinje cell can also be potentiated. Indeed, several experimental studies indicate the presence of postsynaptic LTP in this synapse (Jörntell & Ekerot, 2002; Lev‐Ram et al. 2002; Coesmans et al. 2004), and that the requirements of this process are essentially a mirror image of the requirements for LTD: low as opposed to high calcium, no climbing fibre activation vs. the presence of simultaneous climbing fibre activation, phosphatase inhibitor activity as opposed to phosphatase activity (Jörntell & Hansel, 2006).

As both LTD and LTP of the parallel fibre synapses on Purkinje cells can participate in shaping the synaptic weight distribution, i.e. in shaping the connectivity of the network, an important question is what is the synaptic weight distribution among these synapses in early development, before major learning events have started taking place? At least as compared to rats as early as postnatal day 17–21, the synaptic weight distribution in the adult state is more skewed in that there are a few synapses with much higher weights as well as a clearer tendency for the majority of the parallel fibre synapses to have a very low weight (Valera et al. 2012). This would suggest that parallel fibre synapses from the beginning are given low, but non‐zero, weights and that the effect of learning is to enhance a few and depress a majority of the synapses.

In addition to inducing LTD of active parallel fibre synapses on Purkinje cells, climbing fibre inputs also exist to MLIs (Jörntell & Ekerot, 2002, 2003; Szapiro & Barbour, 2007) where they induce LTP of active parallel fibre synapses (Jörntell & Ekerot, 2002, 2003; Rancillac & Crepel, 2004; Jörntell & Ekerot, 2011), i.e. the opposite effect as compared to the parallel fibre synapses on Purkinje cells. Parallel fibre input without climbing fibre input leads to LTD of the parallel fibre input to the interneurons (Jörntell & Ekerot, 2002, 2003; Rancillac & Crepel, 2004; Jörntell & Ekerot, 2011). At least partly as a consequence of this anti‐parallelism in the parallel fibre plasticity mechanisms, the temporal firing modulation of Purkinje cells and their afferent interneurons can be mirror images of each other in cyclic motions (Jörntell & Ekerot, 2002; Barmack & Yakhnitsa, 2008). Indeed, when the pattern of firing of the climbing fibre input changes due to a re‐routing of the ingrowth of climbing fibres to the cerebellar cortex, the patterns of firing modulation during such cyclic motions change accordingly in the Purkinje cells and interneurons but remain mirror images of each other (Badura et al. 2013).

Climbing fibres may hence have a strong governing role for the shaping of the overall firing modulation in relation to cyclic motions in Purkinje cells, which can be explained by the parallel fibre synaptic plasticity (Jörntell & Hansel, 2006) and the similarity of the information that drives the mossy fibres and the climbing fibres (Figs 2 B and 3). However, as described above, the discharge rate of climbing fibres, their apparent active suppression during movement (Horn et al. 2004), as well as the wide temporal window for inducing LTD around a climbing fibre discharge (Safo & Regehr, 2008; Jirenhed et al. 2013), suggest that climbing fibre‐governed plasticity is insufficient to generate connectivity patterns with the high degree of specificity that would be needed for the integration of information and the coordination of more complex, high‐dimensional systems. In contrast, parallel fibre LTP in Purkinje cells and interneurons could be sufficiently temporally specific to solve this problem even though that remains to be shown.

The cerebellar microcircuitry: associative and activity‐balancing subsystems

In a scenario where spike times are non‐deterministic (Spanne et al. 2014 b) and sparse coding is not a main organizational principle, but the cerebellum is instead considered responsible for forming and linking models in a regressor‐like fashion with a high capacity for generalization (Spanne & Jörntell, 2015), the main function of the cerebellar cortex becomes that of a gigantic associative element. In this associative machine, mossy fibre input/information can contribute to driving or suppressing the Purkinje cell spike output depending on whether the information is mediated directly via the parallel fibre or indirectly via the parallel fibre–inhibitory interneuron. The central point of the cerebellar cortex then becomes to learn which parts of the incoming mossy fibre information should be associated with each other, and with what relative weights (where even negative weights are hence allowed), in relation to the functional contribution of the Purkinje cell (Fig. 4). The inhibitory effect of the Purkinje cell is exerted on the cerebellar nuclear neuron, which in itself can be excited by direct mossy fibre inputs. Hence, if the Purkinje cell is inhibited by inhibitory interneurons during a particular phase of a movement, the cerebellum can in principle generate net excitation of downstream motoneurons, even though that net excitation is dependent on activity in the spinocerebellar neurons that is driven by motor centres. As will be argued below, depending on the complexity of the system controlled, a straightforward function for the cerebellum to find and store useful associations between sensorimotor functions within a very rich set of information is sufficient to achieve complex functionality at the same time as it may still be possible to survive without its contribution, i.e. at least in humans in modern society where the full range of the behavioural repertoire made possible with an intact cerebellum may not be necessary for survival.

Figure 4. Proposed division into functional and activity balancing circuitry in the cerebellum.

A, cerebellar circuitry with the major cellular components, including the feedback control of the inferior olive (IO) from the deep cerebellar nucleus (DCN). Mossy fibres (mf) provide the main driving input of the circuitry and synapse with the granule cells (grc) and the DCN cells. The grcs are under inhibitory control from the Golgi cells (Goc). grc information is forwarded to the molecular layer interneurons (MLI) and the Purkinje cells (PC), where the latter provide the output of the cortex in the format of inhibitory input to the DCN cells. The DCN cells in turn provide the output of the cerebellum, the route by which effects on various motor command structures are exerted. The IO provides the cerebellar cortex with climbing fibres, which have a very powerful synapse on the PCs, but also contact MLIs and Gocs. B, in the proposed interpretation, the functional components of the circuitry can be limited to the MLIs, PCs and DCN output neurons, where the mossy fibre input (via granule cells in the case of the MLIs and the PCs) is the main plastic component that can mediate functional change by altering the weights in the matrix of the mossy fibre‐to‐Purkinje cell connections. Plasticity at other types of synapses contributes primarily to the activity balancing of the circuitry. C, assuming that the main function of MLIs is to make it possible to invert the polarity of the mf–grc input, the MLIs can be eliminated from the ‘effective function’ circuitry if one assumes that the mf–grc synaptic input can be assigned both positive and negative weights in learning processes. The circuitry hence becomes much simplified, but can still achieve complex functions due to the fact that each PC receives input from a massive population of mfs with a wide spectrum of information and patterns of temporal activation (matrix to the left). Also the output of the cerebellum should likewise be considered as a massive population of outputs (matrix to the right), where each individual DCN neuron only carries a fragment of the information necessary for the control of the motor output structures.

For the control of one‐dimensional systems, like the vestibulo‐ocular reflex where individual microzones may control eye movements in one dimension (one muscle) with input from the vestibular organ as a main driver of the motor output cells in the vestibular nuclei (Ito, 1984; Jörntell & Hansel, 2006), the function of the cerebellum becomes that of a relatively simple gain control. The mossy fibre inputs that are relevant for the gain of the movement dimension controlled are specifically potentiated or depressed in order for the reflex to work appropriately (i.e. controlling the gain of the reflex pathway so that the retinal image does not slip as the head moves).

For multi‐dimensional systems, like whole‐body or whole‐limb control, the range of possibilities becomes much wider. A key element is the interaction between the muscle output and the biomechanics of the body. The biomechanics is the ‘filter’ that determines in which spatiotemporal pattern the sensors become activated for a given muscle output (Spanne & Jörntell, 2013). The patterns of sensor activation represent the readout of the effect of the motor command on the body. The pool of spinal interneurons, which have a key position in somatic motor control and are activated by a combination of motor command signals and the sensor feedback, carries this information (Fig. 2 B). At the same time as this information is used to activate the α‐motoneurons, parts of it are also forwarded to the cerebellum through the axon branches of these spinal interneurons via the SCTs and the SRCTs. However, the spinocerebellar information is decomposable into smaller functional units, i.e. the information carried by the individual spinal interneurons. Spinal interneurons are by means of their branching patterns natural synergy controllers, i.e. they individually control multiple muscles (Jankowska, 1992; Santello et al. 2013). What could be achieved via the cerebellum is that different sets of synergy controllers can be linked to each other to form more well‐composed synergies (Bengtsson & Jörntell, 2014), or chains of synergies linked to each other in a particular temporal sequence. Thereby, smoother, faster, more complex and probably also more precise movements can be achieved than what would have been possible to achieve without the cerebellum.

The conclusion from this reasoning is that for the aspects of motor function that we normally associate with the cerebellum, it is probably sufficient to assume that its role is to link the appropriate combinations of mossy fibre inputs with each other on a Purkinje cell‐by‐Purkinje cell basis. The specification of what mossy fibre inputs are relevant for a particular Purkinje cell should depend on its function, i.e. its contribution to ongoing movements via the DCN cells. This function is determined by the motor output structures that the DCN cell targets (Fig. 1), and what particular spinal interneurons it influences via the population of neurons contacted in those output targets. Here it is important to realize that a Purkinje cell is never activated alone. All motor output can be expected to result in a very large number of Purkinje cells, DCN cells and spinal interneurons being activated together alongside all of the motor command structures of the brainstem and motor cortex. Therefore, the effect that the output of any single Purkinje cell will have depends on in which context, i.e. sensorimotor state, the output occurs. An essential contribution from the mossy fibre systems therefore will be to signal the state. As a very high number of states are possible (Spanne & Jörntell, 2015), a main task of the Purkinje cell will be to keep track of which state applies at the moment and to learn which types of mossy fibre signals it should associate with that state. At the same time, the synaptic weight of a mossy–parallel fibre synapse can take only one value, which applies across all states, and if a Purkinje cell is to be able to contribute to motor control across all states it needs to find a set of synaptic weight distributions for all mossy–parallel fibre inputs that applies across all states. This is a very demanding task that requires an enormous memory storage capacity, which may explain why the Purkinje cell dendritic tree has assumed such enormous proportions.

A major issue is how the learning of such multiple states can be achieved at the cellular level. First, it is important that the mossy fibre–parallel fibre system does not follow sparse coding principles. This is because the ability to make learned circuitry structures generalizable, i.e. to be useful across multiple states or contexts, is most effectively provided by a system with a more dense coding strategy. In the case of the cerebellum, this would translate to many or all granule cells participating in all contexts, although each individual granule cell does so with a fine graded signal that depends on the state and where also zero activity counts (Spanne & Jörntell, 2015). Second, it is known that the parallel fibre synaptic inputs to both Purkinje cells and their afferent molecular layer interneurons are bi‐directionally plastic, i.e. they can both be potentiated or depressed (Jörntell & Ekerot, 2002; Rancillac & Crepel, 2004; Jörntell & Hansel, 2006; Dean et al. 2010; Jörntell et al. 2010). In this context, climbing fibres are expected to play an important role, since their activation is a signal for the parallel fibre synapses on the Purkinje cells to become depressed (LTD) and those on the molecular layer interneurons to become potentiated (LTP). However, the activity of the climbing fibre system is typically depressed during active movement (Horn et al. 2004) and the time of activation of individual climbing fibre discharges is widely stochastic (Bengtsson & Jörntell, 2009 a) and does not seem to have the temporal precision in relation to movement execution that would be required of a system carrying the main responsibility for the formation of the high precision synaptic weight distributions that could support the performance of that movement at any level of detail. However, it is essential with a powerful system for down‐regulating the efficacy of synaptic weights in the cerebellum: if all granule cells, i.e. all mossy fibre–parallel fibres, participate in the signalling of all states, the complexity of the learning required becomes very high unless a majority of the parallel fibre synapses are kept silent, i.e. given near‐zero synaptic weight (Spanne & Jörntell, 2015). The climbing fibre induced LTD of parallel fibre synaptic input to the Purkinje cells fulfils this criterion: it is a low‐threshold, powerful form of synaptic plasticity (Launey et al. 2004), which is induced across a wide time‐window relative to the occurrence of climbing fibre input (Safo & Regehr, 2008; Jirenhed et al. 2013) and appears to leave most parallel fibre synapses silent (Ekerot & Jörntell, 2001; Isope & Barbour, 2002; Valera et al. 2012). Finally, although the context that applies at any given moment is likely to be primarily signalled by the mossy fibre systems, it is also possible that the tonic Golgi cell inhibition of granule cells can participate in this signalling, by regulating the excitability of the granule cells according to state (Spanne & Jörntell, 2013).

Rationale for the need of an activity balancing system in the cerebellum

Both Purkinje cells and their target DCN neurons are firing perpetually. During behaviour, their firing can be deeply modulated, but never ceases for any longer duration of time. Since the cerebellum has a pivotal role in motor output, by direct connections to all of the major motor output structures, it becomes important to make sure that this output does not transgress outside an overall range of activity over time. As a practical example, this is likely to be useful to avoid overexcitation of muscles, or to avoid muscles and α‐motoneurons having an excitability too low to become instantly activated when motor command signals arrive. Indeed, a clinical manifestation of stroke lesions in the cerebellum appears to be an altered excitability in peripheral motor axons (Huynh et al. 2014).

Hence, many aspects of the cerebellar neuronal circuitry, in particular the Golgi cells and their inhibition of granule cells, but also the low rate, spontaneous activation of climbing fibre synapses on Purkinje cells could serve as a regulatory system that keeps the overall activity of the tonic output of the cerebellar cortex and nuclei within operative range (Jörntell, 2016). The main arguments for this view is that unlike very young animals (Brickley et al. 1996; Duguid et al. 2015), in the adult state the Golgi cell to granule cell inhibition appears to be almost entirely tonic with little or no fast IPSP effects (Brickley et al. 1996; Jörntell & Ekerot, 2006; Bengtsson et al. 2013). Thereby, they cannot substantially contribute to fast cortical processing, but they can regulate granule cell excitability from one context to another (Spanne & Jörntell, 2013). Golgi cells are by themselves regulated by an inhibitory input from Lugaro cells, which are in turn regulated by serotonin (Dieudonne & Dumoulin, 2000). For the climbing fibre synapses on Purkinje cells, it is clear that apart from contributing to the LTD of parallel fibre synapses, and thereby reducing the overall amount of excitatory synaptic input, they also directly regulate Purkinje cell simple spike output over longer time scales (Bengtsson et al. 2004; Cerminara & Rawson, 2004). Such regulation could, for example, involve the expression of metabotropic glutamate channels at the sites of the parallel fibre synapses activated in conjunction with the climbing fibre synapse (Johansson et al. 2015). Because of the feedback via the inhibitory control of the inferior olive from specific cells in the cerebellar nucleus (Svensson et al. 2006; Chaumont et al. 2013), the output of the Purkinje cells can be used to control their own overall level of firing. Perturbing this feedback control system leads to severe problems. Inactivation of the inferior olive, which can be expected to result in a dramatic increase in the firing of the Purkinje cells being contacted by the corresponding climbing fibres (Bengtsson et al. 2004; Cerminara & Rawson, 2004), leads to severe loss of muscle tone control and motor deficits (Horn et al. 2010, 2013), with the specific deficit being dependent on which particular part of the inferior olive (i.e. which sagittal zones that are affected, Fig. 1) is inactivated.

Cerebellar coding and granule cell function

The circuitry structure portrayed above provides limited opportunity to perform advanced operations to transform the information that the cerebellum receives from the numerous mossy fibre pathways. In particular the absence of dominant intrinsic recurrent connectivity within the cerebellar cortex in principle excludes any other possibility than that of a pure associative function of this network. This arrangement arguably makes sense, since the information that is provided at least by the spinocerebellar systems (Jankowska et al. 2011 b; Shrestha et al. 2012; Spanne & Jörntell, 2013), and also the cuneocerebellar system (Jörntell et al. 2014), represents relatively precise and specific information that it may not be wise to transform to a substantial extent. The fact that the Purkinje cell firing pattern in some cerebellar areas correlates with predictions of limb kinematics across multiple behavioural contexts (Pasalar et al. 2006) could, for example, be explained by these Purkinje cells combining specific information from the SCT/SRCTs. The huge potential for associative function in the cerebellar cortex, and in particular the Purkinje cells, was also one of the first potential functions recognized by early investigators (Eccles et al. 1967; Marr, 1969; Albus, 1971). However, in addition, the very high numbers of granule cells were historically assumed to imply that some major transformation of the mossy fibre information was carried out at this level. In particular, Marr (1969) proposed that a random combination of mossy fibre inputs occurred at the level of the granule cells. Given that each granule cell receives only in the order of four mossy fibre synaptic inputs, the chances that each granule cell could be given a unique condition of activation would be high in this scenario. Further diversification of the granule cell signalling was predicted to be achieved by the Golgi cell inhibition of the granule cells. The idea of the granule cell layer as an expansion recoding unit, i.e. where mossy fibre information from different sources is combined to generate a new and richer set of information, fitted the likewise early idea of the output Purkinje cell as a ‘perceptron’ (Albus, 1971). As a perceptron is an element with a binary activation function, each Purkinje cell was expected to be activated by a unique combination of granule cell input patterns. Because of the uniqueness of the input, each Purkinje cell was expected to be activated primarily during a particular phase of the movement and thereby contribute to motor control. For discriminative functions such as this, where the cerebellum is assumed to work as a classifier of contexts and use that as a means to control motor programme activation, it is important to have as diverse a set of inputs as possible to the associative layer.

Due to the fact that granule cells for many decades remained impossible to record from, these ideas became highly influential. However, over the last decade granule cell recordings in the cerebellum in vivo have emerged (Chadderton et al. 2004; Jörntell & Ekerot, 2006; Arenz et al. 2008; Bengtsson & Jörntell, 2009 b) and major revelations have followed. Clearly, the initial ideas of each granule cell being activated by a unique event and then only firing in a sparse coding mode has proven incorrect (Jörntell & Ekerot, 2006; Ozden et al. 2012; Powell et al. 2015; Spanne & Jörntell, 2015). Furthermore, a binary activation function of a perceptron in the Purkinje cells is in principle impossible to defend given what we now know of the firing patterns in which the Purkinje cells are normally activated (Pasalar et al. 2006; Hewitt et al. 2011). However, the idea of each granule cell being activated by a unique combination of mossy fibre inputs still lingers. It has recently been supported in a couple of studies, which report that different mossy fibre systems converge in a single granule cell (Huang et al. 2013; Chabrol et al. 2015). Against these stand directly contradicting observations on the absence of integration of different types of information in the granule cells (Jörntell & Ekerot, 2006; Bengtsson & Jörntell, 2009 b), and a fundamentally different type of observation, i.e. that input from single mossy fibre pathways are sufficient to activate Purkinje cells (and, naturally, granule cells) (Ekerot & Jörntell, 2001; Geborek et al. 2013 b, 2014). There are two possible ways to reconcile these observations. One is to assume that the convergence between different types of mossy fibre information at the granule cell layer results in too much of a loss of information in systems devoted to the control of multi‐dimensional systems (Jörntell & Ekerot, 2006; Bengtsson & Jörntell, 2009 b; Geborek et al. 2013 b, 2014), whereas in the more low‐dimensional system, such as in eye movement control (Chabrol et al. 2015), this loss can be afforded and could even lead to advantages. The other explanation is that the specific structures that have been labelled for anterograde mossy fibre tracing in the anatomical studies (Huang et al. 2013; Chabrol et al. 2015) are not functionally different, i.e. they convey in principle equivalent information. In the study of Huang et al. (2013), the convergence of the basilar pontine nucleus and the external cuneate nucleus mossy fibres on granule cells may not necessarily mean that they represent functionally different information as the external cuneate nucleus projects to the pontine nuclei, and both structures receive cerebral input (Fig. 2 B). Likewise, Chabrol et al. (2015) studied inputs from the medial vestibular nucleus, the vestibular ganglion and the prepositus hypoglossi nucleus (PrH), where the latter was described as being a system mediating visual inputs. However, similar to the former two systems, PrH is also directly involved in the control of eye movements and hence all three inputs may correspond to motor commands (McFarland & Fuchs, 1992).

Assuming that the main activation of the granule cells comes from mossy fibres containing the same type of information from SCTs (Fig. 2 B), a simulation predicting how granule cells may respond under locomotion (Spanne et al. 2014 a) has striking qualitative similarities with actual granule cell responses during this condition in awake mice (Powell et al. 2015). Hence, current understanding of the granule cells is sufficient to approximately predict how they will be activated under behaviour, given that a general idea of how the afferent spinocerebellar mossy fibres are activated is available.

Concluding statement

For the functional description at the neuronal circuitry level, the present‐day understanding of the cerebellum is arguably unsurpassed in the mammalian brain. Due to the library of knowledge of cellular and synaptic physiological properties, as well as of the overall rules of plasticity that shape the patterns of synaptic connectivity, it is today theoretically possible to compute how the circuitry would respond and adapt to virtually any spatiotemporal pattern of mossy fibre information. However, this is also where today's challenge lies: the cerebellum receives information from a vast set of mossy fibre systems and the nature and spatiotemporal format of that information in relation to different types of behaviour need to be known in order to tell exactly what the cerebellum does in a given situation and in a given region. This issue is related to the larger issue of how information in the brain arises in the first place and how it is represented in the brain once generated.

Hence, for future advances of the field of cerebellar physiology a better characterization of the information forwarded by the different mossy fibre systems is urgent. The realization that most mossy fibre systems can be expected to be activated across many motor contexts, and that cerebellar learning therefore requires synaptic weights to become applicable across these contexts, is also a factor that needs to be taken better into account for any model of cerebellar circuitry function. The division of the cerebellar circuitry into activity balancing and associative components that is proposed here also awaits further experimental testing. Finally, an exciting future prospect is that the emerging understanding of how the sensorimotor mossy fibre information is formatted in order to be processed by the cerebellar circuitry will also lead to insights regarding the format of information from mental processes that is processed in other subdivisions of the cerebellum using a similar generic circuitry structure.

Additional information

Competing interests

None declared.

Funding

The author is supported by Vetenskapsrådet (Swedish Research Council; K2014‐63X‐14780‐12‐3).

Biography

Henrik Jörntell received his PhD degree in Neurophysiology from Lund University. He is currently employed as a senior lecturer at Lund University where he heads the lab ‘Neural Basis for Sensorimotor Control’ at the Department of Experimental Medical Science. His interests are the neurophysiological analysis of neuronal microcircuits involved in movement control, spanning cerebellar, spinal, brainstem and neocortical circuitry as well as models of how these structures interact during movement performance and motor learning.