Extending the bandwidth of long‐term plasticity at the cerebellar input stage

The granule cell layer of the cerebellum, the input stage of the cerebellar cortex, contains certainly as many neurons as the whole cerebral cortex. Since David Marr's hypothesis (Marr, 1969), models and theories of cerebellar information processing have suggested that this huge number of neurons is dedicated to the separation of sensori-motor input patterns provided by mossy fibres, the major excitatory input of the cerebellar cortex. The fundamental notion is that combinatorial sampling of mossy inputs by granule cells amplifies discrepancies between similar patterns carried by overlapping collections of mossy fibres. This notion is supported by the extensive divergence at this stage: the high number of mossy terminals or rosettes per fibre (20–50) and the fact that dendritic claws from up to 50 granule cells contact one rosette in an intricate structure called the glomerulus. Furthermore, each dendritic claw – from the 2–5 branches in a granule cell – receives one input from a single mossy fibre and the granule cell requires at least two mossy fibres inputs to emit a single or a burst of action potentials. In this picture, Golgi cells control the level of granule cell activity via a large axonal plexus invading many glomeruli and mediating both tonic and phasic inhibition. The gain control of the Golgi cell network is triggered both by a feedback inhibition via granule cell excitation or by a feedforward inhibition via direct mossy fibre inputs. Although Marr's predictions on synaptic plasticities were not verified exactly – for example that only synapses between parallel fibres and Purkinje cells are modifiable – it should be emphasized that his fundamental hypothesis of pattern discrimination remains hugely influential forty years later. The recent work of the group of Egidio D’Angelo in Italy, published in The Journal of Physiology (see D’Errico et al. 2009), extends our understanding of granule cell layer computation and provides further support for Marr's hypothesis. In their study, they identified a long-term depression (LTD) of mossy fibre EPSCs induced by low frequency stimulation (1–10 Hz) that is dependent on mGluR receptor activation. This LTD is calcium dependent in that strong buffering of postsynaptic free calcium blocks LTD induction. Increasing the mossy fibre stimulation frequency above 10 Hz leads to a build-up of calcium in granule cell dendrites and an NMDA receptor-dependent long-term potentiation (LTP) of EPSCs is then observed as shown previously by the same group. This induction protocol is similar to BCM (Bienenstock–Cooper–Munro, 1982) rules for plasticity induction observed at hippocampal synapses and opposite to plasticity rules at parallel fibre-to-Purkinje cell synapses where a high calcium concentration leads to a long-term decrease in synaptic efficacy. Interestingly, D’Errico et al. suggest that both mGluR-dependent LTD and NMDA receptor-dependent LTP are expressed presynaptically. Although presynaptic mechanisms are difficult to decouple from complex postsynaptic pathways, particularly in the mossy fibre glomerulus, all classical indicators (coefficient of variation, failure rate and paired pulse ratio) converge to suggest a change of release probability at mossy fibre rosettes after plasticity induction. In a previous study (Mapelli & D’Angelo, 2007), this group showed that short high frequency bursts of mossy fibre stimulation leading to a moderate increase in calcium level induced an NMDA receptor-dependent LTD, then the authors propose that the level of free calcium rather than the origin of calcium increase (mGlu or NMDA receptors) is what determines the sign of the plasticity. The issue now is to understand how calcium level is translated into a retrograde message capable of targeting the presynaptic machinery of vesicle release and modifying in both directions the release probability. One hypothesis could be that two competing transduction pathways regulated by postsynaptic calcium concentration lead to the production of two retrograde messengers – possibly NO and endocannabinoids – controlling different presynaptic targets. What is the relevance of these results for information processing at this input stage? Several studies have now described in vivo the temporal patterns of mossy fibre input to granule cells in the cerebellar cortex of rodents and cats (Chadderton et al. 2004; Jorntell & Ekerot, 2006; Arenz et al. 2008). An interesting point is the wide range of temporal input patterns these groups have observed in their preparation, whether they record in the crus or the vestibular region of the rat or in the C3 zone of the cat. In crus I or II, spontaneous EPSCs in granule cells occur at 4 Hz and bursts of EPSCs can have instantaneous frequencies above 500 Hz; in the vestibular area, frequencies of EPSCs range between 1 Hz and 40 Hz. In the C3 zone of lobules IV and V of the cat, spontaneous activity ranges from 10 to 50 Hz. The fact that LTD is induced by a metabotropic pathway which does not require postsynaptic depolarization allows granule cells to integrate a wider range input frequency (from 1 Hz to 10 Hz for LTD and above 10 Hz for LTP). These bidirectional plasticities may increase the signal-to-noise ratio for input selection by filtering out spurious or uncorrelated inputs. Previous findings from this group suggest that a tight control of the balance between the direct mossy fibre excitatory drive and the feedforward/feedback inhibition via Golgi cells adjusts the balance between LTD and LTP, favouring selected inputs and discarding others. Since 2–5 mossy fibres can contact the same granule cell, it is also likely that coincidence detection mechanisms might decrease the threshold frequency inducing LTP. Furthermore, the presynaptic site of expression of both plasticities suggests that input selection rules might be shared with other granule cells that contact a given glomerulus, allowing one selected mossy fibre to excite a group of granule cells. All these properties would be in full agreement with Marr's hypothesis, implementing and extending the rules for input pattern segregation.