|Home | About | Journals | Submit | Contact Us | Français|
The way the cerebellar granular layer transforms incoming mossy fiber signals into new spike patterns to be related to Purkinje cells is not yet clear. Here, a realistic computational model of the granular layer was developed and used to address four main functional hypotheses: center-surround organization, time-windowing, high-pass filtering in responses to spike bursts and coherent oscillations in response to diffuse random activity. The model network was activated using patterns inspired by those recorded in vivo. Burst stimulation of a small mossy fiber bundle resulted in granule cell bursts delimited in time (time windowing) and space (center-surround) by network inhibition. This burst–burst transmission showed marked frequency-dependence configuring a high-pass filter with cut-off frequency around 100Hz. The contrast between center and surround properties was regulated by the excitatory–inhibitory balance. The stronger excitation made the center more responsive to 10–50Hz input frequencies and enhanced the granule cell output (with spikes occurring earlier and with higher frequency and number) compared to the surround. Finally, over a certain level of mossy fiber background activity, the circuit generated coherent oscillations in the theta-frequency band. All these processes were fine-tuned by NMDA and GABA-A receptor activation and neurotransmitter vesicle cycling in the cerebellar glomeruli. This model shows that available knowledge on cellular mechanisms is sufficient to unify the main functional hypotheses on the cerebellum granular layer and suggests that this network can behave as an adaptable spatio-temporal filter coordinated by theta-frequency oscillations.
The cerebellum, owing to its regular structure, has inspired several theoretical models emphasizing the combinatorial properties of the network (Marr, 1969; Albus, 1971; Tyrrell and Willshaw, 1992; Dean et al., 2010). More recently, simplified spiking models have also been developed (Maex and De Schutter, 1998; Medina and Mauk, 2000). Here, by taking advantage of recent advances in single cell and granular layer circuit physiology, a new computational model is developed that incorporates a much higher level of realism than previously possible. The model has then been used to investigate granular layer spatio-temporal dynamics and evaluate the main functional hypotheses suggested by biological investigations.
In brain slices, intracellular recordings have provided a detailed understanding of membrane and synaptic properties of single neurons (among others D'Angelo et al., 1995; Dieudonné, 1998; Forti et al., 2006; see also below), while multi-electrode array (MEA) and voltage-sensitive-dye imaging (VSD) recordings have revealed relevant aspects of ensemble granular layer activities (Mapelli and D'Angelo, 2007; D'Angelo et al., 2009; Mapelli et al., 2010a,b). Recordings in vivo have shown the behavior of single granular layer neurons (Vos et al., 1999; Chadderton et al., 2004; Jörntell and Ekerot, 2006; Rancz et al., 2007) and their collective activity (Roggeri et al., 2008; Courtemanche et al., 2009) in response to natural stimuli. These results have supported in turn four main functional hypotheses. (1) The granular layer response to mossy fiber bursts should be spatially organized in a center-surround pattern, where the excitatory–inhibitory balance is dominated by excitation in the center and by inhibition in the surrounding areas. (2) The granular layer should generate a time-window effect limiting the duration and intensity of the GrC output (D'Angelo and De Zeeuw, 2009). (3) The granular layer should behave as a high-pass filter allowing patterns over 50Hz to be optimally transmitted (Mapelli et al., 2010b). Finally, (4) sparse GrC random activity should be able to sustain coherent low-frequency oscillations of granular layer activity (Maex and De Schutter, 1998). However, whether and how these aspects integrate into a coherent functional framework has remained unclear.
Among the properties that may contribute to determine the granular layer function, some appear especially relevant. The cardinal structural properties include the glomerular organization of the mossy fiber inputs (Rossi and Hamann, 1998; Mapelli et al., 2009) and feed-forward, feed-back and lateral inhibition (Mapelli and D'Angelo, 2007; Kanichay and Silver, 2008). At the molecular level, special importance has been attributed to NMDA and GABA A (α1 and α6 subunit-containing) receptors (D'Angelo et al., 1995; Cull-Candy et al., 1998; Farrant and Nusser, 2005). These are highly sensitive to neurotransmitter spillover and are suitable to set the appropriate time constants for signal processing in the cerebellar glomerulus. Finally, neurotransmitter release probability at the mf-GrC synapse, which can be tuned by long-term synaptic plasticity (D'Errico et al., 2009), regulates the time course of EPSP temporal summation and spike emission. Thus, these factors need to be taken into account to appropriately simulate granular layer dynamics in the spatial, temporal and frequency domains.
The computational reconstruction of the granular layer reported in this work was based on biophysically realistic models of granule cells (GrCs: D'Angelo et al., 2001; Diwakar et al., 2009) and Golgi cells (GoCs: Solinas et al., 2007a,b) and of their synapses (Nieus et al., 2006; Mapelli et al., 2009). The network, stimulated with patterns inspired to those observed in the mossy fibers (mfs) in vivo, generated center-surround, time-windowing, high-pass filtering and theta-frequency oscillations regulated by neurotransmitter release and NMDA and GABA-A synaptic receptors. Interestingly, the center-surround structure was capable of fine-tuning the delay, number, and frequency of spikes generated by GrCs suggesting that the cerebellum granular layer behaves as a complex spatio-temporal filter, which can be adapted through long-term synaptic plasticity and coordinated by coherent oscillations.
The granular layer network structure was generated on the basis of detailed anatomical and functional information (Eccles et al., 1967; Palkovits et al., 1971, 1972; Hámori and Somogyi, 1983; Jakab and Hámori, 1988; Harvey and Napper, 1991; Korbo et al., 1993; Sultan, 2001; Barmack and Yakhnitsa, 2008) and using models of neurons and synapses including biophysical representations of membrane ionic channels and receptors (D'Angelo et al., 2001; Nieus et al., 2006; Solinas et al., 2007a,b; Diwakar et al., 2009). These neuronal and synaptic models have been extensively validated in previous works using electrophysiological and imaging data. Therefore, we have been able to develop a “realistic network model”, in which the large number of parameters is constrained to biology. In this “bottom-up” approach, the functional properties of the network emerge from the properties of constitutive elements and from their synaptic organization (e.g. see Druckmann et al., 2007; Gleeson et al., 2007; Izhikevich and Edelman, 2008). The appropriateness of network responses was assessed by comparison with MEA and VSD recordings of network activity, which have recently become available for the cerebellum granular layer. Moreover, the availability of input spike patterns and neuronal responses in vivo (Vos et al., 1999; Chadderton et al., 2004; Jörntell and Ekerot, 2006; Rancz et al., 2007; Roggeri et al., 2008) has allowed to simulate granular layer network dynamics under conditions representative of natural activity states.
The network had a size sufficient to reproduce a functionally relevant portion of the cerebellar granular layer, i.e. a cube with 100μm edge length (Table (Table11 and Figure Figure1).1). The model included 315mfs, 4393 neurons (4096 GrCs, 27 GoCs and 270 SCs/BCs) and more than 40000 synapses. The number of cells and synapses was large enough to maintain realistic convergence/divergence ratios (Eccles et al., 1967). On this scale, mf branching was not implemented (see Sultan and Heck, 2003). Moreover, to achieve inhibitory control over GoCs, a partial representation of the SC/BC was also included.
Network connections were constructed using precise rules, yet allowing the number of connections and synaptic weights to show statistical variability (Gaussian distribution: mean =1, s.d. =0.4; see Medina and Mauk, 2000). No systematic differences were observed using different seeds for parameter randomization, so that in several cases the same network configuration was used to facilitate data comparison. Background noise in the network was generated by random spike patterns in mfs and pacemaking in GoCs and SC/BCs (see e.g. Häusser and Clark, 1997; Chadderton et al., 2004; Rancz et al., 2007). Neurons and synapses were endowed with multiple receptor and ionic channel-based mechanisms, allowing an accurate representation of neuronal firing. The synapses were endowed with neurotransmitter diffusion mechanisms and with a representation of vesicle cycling, generating spillover and developing short-term facilitation and depression. However, no molecular noise (e.g. from ionic diffusion, channel gating or receptor binding) or synaptic noise (e.g. from stochastic vesicle fusion) were introduced.
The model was written with NEURON-7.1. The simulation of 3s of activity required about 20h on a Pentium-5 dual-core but just 30min using 80 CPUS on the CASPUR parallel cluster (http://www.caspur.it/en/). A graphical interface was written to represent the data as in MEA and VSD experiments.
The most relevant aspects of granular layer network organization depend on the inhibitory connections, which are therefore considered in detail. Lateral inhibition, as in other brain areas (Buzsáki, 2006), is an important structural property of the circuit. This has been reported since early anatomical investigations (e.g. see Eccles et al., 1967) and has recently been substantiated by electrophysiological recordings (Mapelli and D'Angelo, 2007). Lateral inhibition descended from the disproportionate extension of the GoC axon compared to its dendritic afferent field (i.e., the mossy fibers excited an area including GrC and GoC basal dendrites, but then the GoC axon redistributed inhibition over a much larger area). Two relevant aspects of the inhibitory connections, which have recently been reported in experimental investigations, have been taken into account in the model (Figure (Figure11C).
First, ultrastructural measurements have revealed that each GrC receives on average three GoC inhibitory synapses (Hámori and Somogyi, 1983), but has left open the problem on whether these synapses originated from the same or from different GoCs. Typically, GrC IPSCs can be recruited by raising stimulation intensity (Mapelli et al., 2009), suggesting that three to four independent GoCs are indeed connected. Recordings from GoC-GrC pairs further support this connectivity by showing that the size of the GoC-GrC IPSCs is equivalent to the minimal response generated by single fiber stimulation (Mapelli et al., 2009).
Secondly, since glomeruli receive about 50 dendrites from as many different GrCs, the additional issue is whether a GoC innervates all the GrCs impinging on the same glomerulus. Even minimal stimulation (i.e. one that activates a single synaptic contact) can elicit a direct and an indirect spillover-mediated component in GrC IPSCs (Mapelli et al., 2009). Since spillover is a sign of release on neighboring synapses in the glomerulus (Rossi and Hamann, 1998), a GoC axon should inhibit numerous (if not all) GrC dendrites in the same glomerulus.
Therefore, in addition to lateral inhibition, the GoC-GrC connections were implemented with these topological rules: a GoC was not allowed to innervate a GrC more than once (and therefore did not innervate adjacent glomeruli) and each glomerulus was fully innervated by a single GoC. In fact these rules are a simplification, since do not separate direct from indirect inhibition. An explicit representation of the glomerulus with internal diffusion allowing for independent generation of direct and indirect IPSCs may further improve this description.
The generation of the model network occurred in three steps: (1) calculating the number of constitutive elements, (2) distributing the elements in space, and (3) connecting the elements.
In order to test the impact of the connectivity rules, in some modeling experiments a “mesh” configuration was used. The “mesh” was built after removing the connectivity rules 3c and 3d reported above, so that GoCs were connected to GrC distributed all over the network. Nonetheless, each GrC still received the same number of excitatory and inhibitory synapses. A competitive growth algorithm would be desirable in the future.
The neuron (GrC and GoC) models derived from previous models, which had been carefully tested against available experimental results in slices (D'Angelo et al., 2001; Nieus et al., 2006; Solinas et al., 2007a,b). These models were able to reproduce all the details of spike shape, timing and frequency in response to current injection and synaptic stimulation. The synaptic models were adapted from the original scheme reported by Nieus et al. (2006) and were able to reproduce the kinetics and size of the EPSCs and IPSCs during repetitive synaptic transmission at the different synapses. These models accounted for vesicular dynamics, neurotransmitter spillover and receptor gating (including multiple closed, desensitized and open states) but not for quantal release mechanisms. The dynamics of synaptic responses were fully determined by the kinetic constants of synaptic and neuronal models. Given the short distances traveled by the spikes, axonal conduction times were considered negligible. Transmission delay was 1ms for all the synapses.
In order to conform to in vivo conditions, all models had to be adapted from their original temperature Torig to Tsim =37°C using the correction factor Q10(Tsim–Torig)/10 (Gutfreund et al., 1995; see also Traub and Llinas, 1979; Traub et al., 1991; Vanier and Bower, 1999). We have used: Q10 =3 for ionic channel gating, Q10 =2.4 for receptor gating, Q10 =1.5 for ionic channel permeation, Q10 =1.3 for neurotransmitter diffusion, Q10 =3 for Ca2+ pumps and buffers, Q10 =1.3 (GrC) or 1.7 (GrC) for intracellular Ca2+ diffusion. Following adaptation at 37°C, the models were in matching with recordings at this same temperature (data not shown). The basic properties of GrCs and GoCs embedded into the granular layer network are shown in Figures Figures22–4.
The GrC model was adapted from Nieus et al. (2006) by applying appropriate Q10 corrections. In addition, the GABA leakage conductance was increased by two times (60μS/cm2), the Inward rectifier K+ conductance was increase by 1.5 times (1350μS/cm2) and the leakage reversal potential was adjusted to restoring resting potential to −70mV (see D'Angelo et al., 2001). With this asset, the GrC model properly reproduced responses to current injection at 37°C (data not shown) and spike trains observed in vivo (Chadderton et al., 2004; Jörntell and Ekerot, 2006) reaching maximum firing rates as high as 500Hz (see Figure Figure22B1).
Without needing any further change, the GoC model properly reproduced responses to peripheral stimulation observed in vivo (Vos et al., 1999) and could reach a maximum firing rate of 350Hz (see Figure Figure22B2).
The SC/BC models, in the absence of detailed computational representations, were designed as random spike generators with a basal firing of 18.5Hz (Armstrong and Rawson, 1969; Häusser and Clark, 1997).
The mf-GrC synapses take part to the formation of the cerebellar glomerulus, are glutamatergic and activate AMPA and NMDA receptors. The release, diffusion and ionic receptor mechanisms were the same reported by Nieus et al. (2006). Using a probability of release of 0.6, the model was able to faithfully reproduce postsynaptic currents recorded at 37°C in vitro (Saviane and Silver, 2006) and in vivo (Chadderton et al., 2004; Rancz et al., 2007). The time constant of the recovery from depression, τREC =8ms, was derived from in vivo measurements (Jörntell and Ekerot, 2006) and allowed to reproduce natural dynamics of short-term plasticity (the time constants of presynaptic facilitation and vesicle inactivation were set to τfacil=5ms and τI =1ms, respectively).
The mf-GoC synapses are similar in several respects to the mf-GrC synapses. They are also located within the cerebellar glomerulus (Eccles et al., 1967) and are glutamatergic activating both AMPA and NMDA receptors (Kanichay and Silver, 2008; Cesana et al., 2009). The mf-GoC synapse was adapted from the mf-GrC synapse model (see above) to reproduce a peak postsynaptic current of −66pA (Cesana et al., 2009; see Figures Figures22–5). Release probability and vesicle cycling parameters were set at the same values as at the mf-GrC synapse.
The GrC-GoC synapses are formed by pfs onto GoC apical dendrites in the molecular layer (Palay and Chan-Palay, 1974). These glutamatergic synapses activate AMPA, NMDA and Kainate-receptors (Dieudonné, 1998; Bureau et al., 2000; Misra et al., 2000). During repetitive stimulation, the AMPA current shows synaptic depression while the Kainate and NMDA currents show slow temporal summation. AMPA and NMDA currents were taken from the mf-GrC synapses and the kainate receptor current was modified from the AMPA kinetic scheme. Release probability was 0.1 and vesicle cycling parameters were set at the same values as at the mf-GrC synapse.
The GoC-GrC synapses are GABAergic and impinge on GrC dendrites within the glomeruli. GABAergic neurotransmission was modeled based on Mapelli et al. (2009). The GABA-A receptor schemes comprised channels with fast (α1) and slow (α6) kinetics and GABA spillover generating the transient and sustained components of inhibition observed experimentally. In order to account for experimental results (Mapelli et al., 2009), the parameters describing presynaptic dynamics were: release probability=0.35, τREC=36ms, τfacil =58.5ms and τI=0.1ms, respectively (Mapelli et al., 2009).
The SC/BC-GoC synapses are GABAergic and impinge on the GrC apical dendrites in the molecular layer. The postsynaptic current resulting from a single spike was described by a double-exponential function with τ1 =0.26ms, τ2=15ms, and 1370pS maximum conductance, similar to synapses made by the same neurons onto Purkinje cells (see Jaeger et al., 1997).
The granular layer shows a background activity state, over which the mf inputs generate evoked responses. In brain slice recordings, mfs and GrCs are silent, GoCs show rhythmic spontaneous activity at around 6Hz (Dieudonné, 1998; Forti et al., 2006) and SC/BCs show rhythmic spontaneous activity at around 18.5Hz (Häusser and Clark, 1997). Evoked activity occurs in GrC clusters of about 30μm diameter (Mapelli et al., 2010a), which may represent granular layer functional units (Sultan and Heck, 2003). In the anesthetized rat in vivo, the mfs show a low basal activity (<1Hz), GrCs generate sporadic spikes (<0.1Hz) (Chadderton et al., 2004; Rancz et al., 2007), GoCs show a basal activity at 1–18Hz (mean=8.42, numeric cv=0.43; Vos et al., 1999) and SC/BCs show a basal activity at 47±17Hz (Wang et al., 2009). Following punctuate sensory stimulation, mfs convey high-frequency bursts and GrCs and GoCs respond after short delay generating themselves short bursts at 200–300Hz (Vos et al., 1999, 2000; Chadderton et al., 2004; Jörntell and Ekerot, 2006; Rancz et al., 2007). These responses are clustered in small areas and can be detected by measuring the corresponding local field potentials (Roggeri et al., 2008).
Background activity in the model was generated by the following mechanisms. (i) The mfs were activated with a random spiking activity. Spike timing was drawn from a Poisson distribution (mean=1Hz, cv2 =0.9; see below for cv2 definition). (ii) GoCs were spontaneously active and their spike frequency and cv2 matched the range reported in vivo once mfs and SC/BC were made active themselves. (iii) SC/BCs were endowed with a random spiking activity drawn from a Poisson distribution (mean =18.5Hz, cv2=0.9). It should be noted that SC/BC functions were critical to control the basal activity state of GoCs. When the SC/BC random activity was turned off, the GoCs showed an unnaturally high firing frequency and blocked granular layer signal processing (unpublished observations).
The mf bursts generated by punctuate tactile stimulation are composed of 5–10 spikes with an average frequency around 100Hz and instantaneous frequencies as high as 700Hz in the anesthetized rat (Chadderton et al., 2004; Rancz et al., 2007). More protracted stimuli have been reported to generate longer bursts with frequencies modulated between a few Hz and 500Hz in the behaving monkey (Kase et al., 1980) and up to 1000Hz in the decerebrated cat (Jörntell and Ekerot, 2006). Evoked activity was simulated by stimulating eight contiguous unbranched mfs in the middle of the network (see Figure Figure1),1), which activated ~30μm large GrC clusters. The stimuli consisted in spike bursts of different frequencies and duration, as indicated in the specific result sections (Figures (Figures22–7), overriding the ongoing background activity.
Simulation results were stored for offline analysis and processed using customized MATLAB programs (The MathWorks, Natick, MA, USA).
Single cell activities were analyzed as in ordinary patch-clamp experiments (averages of up to 150 simulations were required to overcome the irregularity generated by background network activity). The auto- and cross-correlograms (e.g. see Figure Figure8)8) were generated using data from 3-s long simulations. In order to display the average firing rate of an individual cell, the histograms were normalized by the total number of cells, by the number of stimulation spikes and by the bin width (1ms).
Spatially organized activity was analyzed as in VSD experiments. Activity images (e.g. see Figure Figure5A)5A) were reconstructed from the membrane potential of individual GrCs generating an average response within a volume. The peak of the cumulative depolarization of GrCs contained within the given volume was in fact a function of spike synchrony and frequency (see Mapelli et al., 2010b). To keep into account image blurring due to light diffraction in VSD imaging recordings (Mapelli et al., 2010b), simulated images were constructed from GrCs located within ±10μm from the focal plane. The spatial profile of responses to focal mf stimulation was constructed by computing the mean depolarization of GrCs located within spherical shells with radius between 5 and 35μm from the core of excitation. The gain of transmission of mf bursts was evaluated by measuring the granular layer output at a given frequency relative to the single-pulse response (Figure (Figure77).
In order to quantify the excitatory–inhibitory balance (E-I balance), the reconstructed VSD images were analyzed as in MEA experiments. The network was stimulated to make a doublet of spikes. The intensity of excitation was measured at the peak of the first spike in control conditions. The intensity of inhibition was estimated by the changes in the second spike caused by GABA receptor blockage (see Mapelli and D'Angelo, 2007; see Figure Figure55).
In this work three definitions of the coefficient of variation were used for consistency with different experimental works: cv=SD/mean, numeric cv=MAD/median, cv2 =mean [2×|ISIn+1 −ISIn|/(ISIn+1 +ISIn)] (Holt et al., 1996).
Once embedded into the network, single neurons receive a continuous barrage of excitatory and inhibitory synaptic inputs driven by network-dependent activity. This can alter short-term facilitation and depression and the threshold and precision of spike firing (Dobrunz and Stevens, 1999; Klyachko and Stevens, 2006). The intrinsic responsiveness and synaptic activation of single GrCs and GoCs was assessed in model neurons embedded into the simulated granular layer network (Figures (Figures22–4), which generated a continuous background activity (see Materials and Methods). Intrinsic responsiveness was assessed through intracellular current injection and synaptic responses were elicited by delivering a 100-Hz burst composed of five impulses over a small bundle of eight adjacent and unbranched mfs. With this stimulus, the response occupied a surface of about 30μm diameter reflecting the spots observed with VSD recordings in slices (Mapelli et al., 2010a,b) and matching the granular layer functional units reported by Sultan and Heck (2003) (the spatial aspects of the response are considered in detail below, see Figures Figures55 and and66).
The GrC model (Figure (Figure2A1)2A1) background activity comprised EPSPs generated by mfs and IPSPs generated by GoCs, while spontaneous firing was almost absent. The GrC model properly reproduced responses to current injection (Figure (Figure2B1;2B1; D'Angelo et al., 1995, 2001). Most salient properties were non-sagging inward rectification in the subthreshold range and regular firing with almost no adaptation during depolarization, with maximum frequencies as high as 500Hz (Figure (Figure2B1).2B1). In response to a brief mf burst, the GrC model showed various responses depending on the number of active excitatory synapses and on the number and timing of IPSPs received through the mf-GoC-GrC (feed-forward) and mf-GrC-GoC-GrC (feed-back) loops (Figure (Figure2C1).2C1). In response to a short mf spike burst, the GrC model generated EPSPs showing short-term depression of the transient component and temporal summation of the protracted component, which could lead to brief spike bursts. The background and evoked activity of GrCs faithfully reproduced patch-clamp recordings in vivo (Eccles et al., 1967; Chadderton et al., 2004; Jörntell and Ekerot, 2006; Rancz et al., 2007).
The GoC model (Figure (Figure2A2)2A2) showed spontaneous activity determined by its pacemaker properties (Dieudonné, 1998; Forti et al., 2006) and modulated by mf, pf and SC/BC synaptic inputs. The GoC maintained a firing rate of 8.4Hz with cv2=0.44, consistent with recordings in vivo (Vos et al., 1999; Solinas et al., 2007b). Upon current injection (Figure (Figure2A2),2A2), the GoC model showed the complex repertoire of dynamic rebounds observed in recordings in slices, including firing with adaptation during depolarization, sagging inward rectification during hyperpolarization, rebound excitation on return to the resting state from hyperpolarization and phase-reset on return to the resting state from burst firing (Solinas et al., 2007a,b). The GoC model reached firing rates as high as 350Hz (Figure (Figure2B2).2B2). In response to a brief mf burst (Figure (Figure2C2),2C2), the GoC model generated EPSPs determined by direct mf-GoC activation and by propagation of activity through the mf-GrC-GoC loop (Cesana et al., 2009). With a weak input (8mfs) the GoC generated a single spike and phase reset the pacemaker. With a strong input (45mfs) the GoC generated a brief burst composed of two to three spikes followed by phase reset (Figure (Figure22C3, inset) as observed following punctuate stimulation in vivo (Vos et al., 1999; Solinas et al., 2007b). The spikes were generated in 1.5ms and the burst reached an initial frequency of over 300Hz (cf. Kanichay and Silver, 2008). The background and evoked activity of GoCs as well as their firing patterns are in keeping with extracellular recordings in vivo (for review and other references see D'Angelo, 2008).
These simulations show that the GrC and GoC models [for GrCs: D'Angelo et al. (2001), Nieus et al. (2006), Diwakar et al., (2009); for GoCs: Solinas et al. (2007a,b)] adapted to 37°C and subjected to background and stimulated network activity, were able to reproduce the typical response patterns observed in vivo [for GrC: Chadderton et al. (2004), Jörntell and Ekerot (2006), Rancz et al. (2007); for GoC: Vos et al. (1999)].
While neuronal responses in slices look stereotyped, the responses of GrCs and GoCs embedded into the network show a remarkable variability. This is due to random background activity, to variations in the number and strength of synaptic contacts and to the assortment of excitatory and inhibitory fibers impinging on a given cell. Figure Figure33 shows the cell-specific contribution of glutamate and GABA-A receptors to dendritic responses of GrCs and GoCs.
Fast excitatory transmission was ensured through AMPA receptor-mediated currents, which showed short-term depression at the mf-GrC and mf-GoC synapses and short-term facilitation at the pf-GoC synapses. Slow excitation was sustained by the NMDA and the kainate receptor-mediated currents, the former being most evident at the mf-GrC synapse and the latter being specifically expressed at the pf-GoC synapse (Dieudonné, 1998; Bureau et al., 2000; Misra et al., 2000; Kanichay and Silver, 2008; Cesana et al., 2009). Thus, whereas all excitatory synapse could efficiently react to spike bursts, only the mf-GrC synapse was sensitive to fast transitions in firing rate. Inhibition was mediated by GABA-A currents both at the GoC-GrC and at the SC/BC-GoC synapses. At the GoC-GrC synapses, GABAergic currents were protracted by the slow kinetics of GABA-A α6 receptors (see below; D'Angelo et al., 1995; Rossi and Hamann, 1998; Mapelli et al., 2009). Therefore, potentially, the network could implement the time-window mechanism (requiring fast GrC responses and delayed GoC reaction through the mf-GoC-GrC or feed-forward inhibitory loop; D'Angelo and De Zeeuw, 2009), high pass filtering (requiring frequency-dependent build-up of the burst response through NMDA currents at the mf-GrC relay), and slow oscillations (through the pf-GoC-GrC or feed-back inhibitory loop). These properties are considered below.
Among the processes controlling the GrC response, there are some that merit specific mention, since they are thought to confer the GrC with specific integration properties: the regulation of synaptic excitation by NMDA and GABA-A receptors and the regulation of neurotransmission by mf-GrC release probability (Figure (Figure4A).4A). In addition to AMPA receptors, GrCs have the most prominent expression of NMDA receptors in the cerebellum (Cull-Candy et al., 1998) and, unique in the brain, express the α6 subunit of GABA-A receptors (Farrant and Nusser, 2005). Both NMDA and GABA-A α6 receptors have slow kinetics and are highly sensitive to neurotransmitter spillover (see also Figure Figure33).
In these simulations, blocking the NMDA receptors markedly reduced EPSP temporal summation delaying and curtailing the GrC response, while blocking GABA-A α1 and GABA-A α6 receptors sorted the opposite effect. The time of occurrence for the 1st, 2nd, 3rd, 4th spikes was 6.5±1.1 (n=20), 14.1±1.3 (n=5), 24.6±1.3 (n=4), 36.4±1.6 (n=4)ms in control; 5.4±0.4 (n=34), 16.0±2.0 (n=35), 24.4±1.6 (n=37), 34.5±1.9 (n=21) ms with α6 receptor block; 5.3±0.5 (n=36), 13.3±1.0 (n=106), 23.6±0.8 (n=34), 33.4±0.6 (n=64) ms with α1 and α6 receptor block. Thus, both α1- and α6-containing GABA-A receptors controlled the number and the regularity of GrC spike discharge.
Mf-GrC release probability can be tuned by long-term synaptic plasticity (Sola et al., 2004; D'Errico et al., 2009) and regulate the rate of short-term depression and EPSP temporal summation (Nieus et al., 2006). In the model, GrC spikes at low release probability were delayed and less numerous compared to those at high release probability (Figure (Figure4A),4A), confirming the experimental observations obtained during LTP and LTD recordings in slices and in vivo (Nieus et al., 2006; Roggeri et al., 2008).
Once considering the whole GrC-GoC circuit, the relative timing of GrC and GoC activity became evident (Figure (Figure4B).4B). In response to an input burst, the GrCs and GoCs coactived by the same mf set fired almost at the same time (it should be noted that the reaction time of GrCs and GoCs conforms to experimental measurements: D'Angelo et al., 1995; Kanichay and Silver, 2008; Cesana et al., 2009). An additional delay was due to the time spent at the GoC-GrC synapse for GABA release and for the opening of GABA channels. This allowed the GrCs to generate a short spike burst before being inhibited.
These observations indicate that the mechanisms and reaction times of the circuit are appropriate to implement spatio- temporal filtering according to the time-window hypothesis (see also Figures Figures22 and and3;3; D'Angelo and De Zeeuw, 2009).
A main organizing principle deriving from MEA (Mapelli and D'Angelo, 2007) and VSD (Mapelli et al., 2010a,b) experiments is that activation of the mf bundle generates alternated areas of excitation and inhibition, which, on average, are organized in a center-surround manner. The origin of this effect has been attributed to the larger extension of the GoC inhibitory field compared to the GoC input through the basal dendrites, but a single center-surround structure was experimentally hard to isolate. To test the center-surround hypothesis, the model was activated through eight neighboring unbranched mfs with 500-Hz bursts composed of just two spikes (similar results were also obtained with longer bursts, data not shown). Mf stimulation caused a doublet of spikes forming a spot of activity that degraded around the core. Then, blocking inhibition increased the intensity and extension of the response. As explained in our previous paper (Mapelli and D'Angelo, 2007), the difference between the response before and after inhibition allowed to quantify the excitatory– inhibitory balance (E-I balance) (Figure (Figure5A,5A, right). The E-I balance was evaluated using the first peak to estimate excitation and the changes of the second peak caused by GABA receptor blockage to estimate inhibition (see Mapelli and D'Angelo, 2007, Materials and Methods and Figure Figure5B).5B). The spatial profile of the E-I balance revealed a Mexican-hat profile indicative of center-surround organization, in which inhibition overcomes excitation around the core (Figure (Figure55C).
The center surround organization bore consequences for the way spikes were transmitted through the granular layer. In the center, where the E-I balance is higher, the depolarization was faster and the inhibitory window closed more slowly, so that the GrCs emitted spike bursts with shorter delay, higher rate and longer duration than in the surround. This effect was particularly evident by comparing the PSTH generated by two exemplar GrCs, one in the center and the other in the surround (Figure (Figure66A).
Inhibition in the model respected two rules derived from experimental observations. Just one GoC axon was allowed to enter into a glomerulus and to form inhibitory synapses on all the GrC dendrites therein. Moreover, a GoC axon entering a glomerulus was prevented from accessing the neighboring glomeruli, which shared GrCs with the first glomerulus. This prevented a GrC from being inhibited twice through the same GoC (see Mapelli et al., 2009) and clustered inhibitory synapse belonging to closed mf-GrC-GoC-GrC and mf-GoC-GrC loops. The extent to which the center- surround organization depended on this synaptic organization of inhibition was assessed by comparison with a mesh-like GoC-GrC connectivity merely based on statistics. In the mesh configuration, late spikes were more common and the time-window effect was less pronounced (Figure (Figure6B).6B). The core remained more excited after the first spike, while the surround appeared less affected (Figure (Figure6C).6C). Thus, time-window and the center-surround effects generated by the feed-forward inhibitory loop were enhanced by the specific topology of inhibitory connections.
These simulations suggest that the center-surround organization benefits of the specific connectivity of the inhibitory synapses and allows for a selective control of spike transfer through the core, as further considered below.
The mfs can discharge at different frequencies (Kase et al., 1980; Van Kan et al., 1993, 1994; Arenz et al., 2008). In the model, the network was stimulated using bursts of five impulses activating eight neighboring mfs repeated at frequencies between 5 and 500Hz (Figure (Figure7A;7A; see Materials and Methods). Raising stimulation frequency from 5 to 500Hz increased temporal summation (Figure (Figure7B)7B) and spike generation in GrCs. As a consequence, the response increased following a sigmoidal gain curve with cut-off frequency around 100Hz (Figure (Figure7C).7C). The frequency- dependent enhancement of excitation matched the high-pass filtering properties of the granular layer observed experimentally (Mapelli et al., 2010b).
The granular layer transmission properties strictly depended on specific synaptic receptors (Figure (Figure7C).7C). When NMDA receptors were blocked, the transmission curve was selectively depressed at low-frequency (below 20Hz, where EPSP temporal summation was more critically dependent on NMDA receptors; cf. D'Angelo et al., 1995) but the high-pass filtering properties were maintained. When GABA-A α6 receptors were blocked, the transmission curve was enhanced especially at low frequency, consistent with the slow time constant of these receptors. When both GABA-A α6 and GABA-A α1 receptors were blocked the transmission curve was enhanced at all frequencies and high-pass filtering was strongly reduced. These results, which are in keeping with experimental observations obtained with VSD imaging (Mapelli et al., 2010b), indicate that the high-pass filtering properties of granular layer depend on phasic GABA-A receptor-dependent transmission (no further changes were obtained by blocking the tonic GABA-A receptor-mediated current, notshown).
The granular layer transmission properties also depended on glutamate release probability at the mf-GrC synapse (Figure (Figure7C).7C). Reducing release probability from the control value 0.6 to 0.2 depressed temporal summation yielding an effect similar to blocking NMDA receptors. Raising release probability to 0.9 enhanced transmission at low-frequency but not at high-frequency (probably because EPSP short-term depression became so strong that temporal summation was prevented).
The frequency dependence of the granular layer response changed moving from the center to the surround of the responding area (Figure (Figure7D).7D). In the center, the high-pass filtering curve arose at lower frequencies and attained a higher maximum gain than in the surround. Moreover, the gain difference between center and surround was more pronounced at high than low input frequency (the ratio between gain measured at 10 and 25μm from the core was 3 at 300Hz, 2.3 at 100Hz and 1.5 at 50Hz; Figure Figure7E).7E). Thus, the center-surround is predicted to generate complex transformations of incoming mf signals.
In vivo, GrCs show rare spontaneous activity (Chadderton et al., 2004; Jörntell and Ekerot, 2006; Rancz et al., 2007). However, in certain circumstances, the ensemble activity generated by the granular layer can take the form of coherent oscillations [7Hz in the rat (Hartmann and Bower, 1998) and 7–25Hz in the monkey (Pellerin and Lamarre, 1997; Courtemanche et al., 2009)]. It was not clear if this sparse GrC background activity could sustain the coherent oscillations.
In these simulations, the model was driven by random mf activity at different frequencies (see Materials and Methods) generating a sparse low-frequency discharge of GrCs (<1Hz/cell; Figure Figure8A)8A) reflecting the basal GrC activity recorded in the anesthetized rat (Chadderton et al., 2004; Rancz et al., 2007) and in the decerebrated cat (Jörntell and Ekerot, 2006). When basal mf activity was low (10Hz) ensemble activity was barely detectable but when basal mf activity was higher (20–40Hz) the granular layer generated coherent oscillations at 7–15Hz. Coherent oscillations were never easily resolved in single GrCs, which maintained a sparse activity, but clearly emerged as a population activity in the spikegrams.
A measure of rhythmicity was provided by the autocorrelograms of groups of GrCs and GoCs, which showed firing at regular periods over hundreds of milliseconds. The cross-correlograms between GrCs and GoCs cells (Figure (Figure8B)8B) also showed coherent oscillations, indicating that activity in the two cell populations were correlated. The coherent granular layer oscillations followed those of GoCs with a 5- to 10-ms lag due to the time spent in the feed-back inhibitory loop.
The granular layer oscillation assumed a frequency correlated with that in the mf input, as revealed by power spectrum analysis (Figure (Figure8C).8C). The output frequency was typically in the 7–15Hz range remaining lower than the input frequency. The oscillation was regulated by several synaptic mechanisms operating in the cerebellar glomerulus (Figure (Figure8Di;8Di; cf. Figure Figure4).4). Blocking GrC NMDA receptors strongly reduced the PSD peak amplitude, while blocking GrC GABA-A α6 receptors markedly increased the PSD peak amplitude, reflecting opposite regulation on the feed-back (GrC-GoC-GrC) loop. Blocking both GrC GABA-A α6 and GABA-A α1 interrupted the feed-back loop and flattened the PSD curve. Another remarkable effect was generated by altering mf-GrC release probability and therefore EPSP temporal summation. Decreasing mf-GrC release probability (p=0.2) reduced the PSD peak amplitude, while increasing mf-GrC release probability (p=0.9) increased the PSD peak amplitude, again reflecting opposite regulation on the feed-back loop. Moreover, (Figure (Figure8Dii),8Dii), when the feed-back loop was enhanced (GABA-A α6 receptor block or p=0.9) the PSD peak frequency was increased, while the opposite occurred when the loop was weakened. Therefore, synaptic mechanisms allowed to fine-tune the intensity and frequency of oscillations.
It should be noted that increasing the strength of the mf-GoC synapse (Figure (Figure8Ei)8Ei) and of the SC/BC-GoC synapse (Figure (Figure8Eii)8Eii) progressively reduced the PSD of oscillations. Thus, oscillations were critically dependent on circuit components that controlled the feed-back inhibitory loop.
This computational investigation provides a formal demonstration that available knowledge on cellular properties and circuit connectivity can explain center-surround and time-window effects during burst transmission and that sparse GrC random activity can sustain coherent low-frequency oscillations (D'Angelo and De Zeeuw, 2009; D'Angelo et al., 2009). The model predicts differential transmission through the center-surround structure reflecting the balance between NMDA and GABA-A receptor activation. The center had a lower cut-off for input signals and generated spike bursts at higher frequency and with shorter delay than the surround. In the center the time–window closed less effectively and contained more spikes than in the surround. The higher mf-GrC release probability in the center than in the surround due to long-term synaptic plasticity (Mapelli and D'Angelo, 2007) would further enhance the contrast between adjacent network areas. The main prediction of the model is therefore that the granular layer circuit behaves as a spatio-temporal filter, whose local properties can be adapted through long-term plasticity and different areas can be synchronized through low-frequency oscillations. It can be envisioned that distributed inputs generate a continuum of center-surround structures performing a complex time- and frequency-dependent transformation of input signals (e.g. Mapelli et al., 2010a) to be relayed to Purkinje cells. These results support the concept that input mossy fiber bursts are first separated and transformed in granular layer sub-circuits and then relayed to Purkinje cells for subsequent integration and pattern recognition (Dean et al., 2010).
The major innovation in this model is that membrane mechanisms are reproduced using Hodgkin–Huxley gating schemes and synaptic transmission using synaptic vesicle cycling schemes, which are based on extended electrophysiological experiments and biophysical analysis in slices (D'Angelo et al., 2001; Nieus et al., 2006; Solinas et al., 2007a,b; Diwakar et al., 2009 and references therein). Moreover, circuit structure is reproduced beyond statistical connectivity, accounting for a sophisticated arrangement of neuronal and synaptic elements (see Mapelli and D'Angelo, 2007; D'Angelo, 2008; D'Angelo and De Zeeuw, 2009; D'Angelo et al., 2009; Mapelli et al., 2009, 2010a). Finally, the simulations accounted for granular layer responses to three major mf discharge patterns inspired by in vivo recordings: localized bursting, frequency-modulated activity and diffused random activity (Kase et al., 1980; Chadderton et al., 2004; Jörntell and Ekerot, 2006; Rancz et al., 2007). In response to single input bursts, the model showed center–surround responses and time windowing (D'Angelo and De Zeeuw, 2009). In response to input trains at different frequencies, the granular layer behaved as a high-pass filter with a rapid growth of the response between 50 and 100Hz, as observed in VSD recordings (Mapelli et al., 2010b). In response to continuous random stimuli, the model generated synchronous oscillations in the theta band (Pellerin and Lamarre, 1997; Hartmann and Bower, 1998; Maex and De Schutter, 1998; Courtemanche et al., 2009).
The model, by being reconstructed from a detailed description of its constitutive elements, could account for the multiple dynamics of granular layer activity observed in vitro and in vivo. The individual GrCs and GoCs showed responses to mf burst stimulation compatible with those elicited in vivo following punctuate stimulation of the whisker pad (Vos et al., 1999; Chadderton et al., 2004; Rancz et al., 2007) or following more prolonged touching on the limbs (Jörntell and Ekerot, 2006). The specific firing patterns depended on the blend of local excitation and inhibition. Moreover, the neurons, although individually maintaining a low and irregular firing rate, could be entrained into coherent oscillatory activity. Therefore, the model, in addition to demonstrate that biophysical properties in vitro are sufficient to reproduce multiple activity patterns in vivo, provides the basis for investigating the contribution of single neurons to network activity and, conversely, network's influence on single neurons (Buzsáki, 2006; Izhikevich and Edelman, 2008).
Previous models [originally the theoretical models of Marr (1969), Albus (1971), Tyrrell and Willshaw (1992), Dean et al. (2010) but also the computational model of Maex and De Schutter (1998), Medina and Mauk (2000)] were based on the statistical properties of the granular layer connectivity. Here we have faced the problem of how GoCs contact GrCs in the glomeruli by using the following rules: a GoC inhibits all the dendrites in a glomerulus and cannot inhibit a GrC more than once. As a corollary, each GrC is inhibited by different (up to 5 with a mean of 4) GoCs and all GrCs sending their dendrites in the same glomerulus have at least one GoC in common. This organization, by increasing the efficiency of feed-forward loops and concentrating inhibition into certain cell subgroups, enhanced center-surround transmission and time windowing. These rules are compatible with original Golgi staining images (Golgi, 1883; Ramón y Cajal, 1887, 1889, 1904; Eccles et al., 1967; Palay and Chan-Palay, 1974) but would require further refinement of the glomerular representation in the model to improve generation of direct and indirect GABAergic responses (Rossi and Hamann, 1998; Mapelli et al., 2009) and to implement glomerular crosstalk (Mitchell and Silver, 2000a,b). It should also be noted that network topologies were not fully resolved in the model, because mf and GoC axon branching on the sagittal plane and pf long-range connections on the transverse plane could not be implemented on the scale of this network (see Wu et al., 1999; Sultan, 2001; Barmack and Yakhnitsa, 2008; D'Angelo, 2008).
The model identified different roles for the main components of inhibition. Lateral inhibition was critical to determine the center-surround effect, feed-forward inhibition to determine the time-windowing effect and feed-back inhibition to determine coherent oscillations. In fact, in the model, the generation of oscillatory activity in response to random mf bursts decreased with the strength of the feed-forward loop, which, in turn, was essential to generate time-windowing in response to localized mf activity. Since the relative strength of the feed-forward and feed-back loops is unknown, different mechanism can be envisaged: (i) the relative strength of feed-back and feed-forward loops is dynamically balanced through the intervention of SC/BC-GoC inhibition (Dumoulin et al., 2001) or through glomerular mechanisms (Mitchell and Silver, 2000a,b, 2003; Mapelli et al., 2009); (ii) the feed-forward loop is independently regulated through pfs originating from GrCs located outside the active beam; (iii) oscillating activity is conveyed into the cerebellum in pre-organized patterns. Although these mechanisms may coexist, the latter is especially interesting since both GrCs and GoCs have resonant properties in the theta-band (D'Angelo et al., 2001; Solinas et al., 2007a,b; D'Angelo and De Zeeuw, 2009). It is therefore possible that theta patterns at the input (around 7Hz in rodents; Hartmann and Bower, 1998) are particularly efficient in generating granular layer responses (Ros et al., 2009).
The model appropriately reproduced the high-pass filtering properties of the granular layer, so that burst patterns over 50Hz were efficiently transmitted while those at lower frequencies were not (Mapelli et al., 2010b). This is in keeping with the proposal that GrCs receive and retransmit high-frequency bursts (Chadderton et al., 2004; Rancz et al., 2007). The higher cut-off observed in the surround than in the center suggests that a frequency-code is spatially implemented through the center-surround structure. The gain control mechanisms based on stochastic resonance (Mitchell and Silver, 2003) may overlay with those considered here and account for the frequency-dependent modulation of incoming signals observed in certain functional conditions (Arenz et al., 2008).
In the model, generation of appropriate spatio-temporal patterns of activity of the cerebellum granular layer required a detailed description of cellular and synaptic properties. The time course of vesicle cycling regulated temporal summation during repetitive neurotransmission (see Figures Figures22 and and4)4) and filtering at the mf-GrC relay (see Figure Figure7).7). Transmission at low frequency was enhanced by the slow time constant of excitation introduced by the NMDA receptors and was reduced by the slow time constant of inhibition introduced by GABA-A α-6 receptors (see Figure Figure7),7), suggesting that the balance of the two is critical to regulate circuit functioning. These same mechanisms also helped maintaining network oscillations in the theta-frequency band (see Figure Figure8;8; D'Angelo et al., 2001; Solinas et al., 2007b). It should be noted that, with unconstrained parameterization, a previous model generated 40Hz oscillations, which are probably non-physiological (Maex and De Schutter, 1998).
In order to assess the hypothesis that the granular layer behaves as an adaptable spatio-temporal filter coordinated by low-frequency oscillations, experimental tests may be combined with further computational investigations. First, the implementation of rules for long-term synaptic plasticity in the model of the mf-GrC synapse (D'Errico et al., 2009) could allow investigating the impact of adaptation on network filtering and pattern recognition. Secondly, the ability of the circuit of generating theta-frequency oscillations in response to repetitive bursting (which would be naturally conveyed by the cerebral cortex; Ros et al., 2009) could be investigated by modulating the resonant properties of GrCs and GoCs models (D'Angelo et al., 2001; Solinas et al., 2007b). Thirdly, the observation that tonic inhibition can control granular layer gain (Mitchell and Silver, 2003) could be assessed by introducing quantal mechanisms of neurotransmitter release in the model of the mf-GrC synapse (Arleo et al., 2010). The integration of the current model into larger modules including the molecular layer and deep cerebellar nuclei will eventually provide a tool for investigating the spatio-temporal filtering hypothesis of the entire cerebellar network (Dean et al., 2010).
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
This work was supported by projects SENSOPAC (FP6-IST 028056) and CYBERRAT (Bio-ICT convergence 216528) of the European Commission and by NEUROIMAGE of CNISM (Consorzio Interuniversitario per le Scienze Fisiche della Materia) to E.D. The computational facility for the most intensive simulations was provided by the parallel cluster CASPUR (Consorzio Interuniversitario per le Applicazioni di Supercalcolo per Università e Ricerca). We thank Dr. S. Masoli for his contribution to parallelization of themodel.
GrC, granule cell; GoC, Golgi cell; SC/BC, stellate cell/basket cell; MLI, molecular layer interneuron; mf, mossy fiber; pf, parallel fiber.