Abstract
The cerebellar cortex computes sensorimotor information from many brain areas through a feedforward inhibitory (FFI) microcircuit between the input stage, the granule cell (GC) layer, and the output stage, the Purkinje cells (PCs). Although in other brain areas FFI underlies a precise excitation versus inhibition temporal correlation, recent findings in the cerebellum highlighted more complex behaviors at GC–molecular layer interneuron (MLI)–PC pathway. To dissect the temporal organization of this cerebellar FFI pathway, we combined ex vivo patch-clamp recordings of PCs in male mice with a viral-based strategy to express Channelrhodopsin2 in a subset of mossy fibers (MFs), the major excitatory inputs to GCs. We show that although light-mediated MF activation elicited pairs of excitatory and inhibitory postsynaptic currents in PCs, excitation (E) from GCs and inhibition (I) from MLIs reached PCs with a wide range of different temporal delays. However, when GCs were directly stimulated, a low variability in E/I delays was observed. Our results demonstrate that in many recordings MF stimulation recruited different groups of GCs that trigger E and/or I, and expanded PC temporal synaptic integration. Finally, using a computational model of the FFI pathway, we showed that this temporal expansion could strongly influence how PCs integrate GC inputs. Our findings show that specific E/I delays may help PCs encoding specific MF inputs.
SIGNIFICANCE STATEMENT Sensorimotor information is conveyed to the cerebellar cortex by mossy fibers. Mossy fiber inputs activate granule cells that excite molecular interneurons and Purkinje cells, the sole output of the cerebellar cortex, leading to a sequence of synaptic excitation and inhibition in Purkinje cells, thus defining a feedforward inhibitory pathway. Using electrophysiological recordings, optogenetic stimulation, and mathematical modeling, we demonstrated that different groups of granule cells can elicit synaptic excitation and inhibition with various latencies onto Purkinje cells. This temporal variability controls how granule cells influence Purkinje cell discharge and may support temporal coding in the cerebellar cortex.
Introduction
In the cerebellar cortex, granule cells (GCs) and molecular layer interneurons (MLIs) define a feedforward inhibitory (FFI) pathway that triggers a reliable excitation/inhibition (E/I) sequence in Purkinje cells (PCs). Direct GC stimulation leads to a very low temporal dispersion between EPSCs and IPSCs in PCs (Eccles et al., 1967; Brunel et al., 2004; Mittmann et al., 2005; Grangeray-Vilmint et al., 2018). This fast and almost invariant sequence is favored by the beam-like arrangement of parallel fibers (PFs), the axon of the GCs, which run perpendicularly to MLIs and PC dendrites in the mediolateral axis. This orthogonal organization, combined with the fact that MLIs are (1) contacted by multiple beams of PFs and (2) inhibit PCs located in the near vicinity along the sagittal plane, underlies both lateral and feedforward inhibition. (Palay and Chan-Palay, 1974; Jörntell et al., 2010; Kim and Augustine, 2021). The cerebellar cortex is organized as an array of parasagittal computational units called microzones in which PCs share similar receptive fields (Apps and Hawkes, 2009; Apps et al., 2018). Therefore, GC axons implement a classical FFI by contacting both MLI and PC dendrites within a given microzone, leading to the modulation of PCs. Several studies demonstrated that FFI sharpens the temporal window for spike discharge, prevents PC discharge saturation, and promotes long-term plasticity at GC–PC synapses, thus extending the dynamic range for GC input encoding (Eccles et al., 1967; Barbour, 1993; Brunel et al., 2004; Mittmann et al., 2005; Isaacson and Scanziani, 2011; Binda et al., 2016; Rowan et al., 2018). However, most of these studies were performed by stimulating either bulks of parallel fibers or GC somas, whereas under physiological conditions FFI is triggered by sensorimotor inputs carried by mossy fibers (MFs), the latter exciting discrete and spatially organized clusters of GCs (Valera et al., 2016; Spaeth and Isope, 2023; Spaeth et al., 2022).
MFs, which originate in precerebellar nuclei in the brainstem and spinal cord, convey sensorimotor, proprioceptive, and vestibular information to the cerebellum. A single MF gives rise to numerous collaterals and contacts many GC clusters in a given lobule (Fig. 1A; Heckroth and Eisenman, 1988; Shinoda et al., 2000; Voogd et al., 2003; Quy et al., 2011). This arrangement leads to spatially and temporally dispersed GC activation, as observed in lobules III–V and IX–X of the cerebellar cortex (Chabrol et al., 2015; Valera et al., 2016; Spaeth et al., 2022). GC clusters make functional synaptic connections with specific groups of PCs or MLIs because of activity-dependent processes (Isope and Barbour, 2002; Valera et al., 2016; Spaeth et al., 2022). Previous studies showed that (1) several and distant GC clusters target the same PC, and (2) PCs and MLIs located in the same microzone have different GC–PC synaptic connectivity maps (Valera et al., 2016). Therefore, the temporal organization of E/I sequences elicited by single MFs might be more complex than the classical FFI. Indeed, in vivo recordings showed that sensory inputs elicit multiple temporal combinations of E/I sequences in PCs (Santamaria et al., 2007; Jelitai et al., 2016; Brown and Raman, 2018). These combinations include classical FFI sequences or inhibition preceding excitation supporting the hypothesis that PCs may combine separate excitatory and inhibitory synaptic inputs originating from spatially distinct GCs clusters, which may lead to a wide range of E/I delays when reaching PCs (Fig. 1A, different scenarios). We therefore investigated the functional basis of these combinations by coupling optogenetic MF stimulation and ex vivo cerebellar recordings. Channelrhodopsin2 was expressed in a subset of MF terminals, allowing selective MF activation and subsequent measurement of E/I delays in recorded PCs. We provide a direct demonstration that the GC-MLI-PC FFI undergoes temporal expansion beyond classical FFI. Finally, using a computational approach, we suggest that this arrangement expands PC coding capacity, which would further increase its ability to process and integrate multiple sensorimotor inputs, strengthening the ability of the cerebellum to contribute to the performance of various and diverse tasks and behaviors in different contexts.
Materials and Methods
All experiments were conducted in accordance with the guidelines of the Ministère de l'Education Supérieure et de la Recherche and the local ethical committee, the Comité Régional en Matière d'Expérimentation Animale de Strasbourg under the agreement delivered to the animal facility Chronobiotron (Unité Mixte de Service 3415, University of Strasbourg).
Precerebellar nuclei recombinant adeno-associated viral–mediated transduction
In vivo stereotaxic injections of recombinant adeno-associated viral (rAAV) particles were performed as previously described (Valera et al., 2016). CD1 male mice, postnatal day (P)21, were anesthetized by a brief exposure to isoflurane 4%, and anesthesia was maintained by intraperitoneal injections of a mixture of ketamine (100 mg/kg), medetomidine (1 mg/kg), and acepromazine (3 mg/kg). rAAV 9/2 particles carrying the cDNA for ChRd2(H134R)-YFP under the hSyn promoter (3.38 1013 GC/ml; Penn Vector Core) were unilaterally injected in the dorsal column nuclei (DoCN e.g., external cuneate nucleus) at an approximate speed of 250 nl/min via a graduated pipette equipped with a piston for manual injections. A final volume of 1.5 µl was delivered by two injections (0.75 µl/injection) separated by 0.2 mm in the anteroposterior direction; after that half of the virus volume was delivered, the pipette was raised up 0.2 mm and maintained in place until the end of the injection. For effective virus diffusion, the pipette was left in place at least 5 min following injection. Injections coordinates were determined using Franklin and Paxinos (2007), and corrections based on tissue markers were applied to counterbalance the variability of the CD1 outbred background (from λ, starting point, AP, 2.73 ± 0.14; L, 1.36 ± 0.08; DV, 5.1 ± 0.09; mean ± SEM). At the end of the injection, antipamezole (1 mg/kg) was administered to the mice via intraperitoneal injection to aid recovery from anesthesia.
Slice preparation
Acute cerebellar transverse slices were prepared from injected mice 3–4 weeks after virus injection (>P62). Mice were anesthetized by a brief exposure to isoflurane 4%, decapitated, and the cerebellum rapidly dissected in ice-cold artificial cerebrospinal fluid (aCSF) bubbled with carbogen (95% O2, 5% CO2) and containing the following (in mm): 120 NaCl, 3 KCl, 26 NaHCO3, 1.25 NaH2PO4, 2 CaCl2, 1 MgCl2, and 16 glucose. Transverse slices 300 µm thick were cut (Microm HM 650 V microtome, Microm) in ice-cold sucrose-based cutting solution bubbled with carbogen (95% O2, 5% CO2) and containing the following (in mm): 246 sucrose, 4 KCl, 26 NaHCO3, 1 CaCl2, 5 MgCl2, 10 glucose, and 1 kynurenic acid. Slices were then transferred in bubbled aCSF at 34°C, and they were allowed to recover for at least 30 min before starting experiments. Following recovery, slices were maintained at room temperature for the rest of the day.
Electrophysiology
Electrophysiology experiments were performed at room temperature in a recording chamber continuously perfused with bubbled aCSF supplemented with the following (in mm): 0.001 strychnine, 0.05 d-APV, 0.0005 DPCPX, 0.001 AM251, 0.001 CGP52432, and 0.002 JNJ16259685. To allow the full illumination of the surface, slices were rapidly mounted on glass coverslips coated with poly-l-lysine (1 mg/ml) right before transferring them to the recording chamber.
Patch-clamp pipettes were pulled from borosilicate glass to a final resistance of 3–4 MΩ when filled with the following solution (in mm): 135 cesium methanesulfonate, 6 NaCl, 1 MgCl2, 10 HEPES, 4 MgATP, 0.4 Na2GTP, 1.5 EGTA, 5 QX314Cl, adjusted to pH 7.3 and 295 mOsm.
Whole-cell patch-clamp recordings were performed with a MultiClamp 700B amplifier (Molecular Devices) and acquired with WinWCP 4.2.1 freeware (John Dempster, Strathclyde Institute of Pharmacy and Biomedical Sciences, University of Strathclyde). Whole-cell currents from PCs were filtered at 2 kHz and digitized at 50 kHz, and series resistance was monitored during the experiments and compensated by 70–80%. Synaptic currents were recorded at various holding potentials (−60 mV, −50 mV, and 0 mV) to isolate excitation and inhibition. Junction potential was not corrected and was estimated at 10 mV (e.g., Vh = −60 mV corresponds to around −70 mV in the cell).
Loose cell-attached recordings were obtained using a MultiClamp 700B amplifier (Molecular Devices) and acquired with WinWCP 4.2.1 freeware. Rosettes were recorded with 5 MΩ glass pipettes (borosilicate), and potential was held at 0 mV for all recordings. The internal pipette solution contained the following (in mm): 120 NaCl, 3 KCl, 10 HEPES, 1.25 NaH2PO4, 2 CaCl2, 1 MgCl2, and 10 glucose (Sigma-Aldrich). Osmolarity and pH were set at 295 mOsm and 7.3, respectively. Recordings were low-pass filtered at 2.6 kHz then sampled at 20–50 kHz. All experiments were performed at room temperature (23°C) using the same bubbled aCSF as for slices preparation. For controls, NBQX (20 μm; Sigma-Aldrich) and lidocaine (1 mm; Sigma-Aldrich) were bath applied, respectively, to block AMPA receptor transmission and action potential propagation.
For GC stimulation, a glass pipette (borosilicate) containing the following (in mm): 120 NaCl, 3 KCl, 10 HEPES, 1.25 NaH2PO4, 2 CaCl2, 1 MgCl2, and 10 glucose (Sigma-Aldrich) was inserted in the deepness of the GC layer. GCs were electrically stimulated (3–5 µA, 100 µs) using a constant current delivered by an isolated stimulator (model DS3, Digitimer).
Photostimulation
Surface mossy fiber blue light-mediated activation was obtained by a two-dimension scan mode of a confocal microscope (FV300, Olympus), equipped with a diode-pumped solid-state blue laser (473 nm, CrystaLaser; stimulation, 50 ms blue light pulses through a 20× objective, 83 × 83 µm scan area), driven by a Programmable Acquisition Protocol Processor (FluoView 300) at successive positions along the longitudinal axis. The stimulation protocol was repeated between three and seven times, and average traces were used for analysis. MFs-dependent EPSCs and IPSCs were isolated by clamping Purkinje cells (voltage command) at –60 mV and 0 mV, respectively.
Unitary stimulations (10 ms duration) for single illumination protocol and loose-cell recordings of MF terminals were performed using a 460 nm blue LED (Prizmatix). The illuminated area was restricted to a single rosette using a Digital Micromirror Devices array (Mosaic, Andor Technology) through a 40× objective (Olympus). Steady illumination power was set at 2–4 mW/mm2. The stimulation protocol was repeated from 10 to 20 times, and average traces were used for analysis.
Simulation
We used a previous model of the GC–MLI–PC pathway that we published in Grangeray-Vilmint et al. (2018) to study the effect of E/I delay and input frequency on the PC output. The model was simulated using the Neural Simulation Tool environment (http://www.nest-simulator.org/).
PC neuron
The Purkinje cell was modeled as a conductance-based point neuron. The membrane voltage dynamics of the PC was given by the following:
Synapse model
Synapses were modeled as a conductance transient as follows:
Model parameters
Short-term dynamics of synapses
To model the short-term dynamics (STD) of the PSP amplitude we used a deterministic model proposed by Tsodyks and Markram (1997) as follows:
The effective synaptic weight was given by the following,
To systematically change the nature of the synapses from facilitating to depressing, we changed the variable Ua while keeping all other variables unchanged. Because we kept all the time constants unchanged in the model, going from facilitation to depression did not affect the frequency dependence of the synapses. The parameter Aa was changed to increase the amplitude of the postsynaptic potential independent of the STD. The values of these various parameters are provided in Table 1.
Data analysis
Electrophysiological recordings were analyzed with custom scripts and routines written in Python 3.9 using the following packages: Pandas 1.3, Scipy 1.6, NumPy 1.19, and Neo 0.9. Complete code is available at https://github.com/ludo67100/MFDeltaLat.
Both synaptic charges and latencies were computed on averaged signals (3–20 stimulations). In experiments at Vh = 0 mV, spontaneous inhibitory currents were recorded as MLIs remain spontaneously active in acute slices. We therefore excluded traces in which spontaneous IPSCs were recorded near the stimulation onset (i.e., ±20 ms from stimulation onset).
The latency was defined as the time difference between stimulation onset and the intersection between a linear fit of the current rising phase and the baseline of the signal (Fig. 2E). ΔLat (i.e., IPSC latency minus EPSC latency) was computed by subtracting IPSC latency from EPSC latency in each experiment. Synaptic charges were measured as the integral of the signal in a 100 ms time window starting from the onset of light stimulation.
For spike gain, we binned the spiking activity across all the 200 trials to obtain a peristimulus time histogram (PSTH; bin width = 2 ms). Next, we measure the mean of the PSTH in the prestimulus duration (i.e., in the interval 50–200 ms). Next, we calculated spike gain as mean and sum of the normalized PSTH (see Fig. 4A) as follows:
We normalize with 200 as we have 200 bins of 2 ms each. We have plotted mean norm, spike gain, but the results are same if we plot sum norm, spike gain (see Fig. 4). We further normalized these spike gain measures with respect to the spike gain measured as E/I delay of +2 ms (see Fig. 4B).
For spike grain trajectory distance, we have two spike gains, one for the three stimulations and one for seven stimulations. We can represent the neuron output as a point in a two-dimensional space spanned by spike gain for three stimulations (x-axis) and seven stimulations (y-axis; see Fig. 5A). To quantify the effect of E/I spike latency, we measured the Euclidian distance between neuron output for two different E/I latencies as follows:
Entropy of the distances
Let dij(f) where i = 1:1:11 and j = i + 1 is the set of all distances for a specific frequency (see above). We first estimated the count histogram of all dij (f). This histogram was normalized to have a unit area. Now we treated this histogram as a probability distribution and estimated the entropy using Shannon's formula as follows:
Statistics
Normality and homoscedasticity of data distributions were assessed with Shapiro–Wilk and Levene statistical tests, respectively. If both criteria were met, the corresponding distributions were compared with independent t test. In other cases, we used the Mann–Whitney U test. Statistical tests were performed with the corresponding package from the scipy.stats module. Alternative and method argument were respectively set to two-sided and auto unless otherwise reported. The p values and test results are reported either in figures and/or in text when applicable. For the ΔLat measurement following direct electrical GC stimulation, the dataset recorded for this experiment has been combined with the dataset from Grangeray-Vilmint et al. (2018).
Results
Stimulating specific MFs and individual rosettes
To physiologically recruit the GC–(MLI)–PC FFI pathway in the cerebellar cortex, we expressed Channelrhodopsin2 in individual MFs by injecting an rAAV 9/2-hSyn-ChRd2(H134R)-YFP in the dorsal column nuclei (DoCN; e.g., external cuneate nucleus), a subset of precerebellar nuclei (Fig. 1B; number of recordings, n = 112; number of cells, n = 28; number of mice, N = 8), which convey proprioceptive and exteroceptive information from the upper limbs. MFs originating in the DoCN send collaterals to many different locations in a given lobule (Quy et al., 2011; Valera et al., 2016), which may activate different groups of GCs. We then photostimulated MFs in acute cerebellar slices while recording PCs close to the lobule midline (lobule III–VIII) using whole-cell patch clamp (Fig. 1A,C; see above, Materials and Methods). Two strategies of photostimulation were used, consisting of (1) a laser scanning method or (2) a patterned illumination (see above, Materials and Methods). The laser scanning method allowed us to illuminate small surfaces of the GC layer (83 × 83 µm, hereafter referred to as surface illumination, 50 ms duration) leading to the activation of many MF rosettes (Fig. 1C). This method mimics the desynchronized activation of several MFs that occurs in vivo, as seen, for example, with the convergence of trigeminal and somatosensory cortical inputs (Ishikawa et al., 2015; Shimuta et al., 2020). Alternatively, we used patterned illumination (see above, Materials and Methods) to selectively illuminate a single rosette (hereafter referred to as single illumination; Fig. 1C). We first addressed whether blue light enables individual MF rosette excitation. YFP fluorescence allowed us to visualize and record MF rosettes in loose cell-attached mode (Barbour and Isope, 2000; Fig. 1D). In all recorded rosettes, illumination triggered a direct depolarization eliciting a unique and reliable action potential (mean delay ± SD, 1.63 ± 0.31 ms; mean jitter ± SD, 0.023 ± 0.03 ms; n = 5; Fig. 1D,E). Application of NBQX and Lidocaine isolated the direct depolarizing current elicited during the length of the light pulse (mean amplitude ± SD, 15.6 ± 9.8 pA, n = 5; Fig. 1D). Therefore, rosette illumination likely elicited reliable transmitter release at the MF–GC synapses (Jackman et al., 2014).
Optogenetic MF stimulation elicits EPSCs and IPSCs in PCs. A, Left, Diagram of the cerebellar cortex microcircuit illustrating how MF inputs may affect the timing of GC–PC and GC–MLI–PC inputs. Sensorimotor information (magenta) is conveyed to several GC clusters targeted by a given MF (green). Right, Such an organization may lead to independent EPSC and IPSC synaptic inputs yielding different types of E/I sequences in PCs. PCL, Purkinje cell layer; WM, white matter. B, Left, rAAVs 9/2 injection in the DoCN allowed Channelrhodopsin-2-YFP expression in MFs. Right, Confocal image of YFP fluorescence in MF rosettes from a transverse cerebellar slice. C, Left, Photostimulation protocols. In surface illumination, several MF terminals are activated sequentially in an 83 × 83 µm area. In single illumination, patterned light illumination selectively activates a single rosette. Top right, Example of acute slice under infrared or 460 nm light illumination. The large and small dashed-blue squares illustrate surface and single illumination, respectively. Scale bar, 40 µm. Bottom right, Example of IPSC (Vh = 0 mV) and EPSC (Vh = −70 mV) postsynaptic currents elicited in PCs following MF photostimulation. D, Example of a loose cell-attached recording of a rosette and corresponding current evoked by photostimulation in normal aCSF (see above, Materials and Methods) in the presence of NBQX or NBQX + Lidocaine (10 μm each). E, Left, Latency of the first spike from stimulation onset elicited in five different rosettes. Large and small markers show average and individual trials, respectively. Right, Variability in spike occurrence (jitter) measured in rosettes. The latency of each spike in every trial was subtracted from the average latency of all the trials (each color represents a different rosette). F, Distribution of excitatory and inhibitory synaptic charges evoked by photostimulation, in both surface (light gray) and single (orange) illumination protocols (EPSQ, U = 1601.0; IPSQ, U = 1645.0; Mann Whitney U test in both, p values in plot). G, Left to right, Absolute latency of EPSCs; absolute latency of IPSCs; ΔLat (i.e., Eonset − Ionset); excitatory and inhibitory synaptic charges sorted by the location of the stimulated MF rosettes within the granule cell layer. Rosettes located within ± 80 µm of the recorded PC axis (orange); distant rosettes (green); rosettes were in the upper/lower part of the granule cell layer (respectively, violet/blue). No significant difference, except between excitatory charges elicited by local versus distal rosettes; Mann Whitney U test, U = 548.0, p value = 0.0212). n.s. (i.e., p value > 0.05, from left to right, U = 476, p = 0.25; U = 410, p = 0.9; t = 1.51, p = 0.14; U = 392, p = 0.69; U = 475.0, p = 0.25; U = 395, p = 0.72; U = 420, p = 0.99; U = 308, p = 0.14; U = 412, p = 0.65; Mann Whitney U test or independent t test). White line, median; box edges, interquartile range; whiskers: minimum/maximum.
MF illumination elicited temporally dispersed E and I
To explore the variability of E/I delay variability on PCs, we recorded EPSCs and IPSCs in PCs on MF activation via single or surface illumination; EPSCs and IPSCs were recorded at Vm = −60 mV and Vm = 0 mV, respectively (i.e., close to reversal potential for inhibition and excitation, respectively; Fig. 1C, right) and MFs were activated by a single photostimulus repeated overtime (3–20 trials). To estimate the total synaptic weight transfer, we calculated the charge [excitatory synaptic charges (EPSQs) and inhibitory synaptic charges (IPSQs)] carried by these evoked currents (mean ± SD, surface illumination, EPSQ, −3.4 ± 3.1 pC; IPSQ, 13.4 ± 10.8 pC, n = 54 recordings, n = 17 cells; single illumination, EPSQ, −3.8 ± 4.6 pC; IPSQ, 15.7 ± 14.8 pC, n = 58 recordings, n = 11 cells; EPSQsingle vs EPSQsurface, U = 1601.0, p = 0.84; IPSQsingle vs IPSQsurface, U = 1645.0, p = 0.64; Mann–Whitney U test in each case; Fig. 1F). MF photostimulation elicited temporally scattered E/I pairs in the recorded PCs (surface illumination, mean latency excitation ± SD, 14.1 ± 3.6 ms; mean latency inhibition ± SD, 18.1 ± 4.3 ms; single rosette, mean latency excitation ± SD, 5.7 ± 1.6 ms; mean latency inhibition ± SD, 7.6 ± 2.3 ms; excitationsurface vs excitationsingle, U = 20.0, p = 2.27.10−19; inhibitionsurface vs inhibitionsingle, U = 11.0, p = 1.40.10−19; Mann–Whitney U test in both cases). We then sorted the data depending on the relative location of stimulated rosettes to the recorded PCs and found that neither latencies nor IPSQs varied with rosette position in the granule cell layer (GCL). However, the evoked EPSCs were larger when rosettes were located in close proximity to the recorded PCs than distant from them (local rosettes, EPSQs, −5.0 ± 4.7 pC; distal rosettes, EPSQs, −3.1 ± 4.4 pC; U = 548.0, p = 0.021, Mann–Whitney U test; Fig. 1G). Excitatory current latencies observed in PCs are determined by (1) MF rosette depolarization and synaptic release at MF–GC synapses, (2) integration time in GCs, (3) conduction time to GC synaptic boutons, and (4) synaptic release at GC–PC synapses. The latencies for inhibition include the additional GC–MLI synapses and the integration time in MLIs. As MLIs contact only PCs in close proximity (Palay and Chan-Palay, 1974; Kim et al., 2014), GC axons contacting both MLIs and PCs should lead to E/I pairs with short and constant relative latencies (i.e., classic FFI; Fig. 1A). However, when we quantified the difference in relative latencies between excitatory and inhibitory synaptic inputs recorded in PCs (ΔLat, i.e., IPSC latency minus EPSC latency), we observed a large temporal spread with single and surface MF illumination. In both protocols, ΔLat ranged from negative to positive values that can exceed ±5 ms, although the spread was more pronounced in surface illumination (surface illumination, from −3.14 to 21 ms, mean ± SD = 3.9 ± 4.0 ms, n = 54; single illumination, from −5.9 to 5.2 ms, mean ± SD = 1.8 ± 1.9 ms, n = 58, U = 1045.0, p = 0.0024, Mann–Whitney U test, Fig. 2A,B). This large variability in ΔLat, which is not explained by the relative position of rosettes within the GCL (Fig. 1G), led us to hypothesize that even when stimulating a single rosette yielding inputs from a single MF, an antidromic action potential invades the entire MF collaterals and several independent GC–PC and GC–MLI–PC pathways are activated at different locations in the GCL (Fig. 1A). Although elicited in a rosette, this antidromic stimulation would recapitulate the journey of the action potential along the MF under physiological conditions. To test this hypothesis, we assessed whether the wide range of recorded ΔLats is the result of multiple GC clusters excited by a MF or could be accounted for by the downstream GC–(MLI)–PC connections. We therefore directly stimulated individual groups of GCs using electrical stimulation (Fig. 2C; see above, Materials and Methods), and we observed short and almost invariant E/I ΔLat (GC_ΔLat = 2 ± 0.74 ms, mean ± SD, n = 23; Fig. 2C) as previously shown by many studies (Vincent and Marty, 1996; Brunel et al., 2004; Mittmann et al., 2005; Grangeray-Vilmint et al., 2018). These invariant ΔLats illustrate the classical FFI (Fig. 1A).
We then used these GC_ΔLats as a proxy for the classical FFI in the MF–GC–MLI–PC pathway. Following MF stimulation, we classified an E/I pair as a classical FFI when ΔLat was in the range GC_ΔLat ± 2 SDs (i.e., 0.52 ms < ΔLat < 3.48 ms, referred to as group 1) and as a nonclassical E/I sequence otherwise (referred to as group 2; Fig. 1A). Group 2 corresponds to FFI elicited by different groups of GCs, whereas in group 3, only inhibition was recorded (Fig. 2D,F). Therefore, we identified that the E/I balance in PC soma is frequently determined by independent clusters of GCs (surface illumination, group 2/3 = 68.8%; single illumination, group 2/3 = 46.8%; Fig. 2G).
MF stimulation elicits a wide range of ΔLat profiles in PCs. A, Histogram of the ΔLats following MF surface illumination. B, Histogram of the ΔLats following single rosette illumination. C, Histogram of the ΔLats following direct electrical stimulation of GCs (direct FFI, GC–ΔLat). The dotted black line represents the average, and gray lines illustrate the interval containing ΔLats considered as classic FFI (i.e., mean ± 2 SD: 0.52 < ΔLat < 3.48 ms). Inset, Diagram illustrating FFI obtained with direct excitation of GCs. D, Merged ΔLat distributions from MF surface illumination (light gray), single rosette illumination (red), and GCs electrical stimulation (blue). The data between dotted lines illustrate ΔLats considered as classic FFI. E, Left, Representative examples of E/I temporal profiles observed during single rosette and surface illumination. Group 1 corresponds to 0.52 < ΔLat < 3.48 ms, illustrating classic FFI. Group 2 corresponds to larger (delayed inhibition) or negative (inhibition first) ΔLat. Blue triangles represent stimulation onset. Insets, Onset of EPSC and IPSC superimposed. Magenta and orange dashed lines show linear fit of current onsets. EPSC (magenta) and IPSC (orange) latencies were measured as the intersection with baseline (black dashed line). F, Illustration of GC clusters eliciting only inhibition (group 3). Excitation only was never observed in our conditions. G, Proportion of the different scenarios observed in all recordings from single (left pie chart) and surface (right pie chart) illumination protocols. H, Illustration of E/I sequence recorded at intermediate potential (Vh = −50 mV, magenta), showing integration of E and I currents in cerebellar PCs following photostimulation of MF rosettes. I, Synaptic charges (at Vh = −50 mV) are negatively correlated with ΔLat (n = 13, slope = −1.45 ± 0.57, intercept = 7.05 ± 1.72, r = −0.61, p value in plot).
ΔLat variability is likely to affect dendritic integration of E/I sequence in PCs. To further evaluate the interactions between E and I, we recorded EPSC and IPSCs at an intermediate holding potential (Vh = −50 mV) in a subset of PCs (Fig. 2H). At this potential, both excitation and inhibition have significant driving force, which lead to reduced net synaptic currents that integrate both E and I. We then observed a negative correlation between ΔLat and total synaptic charges (n = 13, slope = −1.45 ± 0.57, intercept = 7.05 ± 1.72, r = −0.61, p = 0.026, linear least-square regression; Fig. 2I). These findings indicate that short or negative ΔLat minimizes excitation in PCs, whereas long ΔLat favors excitation. Hence, the relative timing between E and I (ΔLat) may control how E/I sequences affect PC discharge.
Excitatory and inhibitory synaptic weights in PCs are correlated
In the forebrain, E/I ratios are equalized (i.e., target neurons receive inhibitory inputs that are scaled to their excitatory ones) because of a particular spatial arrangement of excitatory and inhibitory synapses (Xue et al., 2014). We then sought to determine whether the organization of the MF–GC pathway would also favor homogeneous E/I ratios across PCs in the cerebellar cortex. We evaluated whether excitatory and inhibitory synaptic charges elicited by MF stimulation onto PCs were correlated. In both group 1 and group 2 E/I sequences, EPSQs were linearly correlated with the IPSQs (group 1, slope = 3.2 ± 0.4, intercept = 5.2 ± 1.4, r = 0.72, n = 53, p = 9.2.10−10; group 2, slope = 2.6 ± 0.2, intercept = 3.5 ± 1.6, r = 0.82, n = 59, p = 7.1.10−16; groups 1 and 2, slope = 2.6 ± 0.2, intercept = 5.2 ± 1.0, r = 0.79, n = 112, p = 9.7.10−26; linear least-square regression; Fig. 3A–C, black curves). Thus, independently of whether E/I sequences are mediated by spatially different groups of GCs (group 2) or by a single cluster of GCs (group 1), the excitatory and inhibitory synaptic weights evoked in PCs by GCs targeted by a given MF are paired and lead to homogeneous and scaled E/I balance as described in layer 2/3 pyramidal cells of the cerebral cortex (Xue et al., 2014). In agreement with these results, we found that inhibition was significantly stronger when MF stimulation led to E/I pairs (groups 1 and 2) than when only inhibition was measured (group 3) in both datasets (single protocol, group 1 and 2, 15.7 ± 14.4 pC, n = 58; group 3, 1.1 ± 0.2, n = 4, U = 220.0, p = 0.00,055; surface protocol, groups 1 and 2, 13.4 ± 10.1 pC, n = 54; group 3, 3.2 ± 3.2, n = 10, U = 452.0, p = 0.0008; Mann–Whitney U test in both cases; Fig. 3E).
E/I synaptic weights are correlated. A, Scatter plot of IPSQs against EPSQs for E/I sequences considered as classic FFI (Fig. 2, group 1). Single illumination (orange, slope = 3.6 ± 0.6, intercept = 4.9 ± 1.6, r = 0.76, n = 33, p value in plot); surface illumination (light gray; slope = 2.9 ± 0.7, intercept = 5.0 ± 2.7, r = 0.68, n = 20, p value in plot); linear fit of merged datasets (black; slope = 3.2 ± 0.4, intercept = 5.2 ± 1.4, r = 0.72, n = 53, p value in plot). Colored interval represents 95% CI in all cases. B, Scatter plot of IPSQs against EPSQs from group 2 E/I sequences. Single illumination (orange; slope = 2.9 ± 0.2, intercept = 1.7 ± 2.1, r = 0.93, n = 25, p value in plot), surface illumination (light gray; slope = 2.0 ± 0.5, intercept = 5.9 ± 2.4, r = 0.6, n = 34, p value in plot), linear fit of merged datasets (black; slope = 2.6 ± 0.2, intercept = 3.5 ± 1.6, r = 0.83, n = 59, p value in plot). Colored interval represents 95% CI in all cases. C, Scatter plot of IPSQs against EPSQs from all E/I sequence (groups 1 and 2). Single illumination (orange; slope = 2.7 ± 0.2, intercept = 5.0 ± 1.2, r = 0.89, n = 58, p value in plot), surface illumination (light gray; slope = 2.1 ± 0.4, intercept = 6.2 ± 1.8, r = 0.6, n = 54, p value in plot), linear fit of merged datasets (black; slope = 2.6 ± 0.2, intercept = 5.2 ± 1.0, r = 0.79, n = 112, p value in plot). Colored interval represents 95% CI in all cases. D, Histogram of the synaptic charge E/I ratio (orange, single; gray, surface illumination). Curves show kernel density estimation of distributions. Mann–Whitney test, U = 1461.0, p value in plot. E, Comparison between inhibitory synaptic charges from group 3 versus groups 1 and 2. White line, median; box edges, interquartile range; whiskers, minimum and maximum; filled circles, outliers. Single protocol, groups 1 and 2, 15.7 ± 14.4 pC, n = 58; group 3, 1.1 ± 0.2, n = 4, U = 220.0, p = 0.00,055; surface protocol, groups 1 and 2, 13.4 ± 10.1 pC, n = 54; group 3, 3.2 ± 3.2, n = 10, U = 452.0, p = 0.0008; Mann–Whitney U test in both cases).
Spike gain in PCs relies on ΔLat and frequency of GC and MLI inputs. A, Example of raster plot (top) and peristimulus time histogram (bottom) simulating PC discharge and spike gain (i.e., mean normalized PSTH) in response to seven stimulations delivered at 50 Hz with a fixed ΔLat of + 2 ms. B, Absolute spike gain as a function of various ΔLats (blue curve) with input parameters described in the gray box. The gray box illustrates the range of classical FFI. The black asterisk shows the gain measured in the example data shown in A. C, Spike gain as a function of variable ΔLat (−5 ms to +6 ms) and stimulation frequencies (10–200 Hz) for three (top), five (middle), and seven stimulations protocols (bottom). The black dashed line illustrates classical FFI value when GCs are directly stimulated. The white dashed lines illustrate the range of frequencies for which the variability in spike gain is the largest when ΔLat varied. D, Average spike gain measured across three (blue curves), five (orange curves), or seven (green curves) stimulations delivered at frequencies ranging from 10 to 20 Hz (top), 30 to 75 Hz (middle), and 100 to 200 Hz (bottom). The black vertical dashed line represents the average ΔLat found in classic FFI. The gray box shows the classic FFI interval (i.e., ±2 SD from average; compare Fig. 2).
Variable ΔLat expand PC discharge range in response to biologically relevant GC inputs. A, Trajectories of PC spike gain for three and seven stimulations when ΔLat varies from −5 ms (blue) to +5 ms (red) at various simulation frequencies (10–200 Hz). STDs are defined in Table 1. The magenta line illustrates the distance measured at 40 Hz (see above, Materials and Methods). Yellow circles emphasize the trajectory measured in classic FFI (ΔLat = +2 ms). B, Trajectories of PC spike gain for three and seven stimulations when ΔLat varies from −5 ms (blue) to +5 ms (red) at 40 Hz stimulation frequency. Each line represents one combination of E, I, and STD (12 combinations; see Results; Table 1). C, Trajectories shown in B transformed into Euclidian distances (i.e., blue, no influence of E/I delay, to red, strong influence from E/I delay; see above, Materials and Methods) and calculated for stimulation frequency from 10 to 200 Hz. EI delays largely contribute to modulation of PC discharge in case of low initial spike gain. D, Distribution of distance for each combination of E, I, and STD at different frequencies. E, Entropy of distributions shown in D at varying frequencies.
Variable delays expand PC dynamics
We anticipated that the broad distribution of ΔLat in PC soma could influence PC discharge. We therefore investigated how ΔLat variability in E/I sequences influences the PC firing rate using a computational model of the GC–MLI–PC FFI pathway validated in a previous study (Grangeray-Vilmint et al., 2018; see above, Materials and Methods). In this model, burst of EPSPs from GCs were combined with bursts of IPSPs from MLIs to predict PC rate. Initially, the following parameters were varied: EPSP and IPSP sizes, input frequency, and STD during bursts. We now included ΔLat as a new parameter with the range determined in our experimental dataset. EPSP and IPSP amplitudes as well as STD at GC–PC and MLI–PC synapses were simulated using the deterministic model proposed by Tsodyks and Markram (1997; Table 1) in which a single parameter (U) can be used to systematically change the nature of a synapse from a facilitating to a depressing mode (UE and UI, respectively). In our previous study (Grangeray-Vilmint et al., 2018), we demonstrated that these four parameters (i.e., EPSP and IPSP size, STD of excitation and inhibition) in association with the number of GC burst stimulations strongly influence how GCs control PC discharge. Because, on average, GC–PC synapses facilitate, whereas GC–MLI–PC synapses depress, short GC bursts (three stimulations) leading to a net inhibition can switch to a net excitation when burst duration increases (e.g., seven stimulations; Grangeray-Vilmint et al., 2018). This model of the GC–MLI–PC could reproduce PC discharge following bursts of GC inputs (three, five, and seven stimulations) in various combinations of parameters (Grangeray-Vilmint et al., 2018). We therefore quantified changes in the PC firing rate when the ΔLat parameter is included in the model (see above, Materials and Methods). Variation of PC firing rate after stimulation was monitored using a simple metric, the spike gain, defined as the net effect of the stimulation. The spike gain is negative when GC stimulation leads to omitted spikes in PCs (i.e., net inhibition) during the time window of stimulation and conversely positive when an excess of spike (i.e., net excitation) is simulated (in comparison to baseline activity in both cases).
Using a biologically plausible range of EPSP (1.22 to 1.82 mV at resting potential), IPSP (−0.63 to −1.39 mV at resting potential), STD parameters (facilitation at GC–PC synapses, depression at MLI–PC synapses), and stimulation frequencies (three to seven stimulations at 10–200 Hz; Chadderton et al., 2004; Jörntell and Ekerot, 2006; Arenz et al., 2008; Grangeray-Vilmint et al., 2018), we could reproduce the sample raster plot observed in biological data (Fig. 4A, raster plots; Grangeray-Vilmint et al., 2018). Next, we estimated spike gain by systematically varying ΔLat from −5 ms (I before E) to +6 ms (E before I), input frequency (10–200 Hz), and stimulus burst size (three, five, seven) for different sets of E, I, UE, and UI (Fig. 4A,B; Table 1; see above, Materials and Methods). Figure 4, C and D, shows spike gains for the same set of E, I, UE, and UI parameters as a function of ΔLat and stimulation frequency for different burst durations.
We observed that although spike gain is negative for a short burst of GC stimulation (Fig. 4C, top left), it may become positive for a longer burst (Fig. 4C, bottom left) at all frequencies above 30 Hz and for most ΔLats. Indeed, for short bursts, when inhibition occurs a few milliseconds before the excitation, the impact of EPSP on the spike probability of the neuron is strongly reduced (Fig. 4C, top left, blue area; Fig. 2H,I). By contrast, when inhibition occurs a few milliseconds after the excitation, the spike probability may not be much affected as the neuron may have already spiked when the EPSP peaked, that is, before the peak of the incoming IPSP. Consistent with this reasoning, as we increase the ΔLat from negative to positive, the spike gain increased monotonically with a steady state above ΔLat = 2.5 ms (Fig. 4B,C, left, vertical axis). However, this absolute spike gain increase was strongly affected by burst frequency. This is because excitatory and inhibitory synapses undergo STD (Grangeray-Vilmint et al., 2018), and the amount of facilitation and depression per spike differs for excitatory and inhibitory synapses. These simulations also revealed that ΔLat influence is maximal at a stimulation frequency of 30–75 Hz, as also observed when spike gains were pooled by stimulus frequency bands (Fig. 4C,D). Overall, similar patterns (reduced spike probability when inhibition comes first and monotonic increase in absolute spike gain) were observed when the stimulus consisted of five or seven spikes. (Fig. 4C, left, middle, bottom). For longer stimuli and higher frequencies, the definition of the ΔLat becomes blurred from the second to the last spike, and the influence of ΔLat is attenuated (Fig. 4D, bottom). Figure 4, C and D, confirmed that burst duration (three, five, and seven GC stimulations) strongly affected PC behavior.
To further study the effect of ΔLat and synaptic parameters on PC discharge, we measured spike gains in 12 different combinations of E and I synaptic strengths and two different E facilitations for each ΔLat (i.e., 24 different synaptic strength combinations per E/I delay and frequency). To evaluate these combinations in a simple way, we rendered the results by monitoring ΔLat influence at two burst durations (three and seven stimulations) at which the switch between a net inhibition to a net excitation was observed. (One example for the set of parameters used in Fig. 4 is shown in Fig. 5A and for all combinations at 40 Hz in Fig. 5B.) As shown in Figure 4C, varying ΔLat from negative to positive values increased spike gains; hence, trajectories always move toward the upper right part of the plot. At 40 Hz (Fig. 5B), the length of the trajectories (i.e., the total change in spike gain) is inversely proportional to the initial spike gain, and the strongest is the inhibition for three stimulations, the longer is the trajectory. Therefore, we converted each trajectory into a cumulative Euclidian distance in the two-dimensional space spanned by spike gain for stimulus with three and seven spikes (i.e., the total change in spike gain; see above, Materials and Methods; Fig. 5A), and all 24 combinations were rendered for stimulation frequencies ranging from 20 to 200 Hz (Fig. 5C). In this rendering, the longer the distance, the higher the effect of ΔLat. As shown in Figure 4, midrange frequencies (20–50 Hz) yield the highest distance, particularly for negative and low spike gains. These results clearly show that E/I-delays have a large effect on spike gain in PCs but only at low-frequency bursts. To better illustrate the influence of stimulation frequency on ΔLat effects, we estimated the distribution of distance at different frequencies (Fig. 5D) and measured the entropy of the distribution (Fig. 5E; see above, Materials and Methods). In this analysis, the higher the entropy, the wider the distribution, and therefore the bigger the effect of ΔLat. This revealed that ΔLat affects the spike gain at a wide range of frequencies, but, again, its effect is the strongest at midrange frequencies (20–75 Hz; Fig. 5D,E). Together, our model suggests that ΔLats observed in the experimental dataset can expand the dynamic range of GC influence of PC discharge (spike gain) in a biologically relevant range of frequencies.
Discussion
Here, we have shown that MF stimulation of single or multiple rosettes in the cerebellar cortex leads to diverse sequences of synaptic excitation and inhibition in PCs. Although synaptic strengths appeared to correlate within these sequences, the delay between E and I ranged from negative to positive values, extending the possibilities for a given set of synaptic inputs to influence PC discharge. Although it seems easy to predict the modulation of the PC firing rate in response to nominal synaptic inputs a priori (i.e., GC input equals acceleration and/or MLI input equals deceleration), our experimental results supported by further simulations suggest that it remains crucial to consider the subsequent timing in which such inputs target PCs. As the architecture of the MF–GC–MLI–PC pathway inherently leads to diverse E/I sequences, as summarized in Figure 1A, we therefore propose that the selective combination of GC–PC and GC–MLI–PC synapses allows PCs to distinguish between multiple synaptic inputs based on their relative temporal occurrence (i.e., ΔLat).
The single and surface illuminations evoked temporal E/I decorrelation in the MF–GC–(MLI)–PC pathway
In this study, we addressed the temporal aspect of E and I integration of the MF–GC–MLI–PC pathway in PCs by using single rosette and surface illumination protocols. Single rosette illumination elicited a single action potential and a steady depolarization, allowing for a precise onset of vesicular release at the MF–GC synapses. We postulate that direct, steady depolarization of axon terminals (rosettes), although not physiological, should reliably release more vesicles than a single action potential (Jackman et al., 2014). This approach may overcome weak synaptic weights that can be observed at individual MF–GC synapses (Chabrol et al., 2015) as well as tonic and phasic inhibition by Golgi cells (Jörntell and Ekerot, 2006; Kanichay and Silver, 2008). As our results were measured from the first occurrence of E and I in PCs rather than from the influence of train of stimulation, we therefore considered this technical approach as the most relevant. Despite this direct illumination, delays between E and I in PCs were not only highly variable but also incompatible with classical FFI during which GCs excite both PC and MLIs. Conversely, directly stimulating GCs leads to a very narrow distribution of ΔLat and recapitulates classical FFI previously recorded in many studies both in vivo and in vitro (Eccles et al., 1967; Brunel et al., 2004; Mittmann et al., 2005; Grangeray-Vilmint et al., 2018). Although recording at room temperature may certainly increase E and I absolute latencies, the difference in ΔLat between direct GC stimulation and MF illumination cannot be accounted for by temperature. ΔLat, especially negative ΔLat, can only be accounted for by the inherent spatial organization of the MF–GC pathway (i.e., MFs targeting different groups of GCs). In our conditions the illumination of a single rosette leads to an antidromic propagation of the action potential in all the branches of the MF. Indeed, this behavior mimics physiological properties as incoming action potentials from the precerebellar nuclei likely invade all the MF collaterals in the cerebellum (Fig. 1A). Anatomical reconstructions showed that a single MF gives rise to many distant rosettes in a cerebellar lobule (e.g., more than 100 when considering single MFs from the dorsal column nuclei; Shinoda et al., 2000; Sultan, 2001; Quy et al., 2011; Na et al., 2019), activating discrete and distinct GC clusters that may target the same PCs as shown previously (Valera et al., 2016; Spaeth et al., 2022). This latter configuration is also mimicked by the surface illumination used in our experiments, which showed similar results than single rosette excitation but with a wider range of ΔLat. Indeed, we observed no significant difference in synaptic weights between single and surface illumination in agreement with the hypothesis that single rosette illumination gives rise to a propagating action potential in MF collaterals. Also, the small surface illuminated (80 × 80 µm), which likely excite rosettes belonging to the same MF (Quy et al., 2011), the poor penetration of blue light in the slice (Al-Juboori et al., 2013), and the fraction of MFs expressing Channelrhdopsin2 indicate that surface stimulation excites only a few different MFs. Overall, these findings demonstrate that MFs contact GCs that can eventually target MLIs or PCs independently. Beyond ΔLat, we also observed that MF illumination can lead to pure inhibitory synaptic transmission onto PCs, in agreement with the nonoverlapping GC–PC and GC–MLI connectivity maps described in Valera et al. (2016). In this latter study, by recording either PCs or MLIs at a given location in the cerebellar cortex, we described discrete areas of the GC layer that preferentially connect MLIs or PCs, predicting that GC may lead to only inhibitory inputs to PCs. Altogether, along with the temporal expansion observed at the MF–GC connection (Chabrol et al., 2015), our data predict that E and I can be temporally decorrelated at the GC–PC connection.
In an FFI circuit, when excitatory synapses are facilitatory and inhibitory synapses are depressing, it is optimal to represent input activity as a sparse code in which only a few neurons carry the stimulus-related information (Tauffer and Kumar, 2021). In such a scenario, stimulus-related activity of PCs will be correlated, even if we assume a distribution of synaptic weights. However, in the case where each GC projection on a PC would be associated with a slightly different delay, PC responses could be decorrelated. So in addition to controlling the sign and magnitude of the PC responses, ΔLat distribution can also contribute to a temporal decorrelation of PC responses.
Based on these results, it can be argued that E and I synaptic weights would also be decorrelated. However, we showed that E and I synaptic weights, recorded in PCs are correlated (Fig. 3), as already shown in the cerebral cortex between parvalbumin interneurons and pyramidal cells (Xue et al., 2014). The equalized E/I ratio in the cerebral cortex is activity dependent and spatially regulated independently at individual inhibitory synapses. Therefore, we suggest that similar rules may apply in the cerebellar GC–MLI–PC FFI pathway and that individual GC clusters might be selected by PCs to ensure proper E/I balance in PCs. Note that we activated MFs originating in a single set of precerebellar nuclei (DoCN), so additional experiments, in which different sources of MFs would be activated, are then necessary to generalize these results. Nonetheless, because long-term postsynaptic plasticity rules at the MLI–PC and GC–PC are controlled by climbing fiber inputs (Jörntell and Hansel, 2006; Gao et al., 2012), E/I balance regulation may not depend on MF origins.
Altogether, our findings agree with and support the spatial organization of the GC–PC connectivity maps that correlate with specific locomotor context at the level of individuals (Spaeth et al., 2022). Moreover, and as FFI circuits such as the GC–MLI–PC with synapses are ubiquitous in the brain, these results and insights are likely to be applicable to other brain regions.
Extending temporal processing in the MF–GC–MLI–PC pathway
In the cerebellum, most principal neurons are spontaneously active (e.g., PCs, nuclear cells), and MF inputs constantly convey sensorimotor information at a wide range of frequencies to the GC layer that fires frequently in bursts of action potentials (Arenz et al., 2009; Chabrol et al., 2015). Typically at frequencies above 20 Hz, STD at GC–PC and MLI–PC synapses modulates vesicular release (Perkel et al., 1990; Llano et al., 1991; Atluri and Regehr, 1996; Isope and Barbour, 2002; Valera et al., 2012; Doussau et al., 2017; Grangeray-Vilmint et al., 2018) and shapes E/I balance in PCs on a spike-by-spike basis. However, accurate control of GC firing rates with high-frequency stimulation (>20 Hz) of MF rosette was not possible using our approach. Indeed, although Channelrhodopsin2 (H134R) elicits large photocurrents, it has slow kinetics and is calcium permeant. These features would therefore lead to unreliable stimulation as shown in (Jackman et al., 2014; Rost et al., 2022). Furthermore, we opted for rosette steady stimulation to enhance vesicular release and increase our chances to reliably elicit action potentials in GCs and therefore bypass (1) inhibition of Golgi cells and (2) STD at the MF–GC synapse (Kanichay and Silver, 2008; Chabrol et al., 2015).
We therefore studied how ΔLat variability influences PC discharge using a model of the GC–MLI–PC pathway developed in a previous study (Grangeray-Vilmint et al., 2018). We showed that ΔLat controls E/I balance and PC discharges as dendritic integration of synaptic inputs is deeply affected (Fig. 2H) in a frequency-dependent manner (Figs. 4, 5). We notably identified that ΔLat may have a strong effect on E/I balance and PC discharge only when input occurs in a midrange frequency (20–50 Hz; Figs. 4, 5). Indeed, MLIs fire at medium-range frequencies in vivo (i.e., ∼30 Hz) and are synchronized together via gap junctions (Kim and Augustine, 2021). Many studies have shown that MLI discharge is temporally controlled by electrical and chemical synapses (Mann-Metzer and Yarom, 1999; Alcami and Marty, 2013; Kim et al., 2014; Rieubland et al., 2014; Hoehne et al., 2020). Spikelet propagation in the MLI network ensures synchronization, whereas chemical synapses curtail the integration time window (Hoehne et al., 2020). Altogether, selecting groups of GCs targeting specifically MLIs or PCs or both in combination (Valera et al., 2016; Spaeth et al., 2022) with specific STDs (Chabrol et al., 2015; Dorgans et al., 2019) can enhance both the dynamic range and the resolution for temporal encoding in the MF–GC–PC pathway. Furthermore, we and others demonstrated that long-term plasticity (LTP and LTD) at the GC–PC synapses is gated by MLIs (Binda et al., 2016; Rowan et al., 2018). We then postulate that the independent regulation of GC–MLI–PC synapses can influence plasticity induction from specific cluster of GCs. These features may underlie cerebellar module communication and motor coordination (Apps et al., 2018). Finally, an appealing hypothesis would be that temporal dynamics between two different groups of MF inputs (e.g., from different modalities) might lead to the partial or the total suppression of one input (Fig. 2H,I). These properties might account for one of the main roles of the cerebellum, which is to suppress the expected sensorimotor feedback to process further the unexpected inputs.
Footnotes
This work was supported by the Université de Strasbourg; Centre National de la Recherche Scientifique Grants ANR-2015-CeMod, ANR-2019-MultiMod, and ANR-2019-NetOnTime; and Fondation pour la Recherche Médicale Grant DEQ20140329514 to P.I. L.S. was supported by Fondation pour la Recherche Médicale Fellowship FDT201805005172, and A.K. was supported by Swedish Research Council Grant 2018-03118, Strategic Research Area Neuroscience, Digital Future, and a fellowship from University of Strasbourg Institute for Advanced Studies. We thank Théo Gagneux, Dr. Sophie Reibel-Foisset, Dominique Ciocca, and the staff of the animal facility Chronobiotron at University of Strasbourg for technical assistance, and Dr. Bernard Poulain, Dr. Antoine Valera, Dr. Matilde Cordero-Erausquin, Dr. Didier Desaintjan, and Dr. Frédéric Doussau for critical discussions.
The authors declare no competing financial interests.
- Correspondence should be addressed to Philippe Isope at philippe.isope{at}inci-cnrs.unistra.fr
