Introduction

The cerebellum is implicated in a diverse range of behaviors, extending from simple reflexes to complex functions such as language and social interaction (Kelly et al., 2020; Silveri, 2021; Baumann and Mattingley, 2022). The circuit architecture underpinning this behavioral diversity is relatively simple and conserved across its extent, suggesting that different classes of input are subject to common processing mechanisms. The cerebellar cortex lacks recurrent excitation, and almost all local feedback is mediated via GlyT2-positive inhibitory interneurons (Golgi and Lugaro cells) situated in the granule cell layer. Golgi cells receive direct input from mossy fibers (MFs) and excitatory feedback from parallel fibers (PFs), enabling them to exert feedforward and feedback control over granule cells (Eccles et al., 1966; Dieudonne, 1998; Kanichay and Silver, 2008). Golgi cells are numerically sparse relative to granule cells and possess extended axonal arborizations, meaning each Golgi cell innervates many hundreds or thousands of granule cells. Inhibition via Golgi cells could thus potently impact the transformation of MF signals into PF activity. In vivo measurements confirm that inhibition gates the transformation of sensorimotor information in the granular layer (Chadderton et al., 2004; Duguid et al., 2012), sparsifying responses to sensory input, and facilitating pattern separation at the population level (Fleming et al., 2024). Lugaro cells make up approximately one-third of granule cell layer inhibitory interneurons (Simat et al., 2007) and are known to make synaptic connections onto Golgi cells, molecular layer interneurons and Purkinje cells (Eyre and Nusser, 2016; Dean et al., 2003). Although comparatively little is known about their function in vivo (Prestori et al., 2019), their sensitivity to serotonin levels in the cerebellar cortex (Dieudonne and Dumoulin, 2000) suggests Lugaro cells could regulate information processing in the granule cell layer based on behavioral context. Currently, we lack quantitative information about how changes in Golgi/Lugaro cell activity influences cerebellar cortical dynamics and affects sensorimotor representations in the molecular and Purkinje cell layers. Indeed, the net influence of granule cell layer inhibitory interneurons is not known but may be presumed to alter both the rate and timing of granule cell output. Resultant changes in PF excitation could have net excitatory or inhibitory effects at the level of Purkinje cells and consequently for cerebellar nuclear neurons.

Here we investigate the influence of granule cell layer inhibitory interneurons on cerebellar population activity in the context of whisker movement, to reveal how inhibitory feedback influences sensorimotor transformations and ongoing behavior. Neuronal representations of whisking activity in the lateral cerebellum are widespread (Bosman et al., 2010; Chen et al., 2016, 2017; Brown and Raman, 2018; Zhai et al., 2024) and are partly based on a linear rate code, whereby the firing rates of cerebellar neurons predict upcoming whisker position (Chen et al., 2016). These representations are first seen in patterns of mossy fiber input and are passed forward at each stage of cerebellar cortical processing (Chen et al., 2017). Golgi cell inhibition has the potential to modulate how incoming whisking information is encoded and transformed within the granule cell layer (Gurnani and Silver, 2021; Palacios et al., 2021). However, it is unknown whether, and how, changing granule cell layer inhibition impacts neuronal representations of whisker behavior and indeed whisking behavior itself. We have recorded whisker movements and population activity in the lateral cerebellum using Neuropixels probes in head-fixed mice expressing inhibitory DREADDs (designer receptors exclusively activated by designer drugs) selectively in Golgi and Lugaro cells (i.e., GlyT2-positive neurons) of lobule Crus 1. In these mice, it was possible to downregulate GlyT2-positive cell activity by activating DREADDs with the agonist, clozapine N-oxide (CNO), applied topically on the recording site. Our results reveal a subtle but direct influence of GlyT2-positive cells on network dynamics and consequently, the cerebellar-behavioral loop underlying voluntary whisking.

ResultsPopulation dynamics in lateral cerebellum during voluntary whisking

To understand the contribution of GlyT2-positive interneurons to cerebellar cortical dynamics, we first set out to describe how cerebellar populations represent sensorimotor information while Golgi and Lugaro cell inhibition was intact. We recorded cerebellar population activity using Neuropixels probes in lobule Crus 1 of awake mice and performed spike sorting to isolate single units (see Materials and Methods). We isolated an average of 42.3 (SD = 25.3) cortical units per recording (N = 12), for a total (n) of 508 putative units. To correlate spiking activity with whisker movement, we simultaneously tracked spontaneous whisking via a high-speed camera (Fig. 1a). In all cases, bouts of whisking activity were associated with pronounced alterations in cerebellar spiking activity (Fig. 1b,c). Single cell recordings previously revealed widespread tuning to whisker position in granule cells, inhibitory interneurons, and Purkinje cells (Bosman et al., 2010; Chen et al., 2016, 2017; Brown and Raman, 2018; Zhai et al., 2024). We therefore measured tuning to whisker position among single units of our population recordings. On a unit-by-unit basis, we examined the relationship between firing rate and whisker position, to construct individual tuning curves (see Materials and Methods). Consistent with previous work, we observed robust tuning to whisker position, such that single units exhibited their highest firing rates at preferred angles of whisker deflection.

Within locally recorded populations, we observed substantial diversity in the tuning to whisker position. Indeed, units recorded in the same penetration displayed a broad range of preferred angles and tuning profiles, including monotonic increases and decreases to whisker protraction/retraction (Fig. 1d). We then recombined all single unit spiking responses to calculate population tuning. Whereas individual units exhibited diverse selectivity, the mean tuning function for the population displayed a near-monotonic dependence upon whisker position, with rate changes in proportion to the magnitude of protraction and retraction (Fig. 1e).

To further assess tuning diversity at the population level, we clustered units based on their tuning curves, testing whether we could discriminate discrete functional subclasses of unit. We fitted a spline model with 7 cubic b-spline *s* to each tuning curve (n = 508; N = 12) to capture tuning curve shape and reduce variability due to noise (see Materials and Methods). We then inspected the output of k-means clustering with different numbers of expected clusters. With increasing cluster number, there was no evidence of a change in the sum-of-squared-distances decaying pattern, or elbow, which can be used as an indicator of the optimal number of clusters to be considered. Visual inspection of the mean tuning curves for each cluster revealed a progressive shift of preferred position from the most retracted to the most protracted whisker locations (Fig. 1f). Projection of tuning curves in 3D space using t-SNE, a nonlinear dimensionality reduction technique particularly suited to uncovering clustering structure, further confirmed that different groups of clusters are not well isolated (data not shown). Thus, our analysis does not support discrete classes of tuning curve but rather confirms that cerebellar neurons form a functional continuum in the space of possible tuning curves. At the local population level, Crus 1 cerebellar cortical neurons heterogeneously encode whisking position but represent increasing protractive and retractive movements via accumulating increases in firing rates.

Having established that concerted population activity is required for accurate sensorimotor representations, we performed principal component analysis upon the grouped spike trains from each recording. We initially preserved projections of population activity onto the first three eigenvectors, which describe the three orthogonal axes in neuronal space along which activity varies the most (Fig. 1g). Analysis was restricted around whisking bouts (−2 to +3 s about whisking onset). Projections of population activity revealed clear structure in neuronal space during transitions from quiescence to whisking activity and back again. This structure in turn suggested that population activity contains accurate information about whisking dynamics. We tested this proposal by linearly combining the first three principal components, attempting to reconstruct movement trajectories (see Materials and Methods). Using this approach, it was possible to decode whisking set point during single trials (Fig. 1h). In summary, our results show that the population activity in the lateral cerebellar cortex accurately reflects upcoming slow whisking dynamics, even when approximated within a low-dimensional space.

To quantify the relationship between different principal components and whisking, we computed the normalized cross-correlation between whisker angle and each of the first three principal components (pc) in each recording. Interestingly, pc1–3 showed distinct temporal relationships with respect to whisker movement (Fig. 2a,b). Information related to whisker position contained in pc1 tended to anticipate behavior, whereas information in pc2 and pc3 lagged behavior. Across all recordings (N = 12), the peak signal of pc1 always preceded the whisking signal, indicating that, across recordings, variation in the whisking-aligned average neuronal activity expressed by the pc1 tends to anticipate behavior (one-sample t test; t = 4.72; p = 0.0006; mean = 36 ± 8 ms). On the other hand, the peak signal for pc2 almost always occurred after whisker movement, indicating that whisker-related information contained in pc2 lags whisking activity (t = −4.43; p = 0.001; mean = −195 ± 44 ms). Finally, peak signals for pc3 were broadly distributed around 0 ms (t = −1.38; p = 0.19; mean = −89 ± 65 ms). Notably, pc1 peaks across recordings are narrowly distributed in time; this suggests that pc1 specifically contains precise information about whisking behavior, i.e., upcoming movement, occurring ∼40 ms in the future (Fig. 2c).

Figure 2.

Population representations of whisker movement at different timescales. a, Trial-averaged whisking position (top) and projected population activity (bottom) for one recording. b, Cross-correlations between pc1–3 and whisking activity for the same recording; vertical lines indicate times of peak correlation. c, Cumulative variance explained by pc1–3 across all recordings (N = 12). Only a modest amount of variability in neuronal activity is accounted for by first three principal components. d, Distribution of unit loadings (n = 508) for pc1–3. Kurtosis measurement, to quantify the number of outliers in each distribution, reveals fewer outliers for pc1 than what would be expected if the data were normally distributed (excess kurtosis −2.76). This indicates that information contained in pc1, which best reflects whisking activity, tends to be distributed across neurons. The loading distributions for pc2 and pc3 have respectively a similar or higher number of outliers compared with normally distributed data (excess kurtosis −0.43 and 4.97, respectively), indicating that information in pc2 and pc3 is increasingly concentrated in fewer units. e, Time of peak correlation for pc1–3 across all recordings. The distribution for pc1 is tightly concentrated around ∼40 ms, meaning that information captured by pc1 tends to anticipate whisking activity with high temporal precision (one-sample t test; t = 4.71; p = 0.0006). The information contained in pc2 and pc3 shows the opposite trend, with more variability (pc2: mean −194 ms, t = −4.44, p = 0.0009; pc3: mean −92 ms, t = −1.44, p = 0.17), suggesting that these components might reflect different aspects of behavior. f, Mean decrease in unexplained variance (unexp. var.) across recordings with increasing numbers of principal components (#pcs); standard deviation shaded in gray.

Whisking-related information contained in pc1–3 could originate from a small proportion of units in the population, whose activity may be particularly responsive to whisking; alternatively, information could be more distributed, whereby all units contribute to some extent to the population encoding of behavior. To measure how behavioral information was distributed across the population, we measured the pc loading of each unit (n = 508), indicating the contribution of a unit to the corresponding pc (see Materials and Methods). The loading distribution for each pc was centered around 0 and was roughly symmetric, meaning that units could contribute both positively or negatively to each pc (Fig. 2d). The number of outliers in each distribution was used to quantify the extent to which each pc is dominated by the activity of few units. We calculated the excess kurtosis of each distribution, which is a measure of the frequency of outliers observed, using as a point of reference the frequency expected from normally distributed data. The distribution for pc1 had an excess kurtosis of −2.85, indicating infrequent outliers; this in turn suggests that the contribution of units across recordings to pc1 is broadly distributed within the population. In contrast, the excess kurtosis for pc3 was 4.54, indicating a high frequency of outliers, and thus a subset of units which tend alone to contribute mostly to pc3. The excess kurtosis for pc2 was −0.47, meaning that outliers are neither frequent nor infrequent. Together, these results suggest that for pc1, and thus accurate representation of whisking behavior, information is broadly distributed across the entire population. Finally, we noted that the first three eigenvalues accounted for only a moderate amount of the total variance in the population activity (∼30%) across all recordings (Fig. 2e,f). Therefore pc1–3 provide only a coarse description of population activity, and whisker movement only partially describes neuronal dynamics in Crus 1 despite restricting our analysis to the initial period of whisking activity. Indeed, the large proportion of neuronal variability not captured in the first three principal components supports the proposal that cerebellar cortex may represent additional behaviorally relevant variables.

Chemogenetic downregulation of GlyT2-positive interneurons elevates population activity

Having established the relationship between population activity and slow whisking dynamics under baseline conditions, we next sought to understand how Golgi and Lugaro cell inhibition contributes to these representations in the cerebellar cortex. We adopted a chemogenetic strategy to selectively downregulate GlyT2-positive neuronal activity in an otherwise intact network while simultaneously recording population activity and whisker movement. Manipulation of GlyT2-positive cells was achieved via activation of the inhibitory DREADD, hM4Di, which causes reductions in excitability and neurotransmitter release (Stachniak et al., 2014) in cells expressing these receptors. Localized expression of DREADDs (Fig. 3b) was achieved by injecting the AAV-DIO-hM4D(Gi)-mCherry virus into Crus 1 of GlyT2-Cre mice (cre-recombinase restricted to GlyT2-positive cells in the cerebellar cortex; see Materials and Methods; Kakizaki et al., 2017). DREADDs were activated by local, topical delivery of the agonist, CNO, on top of the Neuropixels penetration site, which itself had been targeted as the site of virus injection (Fig. 3a); this method was used to minimize potential off-site effects of the agonist and afforded within-minute temporal precision to our manipulation. To confirm the fidelity of our approach, we first tested the effects of topical delivery of muscimol (1.7 μg/μl), a GABAA receptor agonist, during Neuropixels recordings from wild-type mice (N = 3). In all cases, within minutes of drug application, the total spike count recorded across the cerebellar cortex started to decrease and reached its minimum within 10–30 min (data not shown). These results supported the viability of applying CNO topically onto cerebellar cortex to locally activate DREADD receptors.

Figure 3.

Chemogenetic inhibition of GlyT2-positive cells increases neuronal activity in cerebellar cortex. a, Left, Targeted expression of hM4Di receptors was achieved via injection of AAV-DIO-hM4Di-mCherry into the lateral cerebellar cortex of GlyT2-Cre mice, selectively expressing Cre-recombinase in GlyT2-positive cells in this brain region. Right, Neuropixels probes were targeted to site of viral injection. During electrophysiological recordings, the exogenous drug, clozapine N-oxide (CNO) was topically applied to activate hM4Di receptors. b, Left, GlyT2-Cre mouse brain selectively expressing hM4Di in GlyT2-positive cells of the lateral cerebellar cortex. Right, Tract left by Neuropixels probe coated with DiI, showing colocalization of the site of recording and hM4Di expression. c, Changes in population spike count before and after CNO/vehicle delivery (at 0 min) normalized by count at −5 min, for three experiment conditions. Pink, GlyT2-Cre mice and CNO (N = 19); blue, C57BL6 mice and CNO (N = 5); green, GlyT2-Cre/C57BL6 mice and saline vehicle only (N = 9). d, Box plots showing pooled data for each experimental condition. e, Multilevel modeling approach to assess effect of drug delivery on cerebellar population activity. Population spike count distributions for each condition (top) were modelled using an inverse-Gamma distribution, described by shape parameter α and scale parameter β. Each parameter was modeled as a linear combination of different coefficients, including two θcond, one for α and one for β, which captured the specific effect of the experimental manipulation on total spike counts. Comparing the empirical distribution with the distribution of posterior predictive checks (samples from the model after fitting) shows that the model captures the overall structure of the data (bottom). f, 94% highest density intervals (hdi) of the contrasts between post- and predrop posterior samples for θcond for α and β. The contrasts highlight a specific effect of GlyT2-positive cell manipulation (GlyT2 + CNO) on both α and β parameters. g, Contrast difference (mean and variance) between post- and predrop of the inverse-Gamma distribution fitted to total spike counts. Bars indicate the 94% hdi of the posterior differences. Left, Contrast difference between GlyT2 + CNO condition and each control condition. Right, Contrast difference between the two control conditions. GlyT2-positive cell manipulation produces positive contrast differences versus both control conditions.

We next examined whether CNO-dependent reduction in GlyT2-positive cell activity had a measurable impact on cerebellar cortical population activity. To do this, we compared changes in cortical spike count before and after CNO administration across three experimental groups. GlyT2-Cre mice, preinjected with AAV9-hSyn-DIO-hM4D(Gi)-mCherry virus into Crus 1, or uninjected C57BL6 mice received topical delivery of CNO (GlyT2-CNO; N = 19 recordings) or saline vehicle (Veh; N = 9 recordings) to the cerebellar cortex. A third group of uninjected wild-type (C57BL6) mice received topical delivery of CNO to the cerebellum (WT-CNO; N = 5 recordings). Initial inspection of the data revealed an increase in normalized spike count in the GlyT2-CNO groups, but not Veh and WT-CNO groups, shortly after application of CNO (Fig. 3c,d). Additionally, when comparing the mean of the normalized postdrop cortical spike counts (5 min bin width), we found a significant difference across conditions (ANOVA; F = 10.91; p = 2.82 × 10−5). However, significant variability in spike counts were observed both within and between experimental groups. Therefore, to confirm that alterations in spike count observed across recordings were due to chemogenetic manipulation of GlyT2-positive cells, we used a multilevel model approach; this approach was warranted by the temporal correlations between repeated measures and clustering by animals. The model described the distribution of spike counts for each time bin with an inverse-Gamma distribution, parameterized by a shape parameter α and a scale parameter β. The model used as explanatory variables the experimental conditions, associated to pre- and postdrug administration periods, to find the α and β parameters that best explained the pre- and postdrug administration data. For each parameter of the gamma distribution (α and β) and each group (WT-CNO, Veh, GlyT2-CNO), the model had a coefficient θ associated to the predrug recording period and one for the postdrug administration period. The contrast between the two coefficients was used to measure the change in spike counts due to the experimental manipulation. To confirm the fidelity of the model, we compared the distribution of all experimentally observed spike counts together with the model fit (Fig. 3e): the overlap between these distributions indicated that the model performed well in capturing the overall structure of the data.

Having confirmed the fidelity of the model, we then compared the contrasts between model coefficients associated with the pre- and postdrug administration periods for each experimental group. These contrasts reflect the effect of the experimental manipulation on spike count distribution. For each experimental group there are two contrasts, one associated with the shape parameter α and one with the scale parameter β of the distribution describing spike counts. Comparing contrasts and highest density intervals (hdi, 94%) for the parameters across groups revealed a specific effect in GlyT2-CNO mice on both α and β (Fig. 3f). To gain a better intuition of the implications of these results, we used the posterior samples of α and β to derived samples for the mean and variance parameters of the inverse-Gamma distribution; the mean and variance offer an alternative parametrization that characterize more intuitively the center of mass and spread of spike counts across recordings for each condition, before and after drug application (see Materials and Methods). Thus, to compare results across condition, we computed the pre- minus postdrug mean and variance contrasts for each experimental condition, and the difference in contrasts between conditions, for each parameter. This analysis revealed that both mean and variance increased after drug application in the GlyT2-CNO group compared with the two control groups (Fig. 3g; mean hdi 0.05–0.46, 0.05–0.49 and variance hdi −0.01 to 0.31, −0.02 to 0.31 for GlyT2-CNO vs WT-CNO and Veh, respectively). It further revealed that in the WT-CNO group, CNO application alone has no detectable effect on network spiking activity (mean hdi −0.17 to 0.21 and variance hdi −0.02 to 0.03 for WT-CNO vs Veh). Together our analyses confirm that chemogenetic inhibition of GlyT2-positive cells causes a global increase in spiking activity within the cerebellar cortex.

Limited influence of GlyT2-positive cell downregulation on sensorimotor representation

Having established that chemogenetic downregulation of GlyT2-positive cells measurably altered cerebellar activity, we investigated how the perturbation affected sensorimotor representations of whisking. Golgi cell inhibition acts via the granule cell layer, but changes in the input–output transformation of granule cells will have knock on effects upon the activity of molecular interneurons and Purkinje cells. Golgi cells control spike output of granule cells through various mechanisms, including changes in gain, i.e., slope of the input-output function (Brickley et al., 1996; Hamann et al., 2002; Mitchell and Silver, 2003; Chadderton et al., 2004). By reducing Golgi cell inhibition, granule cells may become more sensitive to excitatory inputs, increasing their gain. We looked for changes in the sensitivity of cerebellar activity to whisker movement by comparing population tuning curves before and after CNO administration in GlyT2 mice. We observed no striking differences in the overall profile of these curves (Fig. 4a). Post-CNO, population firing rates were marginally higher for some whisker positions, but tuning curves were essentially unchanged. This result suggests that chemogenetic downregulation of GlyT2-positive cells has no substantial effect on gain or overall representation of whisker position at the population level. In principle, altering Golgi inhibition could cause reorganization of the granule cell population code by altering the requirements for granule cell integration of excitatory inputs from different mossy fibers (Marr, 1969). We therefore explored whether perturbed GlyT2-positive cell activity causes substantial reorganization of whisking representation at the single cell level. To do this, we measured the entropy of individual tuning curves before and after CNO/vehicle administration. Entropy is a measure of how informative units are about, in this case, whisking position, with higher-entropy neurons being less informative. Following CNO administration, we observed no changes in cumulative entropy distribution in GlyT2-CNO recordings (Fig. 4b). Interestingly, under control conditions (i.e., Veh and WT-CNO recordings), units tended to have higher entropy values after vehicle/CNO administration, indicating that tuning curves become less informative about whisking position over time. We further compared the distributions of entropy differences pre- and postdrug delivery (Fig. 4c). Under control conditions, tuning curve entropies typically remained unchanged or increase over time (control density skewness, 2.80). CNO administration in GlyT2 mice causes additional reorganization—some units decrease their entropy—leading to an even distribution of entropy changes after drug application (GlyT2-CNO density skewness, 0.21). The cumulative entropy distribution in GlyT2-CNO recordings therefore remains similar across time suggesting that the firing rate of some units becomes more sensitive to changes in whisking position following GlyT2-positive cell manipulation. These results also imply that recording conditions in the control condition are associated with small but progressive losses of representational fidelity. To confirm this proposal, we again measured the correlation between each of the first three population principal components (pc1–3) and whisker position for GlyT2-CNO and control recordings (Fig. 4d,e). We found no evidence for large differences in correlation pre- and postdrug administration. However, there was a general trend for correlations to decrease after drug delivery, except for pc1 and pc2 (i.e., fast and slower representations of whisking) in the GlyT2-CNO condition, where correlations increased slightly. Our results suggest that the fidelity of sensorimotor representation is subject to a mild downward drift under control conditions in the present experimental setting and that GlyT2-positive cell downregulation may slightly improve representation, offsetting this drift. But overall, we did not observe notable changes in the quality of the encoding of whisking behavior by population activity. Together our results indicate that individual sensorimotor tuning of cerebellar neurons is only weakly influenced by chemogenetic perturbation of GlyT2-positive interneurons.

Figure 4.

Weak influence of GlyT2-positive cell perturbation on cerebellar sensorimotor representations. a, Average tuning curves for GlyT2-CNO recordings before (gray) and after (red) topical administration of the drug. b, Entropy cumulative density functions for all units recorded in the GlyT2-CNO (top) and control (bottom) conditions. Under control conditions, but not GlyT2-CNO, unit entropy tends to increase after drug/vehicle delivery (postdrop), suggesting that tuning curves become less informative about whisking position over time. c, Probability mass function of unit entropy differences before and after CNO/vehicle administration (“postdrop” minus “predrop” entropy). Under control conditions, tuning curve entropy tends to remain stable or increase over time (density skewness, 2.79). In the GlyT2-CNO recordings, a fraction units exhibit decreased entropy after drug application (density skewness, 0.21), suggesting that some units become more sensitive to changes in whisking position following GlyT2-positive cell inhibition. d, Comparison of peak pre- and postdrop correlation values between principal components 1–3 (pc1–3) and whisker position for control recordings. e, Comparison of peak pre- and postdrop correlation values between principal components 1–3 (pc1–3) and whisker position for GlyT2-CNO recordings.

GlyT2-positive cell downregulation alters cerebellar dynamics and weakens coupling to whisking behavior

We further explored the impact of GlyT2-positive cell perturbation by comparing trial averaged perievent time histograms (PETH) and whisking activity from pre- and postdrug administration periods (see Materials and Methods). Because cerebellar activity can directly influence whisker movement (Proville et al., 2014), reduced GlyT2-positive cell inhibition could influence both cerebellar dynamics and produce measurable changes in whisking behavior. We first considered cerebellar activity: reductions in Golgi cell inhibition could alter the temporal dynamics of population activity in several ways, acting via tonic and/or phasic inhibition (Brickley et al., 1996; Chadderton et al., 2004; Duguid et al., 2015) and/or affecting Golgi cell synchrony (van Welie et al., 2016). We therefore compared the patterns of population activity at movement onset before and after chemogenetic perturbation. Typically population activity peaks close to the onset of individual whisking bouts (Fig. 1b,c). Therefore, for each recording, we computed the standard deviation of the distribution of peak unit firing times aligned to whisking onset, pre- and postdrug delivery (Fig. 5a; see Materials and Methods). In the GlyT2-CNO mice, but not controls, pre- and postdrug standard deviations were significantly different (Fig. 5b), with significantly lower variation following CNO administration (two-sided Wilcoxon signed-rank test; T = 12; p = 0.017). Thus, GlyT2-positive cell downregulation increased the temporal alignment of peak neuronal activity around whisking onset. These results suggest that decreased GlyT2-positive cell inhibition has an impact on the pacing of neuronal responses around the onset of whisking activity and in particular narrows the time window within which units fire most strongly.

Figure 5.

GlyT2-positive cell inhibition reduces temporal variability in neuronal populations and weakens behavioral coupling at movement onset. a, Peri-event time histograms aligned to onset of whisker movement for pre- (top) and postdrop (bottom) periods for a GlyT2-CNO recording. Gray and red dots indicate the absolute peak of neuronal activity centered around whisking onset for the pre- and postdrop period, respectively. Units ranked according to their peak firing rate. Temporal jitter peak activity decreases after CNO application. b, Temporal dispersion of neuronal population activity before and after drug/vehicle administration (pre- and postdrop) for control (n = 12) and GlyT2-CNO (n = 13) recordings. Each data point represents the standard deviation of the distribution of absolute peak times of neuronal activity for each recorded population. Reduction of GlyT2-positive cell inhibition decreases the temporal dispersion of neuronal activity during whisking initiation (two-sided Wilcoxon signed-rank test; T = 12; p = 0.017). c, Average whisking onsets for two representative GlyT2-CNO recordings before (gray) and after (red) drug administration. Dashed lines represent linear fits of initial whisker protraction. CNO delivery was associated with both increased (top) and decreased (bottom) slope of protraction. d, Change in slope of whisker protraction at movement onset before and after drug/vehicle administration (pre- and postdrop) for control and GlyT2-CNO recordings. e, Box plot showing differences in slope of whisker protraction at movement onset between pre- and postdrug/vehicle administration for control and GlyT2-CNO recordings. The standard deviation of the slope distribution in the GlyT2-CNO condition is higher than controls (Levene's test W = 8.39; p = 0.008) suggesting that decreasing local GlyT2-positive cell inhibition produces variable changes in the dynamics of movement onset across recordings. f, Relationship between change in cerebellar population dynamics and whisker movement following drug/vehicle administration for all recordings. Under control conditions, measured changes in neural activity and movement are small in magnitude and well correlated (r2 = 0.52). In GlyT2-CNO recordings, relationships between cerebellar activity and movement are decoupled (r2 = 0.15).

Finally, we measured the effects of GlyT2-positive cell manipulation on whisking velocity at the onset of movement bouts (see Materials and Methods). For each recording, we used a linear model to approximate the slope of the average whisking bouts during its protraction phase (Fig. 5c); the coefficient for each linear fit was used as a read out for velocity during the movement initiation. Interestingly, unlike control recordings, GlyT2-CNO mice showed changes in their whisking behavior following drug administration. However, the sign of these changes was not consistent across animals, occurring in both increasing and decreasing directions (Fig. 5d). Individual onsets of voluntary whisking thus became faster or slower after drug administration in GlyT2-CNO mice. The overall distribution of changes in whisker onset velocity (“slope contrast”) was therefore significantly broader in the GlyT2-CNO mice compared with control (Fig. 5e). Thus, GlyT2-positive cell downregulation produces measurable but divergent effects on local cerebellar dynamics and whole animal whisking behavior. Local cerebellar dynamics were consistently compressed, leading to a more tightly locked population activity, whereas whisking velocity was either increased or reduced. These results suggest a breakdown in the link between cerebellar dynamics and behavior following GlyT2-positive cell downregulation. We confirmed this by plotting changes in population activity and whisking velocity at movement onset for individual recordings (Fig. 5f). In control recordings, differences in network activity and whisker movement before and after CNO/vehicle delivery are small in magnitude but importantly, are well correlated (r2 = 0.52). In GlyT2-CNO recordings, changes in cerebellar dynamics and movement become decoupled from one another, leading to break down in this correlation (r2 = 0.15). Together our results indicate GlyT2-positive cell inhibition regulates the temporal patterning of population activity and that disruption of this patterning at the local level can have diverse consequences for behavior.

Discussion

Granule cell layer inhibitory interneurons are proposed to play important roles in cerebellar transformation (Marr, 1969; Albus, 1971; D'Angelo et al., 2013;