Abstract
Cerebellar outputs take almost exclusively polysynaptic routes to reach the rest of the brain, impeding conventional tracing efforts. Here we quantify pathways between cerebellar cortex and contralateral thalamic and corticostriatal structures using an anterograde (H129) transsynaptic tracer herpes simplex virus type 1, a retrograde (Bartha) tracer pseudorabies virus, and a whole-brain pipeline for neuron-level analysis using light-sheet microscopy. In ascending pathways, sensorimotor regions contained the most labeled neurons, but higher densities were found in associative areas, including orbital, anterior cingulate, prelimbic, and infralimbic cortex. Ascending paths passed through most thalamic nuclei, especially ventral posteromedial and lateral posterior (sensorimotor), mediodorsal (associative), and reticular (modulatory) nuclei. Retrograde tracing revealed that the majority of descending paths originate from somatomotor cortex. Patterns of ascending influence correlated with anatomical pathway strengths, as measured by brainwide mapping of c-Fos responses to optogenetic inhibition of Purkinje cells. Our results reveal parallel functional networks linking cerebellum to forebrain and suggest that cerebellum is capable of using sensory-motor information to guide both movement and nonmotor functions.
INTRODUCTION
The cerebellum has an increasingly recognized role in nonmotor processing1–3. Patients with cerebellar damage show multiple cognitive and affective symptoms1, and damage at birth leads to autism spectrum disorder (ASD) in half of cases4–6. These observations suggest a broad role for the cerebellum in nonmotor function during development and adulthood.
However, the pathways that mediate these influences are poorly characterized on a whole-brain scale. Of particular interest is the cerebellum’s partnership with neocortex as these two structures are the second-largest and largest divisions, respectively, of most mammalian brains7. The major descending corticocerebellar pathway passes through the pons and the majority of returning ascending fibers pass through the thalamus8, 9, comprising two massive within-brain long-distance pathways10. These descending and ascending pathways are suggested to form closed loops11, giving each cerebellar region one or more specific neocortical partners with which it exchanges information.
Great efforts have been made in mapping the structure of inputs, organization, and outputs of the cerebellum. Two information streams, originating from the cortex and spinal cord, serve as input into the cerebellum, with their targets arranged in zones defined by gene expression and climbing fiber connections12–15. Incoming information through the mossy fiber/granule-cell pathway diverges massively, followed by convergence onto Purkinje cells and deep cerebellar nuclei (DCN)13. However, although classic tracing methodology can label monosynaptic connections along these paths, the same methods cannot trace incoming or outgoing polysynaptic chains of connection.
Given the brain-wide nature of cerebello-cortical pathways, researchers have examined them using large-scale functional approaches. Transcranial magnetic stimulation in humans has demonstrated that the cerebellum influences neocortical excitability16. Functional MRI can attain subcentimeter resolution, detect long-distance correlations17, and when coupled with cerebellar stimulation in mice, demonstrate causal relationships18. Rat cerebellar output to focal regions in the sensorimotor cortex and primate cerebellar output to specific motor and nonmotor regions have been mapped by transsynaptic tracers19, 20. Functional imaging at cellular resolution in nonhuman animals has been made possible by visualizing c-Fos, an immediate-early gene product whose expression is regulated by neural activity. Although useful in demonstrating coupling between distant brain regions, these methods do not provide cellular-resolution information about cerebello-cortical circuits.
Pathways entering and exiting the cerebellum pass through synapses in the brainstem and the cerebellum itself, blocking the passage of most cellular tracer molecules. However, this problem can be overcome using transsynaptically transported viruses21. The H129 strain of herpes simplex virus type 1 (HSV-H129) is an anterograde tracer that can identify long-distance targets of specific brain regions. For retrograde tracing, the Bartha strain of pseudorabies virus (PRV-Bartha) allows efficient synapse-specific transport. Thus recent molecular technology opens a means of mapping the cerebellum to its brainwide information-processing partners.
Conventional histological methods are too laborious for quantifying connectivity in the whole brain at once. But with the recent advent of optimized tissue clearing techniques with light-sheet microscopy22, the same tracing methods can now be scaled to cover entire brains. The resulting imaging datasets can occupy terabytes, creating a need for computationally efficient cell detection and anatomical assignment. These challenges can be addressed using machine learning algorithms to detect neurons and image registration methods to align brains. For the cerebellum, an additional problem is the absence of a reference template: the current field standard, the Allen Brain Atlas, omits the posterior two-thirds of the cerebellum. Any integrative study of cerebellar anatomy and function must therefore start with the creation of a suitable atlas.
In this project we used HSV-H129 to map the cerebellum’s direct ascending outputs to the thalamus and striatum and trisynaptic paths to the neocortex, and PRV-Bartha to map descending paths from neocortex to cerebellum. We then used these measurements to generate a brainwide atlas of cerebellum-forebrain connectivity. We developed an analysis pipeline that allows per-region cell counts to be converted from cell counts to per-volume cell density, giving a measure of relative impact on local circuitry. The potential impact of ascending paths was confirmed using optogenetic stimulation of c-Fos expression. All measurements were referred to a whole-brain atlas that includes the entire cerebellum. Taken together, our results provide a brainwide quantitative map of the cerebellum’s connectivity partners and provide insight into the extent of divergence and convergence in pathways between neocortex and cerebellum, as well as possible cerebellar contributions to whole-brain function and neurocognitive disorders.
RESULTS
HSV-H129 transsynaptic viral labeling reveals distant cerebellar targets
To trace transsynaptic pathways from cerebellum to midbrain and neocortex, we used HSV-H129-VC22 (Figure 1a), an HSV-H129 recombinant virus that expresses enhanced green fluorescent protein (EGFP) targeted by means of a localization sequence to the cell nucleus. Transsynaptic viral tracing yields weaker labeling of cells compared with longer-expression-time strategies such as AAV-driven fluorophore expression. To achieve a high signal-to-noise ratio we used iDISCO+, a method that combines tissue clearing, whole-brain immunostaining using Dako anti-HSV antibody, and light-sheet microscopy.
To determine the optimal timepoints for examining primarily disynaptic (i.e. Purkinje cell to cerebellar nuclear to thalamic) and primarily trisynaptic (Purkinje cell to cerebellar nuclear to thalamic to neocortical) targets, we injected H129-VC22 into the cerebellar cortex of mice and examined tissue between 12 and 89 hours post-injection (hpi; Figure 1b,c). At 54 hpi, labeling was observed in thalamus, with little visible labeling in neocortex (Figure 1c,d), and we used this as the thalamic timepoint (Table 1). Labeling was seen in other midbrain and hindbrain areas, consistent with known monosynaptic anterograde targets of the cerebellar and vestibular nuclei23, 24. Neocortical labeling was first visible at 73 hpi and throughout its extent by 82 hpi; we defined 80 hpi as our neocortical timepoint. These timepoints are consistent with prior studies using transsynaptic tracing with conventional histological-sectioning methods spanning two synapses (60-72 hpi; Refs. 2,25; Table 6) and three synapses (80-96 hpi; Refs. 2,25).
HSV-H129 has previously been reported to have a slow retrograde component26, which might occur via uptake by axon terminals27 followed by transport backward across synapses. To test this, we examined dorsal column nuclei, which are two synaptic steps retrograde from the cerebellar cortex (via mossy fibers) but which receive no known cerebellar-nuclear or vestibular-nuclear innervation (Supplementary Figure 1). At 28-36 hpi, HSV-H129 led to minimal visible labeling in the dorsal column nuclei). We normalized injections by dividing cell density by anterogradely labeled cell density in the DCN. The retrograde:anterograde density ratio for HSV-H129 injections was a median of 0.084 ± 0.032 at 28-36 hpi (estimated SEM, n=15 locations in 5 mice), and 0.096 ± 0.017 at 54 hpi (n=69 locations in 23 mice). To demonstrate the maximum possible amount of retrograde labeling we used PRV (80 hpi) and found a median retrograde:anterograde density ratio of 2.06 ± 0.17 (n=75 locations in 25 mice). Finally, we ascertained where any retrograde HSV-H129 viral uptake would lead via subsequent anterograde spread by examining the Mouselight database28 (http://ml-neuronbrowser.janelia.org/). We found 36 fully-traced brainstem (medulla and pons) neurons with at least one axon projecting into the cerebellum. In all but one of these neurons, the cerebellum was the sole target (Supplementary Figure 2). Thus, to the extent that retrograde transport from injection sites occurs, it would still not lead to alternate noncerebellar pathways. For purposes of transsynaptic tracing we conclude that at our selected timepoints, HSV-H129 acts as an anterograde tracer.
Automated cell detection using a convolutional neural network
Each brain generated a dataset exceeding 100 gigabytes. To automate cell detection, we trained a three-dimensional convolutional neural network (CNN) to recognize neurons. A CNN with U-Net architecture running on a GPU-based cluster was trained by supervised learning using more than 3600 human-annotated centers of cells as ground truth (Figure 1e; Table 2). The performance at different likelihood thresholds was plotted as a receiver-operator curve of precision and recall (Figure 1f), where precision was defined as the number of true positives divided by all positives, and recall was defined as the number of true positives divided by the number of true positives plus false negatives. A threshold likelihood of 0.6 was found to maximize the harmonic mean of precision and recall, a quantity known as the F1 score. Querying the CNN with the testing dataset gave an F1 score of 0.864, nearly the F1 score for human-human concordance, 0.891, indicating that the CNN had successfully generalized to whole-brain datasets.
Generation of the Princeton Mouse Atlas
To overcome past difficulties in registering images taken using different modalities, we devised a two-step procedure to calculate an averaged light-sheet-based brain template for referral to the Allen Brain Atlas (Figure 2). After this procedure, we fitted individual light-sheet brains to this template. The Allen Brain volumetric Atlas (ABA), a field standard, is based on serial two-photon microscopy and lacks a complete cerebellum (Figure 2a). To remedy that lack and to generate a template useful for our light-sheet images, we constructed a Princeton Mouse brain Atlas (PMA; Supplementary Figure 3). To make the PMA compatible with Allen standards, we computed a transform to convert it to Allen Brain CCFv3 space (Figure 2b). We then extended the space using manually-drawn contours to generate a complete, annotated cerebellar anatomy (Figure 2c,d) that included posterior lobules (Figure 2c,d red lines; Supplementary Figure 4).
To quantify the precision of atlas registration, we asked blinded users to find readily identifiable points in our atlas, in four sets of unregistered, affine-only registered, and fully registered volumes (Figure 2b, Supplementary Figure 5). After registration, the median Euclidean distance from complementary points in the PMA was 93 ± 36 µm (median ± estimated standard deviation) to b-spline registered volumes. Blinded users determined points in the same volume twice to establish an intrinsic minimum limit of 49 ± 40 µm. Assuming that uncertainties sum as independent variables, the estimated accuracy of the registration method was √ (932- 492)=79 µm, or 4 voxels.
The cerebellum sends output to a wide range of thalamic targets
We used our automated analysis pipeline, which we named BrainPipe, to quantify cerebello-thalamic connectivity (Figure 3a). We injected 23 brains with H129-VC22 at different sites in the posterior cerebellum (Figure 3b; Supplementary 6,7) and collected brains at 54 hpi, the thalamic timepoint. At this time, the number of neurons per region were widely distributed among contralateral thalamic regions (Figure 3c). The density of neurons observed in neocortical regions was 0.085 ± 0.073 (mean ± standard deviation, 17 regions) times that seen in 80 hpi injections, indicating that sufficient time had elapsed to allow transport to thalamus but not to neocortex. Number of neurons by region were not systematically related to anteroposterior position (rank correlation with anteroposterior position r=+0.05), suggesting that the efficiency of labeling was not strongly dependent on differences in transport distance. For display, the number of neurons for each region were converted to percentage of total per-brain thalamic neurons and coded according to “sensory/motor” and “polymodal association” functionalities based on ABA ontology (Figure 3, yellow/green shading).
The cerebellothalamic tract originates from the cerebellar nuclei and ascends through the superior cerebellar peduncle (also known as brachium conjunctivum), with most axons crossing the midline before reaching the thalamus. We observed vestibular nuclear labeling after cerebellar cortical injections, consistent with a direct projection from the cerebellar cortex (Supplementary Figure 8). Short-incubation experiments showed that vestibular nuclei contained 22% of the total combined vestibular and cerebellar-nuclear cell count. To assess the contribution of these paths to vestibular nuclei, we examined the Mouselight database28 and found 9 Purkinje cells with direct vestibular-projecting axons, 6 of which were found in non-flocculonodular regions (Supplementary Figure 8). Thus a substantial fraction of cerebellar projections to vestibular nuclei arise from widely distributed lobules.
A principal target of cerebellothalamic axons is the thalamic ventral nuclear group23, 29, which include somatosensory and somatomotor nuclei30. Consistent with known DCN projections, we observed strong connectivity to ventromedial (VM) and ventral anterior-lateral (VAL), motor thalamic nuclei31, from vermal lobules (I-VII) and Crus II, and moderate connectivity from Crus I (Figure 3d,e). The ventral posteromedial nucleus (VPM), which conveys somatosensory whisker and mouth information, received diverse input from all but the most extreme posterior cerebellar areas (Figure 3c,d,e) consistent with known interpositus and vestibular nuclear projections24, 32. These findings confirm that cerebellar-injected H129-VC22 labels major known pathways to neocortex via multiple distributed cerebellar-nuclear, vestibular, and thalamic intermediates.
We also observed labeling outside the ventral thalamus, including the thalamic reticular nucleus (TRN), lateral posterior (LP) and mediodorsal (MD) nuclei, and zona incerta33. MD and LP are association thalamic nuclei. MD is engaged during reversal learning34, sends its output to frontal regions, including insular, orbital, and prelimbic cortex35, and is engaged in cognitive and working memory tasks in humans34. Lobule VI, a site of structural abnormality in ASD36, made dense projections to MD (Figure 3d,e). These results suggest a strong role for cerebellum in flexible cognition. LP sends its output to primary somatosensory cortex, primary and secondary motor cortex, and frontal association area35. TRN, unlike other thalamic nuclei, does not project to neocortex, instead sending inhibitory projections to other thalamic nuclei. Thus, major paths from cerebellum to thalamus include both relay nuclei and the other two major classes of nuclei, association (MD, LP) and local modulatory (TRN).
To identify specific topographical relationships, we fitted a generalized linear model (GLM; Figure 3d), using the fraction-by-lobule of the total cerebellar injection as input parameters, and the fraction-by-nucleus of total thalamic expression as output measurements. The GLM revealed a broad mapping of lobules I-X to a variety of thalamic targets, and a more focused pattern of mapping from simplex, crus I and II, paramedian lobule, and copula pyramidis to specific thalamic targets. Hotspots of mapping included lobules I-X to VPM, TRN, MD, VM, VPL, VA-L, paraventricular, anteromedial, and centrolateral; simplex to VPM, LP, dorsal lateral geniculate, medial geniculate, and anterodorsal; crus I to VPM, lateral dorsal, reuniens, paraventricular and anteromedial and central lateral; crus II to zona incerta, VPM, MD, TRN, VA-L, posterior complex, and LP; and paramedian lobule and copula pyramidis to zona incerta, LD, parafascicular, ventral lateral geniculate, and lateral habenula (Table 3).
Direct projections from cerebellar nuclei to thalamus are largely consistent with transsynaptic tracing
As a second, non-transsynaptic approach to characterizing cerebellar projections to thalamus, we injected adeno-associated virus containing the GFP sequence into cerebellar nuclei and characterized the spatial distribution of fluorescent nerve terminals (Figure 4). Injections (n=4) of 125 nl (titer 7×10¹² vg/mL) primarily targeted bilateral dentate nuclei and also reached interposed and fastigial nuclei (Figure 4a; Supplementary Figure 9). Three weeks after injection, animals were sacrificed and brains sectioned and imaged by confocal microscopy (Figure 4b).
Terminals were clearly visible throughout thalamic sites, largely contralateral to the site of injection. Counts of vGluT2 and YFP co-labeled varicosities in thirteen randomly picked regions in VM, VA-L, and CL (each region 100×100×5 microns) were strongly correlated with average YFP brightness for that same region (r=+0.94, t=8.76, p<0.0001; Supplementary Figure 10). Therefore we used summed brightness as a measure of total innervation. Summed brightness was defined as the total fluorescence within a nucleus, summed across all sections where the nucleus was present. Overall, the highest summed brightness was found in ventral thalamic nuclei including VM, VA-L, VPM, and VPL, consistent with previous literature reports and with the density of cells observed in HSV-H129 injections. The nucleus-by-nucleus fluorescence density (i.e. summed brightness divided by the total area covered by the nucleus in the analyzed images) was correlated with the HSV-H129 neuron density averaged across all injections (Figure 4c; log-log correlation r=+0.59, p=0.023). Taken together, these measurements indicate that HSV-H129 injections at the thalamic timepoint capture a representative pattern of projection from cerebellar nuclei to contralateral thalamus.
Cerebellar nuclei project directly to the reticular thalamic nucleus
To identify neurons forming a monosynaptic projection to TRN, we injected into the TRN an adeno-associated virus (AAV) that infects presynaptic terminals and expressed retrogradely along axons to the parent neuronal cell body (AAVrg; Ref. 37). Injection into TRN of AAVrg-hSyn-Chronos-GFP38 induced GFP expression in a subset of neurons contralateral to the injection in the ventrolateral dentate and dorsolateral interpositus nuclei (Figure 4d-g; Supplementary Figure 11), consistent with prior reports39–41 (Supplementary Figure 12). This pattern of retrograde labeling in the cerebellar nuclei using a non-transsynaptic viral tracer (AAVrg) supports the thalamic-timepoint HSV labeling data which indicate a direct projection from cerebellar nuclei to TRN. An injection that missed TRN and instead infected nearby internal capsule (Supplementary Figure 11) did not label neurons in cerebellar nuclei. In summary, AAVrg-based retrograde labeling from TRN confirms a monosynaptic projection from cerebellar nuclei and is consistent with the disynaptic projection from cerebellar cortex uncovered by HSV-H129 injections.
Cerebellar paths to neocortex are strongest in somatomotor regions and densest in frontal regions
To characterize cerebellar paths to neocortex, we examined 33 HSV-injected brains at 80 hpi (Figure 5a,b,c). As expected, the majority of contralateral neocortical neurons were found in somatosensory and somatomotor areas, with additional neurons found at more anterior and posterior locations (Figure 5d). No output projection patterns were identified within subregions of the somatosensory and somatomotor areas (Supplementary Figure 13).
When converted to density, a different pattern of projection density became apparent (Figure 5e). The highest densities of neurons were found in contralateral anterior and medial neocortical regions, with peak regions exceeding 400 neurons per mm3, more than twice the highest density found in somatosensory and somatomotor regions. The most densely labeled regions included infralimbic, orbital, and prelimbic areas but excluded the frontal pole (Figure 5e).
To build a single map from the results of many injections, we fitted a GLM to the data in the same way as for thalamic labeling. Sensorimotor and frontal regions were strongly represented in the model weights. The GLM also sharpened the cerebellocortical topographical relationship (Figure 5d). All injected cerebellar sites showed high weighting in somatomotor and somatosensory cortex. In addition, lobules I-V showed significant weights in anterior cingulate cortex. Weak clusters of connectivity were also visible in visual and retrosplenial cortex. Mean neuron density by primary injection site (Figure 5g) revealed that all injected cerebellar sites sent dense projections to infralimbic cortex. Cerebellar vermis (lobules VI-X) and crus I sent denser projections than other cerebellar injection sites to infralimbic, prelimbic, and orbital cortex (Figure 5g). A similar pattern was observed by taking the maximum of the fraction of neurons across each cerebellar region, where the majority of neurons were found in the somatosensory and somatomotor cortex and a smaller number of neurons were found the retrosplenial, agranular insular, anterior cingulate, and orbital cortex (Supplementary Figure 14).
Cerebellar paths reach reward-based structures in striatum and hypothalamus and project modestly to ventral tegmental area
Among other putative monosynaptic targets of the cerebellar nuclei, an area of renewed focus has been the ventral tegmental area (VTA)42, 43, including a recent report of cerebellar influence over reward processing44. We used our anterograde tracing pipeline to compare the relative projection strengths of contralateral cerebellar paths to thalamus and two midbrain dopaminergic areas, VTA and the substantia nigra (Supplementary Figure 15). We found that the total number of neurons in contralateral VTA45 was considerably lower than in thalamic regions, consistent with known tracing23, 32, 43, 44. Normalized to density per unit volume of the target region, VTA projections were less than one-third as strong as projections to VPM, MD, and TRN. Cell count densities in substantia nigra (SNr and SNc) were even lower than in VTA. In summary, cerebellar projections to VTA constituted a moderate-strength projection, smaller in strength than thalamic pathways but greater than other dopaminergic targets.
Like the VTA, striatal regions are also involved in reward learning. The cerebellar cortex is known to project to basal ganglia trisynaptically via the cerebellar nuclei and thalamus46. Among striatal regions, at our neocortical labeling timepoint we observed the most labeling in the caudate, nucleus accumbens, and cortical amygdala. Labeling was dense in nucleus accumbens, septohippocampal and septofimbrial nuclei as well as central/medial amygdala (Supplementary Figure 16). At the thalamic and neocortical timepoints we also quantified hypothalamic connectivity observing relatively strong expression in the lateral area as well as the periventricular nucleus. We also observed high variability in projection density, likely related to the small volumes of hypothalamic nuclei (Supplementary Figure 17). At both time points, we observed strong labeling in the lateral hypothalamic area, which has been shown to regulate feeding and reward47 and the zona incerta, a well established recipient of DCN output48.
Cerebellum-neocortical paths strongly innervate deep neocortical layer neurons
To investigate the layer-specific contributions of cerebellar paths to neocortex, we examined laminar patterns of expression at the neocortical time point of H129-VC22 injections (Figure 6). To minimize near-surface false positives, 60 µm was eroded from layer 1. In most neocortical areas, the most and densest anterogradely labeled neurons were found in layers 5, layers 6a and 6b (Figure 6b,c). No differences were found among the layer-specific patterns resulting from injections to anterior vermis, posterior vermis, and posterior hemisphere (p>0.95, ANOVA, two-tailed, 3 injection groups).
The layer-specificity of thalamocortical connections varies by neocortical region49, 50. A common motif of thalamocortical projections is strong innervation of layer 6 neurons, especially in sensory regions51, 52. In sensorimotor regions (somatomotor and somatosensory), over 40% of our labeled cells were found in layer 6, a higher fraction than in other categories of neocortex (Figure 6b). To validate these findings, we injected H129-VC22 at midline lobule VIa in Thy1-YFP mice, which express YFP primarily in layer 5. These injections revealed viral labeling in neocortex subjacent to YFP (Supplementary Figure 18).
Sensory regions are known to receive thalamic innervation of layer 4 neurons51, However, classical tracing typically does not identify the cellular target, only the cortical layer where the synapse occurs53. We found that labeled layer 4 neurons comprised only 10% of cells in somatosensory cortex and even less in other sensory regions (gustatory, visceral, temporal, visual). Our results are consistent with the fact that thalamocortical synapses often occur in a more superficial layer than the postsynaptic neuron54.
A different pattern was seen in rhinal cortex, which forms part of the medial temporal system for declarative memory. Rhinal regions (perirhinal, ectorhinal, and entorhinal) had the highest fraction of layer 2/3 neurons (Figure 6c,d,e). This finding recalls the observation that in associative neocortical regions, thalamocortical axons send substantial projections to superficial layers52. Frontal and other association regions showed patterns that were intermediate between sensorimotor and rhinal regions, while infralimbic, prelimbic, orbital, and anterior cingulate cortex also received more and denser projections to layer 1 (Figure 6c,d,e). The share of labeling found in in layer 5 and 6 neurons was higher for frontal nonmotor regions than for other cortical areas. Taken together, our analysis reflects past findings that thalamic influences on neocortex arrive first through superficial and deep layer pathways54 (Figure 6f).
Pseudorabies virus reveals strong descending somatomotor influence
To characterize descending paths from neocortex to the cerebellum, we performed a series of injections of pseudorabies virus Bartha strain (PRV-Bartha), a strain that travels entirely in the retrograde direction (Figure 7a,b,c). In pilot experiments, expression was strong in neocortex at 80 hpi. To isolate layer 5 neurons, whose axons comprise the descending corticopontine pathway, we analyzed neurons registered to deep layers, which comprised 64% of all contralaterally labeled neocortical neurons (Supplementary Figure 19).
Similar to the anterograde tracing results, we found the largest proportion of neurons in somatosensory and somatomotor areas (Figure 7d,f, Supplementary Figure 13). Normalized to volume, neuron densities were highest in somatosensory, somatomotor, and frontal cortex (Figure 7e,g). Two regions identified as sources of corticopontine axons by classical tracing55 were labeled: anterior cingulate areas from injection of lobule VI and VII, and agranular insular cortex from crus II. In addition, retrosplenial and auditory areas were labeled from injection of paramedian lobule and copula pyramidis.
A GLM fitted to the data by the same procedure as the HSV-H129 tracing showed highest weighting in somatomotor, somatosensory, and frontal regions (Figure 7f). Weights in retrosplenial and visual cortex were smaller for vermal injections, and weights in gustatory, agranular insula, and visceral cortex were elevated for simplex and crus II injections. Averaging neuron density by primary injection site revealed all injected cerebellar sites received dense projections from somatomotor and somatosensory cortex. Lobules I-VII and crus II received denser projections from anterior cingulate and prelimbic cortex compared to other cerebellar injection sites. Crus II also received dense projections from infralimbic, agranular insula, gustatory, ectorhinal, and visceral cortex.
Descending corticopontine projections are known to be largely contralateral. To test the extent to which descending paths remain contralateral across multiple synaptic steps, we quantified the ratio of contralateral to ipsilateral cells for PRV-Bartha injections. Contralateral cells outnumbered ipsilateral cells in all major neocortical areas, with average contralateral-to-ipsilateral ratios of 1.4 in frontal cortex, 1.7 in posterior cortex, and 3.2 in somatomotor and somatosensory cortex. Contralateral-to-ipsilateral ratios were higher for hemispheric injection sites than vermal sites (Supplementary Table 1).
Ascending axonal projections of cerebellar nuclei are known to largely decussate to reach contralateral midbrain structures56. For H129-VC22 injections, we observed bilaterality at both the thalamic and neocortical timepoints. At the thalamic timepoint, the mean ratio of contralateral cells to ipsilateral cells was 2.5 in sensorimotor nuclei and 1.0 in polymodal association nuclei. Contralateral-to-ipsilateral ratios were highest for hemispheric injection sites (Supplementary Table 1). Taken together, our HSV-H129 and PRV-Bartha observations suggest that the organization of projections between cerebellum and neocortex is, by total proportion, to sensorimotor cortical areas, most strongly contralateral in pathways that concern movement, and more symmetrically distributed for nonmotor paths.
c-Fos mapping reveals brainwide patterns of activation consistent with transsynaptic tracing
The reciprocal paths we have identified suggest that cerebellum incorporates descending information and influences forebrain processing through diverse thalamocortical paths. To test whether the functional strength of ascending paths was commensurate with their anatomical connection, we measured expression of the immediate early gene c-Fos after optogenetic perturbation of cerebellar activity (Figure 8). c-Fos expression reflects cumulative neural activity and provides an independent means of quantifying long-distance influence. We expressed the hyperpolarizing proton pump ArchT in Purkinje cells by injecting rAAV1-CAG-FLEX-ArchT-GFP into the cerebellar vermis of L7-Cre+/- mice, using L7-Cre-/- mice as controls (Figure 8a). Inactivation of Purkinje cells, which inhibit neurons of the cerebellar nuclei, would be expected to have a net excitatory effect on thalamic and therefore neocortical activity.
Photostimulation led to reductions in Purkinje cell firing rate that were restricted to the light-flash period and led to no perturbations in arm speed during treadmill running (Supplementary Figure 20). After 1 hour photostimulation over lobule VI, either in mice expressing ArchT (Cre+/-) or in nonexpressing controls (Cre-/-), brains were removed and cleared using iDISCO+, then immunohistochemically stained for c-Fos using AlexaFluor-790 as the fluorophore, and analyzed using ClearMap22 (Figure 8b,c; Supplementary Figure 21) for comparison with HSV-H129 tracing (Figure 8d).
Fourteen structures were identified having both significant count differences by a nonparametric t-test and an activation ratio (defined as stimulation-group c-Fos average count divided by control-group average) greater than 2.5 (Figure 8e,f). The strongest activation ratios occurred in the anterior cingulate cortex, centrolateral nucleus of the thalamus, and the nucleus accumbens (Figure 8f). Lobule VI itself also showed elevated c-Fos counts, as expected for pulsed-light inactivation of Purkinje cells57. A voxel-wise t-test on cell count volumes (Supplementary Figure 22) showed strong c-Fos expression in frontal neocortical regions, especially in deep and middle neocortical layers. These findings are consistent with transsynaptic tracing using both automated analysis of cleared tissue (Supplementary Figure 19) and standard tissue sectioning and epifluorescent microscopy (Supplementary Figure 18).
Among neocortical regions, mean c-Fos stimulation-to-control cell density ratios and H129-VC22 expression density was highly correlated (Figure 8g; c-Fos ratio vs. log HSV expression r=+0.66, p=0.004), indicating that brainwide patterns of neural activity coincide with patterns of ascending polysynaptic targets from lobule VI. Subcortical examination of c-Fos brains revealed further broad similarities in expression with H129-VC22 labeling, including pontine nuclei, midbrain, superior colliculi, and hypothalamus (Supplementary Figure 23 and 24). Overall, these data show that c-Fos-based measures of brain activation coincide well with patterns of anatomical projection as measured by transsynaptic viral labeling.
DISCUSSION
We found that ascending synaptic paths from the cerebellum can be classified into three homologous systems serving sensorimotor, flexible cognitive, and modulatory functions (Figure 9). Parallel paths to several of these systems often originated from a single injection site. Well-known sensorimotor regions contained the most connections by total proportion, but nonmotor paths achieved comparable or higher local connection densities. Overall, these paths reached nearly all parts of neocortex after passing through a variety of thalamic, striatal, and midbrain structures.
In both neocortex and thalamus, the majority of total count of neurons labeled by anterograde (HSV-H129) or retrograde (PRV-Bartha) viruses were found in structures classified as sensorimotor, including ventral anterior-lateral (VAL; some portions termed ventrolateral (VL) in older atlases58) and ventromedial (VM) thalamic nuclei59. By per-volume density of labeled neurons, the strongest ascending projections went to anterior cingulate, prelimbic and infralimbic cortex, as well as agranular and orbital areas. We also observed substantial relative labeling in the VPM (sensorimotor), reticular thalamic (modulatory), and mediodorsal (associative) nuclei, providing substrates for a variety of brain functions.
Although paths from cerebellum to thalamus and neocortex have been well-reported in the past, differences from previous approaches may arise for several reasons. Traditional tracers that do not cross synapses emphasize large presynaptic axons and extend only to their terminations, which can be up to hundreds of microns away from postsynaptic neuronal cell 18 bodies. Such a distinction between axonal and transsynaptic tracing may account in part for the strength of our observed projection to TRN neurons, which have elaborate dendritic processes that extend into other thalamic nuclei60. Similarly, neocortical layer 5/6 pyramidal neuron dendrites extend to superficial layers where thalamocortical synapses occur, and presynaptic fiber tracing reveals connectivity to some superficial layers53 while electrophysiological recordings identify cortical recipient neurons in deeper layers54. Our neocortical labeling is consistent with electrophysiological recordings.
A major architectural theme of cerebellar-neocortical interaction is that of closed loops, in which it has been suggested that each region of neocortex may have a small number of principal partners in the cerebellum, and vice versa11. Our work builds upon and broadens that picture. Ascending cerebellar paths are influenced by a variety of both motor and nonmotor descending information, thus forming a bidirectionally communicating loop between the neocortex and cerebellum. Descending pathways from the neocortex were most strongly concentrated in somatomotor and somatosensory cortex. Consistent with recent anterograde tracing from neocortex to mossy fibers61, by far the strongest neocortical paths to cerebellum originated from somatosensory and somatomotor sites. Although that study did not encompass frontal nonmotor neocortex, transsynaptic tracing in rat15 does align with our observation of strong input from frontal areas (prelimbic, infralimbic, insula and orbital cortices) to crus II. Taken together, connectivity patterns observed suggest that the cerebellum receives information from a wide variety of neocortical sites to exert returning influence on a similarly broad range of thalamic and neocortical targets.
Nonmotor functions of the cerebellum
Among the cerebellar injection sites, nonmotor functions have been suggested for lobule VI in the posterior vermis, and crus I and II in the posterior parts of the hemispheres. We found that lobule VI sent strong projections to mediodorsal and paraventricular nucleus of thalamus and to frontal neocortical regions, which serve a wide range of cognitive and emotional functions62, 63. Because the refinement of neural circuitry is 19 activity-dependent64, this projection may also potentially account for the observation that cerebellar perturbation of lobule VI can affect cognitive and social development in rodents2 and humans6, and the association of posterior vermal abnormalities with a high risk of ASD36. Optogenetic stimulation of lobule VI also led to strong activation of c-Fos in the nucleus accumbens (NAc), the main component of the ventral striatum, which is implicated in reward learning and motivation65. This observation is consistent with our observation of NAc labeling at the neocortical timepoint of HSV anterograde tracing (Supplementary Figure 18).
We found that crus I projects to lateral dorsal and paraventricular nuclei of the thalamus, as well as frontal neocortical regions. Crus I has previously been observed to be activated during working memory66. Novel projections from the DCN through lateral dorsal thalamus to regions involved in working memory, the hippocampus and retrosplenial cortex, have been described previously in the mouse67. In mice, disruptions of crus I activity in adulthood or juvenile life lead to deficits in adult flexible behavior2, 68; adult disruption shortens the time constant of a working memory task3. Crus I also projects to paraventricular nucleus62, which provides a possible substrate for the observation that juvenile disruption of crus I leads to long-lasting deficits in social preference2. In addition, we observed dense expression in septal and amygdalar regions consistent with pathways established in rat and cat69.
The cerebellum is known to be a site of sensory gating70, 71. Tracing to thalamus suggests that such gating might contribute to a broader regulatory network. Lobules (I-VII) and crus II sent strong paths to thalamic reticular nucleus (TRN), a known monosynaptic target of the cerebellar nuclei72. Our identification of neurons in the dentate and interpositus cerebellar nuclei that project to TRN agrees with similar studies in rat39, 40 (Supplementary Figure 12) and cat41.
TRN is the only thalamic nucleus that is inhibitory and the only one to send projections exclusively within the thalamus itself. TRN may control sensory gain73 and the flow of information in and out of the neocortex74. TRN also receives a strong descending projection from neocortical layer 674, 75, a site of prominent expression in our work. This descending projection completes an 20 inhibitory loop, and has been suggested to contribute to neocortical oscillations and synchrony76, 77. Our findings add cerebellum as a substantial contributor to this modulatory thalamocortical network.
A pipeline for long-distance transsynaptic mapping
Although many individual projections within these pathways have been previously reported, our work presents a brainwide survey of their relative strength. Polysynaptic transsynaptic tracing studies relied on time-consuming human identification for analysis. More recently, tissue clearing has been used for volumetric histological analyses, with the recent introduction of automated methods for cell identification22. We find that cell counting can be efficient, accurate, and scalable to the whole brain. Our mapping project relied on our BrainPipe pipeline, which combines transsynaptic tracing, whole-brain clearing and microscopy, automated neuron counting, and atlas registration. BrainPipe should be scalable for larger datasets as the resolution of light-sheet microscopy improves. Adapting BrainPipe to other experimental studies requires only a different annotated dataset to train a new convolutional neural network to identify objects of interest. Our pipeline can run on high-performance computing clusters, allowing for faster turnaround of results than other analysis pipelines, such as ClearMap22.
In creating our light-sheet brain atlas, we overcame the general problem of creating a reference atlas for a different imaging modality from the ABA. Our solution consisted of three steps: (1) align individually imaged brains to a single experimental brain serving as the initial template, (2) average the post-aligned brains to obtain a project-specific atlas for precise automated registration, and (3) learn the transform between the project-specific atlas and the field standard. Our basic software package (github.com/PrincetonUniversity/pytlas) is capable of efficiently generating atlases for other imaging modalities as well.
We took two approaches to quantifying relative projection strength: fraction of total expression for an entire parent structure (e.g. whole thalamus or neocortex) or density of cells within a nucleus or region. Fraction of total expression allows for relative comparison of projection 21 strengths within a parent structure, conveying information about the distribution of influence. In contrast, density takes the local target structure into account and provides information about the concentrated influence on a particular recipient target. For example, although the great majority of projections to neocortex are found in somatomotor and somatosensory cortices, the smaller prefrontal areas receive a higher density of projections. This suggests cerebellum’s capacity for influence on prefrontal areas, while smaller in total terms, might still provide a means by which cerebellar disruptions lead to deficits in nonmotor behavior2, 68.
Our HSV-based transsynaptic approach was designed to identify disynaptic paths from cerebellar cortex to thalamus and trisynaptic paths to neocortex. Our control experiments using shorter timepoints (28 and 36 hpi) and brainstem analysis at the disynaptic timepoint indicate that retrograde transport was minimal. However, a variety of longer anterograde paths are possible. The fastigial nuclei have bilateral efferents to the ipsilateral brachium conjunctivum (BC) and, via the uncinate fasciculus to the contralateral BC10, 78 and the cerebellar nuclei project to hindbrain/midbrain targets in addition to thalamus10, 79. Tracing from neurons in the dentate (lateral) nucleus has been shown to have axons crossing the thalamic midline terminating in the ipsilateral VL40, 80 (current atlas definitions mostly of VA-L; Supplementary Figure 12). Indeed, we observed contralateral crossing of axons after AAV injections in the fastigial nucleus (Supplementary Figure 9). Over long distances, where transport time is increasingly dominated by axon-associated transport mechanisms81, HSV-H129 may follow such alternative paths, as well as retrograde paths for incubations of 96 hours or longer. We therefore restricted our analysis to contralateral projections and the shortest necessary incubation times, decreasing risk for asynchronous or retrograde spread. The correlation of the resulting observed labeling with c-Fos activation suggests that our observations reflect major routes by which the cerebellum influences neocortical function.
Our cerebellum-to-thalamus projection data is broadly consistent with the literature, though with more emphasis on sensory (e.g. VPM, DLG) over motor (e.g. VAL) structures. This difference may arise from HSV-H129’s ability to track the summed contributions of multiple pathways with different intermediates. Convergent cerebellar projections to motor and sensory neocortex have been recently characterized in the rat19. Some connectivity to sensory thalamus may arise from vestibular nuclei24, 82, 83. Vestibular nuclei receive a well-known input from Purkinje cells in flocculonodular cerebellum, but we found even more total input from other cerebellar regions. In this way, cerebellar influence over a thalamic or neocortical target may be transmitted via more than one path. Additional contrasts may arise from differences in the uptake, selectivity and spread of tracers84–86. In particular, HSV spreads slower in large axons than small axons87, which may emphasize sensory pathways.
From local cerebellar circuitry to global brain function
The local circuitry of cerebellum is thought to make rapid predictions about future states, which then modulate the activity of other brain regions. In the motor domain, evidence suggests error learning through a supervised learning process. Contextual information in this learning process comes from the mossy fiber-granule cell pathway, and a teaching signal comes from the climbing fiber pathway. This circuitry is homologous across cerebellar regions, with each part of the cerebellar cortex managing a massive convergence of diverse incoming information from a distinct assortment of distant brain regions. Initial fast and early learning occurs at parallel-fiber to Purkinje cell synapses and other synapses88. Cerebellar and vestibular nuclei have their own information-processing and learning rules and serve as an exit gateway to influence other brain regions10, 88. The cerebellum may generate predictions to fine-tune activity across nonmotor functions as well3, 6, 89 as it is composed of many modules, whose processing converges strongly onto cerebellar and vestibular nuclei. Although classically these modules are considered to have specific extracerebellar partners13, our tracing quantifies the extent to which ascending cerebellar output paths from a single site may provide the same predictive information to many targets with diverse functions at the same time.
METHODS
OVERVIEW OF AUTOMATED PIPELINE FOR TRANSSYNAPTIC TRACING
In order to identify and quantify cerebellar connectivity on the long distance scale, we developed a pipeline, BrainPipe, to enable automated detection of transsynaptically labeled neurons using the mostly anterogradely-transported HSV-H12926, identifying cerebellar output targets, and retrogradely-transported PRV-Bartha90, identifying the descending corticopontine pathway, comprised mostly of layer 5 pyramidal neurons91. Mouse brains with cerebellar cortical injections of Bartha or H129 were cleared using iDISCO+. We then imaged the brains using light-sheet microscopy, generating brain volumes with a custom Python package. Next, to ensure accurate anatomical identification across brains, we created a local light-sheet template, the Princeton Mouse Brain Atlas (PMA) and quantified registration performance of individual volumes to the local template. We then determined the transform between the PMA and the Allen Brain Atlas, enabling standardization of our results with the current field standard. Next, to automatically and accurately detect labeled cells, we developed a convolutional neural network whose performance approached that of human classifiers.
ANIMAL EXPERIMENTATION
Experimental procedures were approved by the Institutional Animal Care and Use Committees of Princeton University (protocol number 1943-19), the University of Idaho (protocol number 2017-66), and the Dutch national experimental animal committees (DEC), and performed in accordance with the animal welfare guidelines of the National Institutes of Health (USA) or the European Communities Council Directive (Netherlands).
VIRUS SOURCES
HSV-1 strain H129 recombinant VC22 (H129-VC22) expresses EGFP-NLS, driven by the CMV immediate-early promoter and terminated with the SV40 polyA sequence. To engineer this recombinant, we used the procedure previously described to construct HSV-772, which corresponds to H129 with CMV-EGFP-SV40pA26. We generated plasmid VC22 by inserting into plasmid HSV-772 three tandem copies of the sequence for the c-Myc nuclear localization signal (NLS) PAAKRVKLD92, fused to the carboxy-terminus of EGFP. Plasmid VC22 contains two flanking sequences, one of 1888-bp homologous to HSV-1 UL26/26.5, and one of 2078-bp homologous to HSV-1 UL27, to allow insertion in the region between these genes. HSV-1 H129 nucleocapsid DNA was cotransfected with linearized plasmid VC22 using Lipofectamine 2000 over African green monkey kidney epithelial cell line Vero (ATCC cell line CCL-81), following the manufacturer’s protocol (Invitrogen). Viral plaques expressing EGFP-NLS were visualized and selected under an epifluorescence microscope. PRV-152 (PRV Bartha90), which drives the expression of GFP driven by the CMV immediate-early promoter and terminated with the SV40 polyA sequence, was a gift of the laboratory of Lynn W. Enquist. Adeno-associated virus was obtained from Addgene (https://www.addgene.org).
IN VIVO VIRUS INJECTIONS
Surgery for HSV and PRV injections
Mice were injected intraperitoneally with 15% mannitol in 0.9% saline (M4125, Sigma-Aldrich, St. Louis, MO) approximately 30 minutes before surgery to decrease surgical bleeding and facilitate viral uptake. Mice were then anesthetized with isoflurane (5% induction, 1-2% isoflurane/oxygen maintenance vol/vol), eyes covered with ophthalmic ointment (Puralube, Pharmaderm Florham Park, NJ), and stereotactically stabilized (Kopf Model 1900, David Kopf Instruments, Tujunga, CA). After shaving hair over the scalp, a midline incision was made to expose the posterior skull. Posterior neck muscles attaching to the skull were removed, and the brain was exposed by making a craniotomy using a 0.5 mm micro-drill burr (Fine Science Tools, Foster City, CA). External cerebellar vasculature was used to identify cerebellar lobule boundaries to determine nominal anatomical locations for injection. Injection pipettes were pulled from soda lime glass (71900-10 Kimble, Vineland, NJ) on a P-97 puller (Sutter Instruments, Novato, CA), beveled to 30 degrees with an approximate 10 μm tip width, and backfilled with injection solution.
AAV deep cerebellar nuclear injections
During stereotaxic surgery, mice were anesthetized with isoflurane (PCH, induction: 5%; maintenance: 2.0-2.5%) and received a mannitol injection intraperitoneally (2.33 g/kg in milli-Q) and a Rimadyl injection subcutaneously (5 mg/kg Carprofen 50 mg/ml, Pfizer, Eurovet, in NaCl). Body temperature was kept constant at 37°C with a feedback measurement system (DC Temperature Control System, FHC, Bowdoin, ME, VS). Mice were placed into a stereotactic frame (Stoelting, Chicago laboratory supply), fixing the head with stub ear bars and a tooth bar. DURATEARS® eye ointment (Alcon) was used to prevent corneal dehydration. A 2 cm sagittal scalp incision was made, after which the exposed skull was cleaned with sterile saline. Mice were given 2 small (ر1 mm) craniotomies in the interparietal bone (-2 mm AP relative to lambda; 1.8 mm ML) for virus injection. Craniotomies were performed using a hand drill (Marathon N7 Dental Micro Motor). A bilateral injection of AAV5-Syn-ChR2-eYFP (125 nl per hemisphere, infusion speed ∼0.05 µl/minute) in the AIN was done using a glass micropipette controlled by a syringe. This AAV was used because it gave reliable strong axon terminal labeling and because the animals were also used for another optogenetic study. After slowly lowering the micropipette to the target site (2.2 mm ventral), the micropipette remained stationary for 5 minutes before the start of the injection, and again after finishing the injection. Micropipette was then withdrawn slowly from the brain (ejection speed ∼1 mm/minute). Craniotomies and skin were closed and mice received post-op Rimadyl. Animals were perfused transcardially 3 weeks after viral injection using 4% PFA. Brains were collected post mortem, stained for co-stained for DAPI (0100-20, Southern Biotech, Birmingham AL), coronally sectioned at 40 µm/slice and imaged with an epifluorescent microscope at 20x (Nanozoomer, Hamamatsu, Shizuoka, Japan).
AAV TRN injections
During stereotaxic surgery, mice (1-3 months of age) were anesthetized with isoflurane (VetOne, induction: 3-5%; maintenance: 1.5-2.5%). For analgesic support mice provided oral carprofen ad libitum from the day before and through 24hrs after surgery and given slow release meloxicam (4mg/kg; ZooPharm, Larami, WY). Body temperature was maintained by a warming blanket (Stoelting, Wood Dale, IL) under the animal throughout the surgery. Mice were placed into a stereotactic frame (Kopf, Tujunga, CA), fixing the head with non-rupture ear bars, a tooth bar and nose cone. Puralube® Vet eye ointment (Dechra) was used to prevent corneal dehydration. A sagittal scalp incision was made, after which the exposed skull was cleaned with sterile saline. A single small craniotomy (Ø0.6 mm) was made in the parietal bone (−1.3 mm AP relative to lambda; 2.3 mm to the right of ML) for virus injection. Craniotomies were performed using a stereotaxic-mounted drill (Foredom K.1070 Micromotor Drill). A unilateral 200-300 nl injection of AAVrg-hSyn-Chronos-GFP (9.0×1012; Addgene, Watertown, MA) at an infusion speed of 0.01 μl/minute in the right TRN was done using a glass syringe and needle (Hamilton Company, Franklin, MA). After slowly lowering the needle to the target site (-2.9 mm ventral), the needle remained stationary for 1 minute before the start of the injection, and for 5 min after finishing the injection. The needle was then withdrawn slowly from the brain. Craniotomies and skin were closed using removable staples and mice continued to receive oral carprofen ad libitum for 24hrs post-surgery. Animals were euthanized 20-25 days after viral injection and brains were fixed in 4% PFA. Brains were collected, frozen and coronally sectioned into 40 μm slices, then co-stained with Hoechst 33324 (5 µg/ml; Invitrogen), chicken anti-GFP (1:500; Novus Biologicals; NB100-1614), and rabbit anti-parvalbumin (1:500; ZRB1218; Millipore Sigma), and imaged with a confocal fluorescence microscope at 10X and 20x (Nikon Instruments TiE Inverted Microscope with Yokogawa X1 Spinning Disk) or epifluorescence microscope at 2.5X and 10X (Zeiss Axio Imager.M2).
Transsynaptic viral tracing for tissue clearing (HSV-H129 and PRV-Bartha)
Transsynaptic viral tracing studies used male and female 8-12 week-old C57BL/6J mice (The Jackson Laboratory, Bar Harbor, Maine). Injection solution was prepared by making a 9:1 dilution of virus stock to 0.5% cholera toxin B conjugated to Alexa Fluor 555 in saline (CTB-555, C22843, Sigma-Aldrich; as per Ref. 93). At the timepoints used CTB-555 persisted at the injection site. Pipettes were inserted perpendicular to tissue surface to a depth of approximately 200 µm. Table 4 describes injection parameters for each type of experiment.
Pressure injections delivered 80 to 240 nl into the target location. Consistent with prior literature we observed that minimum injections of 104 PFUs were required for successful HSV-H129 infection94. This viral feature dictated a lower injection size limit. Smaller injections consistently produced unsuccessful primary infections and thus no transsynaptic spread. Unfortunately, this feature also prevented consistent injections of single zones as defined by zebrin staining.
After viral injection, Rimadyl (0.2 ml, 50 mg/ml, Carprofen, Zoetis, Florham Park, NJ) was delivered subcutaneously. At the end of the post-injection incubation period, animals were overdosed by intraperitoneal injection of ketamine/xylazine (ketamine: 400 mg/kg, Zetamine, Vet One, ANADA #200-055; xylazine: 50 mg/kg, AnaSed Injection Xylazine, Akorn, NADA #139-236) and transcardially perfused with 10 ml of 0.1 M phosphate buffer saline (PBS) followed by 25 ml 10% formalin (Fisher Scientific 23-245685). Tissue was fixed overnight in 10% formalin before the iDISCO+ clearing protocol began.
For anterograde transport experiments, incubation times were determined by immunostaining for viral antigens at various timepoints (30, 36, 41, 49, 54, 58, 67, 73, 80, 82 and 89 hours post-injection) the canonical ascending pathway of cerebellar cortex to cerebellar nuclei to thalamus to neocortex. For retrograde transport experiments, incubation times were determined by immunostaining for GFP (48, 60, 72, 78, 81, 84 and 91 hpi) targeting the canonical descending pathway: neocortex to brainstem to cerebellar cortex. We selected timepoints with the goal of achieving sufficient labeling for detection, while minimizing incubation periods, given that with increasing long distance, transport time is increasingly dominated by axon-associated transport mechanisms81, 95–97, leading to labeling of alternative paths and retrograde paths after 96 hours26. Our selected timepoints were shorter than published timepoints (Table 6), and were therefore likely to reduce the degree of supernumerary synaptic spread.
Considerations when using transsynaptic tracing viruses
For practical use in tracing, transsynaptic viral tracers typically require doses close to the infectivity threshold. For H129, the minimum dose is 104 pfus94; consistent with our injections). Although injecting close to the infectivity threshold results in greater infection variability as well as failure to label all second- and third-order neurons, it ensures the most specific transport, protecting against nonspecific nontransneural transport98, 99.
To quantify possible retrograde HSV-H129 transport from the injection site27 (Supplementary Figure 1), we compared retrograde:anterograde density ratios for HSV-H129 and PRV at disynaptic (HSV-H129 54 hpi; PRV 80 hpi) timepoints. We used the dorsal column nuclei, disynaptically retrograde from cerebellar cortex that lacks anterograde projections from DCN, for retrograde spread. We used the DCN, monosynaptic anterograde, for anterograde spread. PRV provides the maximum amount of retrograde labeling.
To minimize unintended further downstream synaptic spread, we performed empirical time series and selected the minimal time that provided sufficient spread. We compared our time points with known literature and found that we never exceeded published timepoints and often had considerably shorter timepoints than other work (Table 6). Transsynaptic tracing is always at risk for loss of synchrony with viral particles traveling along parallel polysynaptic circuits. For example, the HSV-H129 neocortical timepoint will not be 100% thalamocortical target neurons, but may contain a small fraction of “leading-edge” cells that are one step further. However, our shorter-than-published timepoints suggests our studies minimize this risk. This also suggests that tissue clearing is likely a more sensitive method, allowing for unbiased whole-brain observation of viral spread. Previous studies, by nature of manual annotation, could only observe a small subset of brain regions without robust quantification. Every tool, viral or nonviral, has its own differential selectivity for different types of neurons; Alphaherpesviruses (such as PRV and HSV) have been shown to spread faster in sensory peripheral nerves than motor paths87, 100. Other viral tropism differences have been directly studied in Rabies versus AAV84 as well as tropism across AAV serotypes85, 86. Although HSV-H129 has its limitations with different spread characteristics across strains101 is it the most robust tool currently available for anterograde transsynaptic viral tracing. Therefore likely some of differences between our findings and classically established literature could be related to both the use of transsynaptic viruses as well as whole-brain statistical unbiased quantification.
VIRAL TRACING WITH TISSUE SECTIONING AND SLIDE-BASED MICROSCOPY
Viral tracing with classical sectioning-based histology: HSV-772 cerebellar injections
Adult Thy1-YFP male mice (YFP +, n=2, B6.Cg-Tg (Thy1-YFP)HJrs/J, 003782, The Jackson Laboratory, 22 weeks), were prepared for surgery, in a similar fashion as in Transsynaptic viral tracing for tissue clearing (H129 and Bartha). We used the HSV recombinant HSV-772 (CMV-EGFP, 9.02 x 108 PFU/ml; as in Ref. 26), a H129 recombinant that produces a diffusible EGFP reporter. Again, using a 9:1 HSV:CTB-555 injection solution, 350 nl/injection was pressure injected into two mediolateral spots in lobule VIa. Eighty hours post-injection, animals were overdosed using a ketamine/xylazine mixture as described previously. Brains were extracted and fixed overnight in 10% formalin and cut at 50 µm thickness in PBS using a vibratome (VT1000S, Leica). Sections were immunohistochemically blocked by incubating for 1 hour in 10% goat serum (G6767-100ML, Sigma-Aldrich, St. Louis, MO), 0.5% Triton X100 (T8787-50ML, Sigma-Aldrich) in PBS. Next sections were put in primary antibody solution (1:750 Dako Anti-HSV in 2% goat serum, 0.4% Triton X100 in PBS) for 72 hours at 4°C in the dark. Sections were washed in PBS 4 times for 10 minutes each, and then incubated with secondary antibody (1:300 Goat anti-rabbit-AF647 in 2% goat serum, 0.4% Triton X100 in PBS) for two hours. Another series of PBS washes (four times, 10 minutes each) before mounting onto glass microscope slides with a Vectashield mounting agent (H-1000, Vector Laboratories, Burlingame, CA). Sections were fluorescently imaged at 20x (Nanozoomer, Hamamatsu, Shizuoka, Japan) and at 63x with 5 μm z steps (Leica SP8 confocal laser-scanning microscope).
TISSUE CLEARING AND LIGHT-SHEET MICROSCOPY
iDISCO+ tissue clearing
After extraction, brains were immersed overnight in 10% formalin. An iDISCO+ tissue clearing protocol22 was used (Supplemental clearing worksheet). Brains were dehydrated step-wise in increasing concentrations of methanol (Carolina Biological Supply, 874195; 20, 40, 60, 80, 100% in doubly distilled H20 (ddH20), 1 hr each), bleached in 5% hydrogen peroxide/methanol solution (Sigma, H1009-100ML) overnight, and serially rehydrated (methanol: ddH20 100, 80, 60, 40, 20%, 1 hr each). Brains were washed in 0.2% Triton X-100 (Sigma, T8787-50ML) in PBS, then in 20% DMSO (Fisher Scientific D128-1) + 0.3 M glycine (Sigma 410225-50G) + 0.2% Triton X-100/PBS at 37°C for 2 days. Brains were then immersed in a blocking solution of 10% DMSO + 6% donkey serum (EMD Millipore S30-100ml) + 0.2% Triton X-100 + PBS at 37°C for 2-3 days to reduce non-specific antibody binding. Brains were then twice washed for 1 hr/wash in PBS + 0.2% Tween-20 (Sigma P9416-50ML) + 10 µg/ml heparin (Sigma H3149-100KU) (PTwH).
For HSV and c-Fos antibody labeling, brains were incubated with primary antibody solution (see Table 4 for antibody concentrations) consisting of 5% DMSO + 3% donkey serum + PTwH at 37°C for 7 days. Brains were then washed in PTwH at least 5 times (wash intervals: 10 min, 15, 30, 1 hr, 2 hr), immunostained with secondary antibody in 3% donkey serum/PTwH at 37°C for 7 days, and washed again in PTwH at least 5 times (wash intervals: 10 min, 15, 30, 1 hr, 2 hr). Finally, brains were serially dehydrated (methanol: ddH20: 100, 80, 60, 40, 20%, 1 hr each), treated with 2:1 dichloromethane (DCM; Sigma, 270997-2L):methanol and then 100% DCM, and placed in the refractive index matching solution dibenzyl ether (DBE; Sigma, 108014-1KG) for storage at room temperature before imaging.
Light-sheet microscopy for transsynaptic tracing
Cleared brain samples were glued (Loctite, 234796) ventral side down on a custom-designed 3D-printed holder and imaged in an index-matched solution, DBE, using a light-sheet microscope (Ultramicroscope II, LaVision Biotec., Bielefeld, Germany). Version 5.1.347 of the ImSpector Microscope controller software was used. An autofluorescent channel for registration purposes was acquired using 488 nm laser diode excitation and 525 nm emission (FF01-525/39-25, Semrock, Rochester, New York). Injection sites, identified by CTB-555, were acquired at 561 nm excitation and 609 nm emission (FF01-609/54-25, Semrock). Cellular imaging of virally infected cells (anti-HSV Dako B011402-2) was acquired using 640 nm excitation and 680 nm emission (FF01-680/42-25, Semrock). Cellular-resolution imaging was done at 1.63 µm/pixel (1x magnification, 4x objective, 0.28 NA, 5.6 - 6.0 mm working distance, 3.5 mm x 4.1 mm field of view, LVMI-FLuor 4x, LaVision Biotech) with 3×3 tiling (with typically 10% overlap) per horizontal plane. Separate left- and right-sided illumination images were taken every 7.5 micrometers step size using a 0.008 excitation-sheet NA. A computational stitching approach102 was performed independently for left- and right-side illuminated volumes, followed by midline sigmoidal-blending of the two volumes to reduce movement and image artifacts.
REGISTRATION AND ATLAS PREPARATION
Image registration
Most registration software cannot compute transformation with full-sized light-sheet volumes in the 100-200 gigabyte range due to computational limits. Using mid-range computers, reasonable processing times are obtained with file sizes of 300-750 megabytes, which for mouse brain corresponds to 20 µm/voxel. Empirically, we found that light-sheet brain volumes to be aligned (“moving”) resampled to approximately 140% the size of the reference (“fixed”) atlas volume yielded the best registration performance. Alignment was done by applying an affine transformations allowing for translation, rotation, shearing and scaling to generally align with the atlas, followed by b-spline transformation to account for brain-subregion variability among individual brains.
For uniformity among samples, registration was done using the autofluorescence channel, which has substantial autofluorescence at shorter wavelengths useful for registration103. In addition to autofluorescence-to-atlas registration, the signal channel was registered using an affine transformation to the autofluorescence channel to control for minor brain movement during acquisition, wavelength-dependent aberrations, and differences in imaging parameters22.
Affine and b-spline transformations were computed using elastix104, 105; see supplemental Elastix affine and b-spline parameters used for light-sheet volume registration. Briefly, the elastix affine transform allows for translation (t), rotation (R), shearing (G), and scaling (S) and is defined as: where c is a center of rotation and t is a translation. The elastix b-spline transformation allows for nonlinearities and is defined as: Where xk are control points, β3 (x) the B-spline polynomial, pk the b-spline coefficient vectors, Nx, B-spline compact support control points, and σ is the b-spline compact control point-spacing (see Ref. 106, pages 8-10 for reference). For the assignment of cell centers to anatomical locations, we calculated transformations from cell signal space to autofluorescent space (affine only) and autofluorescent space to atlas space (affine and b-spline; Supplementary Figure 25).
Princeton Mouse Atlas generation
To generate a light-sheet atlas with a complete posterior cerebellum, autofluorescent light-sheet volumes from 110 mice (curated to eliminate distortions related to damage, clearing, or imaging) were resampled to an isotropic 20 µm per voxel resolution (Figure 2 and Supplementary Figure 3a). We selected a single brain volume to use as the fixed (template) volume for registration of the other 109 brains and computed the transformations between the other 109 brains and the template brain. The registration task was parallelized from ClearMap22 adapting code for use on a Slurm-based107 computing cluster.
After registration, all brains were pooled into a four-dimensional volume (brain, x, y, z), and the median voxel value at each xyz location was used to generate a single median three-dimensional volume. Flocculi and paraflocculi, which can become damaged or deformed during extraction and clearing, were imaged separately from a subset of 26 brains in which these structures were intact and undeformed. Manual voxel curation sharpened brain-edges in areas where pixel intensity gradually faded. Finally, contrast limited adaptive histogram equalization (skimage.exposure.equalize_adapthist) applied to the resulting volume increased local contrast within brain structures, generating the final PMA (Supplementary Figure 3b and Supplementary Figure 26). We then determined the transformation between the PMA and the Allen Brain CCFv3108 space in order to maintain translatability. Our software for basic atlas creation with an accompanying Jupyter tutorial notebook is available online via github.com/PrincetonUniversity/pytlas. Volumetric projection renderings were made using ImageJ109; 3D project function (Supplementary Figure 3a). The PMA interactive three-dimensional rendering of the PMA is available http://brainmaps.princeton.edu/pma_neuroglancer and can be downloaded from https://brainmaps.princeton.edu/pma_landing_page.
Statistical analysis of registration precision
Precision of registration was measured by quantifying euclidean landmark distances, defined by blinded users (similar to Ref. 110) between the PMA and brains at different stages of registration. Estimated standard deviations are defined as the median absolute deviation (MAD) divided by 0.6745. MADs were calculated with Statsmodels111 0.9.0 (statsmodels.robust.mad). One measurement was considered to be user error and was dropped in the theoretical-limit measurements, as it was over 12 times the median of the other measures.
Generation of 3D printable files
To generate 3D printable Princeton Mouse Atlas files usable for experimental and educational purposes, we loaded volumetric tiff files as surface objects using the ImageJ-based 3D viewer. After resampling by a factor of 2 and intensity thresholding, data were then imported to Blender112, where surfaces were smoothed (Smooth Vertex tool) before finally exporting as stereolithography (stl) files.
AUTOMATED DETECTION OF VIRALLY LABELED CELLS
BrainPipe, an automated transsynaptic tracing and labeling analysis pipeline
Whole-brain light-sheet volumes were analyzed using a new pipeline, BrainPipe. BrainPipe consists of three steps: cell detection, registration to a common atlas, and injection site recovery. For maximum detection accuracy, cell detection was performed on unregistered image volumes, and the detected cells were then transformed to atlas coordinates.
Before analysis, datasets were manually curated by stringent quality control standards. Each brain was screened for (1) clearing quality, (2) significant tissue deformation from extraction process, (3) viral spread from injection site, (4) antibody penetration, (5) blending artifacts related to microscope misalignment, (6) injection site within target location, (7) successful registration, and (8) CNN overlay of detected cells with brain volume in signal channel. Because of the relatively high concentration of antibody solution needed for brain-wide immunohistochemical staining, non-specific fluorescence was apparent at the edges of tissue, i.e. outside of the brain and ventricles, in the form of punctate labeling not of cell origin. We computationally removed a border at the brain edge at the ventricles to remove false positives, at the cost of loss of some true positives (skimage.morphology.binary_erosion, Table 4). For neocortical layer studies, a subregion of the primary somatosensory area: “primary somatosensory area, unassigned” in PMA did not have layer-specific mapping in Allen Atlas space and was removed from consideration.
Injection site recovery and cell detection
Injection sites were identified in H129 studies by co-injecting CTB with virus (Supplementary Figure 27) and in c-Fos studies using ArchT-GFP expression. Post-registered light-sheet volumes of the injection channel were segmented to obtain voxel-by-voxel injection-site reconstructions. Volumes were Gaussian blurred (3 voxels). All voxels below 3 standard deviations above the mean were removed. The single largest connected component was considered the injection site (scipy.ndimage.label, SciPy 1.1.0113). CTB was selected for injection site labelling for transsynaptic tracing as it does not affect the spread of alpha-herpesviruses and its greater diffusion due to its smaller size overestimates the viral injection size by as much as two-fold114, 115. Although typically used as a tracer itself, used during the small window of ∼80 hours post-injection, it did not have time to spread. Supplementary Figure 7 shows the percentage of cerebellum covered by at least one injection in each of the three datasets. Lobules I-III, flocculus, and paraflocculus were not targeted.
Automated detection of transsynaptically labeled neurons
To optimize cell detection for scalability, whole-brain light-sheet volumes (typically 100-150 GB 16-bit volumes) were chunked into approximately 80 compressed 32-bit TIF volumes per brain, with an overlap of 192 x 192 x 20 voxels in xyz between each volume, and stored on a file server.
For deploying the custom-trained cell-detection neural network, the file server streamed the volumes to a GPU cluster for segmentation. Once the segmentation was completed, the soma labels were reconstructed across the entire brain volume from the segmented image on a CPU cluster by calculating the maximum between the overlapping segments of each volume. The reconstructed brain volumes after segmentation were stored as memory-mapped arrays on a file server. Coordinates of cell centers from the reconstructed volumes were obtained by thresholding, using the established threshold from training evaluation, and connected-component analysis. Additionally, measures of detected cell perimeter, sphericity, and number of voxels it spans in the z-dimension were calculated by connected-component analysis for further cell classification if needed. The final output consisted of a comma-separated values file that includes the xyz coordinates as well as measures of perimeter, sphericity, and number of voxels in the z-dimension for each detected cell in the brain volume.
Convolutional neural network training
Supervised learning using CNN is useful in complex classification tasks when a sufficient amount of training data is available. Annotated training volumes were generated by selecting volumes at least 200 x 200 x 50 pixels (XYZ) from full-sized cell channel volumes. To ensure training data were representative of the animal variability across the whole-brain, training volumes were selected from different anatomical regions in different brains with various amounts of labeling (Table 2 for dataset description). Annotations were recorded by marking cell centers using ImageJ109. To generate labeled volumes, Otsu’s thresholding method (skimage.filters.threshold_otsu, Scikit-Image116 0.13.1) was applied within windows (30 x 30 x 8 voxels, XYZ) around each center to label soma. Using annotated volumes, we trained a three-dimensional CNN with a U-Net architecture117, 118 (github.com/PrincetonUniversity/BrainPipe). A 192 x 192 x 20 CNN window size with 0.75 strides was selected. The training dataset was split into a 70% training, 20% validation, and 10% testing subset. Training occurred on a SLURM-based GPU cluster. During training, the CNN was presented with data from the training dataset, and after each iteration its performance was evaluated using the validation dataset. Loss values, which measure learning by the CNN, stabilized at 295,000 training iterations, at which point training was stopped and the CNN was evaluated for performance, as a risk in machine learning is overfitting, i.e. the possibility that the neural network will learn particular training examples rather than learning the category.
Evaluation of CNN
To determine CNN performance on H129 data, we calculated an F1 score (Ref. 119). First, we needed to compare CNN output with our ground truth annotations by quantifying true positives (TP), false negatives (FN), and false positives (FP). We defined human-annotation as ground truth, consistent with the machine learning field120. Our neural network architecture produced a voxel-wise 0 (background) to 1 (cell) probability output. To determine a threshold value for binarization of the continuous 0-1 CNN-output values, F1 scores as a function of thresholds between 0 and 1 were determined (FIgure 1f). Connected-component analysis (scipy.ndimage.label) grouped islands of nonzero voxels to identify each island as a putative cell. Pairwise Euclidean distances (scipy.spatial.distance.euclidean) were calculated between CNN-predicted cell centers and human-annotated ground truth centers. Bipartite matching serially paired closest predicted and ground truth centers, removing each from unpaired pools. Unmatched predicted or ground truth centers were considered FPs or FNs, respectively. Prediction-ground truth center pairs with a Euclidean distance greater than 30 voxels (∼49 µm) were likely inaccurate and not paired.
The F1 score was defined as the harmonic average of precision and recall. Precision is the number of correct positive results divided by the number of all positive results returned by the classifier, i.e. TP/ (TP+FP). Recall is the number of correct positive results divided by the number of all samples that should have been identified as positive, i.e. TP/ (TP+FN). The F1 score reaches its best value at 1 (perfect precision and recall) and worst at 0. Using a 20 voxel cutoff instead of 30 gave 0.849 and 0.875 for human-CNN and human-human F1 scores, respectively. To determine CNN performance metrics, the testing dataset, which the network had yet to be exposed to was finally run using the established threshold producing an F1 score of 0.864. To generate the precision-recall curve, precision and recall values were calculated between thresholds of 0.002 and 0.998 with a step size of 0.002. Values of precision and 1-recall were used to plot the curve. The area-under-curve of the precision-recall curve was calculated using the composite trapezoidal rule (numpy.trapz).
Statistical analysis of transsynaptic tracing data
For initial inspection of thalamic or neocortical neurons, each injected brain was sorted by cerebellar region with the greatest volume fraction of the injection (as in Ref. 2; this region was defined as the primary injection site. Injections from each “primary” region were then pooled and averaged per thalamic nucleus (Figure 3f).
Two primary methods of quantification were used, fraction of all labeled neurons in the thalamus or neocortex, and density within particular structures. Fraction of neurons is defined as the total labelled neuron count within a structure (e.g. VAL) divided by the main parent structure (e.g. thalamus). This number is useful as it gives relative target projection strength relative to other projection strengths within the parent structure. However, this does not provide information in non-relative terms. Density, defined as total labelled neurons divided by volume of the structure, takes into account the relative sizes for each structure, allowing for more absolute comparisons of recipient structures. In anterograde examples, density therefore provides information on the concentration of influence a cerebellar region may have on a target structure.
Generalized linear model analysis
Contribution of each cerebellar meta-lobule to viral spread in each neocortical or thalamic region was fitted to a generalized linear model (GLM) consisting of an inhomogeneous Poisson process as a function of seven targeted cerebellar regions (“meta-lobules”). The predictor variables were xj, where xj is defined as the fraction of the total injection to be found in the j-th meta-lobule, such that Σxj = 1. The outputs to be predicted were yk defined as the fraction of the total number of cells in the entire neocortex (or thalamus) to be found in the k-th region. For the resulting fit coefficients βjk, the change in ŷjk arising from a unit change in xj is eβjk− 1. In figures 4f, 6f, and 8f, the heatmap indicates a measure of confidence, defined as the coefficient (βjk) divided by the coefficient’s standard error.
To determine greater than chance significant weights, we compared significant weights computed from the t-stats of the coefficients with those observed in a shuffle-based null model in which predictors were shuffled uniformly at random (n = 1,000). We found that the true number of positive significant weights is significantly greater than that expected under the null model with a one-sided, non-parametric p < 0.05. In Figure 7, the neocortical region “Frontal pole, cerebral cortex” was excluded from generalized linear model analysis due to zero counts across all brains for the region.
IMMUNOFLUORESCENCE IMAGE AND ANALYSIS
AAV DCN Injections
To visualize YFP labeled fibers and vGluT2-positive terminals in the thalamus, 40 micron thick slices were stained for vGluT2 anti-guinea pig Cy5 (Millipore Bioscience Research reagent 1:2000 diluted in PBS containing 2% NHS and 0.4% Triton). Images were taken using a confocal LSM 700 microscope (Carl Zeiss). Terminals positive to VGluT2 staining were identified and morphologically studied using confocal images that were captured using the following excitation wavelengths: 488 nm (YFP) and 639 nm (Cy5). High resolution image stacks were acquired using a 63X1.4 NA oil objective with 1X digital zoom, a pinhole of 1 Airy unit and significant oversampling for deconvolution (voxel dimension is: 46 nm width x 46 nm length x 130 nm depth calculated according to Nyquist factor; 8-bit per channel; image plane 2048 x 2048 pixels). Signal-to-noise ratio was improved by 2 times line averaging. Image stacks were deconvolved using Huygens software (Scientific Volume Imaging). With a custom-written Fiji-scripts (ImageJ) we identified putative synaptic contacts, i.e. YFP-positive varicosities that colocalized with vGluT2-staining, following the same analysis pipeline as Ref. 59 (script available upon reasonable request). The color channels (YFP and Cy5) of the images were split to get separate stacks. The YFP and Cy5 channels were Gaussian blurred (sigma = 1) and selected by a manually set threshold. A binary open function was done on both images (iterations = 4, count = 2) and objects were removed if their size was <400 pixels (YFP). A small dilatation was done on the red image (iteration = 1, count = 1). With the image calculator an ‘and-operation’ was done using the binary red and green image. The values 255 (white) of the binary YFP image were set to 127. This image and the result of the AND-operation were combined by an OR-operation. The resulting image was measured with the 3D-object counter plugin for volumes and maximum intensities. Only objects containing pixels with an intensity of 255 (overlap) are taken in account for particle analysis. Estimation of synapse density (number of terminals/area μm3) was obtained for each image by dividing the number of terminals by the image area121. Regression between AAV and HSV-H129 density was performed using two-sided Pearson’s regression was performed (R, cor.test).
C-FOS MAPPING EXPERIMENT
c-Fos mapping after optogenetic perturbation
Neural activity has been shown to increase c-Fos, an immediate-early gene product122. Mapping of c-Fos expression used L7-Cre +/- (n=10) and -/- (n=8) mice (males, B6; 129-Tg (Pcp2-cre)2Mpin/J, 004146, The Jackson Laboratory, Bar Harbor, Maine, bred in-house, 56 days or older). L7-Cre mice express Cre recombinase exclusively in Purkinje neurons123. rAAV1-CAG-FLEX-ArchT-GFP (UNC Vector Core, deposited by Dr. Ed Boyden, 4×1012 vg/ml, AV5593B lot number, 500 nl/injection 250 µm deep perpendicular to tissue) was pressure injected into four locations in lobule VIa/b.
Unlike transsynaptic tracing, where each individual animal can be used to test connectivity of a different cerebellar region, this experimental paradigm requires targeting the same cerebellar region (Lobule VI) to achieve sufficient statistical power. To ensure adequate power in this experiment our sample sizes were at least double the size in the original studies developing this methodology22. We selected lobule VI as the target, given prior nonmotor findings associated with this lobule2.
After virus injection, a cover slip (round 3 mm, #1 thickness, Warner Instruments 64–0720) was used to cover the craniotomy and a custom titanium plate for head fixation124 was attached using dental cement (S396, Parkell, Brentwood, NY). Mice were allowed to recover after surgery for 4 weeks and then were habituated to a head-fixed treadmill124 for three days, 30 minutes per day. On the last day of habituation, ArchT-GFP expression was confirmed using wide-field fluorescence microscopy. The following day, mice were again placed on the treadmill and a 200 µm fiber (M200L02S-A, Thorlabs, Newton, NJ) was placed directly over the cranial window for optogenetic stimulation with 532 nm laser (1 Hz, 250 ms pulse-width, 56 mW, 1 hr, GR-532-00200-CWM-SD-05-LED-0, Opto Engine, Midvale, UT). We determined the appropriate stimulation power using test animals, prepared in the same manner as previously described, but also with electrophysiological recordings. We titrated our stimulus on these test animals (not included in manuscript cohort) to ensure it did not produce movements while producing reliable silencing of Purkinje cells during light stimulus, but without significant silencing after termination of the light-stimulus (Supplementary Figure 20). The experimental configuration delivered light from illumination from outside the brain, which was therefore attenuated through the air, coverslip, and brain tissue, leading to light scattering and heat dissipation. This made power requirements higher than other published studies18.
We selected the comparison between Cre +/- and Cre -/- animals to ensure our control animals (no Cre, no channelrhodopsin) received the same surgery, injection, coverslip placement, and head mount, and were placed on the same wheel and received the laser placement and activation. This controlled for differences in experimental stress. These animals were also cage mates (mixed Cre +/- and Cre -/- in each cage) since birth and provided natural blinding of condition to the experimenter.
Mice were then individually placed into a clean cage, kept in the dark for one hour, and perfused as described previously. Brains were fixed overnight in 10% formalin (4% formaldehyde) before beginning the iDISCO+ clearing protocol. Both ArchT-expressing mice and non-expressing mice received cranial windows, habituation, and photostimulation.
For behavioral quantification (Supplementary Figure 20), videos were imported into ImageJ using QuickTime for Java library, and images converted into grayscale. Timing of optogenetic stimulation was confirmed by analysis of pixel intensity over optical fiber connection to implanted cannulae. Forelimb kinematic data and treadmill speed were analyzed by Manual Tracking plugin. For arm movement stimulation movements with forearm moving forward (initial positive slope) from forearm moving backwards (initial negative slope).
Electrophysiological confirmation of ArchT expression in Purkinje cells
To confirm that ArchT was optically activatable in Purkinje cells, photostimulation was done during patch clamp recording in acutely prepared brain slices. Brain slices were prepared from three 10 week-old male Pcp2-cre mice (B6.Cg-Tg (Pcp2-cre)3555Jdhu/J, 010536, The Jackson Laboratory), two weeks after injection with rAAV1-CAG-FLEX-ArchT-GFP. Mice were deeply anesthetized with Euthasol (0.06 ml/30g), decapitated, and the brain removed. The isolated whole brains were immersed in ice-cold carbogenated NMDG ACSF solution (92 mM N-methyl D-glucamine, 2.5 mM KCl, 1.25 mM NaH2PO4, 30 mM NaHCO3, 20 mM HEPES, 25 mM glucose, 2 mM thiourea, 5 mM Na-ascorbate, 3 mM Na-pyruvate, 0.5 mM CaCl2, 10 mM MgSO4, and 12 mM N-acetyl-L-cysteine, pH adjusted to 7.3–7.4). Parasagittal cerebellar brain slices 300 μm) were cut using a vibratome (VT1200s, Leica Microsystems, Wetzlar, Germany), incubated in NMDG ACSF at 34°C for 15 minutes, and transferred into a holding solution of HEPES ACSF (92 mM NaCl, 2.5 mM KCl, 1.25 mM NaH2PO4, 30 mM NaHCO3, 20 mM HEPES, 25 mM glucose, 2 mM thiourea, 5 mM Na-ascorbate, 3 mM Na-pyruvate, 2 mM CaCl2, 2 mM MgSO4 and 12 mM N-acetyl-L-cysteine, bubbled at room temperature with 95% O2 and 5% CO2). During recordings, slices were perfused at a flow rate of 4–5 ml/min with a recording ACSF solution (120 mM NaCl, 3.5 mM KCl, 1.25 mM NaH2PO4, 26 mM NaHCO3, 1.3 mM MgCl2, 2 mM CaCl2 and 11 mM D-glucose) and continuously bubbled with 95% O2 and 5% CO2.
Whole-cell recordings were performed using a Multiclamp 700B (Molecular Devices, Sunnyvale, CA) using pipettes with a resistance of 3–5 MΩ filled with a potassium-based internal solution (120 mM potassium gluconate, 0.2 mM EGTA, 10 mM HEPES, 5 mM NaCl, 1 mM MgCl2, 2 mM Mg-ATP and 0.3 mM Na-GTP, pH adjusted to 7.2 with KOH). Purkinje neurons expressing YFP were selected for recordings. Photostimulation parameters used were 525 nm, 0.12 mW/mm², and 250 ms pulses at 1 Hz.
Light-sheet microscopy for c-Fos imaging
Opaque magnets (D1005A-10 Parylene, Supermagnetman, Pelham, AL) were glued to ventral brain surfaces in the horizontal orientation and imaged using a light-sheet microscope as described previously. Version 5.1.293 of the ImSpector Microscope controller software was used. ArchT-GFP injection volumes were acquired using the 561 nm excitation filter. Cellular imaging of c-Fos expressing cells was acquired using 640 nm excitation filter at 5.0 µm/pixel (1x magnification, 1.3x objective, 0.1 numerical aperture, 9.0 mm working distance, 12.0 x 12.0 mm field of view, LVMI-Fluor 1.3x, LaVision Biotech) with a 3 µm step-size using a 0.010 excitation NA. This resolution was selected to allow whole-brain imaging using ClearMap without tiling artifacts. To speed up acquisitions, the autofluorescence channel and injection channels were acquired separately with a shorter exposure time than the cell channel. The left and right horizontal focus was shifted towards the side of the emitting sheet. Left and right images were then sigmoidally blended before analysis. In order to maximize field of view, some olfactory areas were not completely represented in images and were removed from analysis. Five brains were reimaged a second time due to ventricular imaging artifacts.
Automated detection of c-Fos expressing cells
Detection of c-Fos expressing cells after optogenetic stimulation was done using ClearMap software for c-Fos detection22 modified to run on high performance computing clusters (“ClearMapCluster”, see Table 5 for analysis parameters). Cell detection parameters were optimized by two users iterating through a set of varying ClearMap detection parameters and selecting those that minimized false positives while labelling only c-Fos positive neurons with high signal-to-noise ratio.
Statistical analysis of c-Fos data
Cell and density heat maps and p-value maps were generated using ClearMap. Projected p-value maps were generated by binarizing the p-value maps and counting non-zero voxels in z; color bar thresholding displayed greater than 25% for coronal and 27% for sagittal sections of the z-distance. Injection sites were segmented and aligned in the manner described previously. Activation ratio was defined as the mean number of cells in an anatomical area across experimental brains divided by the mean number of cells in the same anatomical area in control brains. To compare the c-Fos activation data with transsynaptic tracing data across the major divisions in the neocortex, linear-least squares regression (scipy.stats.linregress, two-sided) were calculated using mean viral-labeling neocortical densities with H129-VC22 injections (80 hpi) were compared with the mean cell density ratio of c-Fos stimulation vs control groups.
SOFTWARE
Data analysis pipelines were run using custom code written for Python 3+ (available at github.com/PrincetonUniversity/BrainPipe and github.com/PrincetonUniversity/ClearMapCluster) Unless otherwise noted, analyses and plotting were performed in Python 2.7+. DataFrame manipulations were done using Numpy125 1.14.3 and Pandas126 0.23.0. Plotting was done with Matplotlib127 2.2.2 and Seaborn128 0.9.0. Image loading, manipulation and visualization was done using Scikit-Image116 0.13.1 and SimpleITK129 1.0.0. SciPy113 1.1.0 was used for statistical analyses. Clustering analysis was performed using Seaborn128 0.9.0 and Scikit-Learn130 0.19.1 was used for hierarchical agglomerative clustering (average metric, Ward’s method). Coefficients and standard errors for the generalized linear model were obtained by fitting the model using the statsmodels 0.9.0 package in Python 3.7.1 (as in Ref. 2). The Mann-Whitney U test (two-tailed; scipy.stats.mannwhitneyu, SciPy113 1.1.0) was used to determine statistical significance between control and experimental brain regions in c-Fos studies.
AUTHOR CONTRIBUTIONS
T.P., M.K., H.-J.B., and S.W. conceived and designed the experiments. T.P., D.B., and J.V. performed virus injections and prepared tissue. Z.D. and T.P. imaged tissue and ran the computational data analysis pipeline for whole-brain imaging data. T.P., Z.D., and H.-J. B. performed subsequent data analysis and prepared figures. E.E. designed and provided HSV vectors. K.U.V. and T.P. designed algorithms for image analysis. M.K., J.L., and T.P. performed optogenetics experiments. E.H. and B.R. performed AAVrg experiments, B.R. collected and analyzed images. H.-J. B., N. de O. and F.H. performed AAV experiments and collected and analyzed images. T.P. and S.W. wrote the initial draft of the manuscript, which was edited by all authors.
COMPETING INTERESTS
The authors declare that they have no competing interests.
DATA AVAILABILITY
The datasets generated and analyzed in the current study are available from the corresponding author upon request.
CODE AVAILABILITY
All experimental and analysis code is available at the links provided in the Methods section.
ACKNOWLEDGMENTS
We thank Aleksandra Badura for advice on experimental design, Lynn W. Enquist for discussion and for the generous gift of PRV-Bartha 152 (CNNV, P40 OD010996), James Gornet for neural network implementation assistance, Austin Hoag for software design and management, Nicolas Renier and Kelly Seagraves for tissue-clearing optimization, Stephan Thiberge for microscopy help, and Shruthi Deivasigamani, Joseph Gotto, Joyce C. Lee, Laura A. Lynch, Caroline Jung, Dafina Pacuku, and Thaddeus Weigel for technical assistance. This work was supported by NIH R01 NS045193, R01 MH115750, and U19 NS104648 (S.W.), F31 NS089303 (T.P.), P40 OD010996 (E.E.), R21 DC018365 (B.R.), and P20GM103408 (E.H. and B.R.), Netherlands Organization for Scientific Research - Veni ZonMW, 91618112 (H.-J.B), Erasmus MC Fellowship 106958 (H.-J.B), and the New Jersey Council on Brain Injury Research (J.V.) .
Footnotes
Manuscript text revised after reviews; Figures 3-8 updated; Supplementary figures added
REFERENCES
- 1.↵
- 2.↵
- 3.↵
- 4.↵
- 5.
- 6.↵
- 7.↵
- 8.↵
- 9.↵
- 10.↵
- 11.↵
- 12.↵
- 13.↵
- 14.
- 15.↵
- 16.↵
- 17.↵
- 18.↵
- 19.↵
- 20.↵
- 21.↵
- 22.↵
- 23.↵
- 24.↵
- 25.
- 26.↵
- 27.↵
- 28.↵
- 29.↵
- 30.↵
- 31.↵
- 32.↵
- 33.↵
- 34.↵
- 35.↵
- 36.↵
- 37.
- 38.↵
- 39.↵
- 40.↵
- 41.↵
- 42.↵
- 43.↵
- 44.↵
- 45.↵
- 46.↵
- 47.↵
- 48.↵
- 49.↵
- 50.↵
- 51.↵
- 52.↵
- 53.↵
- 54.↵
- 55.↵
- 56.↵
- 57.↵
- 58.↵
- 59.↵
- 60.↵
- 61.↵
- 62.↵
- 63.↵
- 64.↵
- 65.↵
- 66.↵
- 67.↵
- 68.↵
- 69.↵
- 70.↵
- 71.↵
- 72.↵
- 73.↵
- 74.↵
- 75.↵
- 76.↵
- 77.↵
- 78.↵
- 79.↵
- 80.↵
- 81.↵
- 82.↵
- 83.↵
- 84.↵
- 85.↵
- 86.↵
- 87.↵
- 88.↵
- 89.↵
- 90.↵
- 91.↵
- 92.↵
- 93.
- 94.↵
- 95.↵
- 96.
- 97.↵
- 98.↵
- 99.↵
- 100.↵
- 101.↵
- 102.↵
- 103.↵
- 104.↵
- 105.↵
- 106.
- 107.↵
- 108.↵
- 109.↵
- 110.↵
- 111.↵
- 112.↵
- 113.↵
- 114.↵
- 115.↵
- 116.↵
- 117.↵
- 118.↵
- 119.
- 120.↵
- 121.↵
- 122.↵
- 123.↵
- 124.↵
- 125.↵
- 126.↵
- 127.↵
- 128.↵
- 129.↵
- 130.↵
- 131.
- 132.
- 133.
- 134.
- 135.
- 136.
- 137.
- 138.
- 139.
- 140.
- 141.
- 142.
- 143.
- 144.
- 145.
- 146.
- 147.
- 148.
- 149.