Quantitative organization of the excitatory synapses of the primate cerebellar nuclei: further evidence for a specialized architecture underlying the primate cerebellum

All animal experiments were carried out in accordance with the Society for Neuroscience and local German authorities (approved by the regional authorities (Regierungspräsidium Tübingen). Brain tissue from two male adult macaques was used in this study (D98: body weight 17 kg and 18 years old, H01: body weight 9 kg and 13 years old). Following premedication with a mixture of ketamine (20 mg/100 g body weight), xylazine (2 mg/100 g body weight) deep anesthesia was induced with pentobarbital (400 mg/kg). The animals were then perfused transcardially by 0.1 M PB and then by 4% ice-cold paraformaldehyde in 0.1 M phosphate buffer (PB) at pH 7.4. The brain was immediately dissected out of the skull and then cryoprotected in an ascending concentration of sucrose (10%, 20%, and 30% in 0.1 M PB). The cerebellum was removed from the brainstem and mounted on a microtome with the lateral side facing the freezing platform. Sections of 40 µm (D98) or 50 µm (H01) were serially acquired and stored in 0.1 M PB. Immunofluorescence staining was carried out on free floating sections. Prior to primary antibody incubation, sections were washed three times in 0.1 M PB for 5 min each time and then blocked in 0.1 M PB with 10% horse serum (PAA Laboratories, Coelbe, Germany) and 0.3% Triton X-100 at room temperature for 1 h.

The concentration and incubation times were tested for each antibody in single stains. The optimized conditions were then used for the multiple stains. We performed a quadruple staining protocol for vGluT1, vGluT2, Purkinje Cell Protein 2(PCP2), and potassium channel KCNC3/Kv3.3 on H01, and a further quadruple staining protocol for vGluT1, vGluT2, MAP2 and KCNC3/kv3.3 on D98. Data of the Kv3.3 were used for a subsequent publication. The sections were first incubated with goat vGluT1 (sc-13320, Santa Cruz Biotechnology, Texas) at 1:1000 and 4 °C for 16 h, followed by 3 × washes in 0.1 M PB and then incubated with donkey anti goat Alexa Fluor 488 at 1:500 for 2 h at room temperature. The sections were then washed three times in PB before incubation with a combination of three antibodies of vGluT2 (1:1000, lot number: 135404, Synaptic systems, North Saanich British Columbia), Kv3.3 (1:500, lot number: APC-102, Alomone labs, Israel) and PCP2 (1:200, lot number: sc-137064, Santa Cruz Biotechnology, Texas) or MAP2 (1:200, lot number: M1406, Sigma Aldrich) at 4 °C for 16 h. The sections were then washed and incubated with the combinations of the three secondary antibodies: goat anti-guinea pig Alexa Fluor 633 (1:500, A21105, Invitrogen, California), goat anti-rabbit Cy3 (1:500, lot number: 81-6115, Invitrogen, California) and goat anti-mouse Alexa Fluor 405 (1:500, lot number: A31553, Invitrogen, California) for 2 h at room temperature. Sections were then washed and mounted in glycerol on glass slides with Mowiol 4–88 (Merck, Darmstadt, Germany). The slides were stored at 4 °C.

To test for vGluT2 antigen–antibody specificity, we mixed 1 μl vGluT2 antibody with 2 μl of the peptide that was used for immunization and was provided by the supplier. This consisted of Strep-Tag fusion protein (amino acids 510–582 of the rat vGluT2) in 500 μl 0.5 M PB solution with 2% horse serum and 0.1% tritonX-100. The mixture was rotated at room temperature for 1 h to allow for full interaction of the peptide with the antibody. As a control solution we used 1 μl vGluT2 antibody diluted in 500 μl 0.5 M PB solution with 2% horse serum and 0.1% tritonX-100. Two adjacent cerebellar sections from the same series of monkey H01 were washed and blocked. One section was incubated in the antigen–antibody mixture and the other in the control solution for 48 h at 4° with constant rotation. The two sections were washed and incubated with goat anti-guinea pig Alexa Flour 633 1:1000 for 2 h at room temperature. The sections were washed, mounted, and scanned with laser confocal microscope LSM 510. VGluT2 channel was scanned under the excitation of 633 nm and emission at 650–750 nm, while lipofuscin channel was scanned under the excitation of 405 nm and emission at 650–750 nm.

Images were acquired on a laser scanning confocal microscope (LSM 510, Carl Zeiss, Jena, Germany). We took overview images of vGluT1 staining for every section with 488 nm excitation and emission band pass filtering 505–550 nm at low magnification (10 × objective). Probe positions were determined by first marking an identifiable origin (x, y coordinates: 0, 0) and then determining the position of the other probes in the DCN region at a regular spacing of 1000 μm × 1000 μm. The location of the origin point was chosen from an easily identifiable structure (i.e., vessels) within a core region of the DCN. However, this location differs from slice to slice and is not related to the structures to be analyzed. A z-stack was acquired using a 63 × (NA1.4, Oil immersion) objective, applying a 2 × zoom for each probe. The pinhole was set to one airy unit and subsequently optimized for every detection channel to achieve an equal optical slice thickness on all channels. XY voxel size was set to 0.14 μm × 0.14 μm; image digital resolution was set to 512 × 512 pixels. The z-stack step size was 0.32 μm and we took an average of 30 optical sections. We stained and analyzed 13 cerebellar sections (5 slices from D98 and 8 slices from H01) and 168 probes in total were sampled (90 from D98, 78 from H01). The fluorescence emitted by lipofuscin granules were recorded with excitation wavelength at 405 nm and emission wavelength at 650–750 nm. Details of antibodies, excitation wavelength and emission filters are listed in Table 1.

We also determined tissue shrinkage in section thickness by taking four random positions within DCN region for each slice and determining the upper and lower section borders by the staining signal for vGluT1 at a magnification of 20x (NA0.8, dry). Fifty-two probes were taken from 13 sections. The mean thickness of the slices after mounting was 24.6 μm for D98 and 26 μm for H01, while the original sectioning thickness was 40 μm and 50 μm, respectively. Classification of the macaca DCN was based on the description in an earlier paper (Hamodeh et al. 2014).

Lipofuscin has a broader excitation and emission spectrum than commercial fluorophores which are designed with narrow spectra. We acquired the signals from our own fluorophores by regular microscopic settings while simultaneously recording lipofuscin fluorescence signal in an additional channel and subtracted the lipofuscin fluorescence during post-processing.

We performed surface reconstruction to get boundaries of lipofuscin granules using the Imaris software (Bitplane). The lipofuscin surfaces were smoothed at 0.279 μm and the background was subtracted using the Ridler and Calvard (R–C) algorithm (Ridler and Calvard 1978) to determine the best thresholding level for this channel. The thresholding level was about 35/255. The reconstructed lipofuscin surfaces were subsequently used to mask the vGluT1 and vGluT2 channel. The voxels inside the lipofuscin mask were set to 0, while the voxels outside the lipofuscin mask remained unchanged. The background of the masked vGluT1 channel was subtracted. The surface bounding of vGluT1 boutons was smoothed at 0.15 μm and the diameter of the largest sphere fitting into the objects was taken at 0.5 μm. The masked vGluT2 channel was similarly surface-rendered, except that we fitted a Gaussian filter (SD = 0.3 μm) beforehand and used a user-defined fixed threshold of 68/255. The threshold for masked vGluT1 channel was determined by the Ridler and Calvard (R–C) algorithm and the resulting mean threshold was 20.1/255 for D98 and 26.2/255 for H01. We used Gaussian filter and user-defined threshold for the masked vGluT2 channel to effectively exclude the background and DCN neuron somata (Mao et al. 2018). The surface-reconstructed boutons were further filtered according to size by excluding surfaces with less than ten voxels (Haass-Koffler et al. 2012).

The number of surface reconstructed boutons for each probe was exported and normalized to the corrected volume of probe in question. First, we obtained the (uncorrected) section thickness by adding the average diameter of the contours, 1.12 μm for vGluT1 and 1.19 μm for vGluT2 to account for boutons being only partly contained within a section (Braitenberg and Schüz 1998). The uncorrected probe volume was then obtained by multiplying the probe length (71.4 μm) and the probe width (71.4 μm) and the corrected depth (varying from probe to probe). Tissue shrinkage affected the thickness of the section and was, therefore, taken into consideration. To determine the extent of the shrinkage, four random positions were chosen within the DCN region for each stain section. All the values were then averaged and the cerebellar sections from D98 were shrunk to 24.6 μm thickness, while those of H01 were shrunk to a thickness of 26 μm, the original sectioning thicknesses being 40 μm for D98 and 50 μm for H01. Finally, we divided the uncorrected probe volume with 24.6/40 for D98 and 26/50 for H01 to obtain the corrected probe volumes.

The quantification of co-labelling of vGluT1 and 2 was performed by obtaining the intersection of the surface boundaries of the 2 vGluT channels (vGluT1∩2). However, we could only perform channel subtractions in Imaris. We, therefore, subtracted vGluT1 from the vGluT1-only channel. The vGluT1-only channel was obtained for each probe by subtracting the vGluT2 channel from the original vGluT1 channel for the rats and from the lipofuscin cleaned vGluT1 channel in the case of the monkey. The number of the co-labelled boutons was divided by the vGluT1 bouton number or the vGluT2 bouton number to ascertain the co-labelled percentage for vGluT1 boutons or vGluT2 boutons. The percentages were compared between different nuclei and between the two species.

The vGluT1 and 2 density and volume values were checked for normal distribution and variance homogeneity. Quantile–quantile plots (qqnorm in R, http://​www.​r-project.​org) and Shapiro-tests (package lme4 in R) indicated normal distribution for vGluT1 and vGluT2 volume. Levene’s tests (package Rcmdr in R) indicated equal variance for vGluT1 and vGluT2 volume. The vGluT1 and vGluT2 density values were power transformed by transformTukey (package rcompanion in R) to secure normal distribution and equal variance. Since multiple data points were obtained in each subject (Aarts et al. 2014), a linear mixed effect model was used to accommodate the dependence of the data. The linear mixed effects model via restricted maximum likelihood estimations were performed using lme4 package in R. The response was the transformed density or non-transformed volume (for both vGluT1 and 2). The nuclei variance was modelled as a fixed effect, while subject variance was modelled as a random effect in the model. We also constructed a null model with only the random effect. We then compared the two models using a likelihood ratio test to ascertain whether these two models differed significantly from each other. Post hoc multiple comparisons (Tukey contrast) with Holm–Bonferroni correction were performed. The significance level α was set to 0.05. We also plotted the data with intensity color-coded scatter plots using matlab scripts (Eilers and Goeman 2004).

The vGluT2 immunofluorescence showed a puncta pattern organized in a linear beaded fashion in the molecular layer (ML) and glomeruli-like pattern in the granule cells layer (GCL) of the cerebellar cortex (small arrows in Fig. S1a). There are also granular structures in the Purkinje cells layer (PCL) as described for lipofuscin in the literature [big arrow in Fig. S1a, (Brizzee et al. 1974)]. Not surprisingly, these structures are obtained again under our setting which is used for lipofuscin fluorescence (big arrow, Fig. S1b). The staining pattern of vGluT2 after lipofuscin channel subtraction is consistent with our previous study in the rat using the same antibody (Fig. S1c) (Mao et al. 2018). Preabsorption of vGluT2 antibody with its corresponding peptide abolished the vGluT2 signals in the cortex, pointing to the specificity of the vGluT2 antibody (Fig. S1d, e, f). Similarly, the vGluT2 antibody is also specific in the DCN region (Fig. S1 g–l). We also controlled the multiple staining patterns in the cerebellar cortex (molecular layer) and DCN of the monkey. As anticipated, parallel fibers were stained by vGluT1 only, whereas climbing fiber presynapses were stained by vGluT2 only (Fig. 1a, b). The vGluT1 and vGluT2 were separated from each other in the molecular layer, as the negative Pearson’s coefficient indicates (Fig. 1c). We also compared the staining of vGluT1 and 2 with the PCP2 staining in the DCN and found them to be segregated as in the molecular layer (Fig. 1d, e). However, unlike in the cortex, the vGluT1 and vGluT2 show some colocalization (Pearson’s coefficient 0.47). These co-labelled boutons may come from mossy fibers that express both vGluT1 and vGluT2 (Boulland et al. 2004). A comparison of labelling intensity within a voxel confirmed the different patterns; in contrast to the DCN there was no above-threshold intensity co-labelling of vGluT1 and 2 in the molecular layer (Fig. 1c, f). The spatial relationship of the vGluT1 and 2 labelled puncta with their postsynaptic partner, the dendrites from DCN neurons, was also examined in a quadruple staining environment. Most of the vGluT1 and 2 labelled puncta can be observed around the dendrites (Fig. 1g, h). Again, the colocalization of vGluT1 and vGluT2 in the DCN can be seen in this example (Fig. 1i).

To obtain the quantitative data of vGluT1+ and vGluT2+ bouton density and compare between nuclei, the lipofuscin signal was first detected and eliminated from the vGluT1 and vGluT2 channels. First, we verified the lipofuscin excitation and emission characteristics in unstained macaque cerebellar sections (see supplementary material and Fig. S2). Then we validated our lipofuscin masking approach in the vGluT1 and vGluT2 channels and this is summarized in the supplementary material (see also Fig S2–4).

The structures that are labelled by vGluT1, vGluT2 or both are shown in different nuclei (Fig. 2). Double-labelled structures can be found in all the different nuclei. However, the vGluT1+ boutons are sparser in MN and AIN than in LN and PIN. Nevertheless, the vGluT2+ boutons are more similar between nuclei (Fig. 2a, c, e, g). The surface analysis allowed for a detailed visualization of vGluT1 and vGluT2-labelled structures and quantification of the number and volume of the boutons as well as the co-labelled boutons (Fig. 2b, d, f, h).

We quantified the vGluTs bouton density in the different DCN. The vGluT1+ bouton density is higher in the PIN and LN, with 1.97 × 10⁶ (SEM = 1.78 × 10⁵) and 1.84⁶/mm³ (SEM = 4.96 × 10⁵), and lower in AIN and MN, 1.21 × 10⁶ (SEM = 3.04 × 10⁵) and 1.12 × 10⁶/mm³ (SEM = 3.3 × 10⁵) (ANOVA of mixed effect model, df: 3, Chi-square: 29.38, p 1.87 × 10⁻⁶). The vGluT2+ bouton density differed less intensely than the vGluT1 between subnuclei, although there was still a significant difference (ANOVA of mixed effect model, df: 3, Chi-square: 7.85, p 0.049), mainly due to the lower density of vGluT2 in AIN. The mean density for vGluT2 was 1.7⁶/mm³ (SEM = 8.29 × 10⁴) for the total DCN. The vGluT1 and 2+ bouton density data were power transformed to secure normal distribution and equal variance to satisfy the criterion for the statistic tests (Fig. S5a, c). The density and the volume of vGluT1+ and vGluT2+ boutons are shown for the four nuclei (Fig. 3a, b). The co-labelling of vGluT1 and vGluT2 is comparable in the AIN, PIN and LN, around 30%, with slightly lower values in MN (Fig. 3c). However, both the rat vGluT1 and vGluT2 co-labelling, which were obtained from our previous study, are significantly lower in the rat DCN than in the macaque DCN (Fig. 3d). The mean vGluT1+ and 2+ bouton volume is around 0.7 μm³ and 0.9 μm³, respectively (Fig. 3b). The vGluT1 volume was slightly larger in the AIN (0.78 μm³) and LN (0.8 μm³) than in the other DCN (MN: 0.61 and PIN: 0.70 μm³) and the vGluT2 volumes was larger in the LN (0.97 μm³) than in the other DCN (PIN: 0.84, AIN: 0.89, MN: 0.86 μm³), albeit with a larger variability. A one-way ANOVA was used to test the variance by different nuclei classification and the results are summarized in Table 2.

We also scrutinized our reconstructed vGluT1 and 2 boutons for any cluster outliers, such as axonal regions where the vGluTs are detected on their way to the presynaptic terminals (Mao et al. 2018). These elongated fragments would have a large ellipticity prolate value (close to 1) and a small diameter. Figure 4 displays density color-coded scatter plots of vGluT1 and 2 presynapse diameter vs. the shape parameter ellipticity prolate for the different DCN. Data show one large cluster with diameters larger than 0.5 μm and ellipticity prolate larger than 0.2 and one small cluster with ellipticity prolate smaller than 0.2. Based on these values they cannot be considered as axon fragments, because of their round shape. The number of boutons in the small cluster in the MN, AIN, PIN and LN amounted to 2.9%, 3.8%, 4.1%, 3.8% for vGluT1 and 4.2%, 3.8%, 3.6%, 4.2% for vGluT2, respectively. These boutons may come from a distinct population, e.g., the intrinsic DCN glutamatergic synapses.

We calculated the mean bouton number per DCN neuron on the basis of our vGluT1 and vGluT2 labelled boutons in the current study and the neuron density in our previous study (Hamodeh et al. 2014, 2017). The vGluT1 and 2 bouton number per neuron is similar in rat and monkey. However, the vGluT1 bouton number per neuron is significantly lower in the macaque LN/dentate (Fig. 5a). The total excitatory synaptic number per neuron was calculated by subtracting the co-labelled vGluT1 bouton number per neuron from the total vGluT1 bouton number per neuron and adding the total vGluT2 bouton number per neuron (Fig. 5c, d). The total bouton number per neuron in macaque DCN is lower than in the rat, except for the MN. Within the macaque, PIN has the largest bouton number per neuron (Fig. 5c, d).

One explanation of the comparable values of excitatory synapses per neuron could be due to the modular organization of the cerebellum with similar number of units/neurons per module. However, other, potentially confounding factors may have led to a reduction of vGluT immunostaining in the monkey. The antibodies we used were optimized to stain the rodent’s vGluT c-terminal (unspecified peptide length for vGluT1 and amino acids 510–582 for vGluT2). A comparison of the amino acid sequence of the two transporters between rat and macaca mulatta shows that vGluT1 has an essentially conserved sequence, whereas the vGluT2 shows multiple amino acid changes in the last 72 amino acids. It is, therefore, possible that the vGluT2 immunostaining in the monkey tissue could be altered, which would also explain the higher threshold used for vGluT2 compared to the rat. However, a comparison of the staining pattern within the molecular layer did not show colocalization at above-threshold intensities for the vGluT2+ only climbing fibers and the vGluT1+ only parallel fibers. Our calculation of the individual excitatory inputs per neuron was based on different individuals and this may have introduced additional variability in our results. We used systematic unbiased sampling in both cases and tried to minimize the influence of such bias. Finally, our two monkeys, aged 13 and 18 years, may have already shown signs of age-related synapse elimination. As anticipated, our quantification of the lipofuscin did show a higher lipofuscin density in the older monkey (Gilissen et al. 2001). However, we did not observe systematically lower vGluT densities in the older monkey (D98 vs. H01:1.5 × 10⁶ vs. 2.3 × 10⁶/mm³ for vGluT1 and 1.8 × 10⁶ vs. 1.2 × 10⁶/mm³ for vGluT2). In summary, our data could point to the importance of a rather constant number of excitatory synapses per DCN neuron, reflecting a limited number of input units per DCN neuron.

One unexpected finding of our study is that the monkey DCN contain a higher percentage of vGluT1 and 2 co-labelled presynapses than those of the rat DCN (30 vs. 15%). This is probably due to the fact that the density of vGluT2 in the monkey is higher than in the rat. The cerebellum is one of the few brain regions that shows the presence of both kinds of vGluTs (in climbing and mossy fibers), as well as presynapses with both kinds of vGluTs in a subset of mossy fibers. Developmental studies have shown that, during the early stages of development, vGluT2 synapses change to vGluT1 following the switch of vGluT2 mRNA expression to vGluT1 mRNA expression in the neuron somata (Miyazaki et al. 2003). This suggests that a change occurs at the transcription and translation level of protein synthesis. This could thus imply that the mossy fibers retain a status comparable to that of the early stages of development, allowing for prolonged modification from vGluT2 to vGluT1. A different scenario emerges if we look at a recent learning/plasticity study in the DCN. The study by Boele and colleagues showed an increase in basilar pontine nuclei vGluT2 mossy fiber presynapses in the DCN following Pavlovian eyeblink conditioning (Boele et al. 2013). So far, it remains unclear whether this kind of learning in the DCN requires the sprouting and formation of new vGluT2 synapses, or whether it is achieved by an increase in vGluT2 levels in vGluT1 mossy fiber collaterals. The latter has not been described so far and, therefore, currently appears to be the less likely scenario.

A further new finding of our study is that there is a lower number of vGluT1+ excitatory synapses per neuron in the LN/dentate of the rhesus monkey compared to the rat’s LN/dentate. This result tallies well with the lower dendritic length per neuron that we already reported in the monkey LN/dentate (Hamodeh et al. 2017) and lends further support to our model of hyposcalled dendrites that underlie the special architecture of the LN/dentate. The hyposcaled dendritic arbors (Hamodeh et al. 2017) and vGluT1+ inputs (this study) could allow for an increase in the number of independent/segregated parasagittal strips in the cerebellar hemispheres connected to the LN/dentate, which would confirm a recent prediction (Jorntell 2017). A theory proposed by van Essen (1997) on how and why structures in the central nervous system tend to fold was based on the amount of mechanical tension that arises from the connectivity pattern of a given brain region (Van Essen 1997). If we assume that the hyposcalled dendrites with smaller region of influence and with less excitatory synaptic connections would lead to less tightly connected tissue, then these less tightly connected regions would lead to a reduction of tension, which in turn would lead to some indentation and folding at these sites during the ontogenetic and phylogenetic volume increase of the nucleus. These changes could then first lead to the emergence of a torus-like or cup-shaped nucleus, as indeed the rhesus monkey LN/dentate exhibits (Sultan et al. 2010), whereas, in the nucleus with more similar connectivity tension would be basically the same all across the nucleus and would help in preserving a globose structure.