Anatomical clustering of synapses on developing dendrites

_________________________________

Anatomical clustering of synapses on

developing dendrites

Report internship 1 26/08/2017

A.M.Heuvelmans 11093714

Brain and Cognitive Sciences University of Amsterdam

Alexandra Leighton & Nawal Zabouri Co-assessor: Harm Krugers

Synapse & Network Development – Christian Lohmann Netherlands Institute for Neuroscience

2 1. ABSTRACT

Recent research has proven that dendrites, in contrast to what were assumed to be passive structures receiving input, have active compartments. These compartments sum inputs in a supralinear fashion, thereby increasing the influence on somatic firing. These active events enlarge the computational power of pyramidal neurons. The active dendritic properties demand a spatiotemporal organization of inputs onto dendritic compartments. Recent studies have shown that synapses in close proximity to each other are more likely to be co-active (Winnubst et.al. 2015). This occurred already during development and the functionally spatiotemporal organization of synaptic inputs was shown to be depending on spontaneous activity and N-aspartate-D-methyl receptor signalling (Kleindienst et.al. 2011; Takahashi et.al. 2012). In this study, clustering of synaptic inputs onto layer 2/3 pyramidal neurons during development was shown to be present. The composition of these anatomical clusters has not been investigated previously. Broadly, two types of synaptic inputs onto layer 2/3 pyramidal neurons in the visual cortex can be distinguished: thalamocortical projections and corticocortical projections. Both thalamocortical- and corticocortical projections were shown to be clustered on developing dendrites in single identity clusters. Additionally mixed clustering was observed, where both projections were spatially clustered. Data was collected from cells that were functionally imaged in vivo which grants an opportunity of directly relating functional and anatomical clustering. Anatomical clustering increases computational power of dendrites. Understanding the computational properties of dendrites is important to increase our understanding of neural network communication.

3 2. INTRODUCTION

Without any prior experience of external stimuli, humans are able to receive, process and react to these external stimuli from the moment they are born. This capability requires the development of complex neural networks already before birth. An important and interesting question is how the brain is structuring and preparing networks during development to be able to execute its complex cortical functions.

During development, before sensory systems rely on sensory inputs, spontaneous activity shapes neuronal circuits (Ben-Ari et.al., 1989; Galli & Maffei, 1988; Shatz & Stryker, 1988). These bursts of intrinsically generated, synchronized activity activate a large group of neurons. This activity spreads in a wave like fashion to neighbouring neurons. The synchronous activity events are involved in the formation of synaptic connections, thereby organizing functional neural networks and refining retinotopic mapping (Katz & Shatz, 1996). With maturation, and the use of sensory organs, the brain starts to rely on the sensory inputs (Katz & Shatz, 1996). Visual input is processed via a specific route. Input onto the retina is projected to the primary visual cortex (V1) via the lateral geniculate nucleus of the thalamus. Interestingly, the retinotopic organization of the visual input in the retina is also present in the thalamus and V1. The development of this topographic representation of the visual input has been found to depend on spontaneous neuronal activity (Cang et. al. 2005). It was shown that genetical disruption of the waves of spontaneous activity during the first postnatal week in mice disturbed retinotopic mapping (Cang et.al. 2005). Because mice are born with their eyes closed and eye opening only occurs at the end of the second postnatal week, their visual system can serve as a model to study development of neuronal networks before sensory input is received.

Within the complex neuronal networks of the brain, communication is an essential basic feature of neurons. Networks consist of excitatory and inhibitory neurons that are interconnected with each other. The pyramidal neuron is an excitatory neuron that is mainly found in structures that are associated with advanced cognitive functions (Spruston, 2008). Pyramidal neurons can be characterized by their typical pyramidally shaped soma. They have a single axon and distinct long, apical, and short and highly branched basal dendritic trees. The dendrites of these neurons are the main structures receiving excitatory inputs both from cortical and thalamic areas (Spruston, 2008). These inputs are integrated and send towards the soma, the main processing body of the cell. To profoundly understand the function of a pyramidal neuron, it is critical to know how the inputs at the dendrite are integrated to produce an output in the form of an action potential. For a long time, it has been thought that integration of dendritic inputs occurred through summation in a linear fashion, regardless of their location in the dendritic tree. If the sum of these inputs would reach a certain threshold when arriving at the soma, an action potential will be generated that is transmitted via the axon. This model is known as the integrate-and-fire model (Abbott, 1999). In this model, individual synapses would have a low influence on action potential initiation due to constraints of the dendritic tree, especially synapses more distant from the soma. (Stuart & Spruston, 1998; Spruston, 2008, Larkum & Nevian, 2008). More recent studies, however, revealed that, instead of being a passive part of the neuron, dendrites have local, active compartments that process excitatory input nonlinearly, producing dendritic spikes (Na+_{, Ca}2+ _{and NMDA-spikes). (Schiller et.al. 1997; Branco & Häusser, 2010;} Häusser et.al. 2000; Rhodes, 2006; Nevian et.al. 2007, Winnubst & Lohmann, 2012). N-methyl-D-aspartate (NMDA) spikes are regenerative spike events that are the result of the activation of NMDA receptor channels along a 10-20 µm stretch of dendrite (Losonczy & Magee, 2006; Nevian et.al. 2007).

4 The potential of the NMDA spike is much larger than the linear summation of the involved synaptic signals and is much more effectively transmitted towards the soma. Although one spike might not be sufficient to trigger the somatic firing, the supralinear summation of synaptic inputs does increase the synaptic influence on the generation of an action potential, and thereby greatly enhances the computational capacity of a neuron (Larkum & Nevian, 2008; Piorazi & Mel, 2001; Häusser & Mel, 2003). Recent research even proposed local plasticity to occur within the dendritic compartments influencing local dendritic excitability (Losonczy et.al. 2008; Magee & Johnston, 2006; Frick et al. 2004). For local dendritic spikes to be generated, synapses should be located spatially close to each other, instead of synapses being randomly distributed along the dendrite. This suggests a structural dendritic organization with a high spatiotemporal correlation between synapses. Besides non-random synaptic density, synaptic strength was shown to be non-random, where a few synapses in close proximity were stronger than surrounding synapses (Song et.al. 2005; Katz et.al. 2009). In addition to the spatial proximity, neighbouring synapses should be synchronously activated for dendritic spike generation. Dendrites should thus have spatiotemporal synaptic clusters. The first evidence for synaptic clustering in vivo was found by McBride and colleagues, who showed experience dependent clustering of co-active synaptic inputs in the barn-owl auditory system (McBride et.al. 2008). Additionally, in vivo and ex vivo functional studies performed in hippocampus, barrel cortex and visual cortex, during adulthood and/or development supported the finding that synaptic inputs onto dendrites are arranged in clusters (Kleindienst et.al. 2011; Takahashi et.al. 2012; Winnubst et.al. 2015). These studies have shown that, synapses in close proximity of each other were more likely to be coactive than more distant synapses. Importantly, spontaneous action potential firing and NMDA-receptor signalling at the synapse were shown to be essential for functional synaptic clustering along the dendrite as with the abolishment of either of these two, neurons failed to establish synaptic clustering (Kleindienst et.al. 2011; Takahashi et.al. 2012). Two types of spontaneous activity can be discriminated; high participation events (H-events) and low participation events (L-(H-events) (Siegel et.al. 2012). H-events are highly synchronous events spreading through gap junctions, being independent of retinal input in which virtually all cells (> 80%) participate. They are thought to play a role in homeostatic downregulation of synaptic weights (Siegel et.al. 2012; Turrigiano & Nelson 2004). The L-events on the other hand do depend on retinal wave transmission and are thought to be ‘test runs’, crucial for shaping the spatiotemporal characteristics of neuronal connections. It is thus these events that are thought to be involved in the arrangement of synapses in clusters along the dendrite (Siegel et.al. 2012).

The establishment of clusters is thought to depend on a local ‘out of sync – lose your link’ plasticity mechanism, in which transmission reliability and frequency of locally desynchronized synapses decreases (Winnubst et.al. 2015). This low transmission reliability was shown to be predictive for synaptic pruning. Spontaneous activity might therefore establish synaptic clustering by removing synapses that are depressed due to dyssynchronous firing with their neighbouring synapses and stabilizing the synapses that are highly coactive (Wiegert and Oertner, 2013; Winnubst et.al. 2015). Functional synaptic clusters, receiving correlated information, are thus thought to be established during development depending on spontaneous activity and NMDA-receptor signalling, being further modified during learning (Kastellakis et.al. 2015; Fu et.al. 2012; Winnubst & Lohmann, 2012). Synaptic pruning being dependent on synaptic co-activity proposes that synaptic clustering should also be present on the anatomical level, which was also shown in hippocampus, motor cortex, somatosensory cortex and auditory system in several studies (Druckmann et.al. 2014; Fu et.al. 2012; McBride et.al.

5 2008; Kastellakis et.al. 2015, Schoonover et.al. 2014). Anatomical clustering has however not yet been shown in the visual cortex.

Although there is evidence for both functional and anatomical clustering of excitatory synapses during development, studies on the direct relationship between the two are very limited. Moreover, there is no knowledge yet on the origin of the excitatory synaptic projections being clustered along dendrites. Layer 2/3 pyramidal neurons in the mouse visual cortex receive excitatory inputs from different regions. These can be separated into two main sources; thalamocortical (TC) and corticocortical (CC) projections. The information processing capacity of clusters will depend on their architecture. As there are two main input sources for excitatory synapses, different cluster arrangements are possible. Clusters can be composed of a single identity, having either TC or CC synapses, thus receiving similar information which increases its probability to elicit action potential firing. This, single identity type of clustering has been shown in the layer 4 and 5 neurons in the mouse somatosensory cortex, where clusters of TC synapses were identified (Rah et.al. 2013; Schoonover et.al. 2014). However, there is also the possibility of finding more complex clusters that encompass both TC and CC projections. Finding these ‘mixed’ clusters would suggest an intermediate step of signal integration, combining input from different origins before signal transmission to the soma, thereby increasing its information integrating capacity. The higher computational power of dendrites with excitatory synapses being clustered, increasing the chance of action potential firing, suggests a possibly strategic placement of inhibitory synapses with respect to the excitatory synaptic clusters.

The aim of this study is to investigate if synaptic clusters could be identified on dendrites of layer 2/3 pyramidal neurons in the visual cortex of mice and how these clusters would be arranged (CC, TC or mixed identity clusters). In a large subset of the analysed dendrites, anatomical clustering was observed. CC, TC and mixed clusters were identified. Interestingly, dendrites with single identity clusters were shown to be clustered for either CC or TC projections but not for both. The ability to post hoc analyse an in vivo functionally recorded neuron allows new research to correlate functional to anatomical data, thereby granting increased understanding of synaptic connectivity and its refinement. Single and mixed identity clusters were found to be present on dendrites of layer 2/3 pyramidal neurons in the developing mouse visual system.

3. METHODS

3.1 Functional and anatomical animals

For the acquisition of functional data, E16 C57BL/6 mice were in utero electroporated with 2mg/mL GCaMP6s and 2mg/mL DsRed, to transfect pyramidal neurons in layer 2/3 of the visual cortex. Using the protocol previously described by Siegel et.al. (2012), in vivo two-photon Ca2+_{-imaging and} electrophysiological data were collected by Dr. Juliette Cheyne and Alexandra Leighton. For the collection of anatomical data, E16 mice were in utero electroporated with GFP, to transfect neurons in layer 2/3 of the visual cortex.

Upon functional data collection, animals were transcardially perfused with phosphate-buffered saline (PBS), followed by 4% paraformaldehyde in PBS. For anatomical data, transcardial perfusions were performed at 3 different ages; P8, P12 and P15. The brains were removed from the animal. If needed, postfixation in 4% paraformaldehyde in PBS was performed. For both hemispheres, the cerebral cortex was dissected and flattened. Flattened cortices were first immersed in 15% sucrose after which they were stored in 25% sucrose for cryoprotection.

6 3.2 Tissue pre-treatment

To enhance penetration of solutions used during immunohistochemical staining, the cortices were subjected to 10 freeze/thaw cycles while in 25% sucrose. As the expression of GCaMP6s and DsRed would interfere with the fluorescent signals used for staining different synaptic markers, the tissue fluorescence was bleached before staining. To avoid tissue damage due to the formation of gas bubbles during the bleaching reaction and to reduce the severity of the bleaching reaction, the flattened cortex was first dehydrated using increasing methanol concentrations. Dehydration was performed by consecutive immersion in 50% methanol in PBS, 80% methanol in PBS, and twice in 100% methanol, at room temperature for 1hr each. Subsequently, the cortex was immersed for 10 minutes at room temperature in bleaching solution: 500 µL DMSO, 500 µL 30% H2O2, 475 µL H2O and 25 µL 0.1M NaOH. Bleaching of the fluorescent signal was terminated by two 10-minute washes in 100% methanol. Rehydration was performed by consecutive immersion in 80% methanol in PBS, 50% methanol in PBS and PBS for 1hr each at room temperature.

3.3 Immunohistochemical staining

All primary and secondary antibody solutions were prepared in blocking solution. Blocking solution for whole cortex and 400 µm section staining were prepared with of PBS with 0.1% TritonX and 0.01% azide (PBST), to which 5% Normal horse serum (NHS), 1% TritonX and 1% saponin was added. Blocking solution for staining on sections thinner than 400µm consisted of PBST containing 5% NHS, 0.5% TritonX and 0.5% saponin.

In utero electroporation - GCaMP6S In utero electroporation - GFP In vivo recordings - Whole-cell patch clamp Immunostaining - GFP - vGlut1 - vGlut2 - Streptavidin Optical clearing - SeeDB2S Confocal imaging - Tilescan - Dendritic segments - z-stack, 40x oil objective Cluster analysis - Anatomical clustering - Functional clustering

Figure 1 Workflow of the protocol for the quantification of clustering. Dashed boxes are procedures not executed by the student.

7 2.3.1. Immunohistochemical staining of functional and anatomical data

For visualization of the dendrite, an anti-GFP antibody was used, which stains both cells expressing GCaMP6s and GFP. To distinguish between excitatory synapses with a thalamocortical (TC) or corticocortical (CC) origin, two different excitatory markers were used; vesicular glutamate transporter 1 (vGlut1) was used to stain corticocortical projections, and vesicular glutamate transporter 2 (vGlut2) was used as a marker for thalamocortical projections (Fremaeau et.al. 2001; Fujiyama et.al. 2003). For the cortices of which functional data was collected, immunohistochemical staining was performed on the flattened cortex to keep the patched cell intact. For purely anatomical data, the GFP expressing cortices were sectioned into 400 µm thick sections. Flattened cortices or sections were incubated in primary antibody solution (Table 1). Flattened cortices were incubated for 3 days at 37°C and 3 days at 5°C, and 400µm thick sections were incubated for 3 days at 37°C. Cortices/sections were washed 4 times 15 minutes in PBST. Next, cortices/sections were incubated in secondary antibody solution (Table 2). Flattened cortices were incubated for 4 days at 37°C. 400µm thick sections were incubated overnight at 37°C. To the secondary antibody solution of cortices from which functional data was collected, streptavidin-650 was added to visualize biocytin, with which the patched cell was filled after data collection. Because the streptavidin antibody causes high noise levels, 3% bovine serum was added to the blocking solution for the incubation with secondary antibody solution. After incubation with the secondary antibody solution, tissue was washed 4 times 15 minutes in PBST.

Initially the aim was to include GAD65/67 as a marker for inhibitory synapses but the antibody combination did not work (S1).

Flattened cortex staining

Marker Primary antibody (host-α-target) Concentration Manufacturer

GFP Rabbit-α-GFP 1/300 Abcam #ab6556

vGlut1 Guinea pig-α-vGlut1 1/300 Millipore #AB5905

vGlut2 Chicken-α-vGlut2 1/300 Synaptic Systems #135416

400 µm section staining

GFP Rabbit-α-GFP 1/1000 Abcam #ab6556

vGlut1 Guinea pig-α-vGlut1 1/500 Millipore #AB5905

vGlut2 Chicken-α-vGlut2 1/500 Synaptic Systems #135416

Flattened cortex and 400 µm section staining

Marker Secondary antibody (host-α-target) Concentration Manufacturer

GFP Goat-α-Rabbit F(ab)2 (Alexa 488) 1/500 Jackson ImmunoResearch #111-546-003

vGlut1 Donkey-α-Guinea Pig F(ab)2 (Alexa 594) 1/500 Jackson ImmunoResearch #706-586-148

vGlut2 Donkey-α-Chicken F(ab)2 (Cy3) 1/500 Jackson ImmunoResearch #703-166-155

Only added in solution used for cortices of functional data

Biocytin Streptavidin-650 F(ab)2 1/1000 ThermoFisher #84547

3.4 Tissue clearing

To allow high-resolution fluorescence imaging, optical clearing was performed with the SeeDB2S protocol. With this method, the refractive index of the fixed tissue is matched to that of immersion oil (1.518), thereby minimizing light scattering and spherical aberrations, while preserving morphology and protein fluorescence (Sigal et.al. 2015; Ke et.al. 2013; Ke et.al. 2016). The tissue, either flattened

Table 2 Primary antibody solution for anatomical data of flattened cortex and 400µm sections

8 cortex or 400 µm sections, was cleared in 4 steps. In the first step, the cortex was immersed in a solution with 300 µL saponin 20%, 1700 µL H2O and 1000 µL of 350 mg Histodenz per 1 mL H2O. The second step contained 300 µL saponin 20%, 1200 µL H2O and 1500 µL of 350 mg Histodenz per 1 mL H2O. The third step contained 300 µL saponin 20% and 2700 µL of 350 mg Histodenz per 1 mL H2O. The fourth step contained 300 µL saponin and 2700 µL SeeDB2S (10 g Histodenz in 3.78 ml 10mM Tris-1mM EDTA). The tissue was incubated for 1 hr in step one, two and three and overnight in step 4. 3.5 Confocal microscopy

Imaging of the tissue was performed on a Leica TCS SP8 confocal microscope, equipped with a white fiber laser source. The cleared cortex was mounted in a medium consisting of 9.5 g Histodenz and 0.4 g Mowiol in 3.78 mL 10mM Tris-1mM EDTA. The refractive index of this mounting medium is 1.521 which is very close to the 1.518 of immersion oil. High quality coverslips of 170µm were used to reduce variability within and between coverslips, thereby minimizing differences in aberrations of the light. To image dendrites, a 40x oil immersion objective with a NA of 1.3 was used. A z-stack of a dendritic segment was imaged with a zoom of 4.0, z-step size of 100 nm. Because image deconvolution is performed on the data, oversampling was needed. For this reason, image width was kept constant over all imaged dendritic segments to assure a pixel size of 49 x 49 nm in all images. In order to minimize bleaching of the fluorescent signal, laser power was kept low (1-5%) for all four excitation channels. Besides that, it was aimed to keep the size of the z-stack below 6 µm.

To find the functionally analysed cell, a tile scan was performed which was scanned for the GFP signal, present in cells and their dendrites, and the streptavidin signal, which was present in the functionally analysed cell. To perform a tile scan of the entire cortex, a 20x water objective was used. This objective has a higher working distance than the 20x oil objective, which increases the chance of finding the cell when the cell was located deeper in the tissue.

3.6 Data analysis 3.6.1. Deconvolution

During microscopic imaging, an image is convoluted; the image is degraded by blurring and noise. As a result, imaged objects look different from their real, biological size and shape (Beliën & Wouterlood, 2011). To restore the convoluted images, the z-stacks were processed using the deconvolution tool in Huygens Pro software (Huygens Professional, Scientific Volume Imaging, Hilversum, The Netherlands). Deconvolution is done using a theoretical point-spread function (PSF). Using an iterative classic maximum likelihood estimation, background noise is reduced and resolution is increased based on the theoretical PSF. As a result, the image will look more similar to its biological appearance. The highest gain of resolution is in the z-direction.

3.6.2. Synapse identification

Using the ‘Object Analyzer’ tool in Huygens software, synaptic contacts on the dendrite were identified. In the Object Analyzer, a surface rendering for each channel, dendrite, vGlut1 and vGlut2, was produced separately. Individual objects were distinguished from the background using a threshold criterion. The threshold of the dendrite was chosen by making the main dendrite object as continuous as possible without having parts attached that were not part of the dendrite. The latter was judged by going through the raw data. The main dendrite stretch was then selected for object analysis and all

9 other objects in the dendrite channel were discarded. The threshold for the synaptic objects was set in an unbiased manner by choosing the threshold value giving the maximum number of individual objects (Wouterlood et.al. 2008; Wouterlood et.al. 2003). For both vGlut1 and vGlut2, the garbage volume level was set to 0.072 µm3_{, thereby discarding all objects smaller than this volume. This number} was based on previous data where the minimum size of a vGlut2 synaptic terminal, including fluorescent spread, was stated to be at 0.075µm3 _{(Owe et.al. 2013).}

Watershed segmentation was applied to the surface rendering of the synaptic markers to get a better separation of objects that are merged due to the thresholding method. Objects merging due to the thresholding method is likely due to the close proximity of some synaptic contacts which makes their PSFs interfering with each other, even after deconvolution. The watershed segmentation separates merged objects at their local minimum. A watershed sigma value of σ = 0.1 was used. This setting adds a gaussian filter to smoothen intensities in order to avoid over-segmentation of the data.

Although the images are deconvoluted, the 3D reconstructed object will still be larger than the real biological object. As a result, when two membranes are in true apposition, although the biological objects will not touch, the 3D reconstructed objects will still show a partial overlap of voxels. This can be used as a criterion to identify synapses contacting onto the dendrite (Beliën & Wouterlood, 2011). In the Huygens software, the objects colocalizing with the dendrite were identified with a filtering step. Initially, data was filtered on all objects being in contact with the dendrite: >0 voxels overlap. From the distribution of the voxel overlap was decided how many voxels overlap would be considered as a real synaptic contact. This distribution showed a flattening of number of synapses having a number of voxels overlapping with the dendrite higher than 70 voxels, while a lot more objects showed a number of voxels overlapping with the dendrite below this value (Figure 2). The packing of neurons and their processes in the central nerve system is extremely dense. As a result, presynaptic terminals are found around the dendrite that will be synapsing onto structures surrounding the dendrite instead of on the dendrite of interest. Although these presynaptic terminals are thus not forming a synaptic contact onto the dendrite of interest, the signal of these objects can overlap with the dendrite due to their close proximity (Beliën & Wouterlood, 2011). Because the number of overlapping voxels these surrounding terminals have with the dendrite is expected to be lower than of the real synaptic contacts, objects having < 70 voxels overlapping with the dendrite were discarded from further analysis. (Another approach using a marker for post-synaptic terminals was tried but did not work (S2))

The identified contacts were further analysed in MATLAB (Matworks, Natick, MA, USA). The 3D surface rendering was first transformed into a 2D image. To produce a shortest distance matrix, the shortest distance along the dendrite, between all pairs of synaptic contacts, was calculated using the x- and y-coordinates of the centre of mass of the synaptic objects and a binary image of the dendrite.

10 3.6.3. Cluster analysis

Anatomical clustering of synapses was quantified with two different clustering measurements. 3.6.3.1. Nearest Neighbour

For the first measurement, the distance to the nearest neighbour (NN) of either the same (vGlut1-vGlut1 or vGlut2-vGlut2) or the mixed identity ((vGlut1-vGlut1-vGlut2) was extracted from the shortest distance matrix. The median NN distance was then calculated per dendrite. For each dendrite, synapse identity (vGlut1 and vGlut2) was randomized 5000 time, where the synapse identities were shuffled while maintaining both the spatial location and the relative density of vGlut1 and vGlut2. For each of the 5000 simulations, the median NN distance was calculated. we were able to extract the shortest path distances from the shortest distance matrix for NN distance calculations of the simulation. In addition, the spatial location of synapses along the dendrite was also randomized in 5000 simulations where each the synapse was randomly relocated along the dendrite. Again, the relative density of vGlut1 and vGlut2 was maintained. After the relocation of synapses, the distances between all synapse pairs had to be recalculated to find the NN distance for each synapse. The algorithm calculating the shortest path distance between synapses takes a very long time (days to weeks depending on the length of the dendrite). Hence, for each spatial randomization, geometric distances between all synapse pairs were calculated and used to establish the median NN distance per dendrite per simulation. We calculated the difference between the shortest path distance and the geometric between the synapse pairs (shortest path distance – geometric distance) and plotted these against the shortest path distance (Figure 3), highlighting that for close objects the difference between the shortest path distances and the geometric distances negligible .

11 Both for identity randomization and spatial randomization, the median NN distance of the real data was compared to a distribution of the 5000 median values from the randomized data. The p-value was calculated by determining the proportion of medians from the randomized dataset being smaller than the median of the real dataset (Equation 1).

Equation 1: 𝑝 − 𝑣𝑎𝑙𝑢𝑒 = ∑(𝑟𝑎𝑛𝑑𝑜𝑚𝑖𝑧𝑒𝑑 𝑚𝑒𝑑𝑖𝑎𝑛 ≤𝑟𝑒𝑎𝑙 𝑚𝑒𝑑𝑖𝑎𝑛 )

# 𝑆𝑖𝑚𝑢𝑙𝑎𝑡𝑖𝑜𝑛𝑠

2.6.3.2. Clustering coefficient

In previous literature on in vivo and ex vivo data, synapses within 12µm distance from another synapse were more likely to be co-active (Winnubst et.al. 2015). Post hoc tissue fixation, preparation, immunohistochemistry and tissue clearing cause shrinkage of the tissue. To calibrate the amount of tissue shrinkage, images obtained in the in vivo recordings and post hoc immunolabeling were compared. The disparity between the in vivo and immunolabeled images of the distance between two branch points was 60-70%. This would mean that the clustering of synapses, having a high co-activity within 12 µm distance, would correspond to a distance of 8 µm in the post hoc analyses.

The second measurement for clustering, the clustering coefficient (CC), was calculated by dividing the number of synapses within a distance of 8 µm on each site of a synaptic contact by the total number of synapses of a specific identity on the dendrite. Again this was done within or between identities (single identity; vGlut1-vGlut1 or vGlut2-vGlut2 and mixed identity). For single identity clusters, the number of synapses of the same identity were compared to the total number of synapses of that same identity along the entire dendrite. For the mixed identity clusters, the number of synapses within 8µm of both identities were compared to the total number of synapses on the dendrite.

Figure 3 Shortest path distance between all synapse pairs plotted against the difference between the shortest path distance and the geometric distance (Shortest path distance – geometric distance).

12 5000 identity randomizations and spatial randomizations were performed as described above. For each simulation, the median CC was calculated. The median CC of the real data was compared to the distribution of the median CC from the 5000 randomizations. The p-value was calculated by determining the proportion of medians from the randomized dataset being larger than the median of the real dataset (Equation 2).

Equation 2: 𝑝 − 𝑣𝑎𝑙𝑢𝑒 = ∑(𝑟𝑎𝑛𝑑𝑜𝑚𝑖𝑧𝑒𝑑 𝑚𝑒𝑑𝑖𝑎𝑛≥ 𝑟𝑒𝑎𝑙 𝑚𝑒𝑑𝑖𝑎𝑛 )

# 𝑆𝑖𝑚𝑢𝑙𝑎𝑡𝑖𝑜𝑛𝑠

4. RESULTS

4.1. Inhibitory synapse labelling

To include data about the location of inhibitory synapses with respect to clusters of excitatory synapses, the staining protocol was adjusted by including a marker for inhibitory pre-synaptic terminals (S1). While the first test with this antibody combination worked, any of the other stainings performed later showed high levels of colocalization of vGlut1 signal with the GFP signal staining of the dendrite. As vGlut1 is a protein present in synaptic vesicles at the pre-synaptic site, it should have shown a punctate staining outside of the dendrite (Figure 5B). However, the vGlut1 signal was strongly picked up homogeneously inside the dendrite (Figure 4). Multiple tests were performed in order to find what was causing this colocalization (S3).

Figure 4 Illustration of the problem with immunolabeling; colocalization of vGlut1 with the dendrite. A) single z-plane of a dendrite in which all channels are merged. The different colours represent: red = vGlut2, green = vGlut1, grey = GAD65/67, blue = GFP (dendrite). B-D, single channel image of the same z-plane as in A (B = vGlut2, C = GAD65/67, D = vGlut1 and E = GFP). Scale bar = 5µm

13 Because the vGlut1 was initially imaged in the 488 channel, it was hypothesized that remaining GFP signal that would still be present after bleaching was picked up inside the dendrite, causing the co-localization. A bleached section incubated in blocking solution without antibodies was therefore imaged and indeed, dendrites were visible in the 488 channel (Table S3A.1, S3A.2).

However, when changing the labelling scheme to, GFP being labelled with a 488-conjugated secondary antibody and vGlut1 with a cy3-conjugated secondary antibody, the colocalization of vGlut1 inside the dendrite was still present (Table S3A.3). Although the fluorescent spectra of the 488 and the cy3 fluorophores partly overlap, detection windows were critically set and channels were imaged sequentially. The colocalization could therefore not be explained by bleed through of one of the fluorophores into the others channel. This was checked by labelling the vGlut2 with a secondary antibody conjugated to the cy3 fluorophore and here, channels were well separated and no bleed through of the vGlut2 signal into the GFP channel was found (Table S3A.4).

Changes in the protocol, such as decreasing the incubation time, concentration of the primary vGlut1 antibody, incubation temperature and removing saponin from the blocking solution also did not resolve the issue of vGlut1 being colocalized with the dendrite (Table S3A.4-S3A.8).

To test if the either the primary or the secondary vGlut1 antibody was binding to the dendrite, bleached and non-bleached sections were stained without an antibody against GFP but with an anti- vGlut1 antibody(Table S3A.9, S3A.10). Here, no colocalization of vGlut1 with the dendrite was shown. When only the secondary antibody for GFP labelling was left out of the antibody solutions and the primary antibody against GFP was thus incubated on non-bleached sections, colocalization of vGlut1 with the dendrite was detected (Table S3A.11, S3A.12). It could thus be concluded that the primary and/or secondary vGlut1 antibody were interacting with the Sh-α-GFP antibody. Tests on which of the two vGlut1 antibodies (primary or secondary) was interacting with the Sh-α-GFP antibody gave no exclusion of one of them. When the primary antibody was changed to a Guinea Pig-α-vGlut2 antibody and the same secondary antibody was used against the Guinea Pig, colocalization of the vGlut2 signal with the Sh-α-GFP was detected (Table S3A.13). Although one might think that the secondary antibody would thus cause the colocalization by interacting with Sh-α-GFP, and not the primary antibody, a test in which the secondary antibody was changed from the Goat-α-Guinea Pig F(ab)2 (Cy3) antibody to a Donkey-α-Guinea-Pig F(ab)2 (Alexa 594) antibody, also showed co-localization of vGlut1 in the dendrite (Table S3A.14). Both the primary and the secondary antibody used to label vGlut1 thus show problems in combination with the Sh-α-GFP for dendrite labelling.

Different vGlut1 antibodies were tested to find out if the quality of one of the antibodies was sufficient to replace the Gp-α-vGlut1 antibody, which was not the case. Antibody penetration into the tissue was low and the signal had high noise levels (Table S3A.15-S3A.17).

Combining the primary and secondary antibody used to label vGlut1 with a rabbit-α-GFP antibody did not cause the colocalization of vGlut1 with the dendrite thus confirming the interaction with the Sh-α-GFP antibody. However, using the rabbit as a host for labelling the dendrite excludes the Rabbit-α-GAD65/67 antibody from the labelling scheme. The quality of a Mouse-α-GAD65 antibody was tested but this antibody did not work properly with the protocol (TableS3A.18). Also vgat, another inhibitory synapse marker, was tried but this antibody does not work on methanol treated tissue.

The labelling of inhibitory synapses was thus not possible and further stainings were performed according to the labelling scheme described in the Methods section (Table 1 and 2, Figure 5)

14 4.2. Control experiments

For the cells that had been recorded in vivo, the penetration depth of both the primary and secondary antibodies into the tissue is important. Even though tissue was permeabilized using Triton-X100 and saponin, the facilitation of antibody penetration as a result of this will never be sufficient to label the entire flattened cortex. To check if immunolabeling in layer2/3 of the cortex, was sufficient, the flattened cortex was first stained for either vGlut1 or vGlut2 with the standard staining protocol (incubation time and concentration of the antibodies). After this, the flattened cortex was sectioned in 100µm thick coronal sections, which were re-stained for the same marker (S4). For both vGlut1 and vGlut2, strong labelling at superficial layers was observed. Penetration depth for vGlut1 was lower than for vGlut2 but for both labels, density was sufficient at a depth of 150-200µm from the surface, at which dendrites were typically imaged (Figure 6).

To check if the there was no bleaching of the signal at the bottom of the z-stack, the synapse density of the top 20 frames was compared to the synapse density in the bottom 20 frames. No differences were observed in the number of objects detected in the first 20 frames versus the last 20 frames of the same z-stack. Bleaching of the signal during confocal imaging thus does not affect object analysis. A

B

Figure 5 Illustration of a confocal image of a dendrite. A) Maximum projection of the dendrite (GFP) reconstructed from two adjacently imaged parts of the dendrite. The two adjacent parts are marked by the dotted boxes around it. B) Single z-plane of the raw image of one of the two dendrite parts; green = vGlut1, red = vGlut2, blue = GFP (dendrite). Scale bar A = 5µm, B = 2.8 µm

15

A

B

Figure 6 False negative controls to check for vGlut1 (A) and vGlut2 (B) antibody penetration into the tissue. Where red and green co-localize, synaptic terminals were stained in both whole cortex and sections. Where only green punctates are stained, antibodies did not penetrate into the tissue to this depth. Red = labelling from whole cortex staining, green = labelling from section staining. Scale bar A top = 20µm, A bottom = 16µm, B top = 30 µm, B bottom = 15 µm

16 4.3 Finding the cell that was patched in vivo

A tile scan for 2 markers was made of the entire cortex. The first marker was the stained GFP as a label for the cells and dendrites. After the cell was imaged in vivo, it was filled with biocytin. Streptavidin, an antibody binding to biocytin, was used as the second marker to find the biocytin labelled cell. Where the GFP and streptavidin signal co-localized, the cell that was measured in vivo could be found (Figure 7). Subsequently, the imaged dendrite of that particular cell was searched for and imaged for anatomical analysis of synaptic inputs on this dendrite. Further analysis described were done on data that was previously obtained in the lab (For methods, see S5)

4.4. Synapse density

The synapse density was calculated being the number of synapses per micron of dendrite (Figure 8). This was calculated for vGlut1 and vGlut2 synapses separately and for each age. Both for vGlut1 and vGlut2 follow a similar pattern of synapse density with maturity, in which the synapse density first decreases after which it increases again. For vGlut1 this pattern seems to start earlier than for vGlut2. However, no younger ages were imaged.

Figure 7 Co-localization of the biocytin and GFP signal is the cell that was imaged in vivo. Red = biocytin, green = GFP Scale bar 20 µm 0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 P8 P9 P10 P11 P12 n o syn /µ m d en d rite

vGlut1 density vGlut2 density

Figure 8 Synapse density (number of synapses/µm dendrite) per age for vGlut1 (green) and vGlut2 (red) separately. 1 brain per age. N(number of dendrites): P8, N = 5, P9, N = 8, P10, N = 5, P11, N = 6, P12 N = 7

17 4.5. Distance distribution

When synapses would be randomly distributed along the dendrite, the number of pairs having a longer distance between them would decrease in a linear like fashion (Figure 9A). Most dendrites showed this distance distribution. However, at clustered dendrites, a high number of synapse pairs are expected to be close to each other, as well as very far distances, with a drop in between, which are the ‘empty’ areas in between clusters (Figure 9B). Of the 12 dendrites for which this distance distribution of the shortest path distance was analysed, 3 dendrites clearly showed this bimodal distribution of distances between synapse pairs. The distance of the first peak was for all 3 dendrites just below 8 microns after which there was a drop which would correspond to the 8 micron distance in post-hoc analysis within which synapses were more likely to be co-active.

4.6. CLUSTER ANALYSIS

4.6.1 Nearest Neighbour cluster analysis

For every synaptic contact of a dendrite, the NN distance was calculated, for both the distance to a contact’s nearest neighbour of the same and the other identity (vGlut1 and vGlut2). The median NN distance was calculated for the real dataset for all possible types of clustering (single and mixed). Subsequently, the identities of the synapses were randomized 5000 times, keeping the relative density between vGlut1 and vGlut2 contacts the same. For each of these 5000 simulations, the median NN distance was calculated, again for all types of clustering. A distribution of the 5000 median NN distances was plotted for each type of clustering. The median NN distance from the real data was compared to the distribution of randomized median NN distances (Figure 10).

From the dataset of 31 dendrites, 9 dendrites showed significant NN clustering; 3 dendrites for vGlut1 synapse pairs (CC-CC), 2 dendrites for vGlut2 synapse pairs (TC-TC), and 4 dendrites for mixed synapse pairs (CC-TC) (Figure 10B). Additionally, another five dendrites showed a tendency towards clustering, having a p-value <0.10. In all cases, only one type of clustering, vGlut1, vGlut2 or Mixed pairs, was significant in a single dendrite.

A B

Figure 9 Distribution of distances between synapse pairs on one dendrite. A) Example of a distance distribution of a dendrite with a random (not clustered) synapses distribution. B) Example of a bimodal distance distribution of a dendrite that is clustered.

18 In contrast to NN clustering, 9 dendrites had a p-value >0.95, indicating synapses to significantly avoid each other. This significant avoidance of synapses will be indicated as anti-clustering (Figure 10C). Also, 7 dendrites showed a tendency towards anti-clustering (p > 0.90).

A

B

C

Figure 10 left: overlay of the markers on the maximum projection of the dendrite. Grey = dendrite, green = vGlut1, red = vGlut2. Right: distribution of median nearest neighbor distance in micron of the dendrite on the left. Grey = distribution of all medians from 5000 simulations, red vertical line = median of the real datsaset for vGlut2 synapse pairs, yellow vertical line = median of the real dataset for mixed synapse pairs. A) Example of a dendrite and the median distribution, with random vGlut2 synapse distribution along the dendrite (NN p = 0.5802). B) Example of a dendrite and the median distribution, with significant nearest neighbor clustering for mixed synapse pairs (NN p = 0.0382). C) Example of a dendrite and the median distribution, with a tendency for avoidance (anti-clustering) between mixed synapse pairs (NN p = 0.9142). Scale bar = 10µm

19 Interestingly, when comparing single identity clustering to mixed identity clustering, the dendrites that showed single identity clustering or a tendency towards this, appeared to be anti-clustered for mixed identity pairs or have a tendency towards this (Figure 11A, black bar). For the opposite, significant NN clustering for mixed identity pairs and anti-clustering in one of the single identity pairs, this pattern was observed for 4 of the 5 dendrites. However, anti-clustering occurred in either vGlut1 or vGlut2 but not in both (Figure 11A yellow bar). Remarkably, dendrites never showed significant clustering for both vGlut1 and vGlut2 in the same dendrite (Figure 11B).

For each type of clustering, 31 separate tests were performed. With the threshold for significant clustering at 0.05, 1.5 false positive results are likely to be found. However, for all types of clustering, the number of clustered dendrites were higher than 2 and significant results are thus not likely to be produced by chance.

Nearest Neighbour clustering analysis with spatial randomization did not show any significant results.

4.6.2. Clustering Coefficient

For every synaptic contact of a dendrite, the Clustering Coefficient (CC) was calculated as a second measure for synaptic clustering. Single identity clusters were calculated as the proportion of synapses of the same identity within 8µm relative to the total number of the same identity synapses on the dendrite. For mixed identity clusters, the number of synapses within 8µm distance of each synapse was divided by the total number of synapses on the dendrite. The median CC was calculated for all possible types of clustering. In 5000 simulations, the identity of synapses on the dendrite was Figure 11 Data from nearest neighbour analysis with identity randomization A) P-values of the single identity nearest neighbour clustering analysis (green is vGlut1, red = vGlut2) plotted against the p-values of the mixed identity nearest neighbour clustering analysis. Yellow bar = dendrites with significant nearest neighbour clustering (P < 0.05) or a tendency towards clustering (P < 0.10) for mixed synapse pairs (5 dendrites). Black bar = dendrites with significant nearest neighbour clustering (P < 0.05) or a tendency towards clustering (P < 0.10) for single identity synapse pairs (5 dendrites clustered for vGlut1 and 4 dendrites clustered for vGlut2). B) P-value of the nearest neighbour clustering analysis for vGlut1 plotted against the p-value of the nearest neighbour clustering analysis for vGlut2. Green bar = dendrites with significant nearest neighbour clustering (P < 0.05) or a tendency towards clustering (P < 0.10) for vGlut1. Red bar = dendrites with significant nearest neighbour clustering (P < 0.05) or a tendency towards clustering (P < 0.10) for vGlut2)

20 randomized and a distribution of 5000 median CC values was plotted for both single identity CCs (vGlut1 and vGlut2).

The CC of the real data was then compared to the distribution of randomized median CCs (Figure 12) . For mixed clusters analysis with identity randomization could not be done as the number of synapses within 8µm will remain constant when randomizing synaptic identities.

A

B

C

Figure 12 left: overlay of the markers on the maximum projection of the dendrite. Grey = dendrite, green = vGlut1, red = vGlut2. Right: distribution of the median clustering coefficient for the dendrite on the left. Grey = distribution of all medians from 5000 simulations, green vertical line = median of the real dataset for vGlut1 synapses, red vertical line = median of the real datsaset for vGlut2 synapses. A) Example of a dendrite and the median distribution, with random vGlut1 synapse distribution along the dendrite (CC p =0.5778). B) Example of a dendrite and the median distribution, with a significant clustering coefficient for vGlut2 synapses (CC p = 0.0282). C) Example of a dendrite and the median distribution, with significant avoidance (anti-clustering) between vGlut1 synapses (CC p = 0.968). Scale bar = 10µm

21 12 dendrites were significantly clustered when the median CC was compared to the median distribution of identity randomized simulations. Of these, 7 are vGlut1 clusters and 5 are vGlut2 clusters. An additional 10 dendrites showed a tendency for clustering. Again, dendrites appeared to be clustered for either vGlut1 or vGlut2 but not both, except for one dendrite being clustered for vGlut1 and showing a tendency towards clustering for vGlut2 (Figure 13).

Comparing the median CC of the real data to a distribution of median CCs of 5000 spatially randomized simulations resulted in 15 dendrites being clustered and 3 dendrites that showed a tendency towards clustering. Mixed clustering was observed on 9 dendrites. In most dendrites with mixed clusters, single identity clusters were also observed. However these dendrites are not adopted as having single identity clusters as these single identity clusters were part of the mixed clusters on these dendrites (Figure 14A yellow bar). 8 dendrites were clustered for a single identity, again being either vGlut1 or vGlut2 but not both, except for one dendrite (Figure 14B).

4.6.3. Comparison of the 3 methods

The correspondence of the clustering methods was investigated by focusing only on significant clustering for all the methods and all the cluster compositions (Figure 15). Of the 44 times, a significant clustering value was obtained, more than half of them showed only clustering for one type of measurement (either NN or CC with either identity or spatial randomization). In 2 cases, NN clustering with identity randomization (NNI) and CC for both identity (CCI) and spatial (CCS) randomization were all three shown to be significant. Not much correspondence was found between the NN and the CC. 16 dendrites showed significant CCs for both randomizations.

0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1 0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1 p -v alu e vG lut 2 p-value vGlut1

Figure 13 Data from clustering coefficient with identity randomization. P-value of the clustering coefficient analysis for vGlut1 plotted against the p-value of the clustering coefficient analysis for vGlut2. Green bar = dendrites with a significant clustering coefficient (P < 0.05) or a tendency towards clustering (P < 0.10) for vGlut1. Red bar = dendrites with a significant clustering coefficient (P < 0.05) or a tendency towards clustering (P < 0.10) for vGlut2

22

Clustered for a single measurement

Custered for NNI and CCI Clustered for NNI and CCS Clustered for CCI and CCS Clustered for all 3 measurement

Figure 15 Correspondence of significance between the different measurements for clustering. NNI = nearest neighbour identity randomization, CCI = clustering coefficient identitiy randomization, CCS = clustering coefficient spatial

randomization. Clustering of a single measurement = 23, Clustered for NNI and CCI = 2, Clustering for NNI and CCS = 1, Clustering for CCI and CCS = 16, Clustering for all 3 measurements = 2

Figure 14 Data from clustering coefficient with spatial randomization A) P-values of the single identity clustering coefficient analysis (green is vGlut1, red = vGlut2) plotted against the p-values of the mixed identity clustering coefficient analysis. Yellow bar = dendrites with a significant clustering coefficient (P < 0.05) or a tendency towards clustering (P < 0.10) for mixed clusters (9 dendrites). Black bar = dendrites with a significant clustering coefficient (P < 0.05) or a tendency towards clustering (P < 0.10) for single identity clusters (5 dendrites clustered for vGlut1 and 5 dendrites clustered for vGlut2). B) P-value of the clustering coefficient analysis for vGlut1 plotted against the p-P-value of the clustering coefficient analysis for vGlut2. Green bar = dendrites with a significant clustering coefficient (P < 0.05) or a tendency towards clustering (P < 0.10) for vGlut1. Red bar = dendrites with a significant clustering coefficient (P < 0.05) or a tendency towards clustering (P < 0.10) for vGlut2).

23 5. DISCUSSION AND CONCLUSION

To increase the computational capacity of dendrites, they should contain active compartments where synapses are arranged with high spatiotemporal correlation. In this research, two methods were used to determine the spatial arrangement of synapses on dendrites of layer 2/3 pyramidal neurons in de visual cortex during development. Indeed different types of clustering were observed in 25 dendrites from a set of 31 dendrites. Remarkably, dendrites clustered with a single identity appeared to have either CC or TC clusters. Additionally, dendrites with single identity clustering showed avoidance between the two identities. These findings suggest that synapses are specifically arranged in a single type of clustering on dendritic compartments. In addition, a subset of dendrites showed a tendency towards clustering. Dendrites were analysed during development, where refinement of synaptic connections is an ongoing process. Dendrites that showed a tendency towards clustering might therefore be in the process of pruning synaptic contacts depending a local ‘out of sync – lose your link’ plasticity mechanism that has been described before (Winnubst et.al. 2015).

5.1 Methodology

Several control experiments were done to investigate if methodological limitations would have an influence on the data that was collected during the experiments. No significant loss of fluorescent signal was found when comparing the top frames of a dendritic z-stack to the bottom frames. Fluorescent bleaching of the labels used in our protocol was thus not problematic for our data analysis. Also, penetration depth of the antibodies was evaluated. Although for labelling was especially strong at the superficial layers and decreased as a function of depth, labelling was still sufficient at cortical depths corresponding to the depth at which dendrites from layer 2/3 neurons were imaged. Penetration of the vGlut2 antibody was better than of the vGlut1 antibody. This discrepancy could possibly be explained by differences in protein expression. The more densely expressed vGlut1 protein might limit antibody penetration further into the tissue.

The density of both vGlut1 and vGlut2 synaptic contacts onto the dendrite followed a typical pattern in which the synapse density first showed a decrease, after which the synapse density increased again. This pattern of vGlut2 seems to be shifted to occur at a later age with respect to vGlut1. However, synapse density was based on dendrites from only one brain per age, which doesn’t allow reliable statistical analysis of the data. Although brains are stained with the same protocol, the pattern observed in this data could also be due to differences in the quality of the immunohistological staining. In order to conclude anything about the pattern that was observed, more animals should be added to the analysis, after which proper statistics could be performed. Furthermore, to establish if the this pattern starts earlier for vGlut1 synapse density, younger animals should be added to the dataset. Previous literature described that the proportion of thalamocortical inputs onto dendrites was small (+/- 15%) (Benshalom & White, 1986). In contrast to this, synaptic density of thalamocortical inputs in our dataset was a lot higher than this percentage. Although antibody penetration was sufficient, limited penetration depth of the vGlut1 antibody might still cause undetected corticocortical synapses, resulting in a synapse density biased towards higher proportions of thalamocortical projections. Another methodological aspect that should be carefully considered is the detection of synaptic contacts onto the dendrite. Synaptic contacts were identified based on the assumption that the signal of two opposing fluorescently labelled objects will partly overlap (Beliën & Wouterlood, 2011). For synapse identification, a filtering step was performed, keeping only objects with co-localization of at

24 least 70 voxels with the dendrite. Previous literature used a criterion of at least 100 overlapping voxels, which was based on Abbe-type diffraction (Beliën & Wouterlood, 2011). In other studies, a less stringent criterion was used, including all objects having at least 1 voxel overlapping with the dendrite (Schoonover, 2014; Ausdenmoore et.al. 2011). However, due to the dense packing of CNS neurons and their processes, the fluorescent signal of objects can overlap without evidence for synaptic transmission. Our voxel overlap criterion was based on a distribution of voxel overlap after object analysis with a filtering step for all objects contacting the dendrite with at least 1 overlapping voxel. The flattening of the distribution occurred at approximately 70 overlapping voxels which was for our analysis adapted as the filtering criterion for synapse identification. Another approach on determining the filtering criterion was tried. This approach aimed to calibrate the number of overlapping voxels using the post-synaptic marker Homer1 (S2). This, however, added a similar question to the list: how do we define a homer1 punctate as belonging to the dendrite of interest? This approach did therefore not work. Finding an optimal filtering criterion for synapse identification is extremely complicated. Dendrites receive inputs at all sides of the dendrite. The radial and axial resolution of confocal data are different and therefore, the Abbe-type diffraction, on which the synapse identification is based, will be different based on the localization of the synapse with respect to the dendrite in our image. Alignment of the anatomical data to functional data of the same dendritic segment might give more insights in the reliability of our thresholding criterion. Finding a more reliable method for identifying synapses would be helpful but until a better method has been established, synapse identification will remain an important discussion point.

The assumption that vGlut1 synapses represent corticocortical projections and vGlut2 synapses represent thalamocortical projections holds true in the adult mouse brain (Fremeau et.al. 2001; Fujiyama et.al. 2003). However, during postnatal development, occasionally, vGlut1 and vGlut 2 were reported to colocalize (Nakamura et.al. 2005; Nakamura et.al. 2007). However, colocalization of vGlut1 and vGlut2 punctates was observed very rarely in our data and it is unlikely that significant clustering of mixed identities is caused by the colocalization of vGlut1 and vGlut2 in the same synaptic terminal. Two clustering measurements were calculated from the data. The median value of the dataset was subsequently compared to a distribution of medians obtained by 5000 randomizations of the dataset A false positive rate of 5% was used. Two types of randomization were performed. The shuffling of identities had been described before in a study where nearest neighbour clustering of thalamocortical projections in layer 4 of the barrel cortex was found (Schoonover, 2014). Spatial randomization was done by relocating synapses along the dendrite, and for each simulation, geometric distances between synapse pairs were calculated. The correspondence between NN clustering and CC clustering is low. This suggests that the two methods indicate different types of clustering. While the median nearest neighbour values were always very short distances (0.5-3 µm), the clustering coefficient includes a larger area of 8µm around an object. In addition, the nearest neighbour measurement only looks at the spatial relation between 2 synaptic contacts, while the clustering coefficient counts all synaptic contacts within a surrounding region of an object. The two methods thus represent different cluster sizes which might explain their low correspondence. Including a second nearest neighbour in the analysis might result in a higher correspondence between the two measurements.

The two types of randomization are thought explain different levels of clustering. While spatial randomization will tell us if synapses are spatially clustered on dendrites, the measurement will be less convincing for determining the composition of these clusters. Identity randomizations will give more

25 reliable results on the composition of clusters while its say about spatial relationships will be more debatable. A combination of the two randomizations increases the accuracy of the clustering analysis from which both spatial and compositional conclusions can be drawn.

Local changes of synaptic density along a dendritic strength can influence the results for the clustering coefficient. In our dataset, occasionally, dendrites showed a bias in synaptic density towards one side of the dendritic stretch. A better method might therefore be to compare the number of synapses in the ‘cluster region’ of 8 µm around a synapse to an adjacent area of 8µm on each side of the cluster region. In this way, the problem of local density differences could be tackled. This method could however not be used on the data in this study. The vast majority of the dendritic segments analysed had a length of approximately 70µm. For the clustering coefficient to be calculated compared to local densities, the entire region of 16 microns on each side of the dendrite should be located on the dendrite. Therefore, synapses located within 16µm from the end of a dendritic segment have to be excluded from analysis. Collection of additional data in this project should thus be done by imaging multiple adjacent dendritic segments of 70µm, which can be reconstructed into one dendritic stretch, as was shown in figure 5.

5.2. The role of synaptic clustering

The classical view on dendrites as the passive, input receiving, structure of neurons has been contradicted by recent research, in which dendrites were shown to possess active compartments, in which excitatory inputs are summed in a non-linear fashion (Winnubst & Lohmann, 2012). In the hippocampus, integration of co-active inputs localized in close proximity, initiated dendritic spikes, resulting in a temporally precise and stable action potential output at the soma (Ariav et.al. 2003, Gasparini & Magee, 2006). The computational purpose of the active dendritic compartments will depend on the composition of these spatiotemporally clustered synapses. Single identity clusters would enhance the influence of a specific region on action potential firing. Single identity clustering would be favourable for sparse connections from similar input regions, as their clustering would enhance the contribution to action potential firing. Only a small proportion (approximately 15%) of synaptic inputs to L4 cortical neurons were shown to be represented by thalamocortical projections (Benshalom & White, 1986). In addition, the synaptic strength of thalamocortical projections onto layer 4 neurons was shown to be equal to the strength of corticocortical projections in acute slices (Schoonover et.al. 2014; Bruno & Sakmann, 2006). The localization of thalamocortical synapses was shown to be slightly biased towards areas close to the soma. This, however, did not explain the increased influence on action potential firing. As the above described features of thalamocortical projections did not enhance synaptic influence on somatic firing, it would be advantageous for thalamocortical projections to be spatiotemporally clustered. Clustering of thalamocortical projections on pyramidal neurons in layer 4 of the rat barrel cortex was shown before (Schoonover et.al. 2014). Thalamocortical clustering was supported by our results, in which a subset of dendrites showed this type of single identity clustering. The proportion of thalamocortical synapses in our dataset was however bigger than the 15% that was shown by Benshalom and White. The high reliability of thalamic signal transmission to cortical area might thus be explained by the clustering of the inputs onto cortical dendrites.

Clustering of corticocortical projections onto dendrites has not been reported before. Corticocortical projections can originate from different locations in the cortex. The projections can come from a different layer in the same column but also horizontal connections from neurons in the same layer and

26 projections coming from the corpus callosum are received by layer 2/3 neurons (Gilbert & Wiesel, 1982; Czeiger & White, 1993; Nieuwenhuys, 1994). Studies using retrograde and anterograde tracers, showed that long-range horizontal connections are clustered in the cortex, connecting columns with similar orientation specificity in the cat visual system (Gilbert & Wiesel, 1998; Katz & Callaway 1990). Clustering of corticocortical projections was found in the analysed dataset. This clustered structure might enhance communication within or between cortical areas.

For single identity, anatomical clusters to be functionally relevant, the activation of the clustered inputs must be temporally correlated. Neighbouring neurons in the LGN with an overlapping receptive field were previously shown to elicit synchronous spiking activity. In addition, these neighbouring cells projected onto a common neuron in the cortex (Alonso et.al. 1996). Summation of synchronous events was shown to effectively drive postsynaptic spiking (Singer & Gray, 1995, Usrey et.al. 2000). Together, spatially and temporally clustered inputs will provide a highly efficient signal transmission mechanism. The likelihood of mixed inputs being clustered on dendrites was reduced on dendrites where inputs from a single source were clustered. On these dendrites, inputs from different sources even showed significant avoidance. This could simply be a result of spatial competition. The finding that dendritic segments had clusters composed of projections from a single source suggests a local supralinear summation of these inputs, which will be transmitted towards the soma where the signal will be integrated with inputs from different dendritic segments. Single identity clustering proposes the soma as the centre for integrating information.

However, clusters containing both corticocortical and thalamocortical projections were detected in our dataset. Although the likelihood of highly precise spatiotemporal correlation between inputs from different sources is small, functional co-activity of mixed clusters could greatly enhance information-processing capacity of neuron, by serving as an intermediate integration step where action potential firing would only be enhanced when inputs from different origins are spatiotemporally combined. 5.3. Future perspectives

The establishment of synaptic clustering during development raises an interesting question on how clustering evolves. Our dataset had dendrites from 5 brains with different ages (P8-12). Clustering was detected in all ages and no specific pattern in the development of clusters could be observed. However, anatomical data collection in animals of 3 different ages (P8, P12 and P15) was started. More data should be added and analysis should be performed. It would be interesting to investigate how the establishment of synaptic clusters evolves up to a state where the system is able to react to environmental stimuli. The vGlut2 synaptic density being higher than stated in literature suggests that pruning is possibly an ongoing process in the ages that were studied in our project. Functional studies on synaptic clustering suggested a relationship between increasing synapse density and spatially sharper defined synaptic clusters with age (Kleindienst et.al. 2011; Takahashi et.al. 2012, Govindarajan et.al. 2006; Winnubst & Lohmann 2012). While in the third postnatal week, high co-activity of synapses in close proximity was confined to an 8 µm distance (Takahashi et.al. 2012), earlier in development, this distance was shown to be 16µm (Kleindienst et.al. 2011). Additionally, refinement of clustered retinal projections to the thalamus was shown to increase from P8 onwards in mice (Hong et.al. 2014). If the refinement of thalamocortical projections follows this timeline, clustering of thalamocortical projections might also evolve from p8 onwards. Functional co-activity, however, had already been observed at this age in vivo, thereby selectively maintaining these co-active projections (Winnubst et.al. 2015). The proposed ‘out of sync – lose your link’ plasticity