Enabling 3D Multiplexing Spatial Omics Workflows in Neuroscience

List of pre-synaptic markers

A description of the markers for this case study is described in Figure 3. In Figure 1 you can see a simplified version of the localization of some of these proteins in the active zone of the synapse.

The coronal brain tissue section (Figure 2) used in this case study was obtained from a 10-month-old adult mouse from an NGS background.

Staining

Cyclical staining of the tissue section was accomplished with the Akoya PhenoCycler system (formerly CODEX®). The imaging was performed with the ZEISS LSM 980 confocal microscope using the ZEN acquisition software. The staining and imaging rounds were synchronized through custom macros written in the ZEISS ZEN “macro” environment (based on python code), and a script provided by Akoya (also written in python). Please contact ZEISS for more details on how to implement the interface between a ZEISS Imaging System based on ZEN and the Akoya PhenoCycler.

For this application, ten rounds of hybridization and imaging were implemented, with subsequent stripping and washing in between rounds. Two blanks were acquired at the beginning and end of the session to acquire background controls. In each cycle, DAPI staining was included and was used as a reference and fiducial channel for the downstream image alignment, registration, and analysis. In each cycle, three other fluorescent markers were included (Figure 4), an Alexa Fluor 488 tagged marker, an Alexa Fluor 555 marker, and an Alexa Fluor 647 marker.

To image the brain slice, a z-stack was set up for a region of 4x9 tiles which corresponds to an area of 1100x497x12 µm in size. The ZEISS 63x Plan-Apochromat 1.4 N.A. oil immersion objective was used. This generates a pixel size of 70 nm. The z-step size for the z-stack was 0.3 µm. These dimensions were optimal to achieve the full resolution improvement possible utilizing the ZEN LSM Plus processing feature for confocal data. LSM Plus utilizes a Wiener filter-based deconvolution to improve the resolution of the confocal data by 1.4x over what is traditionally obtained using a 1 A.U. pinhole for acquisition. LSM Plus additionally improves the overall signal-to-noise ratio of the images, and therefore, no averaging was needed. After acquisition, the images were background corrected using the smoothed images acquired from the blank cycles.

In these complex multiplexing applications with multiple rounds of staining, it is often the case that only certain markers are needed in a given analysis set. For the analysis shown in the publication above and the example shown here, the following markers were combined into a final 3D multiplex data set:

Segmentation of synapses

For segmentation of the synaptic puncta, the four pre-synaptic markers were available (Piccolo, Syntaxin, Synaptobrevin, and Synaptophysin). In the absence of the post-synaptic marker or the pan-synaptic staining, the four channels with the pre-synaptic markers were averaged. For this, the intensities in each channel were measured separately on the original data. Initially, the images were normalized to 1-99% of their intensities. By normalizing the images to 1-99% of their intensities, the brightest and darkest 1% of pixels are saturated, while the remaining pixels are rescaled to fit within the remaining 98% of the intensity range. This helps to enhance the contrast between different structures and features in the image, making it easier to identify and segment specific structures such as synaptic puncta.

Then, the background was corrected by subtracting a blurred version of the image from the original data for each channel separately. The blur diameter was set to 100 pixels. Next, to enhance the puncta in the image, a top-hat filter (‘Preserve bright objects’ operator in ZEISS arivis Pro) with a disk structural element of 8 pixels was applied. Then, the data from the four pre-synaptic markers was averaged together and segmented using the Blob finder operator. This algorithm includes a combination of automatic seed finding based on structural information of an object map, and a watershed algorithm. This operation uses the Gaussian scale to find the object seeds and a watershed algorithm to identify object boundaries.

Finally, the segmented synapse objects were filtered by size with the lower limit of 5 voxels, corresponding to 70 nm³.

Allocation of cell and synapses to anatomical regions (process +summary table)

Once the segmentation for all components was complete, the final pipeline consisted of importing the anatomical regions, synapses, GFAP and MAP segments. Due to the very large number of synapses (1.7M), a new custom python operator was implemented, to randomly sub-sample 10 thousand synapses out of the total number of 1.7M synapses, detected in the image [The script and the application note is available here]. The subsetted synapses were allocated into the three anatomical structures (cortex, corpus callosum, and striatum), as well as striosomes.

In Figure 17, the number of cells, and volume for each anatomical region is summarized. Additionally, the total number of synapses and the percentage found in each region is also displayed.

Reduced dimensionality analysis

To visualize this very complex dataset, t-SNE (t-distributed Stochastic Neighbor Embedding) analysis was performed. This is a nonlinear dimensionality reduction technique used to visualize high-dimensional data in low-dimensional space. It achieves this by representing each data point as a probability distribution in the low-dimensional space and minimizing the difference between the pairwise similarities of the high-dimensional data and the pairwise similarities of the low-dimensional representation.

In this case study, dimensionality reduction analysis was conducted twice, once for the cells and once for the synapses. For the synapsis analysis, 28 features (XYZ coordinates, mean intensity levels, signal-to-noise ratios and integrated densities from six channels and puncta volume, their randomly subsampled neighborhood densities, and distances to GFAP and MAP2, and their anatomical regions [cortex, corpus callosum, striatum, striosomes]) for each synaptic puncta were used as the input to t-SNE. Only the 10K subsetted synapses were used for the expression profile analysis. For the cell analysis, 27 features were analyzed (XYZ coordinates, mean intensity levels, signal-to-noise ratios and integrated densities from six channels, their neighborhood densities, and distances to GFAP and MAP2, and their anatomical regions [cortex, corpus callosum, and striatum]).

To ensure that the visualization is robust to differences in scale and variability across the different features, the features were pre-processed by being log-transformed and normalized to have a standard deviation of one and minimum of zero before applying t-SNE. The logarithmic transformation reduces the effects of extreme values or outliers. This normalization is done to transform the data so that it has a standardized scale across different features, which will make their visualization more accurate and efficient. Standard deviation is a measure of how spread out the data are, with higher values indicating more variability. By setting the standard deviation to one, the data are rescaled so that each feature has a similar level of variability. Additionally, by setting the minimum value to zero, the data are shifted so that the lowest value for each feature is zero, which can simplify the interpretation of the data and make it easier to compare across different features. The t-SNE analysis was performed using scikit-learn 1.3.0 in Python 3.11.4 with the exact method, perplexity parameter equal to 40 and 5000 iterations.

In this case study, we have demonstrated example analyses for 3D multiplexing spatial omics data based on cellular and synaptic markers in the mouse brain.

These examples can be used to design analyses for 3D registration of multiplexing data, the segmentation of objects using AI, and the cell neighborhood and reduced dimensionality analysis. The registration algorithm aligns images from different experiments, while segmentation and allocation of components enable the analysis of their distribution across different anatomical regions. Cell neighborhood analysis provides insights in cellular organization, whereas reduced dimensionality analysis using t-SNE helps visualize high-dimensional data in low-dimensional space. These steps were performed using ZEISS arivis Pro and custom python scripts, resulting in a comprehensive analysis of the complex dataset.

In this experiment, the spatial distribution and organization of synapses, cells, and astrocytes in different brain regions was investigated. Various observations were made regarding the density and distribution of cells and synapses in the mouse brain. The expression of various presynaptic markers and cell-specific proteins were also evaluated, and the intricate expression patterns were discussed. These analyses serve as a valuable blueprint for users to follow with their own three-dimensional spatial omics data. The techniques used in this study can be applied across a wide range of applications in neuroscience, cancer research, immunology, and single-cell analysis. By exploring cell heterogeneity and clustering, characterizing cell-cell interactions, and identifying biomarkers, these type of analyses can provide a deeper understanding of complex biological systems and disease mechanisms.