Overview of MIKAIA’s Spatial Analysis Apps for Multiplexed Immunofluorescence Slides

A strong suit of MIKAIA^® is its wide portfolio of analysis apps for multiplexed immunofluorescence (mIF) / spatial proteomics. It can be used for both low-plex and high-plex panels and is compatbile with image formats produced by many microscopes, scanners, and instruments including Ome-tiff (e.g., Nanostring, Lunaphore, …), multiple single-channel TIFFs, Akoya/CODEX Qptiff, Zeiss CZI, Hamamatsu NDPIS, Leica SVS, Olympus VSI, …

MIKAIA^® can be used to analyze small field of views, entire whole-slide-images or TMAs (Programming-free Single Cell Analysis of CODEX TMA with MIKAIA), but also multiwell plates (High Content Imaging: Analyzing Ex-Vivo Drug Response Assays with MIKAIA). Also, batch analysis of entire datasets is possible without any coding.

This article demonstrates for a 20-plex CODEX whole-slide-scan of a tonsil MIKAIA’s cell segmentation, cell phenotyping, and spatial analysis capabilities. Each step is shown in a small subtitled video (30-90 seconds each) below.

AI Cell segmentation

Unbiased Phenotyping via Clustering

We outline an arbitrary region of interest (yellow annotation), select it, and then click the analyze-“RoI” button to kick off the analysis for this region with the FL Cell Analysis App. The app is configured to run nuclei segmentation on the DAPI channel. When a membrane marker is available, then this can alternatively be used. Following cell segmentation, the app measures each marker’s mean intensity for each cell. Here, phenotyping by k-means clustering is selected. Based on the marker intensities, each cell is assigned to a cluster (classified). The cluster explorer diagram (2D heatmap) visualizes the cluster-defining markers.

The interactive scatter plot supports t-SNE, UMAP, or PCA. Cells can be directly selected via the plot. By clicking on a point, the viewer will center in on the corresponding cell. This is convenient to inspect cells at the cluster boundaries or cluster centers.

Targeted Phenotyping

Here, cells are segmented in the same way as in the video above. Again, each marker’s intensity is measured per cell. Then, the marker expression (on or off) is determined per cell by checking if the intensity is above or below a threshold. The threshold (per marker) here is determined automatically at the end when the intensity for the whole set of cells is available. If you disagree with the automatically selected threshold, you can repeat the analysis and instead select a manual threshold.

The FL Cell Analysis app was configured to generate two kinds of annotations: “by marker” and “by-coexpression”.

• By marker: This group contains one class per marker, e.g., “CD3”, “CD8”, “PCK”, … The annotation classes contain all cells that were found to be positive for this marker (mean intensity > threshold). Since the same cell can express multiple markers, these annotation classes should be viewed one at a time.
• By coexpression: This group contains one class per cell type. This set of classes can be viewed all at the same time since each cell is in exactly one class. When no cell type map is provided, each encountered combination of markers is considered a cell type. The auto-generated name is composed of a deterministic prefix plus list of expressed markers, e.g., “ABBZ (CD3+, CD8+)”. Optionally, a cell type map can be configured where meaningful cell type names can be defined along with the identifying markers. Subtypes can be automatically created for functional markers (e.g., proliferation marker).

Spatial analysis

After cells have been detected and phenotyped, various types of spatial analyses can be conducted to reveal spatial patterns or investigate spatial relationships.

Cell-Cell-Connections and Cell Contacts

The Cell-Cell-Connections App uses the Delauney triangulation algorithm to connect each cell with its direct neighbor cells. The app can consider all cell types or only a subset. Attention: Make sure not to mix “by marker” and “by coexpression” cell classes. Cell-cell-connections are indicated by line annotations that are grouped into annotation classes defined by which two cell types are connected (e.g., ” CD3 | CD8″). Connections between neighboring cells of the same type can be included or excluded. The app then computes statistics for each cell type and cell-cell-connections type:

• Abundance: Absolute number of connections by type
• Relative abundance: Per cell type, what is the relative composition of neighboring cell types?
• Bystander analysis: On average, how many neighbors of type XYZ has a particular cell type?
• Distance: What is the average distance (and standard deviation) in µm between two neighboring cell types?

Cellular Neighborhoods

The Cellular Neighborhood App focuses in on each cell and then counts the cell types in its vicinity (neighborhood). The neighborhood can be defined by a radius (e.g., 50 µm), by the k-nearest neighbors, or by both constraints simultanously. Based on this analysis, a distance histogram diagram illustrates for a given cell type the average cell type composition in growing distances, e.g., at 0-10 µm distance, 10-20 µm distance, and so on.

Going further, this app can then feed the neighborhood compositions into a clustering algorithm that then identifies multiple types of “Cellular Neighborhoods” (CN). A 2d heatmap diagram illustrates for each CN the defining cell types, which should help figure out the biological type of that particular CN. It is then interesting to see if the presence or abundance of specific CNs is of prognostic interest or perhaps the neighborhood of CNs.

Proximity Analysis (Cell-to-interface or cell-to-cell)

The Proximity Analysis App can find the shortest path between two objects and then draw a line. A use case in oncology would be to measure distances from immune cells or tumor buds to the tumor invasive front. In the video above, the distance from cells to the germinal center outline is measured. When the source or target annotation is a polygon or path, then the shortest path to any point on that polygon is computed. The app can also be used to measure distances between two cell types or two sets of cell types (in the video: B cells to T cells).

Interactive Density Heatmap

The density heatmap is a feature available from the main toolbar, it is not an analysis app. A heatmap can be created that visualizes the presence of a particular annotation type or of a set of annotation types. Interactively, the heatmap’s smoothness, window size, and sensitivity can be configured. The annotations or image can be hidden and the background color can be switched from black to white or any other color in order to generate screenshots for a publication or report.

Grid Analysis (spatial heterogeneity)

Similar to the density heatmap feature, the Grid Analysis App overlays a virtual grid. Per tile in the grid, it computes a metric. Possible metrics are: abundance of cells, ratio of cells A over cells B. It can also operate on tissue masks instead of cells and measure the area or ratio of areas (e.g., tumor-stroma-ratio). Additionally, cutoffs can be provided for grading each tile based on the computed metric. The histogram over the graded tiles represents a good one-dimensional profile of the analyzed ROI that might be prognostically meaningful.

Grouping nearby cells into clusters

The Spatial Clustering App groups nearby cells into clusters. It starts by centering in on a cell of the target class and then checks if another cell is within a given user-defined radius, say 20 µm. If this is the case, these nearby cells form a new cluster. Recursively, the algorithm then centers on the new cells at the cluster border and again checks for further nearby cells. It continues for as long as nearby cells exist and so the cluster grows and grows. When no more nearby target cells are present, the cluster is finished and the cluster outline and area are determined. When the algorithm is done looking at all cells, a histogram over all clusters is computed. The histogram profile indicates whether a cell type forms few large clusters, many medium sized clusters or distributes accross the slide and does not form any clusters. Optionally, a cluster can be required to contain a minimum number of cells, and then the ratio of cells that are part of a cluster is another potentially meaningful attribute.