--- title: "Tools for IMC data analysis" date: "`r BiocStyle::doc_date()`" package: "`r BiocStyle::pkg_ver('imcRtools')`" author: - name: Nils Eling affiliation: - Department for Quantitative Biomedicine, University of Zurich - Institute for Molecular Health Sciences, ETH Zurich email: nils.eling@dqbm.uzh.ch - name: Lasse Meyer affiliation: - Department for Quantitative Biomedicine, University of Zurich - Institute for Molecular Health Sciences, ETH Zurich email: lasse.meyer@@uzh.ch - name: Daniel Schulz affiliation: - Department for Quantitative Biomedicine, University of Zurich - Institute for Molecular Health Sciences, ETH Zurich email: daniel.schulz@@uzh.ch output: BiocStyle::html_document: toc_float: yes bibliography: library.bib abstract: | This R package supports the handling and analysis of imaging mass cytometry and other highly multiplexed imaging data. The main functionality includes reading in single-cell data after image segmentation and measurement, data formatting to perform channel spillover correction and a number of spatial analysis approaches. First, cell-cell interactions are detected via spatial graph construction; these graphs can be visualized with cells representing nodes and interactions representing edges. Furthermore, per cell, its direct neighbours are summarized to allow spatial clustering. Per image/grouping level, interactions between types of cells are counted, averaged and compared against random permutations. In that way, types of cells that interact more (attraction) or less (avoidance) frequently than expected by chance are detected. vignette: | %\VignetteIndexEntry{"Tools for IMC data analysis"} %\VignetteEncoding{UTF-8} %\VignetteEngine{knitr::rmarkdown} editor_options: markdown: wrap: 72 chunk_output_type: console --- ```{r, echo=FALSE, results="hide"} knitr::opts_chunk$set(error=FALSE, warning=FALSE, message=FALSE, crop = NULL) library(BiocStyle) ``` ```{r library, echo=FALSE} library(imcRtools) ``` # Introduction This vignette gives an introduction to handling and analyzing imaging mass cytometry (IMC) and other highly-multiplexed imaging data in R. The `imcRtools` package relies on expression and morphological features extracted from multi-channel images using corresponding segmentation masks. A description of data types and segmentation approaches can be found below ([data types](#dataTypes), [segmentation](#segmentation)). However, due to shared data structures, the functionalities of the `imcRtools` package are applicable to most highly multiplexed imaging modalities. ## Overview The `imcRtools` package exports functions and example data to perform the following analyses: 1. [Read in pre-processed data](#read-in-data) 2. [Perform spillover correction for IMC data](#spillover) 3. [Build](#build-graph) and [visualize](#viz-cells) spatial graphs 4. [Aggregate across neighbouring cells](#aggregate-neighbors) for spatial clustering 5. [Detect spatial patches](#patch-detection) of similar cell-types 6. [Test the attraction or avoidance between celltypes](#test-neighborhood) To highlight the usability of these function, the `imcRtools` package also exports a number of [test data files](#test-data). ## Highly multiplexed imaging Highly multiplexed imaging techniques [@Bodenmiller2016] such as imaging mass cytometry (IMC) [@Giesen2014], multiplexed ion beam imaging (MIBI) [@Angelo2014] and cyclic immunofluorescence techniques [@Lin2018, @Gut2018] acquire read-outs of the expression of tens of protein in a spatially resolved manner. After technology-dependent data pre-processing, the raw data files are comparable: multi-channel images for the dimensions `x`, `y`, and `c`, where `x` and `y` define the number of pixels (`x * y`) per image and `c` the number of proteins (also refered to as "markers") measured per image. More info on the [data types](#dataTypes) and [further pre-processing](#segmentation) can be found below. Increased multiplexity compared to epitope-based techniques is achieved using single-cell resolved spatial transcriptomics techniques including MERFISH [@Chen2015] and seqFISH [@Lubeck2014]. However, data produced by these techniques required different pre-processing steps. Nevertheless, analysis performed on the single-cell level is equally applicable. ### Imaging mass cytometry IMC [@Giesen2014] is a highly multiplexed imaging approach to measure spatial protein abundance. In IMC, tissue sections on slides are stained with a mix of around 40 metal-conjugated antibodies prior to laser ablation with $1\mu{}m$ resolution. The ablated material is transferred to a mass cytometer for time-of-flight detection of the metal ions [@Giesen2014]. At an ablation frequency of 200Hz, a 1mm x 1mm region can be acquired within 1.5 hours. ### Data types {#dataTypes} In the case of IMC, the raw data output are .mcd files containing all acquired regions per slide. During image pre-prcoessing these files are converted into individual multi-channel .tiff and [OME-TIFF](https://docs.openmicroscopy.org/ome-model/5.6.3/ome-tiff/) files. These file formats are supported by most open-source and commercial image analysis software, such as [Fiji](https://imagej.net/software/fiji/), [QuPath](https://qupath.github.io/) or [napari](https://napari.org/). In R, these .tiff files can be read into individual [Image](https://bioconductor.org/packages/release/bioc/vignettes/EBImage/inst/doc/EBImage-introduction.html#3_Image_data_representation) objects or combined into a [CytoImageList](https://www.bioconductor.org/packages/release/bioc/vignettes/cytomapper/inst/doc/cytomapper.html#5_The_CytoImageList_object) object supported by the [cytomapper](https://www.bioconductor.org/packages/release/bioc/html/cytomapper.html) package. ### Segmentation and feature extraction {#segmentation} The pixel resolution of most highly multiplexed imaging technologies including IMC support the detection of cellular structures. Therefore, a common processing step using multi-channel images is object segmentation. In this vignette, objects are defined as cells; however, also larger scale structures could be segmented. There are multiple different image segmentation approaches available. However, `imcRtools` only supports direct reader functions for two segmentation strategies developed for highly multiplexed imaging technologies: 1. The [ImcSegmentationPipeline](https://github.com/BodenmillerGroup/ImcSegmentationPipeline) has been developed to give the user flexibility in how to perform channel selection and image segmentation. The underlying principle is to train a pixel classifier (using [ilastik](https://www.ilastik.org/)) on a number of selected channels to perform probabilistic classification of each pixel into: background, cytoplasm and nucleus. Based on these classification probabilities a [CellProfiler](https://cellprofiler.org/) pipeline performs cell segmentation and feature extraction. 2. A containerized version of this pipeline is implemented in [steinbock](https://github.com/BodenmillerGroup/steinbock). `steinbock` further supports segmentation via the use of `Mesmer` from the `DeepCell` library [@Greenwald2021]. While the output of both approaches is structured differently, the exported features are comparable: 1. per cell: channel intensity, morphology and location 2. cell-cell interactions exported as graph By combining these with availabel channel information, the data can be read into a `r BiocStyle::Biocpkg("SpatialExperiment")` or `r BiocStyle::Biocpkg("SingleCellExperiment")` object (see [below](#read-in-data)). # Example data {#test-data} The `imcRtools` package contains a number of example data generated by the Hyperion imaging system for different purposes. The following section gives an overview of these files. ## For spillover correction To highlight the use of the `imcRtools` package for spillover correction, we provide four .txt files containing pixel intensities of four spotted metals. Please refer to the [spillover correction](#spillover) section for full details. These files are accessible via: ```{r access-spillover-files} path <- system.file("extdata/spillover", package = "imcRtools") list.files(path, recursive = TRUE) ``` ## Raw data in form of .txt files IMC generates .mcd files storing the raw data or all acquired regions of interest (ROI). In addition, the raw pixel values are also stored in individual .txt files per ROI. To highlight reading in raw data in form of .txt files, the `imcRtools` contains 3 sample acquisitions: ```{r access-txt-files} txt_files <- list.files(system.file("extdata/mockData/raw", package = "imcRtools")) txt_files ``` ## ImcSegmentationPipeline output data IMC data preprocessing and segmentation can be performed using the [ImcSegmentationPipeline](https://github.com/BodenmillerGroup/ImcSegmentationPipeline). It generates a number of `.csv` files containing object/cell-specific and image-specific metadata. The `imcRtools` package exports the `read_cpout` function as convenient reader function for outputs generated by the `ImcSegmentationPipeline`. For demonstration purposes, `imcRtools` contains the output of running the pipeline on a small example dataset: ```{r imcsegmentationpipeline-data} path <- system.file("extdata/mockData/cpout", package = "imcRtools") list.files(path, recursive = TRUE) ``` ## steinbock output data The [steinbock](https://github.com/BodenmillerGroup/steinbock) pipeline can be used to process, segment and extract features from IMC data. For more information, please refer to its [documentation](https://bodenmillergroup.github.io/steinbock/). To highlight the functionality of `imcRtools` to read in single-cell data generated by `steinbock`, we provide a small toy dataset available at: ```{r steinbock-data} path <- system.file("extdata/mockData/steinbock", package = "imcRtools") list.files(path, recursive = TRUE) ``` The example data related to the `ImcSegmentationPipeline` and `steinbock` can also be accessed [online](https://github.com/BodenmillerGroup/TestData/tree/main/datasets/210308_ImcTestData). # Read in IMC data {#read-in-data} The `imcRtools` package supports reading in data generated by the [ImcSegmentationPipeline](https://github.com/BodenmillerGroup/ImcSegmentationPipeline) or [steinbock](https://github.com/BodenmillerGroup/steinbock) pipeline. To read in the outpout data into a `r BiocStyle::Biocpkg("SpatialExperiment")` or `r BiocStyle::Biocpkg("SingleCellExperiment")`, the `imcRtools` package exports the `read_cpout` function. By default, the single-cell data is read into a `r BiocStyle::Biocpkg("SpatialExperiment")` object. Here, the extracted channel- and cell-specific intensities are stored in the `counts(spe)` slot. All morphological features are stored in `colData(spe)` and the spatial locations of the cells are stored in `spatialCoords(spe)`. The interaction graph is stored in `colPair(spe, "neighbourhood")`. Alternatively, the data can be read into a `r BiocStyle::Biocpkg("SingleCellExperiment")` object. The only difference is the lack of `spatialCoords(sce)`. Here, the spatial coordinates are stored in `colData(spe)$Pos_X` and `colData(spe)$Pos_Y`. ## Read in CellProfiler output The [ImcSegmentationPipeline](https://github.com/BodenmillerGroup/ImcSegmentationPipeline) produces a number of output files. By default, all single-cell features are measured and exported. To browse and select the features of interest, the `imcRtools` package provides the `show_cpout_features` function: ```{r show_cpout_features} path <- system.file("extdata/mockData/cpout", package = "imcRtools") show_cpout_features(path) ``` By default, `read_cpout` will read in the mean intensity per channel and cell from "hot pixel" filtered image stacks specified via `intensities = "Intensity_MeanIntensity_FullStackFiltered"`. Please refer to `?read_cpout` for the full documentation. ```{r read_cpout} cur_path <- system.file("extdata/mockData/cpout", package = "imcRtools") # Read as SpatialExperiment (spe <- read_cpout(cur_path, graph_file = "Object_relationships.csv")) # Read as SingleCellExperiment (sce <- read_cpout(cur_path, graph_file = "Object_relationships.csv", return_as = "sce")) ``` ## Read in steinbock output Single-cell data and all associated metadata (e.g. spatial location, morphology and interaction graphs) as produced by the [steinbock](https://github.com/BodenmillerGroup/steinbock) pipeline can be read in using the `read_steinbock` function: ```{r read_steinbock} cur_path <- system.file("extdata/mockData/steinbock", package = "imcRtools") # Read as SpatialExperiment (spe <- read_steinbock(cur_path)) # Read as SingleCellExperiment (sce <- read_steinbock(cur_path, return_as = "sce")) ``` For more information, please refer to `?read_steinbock`. ## Read raw .txt files into Image objects For reading in and visualization of multi-channel images and segmentation masks, please refer to the `r BiocStyle::Biocpkg("cytomapper")` package. The `imcRtools` package however supports reading in raw .txt files generated by the Hyperion imaging system into a `CytoImageList` object; a data container exported by `cytomapper`. The user needs to provide a path from which all .txt files will be read in: ```{r read-txt-1} path <- system.file("extdata/mockData/raw", package = "imcRtools") cur_CytoImageList <- readImagefromTXT(path) cur_CytoImageList ``` By specifying the `pattern` argument, individual or a subset of files can be read in. For more information, please refer to `?readImagefromTXT`. # Spillover correction {#spillover} When acquiring IMC images, pixel intensities can be influenced by spillover from neighboring channels. To correct for this, Chevrier *et al.* have developed a staining protocol to acquire individually spotted metal isotopes [@Chevrier2017]. Based on these measurements, spillover into neighboring channels can be quantified to correct pixel intensities. The `imcRtools` package provides helper functions that facilitate the correction of spillover for IMC data. For a full tutorial, please refer to the [IMC data analysis book](https://bodenmillergroup.github.io/IMCDataAnalysis/spillover-correction.html). ## Read in the single-spot acquisitions In the first step, the pixel intensities of individually spotted metals need to be read into a `SingleCellExperiment` container for downstream use with the `r BiocStyle::Biocpkg("CATALYST")` package. For this, the `readSCEfromTXT` function can be used: ```{r read-single-metals} path <- system.file("extdata/spillover", package = "imcRtools") sce <- readSCEfromTXT(path) sce ``` Here, the example metal spot files are read in. The spot information are stored in the `colData(sce)` slot and channel information are stored in `rowData(sce)`. Each column represents a single pixel. ## Quality control on single-spot acquisitions In the next step, it is crucial to identify potentially mislabeled spots or spots with low pixel intensities. The `imcRtools` package exports the `plotSpotHeatmap` function, which visualizes the aggregated (default `median`) pixel intensities per spot and per metal: ```{r plotSpotHeatmap, fig.width=5, fig.height=5} plotSpotHeatmap(sce) ``` Here, high median pixel intensities can be observed in each spot and their corresponding channels (visualized on the `log10` scale by default). To quickly identify spot/channel combinations with low signal, the `threshold` parameter can be set: ```{r plotSpotHeatmap-2, fig.width=5, fig.height=5} plotSpotHeatmap(sce, log = FALSE, threshold = 200) ``` ## Consecutive pixel binning If pixel intensities are low, spillover estimation might not be robust. Therefore, the `binAcrossPixels` function can be used to sum consecutive pixels and enhance the acquired signal. This step is optional for spillover estimation. ```{r pixel-binning, fig.width=5, fig.height=5 } sce2 <- binAcrossPixels(sce, bin_size = 5) plotSpotHeatmap(sce2, log = FALSE, threshold = 200) ``` ## Pixel filtering Prior to spillover estimation, the `r BiocStyle::Biocpkg("CATALYST")` package provides the `assignPrelim`, `estCutoffs` and `applyCutoffs` functions to estimate the spotted mass for each pixel based on their channel intensities. For more information on the spillover estimation and correction, please refer to the [CATALYST vignette](https://bioconductor.org/packages/release/bioc/vignettes/CATALYST/inst/doc/preprocessing.html#compensation). This estimation can be used to identify pixels that cannot be easily assigned to their spotted mass, potentially indicating pixels with weak signal. To remove these pixels, the `filterPixels` function can be used. This function further removes pixels assigned to masses, which only contain very few pixels. ```{r assign-pixels} library(CATALYST) bc_key <- as.numeric(unique(sce$sample_mass)) assay(sce, "exprs") <- asinh(counts(sce)/5) sce <- assignPrelim(sce, bc_key = bc_key) sce <- estCutoffs(sce) sce <- applyCutoffs(sce) # Filter out mislabeled pixels sce <- filterPixels(sce) table(sce$bc_id, sce$sample_mass) ``` ## Estimating the spillover matrix Finally, the pre-processed `SiingleCellExperiment` object can be used to generate the spillover matrix using the `CATALYST::computeSpillmat` function: ```{r estimate-spillover} sce <- computeSpillmat(sce) metadata(sce)$spillover_matrix ``` This spillover matrix can be directly applied to compensate the summarized pixel intensities per cell and per channel as described [here](https://bioconductor.org/packages/release/bioc/vignettes/CATALYST/inst/doc/preprocessing.html#compcytof-compensation-of-mass-cytometry-data). # Spatial analysis The following section will highlight functions for spatial analyses of the data. ## Constructing graphs {#build-graph} When following the `ImcSegmentationPipeline` or `steinbock` and reading in the data using the corresponding functions, the generated graphs are automatically stored in the `colPair(spe, "neighborhood")` slot. Alternatively, the `buildSpatialGraph` function in the `imcRtools` package constructs interaction graphs using either (i) cell-centroid expansion, (ii) k-nearest neighbor search or (iii) delaunay triangulation. ```{r buildSpatialGraph, message=FALSE} library(cytomapper) data("pancreasSCE") pancreasSCE <- buildSpatialGraph(pancreasSCE, img_id = "ImageNb", type = "expansion", threshold = 20) pancreasSCE <- buildSpatialGraph(pancreasSCE, img_id = "ImageNb", type = "knn", k = 5) pancreasSCE <- buildSpatialGraph(pancreasSCE, img_id = "ImageNb", type = "delaunay") colPairNames(pancreasSCE) ``` When setting `type = "knn"`, by default a directional graph will be build. Setting `directed = FALSE` will create bi-directional edges for each pair of cells that are connected by at least one edge in the directed setting. ## Graph/cell visualization {#viz-cells} The cells' locations and constructed graphs can be visualized using the `plotSpatial` function. Here, cells are referred to as "nodes" and cell-cell interactions are referred to as "edges". All visual attributes of the nodes and edges can be set. Either by specifying a variable in `colData(spe)`, a marker name or a single entry using the `*_fix` parameters. By default the `plotSpatial` function will visualize equal physical units on the x- and y-axis with an aspect ratio of 1. The example data are located in different regions of an image and we therefore set `scales = "free"` for simpler visualization. ```{r plotSpatial} library(ggplot2) library(ggraph) plotSpatial(pancreasSCE, img_id = "ImageNb", node_color_by = "CellType", node_shape_by = "ImageNb", node_size_by = "Area", draw_edges = TRUE, colPairName = "knn_interaction_graph", directed = FALSE, scales = "free") # Colored by expression and with arrows plotSpatial(pancreasSCE, img_id = "ImageNb", node_color_by = "PIN", assay_type = "exprs", node_size_fix = 3, edge_width_fix = 0.2, draw_edges = TRUE, colPairName = "knn_interaction_graph", directed = TRUE, arrow = grid::arrow(length = grid::unit(0.1, "inch")), end_cap = ggraph::circle(0.05, "cm"), scales = "free") # Subsetting the SingleCellExperiment plotSpatial(pancreasSCE[,pancreasSCE$Pattern], img_id = "ImageNb", node_color_by = "CellType", node_size_fix = 1, draw_edges = TRUE, colPairName = "knn_interaction_graph", directed = TRUE) ``` The returned object can be further modified using the `ggplot2` logic. This includes changing the node color, shape and size using `scale_color_*`, `scale_shape_*` and `scale_size_*`. Edge attributes can be altered using the `scale_edge_*` function exported by `ggraph`, ## Neighborhood aggregation {#aggregate-neighbors} The `aggregateNeighbors` function can be used to aggregate features of all neighboring cells for each individual cell. This function operates in two settings. 1. `metadata`: when aggregating by cell-specific metadata, the function computes the relative frequencies of all entries to `colData(sce)[[count_by]]` within the direct neighborhood of each cell.\ 2. `expression`: the expression counts of neighboring cells are aggregated using the specified `statistic` (defaults to `mean`). Each cell's neighborhood is defined as endpoints of edges stored in `colPair(sce, colPairName)`. ```{r aggregateNeigbors} pancreasSCE <- aggregateNeighbors(pancreasSCE, colPairName = "knn_interaction_graph", aggregate_by = "metadata", count_by = "CellType") head(pancreasSCE$aggregatedNeighbors) pancreasSCE <- aggregateNeighbors(pancreasSCE, colPairName = "knn_interaction_graph", aggregate_by = "expression", assay_type = "exprs") head(pancreasSCE$mean_aggregatedExpression) ``` The returned entries can now be used for clustering to group cells based on their environment (either by aggregated categorical features or expression). ```{r aggregateNeigbors-clustering} set.seed(22) cur_cluster <- kmeans(pancreasSCE$aggregatedNeighbors, centers = 3) pancreasSCE$clustered_neighbors <- factor(cur_cluster$cluster) # Visualize CellType and clustered_neighbors plotSpatial(pancreasSCE, img_id = "ImageNb", node_color_by = "CellType", node_size_fix = 4, edge_width_fix = 2, edge_color_by = "clustered_neighbors", draw_edges = TRUE, colPairName = "knn_interaction_graph", directed = FALSE, nodes_first = FALSE, scales = "free") + scale_color_brewer(palette = "Set2") + scale_edge_color_brewer(palette = "Set1") # Visualize clustered_neighbors plotSpatial(pancreasSCE, img_id = "ImageNb", node_color_by = "clustered_neighbors", node_size_fix = 4, edge_width_fix = 1, draw_edges = TRUE, colPairName = "knn_interaction_graph", directed = FALSE, nodes_first = FALSE, scales = "free")+ scale_color_brewer(palette = "Set1") ``` ## Spatial context analysis {#spatial-context} The single cell assignments derived from clustering cells based on their environment can be interpreted as **cellular neighborhoods (CNs)**, which can represent sites of unique local processes [@Schurch2020]. Downstream of CNs, imcRtools exports three functions to detect and analyze the **spatial context (SC)** of each cell. 1. `detectSpatialContext`: for the function to detect SCs 2. `filterSpatialContext`: for the function to filter SCs 3. `plotSpatialContext`: for the function to plot SC graphs The term SC was coined by Bhate and colleagues [@Bhate2022] and describes tissue regions in which distinct CNs may be interacting, which can lead to specialized local biological events. The `detectSpatialContext` function relies on CN fractions for each cell in a spatial interaction graph (originally a k-nearest neighbor (KNN) graph). We can retrieve the CN fractions using the above-described `buildSpatialGraph` and `aggregateNeighbors` functions. The window size (k for KNN) for `buildSpatialGraph` should reflect a length scale on which biological signals can be exchanged and depends, among others, on cell density and tissue area. In view of their divergent functionality, we recommend to use a larger window size for SC (interaction between local processes) than for CN (local processes) detection. Subsequently, the CN fractions are sorted from high-to-low and the SC of each cell is assigned the minimal combination of SCs that additively surpass a user-defined threshold. The default threshold of 0.9 aims to represent the dominant CNs, hence the most prevalent signals, in a given window. For more details, please refer to [@Bhate2022]. ```{r detectSpatialContext} # Generate k-nearest neighbor graph pancreasSCE <- buildSpatialGraph(pancreasSCE, img_id = "ImageNb", type = "knn", name = "knn_spatialcontext_graph", k = 15) colPairNames(pancreasSCE) # Aggregate based on clustered_neighbors pancreasSCE <- aggregateNeighbors(pancreasSCE, colPairName = "knn_spatialcontext_graph", aggregate_by = "metadata", count_by = "clustered_neighbors", name = "aggregatedNeighborhood") # Detect spatial contexts pancreasSCE <- detectSpatialContext(pancreasSCE, entry = "aggregatedNeighborhood", threshold = 0.9, name = "spatial_context") # Define SC color scheme col_SC <- setNames(c("#A6CEE3", "#1F78B4", "#B2DF8A", "#33A02C", "#FB9A99", "#E31A1C", "#FDBF6F"), sort(unique(pancreasSCE$spatial_context))) # Visualize spatial contexts on images plotSpatial(pancreasSCE, img_id = "ImageNb", node_color_by = "spatial_context", node_size_fix = 4, edge_width_fix = 1, draw_edges = TRUE, colPairName = "knn_spatialcontext_graph", directed = FALSE, nodes_first = FALSE, scales = "free") + scale_color_manual(values = col_SC) ``` After SC assignment for each individual cell, the `filterSpatialContext` function allows to filter detected SCs based on user-defined thresholds for number of group entries (usually image or patient ID) and/or total number of cells per SC. In addition to a new column entry to the `colData(object)`, the function also returns a `data.frame` entry to `metadata(object)` containing filtered group and cell counts per SC. ```{r filterSpatialContext} # Filter spatial contexts # By number of group entries pancreasSCE <- filterSpatialContext(pancreasSCE, entry = "spatial_context", group_by = "ImageNb", group_threshold = 2, name = "spatial_context_filtered_group") metadata(pancreasSCE)$filterSpatialContext plotSpatial(pancreasSCE, img_id = "ImageNb", node_color_by = "spatial_context_filtered_group", node_size_fix = 4, edge_width_fix = 1, draw_edges = TRUE, colPairName = "knn_spatialcontext_graph", directed = FALSE, nodes_first = FALSE, scales = "free") + scale_color_manual(values = col_SC) # By total number of cells pancreasSCE <- filterSpatialContext(pancreasSCE, entry = "spatial_context", group_by = "ImageNb", cells_threshold = 15, name = "spatial_context_filtered_cells") metadata(pancreasSCE)$filterSpatialContext plotSpatial(pancreasSCE, img_id = "ImageNb", node_color_by = "spatial_context_filtered_cells", node_size_fix = 4, edge_width_fix = 1, draw_edges = TRUE, colPairName = "knn_spatialcontext_graph", directed = FALSE, nodes_first = FALSE, scales = "free") + scale_color_manual(values = col_SC) ``` Lastly, the `plotSpatialContext` function plots directed *SC graphs*, akin to *CN combination maps* in [@Bhate2022], based on symbolic edge-lists and vertex metadata, which operates on cohort-level. ```{r plotSpatialContext} ## Plot spatial context graph # Default plotSpatialContext(pancreasSCE, entry = "spatial_context", group_by = "ImageNb") # Colored by name and size by n_cells plotSpatialContext(pancreasSCE, entry = "spatial_context", group_by = "ImageNb", node_color_by = "name", node_size_by = "n_cells", node_label_color_by = "name") # Colored by n_cells and size by n_group plotSpatialContext(pancreasSCE, entry = "spatial_context", group_by = "ImageNb", node_color_by = "n_cells", node_size_by = "n_group", node_label_color_by = "n_cells")+ scale_color_viridis() ``` The returned object can be further modified using the `ggplot2` logic. This includes changing the node color and size using `scale_color_*` and `scale_size_*`. Edge attributes can be altered using the `scale_edge_*` function exported by `ggraph`. Furthermore, setting `return_data = TRUE` returns the symbolic edge-list and vertex metadata used for graph construction in a `list` of two `data.frames`. ```{r plotSpatialContext - return data} # Return data plotSpatialContext(pancreasSCE, entry = "spatial_context", group_by = "ImageNb", return_data = TRUE) ``` ## Community detection In addition to cellular neighborhood and spatial context analysis, `imcRtools` exports the `detectCommunity` function to detect the spatial community of each cell as proposed by [@Jackson2020]. Here, each cell is clustered based on its interactions as defined by a spatial object graph. In more detail, the spatial community detection procedure is as follows: 1. Create an igraph object from the edge list stored in \code{colPair(object, colPairName)}. 2. Perform community detection using the specified \code{cluster_fun} algorithm (defaults to "louvain"). 3. Store the community IDs in a vector and replace all communities with a size smaller than \code{size_threshold} by NA. ```{r detectCommunity} ## Detect spatial community set.seed(22) pancreasSCE <- detectCommunity(pancreasSCE, colPairName = "expansion_interaction_graph") plotSpatial(pancreasSCE, img_id = "ImageNb", node_color_by = "spatial_community", node_size_fix = 4, edge_width_fix = 1, draw_edges = TRUE, colPairName = "expansion_interaction_graph", directed = FALSE, nodes_first = FALSE, scales = "free") ## Detect spatial community - specify size_threshold set.seed(22) pancreasSCE <- detectCommunity(pancreasSCE, colPairName = "expansion_interaction_graph", size_threshold = 20) plotSpatial(pancreasSCE, img_id = "ImageNb", node_color_by = "spatial_community", node_size_fix = 4, edge_width_fix = 1, draw_edges = TRUE, colPairName = "expansion_interaction_graph", directed = FALSE, nodes_first = FALSE, scales = "free") ``` It is also possible to use different community detection algorithms from the `igraph`package. ```{r detectCommunity-walktrap} ## Detect spatial community - walktrap community detection set.seed(22) pancreasSCE <- detectCommunity(pancreasSCE, colPairName = "expansion_interaction_graph", cluster_fun = "walktrap") plotSpatial(pancreasSCE, img_id = "ImageNb", node_color_by = "spatial_community", node_size_fix = 4, edge_width_fix = 1, draw_edges = TRUE, colPairName = "expansion_interaction_graph", directed = FALSE, nodes_first = FALSE, scales = "free") ``` Moreover, the user can specify \code{group_by} to perform spatial community detection separately for all unique entries to \code{colData(object)[,group_by]} e.g. for tumor and non-tumor cells. ```{r detectCommunity-bygroup} ## Detect spatial community - specify group_by pancreasSCE <- detectCommunity(pancreasSCE, colPairName = "expansion_interaction_graph", group_by = "CellType", size_threshold = 10, BPPARAM = BiocParallel::SerialParam(RNGseed = 22)) plotSpatial(pancreasSCE, img_id = "ImageNb", node_color_by = "spatial_community", node_size_fix = 4, edge_width_fix = 1, draw_edges = TRUE, colPairName = "expansion_interaction_graph", directed = FALSE, nodes_first = FALSE, scales = "free") ``` ## Border cells To exclude cells that are close to the image border, the `imcRtools` package exports the `findBorderCells` function. ```{r findBorderCells} pancreasSCE <- findBorderCells(pancreasSCE, img_id = "ImageNb", border_dist = 10) plotSpatial(pancreasSCE[,!pancreasSCE$border_cells], img_id = "ImageNb", node_color_by = "CellType", node_size_fix = 4, edge_width_fix = 2, edge_color_by = "clustered_neighbors", draw_edges = TRUE, colPairName = "knn_interaction_graph", directed = FALSE, nodes_first = FALSE, scales = "free") + scale_color_brewer(palette = "Set2") + scale_edge_color_brewer(palette = "Set1") ``` Excluding border cells can be useful when incorrectly connected cells are observed at image borders. ## Patch detection {#patch-detection} An alternative and supervised way of detecting regions with similar types of cells is available via the `patchDetection` function. Here, the user defines which cells should be used for patch detection via the `patch_cells` parameter. A patch is defined as a set of cells as defined by `patch_cells` which are weakly connected in the graph in `colPair(object, colPairname)`. Below, the `patchDetection` function is demonstrated using the previously computed `expansion` graph and defining cells of `celltype_B` as the cells of interest. Here, the function additionally draws a concave hull around the detected patch, expands the hull by 20$\mu{}m$ and defines all cells within this expanded hulls as patch cells. ```{r patchDetection} pancreasSCE <- patchDetection(pancreasSCE, patch_cells = pancreasSCE$CellType == "celltype_B", colPairName = "expansion_interaction_graph", expand_by = 20, img_id = "ImageNb") plotSpatial(pancreasSCE, img_id = "ImageNb", node_color_by = "patch_id", scales = "free") ``` Patches that only consist of 1 or 2 cells cannot be expanded. For each patch larger than 2 cells, the spatial area can be computed using the `patchSize` function: ```{r patchSize} patchSize(pancreasSCE) ``` ## Distance statistics to cells of interest To calculate the min, max, mean or median distance for each cell to a given cell type or class of cells of interest, the function `distToCells` is available. Cells of interest are defined via the `x_cells` parameter as `logical` and distances to for all cells to those cells will be reported in a new column in the `colData` of the `SingleCellExperiment`. If the cells of interest form patches (many cells of the same type next to each other) or similarly if a patch detection has previously been performed the positive distances reflect the distances from cells outside of patches (or non cells of interest) to the patch cells (or cells of interest) and the negative distances reflect the distances from cells inside the patches to the cells outside the patches. For `statistics` = "min" this corresponds to the distance from the cells outside to the closest point on the patch border and from the cells inside to the closest point on the patch border. If `return_neg` is set to `FALSE`, distances from the cells of interest (or cells inside the patches) to non cells of interest are computed as positive distances. ```{r distToCells} pancreasSCE <- distToCells(pancreasSCE, x_cells = pancreasSCE$CellType == "celltype_B", coords = c("Pos_X","Pos_Y"), statistics = "min", name = "minDist", img_id = "ImageNb") pancreasSCE <- distToCells(pancreasSCE, x_cells = pancreasSCE$CellType == "celltype_B", coords = c("Pos_X","Pos_Y"), statistics = "mean", name = "meanDist", img_id = "ImageNb") plotSpatial(pancreasSCE, img_id = "ImageNb", node_color_by = "minDist", scales = "free") + scale_color_viridis() plotSpatial(pancreasSCE, img_id = "ImageNb", node_color_by = "meanDist", scales = "free") + scale_color_viridis() ``` ## Neighborhood permutation testing {#test-neighborhood} The following section describes how to observe and test the average number of interactions between cell labels (e.g. cell-types) within grouping levels (e.g. images). For full descriptions of the testing approaches, please refer to [Shapiro et al., Nature Methods](https://www.nature.com/articles/nmeth.4391) [@Shapiro2017] and [Schulz et al., Cell Systems](https://www.sciencedirect.com/science/article/pii/S2405471217305434) [@Schulz2018] The `imcRtools` package exports the `countInteractions` and `testInteractions` functions, which summarize all cell-cell interactions per grouping level (e.g. image). As a result, a table is returned where each row represents one of all possible cell-type/cell-type interactions among all grouping levels. Missing entries or `NA`s indicate missing cell-type labels for this grouping level. The next section gives details on how interactions are summarized. ### Summarizing interactions The `countInteractions` function counts the number of edges (interactions) between each set of unique cell labels per grouping level. Simplified, it counts for each cell of type A the number of neighbors of type B. This count is averaged within each unique grouping level (e.g. image) in four different ways: 1. `method = "classic"`: The count is divided by the total number of cells of type A. The final count can be interpreted as "How many neighbors of type B does a cell of type A have on average?" 2. `method = "histocat"`: The count is divided by the number of cells of type A that have at least one neighbor of type B. The final count can be interpreted as "How many many neighbors of type B has a cell of type A on average, given it has at least one neighbor of type B?" 3. `method = "patch"`: For each cell, the count is binarized to 0 (less than `patch_size` neighbors of type B) or 1 (more or equal to `patch_size` neighbors of type B). The binarized counts are averaged across all cells of type A. The final count can be interpreted as "What fraction of cells of type A have at least a given number of neighbors of type B?" 4. `method = "interaction"`: The count is divided by the total number of interactions from cell type A. The final count can be interpreted as the fraction of interactions of cell type A that occur with cell type B. The `countInteractions` returns a `DataFrame` containing the summarized counts (`ct`) for all combinations of `from_label`, `to_label` and `group_by`. ```{r countInteractions} out <- countInteractions(pancreasSCE, group_by = "ImageNb", label = "CellType", method = "classic", colPairName = "knn_interaction_graph") out ``` ### Testing for significance In the next instance, one can test if the obtained count is larger or smaller compared to what is expected from a random distribution of cell labels. For this, the `testInteractions` function permutes the cell labels `iter` times and counts interactions as described above. This approach generates a distribution of the interaction count under a random distribution of cell labels. The observed interaction count is compared against this Null distribution to derive empirical p-values: `p_gt`: fraction of perturbations equal or greater than the observed count `p_lt`: fraction of perturbations equal or less than the observed count Based on these empirical p-values, the `interaction` score (attraction or avoidance), overall `p` value and significance by comparison to `p_treshold` (`sig` and `sigval`) are derived. All results are returned in form of a `DataFrame`. We set the seed within the `SerialParam` (or `MulticoreParam`) function for reproducibility. ```{r testInteractions} out <- testInteractions(pancreasSCE, group_by = "ImageNb", label = "CellType", method = "classic", colPairName = "knn_interaction_graph", BPPARAM = BiocParallel::SerialParam(RNGseed = 123)) out ``` ### Visualizing interactions The `plotInteractions` function is used to visualize interaction networks from an edge list, typically the output of the `countInteractions`and `testInteractions` functions. ```{r plotInteractions} plotInteractions(out, pancreasSCE, group_by = "ImageNb", label = "CellType") ``` Before visualization, the edge list can be filtered to focus on relevant cell-cell interactions, groupings, and significance. In the example below we retain interactions whose fraction of significant pairs (based on sigval) exceeds 75% and tag the edges as attraction or repulsion. ```{r plotInteractions-filter data} library(dplyr) out <- out %>% as.data.frame() %>% mutate(pair_id = paste(from_label, to_label, sep = "__"), is_sig = !is.na(sig)) summary_df <- out %>% group_by(pair_id) %>% summarise( sum_sigval = sum(sigval, na.rm = TRUE), total_count = sum(is_sig), .groups = "drop" ) %>% mutate( fraction_sig = sum_sigval / total_count, int_type = case_when( fraction_sig > 0.75 ~ "attraction", fraction_sig < -0.75 ~ "repulsion", TRUE ~ NA_character_ ) ) %>% filter(!is.na(int_type)) subset_out <- out %>% semi_join(summary_df, by = "pair_id") %>% left_join(summary_df %>% select(pair_id, int_type), by = "pair_id") # Now we can visualise only these filtered edges plotInteractions(subset_out, pancreasSCE, group_by = "ImageNb", label = "CellType") ``` You can customize node and edge aesthetics directly using `plotInteractions` arguments. The plot can also be modified further using `ggplot2` syntax. Node properties such as color, shape, and size can be controlled using `scale_color_*`, `scale_shape_*` and `scale_size_*`. Edge aesthetics can be modified using `scale_edge_*` functions from `ggraph`. To color edges based on subsetted significant interaction type (attraction, repulsion), specify the `int_type` column via `edge_color_by`, and use `scale_edge_colour_manual()` for custom colors. You can also control edge thickness via `edge_width_by`. For example, "ct" represents the summarized interaction count derived from `countInteractions` or `testInteractions`, averaged per cell-pair and grouping level. In our example, the "ct" value refers to the average neighboring cells of cell type B from cell type A averaged across all images. ```{r plotInteractions-edge aesthetics} library(ggraph) edge_color <- setNames(c("blue", "red"),c("attraction", "repulsion")) plotInteractions(subset_out, pancreasSCE, group_by = "ImageNb", label = "CellType", edge_color_by = "int_type", edge_width_by = "ct") + scale_edge_colour_manual(values = edge_color) ``` Control node properties via: `node_color_by`, `node_size_by`, `node_label_color_by`. ```{r plotInteractions-node aesthetics} plotInteractions(out, pancreasSCE, group_by = "ImageNb", label = "CellType", node_color_by = "name", node_label_color_by = "name", node_size_by = "n_cells") # Define custom colors for node and node_label node_color <- setNames(c("red", "blue", "green"), c("celltype_A", "celltype_B", "celltype_C")) plotInteractions(out, pancreasSCE, group_by = "ImageNb", label = "CellType", node_color_by = "name", node_label_color_by = "name", node_size_by = "n_cells")+ scale_color_manual(values = node_color) ``` The layout of the graph can be customized using the `graph_layout`argument. For more information see \link[ggraph]{ggraph}. ```{r plotInteractions-layout} # Default is "circle" plotInteractions(out, pancreasSCE, group_by = "ImageNb", label = "CellType") # Chord diagram layout plotInteractions(out, pancreasSCE, group_by = "ImageNb", label = "CellType", graph_layout = "chord") # Linear layout plotInteractions(out, pancreasSCE, group_by = "ImageNb", label = "CellType", graph_layout = "linear") ``` Furthermore, setting `return_data = TRUE` returns the symbolic edge-list, along with any specified edge aesthetic parameters and vertex metadata used for graph construction, as a `list` of two `data.frames`. ```{r plotInteractions-return data} # return data plotInteractions(out, pancreasSCE, "CellType", "ImageNb", return_data = TRUE) ``` # Contributions Large chunks of the code of the `imcRtools` package is based on the work of [Vito Zanotelli](https://github.com/votti). Vito has co-developed the spillover correction approach and translated the interaction testing approach developed by [Denis Shapiro](https://github.com/DenisSch) and [Jana Fischer](https://github.com/JanaFischer) into R (formerly known as the [neighbouRhood](https://github.com/BodenmillerGroup/neighbouRhood) R package). Jana has furthermore added the "patch" method for interaction counting and testing. [Tobias Hoch](https://github.com/toobiwankenobi) has written the first version of reading in the `ImcSegmentationPipeline` output and the `patchDetection` function. [Daniel Schulz](https://github.com/SchulzDan) has build the `aggregateNeighbors` and `minDistToCells` functions and contributed to developing the spatial clustering approach. [Lasse Meyer](https://github.com/lassedochreden) has implemented the functions for spatial context analysis and developed the `detectCommunity` function. [Marlene Lutz](https://github.com/marlene1) has contributed the "interaction" method for interaction counting and testing, and implemented the `plotInteractions` function. # Session info {.unnumbered} ```{r sessionInfo, echo=FALSE} sessionInfo() ``` # References