2.1 Pre-processing

In this step, one or multiple selected FCS files will be loaded into the R environment using the read.FCS function of flowCore package. Expression values of selected markers are then extracted and transformed with a proper transformation method (including cytofAsinh (suggest for CyTOF data), autoLgcl (suggest for FCM data), logicle and arcsinh), expression matrixes from multiple files are then combined into a single matrix. For multiple FCS files, cytofkit provides four data combining methods including i) ceil which samples up to a user specified number of cells without replacement from each FCS file then combine, ii) all which takes all cells from each FCS file then combine, iii) min which samples a minimum number of cells among all FCS files from each FCS file then combine and iv) fixed which samples an user specified number of cells (with replacement when the total number of cell in the file is less than the specified number) from each FCS file then combine. After combination, the whole dataset will be stored in a data matrix, with row names represented as filename_cellID, column mames represented as marker_name. This step is implemented in funciton cytof_exprsMerge, read the details of this funcion:

?cytof_exprsMerge

2.2 Cell Subset Detection

The subset detection is performed by clustering algorithms. cytofkit is shipped with two state-of-the-art clustering methods DensVM and PhenoGraph and an in-house developed clustering algorithm ClusterX. Among them, DensVM and ClusterX are density-based clustering algorithms estimated from the t-SNE embedded map. While PhenoGraph is a graph based clustering algorithm performed directly on the high dimensional data. For version later than 1.4.4, we plugged in the FlowSOM clustering algorithm [6]. Check the function cytof_cluster for implementation of this step.

?cytof_cluster

2.3 Cell Subset Visualization and Interpretation

Three dimensionality reduction methods are integrated into cytofkit for visualizing the high dimensional mass cytometry data into an information-condensed and relationship-preserved two-dimensional map. These included one linear transformation method PCA and two non-linear transformation methods ISOMAP and t-SNE. PCA is a widely used technique to remove the linear correlation between variables. After PCA, data is represented by principal components, and those two or three principal components with highest variance are used to visualize the data. ISOMAP is used to embed the high dimensional data points into low dimensions while preserving the intrinsic geometry of the data structure. t-SNE finds the two-dimensional representation of high-dimensional data that best preserve both local and global geometry. After dimensional reduction, cytofkit plots the cell points into these transformed two-dimensional map with point color representing its cell type detected from last step, and point shape representing which sample the cell is from. To summarize the expression profile of each cell subset, a heat map can be generated to visualize the mean or median expression for each marker in each cell type. This heat map can greatly facilitate the interpretation of known cell types with prior knowledge of their special marker expression features as well as detection of new cell types with novel expression patterns. Check the function cytof_dimReduction for implementation of this step.

?cytof_dimReduction

2.4 Inference of Subset Progression

Instead of inferring the subset progression using all cells, we perform the progression path estimation based on the detected subsets of cells. Specifically, we down sample the number of cells to a fixed number for each cell subset, thus to make every subset has the equal number of cells. This removes the potential dominant influence from cell subtypes with most number of cells. Then Diffusion Map or ISOMAP is used to infer the relationships among the down sampled subsets. Diffusion Map and ISOMAP are both able to compare the overall relatedness of cells and embed the cells taking into account the nonlinear geodesic distances between cells. By plotting the marker expression along the Diffusion Map or ISOMAP dimension 1 or dimension 2 with lowess regression, markers with expression patterns that highly correlated with progression can be defined as regulatory markers. Check the function cytof_progression for implementation of this step.

?cytof_progression

2.5 Post-processing

The cluster results can be visualized with dimension transformed data from t-SNE, PCA and ISOMAP. The cluster coordinates, together with the t-SNE, PCA and ISOMAP coordinates, will be stored to the FCS files as additional parameters and for post analysis, all result files and plots can be saved using the function cytof_writeResults.

?cytof_writeResults

Among the saved results, an R data object will be saved with suffix of .RData, this is specially for loading to the shiny web APP for interactive visualization and exploration of the analysis results. The shiny APP can be launched with command cytofkitShinyAPP() or through this link https://cytofkit.shinyapps.io/shiny/.