fsctype - Modification of the Original ScType Method for Barcode Annotation using the KNN Algorithm
The fsctype method is based on the original ScType method used for cluster annotation of single-cell datasets using known markers.
This variation uses the data.table package to substantially speed up calculations and also makes use of the future package for parallel execution.
A typical dataset of 10,000 barcodes or less will run in under 10 seconds in a memory efficient way. You should only rely on parallel execution if your dataset is larger than 50,000 barcodes as the speed up is minimal otherwise and not noticeable.
This method is different than the original in that instead of relying on precomputed clusters for annotation, it traverses each barcode in a neighborhood graph and uses the k nearest-neighbors to assign a per-cell annotation using the original ScType calculation.
You can install all required packages directly from CRAN
install.packages(c('Seurat', 'dplyr', 'ggplot2', 'data.table', 'igraph'))
or via mamba and the environment file that is provided.
mamba env create -f sctype.yml
This method is highly sensitive to the marker genes used. The more specific your markers are the better the annotation will be. You can use the marker genes from the original ScType database or provide your own, as long as the input marker list is created to look like the output of gene_sets_prepare.R from the original method. In other words, you need a nested named list with at least positive markers for your cell types or regions of interest.
Your dataset must be processed up to the shared neighborhood graph calculation or some igraph object with similarity weights for edges.
pkgs <- c('Seurat', 'dplyr', 'ggplot2', 'data.table', 'igraph')
invisible(lapply(pkgs, require, character.only=TRUE))
object <- CreateSeuratObject(counts=Read10X_h5('filtered_feature_bc_matrix.h5'))
object <- object %>%
NormalizeData() %>%
FindVariableFeatures() %>%
ScaleData(verbose=FALSE) %>%
RunPCA(verbose=FALSE) %>%
FindNeighbors(dims=seq(10)) %>%
RunUMAP(dims=seq(10))
Once you have run FindNeighbors, you can use this graph downstream for fsctype as it relies on the k nearest neighbors of each cell for annotation. The annotation_markers_list.rds is a created marker list as defined by gene_sets_prepare.R. You can use this from the original ScType method or create your own.
source('https://raw.githubusercontent.com/shahrozeabbas/fsctype/main/R/fsctype.R')
markers <- readRDS('annotation_markers_list.rds')
counts <- as.data.frame(
object %>%
Seurat::ScaleData(verbose=FALSE, features=unique(unlist(markers))) %>%
Seurat::GetAssayData(slot='scale.data')
)
cells <- colnames(object)
graph <- object %>% get_graph()
k <- round(sqrt(ncol(object) * 0.5))
predictions <- fsctype(barcodes=cells, graph=graph, counts=counts, markers=markers, n_neighbors=k)
cell_type <- predictions[, prediction]
names(cell_type) <- predictions[, barcodes]
object <- object %>% AddMetaData(metadata=factor(cell_type), col.name='cell_type')
ggplot2::ggsave(
width=12, height=8, filename='fsctype_predictions_umap.png',
plot=DimPlot(object, group.by='cell_type', label=TRUE, repel=TRUE, pt.size=1)
)