Single-cell RNA-sequencing cell-type identification that leverages reference data to combine information across thousands of genes, learn probabilistic barcode representations of cell-types, and probabilistically assign cell-type identity to unknown cells.
Our code requires the packages sads and BiocParallel. We tested our code in R version 3.6.2 and 4.1.3. We are currently working on a package. In the meantime, our code can be used by downloading the files cell_type_identification4_clean.R and params_EM_81020.rda from this repository into the same directory, and calling source('cell_type_identification4_clean.R'). This should only take a few seconds.
To learn a barcode representation for cell-types in reference data, given a numeric matrix ref in genes by cells format with Ensembl IDs as rownames, and a vector of cell-type annotations labels:
fit <- trainAllReference(ref,labels)
barcodes <- getBarcode(fit)
By default, this will only fit to a subset of discriminatory genes for classification. If you would like to examine these barcodes for purposes other than classification, for example to identify marker genes, set discrim_only=FALSE in trainAllReference.
To identify unknown cells in a numeric matrix target in genes by cells format with Ensembl IDs as rownames, from a set of possible reference cell-types:
target_labels <- classifyTarget(target,fit)
If you would like cells to be flagged as "other" if they are not sufficiently similar to any of the cell-types present in the reference data, set other=T in classifyTarget.
To obtain probabilistic classifications for unknown cells:
target_probs <- classifyTarget(target,fit,return.probs=T)
Marker genes can be identified for a given cell-type either relative to other cell-types present in the data:
markers_relative <- findMarkers(cell_type_name,barcodes,thresh_up=0.95,thresh_below=0.05)
Or compared to global gene expression patterns from many cell-types and tissues:
markers_global <- findMarkers(cell_type_name,barcodes,thresh_up=0.99,thresh_below=0.1,relative=F)
We currently support human, UMI count data.
The Rmarkdown file PBMCs.Rmd provides links to example data from 10X Genomics and demonstrates training, classification, and marker identification. Note that this data is subsetted to a smaller amount of cells than described in the manuscript for the purposes of providing a lightweight demonstration. The demo requires the packages org.Hs.eg.db and Seurat for loading in and processing the data. This file assumes that the data have been downloaded from the provided links (select the filtered gene-by-cell count matrices). The directories in the Read10X commands in the Rmarkdown file may need to be updated depending on where and with what name the data files were saved to your system. Running this file should take approximately five minutes or less. The expected output is shown in PBMCs.pdf, which resulted from running this Rmarkdown file in R version 4.1.3.
Data used for benchmarking can be found here.