SCRAM is available from https://github.com/akdess/scramUpdated version of CaSpER: is here: https://github.com/akdess/scram This vignette shows the basic steps for running SCRAM.
Install latest version from GitHub (requires devtools package):
if (!require("devtools")) {
install.packages("devtools")
}
devtools::install_github("akdess/scram", dependencies=TRUE, build_vignettes=FALSE)
devtools::install_github("akdess/casper_0.2.0", dependencies=TRUE, build_vignettes=FALSE)We first save the R object in h5ad for predicting cell-types on our trained deep learning models. We next predict the cell types on our data using our pretrained deep learning models.
cd example;
<path_to_scripts_folder>/run_conversion.sh glioma_seuratObj.rda ./
python3 nn_classifier_pretrained.py 041524_glioma ./adata.h5ad --normalize_test <path to nn_models folder> ./
We next load the seuratobject again in R with the predicted model outcomes.
setwd("example")
load("glioma_seuratObj.rda")
project<-"example"scram_obj<- CreateSCRAMObject(seurat_obj=seuratObj, organism="human", min_support=0.1, max_set_size=50) scram_obj<-createCellTypeMatrix (object=scram_obj, nn_path="./041524_glioma_multipleNeuralNetworks/", prob_thr=0.9, refs=c('suva_idh_a_o', 'hpa_brain_simple', 'allen_class_label_main', 'allen_neurons_only', 'TissueImmune', 'aldinger', 'codex', 'suva', 'bhaduri_withAge', 'dirks_primary_gbm_combined'), run="041524_glioma", pretrained=T)
Because tumor cells exhibit a wide range of transcriptional states, we employ redundant and stringent approaches to annotate tumor cells using 3 modular components: (1) marker-expression modeling, (2) genotyping of CNVs on all cells (3) RNA-inferred mutational profiling of known glioma mutations (i.e. IDH1, EGFR).
To estimate a “clean” set of CNV calls that can provide reliable CNV-based tumor scores, we use a pure tumor pseudobulk sample.
The CNV calling on the pseudobulk samples is performed using our updated CNV calling algorithm, CaSpER+, for each patient. CaSpER+ CNV calls are used as the ground truth large-scale CNV calls for each patient.
After CNVs are identified from the pseudobulk sample, we genotype the set of CNVs on all cells and generate a binary matrix that represents the existence of CNVs on the cells, i.e., CNV_(i,j).
SeuratObj should have orig.ident with sample ids, and another metadata with id "tumorType" with "Normal" annotation for control cells
### single cell level CNV calling takes (~5 hours for 200K cells) long for large scRNA-Seq datasets. ### the output is provided under cnv_casper folderscram_obj<- runCASPER_Bulk_Scell(object=scram_obj, sampleCol="orig.ident", project)We performed RNA-inferred rare deleterious (COSMIC-reported and dbSNP, <0.1% frequency) mutational profiling via our recently developed XCVATR tool. Below code shows how to generate SNV input for SCRAM using XCAVTR output:
#### read XCVATR output### the output is saved under XCVATR foldervar_matrix<- readXCVATROutput(XCVATR_folderPath="./XCVATR/", name_mapping)load("cnv_casper/example_BULK_finalChrMat_thr_1.rda")
scram_obj<- runFinalTumorAnnotation(object=scram_obj, loadNumbat=F, loadCasper=T, finalChrMat_bulk=finalChrMat_bulk, loadXCVATR=T, sampleCol="orig.ident", project="example", model_genes=c("PDGFRA" ,"EGFR" ,"SOX2" ))
We summarized co-occurring cell types using a frequent itemset rule mining approach. CNV and SNV calls were added to provide an integrated transcriptomic and genomic summary for each cell.
An example SCRAM output for a single cell is given as “glioma stem cell, mature neuron, synaptic neuron, oligodendrocyte precursor cell, chr1p_deletion, chr19q_deletion + IDH1:2:208248389 mutation”.
We used the tumour and host cell assignments of the previous steps to integrate co-occurring tumour and host cell features.
scram_obj<- runSCRAM (object=scram_obj) writeSCRAMresults(object=scram_obj,project)