2. A simple workflow of the scGPS:
The purpose of this workflow is to solve the following task:
- Given a mixed population with known subpopulations, estimate transition scores between these subpopulation.
2.1 Create scGPS objects
# load mixed population 1 (loaded from day_2_cardio_cell_sample dataset,
# named it as day2)
library(scGPS)
day2 <- day_2_cardio_cell_sample
mixedpop1 <- new_scGPS_object(ExpressionMatrix = day2$dat2_counts,
GeneMetadata = day2$dat2geneInfo, CellMetadata = day2$dat2_clusters)
# load mixed population 2 (loaded from day_5_cardio_cell_sample dataset,
# named it as day5)
day5 <- day_5_cardio_cell_sample
mixedpop2 <- new_scGPS_object(ExpressionMatrix = day5$dat5_counts,
GeneMetadata = day5$dat5geneInfo, CellMetadata = day5$dat5_clusters)2.2 Run prediction
# select a subpopulation
c_selectID <- 1
# load gene list (this can be any lists of user selected genes)
genes <- training_gene_sample
genes <- genes$Merged_unique
# load cluster information
cluster_mixedpop1 <- colData(mixedpop1)[,1]
cluster_mixedpop2 <- colData(mixedpop2)[,1]
#run training (running nboots = 3 here, but recommend to use nboots = 50-100)
LSOLDA_dat <- bootstrap_prediction(nboots = 3, mixedpop1 = mixedpop1,
mixedpop2 = mixedpop2, genes = genes, c_selectID = c_selectID,
listData = list(), cluster_mixedpop1 = cluster_mixedpop1,
cluster_mixedpop2 = cluster_mixedpop2, trainset_ratio = 0.7)
names(LSOLDA_dat)
#> [1] "Accuracy" "ElasticNetGenes" "Deviance"
#> [4] "ElasticNetFit" "LDAFit" "predictor_S1"
#> [7] "ElasticNetPredict" "LDAPredict" "cell_results"2.3 Summarise results
# summary results LDA
sum_pred_lda <- summary_prediction_lda(LSOLDA_dat = LSOLDA_dat, nPredSubpop = 4)
# summary results Lasso to show the percent of cells
# classified as cells belonging
sum_pred_lasso <- summary_prediction_lasso(LSOLDA_dat = LSOLDA_dat,
nPredSubpop = 4)
# plot summary results
plot_sum <-function(sum_dat){
sum_dat_tf <- t(sum_dat)
sum_dat_tf <- na.omit(sum_dat_tf)
sum_dat_tf <- apply(sum_dat[, -ncol(sum_dat)],1,
function(x){as.numeric(as.vector(x))})
sum_dat$names <- gsub("ElasticNet for subpop","sp", sum_dat$names )
sum_dat$names <- gsub("in target mixedpop","in p", sum_dat$names)
sum_dat$names <- gsub("LDA for subpop","sp", sum_dat$names )
sum_dat$names <- gsub("in target mixedpop","in p", sum_dat$names)
colnames(sum_dat_tf) <- sum_dat$names
boxplot(sum_dat_tf, las=2)
}
plot_sum(sum_pred_lasso)
plot_sum(sum_pred_lda)
# summary accuracy to check the model accuracy in the leave-out test set
summary_accuracy(object = LSOLDA_dat)
#> [1] 61.39535 68.57143 60.74766
# summary maximum deviance explained by the model
summary_deviance(object = LSOLDA_dat)
#> $allDeviance
#> [1] "10.55" "5.84" "11.36"
#>
#> $DeviMax
#> dat_DE$Dfd Deviance DEgenes
#> 1 0 5.84 genes_cluster1
#> 2 1 5.84 genes_cluster1
#> 3 2 5.84 genes_cluster1
#> 4 3 5.84 genes_cluster1
#> 5 remaining DEgenes remaining DEgenes remaining DEgenes
#>
#> $LassoGenesMax
#> NULL
