Earlier we demonstrated How a perfect single-cell subpopulation random forest separator is refined ,as well as LASSO regression can also be used for single-cell classification The two machine learning algorithms can be used as single-cell classifiers, and the effect is leveraged. And also tried a variety of machine learning algorithms, such as: SVM single-cell classifier not lost to LASSO
Whether it is random forest, LASSO regression, or support vector machine, their models are a bit abstract, and it is not easy to explain clearly intuitively. But the decision tree model we are going to introduce next is different.
Train a decision tree model
First, copy paste the previous How a perfect single-cell subpopulation random forest separator is refined , you can divide the single-cell expression matrix into training set and test set, then simply install and load the rpart package, and run the rpart function inside:
library('rpart.plot')
library('rpart')
library('survival')
df = cbind(target=target,as.data.frame(predictor_data))
head(colnames(df))
multivariate <- paste(colnames(df)[-1],collapse = '-')
s <- paste0('target ~ ', multivariate )
s # as.formula(s)
fit <- rpart( target ~. ,
data=df, method="class" )
summary(fit)
save(fit,file = 'rpart_output.Rdata')
It can be seen that the expression matrix we input is the expression information of 2000 genes in nearly 2000 cells:
> dim(predictor_data)
[1] 1977 2000
> predictor_data[1:4,1:4]
ISG15 CPSF3L MRPL20 ATAD3C
AAACATACAACCAC -0.8353028 -0.2741145 1.5056616 -0.04970561
AAACATTGAGCTAC -0.8353028 -0.2741145 -0.5625993 -0.04970561
AAACATTGATCAGC 0.3922351 -0.2741145 1.2450489 -0.04970561
AAACCGTGCTTCCG 2.2197621 -0.2741145 -0.5625993 -0.04970561
Our decision tree model is to combine these 2000 genes to divide the classification of cells. Let's simply visualize this effect:
library(rpart.plot);
rpart.plot(fit, branch=1, branch.type=2, type=2, extra=102,
shadow.col="gray", box.col="green",
border.col="blue", split.col="red",
split.cex=1.2 )
As shown below, although there are 6 unicellular subpopulations, only 5 genes are enough to distinguish them:
5 genes are enough to tell them apart
Look at the model effect on the test set
Similarly, the trained model also needs to see the effect in another data set:
test_outputs = predict(fit, as.data.frame(test_expr))
head( test_outputs )
pred_y = colnames(test_outputs)[apply(test_outputs, 1, which.max)]
pred_y = factor(pred_y,levels = levels(test_y))
pdf('rpart-performance.pdf',width = 10)
gplots::balloonplot(table(pred_y,test_y))
dev.off()
It can be seen that the problem is still the mixing of CD8 and NK cells, as well as the mixing of CD4 and CD8, which is currently unsolvable:
Incorporation of CD8 and NK cells
We can simply visualize the 5 genes of the previous decision tree model:
library(Seurat)
sce=CreateSeuratObject(counts = t(predictor_data) )
Idents(sce) = target
cg = c('FTL','HLA-DRA','NKG7','FCER1G','CST3')
VlnPlot(sce,features = cg ,
stack = T)
ggplot2::ggsave('rpart_models_VlnPlot.pdf' )
As follows:
The 5 Genes of a Decision Tree Model
Compared with the previous decision tree model:
- It is indeed the FTL gene that distinguishes monocytes from other cells
- Then HLA-DRA can distinguish B cells and dendritic cells from other cells, and the distinction between B cells and dendritic cells depends on CST3
- Then CD4 in T cells is distinguished by NKG7, and then CD8 and NK cells are distinguished by FCER1G
Such a model is very easy to explain, compared to the previous random forest, LASSO regression, support vector machine.