In raguirreg/DeconCell: DeconCell

Introduction

DeconCell is an r package containing models for predicting the proportions of circulating immune cell subpopulations using bulk gene expression data from whole blood. Models were built using an elastic net and training in 95 healthy dutch volunteers from the 500FG cohort with FACS quantification of 73 circulating cell subpopulations as described in our previous publication. For additional details on methods and results please go our manuscript.

Install the package from github

library(devtools)
#install_github("raguirreg/DeconCell")

Pre-processing example data

Let's load and pre-process our example data. These are 5 samples with > ~40k genes quantified. These are gene read counts, we need to approximate the example data to a normal-like distribution and account for library sizes. In order to do this, we use the dCell.expProcessing function. This function will perform a TMM normalization (as described in the edgeRpackage) a log2(counts+1) and scale (z-transformation) per gene.

library(DeconCell)
library(edgeR)
library(tidyverse)
library(ghibli)


data("count.table")
dCell.exp <- dCell.expProcessing(count.table, trim = TRUE)

Prediction of cell propotions

data("dCell.models")
prediction <- dCell.predict(dCell.exp, dCell.models, res.type = "median")
head(prediction$dCell.prediction)
head(prediction$Evaluation)

Correlation coeficient between of predicted and measured values

data("cell.proportions")
library(reshape2)
library(ggplot2)
data("dCell.names")
pData <- data.frame(PearsonCor= diag(cor(cell.proportions, prediction$dCell.prediction)), 
                    CTs = dCell.names[colnames(cell.proportions), "finalName"], 
                    Subpop = dCell.names[colnames(cell.proportions), "broadSubpopulations"])

ggplot(pData, aes(y=PearsonCor , x= CTs, fill=Subpop))+
  geom_bar(stat="identity", alpha=0.8)+
  geom_hline(yintercept = 0.5, alpha=0.5, color="red")+
  coord_flip()+
  scale_fill_brewer(palette = "Dark2")+
  theme_bw()+
  theme(legend.position = "bottom", text=element_text(size=11, family="Helvetica"))

Generate deconCell models for predicting new cell proportions using gene expression data.

An important functionality of DeconCell is its capacity to generate novel model to later predict the proportions of cell types within a bulk tissue using solely gene expression derived from the bulk tissue itself. To illustrate this we will make use of the publicly available data from the DeconRNASeq package. As this package states:

"Our demo uses a simulated example data set, which can be accessed using the code given below"

library(DeconRNASeq)
data(multi_tissue)

## remove colums that are not needed.
datasets <- x.data[,2:11]
signatures <- x.signature.filtered.optimal[,2:6]
proportions <- fraction
exp <- datasets

As the package indicates this is "real data" which has been mixed in silico, therefore the proportions of each of the different cell types composing the "bulk" expression are known.

For the mixtures, there are 28745 genes. And we have 10 samples. In silico mixed data were simulated using ([2]) data, with disparate proportions drawn from random numbers. The mixing proportions used by each type of tissue are shown in the following. It should also be noted that we investigated the influence of extremely low numbers of contaminating cell types (<2 percent).

We will run DeconCell for each of the proportions using 60% of the samples for training our models.

set.seed(1121)
sampled.train <- sample(colnames(exp), size = 6, replace = FALSE)
#use the rest of the samples for testing the models
sampled.test <- colnames(exp)[which(colnames(exp) %in% sampled.train == FALSE)]

new.dCell.models <- dCell.run(exp = exp[,sampled.train], 
                              proportions = proportions[sampled.train,], 
                              iterations = 5)

Testing the prediction of the new models.

Now we will use \code{dCell.predict} function to use the newly created models to predict the proportions on the defined test set (\code{sampled.test}) In the \code{DeconRNASeq} vignette, the author use the Root Mean Square Error (RMSE), which is the standard deviation from the residuals, as a measure of prediction performance.

test.prediction <- dCell.predict(exp[,sampled.test],
                                 dCell.models= new.dCell.models$deconCell.models.per.CT, 
                                 res.type = "median", custom = TRUE)
# we use custom=TRUE to keep the original names of the proportions.

# reshape the data for plotting 
pData <- reshape2::melt(as.matrix(proportions[sampled.test,]))
pData$Predicted <- reshape2::melt(as.matrix(test.prediction$dCell.prediction))$value

## Function to calculate the Root Mean Square Error
rmse.calculate <- function(x, x.pred){
  sqrt(mean((x - x.pred)^2))
}

tissues <- as.character(unique(pData$Var2))
rmse.per.tissue <- sapply(tissues, function(x){rmse.calculate(pData$value[which(pData$Var2 == x)], pData$Predicted[which(pData$Var2 == x)])})

pData$RMSE <- rmse.per.tissue[as.character(pData$Var2)]
pData$RMSE <- paste0("RMSE= ", format(pData$RMSE,digits= 3))

decon.cell.tissue.plot <- ggplot(pData, aes(x= value, y=Predicted))+
                          geom_point(alpha= 0.9, size=1.5, aes(color= Var2))+
                          facet_grid(facets = ~Var2+RMSE, scales = "free")+
                          geom_smooth(method='lm', lwd=0.5,aes(color= Var2, alpha= 0.5))+
                          ylab("Decon-cell predicted \n tissue proportions")+
                          xlab("Tissue proportions")+
                          scale_color_manual(values = ghibli_palette("KikiMedium")[1:5])+
                          theme_bw()+
                          theme(text = element_text(family = "Helvetica", size = 10), legend.position = "none")


plot(decon.cell.tissue.plot)

As seen in the plot above, we have been able to accurately predict the proportions of the tissues within the bulk expression data by using the models generated with \code{dCell.run}.

raguirreg/DeconCell documentation built on May 29, 2019, 8:05 a.m.