Transcriptomic deconvolution in cancer and other heterogeneous tissues remains challenging. Available methods lack the ability to estimate both component-specific proportions and expression profiles for individual samples. We develop a three-component deconvolution model, DeMixT, for expression data from a mixture of cancerous tissues, infiltrating immune cells and tumor microenvironment. DeMixT is a software package that performs deconvolution on transcriptome data from a mixture of two or three components.
DeMixT is a frequentist-based method and fast in yielding accurate estimates of cell proportions and compart-ment-specific expression profiles for two-component \and three-component deconvolution problem. Our method promises to provide deeper insight into cancer biomarkers and assist in the development of novel prognostic markers and therapeutic strategies.
The function DeMixT is designed to finish the whole pipeline of deconvolution for two or three components. The newly added DeMixT_GS function is designed to estimates the proportions of mixed samples for each mixing component based on a new approach to select genes more effectively that utilizes profile likelihood. DeMixT_DE function is designed to estimate the proportions of all mixed samples for each mixing component based on the gene differential expressions to select genes. DeMixT_S2 function is designed to estimate the component-specific deconvolved expressions of individual mixed samples for a given set of genes.
The DeMixT R-package builds the transcriptomic deconvolution with a couple of novel features into R-based standard analysis pipeline through Bioconductor. DeMixT showed high accuracy and efficiency from our designed experiment. Hence, DeMixT can be considered as an important step towards linking tumor transcriptomic data with clinical outcomes.
Different from most previous computational deconvolution methods, DeMixT has integrated new features for the deconvolution with more than 2 components.
Joint estimation: jointly estimate component proportions and expression profiles for individual samples by requiring reference samples instead of reference genes; For the three-component deconvolution considering immune infiltration, it provides a comprehensive view of tumor-stroma-immune transcriptional dynamics, as compared to methods that address only immune subtypes within the immune component, in each tumor sample.
Efficient estimation: DeMixT adopts an approach of iterated conditional modes (ICM) to guarantee a rapid convergence to a local maximum. We also design a novel gene-set-based component merging approach to reduce the bias of proportion estimation for three-component deconvolutionthe.
Parallel computing: OpenMP enable parallel computing on single computer by taking advantage of the multiple cores shipped on modern CPUs. The ICM framework further enables parallel computing, which helps compensate for the expensive computing time used in the repeated numerical double integrations.
The DeMixT package is compatible with Windows, Linux and MacOS. The
user can install it from Bioconductor
:
if (!require("BiocManager", quietly = TRUE))
install.packages("BiocManager")
BiocManager::install("DeMixT")
For Linux and MacOS, the user can also install the latest DeMixT from GitHub:
if (!require("devtools", quietly = TRUE))
install.packages('devtools')
devtools::install_github("wwylab/DeMixT")
Check if DeMixT is installed successfully:
# load package
library(DeMixT)
Note: DeMixT relies on OpenMP for parallel computing. Starting from R 4.00, R no longer supports OpenMP on MacOS, meaning the user can only run DeMixT with one core on MacOS. We therefore recommend the users to mainly use Linux system for running DeMixT to take advantage of the multi-core parallel computation.
The following table shows the functions included in DeMixT.
Table Header | Second Header |
---|---|
DeMixT | Deconvolution of tumor samples with two or three components. |
DeMixT_GS | Estimates the proportions of mixed samples for each mixing component based on a new approach to select genes that utilizes profile likelihood. |
DeMixT_DE | Estimates the proportions of mixed samples for each mixing component. |
DeMixT_S2 | Deconvolves expressions of each sample for unknown component. |
Optimum_KernelC | Call the C function used for parameter estimation in DeMixT. |
DeMixT_Preprocessing | Preprocessing functions before running DeMixT. |
Let \(Y_{ig}\) be the observed expression levels of the raw measured data from clinically derived malignant tumor samples for gene \(g, g = 1, \cdots, G\) and sample \(i, i = 1, \cdots, My\). \(G\) denotes the total number of probes/genes and \(My\) denotes the number of samples. The observed expression levels for solid tumors can be modeled as a linear combination of raw expression levels from three components: \[ {Y_{ig}} = \pi _{1,i}N_{1,ig} + \pi _{2,i}N_{2,ig} + (1 - \pi_{1,i} - \pi _{2,i}){T_{ig}} \label{eq:1} \]
Here \(N_{1,ig}\), \(N_{2,ig}\) and \({T_{ig}}\) are the unobserved raw expression levels from each of the three components. We call the two components for which we require reference samples the \(N_1\)-component and the \(N_2\)-component. We call the unknown component the T-component. We let \(\pi_{1,i}\) denote the proportion of the \(N_1\)-component, \(\pi_{2,i}\) denote the proportion of the \(N_2\)-component, and \(1 - \pi_{1,i}-\pi_{2,i}\) denote the proportion of the T-component. We assume that the mixing proportions of one specific sample remain the same across all genes.
Our model allows for one component to be unknown, and therefore does not require reference profiles from all components. A set of samples for \(N_{1,ig}\) and \(N_{2,ig}\), respectively, needs to be provided as input data. This three-component deconvolution model is applicable to the linear combination of any three components in any type of material. It can also be simplified to a two-component model, assuming there is just one \(N\)-component. For application in this paper, we consider tumor (\(T\)), stromal (\(N_1\)) and immune components (\(N_2\)) in an admixed sample (\(Y\)).
Following the convention that \(\log_2\)-transformed microarray gene expression data follow a normal distribution, we assume that the raw measures \(N_{1,ig} \sim LN({\mu _{{N_1}g}},\sigma _{{N_1}g}^2)\), \(N_{2,ig} \sim LN({\mu _{{N_2}g}},\sigma _{{N_2}g}^2)\) and \({T_{ig}} \sim LN({\mu _{Tg}}, \sigma _{Tg}^2)\), where LN denotes a \(\log_2\)-normal distribution and \(\sigma _{{N_1}g}^2\),\(\sigma _{{N_2}g}^2\), \(\sigma _{Tg}^2\) reflect the variations under \(\log_2\)-transformed data. Consequently, our model can be expressed as the convolution of the density function for three \(\log_2\)-normal distributions. Because there is no closed form of this convolution, we use numerical integration to evaluate the complete likelihood function (see the full likelihood in the Supplementary Materials in [1]).
DeMixT estimates all distribution parameters and cellular proportions and reconstitutes the expression profiles for all three components for each gene and each sample. The estimation procedure (summarized in Figure 1b) has two main steps as follows.
Obtain a set of parameters \(\{\pi_{1,i}, \pi_{2,i}\}_{i=1}^{My}\), \(\{\mu_T, \sigma_T\}_{g=1}^G\) to maximize the complete likelihood function, for which \(\{\mu_{N_{1,g}}, \sigma_{N_{1,g}}, \mu_{N_{2,g}}, \sigma_{N_{2,g}}\}_{g=1}^G\) were already estimated from the available unmatched samples of the \(N_1\) and \(N_2\) component tissues. (See further details in our paper.)
Reconstitute the expression profiles by searching each set of \(\{n_{1,ig}, n_{2,ig}\}\) that maximizes the joint density of \(N_{1,ig}\), \(N_{2,ig}\) and \(T_{ig}\). The value of \(t_{ig}\) is solved as \({y_{ig}} - {{\hat \pi }_{1,i}}{n_{1,ig}} - {{\hat \pi }_{2,i}}{n_{2,ig}}\).
These two steps can be separately implemented using the function DeMixT_DE or DeMixT_GS for the first step and DeMixT_S2 for the second, which are combined in the function DeMixT(Note: DeMixT_GS is the default function for first step).
Since version 1.8.2, DeMixT added simulated normal reference samples, i.e., spike-in, based on the observed normal reference samples. It has been shown to improve accuracy in proportion estimation for the scenario where a dataset consists of samples where true tumor proportions are skewed to the high end.
data("test.data.2comp")
# res.GS = DeMixT_GS(data.Y = test.data.2comp$data.Y,
# data.N1 = test.data.2comp$data.N1,
# niter = 30, nbin = 50, nspikein = 50,
# if.filter = TRUE, ngene.Profile.selected = 150,
# mean.diff.in.CM = 0.25, ngene.selected.for.pi = 150,
# tol = 10^(-5))
load('Res_2comp/res.GS.RData')
## PiN1 PiT
## Sample 1 0.5955120 0.4044880
## Sample 2 0.2759014 0.7240986
## Sample 3 0.5401655 0.4598345
## Sample 4 0.4497041 0.5502959
## Sample 5 0.6516980 0.3483020
## Sample 6 0.4365191 0.5634809
## [1] "Gene 418" "Gene 452" "Gene 421" "Gene 112" "Gene 154" "Gene 143"
data("test.data.2comp")
# res.S2 <- DeMixT_S2(data.Y = test.data.2comp$data.Y,
# data.N1 = test.data.2comp$data.N1,
# data.N2 = NULL,
# givenpi = c(t(res.S1$pi[-nrow(res.GS$pi),])), nbin = 50)
load('Res_2comp/res.S2.RData')
## Sample 1 Sample 2 Sample 3 Sample 4 Sample 5
## Gene 1 18.857446 60.727041 159.878946 92.031635 40.873852
## Gene 2 2.322481 3.390938 2.406093 2.558962 2.438189
## Gene 3 48.843631 208.166410 66.986239 38.107580 460.556751
## Sample 1 Sample 2 Sample 3 Sample 4 Sample 5
## Gene 1 59.37087 71.80492 74.1755 73.55878 72.96267
## Gene 2 107.66874 131.20005 113.6376 120.35924 125.28224
## Gene 3 513.43184 669.79145 613.3042 491.09308 741.76507
## MuN1 MuT
## Gene 1 6.166484 5.924321
## Gene 2 6.677594 2.974551
## Gene 3 9.329628 7.396647
## SigmaN SigmaT
## Gene 1 0.2222914 1.127726
## Gene 2 0.2319681 1.614169
## Gene 3 0.1881647 1.320477
In the simulation,
## Simulate MuN and MuT for each gene
MuN <- rnorm(G, 7, 1.5)
MuT <- rnorm(G, 7, 1.5)
Mu <- cbind(MuN, MuT)
## Simulate SigmaN and SigmaT for each gene
SigmaN <- runif(n = G, min = 0.1, max = 0.8)
SigmaT <- runif(n = G, min = 0.1, max = 0.8)
## Simulate Tumor Proportion
PiT = truncdist::rtrunc(n = My,
spec = 'norm',
mean = 0.55,
sd = 0.2,
a = 0.25,
b = 0.95)
## Simulate Data
for(k in 1:G){
data.N1[k,] <- 2^rnorm(M1, MuN[k], SigmaN[k]); # normal reference
True.data.T[k,] <- 2^rnorm(My, MuT[k], SigmaT[k]); # True Tumor
True.data.N1[k,] <- 2^rnorm(My, MuN[k], SigmaN[k]); # True Normal
data.Y[k,] <- pi[1,]*True.data.N1[k,] + pi[2,]*True.data.T[k,] # Mixture Tumor
}
where \(\pi_i \in (0.25, 0.95)\) is from truncated normal distribution. In general, the true distribution of tumor proportion does not follow a uniform distribution between \([0,1]\), but instead skewed to the upper part of the interval.
# ## DeMixT_DE without Spike-in Normal
# res.S1 = DeMixT_DE(data.Y = test.data.2comp$data.Y,
# data.N1 = test.data.2comp$data.N1,
# niter = 30, nbin = 50, nspikein = 0,
# if.filter = TRUE,
# mean.diff.in.CM = 0.25, ngene.selected.for.pi = 150,
# tol = 10^(-5))
# ## DeMixT_DE with Spike-in Normal
# res.S1.SP = DeMixT_DE(data.Y = test.data.2comp$data.Y,
# data.N1 = test.data.2comp$data.N1,
# niter = 30, nbin = 50, nspikein = 50,
# if.filter = TRUE,
# mean.diff.in.CM = 0.25, ngene.selected.for.pi = 150,
# tol = 10^(-5))
# ## DeMixT_GS with Spike-in Normal
# res.GS.SP = DeMixT_GS(data.Y = test.data.2comp$data.Y,
# data.N1 = test.data.2comp$data.N1,
# niter = 30, nbin = 50, nspikein = 50,
# if.filter = TRUE, ngene.Profile.selected = 150,
# mean.diff.in.CM = 0.25, ngene.selected.for.pi = 150,
# tol = 10^(-5))
load('Res_2comp/res.S1.RData'); load('Res_2comp/res.S1.SP.RData');
load('Res_2comp/res.GS.RData'); load('Res_2comp/res.GS.SP.RData');
This simulation was designed to compare previous DeMixT resutls with DeMixT spike-in results under both gene selection method.
res.2comp = as.data.frame(cbind(round(rep(t(test.data.2comp$pi[2,]),3),2),
round(c(t(res.S1$pi[2,]),t(res.S1.SP$pi[2,]), t(res.GS.SP$pi[2,])),2),
rep(c('DE','DE-SP','GS-SP'), each = 100)), num = 1:2)
res.2comp$V1 <- as.numeric(as.character(res.2comp$V1))
res.2comp$V2 <- as.numeric(as.character(res.2comp$V2))
res.2comp$V3 = as.factor(res.2comp$V3)
names(res.2comp) = c('True.Proportion', 'Estimated.Proportion', 'Method')
## Plot
ggplot(res.2comp, aes(x=True.Proportion, y=Estimated.Proportion, group = Method, color=Method, shape=Method)) +
geom_point() +
geom_abline(intercept = 0, slope = 1, linetype = "dashed", color = "black", lwd = 0.5) +
xlim(0,1) + ylim(0,1) +
scale_shape_manual(values=c(seq(1:3))) +
labs(x = 'True Proportion', y = 'Estimated Proportion')
In this simulation,
G <- G1 + G2
## Simulate MuN1, MuN2 and MuT for each gene
MuN1 <- rnorm(G, 7, 1.5)
MuN2_1st <- MuN1[1:G1] + truncdist::rtrunc(n = 1,
spec = 'norm',
mean = 0,
sd = 1.5,
a = -0.1,
b = 0.1)
MuN2_2nd <- c()
for(l in (G1+1):G){
tmp <- MuN1[l] + truncdist::rtrunc(n = 1,
spec = 'norm',
mean = 0,
sd = 1.5,
a = 0.1,
b = 3)^rbinom(1, size=1, prob=0.5)
while(tmp <= 0) tmp <- MuN1[l] + truncdist::rtrunc(n = 1,
spec = 'norm',
mean = 0,
sd = 1.5,
a = 0.1,
b = 3)^rbinom(1, size=1, prob=0.5)
MuN2_2nd <- c(MuN2_2nd, tmp)
}
## Simulate SigmaN1, SigmaN2 and SigmaT for each gene
SigmaN1 <- runif(n = G, min = 0.1, max = 0.8)
SigmaN2 <- runif(n = G, min = 0.1, max = 0.8)
SigmaT <- runif(n = G, min = 0.1, max = 0.8)
## Simulate Tumor Proportion
pi <- matrix(0, 3, My)
pi[1,] <- runif(n = My, min = 0.01, max = 0.97)
for(j in 1:My){
pi[2, j] <- runif(n = 1, min = 0.01, max = 0.98 - pi[1,j])
pi[3, j] <- 1 - sum(pi[,j])
}
## Simulate Data
for(k in 1:G){
data.N1[k,] <- 2^rnorm(M1, MuN1[k], SigmaN1[k]); # normal reference 1
data.N2[k,] <- 2^rnorm(M2, MuN2[k], SigmaN2[k]); # normal reference 1
True.data.T[k,] <- 2^rnorm(My, MuT[k], SigmaT[k]); # True Tumor
True.data.N1[k,] <- 2^rnorm(My, MuN1[k], SigmaN1[k]); # True Normal 1
True.data.N2[k,] <- 2^rnorm(My, MuN2[k], SigmaN2[k]); # True Normal 1
data.Y[k,] <- pi[1,]*True.data.N1[k,] + pi[2,]*True.data.N2[k,] +
pi[3,]*True.data.T[k,] # Mixture Tumor
}
where \(G1\) is the number of genes that \(\mu_{N1}\) is close to \(\mu_{N2}\).
data("test.data.3comp")
# res.S1 <- DeMixT_DE(data.Y = test.data.3comp$data.Y, data.N1 = test.data.3comp$data.N1,
# data.N2 = test.data.3comp$data.N2, if.filter = TRUE)
load('Res_3comp/res.S1.RData');
res.3comp= as.data.frame(cbind(round(t(matrix(t(test.data.3comp$pi), nrow = 1)),2),
round(t(matrix(t(res.S1$pi), nrow = 1)),2),
rep(c('N1','N2','T'), each = 20)))
res.3comp$V1 <- as.numeric(as.character(res.3comp$V1))
res.3comp$V2 <- as.numeric(as.character(res.3comp$V2))
res.3comp$V3 = as.factor(res.3comp$V3)
names(res.3comp) = c('True.Proportion', 'Estimated.Proportion', 'Component')
## Plot
ggplot(res.3comp, aes(x=True.Proportion, y=Estimated.Proportion, group = Component, color=Component, shape=Component)) +
geom_point() +
geom_abline(intercept = 0, slope = 1, linetype = "dashed", color = "black", lwd = 0.5) +
xlim(0,1) + ylim(0,1) +
scale_shape_manual(values=c(seq(1:3))) +
labs(x = 'True Proportion', y = 'Estimated Proportion')
Here, we use a subset of the bulk RNAseq data of prostate adenocarcinoma (PRAD) from TCGA (https://portal.gdc.cancer.gov/) as an example. The analysis pipeline consists of the following steps:
The raw read counts for the tumor and normal samples from TCGA PRAD are downloaded from TCGA data portal. One can also generate the raw read counts from fastq or bam files by following the GDC mRNA Analysis Pipeline.
Load input data (available at PRAD.RData)
Three data are included in the PRAD.RData
file.
PRAD
: Read counts matrix (gene x sample) with genes as
row names and sample ids as column names.Normal.id
: TCGA ids of PRAD normal samples.Tumor.id
TCGA ids of PRAD tumor samples.A glimpse of PRAD
:
## TCGA-CH-5761-11A TCGA-CH-5767-11B TCGA-EJ-7115-11A TCGA-EJ-7123-11A
## TSPAN6 3876 7095 5542 2747
## TNMD 14 51 13 24
## DPM1 1162 2665 1544 1974
## SCYL3 777 1517 1096 1231
## C1orf112 136 343 214 280
## FGR 230 511 263 755
## TCGA-EJ-7125-11A
## TSPAN6 8465
## TNMD 63
## DPM1 2984
## SCYL3 1514
## C1orf112 339
## FGR 262
## Number of genes: 59427
## Number of normal sample: 20
## Number of tumor sample: 30
Conduct data cleaning and normalization before running DeMixT.
PRAD = PRAD[, c(Normal.id, Tumor.id)]
selected.genes = 9000
cutoff_normal_range = c(0.1, 1.0)
cutoff_tumor_range = c(0, 2.5)
cutoff_step = 0.1
preprocessed_data = DeMixT_preprocessing(PRAD,
Normal.id,
Tumor.id,
selected.genes,
cutoff_normal_range,
cutoff_tumor_range,
cutoff_step)
PRAD_filter = preprocessed_data$count.matrix
sd_cutoff_normal = preprocessed_data$sd_cutoff_normal
sd_cutoff_tumor = preprocessed_data$sd_cutoff_tumor
cat("Normal sd cutoff:", preprocessed_data$sd_cutoff_normal, "\n")
cat("Tumor sd cutoff:", preprocessed_data$sd_cutoff_tumor, "\n")
cat('Number of genes after filtering: ', dim(PRAD_filter)[1], '\n')
The function DeMixT_preprocessing
identifies two
intervals based on the standard deviation of the log-transformed gene
expression for normal and tumor samples, respectively, within the
pre-defined ranges (cutoff_normal_range
and
cutoff_tumor_range
). In this example, we choose to select
about 9000 genes before running DeMixT with the GS (Gene Selection)
method to ensure that our model-based gene selection maintains good
statistical properties.
DeMixT_preprocessing
outputs a list object called
preprocessed_data
which contains:
preprocessed_data$count.matrix
: Preprocesssed count
matrixpreprocessed_data$sd_cutoff_normal
: Actual cut-off
value when desired number of genes are selected for normal samplespreprocessed_data$sd_cutoff_tumor
: Actual cut-off value
when desired number of genes are selected for tumor samplesTo optimize the parameters in DeMixT
for input data, we
recommend testing an array of combinations of number of spike-ins and
number of selected genes.
The number of CPU cores used by the DeMixT
function for
parallel computing is specified by the parameter nthread
.
By default,
nthread = total_number_of_cores_on_the_machine - 1
. Users
can adjust nthread
to any number between 0 and the total
number of cores available on the machine. For reference,
DeMixT
takes approximately 3-4 minutes to process the PRAD
data in this tutorial for each parameter combination when
nthread
is set to 55.
# Due to the random initial values and the spike-in samples used in the DeMixT function,
# we recommand that users set seeds to ensure reproducibility.
# This seed setting will be incorporated internally in DeMixT in the next update.
set.seed(1234)
data.Y = SummarizedExperiment(assays = list(counts = PRAD_filter[, Tumor.id]))
data.N1 <- SummarizedExperiment(assays = list(counts = PRAD_filter[, Normal.id]))
# In practice, we set the maximum number of spike-in as min(n/3, 200),
# where n is the number of samples.
nspikesin_list = c(0, 5, 10)
# One may set a wider range than provided below for studies other than TCGA.
ngene.selected_list = c(500, 1000, 1500, 2500)
for(nspikesin in nspikesin_list){
for(ngene.selected in ngene.selected_list){
name = paste("PRAD_demixt_GS_res_nspikesin", nspikesin, "ngene.selected",
ngene.selected, sep = "_");
name = paste(name, ".RData", sep = "");
res = DeMixT(data.Y = data.Y,
data.N1 = data.N1,
ngene.selected.for.pi = ngene.selected,
ngene.Profile.selected = ngene.selected,
filter.sd = 0.7, # We recommand to use upper bound of gene expression standard deviation
# for normal reference. i.e., preprocessed_data$sd_cutoff_normal[2]
gene.selection.method = "GS",
nspikein = nspikesin)
save(res, file = name)
}
}
Note: We use a profiling likelihood-based method to
select genes, during which we calculate confidence intervals for the
model parameters using the inverse of the Hessian matrix. When the input
data (e.g., gene expression levels from spatial transcriptomic data) is
sparse, the Hessian matrix will contain infinite values, hence those
confidence intervals can’t be calculated. In this case, gene selection
will be performed through differential expression analysis (identical to
DeMix_DE
). This alternative is automatically performed
inside DeMix_GS
when the above situation happens.
PiT_GS_PRAD <- c()
row_names <- c()
for(nspikesin in nspikesin_list){
for(ngene.selected in ngene.selected_list){
name_simplify <- paste(nspikesin, ngene.selected, sep = "_")
row_names <- c(row_names, name_simplify)
name = paste("PRAD_demixt_GS_res_nspikesin", nspikesin,
"ngene.selected", ngene.selected, sep = "_");
name = paste(name, ".RData", sep = "")
load(name)
PiT_GS_PRAD <- cbind(PiT_GS_PRAD, res$pi[2, ])
}
}
colnames(PiT_GS_PRAD) <- row_names
This step saves the deconvolution results (PiT) into a dataframe with columns named after the combination of the number of spike-ins and number of genes selected. Then one can calculate and plot the pairwise correlations of estimated tumor proportions across different parameter combinations as shown below.
pairs.panels(PiT_GS_PRAD,
method = "spearman", # correlation method
hist.col = "#00AFBB",
density = TRUE, # show density plots
ellipses = TRUE, # show correlation ellipses
main = 'Correlations of Tumor Proportions with GS between Different Parameter
Combination',
xlim = c(0,1),
ylim = c(0,1))