Contents

1 Introduction

High-throughput technologies have revolutionized the genomics research. The early applications of the technologies were largely on cell lines. However, there is an increasing number of larger-scale, population level clinical studies in recent years, hoping to identify diagnostic biomarkers and therapeutic targets. The samples collected in these studies, such as blood, tumor, or brain tissue, are mixtures of a number of different cell types. The sample mixing complicates data analysis because the experimental data from the high-throughput experiments are weighted averages of signals from multiple cell types. For these data, traditional analysis methods that ignores the cell mixture will lead to results with low resolution, biased, or even errorneous results. For example, it has been discovered that in epigenome-wide association studies (EWAS), the mixing proportions can be confounded with the experimental factor of interest (such as age). Ignoring the cell mixing will lead to false positives. On the other hand, cell type specific changes under different conditions could be associated with disease pathogenesis and progressions, which are of great interests to researchers.

For heterogeneous samples, it is possible to profile the pure cell types through experimental techniques. They are, however, laborious and expensive that cannot be applied to large scale studies. Computational tools for analzying the mixed data have been developed for proportion estimation and cell type specific signal detection. There are two fundamental questions in this type of analyses:

  1. How to estimate mixing proportions?

There are a number of existing methods devoted to solve this question. These methods mainly can be categorized to two groups: reference-based (require pure cell type profiles) and reference-free (does not require pure cell type profiles). It has been found that reference-based deconvolution is more accurate and reliable than reference-free deconvolution. However, the reference panels required for reference-based deconvolution can be difficult to obtain, thus reference-free method has wider application.

  1. with available mixing proportions, how to detect cell-type specific DE/DM?

TOAST is a package designed to answer these questions and serve the research communities with tools for the analysis of heterogenuous tissues. Currently TOAST provides functions to detect cell-type specific DE/DM, as well as differences across different cell types. TOAST also has functions to improve the accuracy of reference-free deconvolutions through better feature selection. If cell type-specific markers (or prior knowledge of cell compositions) are available, TOAST provides partial reference-free deconvolution function, which is more accuracte than RF methods and works well even for very small sample size (e.g.<10).

2 Installation and quick start

2.1 Install TOAST

To install this package, start R (version “3.6”) and enter:

if (!requireNamespace("BiocManager", quietly = TRUE))
    install.packages("BiocManager")

BiocManager::install("TOAST") 

2.2 How to get help for TOAST

Any TOAST questions should be posted to the GitHub Issue section of TOAST homepage at https://github.com/ziyili20/TOAST/issues.

2.3 Quick start on detecting cell type-specific differential signals

Here we show the key steps for a cell type-specific different analysis. This code chunk assumes you have an expression or DNA methylation matrix called Y_raw, a data frame of sample information called design, and a table of cellular composition (i.e. mixing proportions) called prop. Instead of a data matrix, Y_raw could also be a SummarizedExperiment object. If the cellular composition is not available, the following sections will discuss about how to obtain mixing proportions using reference-free deconvolution or reference-based deconvolution.

Design_out <- makeDesign(design, Prop)
fitted_model <- fitModel(Design_out, Y_raw)
fitted_model$all_coefs # list all phenotype names
fitted_model$all_cell_types # list all cell type names
# coef should be one of above listed phenotypes
# cell_type should be one of above listed cell types
res_table <- csTest(fitted_model, coef = "age", 
                    cell_type = "Neuron", contrast_matrix = NULL)
head(res_table)

3 Example dataset

TOAST provides two sample dataset.

The first example dataset is 450K DNA methylation data. We obtain and process this dataset based on the raw data provided by GSE42861. This is a DNA methylation 450K data for Rheumatoid Arthiritis patients and controls. The original dataset has 485577 features and 689 samples. We have reduced the dataset to 3000 CpGs for randomly selected 50 RA patients and 50 controls.

library(TOAST)
## Loading required package: EpiDISH
## Loading required package: limma
## Loading required package: nnls
## Loading required package: quadprog
## Registered S3 method overwritten by 'GGally':
##   method from   
##   +.gg   ggplot2
data("RA_100samples")
Y_raw <- RA_100samples$Y_raw
Pheno <- RA_100samples$Pheno
Blood_ref <- RA_100samples$Blood_ref

Check matrix including beta values for 3000 CpG by 100 samples.

dim(Y_raw) 
## [1] 3000  100
Y_raw[1:4,1:4]
##            GSM1051525 GSM1051526 GSM1051527  GSM1051528
## cg14521995  0.8848926  0.8654487  0.8172092 0.004429362
## cg11738485  0.9306579  0.9189274  0.5486962 0.039545301
## cg06193597  0.1388632  0.7127654  0.6925506 0.677185017
## cg14323910  0.8282483  0.8528023  0.8449638 0.828873689

Check phenotype of these 100 samples.

dim(Pheno)
## [1] 100   3
head(Pheno, 3)
##            age gender disease
## GSM1051525  67      2       1
## GSM1051526  49      2       1
## GSM1051527  53      2       1

Our example dataset also contain blood reference matrix for the matched 3000 CpGs (obtained from bioconductor package FlowSorted.Blood.450k.

dim(Blood_ref)
## [1] 3000    6
head(Blood_ref, 3)
##                 CD8T      CD4T        NK     Bcell      Mono      Gran
## cg14521995 0.9321049 0.9245206 0.9184654 0.9178081 0.8902820 0.9314544
## cg11738485 0.3745548 0.2916655 0.3144788 0.2985633 0.3027369 0.2911957
## cg06193597 0.4157621 0.4292540 0.4104737 0.4335429 0.4789953 0.4747334

The second example dataset is microarray gene expression data. We obtain and process this dataset based on the raw data provided by GSE65133. This microarary data is from 20 PBMC samples. The original dataset has 47323 probes. We mapped the probes into 21626 genes and then further reduced the dataset to 511 genes by selecting the genes that have matches in reference panel.

data("CBS_PBMC_array")
CBS_mix <- CBS_PBMC_array$mixed_all
LM_5 <- CBS_PBMC_array$LM_5
CBS_trueProp <- CBS_PBMC_array$trueProp
prior_alpha <- CBS_PBMC_array$prior_alpha
prior_sigma <- CBS_PBMC_array$prior_sigma

Check the PBMC microarray gene expression data and true proportions

dim(CBS_mix) 
## [1] 511  20
CBS_mix[1:4,1:4]
##       X17.002 X17.006 X17.019 X17.023
## ABCB4    96.0  107.50  110.00    92.3
## ABCB9    98.3  109.75  103.85    92.1
## ACAP1   196.8  217.80  351.00   140.7
## ACHE     92.7   97.20   87.10    87.1
head(CBS_trueProp, 3)
##            Bcells      CD8T      CD4T    NKcells      Monos
## 17-002 0.16201354 0.2364636 0.2453469 0.17533841 0.18083756
## 17-006 0.05279853 0.3279443 0.4698660 0.09106723 0.05832395
## 17-019 0.21897143 0.2041143 0.1468571 0.33874286 0.09131429

Check reference matrix for 5 immune cell types

head(LM_5, 3)
##          BCells       CD8T        CD4T   NK cells  Monocytes
## ABCB4 283.22884    4.31128    6.685426   9.119776   6.202496
## ABCB9  18.84917   24.22372   34.725041  19.129933  20.309426
## ACAP1 268.46349 1055.61338 1017.711114 450.109326 190.879024

Check prior knowledge for the 5 cell types

prior_alpha
## [1] 0.09475324 0.23471057 0.33231687 0.09689958 0.24131974
prior_sigma
## [1] 0.0996325 0.1441782 0.1602440 0.1006351 0.1455614

The third example dataset is a list containing two matrices, one of which is methylation 450K array data of 3000 CpG sites on 50 samples, the other is methylation 450K array data of 3000 matched CpG sites on three immune cell types. The first dataset is generated by simulation. It originally has 459226 features and 50 samples.We reduce it to 3000 CpGs by random selection.

data("beta_emp")
Ybeta = beta_emp$Y.raw
ref_m = beta_emp$ref.m

Check matrix including beta values for 3000 CpG by 50 samples.

dim(Ybeta) 
## [1] 3000   50
Ybeta[1:4,1:4]
##                  [,1]       [,2]       [,3]       [,4]
## cg08752431 0.76611838 0.76117075 0.79835014 0.79239766
## cg14555682 0.09677872 0.08096398 0.11906131 0.10553938
## cg23086843 0.88824253 0.88740138 0.92060008 0.92240033
## cg20308511 0.02863965 0.03976921 0.04269377 0.02678175

Check reference matrix for 3000 CpGs by three immune cell types

head(ref_m, 3)
##                  CD4T      CD8T     BCell
## cg08752431 0.77118808 0.7546620 0.7589832
## cg14555682 0.09960192 0.1201291 0.1006569
## cg23086843 0.90647715 0.8810441 0.9065662

4 Estimate mixing proportions

If you have mixing proportions available, you can directly go to Section 5.

In many situations, mixing proportions are not readily available. There are a number of deconvolution methods available to solve this problem. To name a few:

In addition, CellMix package has summarized a number of deconvolution methods and is a good resource to look up.

Here we demonstrate two ways to estimate mixing proportions, one using RefFreeEWAS (Houseman et al. 2016), representing the class of reference-free methods, and the other using EpiDISH (Teschendorff et al. 2017) as a representation of reference-based methods.

We also provide function to improve reference-free deconvolution performance in Section 4.3, which works for both gene expression data and DNA methylation data. The example in Section 4.3 demonstrates the usage of this. Note that we have only 3000 features in the Y_raw from RA_100samples dataset, thus the proportion estimation is not very accurate. Real 450K dataset should have around 485,000 features. More features generally lead to better estimation, because there are more information in the data.

In Secion 4.4, we demonstrate the usage of partial reference-free (PRF) deconvolution. Compared to RB methods, PRF does not require reference panel thus can be more wdiely applied. Compared to RF methods, PRF uses additional biological information, which improves the estimation accuracy and automatically assign cell type labels.

4.1 Reference-based deconvolution using least square method

  1. Select the top 1000 most variant features by findRefinx(). To select the top features with largest coefficients of variations, one can use findRefinx(..., sortBy = "cv"). Default sortBy argument is "var". Here, instead of a data matrix, Y_raw could also be a SummarizedExperiment object.
refinx <- findRefinx(Y_raw, nmarker = 1000)
  1. Subset data and reference panel.
Y <- Y_raw[refinx,]
Ref <- as.matrix(Blood_ref[refinx,])
  1. Use EpiDISH to solve cellular proportions and use post-hoc constraint.
library(EpiDISH)
outT <- epidish(beta.m = Y, ref.m = Ref, method = "RPC")
estProp_RB <- outT$estF

A word about Step 1

For step 1, one can also use findRefinx(..., sortBy = "cv") to select features based on coefficient of variantion. The choice of sortby = "cv" and sortBy = "var" depends on whether the feature variances of your data correlates with the means. For RNA-seq counts, the variance-mean correlation is strong, thus sortBy = "cv" is recommended. For log-counts, the variance-mean correlation largely disappears, so both sortBy = "cv" and sortBy = "var" would work similarly. In DNA methylation data, this correlation is not strong, either sortBy = "cv" or sortBy = "var" can be used. In this case, we recommend sortBy = "var" because we find it has better feature selection for DNA methylation data than sortBy = "cv" (unpublished results).

refinx = findRefinx(Y_raw, nmarker=1000, sortBy = "var")

4.2 Reference-free deconvolution using RefFreeEWAS

  1. Similar to Reference-based deconvolution we also select the top 1000 most variant features by findRefinx(). And then subset data.
refinx <- findRefinx(Y_raw, nmarker = 1000)
Y <- Y_raw[refinx,]
  1. Do reference-free deconvolution on the RA dataset.
K <- 6
outT <- myRefFreeCellMix(Y, mu0=myRefFreeCellMixInitialize(Y, K = K))
estProp_RF <- outT$Omega
  1. Comparing the reference-free method versus reference-base method
# first we align the cell types from RF 
# and RB estimations using pearson's correlation
estProp_RF <- assignCellType(input=estProp_RF,
                             reference=estProp_RB) 
mean(diag(cor(estProp_RF, estProp_RB)))
## [1] 0.1967946

4.3 Improve reference-free deconvolution with cross-cell type differential analysis

Feature selection is an important step before RF deconvolution and is directly related with the estimation quality of cell composition. findRefinx() and findRefinx(..., sortBy = "var") simply select the markers with largest CV or largest variance, which may not always result in a good selection of markers. Here, we propose to improve RF deconvolution marker selection through cross-cell type differential analysis. We implement two versions of such improvement, one is for DNA methylation microarray data using myRefFreeCellMix originally from R package RefFreeEWAS, the other one is for gene expression microarray data using deconf from CellMix package. To implement this, CellMix need to be installed first.

4.3.1 Improved-RF with myRefFreeCellMix

  1. Load TOAST package.
library(TOAST)
  1. Do reference-free deconvolution using improved-RF implemented with RefFreeCellMix. The default deconvolution function implemented in csDeconv() is RefFreeCellMix_wrapper(). Here, instead of a data matrix, Y_raw could also be a SummarizedExperiment object.
K=6
set.seed(1234)
outRF1 <- csDeconv(Y_raw, K, TotalIter = 30, bound_negative = TRUE) 
  1. Comparing udpated RF estimations versus RB results.
## check the accuracy of deconvolution
estProp_RF_improved <- assignCellType(input=outRF1$estProp,
                                      reference=estProp_RB) 
mean(diag(cor(estProp_RF_improved, estProp_RB)))
## [1] 0.2254084

A word about Step 2

For step 2, initial features (instead of automatic selection by largest variation) can be provided to function RefFreeCellMixT(). For example

refinx <- findRefinx(Y_raw, nmarker = 1000, sortBy = "cv")
InitNames <- rownames(Y_raw)[refinx]
csDeconv(Y_raw, K = 6, nMarker = 1000, 
         InitMarker = InitNames, TotalIter = 30)

A word about bounding the negative estimators

Since all the parameters represent the mean observation levels for each cell type, it may not be reasonable to have negative estimators. As such, we provide options to bound negative estimated parameters to zero through the bound_negative argument in csDeconv() function. Although we find bounding negative estimators has minimum impact on the performance, the users could choose to bound or not bound the negative values in the function. The default value for bound_negative is FALSE.

4.3.2 Improved-RF with use-defined RF function

In order to use other RF functions, users can wrap the RF function a bit first to make it accept Y (raw data) and K (number of cell types) as input, and return a N (number of cell types) by K proportion matrix. We take myRefFreeCellMix() as an example. Other deconvolution methods can be used similarly.

mydeconv <- function(Y, K){
     if (is(Y, "SummarizedExperiment")) {
          se <- Y
          Y <- assays(se)$counts
     } else if (!is(Y, "matrix")) {
          stop("Y should be a matrix
               or a SummarizedExperiment object!")
     }
     
     if (K<0 | K>ncol(Y)) {
         stop("K should be between 0 and N (samples)!")
     }
     outY = myRefFreeCellMix(Y, 
               mu0=myRefFreeCellMixInitialize(Y, 
               K = K))
     Prop0 = outY$Omega
     return(Prop0)
}
set.seed(1234)
outT <- csDeconv(Y_raw, K, FUN = mydeconv, bound_negative = TRUE)

4.4 Partial reference-free deconvolution (TOAST/-P and TOAST/+P)

Similar to DSA, our PRF method requires the knowledge of cell type-specific markers. Such markers can be selected from pure cell type gene expression profiles from same or different platforms (through function ChooseMarker()). They can also be manually specified (see function manual ?MDeconv for more explanation). The prior knowledge of cell compositions are optional, but highly recommended. We find prior knowledge of cell compositions (alpha and sigma) help calibrate the scales of the estimations, and reduce estimation bias. Such information can be estimated from previous cell sorting experiments or single cell study. We currently provide prior knowledge for five tissue types: “human pbmc”,“human liver”, “human brain”, “human pancreas”, “human skin”, which can be directly specified in MDeconv() function.

4.4.1 Choose cell type-specific markers

We provide functions to choose cell type-specific markers from pure cell type profiles or single cell RNA-seq data. Here we demonstrate how to select markers from PBMC pure cell type gene expression profile.

## create cell type list:
CellType <- list(Bcells = 1,
                 CD8T = 2,
                 CD4T = 3,
                 NK = 4,
                 Monocytes = 5)
## choose (up to 20) significant markers 
## per cell type
myMarker <- ChooseMarker(LM_5, 
                         CellType, 
                         nMarkCT = 20,
                         chooseSig = TRUE,
                         verbose = FALSE)
lapply(myMarker, head, 3)
## $Bcells
## [1] "BANK1"  "MS4A1"  "IGLL3P"
## 
## $CD8T
## [1] "CD8B" "CD8A" "GZMK"
## 
## $CD4T
## [1] "IL9"   "CTLA4" "IL3"  
## 
## $NK
## [1] "KIR3DL2" "IL18RAP" "KLRF1"  
## 
## $Monocytes
## [1] "FCN1"   "P2RY13" "NCF2"

4.4.2 PRF deconvolution without prior (TOAST/-P)

resCBS0 <- MDeconv(CBS_mix, myMarker,
                epsilon = 1e-3, verbose = FALSE)
## Deconvolution without prior information.
diag(cor(CBS_trueProp, t(resCBS0$H)))
## [1] 0.5333925 0.5647335 0.7027891 0.6484607 0.7116513
mean(abs(as.matrix(CBS_trueProp) - t(resCBS0$H)))
## [1] 0.1313198

4.4.3 PRF deconvolution with prior (TOAST/+P)

We allow manually input the prior knowledge of all cell types, or select from currently supported tissues (“human pbmc”,“human liver”, “human brain”, “human pancreas”, “human skin”). Note that order of cell types in prior knowledge here should match the order in marker list.

Here is an example of manually specifying alpha and sigma:

prior_alpha <- c(0.09475, 0.23471, 0.33232, 0.0969, 0.24132)
prior_sigma <- c(0.09963, 0.14418, 0.16024, 0.10064, 0.14556)
names(prior_alpha) <- c("B cells", "CD8T", "CD4T",
                        "NK cells", "Monocytes")
names(prior_sigma) <- names(prior_alpha) 

Here is to see alpha and sigma for supported tisuses using GetPrior():

thisprior <- GetPrior("human pbmc")
thisprior
## $alpha_prior
##   B cells      CD8T      CD4T  NK cells Monocytes 
##   0.09475   0.23471   0.33232   0.09690   0.24132 
## 
## $sigma_prior
##   B cells      CD8T      CD4T  NK cells Monocytes 
##   0.09963   0.14418   0.16024   0.10064   0.14556

Deconvolution using manually input alpha and sigma:

resCBS1 <- MDeconv(CBS_mix, myMarker,
                alpha = prior_alpha,
                sigma = prior_sigma,
                epsilon = 1e-3, verbose = FALSE)
## Deconvolution with prior infromation.
diag(cor(CBS_trueProp, t(resCBS1$H)))
## [1] 0.5545308 0.5627714 0.6302584 0.6841369 0.7101932
mean(abs(as.matrix(CBS_trueProp) - t(resCBS1$H)))
## [1] 0.079472

For supported tissues, you can directly specify tissue type as alpha input:

resCBS2 <- MDeconv(CBS_mix, myMarker,
                   alpha = "human pbmc",
                   epsilon = 1e-3, verbose = FALSE)
## Deconvolution with prior infromation.
diag(cor(CBS_trueProp, t(resCBS2$H)))
## [1] 0.5545308 0.5627714 0.6302584 0.6841369 0.7101932
mean(abs(as.matrix(CBS_trueProp) - t(resCBS2$H)))
## [1] 0.079472

4.5 Complete deconvolution using a geometric approach

Tsisal is a complete deconvolution method which estimates cell compositions from DNA methylation data without prior knowledge of cell types and their proportions. Tsisal is a full pipeline to estimate number of cell types, cell compositions, find cell-type-specific CpG sites, and assign cell type labels when (full or part of) reference panel is available.

Here is an example of manually specifying K and reference panel:

out = Tsisal(Ybeta,K = 4, knowRef = ref_m)
out$estProp[1:3,1:4]
head(out$selMarker)

Here is an example where both K and reference panel are unknown:

out = Tsisal(Ybeta,K = NULL, knowRef = NULL, possibleCellNumber = 2:5)
out$estProp[1:3,1:out$K]
head(out$selMarker)
out$K

Here is an example where K is unknown and reference panel is known:

out = Tsisal(Ybeta, K = NULL, knowRef = ref_m, possibleCellNumber = 2:5)
out$estProp[1:3,1:out$K]
head(out$selMarker)
out$K

5 Detect cell type-specific and cross-cell type differential signals

The csDE/csDM detection function requires a table of microarray or RNA-seq measurements from all samples, a table of mixing proportions, and a design vector representing the status of subjects.

We demonstrate the usage of TOAST in three common settings.

5.1 Detect cell type-specific differential signals under two-group comparison

  1. Assuming you have TOAST library and dataset loaded, the first step is to generate the study design based on the phenotype matrix. Note that all the binary (e.g. disease = 0, 1) or categorical variable (e.g. gender = 1, 2) should be transformed to factor class. Here we use the proportions estimated from step 4.3.1 as input proportion.
head(Pheno, 3)
##            age gender disease
## GSM1051525  67      2       1
## GSM1051526  49      2       1
## GSM1051527  53      2       1
design <- data.frame(disease = as.factor(Pheno$disease))

Prop <- estProp_RF_improved
colnames(Prop) <- colnames(Ref) 
## columns of proportion matrix should have names
  1. Make model design using the design (phenotype) data frame and proportion matrix.
Design_out <- makeDesign(design, Prop)
  1. Fit linear models for raw data and the design generated from Design_out(). Y_raw here is a data matrix with dimension P (features) by N (samples). Instead of a data matrix, Y_raw could also be a SummarizedExperiment object.
fitted_model <- fitModel(Design_out, Y_raw)
# print all the cell type names
fitted_model$all_cell_types
## [1] "CD8T"  "CD4T"  "NK"    "Bcell" "Mono"  "Gran"
# print all phenotypes
fitted_model$all_coefs
## [1] "disease"

TOAST allows a number of hypotheses to be tested using csTest() in two group setting.

5.1.1 Testing one parameter (e.g. disease) in one cell type.

For example, testing disease (patient versus controls) effect in Gran.

res_table <- csTest(fitted_model, 
                    coef = "disease", 
                    cell_type = "Gran")
## Test the effect of disease in Gran.
head(res_table, 3)
##                  beta   beta_var         mu effect_size f_statistics
## cg03999583 -0.9336162 0.03923521  1.2339749   -1.216966     22.21573
## cg04021544 -0.6798894 0.02899288  0.8520754   -1.327570     15.94356
## cg07755735  0.7132178 0.03332731 -0.1296480    3.142470     15.26315
##                 p_value        fdr
## cg03999583 9.065204e-06 0.02719561
## cg04021544 1.349669e-04 0.15106724
## cg07755735 1.831638e-04 0.15106724
Disease_Gran_res <- res_table

5.1.2 Testing one parameter in all cell types.

For example, testing the joint effect of age in all cell types:

res_table <- csTest(fitted_model, 
                    coef = "disease", 
                    cell_type = "joint")
head(res_table, 3)

Specifying cell_type as NULL or not specifying cell_type will test the effect in each cell type and the joint effect in all cell types.

res_table <- csTest(fitted_model, 
                    coef = "disease", 
                    cell_type = NULL)
lapply(res_table, head, 3)

## this is exactly the same as
res_table <- csTest(fitted_model, coef = "disease")

5.1.3 Testing one parameter in all cell types by incorporating DE/DM state correlation among cell types

Some cell types may show DE/DM state correlation. We can check the existence of such correlation by plotting the -log10 transformed p-value from TOAST result.

res_table <- csTest(fitted_model, coef = "disease",verbose = F)
pval.all <- matrix(NA, ncol= 6, nrow= nrow(Y_raw))
feature.name <- rownames(Y_raw)
rownames(pval.all) = feature.name
colnames(pval.all) = names(res_table)[1:6]
for(cell.ix in 1:6){
  pval.all[,cell.ix] <- res_table[[cell.ix]][feature.name,'p_value']
}
plotCorr(pval = pval.all, pval.thres = 0.05)
## Detect input of pval.thres; Use pval.thres to calculate odds ratio.
## -log10(pval) threshold for each cell type:
##  1.303 1.309 1.302 1.302 1.307 1.302