1 Introduction

There are several approaches available to adjust for differents in the relative proportion of cell types in whole blood measured from DNA methylation (DNAm). For example, reference-based approaches require the use of reference data sets made up of purified cell types to identify cell type-specific DNAm signatures. These cell type-specific DNAm signatures are used to estimate the relative proportions of cell types directly, but these reference data sets are laborious and expensive to collect. Furthermore, these reference data sets will need to be continuously collected over time as new platform technologies emerge measuring DNAm because the observed methylation levels for the same CpGs in the same sample vary depending the platform technology.

In contrast, there are reference-free approaches, which are based on methods related to surrogate variable analysis or linear mixed models. These approaches do not provide estimates of the relative proportions of cell types, but rather these methods just remove the variability induced from the differences in relative cell type proportions in whole blood samples.

Here, we present a statistical model that estimates the cell composition of whole blood samples measured from DNAm. The method can be applied to microarray or sequencing data (for example whole-genome bisulfite sequencing data, WGBS, reduced representation bisulfite sequencing data, RRBS). Our method is based on the idea of identifying informative genomic regions that are clearly methylated or unmethylated for each cell type, which permits estimation in multiple platform technologies as cell types preserve their methylation state in regions independent of platform despite observed measurements being platform dependent.

2 Getting Started

Load the methylCC R package and other packages that we’ll need later on.

library(FlowSorted.Blood.450k)
library(methylCC)
library(minfi)
library(tidyr)
library(dplyr)
library(ggplot2)

3 Data

3.1 Whole Blood Illumina 450k Microarray Data Example

# Phenotypic information about samples
head(pData(FlowSorted.Blood.450k))
## DataFrame with 6 rows and 8 columns
##          Sample_Name      Slide       Array               Basename    SampleID
##          <character>  <numeric> <character>            <character> <character>
## WB_105        WB_105 5684819001      R01C01 idat/5684819001_R01C01         105
## WB_218        WB_218 5684819001      R02C01 idat/5684819001_R02C01         218
## WB_261        WB_261 5684819001      R03C01 idat/5684819001_R03C01         261
## PBMC_105    PBMC_105 5684819001      R04C01 idat/5684819001_R04C01         105
## PBMC_218    PBMC_218 5684819001      R05C01 idat/5684819001_R05C01         218
## PBMC_261    PBMC_261 5684819001      R06C01 idat/5684819001_R06C01         261
##          CellTypeLong    CellType         Sex
##           <character> <character> <character>
## WB_105    Whole blood         WBC           M
## WB_218    Whole blood         WBC           M
## WB_261    Whole blood         WBC           M
## PBMC_105         PBMC        PBMC           M
## PBMC_218         PBMC        PBMC           M
## PBMC_261         PBMC        PBMC           M
# RGChannelSet
rgset <- FlowSorted.Blood.450k[,
                pData(FlowSorted.Blood.450k)$CellTypeLong %in% "Whole blood"]

4 Using the estimatecc() function

4.1 Input for estimatecc()

The estimatecc() function must have one object as input:

  1. an object such as an RGChannelSet from the R package minfi or a BSseq object from the R package bsseq. This object should contain observed DNAm levels at CpGs (rows) in a set of \(N\) whole blood samples (columns).

4.2 Running estimatecc()

In this example, we are interested in estimating the cell composition of the whole blood samples listed in the FlowSorted.Blood.450k R/Bioconductor package. To run the methylcC::estimatecc() function, just provide the RGChannelSet. This will create an estimatecc object. We will call the object est.

set.seed(12345)
est <- estimatecc(object = rgset) 
est
## estimatecc: Estimate Cell Composition of Whole Blood 
##                 Samples using DNA methylation
##    Input object class:  RGChannelSet 
##    Reference cell types:  Gran CD4T CD8T Bcell Mono NK 
##    Number of Whole Blood Samples:  6 
##    Name of Whole Blood Samples:  WB_105 WB_218 WB_261 WB_043 WB_160 WB_149

To see the cell composition estimates, use the cell_counts() function.

cell_counts(est)
##             Gran       CD4T       CD8T      Bcell       Mono         NK
## WB_105 0.4242292 0.16915420 0.09506568 0.04187765 0.08357502 0.18609822
## WB_218 0.4906710 0.15471447 0.00000000 0.04979116 0.14346117 0.16136217
## WB_261 0.5476117 0.11895815 0.14007846 0.01725995 0.08869797 0.08739378
## WB_043 0.5038143 0.12420228 0.08031593 0.06515287 0.07218653 0.15432807
## WB_160 0.6803254 0.07139726 0.04965732 0.00000000 0.09526148 0.10335854
## WB_149 0.5375962 0.14902349 0.10814235 0.03227085 0.06111685 0.11185025

4.3 Compare to minfi::estimateCellCounts()

We can also use the estimateCellCounts() from R/Bioconductor package minfi to estimate the cell composition for each of the whole blood samples.

sampleNames(rgset) <- paste0("Sample", 1:6)

est_minfi <- minfi::estimateCellCounts(rgset)
est_minfi
##               CD8T      CD4T          NK      Bcell       Mono      Gran
## Sample1 0.13967126 0.1581874 0.137528672 0.07040633 0.06383445 0.4835306
## Sample2 0.05797617 0.1751543 0.072686689 0.09859270 0.12429750 0.5228217
## Sample3 0.12091718 0.1531062 0.029632651 0.05447982 0.06775822 0.6064806
## Sample4 0.10438514 0.1709784 0.024322195 0.11447040 0.05233508 0.5700027
## Sample5 0.03775465 0.1465998 0.003996696 0.04767462 0.07452444 0.7069746
## Sample6 0.06568804 0.1873355 0.054344189 0.07039282 0.05196750 0.5932074

Then, we can compare the estimates to methylCC::estimatecc().

df_minfi = gather(cbind("samples" = rownames(cell_counts(est)),
                        as.data.frame(est_minfi)),
                  celltype, est, -samples)

df_methylCC = gather(cbind("samples" = rownames(cell_counts(est)),
                           cell_counts(est)),
                     celltype, est, -samples)

dfcombined <- full_join(df_minfi, df_methylCC, 
                               by = c("samples", "celltype"))

ggplot(dfcombined, aes(x=est.x, y = est.y, color = celltype)) +
    geom_point() + xlim(0,1) + ylim(0,1) +
    geom_abline(intercept = 0, slope = 1) +
    xlab("Using minfi::estimateCellCounts()") + 
    ylab("Using methylCC::estimatecc()") +
    labs(title = "Comparing cell composition estimates")

We see the estimates closely match for the six cell types.

5 SessionInfo

sessionInfo()
## R version 4.0.0 RC (2020-04-19 r78255)
## Platform: x86_64-pc-linux-gnu (64-bit)
## Running under: Ubuntu 18.04.4 LTS
## 
## Matrix products: default
## BLAS:   /home/biocbuild/bbs-3.12-bioc/R/lib/libRblas.so
## LAPACK: /home/biocbuild/bbs-3.12-bioc/R/lib/libRlapack.so
## 
## locale:
##  [1] LC_CTYPE=en_US.UTF-8       LC_NUMERIC=C              
##  [3] LC_TIME=en_US.UTF-8        LC_COLLATE=C              
##  [5] LC_MONETARY=en_US.UTF-8    LC_MESSAGES=en_US.UTF-8   
##  [7] LC_PAPER=en_US.UTF-8       LC_NAME=C                 
##  [9] LC_ADDRESS=C               LC_TELEPHONE=C            
## [11] LC_MEASUREMENT=en_US.UTF-8 LC_IDENTIFICATION=C       
## 
## attached base packages:
## [1] stats4    parallel  stats     graphics  grDevices utils     datasets 
## [8] methods   base     
## 
## other attached packages:
##  [1] IlluminaHumanMethylation450kanno.ilmn12.hg19_0.6.0
##  [2] IlluminaHumanMethylation450kmanifest_0.4.0        
##  [3] ggplot2_3.3.0                                     
##  [4] dplyr_0.8.5                                       
##  [5] tidyr_1.0.2                                       
##  [6] methylCC_1.3.0                                    
##  [7] FlowSorted.Blood.450k_1.25.0                      
##  [8] minfi_1.35.0                                      
##  [9] bumphunter_1.31.0                                 
## [10] locfit_1.5-9.4                                    
## [11] iterators_1.0.12                                  
## [12] foreach_1.5.0                                     
## [13] Biostrings_2.57.0                                 
## [14] XVector_0.29.0                                    
## [15] SummarizedExperiment_1.19.0                       
## [16] DelayedArray_0.15.0                               
## [17] matrixStats_0.56.0                                
## [18] Biobase_2.49.0                                    
## [19] GenomicRanges_1.41.0                              
## [20] GenomeInfoDb_1.25.0                               
## [21] IRanges_2.23.0                                    
## [22] S4Vectors_0.27.0                                  
## [23] BiocGenerics_0.35.0                               
## [24] knitr_1.28                                        
## [25] BiocStyle_2.17.0                                  
## 
## loaded via a namespace (and not attached):
##   [1] colorspace_1.4-1          bsseq_1.25.0             
##   [3] ellipsis_0.3.0            siggenes_1.63.0          
##   [5] mclust_5.4.6              base64_2.0               
##   [7] farver_2.0.3              bit64_0.9-7              
##   [9] AnnotationDbi_1.51.0      xml2_1.3.2               
##  [11] R.methodsS3_1.8.0         codetools_0.2-16         
##  [13] splines_4.0.0             scrime_1.3.5             
##  [15] Rsamtools_2.5.0           annotate_1.67.0          
##  [17] dbplyr_1.4.3              R.oo_1.23.0              
##  [19] HDF5Array_1.17.0          BiocManager_1.30.10      
##  [21] readr_1.3.1               compiler_4.0.0           
##  [23] httr_1.4.1                assertthat_0.2.1         
##  [25] Matrix_1.2-18             limma_3.45.0             
##  [27] htmltools_0.4.0           prettyunits_1.1.1        
##  [29] tools_4.0.0               gtable_0.3.0             
##  [31] glue_1.4.0                GenomeInfoDbData_1.2.3   
##  [33] rappdirs_0.3.1            doRNG_1.8.2              
##  [35] Rcpp_1.0.4.6              vctrs_0.2.4              
##  [37] multtest_2.45.0           preprocessCore_1.51.0    
##  [39] nlme_3.1-147              rtracklayer_1.49.0       
##  [41] DelayedMatrixStats_1.11.0 xfun_0.13                
##  [43] stringr_1.4.0             plyranges_1.9.0          
##  [45] lifecycle_0.2.0           gtools_3.8.2             
##  [47] rngtools_1.5              XML_3.99-0.3             
##  [49] beanplot_1.2              zlibbioc_1.35.0          
##  [51] MASS_7.3-51.6             scales_1.1.0             
##  [53] BSgenome_1.57.0           hms_0.5.3                
##  [55] rhdf5_2.33.0              GEOquery_2.57.0          
##  [57] RColorBrewer_1.1-2        yaml_2.2.1               
##  [59] curl_4.3                  memoise_1.1.0            
##  [61] biomaRt_2.45.0            reshape_0.8.8            
##  [63] stringi_1.4.6             RSQLite_2.2.0            
##  [65] genefilter_1.71.0         permute_0.9-5            
##  [67] GenomicFeatures_1.41.0    BiocParallel_1.23.0      
##  [69] rlang_0.4.5               pkgconfig_2.0.3          
##  [71] bitops_1.0-6              nor1mix_1.3-0            
##  [73] evaluate_0.14             lattice_0.20-41          
##  [75] purrr_0.3.4               Rhdf5lib_1.11.0          
##  [77] labeling_0.3              GenomicAlignments_1.25.0 
##  [79] bit_1.1-15.2              tidyselect_1.0.0         
##  [81] plyr_1.8.6                magrittr_1.5             
##  [83] bookdown_0.18             R6_2.4.1                 
##  [85] magick_2.3                DBI_1.1.0                
##  [87] withr_2.2.0               pillar_1.4.3             
##  [89] survival_3.1-12           RCurl_1.98-1.2           
##  [91] tibble_3.0.1              crayon_1.3.4             
##  [93] BiocFileCache_1.13.0      rmarkdown_2.1            
##  [95] progress_1.2.2            grid_4.0.0               
##  [97] data.table_1.12.8         blob_1.2.1               
##  [99] digest_0.6.25             xtable_1.8-4             
## [101] R.utils_2.9.2             illuminaio_0.31.0        
## [103] openssl_1.4.1             munsell_0.5.0            
## [105] askpass_1.1               quadprog_1.5-8