CellsPickMe

Maggie Fu

The University of British Columbia
maggie.fu@bcchr.ca

Introduction

CellsPickMe is a streamlined R package for cell type proportion prediction in a heterogeneous tissue (i.e. deconvolution) based on DNA methylation (DNAme) data. As cell identity is crucially linked to cell function and DNAme profiles, estimation of cell type proportion is instrumental to understanding major factors driving DNAme variability. Accounting for inter-individual cellular heterogeneity is also vital in epigenome-wide association studies to accurately evaluate the signals associated with the variables of interest.

The CellsPickMe package takes DNAme data generated from Illumina microarray and predict its cellular composition based on cell types available in the reference profiles. Notably, for each population of cells, the package “picks” DNAme features that best predict cellular identities with machine learning algorithms to improve deconvolution performance. Currently, the algorithm is compatible for peripheral blood, cord blood, saliva, and brain (neuron vs non-neuron). The table below illustrate the reference datasets available in the CellsPickMe package for deconvolution. We curated the UniBlood references (7, 13, and 19) to address the current challenge in deconvoluting longitudinal and pediatric blood samples. Refer to the [UniBlood Reference Creation][UniBlood Reference Creation] article for details of the UniBlood references.

Available reference datasets for tissues in CellsPickMe

Tissue	Reference dataset	ID for the getRef() function	DOI / Bioconductor Package
Cord and adult peripheral blood	UniBlood7	UniBlood7	Available soon
Cord and adult peripheral blood	UniBlood13	UniBlood13	Available soon
Cord and adult peripheral blood	UniBlood19	UniBlood19	Available soon
Adult peripheral blood	IDOL extended (2022)	Extended	LA Salas et al. (2022) Enhanced cell deconvolution of peripheral blood using DNA methylation for high-resolution immune profiling. Nat Commun. 13, 761. doi: 10.1038/s41467-021-27864-7 FlowSorted.BloodExtended.EPIC
Adult peripheral blood	IDOL (2016)	IDOL	DC Koestler et al. (2016). Improving cell mixture deconvolution by identifying optimal DNA methylation libraries (IDOL). BMC bioinformatics. 17, 120. doi: 10.1186/s12859-016-0943-7 FlowSorted.Blood.EPIC
Adult peripheral blood	Reinius (2012)	Reinius	LE Reinius et al. (2012). Differential DNA methylation in purified human blood cells: implications for cell lineage and studies on disease susceptibility. PloS one. 7(7), e41361. doi: 10.1371/journal.pone.0041361 FlowSorted.Blood.450k
Cord blood	Cord (2018)	Cord	K Gervin et al. (2019). Systematic evaluation and validation of reference and library selection methods for deconvolution of cord blood DNA methylation data. Clinical epigenetics. 11, 1-15. doi: 10.1186/s13148-019-0717-y FlowSorted.CordBlood.450k
Saliva	Middleton (2022)	Middleton	LY Middleton et al. (2022). Saliva cell type DNA methylation reference panel for epidemiological studies in children. Epigenetics, 17(2), 161-177. doi: 10.1371/journal.pone.0041361 BeadSorted.Saliva.EPIC
Brain	DLPFC (2012)	DLPFC	J Guintivano et al. (2013). A cell epigenotype specific model for the correction of brain cellular heterogeneity bias and its application to age, brain region and major depression. Epigenetics, 8(3), 290-302. doi: 10.4161/epi.23924 FlowSorted.DLPFC.450k

The figure below focused on our optimization of the deconvolution pipeline in the blood tissue. The four main steps of cellular deconvolution are reference selection, data normalization, feature selection, and regression. The function that calls each of the step is bolded on the left, and the detail of each function is described in the Usage section.

CellsPickMe Pipeline, created with BioRender.com

Cell Type Deconvolution

In addition to improved methods for feature selection to optimize prediction process, the package also includes several features to assess prediction accuracy.

The CellsPickMe package includes 5 core functions:

• getRef() finds and load the tissue-specific reference data set
• combData()combines the reference and the user’s sample data sets together
• pickProbes() picks features that best distinguish cell types (multiple feature selection methods available)
• clusterScore() assesses the performance of selected features in clustering cell types in the reference
• predictCT() estimate cell type proportions based on selected probes, with the ability to evaluate

The pickProbes() function can be computationally intensive, depending on the methods for feature selection. Parallel computing is available and highly recommended for improved efficiency. Detailed usage is available here

The following data are required for performing the cell type proportion prediction with CellsPickMe: - DNAme data (RGChannelSet Object is recommended, beta matrix is accepted as well) * An appropriate reference can be obtained with the getRef() function. Otherwise the user can specify their own reference dataset as well (needs to be a RGChannelSet Object with CellTypes information in the pData)

---

Installation

library(devtools)
devtools::install_github("maggie-fu/CellsPickMe")

Usage

The CellsPickMe package takes DNA methylation data generated from Illumina microarray and predict its cellular composition based on cell types available in the reference profiles. Currently, the algorithm is compatible for peripheral blood, cord blood, saliva, and brain (neuron vs non-neuron).

Obtain reference dataset

From the available list of reference datasets (“Reinius”, “IDOL”, “Extended”, “Cord”, “UniBlood7”, “UniBlood13”, “UniBlood19”, “DLPFC”, and “Middleton”), select one that is appropriate for your sample (based on tissue and sample age)

library(CellsPickMe)

# Request the UniBlood19 reference with no normalization
ref_dat <- getRef(ref = "IDOL", normType = "None")

Normalize sample and reference datasets together

Normalize user’s sample and reference data sets together to reduce batch effect and improve prediction accuracy. Option for normType includes “Noob”, “Funnorm”, “Quantile”, “Quantile.b”, “None”, with the first 3 options being exclusively for RGChannelSet objects, and Quantile.b for beta matrix.

# Load example blood cell mixture, subsetted from the IDOL dataset (GSE110554)
test_dat <- CellsPickMe::IDOL_mixed_cells
 
# Combine sample and reference data sets together, followed by normalization (if selected)
comb_dat <- combData(dataset = test_dat, 
                     reference = ref_dat$reference, 
                     cellTypes = ref_dat$cellTypes, 
                     class = "rgset",               #c("rgset", "betas")
                     normType = "None")             # c("Noob", "Funnorm", "Quantile", "Quantile.b", "None")

Pick features that best distinguish cell type

The CellsPickMe package supports feature selection with either the traditional T-test, or with machine-learning-based methods such as elastic net and random forest to obtain a curated list of features that are highly predictive of cell types.

Method	ID for the caretMods argument	Description
LASSO	lasso	Use L1 regularization to optimize shrinkage of coefficients uncorrelated with outcome of interest.
elastic net	EN	Elastic Net is a linear regression-based method that can perform feature selection by using L1 and L2 regularization.
random forest	RF	Random Forest is a decision tree-based ensemble method that can perform feature selection by evaluating the importance of each feature in the model. The Gini importance or Mean Decrease Impurity (MDI) are two ways to calculate feature importance in Random Forest.
boosted logistic regression	BLR	Boosted logistic regression is a machine learning algorithm that combines logistic regression with boosting, a technique for iteratively improving the performance of weak learners. # In boosted logistic regression, a weak logistic regression model is trained on the data, and then additional weak logistic regression models are trained to correct the errors of the previous models. Each subsequent model focuses more on the data points that were misclassified by the previous models, thus improving the overall accuracy of the model.
classification and regression tree	CART	Decision Trees can perform feature selection by evaluating the information gain or Gini impurity of each feature when deciding how to split the data. The important features are those that lead to the highest reduction in impurity or the highest information gain.
gradient boosted machine	GBM	GBM is another ensemble-based method that can perform feature selection by iteratively building trees that focus on predicting the errors of previous trees. In this process, GBM can identify important features that improve the model’s performance.

# Pick probes with T tests
probes <- pickProbes(dataNormed = comb_dat, 
                     probeList = "Ttest", #c("Ttest", "Caret_CV", "Caret_LOOCV", "IDOL")
                     probeSelect = "both", #c("both", "any")
                     nProbes = 100, # number of probes to pick for each cell type
                     p.val = 0.05,  # max pval
                     min.delta.beta = 0.05, # min delta beta
                     plotRef = T, # plot heatmap?
                     verbose = T)

### Set up server for parallelization - run the code if picking probes with Caret
library(doParallel)
cl <- makeCluster(detectCores() - 1) # change as needed
registerDoParallel(cl)

# Pick probes with repeated cross validation with lasso and elastic net
probes <- pickProbes(dataNormed = comb_dat, 
                     probeList = "Caret_LOOCV", #c("Ttest", "Caret_CV", "Caret_LOOCV", "IDOL")
                     caretMods = c("lasso", "EN"),  #c("lasso", "EN", "BLR", "CART", "RF", "GBM")
                     filterK = 1000, # number of probes to put into the predictor for each cell type
                     seed = 1, 
                     plotRef = F, # plot heatmap?
                     verbose = F)

# FilterK results used as input for feature selection
head(probes$coefs$probeList$CD8T$tTestTopK)

# Picked probes with EN and the estimated coefficients
head(probes$coefs$probeCoefs$EN)

# Picked probes with EN for each cell type and the cross validation performance 
head(probes$coefs$probeList$CD8mem$EN$coefs)

Assess clustering stability

To further evaluate the performance of the selected probes, pvClust can be applied to assess whether the picked probes can be used to generate the correct cluster (cell type labeling) in reference data

clustAU <- identClust(dataNormed = comb_dat,
                      probes = probes,
                      parallel = TRUE)

Estimate cell type proportion

Finally, the selected features are used to estimate cell type proportions in the sample data set. We also incorporate the CETYGO score (see Reference) to estimate prediction performance even in the absence of a validation cohort / ground truth cell count.

out <- predictCT(dataNormed = comb_dat, 
                 probes = probes, 
                 method = "CP",  #c("CP", "RPC", "SVR")
                 removenRBC = F, # remove nRBC?
                 verbose = T, 
                 cetygo = T) # CETYGO to assess reference appropriateness (RMSE evaluation)

Citation

The manuscript detailing CellsPickMe and its use is currently under preparation. For more information about this please contact Maggie Fu at maggie.fu@bcchr.ca.

References

Depending on the options you used, please consider citing the following references as this package is built on their data / code / papers.

1. DS Vellame et al. (2023). Uncertainty quantification of reference-based cellular deconvolution algorithms. Epigenetics 18, 1: 2137659. doi: 10.1080/15592294.2022.2137659

   • Please cite this paper if you set the parameter cetygo = T for predictCT()

2. LE Reinius et al. (2012). Differential DNA methylation in purified human blood cells: implications for cell lineage and studies on disease susceptibility. PloS one. 7(7), e41361. doi: 10.1371/journal.pone.0041361

   • Please cite this paper if you used this reference dataset, i.e. getRef(ref = "Reinius")

3. DC Koestler et al. (2016). Improving cell mixture deconvolution by identifying optimal DNA methylation libraries (IDOL). BMC bioinformatics. 17, 120. doi: 10.1186/s12859-016-0943-7

   • Please cite this paper if you used this reference dataset, i.e. getRef(ref = "IDOL")

4. LA Salas et al. (2022) Enhanced cell deconvolution of peripheral blood using DNA methylation for high-resolution immune profiling. Nat Commun. 13, 761. doi: 10.1038/s41467-021-27864-7

   • Please cite this paper if you used this reference dataset, i.e. getRef(ref = "Extended")

5. K Gervin et al. (2019). Systematic evaluation and validation of reference and library selection methods for deconvolution of cord blood DNA methylation data. Clinical epigenetics. 11, 1-15. doi: 10.1186/s13148-019-0717-y

   • Please cite this paper if you used this reference dataset, i.e. getRef(ref = "Cord")

6. LY Middleton et al. (2022). Saliva cell type DNA methylation reference panel for epidemiological studies in children. Epigenetics, 17(2), 161-177. doi: 10.1371/journal.pone.0041361

   • Please cite this paper if you used this reference dataset, i.e. getRef(ref = "Middleton")

7. J Guintivano et al. (2013). A cell epigenotype specific model for the correction of brain cellular heterogeneity bias and its application to age, brain region and major depression. Epigenetics, 8(3), 290-302. doi: 10.4161/epi.23924

   • Please cite this paper if you used this reference dataset, i.e. getRef(ref = "DLPFC")

8. TJ Triche, et al. (2013). Low-level processing of Illumina Infinium DNA Methylation BeadArrays. Nucleic Acids Res. 41, e90. doi: 10.1093/nar/gkt090.

   • Please cite this paper if you used this normalization method, i.e. getRef(normType = "Noob") or combData(normType = "Noob")

9. JP Fortin et al. (2014). Functional normalization of 450k methylation array data improves replication in large cancer studies. Genome Biology 15, 503. doi: 10.1186/s13059-014-0503-2.

   • Please cite this paper if you used this normalization method, i.e. getRef(normType = "Funnorm") or combData(normType = "Funnorm")

10. N Touleimat and J Tost. (2012). Complete pipeline for Infinium Human Methylation 450K BeadChip data processing using subset quantile normalization for accurate DNA methylation estimation. Epigenomics 4, 325-341. doi: 10.2217/epi.12.21

    • Please cite this paper if you used this normalization method, i.e. getRef(normType = "Quantile") or combData(normType = "Quantile")