DNA methylation at CpG dinucleotides is one of the most extensively studied epigenetic marks. With technological advancements, geneticists can profile DNA methylation with multiple reliable approaches. However, different profiling platforms can differ substantially in the CpGs they assess, consequently hindering integrated analysis across platforms. Here, we present CpG impUtation Ensemble (CUE), which leverages multiple classical statistical and modern machine learning methods, to impute from the Illumina HumanMethylation450 (HM450) BeadChip to the Illumina HumanMethylationEPIC (HM850) BeadChip.

CUE is maintained by Gang Li [[email protected]] and Laura Raffield [[email protected]].

• Version 0.8.9 released

From this study, we provide two sets of imputation models: one for whole blood and the other for placenta. Investigators can therefore complete their own imputation of placental or whole blood HM850 CpG sites using their own HM450 data, without access to their own reference panel. Our method is also applicable to imputation in other tissues, provided the user can provide a reference dataset assayed on both HM450 and HM850.

CUE package can be directly installed as following:

git clone https://github.com/GangLiTarheel/CUE.git
cd CUE

Please install the following version of R and R packages before performing imputation.

• R version 3.6.0 (2019-04-26)

• class_7.3-15

• xgboost_0.82.1

• randomForest_4.6-14

• parallel

Note: to install an R packages, simply run install.packages("Name_of_package"). For example,

install.packages("xgboost")

Note: you can also install the version of xgboost we used from the source code: https://cran.r-project.org/src/contrib/Archive/xgboost/xgboost_0.82.1.tar.gz

library(randomForest) # RF
library(xgboost) # XGBoost
library("class") # KNN
source("R/refund_lib.R")

library(parallel) # Use this package if you need parallel computing to accelerate computation speed

Note: The compressed pre-trained models need around ~ 58 GB (ELGAN ~36 GB; and PTSD ~21 GB) memory of storage.

Please save the above pretrained models (tar.gz files) in the same directory of your CUE packages.

To extract the pre-trained models:

Make two directories for two sets of pre-trained models:

pwd # this should be your root directory of CUE
mkdir PTSD
mkdir ELGAN

Download the pre-trained models amd move the PTSD and ELGAN pretrained models to the corresponding folder.

ftp://yunlianon:[email protected]/CUE/PTSD_model.tar.gz

ftp://yunlianon:[email protected]/CUE/ELGAN_model.tar.gz

cd PTSD
tar -xf PTSD_model.tar.gz
cd ../

cd ELGAN
tar -xf ELGAN_model.tar.gz
cd ../

Here we show the imputation for a toy dataset with three samples, 248K HM450K probes. The reason we use only 248K probes is that when we trained models, we only retain the 248,421 HM450 CpG sites (sites overlapping between ELGAN and PTSD, and without missingness in our samples) to train since we don't want the incompleteness or the pre-imputation on HM450 affects our evaluations on HM850 imputation. The full list of 248K HM450K probes can be found in Probes.RData.

## Set the working directory
setwd(“/YOUR_DIRECTORY/CUE”)
# Replace YOUR_DIRECTORY with your directory where you downloaded CUE.

root_dir = getwd()


## load sample dataset
load("PTSD/Sample_Dataset.RData")
X<-sample_data                            # Input HM450
m<-dim(X)[2] # number of samples          # Number of smaples

## load required datasets 
load("PTSD/Annotations.RData")                # load Annotation data
load("Data/PTSD.neighbors.RData")             # load All neighbors
load("Data/Probes.RData")                     # load All probes' names

## load required datasets for PTSD imputation
load("Data/PTSD_CpG_Best_method_list.RData")  # load the best method list 
load("PTSD/RF/best_RF.RData")                 # load neighbors for RF
load("PTSD/XGB/best_XGB.RData")               # load neighbors for XGB
load("PTSD/KNN/best_KNN.RData")               # load neighbors for KNN
load("PTSD/TCR/best_TCR.RData")               # load neighbors for PFR

## load all the required packages
library(randomForest)
library(xgboost)
library(xgboost)
source("R/refund_lib.R")

## load the CUE imputation function
source("R/impute.R")

## Check the input
CUE_check(X)                        # check if the inuput X match the requirements

## TCR required input	
temp <- TCR.input(X)
X_logit = temp[[1]]                 # Logit transformation make it better follows Gaussian distribution
test.funcs <- temp[[2]]             # Pre-calculated sample-specific density function

## Imputation (Single CPU might takes 4-5 days to imputation for a middle sized methylation dataset (n=~100 samples).)
m.imputed<-CUE.impute(X,m,"PTSD")

## Save the imputed probes as y_impute.RData
setwd(root_dir)
save(m.imputed,file="y_impute.RData")

Note: (a) we impute all 339K HM850 specific probes which had complete data in our reference whole blood and placenta datasets. Users of CUE must use the following quality control steps to retain the well imputed probes only for use in subsequent analysis (such as epigenome wide association studies).

(b) we provided the corresponding annotation files within PTSD datasets ("PTSD/Annotations.RData") for building funtional predicitors for penalized function regression models. The annntation files can also be downloaded from the Illumina website.

File link: HM850

ftp://webdata2:[email protected]/downloads/productfiles/methylationEPIC/infinium-methylationepic-v-1-0-b4-manifest-file-csv.zip

HM450

We provdies two sets of QC+ probes list for two datsets with the following thresholds:

QC for ELGAN: RMSE<0.1 and Accruracy > 90%;

QC for PTSD: RMSE<0.05 and Accruracy > 95%.

QC_probes<-read.csv("Placenta (ELGAN)QC_probe_list.csv") # for Placenta
#QC_probes<-read.csv("Whole Blood (PTSD)QC_probe_list") # for whole blood
load("y_impute.RData")
QC_probes<-unlist(QC_probes)
m.QC <- m.imputed[,paste(QC_probes)]
save(m.QC, file="m.imputed.QC.RData")

(Optional): We also provide an shiny app, CUE_QC.R, an visualization tool to select the QC+ probe list.

runApp('CUE_QC.R')

The shiny app will save the list of well imputed probes. Users can use the previous code to subset the output probes.

The final QC+ imputed HM850 probes will be saved as "m.imputed.QC.RData". See the output:

## Output : DNA methylation matrix
dim(m.QC)
# [1]      3 238090
head(m.QC[,1:5])
#          cg16269199 cg22519184  cg24063007 cg10692041 cg02339369
# sample_1  0.7452506  0.3634424 0.012868836  0.9411898  0.9493930
# sample_2  0.7341876  0.3567242 0.006375319  0.9145329  0.9557632
# sample_3  0.7422996  0.3668391 0.012483532  0.9411898  0.9493930

Li G, Raffield L, Logue M, Miller MW, Santos HP Jr, O'Shea TM, Fry RC, Li Y. CUE: CpG impUtation ensemble for DNA methylation levels across the human methylation450 (HM450) and EPIC (HM850) BeadChip platforms. Epigenetics. 2021 Aug;16(8):851-861. doi: 10.1080/15592294.2020.1827716. Epub 2020 Oct 4. PMID: 33016200; PMCID: PMC8330997.