Genome-Wide DNA Methylation Sites Screening by Use of Training and Testing Samples

Description

ttScreening is a screening process to filter out non-informative DNA methylation sites by applying (ordinary or robust) linear regressions to training data, and the results are further examined using testing samples. Surrogate variables are included to account for unknown factors. This package also contains two additional functions, Rare_Screening and firth_screening, for predictor screening in high-dimensional settings when the outcome is binary and is rare (binary with low prevalence). These methods were developed to handle challenges while screening predictors such as separation and small event counts and high-dimensionality, which are common in genomic and epigenomic studies.

Details

This package utilizes training and testing samples to filter out uninformative DNA methylation sites for continuous and rare binary outcomes. Surrogate variables (SVs) of DNA methylation are included in the filtering process to explain unknown factor effects.

Author(s)

Meredith Ray, Mohammad Abrar, Yu Jiang, Xin Tong, Hongmei Zhang

Maintainer: Meredith Ray <maray@memphis.edu>

References

Ray MA, Tong X, Lockett GA, Zhang H, and Karmaus WJJ. (2016) “An Efficient Approach to Screening Epigenome-Wide Data”, BioMed Research International.

Leek JT and Storey JD. (2007) “Capturing heterogeneity in gene expression studies by ‘Surrogate Variable Analysis’.” PLoS Genetics, 3: e161.

Rare_Screening: resampling-based screening with limma

Description

Rare_Screening: a resampling-based method that repeatedly samples cases and controls with replacement and applies bayes moderated linear modeling (limma) to identify predictors that are consistently significant. The function runs a rare-event resampling procedure on user provided data and returns the selected predictors along with iteration-wise selection counts.

Dr. Abrar maintains and is the author of this functions, for questions contact him at abrar0708@gmail.com.

Usage

Rare_Screening(predictor_list, Outcome, iteration, cut)

Arguments

Value

A list with:

• final_selection: character vector of selected predictor IDs (selection count >= cut)

• counts: integer vector of selection counts for all predictors (names = predictors)

• sel_mat: p x iteration matrix of 0/1 selections per iteration

• selected: integer indices of the predictors that met cut

Examples

## Not run: 
# predictor_list: n x p matrix; Outcome: 0/1 vector of length n
res <- Rare_Screening(predictor_list, Outcome, iteration = 100, cut = 70)
head(res$final_selection)

## End(Not run)

ttScreening with Firth bias-reduced logistic regression + multiple testing

Description

firth_screening: a two-stage approach combining ttScreening with Firth’s bias-reduced logistic regression. This method is designed to handle small sample sizes and rare outcomes by reducing bias in logistic regression coefficients.

Two-stage screening: (1) Pre-screen each predictor with bias-reduced logistic regression (brglm2). (2) For predictors passing pre-screen, run repeated train/test splits and count how often the test p-value < 0.05. Also report Bonferroni and FDR sets from the pre-screen p-values.

Dr. Abrar maintains and is the author of this functions, for questions contact him at abrar0708@gmail.com.

Usage

firth_screening(
  predictor_list,
  Outcome,
  iteration,
  train_frac = 0.67,
  alpha = 0.05,
  firth_count_threshold = 50,
  verbose = TRUE
)

Arguments

Value

A list with:

• firth_tt: character vector of predictor IDs selected by repeated ttScreening (count >= firth_count_threshold)

• Bonferroni: character vector selected by Bonferroni-adjusted p < 0.05 (pre-screen)

• FDR: character vector selected by FDR-adjusted p < 0.05 (pre-screen)

• prescreen_p: named numeric vector of pre-screen p-values

• counts: named integer vector of ttScreening success counts

Examples

## Not run: 
res <- firth_screening(
  predictor_list, Outcome,
  iteration = 100, train_frac = 0.67, alpha = 0.05, firth_count_threshold = 50
)
length(res$firth_tt); head(res$firth_tt)

## End(Not run)

Adjusted irwsva.build which builds surrogate variables from gene expression data

Description

This function is directly modified from the original irwsva.build() in the SVA package. It was noticed that under certain circumstances a subscript out of bounds error would occur while running the SVA function. Therefore, this modified code has a single line altered that conditionally uses the generic singular decomposition, svd(), instead of fast singular decomposition, fast.svd().

Usage

irwsva.build2(dat, mod, mod0 = NULL, n.sv, B = 5)

Arguments

Details

See http://www.bioconductor.org/packages/release/bioc/manuals/sva/man/sva.pdf

Value

Note

sva Vignette http://www.biostat.jhsph.edu/~jleek/sva/

Author(s)

Original irwsva.build: Jeffrey T. Leek <jleek@jhsph.edu>, John Storey jstorey@princeton.edu

References

Original sva: Leek JT and Storey JD. (2008) A general framework for multiple testing dependence.Proceedings of the National Academy of Sciences, 105: 18718-18723.

Leek JT and Storey JD. (2007) Capturing heterogeneity in gene expression studies by surrogate variable analysis. PLoS Genetics, 3: e161.

Adjusted num.sv which estimates the number of important surrogate variables from a gene expression data set.

Description

This function is directly modified from the orginal num.sv() in the SVA package on Bioconductor. This function has the tolerance level in the fast.svd() function set back to its orginal default instead of 0.

Usage

num.sv2(dat, mod, method = c("be", "leek"), vfilter = NULL, 
B = 20, sv.sig = 0.1, seed = NULL)

Arguments

Details

See http://www.bioconductor.org/packages/release/bioc/manuals/sva/man/sva.pdf

Value

Note

sva Vignette http://www.biostat.jhsph.edu/~jleek/sva/

Author(s)

Original num.sv: Jeffrey T. Leek <jleek@jhsph.edu>, John Storey jstorey@princeton.edu

References

Original sva: Leek JT and Storey JD. (2008) A general framework for multiple testing dependence.Proceedings of the National Academy of Sciences, 105: 18718-18723.

Leek JT and Storey JD. (2007) Capturing heterogeneity in gene expression studies by surrogate variable analysis. PLoS Genetics, 3: e161.

The adjusted sva code using irwsva.build2

Description

This function is the modified SVA function in which it uses the irwsva.build2 function rather than the irwsva.build function to build the surrogate variables. Thus, only a single line has been altered from the original surrogate variable anaysis function.

Usage

sva2(dat, mod, mod0 = NULL, n.sv = NULL, method = c("irw", "two-step"), 
	vfilter = NULL, B = 5, numSVmethod = "be")

Arguments

Details

See http://www.bioconductor.org/packages/release/bioc/manuals/sva/man/sva.pdf

Value

Note

sva Vignette http://www.biostat.jhsph.edu/~jleek/sva/

Author(s)

Original sva: Jeffrey T. Leek <jleek@jhsph.edu>, John Storey jstorey@princeton.edu

References

Original sva: Leek JT and Storey JD. (2008) A general framework for multiple testing dependence. Proceedings of the National Academy of Sciences, 105: 18718-18723.

Leek JT and Storey JD. (2007) Capturing heterogeneity in gene expression studies by surrogate variable analysis. PLoS Genetics, 3: e161.

A screening process built upon training and testing samples

Description

A screening process to filter out non-informative DNA methylation sites by applying (ordinary or robust) linear regressions to training data, and the results are further examined using testing samples. Surrogate variables are included to account for unknown factors.

Usage

ttScreening(y = y, formula, imp.var, data, B.values=FALSE,iterations = 100, 
sva.method = c("two-step", "irw"), cv.cutoff = 50, n.sv = NULL, 
train.alpha = 0.05, test.alpha = 0.1, FDR.alpha = 0.05, Bon.alpha = 0.05,
percent = (2/3),linear = c("robust", "ls"), vfilter = NULL, B = 5,
 numSVmethod = "be", rowname = NULL, maxit=20)

Arguments

Details

See lm or glm for details.

Value

References

Meredith Ray, Xin Tong, Hongmei Zhang, and Wilfred Karmaus. (2014) "DNA methylation sites screening with surrogate variables", unpublished manuscript.

Leek JT and Storey JD. (2007) "Capturing heterogeneity in gene expression studies by ‘Surrogate Variable Analysis’." PLoS Genetics, 3: e161.

Examples

## Not run: 
library(mvtnorm)
nsub=600
imp=100
num=2000

set.seed(1)
x1= rnorm(nsub,1,1)
size1<-rmultinom(1,nsub,c(0.15,0.25,0.25,0.35))
x2= matrix(sample(c(rep(0,size1[1,]),
			rep(1,size1[2,]),
			rep(2,size1[3,]),
			rep(3,size1[4,])),replace=F),byrow=250,ncol=1)
			
sur1<-rnorm(nsub,0,5)
sur2<-rnorm(nsub,3,1)
sur3<-rnorm(nsub,0,1)
sur4<-rnorm(nsub,2,4)
sur5<-rnorm(nsub,0,3)

sigma1<-matrix(0,nrow=num,ncol=num)
diag(sigma1)<-1.5

beta0<-0.5
beta1<-0.3
beta2<-0.3
beta3<-0.3

sbeta1<-rnorm(1,0.5,0.01)
sbeta2<-rnorm(1,0.5,0.01)
sbeta3<-rnorm(1,0.5,0.01)
sbeta4<-rnorm(1,0.5,0.01)
sbeta5<-rnorm(1,0.5,0.01)

#beta matrix#
beta<-as.matrix(cbind(beta0,beta1,beta2,beta3,sbeta1,sbeta2,sbeta3,sbeta4,sbeta5))
beta.no2<-as.matrix(cbind(beta0,beta1,beta3,sbeta1,sbeta2,sbeta3,sbeta4,sbeta5))
beta.sur<-as.matrix(cbind(sbeta1,sbeta2,sbeta3,sbeta4,sbeta5))
#design matrix#
X<-as.matrix(cbind(rep(1,length(x1)),x1,x2,x1*x2,sur1,sur2,sur3,sur4,sur5))
X.no2<-as.matrix(cbind(rep(1,length(x1)),x1,x1*x2,sur1,sur2,sur3,sur4,sur5))
X.sur<-as.matrix(cbind(sur1,sur2,sur3,sur4,sur5))
#mu matrix#
imp1.mu<-matrix(rep(X%*%t(beta),9),nrow=nsub,ncol=(imp*0.9))
imp2.mu<-matrix(rep(X.no2%*%t(beta.no2),1),nrow=nsub,ncol=(imp*0.1))
noimp.mu<-matrix(rep(X.sur%*%t(beta.sur),num-imp),nrow=nsub,ncol=num-imp)
mu.matrix=cbind(imp1.mu, imp2.mu, noimp.mu)
error<-rmvnorm(nsub,mean=rep(0,num),sigma=sigma1,method = "chol")
y<-t(mu.matrix+error)


runs<-ttScreening(y=y,formula=~x1+x2+x1:x2,imp.var=3,data=data.frame(x1,x2),sva.method="two-step",
			B.values=FALSE,iterations=100,cv.cutoff=50,n.sv=NULL,train.alpha=0.05,
			test.alpha=0.05,FDR.alpha=0.05,Bon.alpha=0.05,percent=(2/3),linear= "ls",
			vfilter = NULL, B = 5, numSVmethod = "be",rowname=NULL,maxit=20)

runs$TT.output
runs$FDR.output
runs$Bon.output

## End(Not run)