GhostKnockoff.fit {GhostKnockoff}

Feature importance score generator

Description

Generate original and knockoff feature importance scores given original Z-scores from multiple studies.

Usage

GhostKnockoff.fit(Zscore_0,n.study,fit.prelim,gamma=1,weight.study=NULL)

Arguments

Zscore_0

A p*K Z-score matrix, where p is the number of variants and K is the number of studies. Variants not observed in the study should be coded as NA.

n.study

A vector of length K, where each element is the study sample size.

fit.prelim

The output of function "GhostKnockoff.prelim".

gamma

The estimated study dependency. See functon "GhostKnockoff.GetGamma".

weight.study

The weights to combine multiple studies. The default is based on sample size. The optimal weights can be estimated using function "GhostKnockoff.prelim.Meta".

Value

GK.Zscore_0

A vector of length p, where each element is weighted combination of original Z-scores

GK.Zscore_k

A p*M matrix, where each column is a knockoff copy of GK.Zscore_0.

T_0

Feature importance score of original data: square of GK.Zscore_0.

T_k

Feature importance score of knockoff copies: square of GK.Zscore_k.

Examples

# We use genetic data as an example
library(GhostKnockoff)

# load example vcf file from package "seqminer", this serves as the reference panel
vcf.filename = system.file("vcf/1000g.phase1.20110521.CFH.var.anno.vcf.gz", package = "seqminer")

## this is how the actual genotype matrix from package "seqminer" looks like
example.G <- t(readVCFToMatrixByRange(vcf.filename, "1:196621007-196716634",annoType='')[[1]])
example.G <- example.G[,apply(example.G,2,sd)!=0]
example.G <- example.G[,1:100]

# compute correlation among variants
cor.G<-matrix(as.numeric(corpcor::cor.shrink(example.G)), nrow=ncol(example.G))

# fit null model
fit.prelim<-GhostKnockoff.prelim(cor.G,M=5,method='asdp',max.size=500)

### if only one study is involved
Zscore_0<-as.matrix(rnorm(nrow(cor.G))) # hypothetical Z-scores
Zscore_0<-Zscore_0+rbinom(nrow(cor.G),size=2,0.1) # set causal
n.study<-c(5000)

# knockoff scores for each block, this can be run in parallel too
GK.stat<-GhostKnockoff.fit(Zscore_0,n.study,fit.prelim,gamma=1,weight.study=NULL)

### if multiple studies are involved, for a meta-analysis

# compute study correlation
Zscore_0<-cbind(rnorm(nrow(cor.G)),rnorm(nrow(cor.G))) # hypothetical Z-scores
Zscore_0<-Zscore_0+rbinom(nrow(cor.G),size=2,0.1) # set causal
cor.study<-GhostKnockoff.GetCorStudy(Zscore_0,fit.prelim)

# note that all steps above can be run in parallel for nonoverlapping blocks of the genome.
# Then the overall study correlation can be computed by averaging cor.study of all blocks.

# compute optimal weights and study dependency using overall study correlaton
n.study<-c(5000,7500)
Meta.prelim<-GhostKnockoff.prelim.Meta(cor.study, n.study)
gamma<-Meta.prelim$gamma
weight.study<-Meta.prelim$w_opt

# knockoff scores for each block, this can be run in parallel too
GK.stat<-GhostKnockoff.fit(Zscore_0,n.study,fit.prelim,gamma=gamma,weight.study=weight.study)