minPtest: A gene region-level testing procedure for each candidate gene...

In minPtest: Gene region-level testing procedure for SNP data, using the min P test resampling approach

Description

Permutation-based p-values estimation via min P test, a gene region-level summary for each candidate gene. The gene region-level summary assesses the smallest p-trend within each gene region comparing cases and controls. The min P test is permutation-based method that can be based on different univariate tests per SNP. Inference is based on the permutation distribution of the ordered p-values from the marginal tests of each SNP. Potentially accelerated by parallelization, if a compute cluster or a multicore computer is available.

Usage

minPtest(y, x, SNPtoGene, formula = NULL, cov = NULL, matchset = NULL,
         permutation = 1000, seed = NULL, subset = NULL,
         parallel = FALSE, ccparallel = FALSE,
         trace = FALSE, aggregation.fun = min,
         adj.method=c("bonferroni","holm","hochberg",
                      "hommel","BH","BY","fdr","none"),
         ...)

Arguments

y	a numeric response vector coded with 0 (coding for controls) and 1 (coding for cases) of length n.
x	a numeric n * p matrix of covariates (i.e. SNPs) containing the genotypes coded by 0, 1 and 2. Thus, each column is assumed to represent one of the SNPs with corresponding column names. Detail for SNP coding are given below.
SNPtoGene	a mapping matrix of dimension p x 2 comprising SNP names (first column) which are same as the column names of x, and the gene names (second column) on which the SNPs are located.
formula	(optional) for unconditional or conditional logistic regression, respectively, with or without covariates other than SNPs. A symbolic description of the model to be fitted including covariates only, see glm or clogistic, the latter requires library Epi, else the default method Cochran Armitage Trend Test, which requires library scrime, is fitted. Details of model specification are given below.
cov	(optional) a n * q matrix containing the covariates for adjustment with corresponding column names.
matchset	(optional) a numeric vector of length n containing matching numbers, needed for conditional logistic regression.
permutation	number of permutations employed to obtain a null distribution.
seed	(optional) vector of length permutation. Allows reproducibility even when running in parallel and for different numbers of parallel processes.
subset	an optional vector specifying a subset of observations to be used in the fitting process.
parallel	indicates whether computation in the permuted data sets should be performed in parallel using package parallel. If TRUE, the parallelization requires at least two cores. A value larger than 1 is taken to be the number of cores.
ccparallel	logical value indicates whether computation should be performed in parallel on a compute cluster, using package snowfall. If TRUE the initialization function of this package, sfInit(), should be called before calling minPtest. See Details.
trace	logical value indicating whether progress in estimation should be indicated by printing the number of permutation that is currently used. (ignored if running in parallel via snowfall).
aggregation.fun	function that is used to combine the trend p-values over multiple loci within a gene region. By default the minimum ("min") is applied to obtain candidate gene region-level summaries. Any other function to integrate the p-values into one single test statistic can be used, e.g., median or different functions designed by the user.
adj.method	correction method for multiple hypothesis testing. By default the Bonferroni method ("bonferroni") is used. Any other correction method as in p.adjust can be used.
...	Further arguments for aggregation.fun. In case of NA/NaN within the evaluated marginal trend p-values for the SNPs from the original data (psnp) or within the permuted trend p-values in the permutation samples (psnpperm), na.rm=TRUE has to be specified.

Details

The idea of the gene region-level summary, using the min P test procedure (Westfall and Young, 1993; Westfall et al., 2002; Chen et al., 2006), is to identify candidate genes by assessing the statistical significance of the smallest p-trend from a set of SNPs (single nucleotide polymorphisms) within each gene region comparing cases and controls by permutation-based resampling methods. A SNP occurs when a single nucleotide, (A), (T), (C) or (G), in the genome differs between individuals and, in addition, this variation, substitution of one nucleotide for another, occurs in more than 1% of a population. A SNP can take three possible values (genotypes): either there is no SNP variant in comparison to some reference coding (homozygous reference (0)) or the SNP variant occurs on one of the two base pair positions (heterozygous (1)), or both base pairs have a variant comparing to the reference coding. minPtest permits to include, instead of the genotypes 0, 1 and 2, also combined carrier SNPs, e.g. coding 0 and 1 (1 + 2).

Computation of the min P test is based on the marginal trend p-values for a set of univariate SNP disease association and the trend p-values for the permutation samples for each SNP. The minPtest package brings together three different kinds of tests to compute such p-values that are scattered over several R packages, and automatically selects the one most appropriate for the design at hand. In any case a response vector y, a SNP matrix x and a mapping matrix SNPtoGene are required. Then the default, a Cochran Armitage Trend Test (Cochran, 1954; Armitage, 1955), is automatically fitted to compute p-values. The Cochran Armitage Trend Test does not depend on covariates and matching scenario. Additionally adding a formula, see also glm from package base, and a covariate matrix cov an unconditional logistic regression is fitted. Unconditional logistic regression can be used without or with covariates for adjustment; either formula=y~1 or formula=y~cov1+cov2+.... The former does not need any information relative to covariates and matching scenario. However, the latter is general for frequency matching with the inclusion of matching variables for adjustment specified in the covariate matrix cov. Providing a matchset, as in the case of 1:1; 1:2 etc. matching, and a formula, see also clogistic from package Epi, a conditional logistic regression is fitted. Conditional logistic regression can be used without or with covariates for adjustment; either formula=y~1 or formula=y~cov1+cov2.... In the latter case covariates other than matching variables can be used and have to be specified in the covariate matrix cov. In general, there are two possibilities to specify the formula, first if no covariates are used for adjustment, the formula has to be written as y~1 without specifying the covariate matrix cov. Second if covariates other than SNPs are used for adjustment, the formula has to be written as response vector y on the left of a ~ operator, and the clinical covariates on the right, as well as a covariate matrix has to be specified.

If SNPs genotypes are coded by 0, 1 and 2, they are included as continuous variables in the logistic regression models. If SNPs are coded as carrier SNPs 0 and 1, they are included as binary variables in the logistic regression models. If covariates are used for adjustment, the column names of the covariate matrix cov have to be specified as used in the formula specification, to link the formula with the covariate matrix cov.

Missing SNP genotypes in x or, if used, missing values in cov are accounted for, as each marginal test makes use of the available data for that SNP in x and for that covariate in cov only. The minPtest uses all subjects with available data for each SNP (and covariates) when fitting Cochran Armitage Trend Test or unconditional logistic regression. Note that in conditional logistic regression, the matched subjects are removed together in case of 1:1 matching. In the 1:2 matching scenario, matched subjects are removed when the missing occurs in a case, otherwise when a missing occurs in one control, only that control is removed.

Concerning parallelization on a compute cluster, i.e. with argument ccparallel=TRUE, there are two possibilities to run minPtest:

1. Start R on a commandline with sfCluster (Knaus et al., 2009) and preferred options, e.g. number of cpus. The initialization function of package snowfall, sfInit(), should be called before calling minPtest.

2. Use any other solutions supported by snowfall. Argument ccparallel has to be set to TRUE and number of cpus can be chosen in the sfInit() function.

sfCluster is a Unix tool for convenient management of R parallel processes. It is available at www.imbi.uni-freiburg.de/parallel, with detailed information.

A print function returns a short overviews of the results. The print function describes the number of subjects included in the analysis, which method is used by the package, briefing of the number of genes, the number of SNPs, the number of missings in the SNP matrix x and the number of permutations used for the fit. A summary.minPtest and a plot.minPtest function are available.

Value

An object of class 'minPtest', which is a list containing the following components:

minp	nrgene * 1 matrix of permutation-based p-values of the min P test for each candidate gene.
p.adj.minp	nrgene * 1 matrix of corrected permutation-based p-values for each candidate gene.
psnp	nrsnp * 1 matrix of marginal trend p-values for each SNP from the original data set.
p.adj.psnp	nrsnp * 1 matrix of corrected marginal trend p-values for each SNP from the original data set.
psnpperm	nrsnp * n.permute matrix of permuted trend p-values for each SNP in each permutation step.
zgen	nrgene * 1 matrix of min P test statistics for each candidate gene from the original data set.
zgenperm	nrgene * n.permute matrix of permuted min P test statistics for each candidate gene in each permutation step.
n	number of subjects in the original data set.
nrsnp	number of SNPs in the original data set.
nrgene	number of genes in the original data set.
snp.miss	number of missings in the SNP matrix x.
n.permute	number of permutations.
method	used method.
call	call.
SNPtoGene	the mapping matrix of dimension p x 2 comprising of SNP names (first column) and names of the genes (second column) on which the SNPs are located.

Author(s)

Stefanie Hieke [e-mail]

References

Armitage,P. (1955). Tests for linear trends in proportions and frequencies. Biometrics, 11(3), 375-386.

Chen,B.E. et al. (2006). Resampling-based multiple hypothesis testing procedures for genetic case-control association studies. Genetic Epidemiology, 30, 495-507.

Cochran,W.G. (1954). Some methods for strengthening the common chi-squared tests. Biometrics, 10(4), 417-451.

Knaus,J. et al. (2009). Easier parallel computing in R with snowfall and sfCluster. The R Journal, 1, 54-59.

Westfall,P.H. et al.(2002). Multiple tests for genetic effects in association studies. Methods Mol Biol, 184, 143-168.

Westfall,P.H. and Young,S.S. (1993). Resampling-Based Multiple Testing: Example and Methods for p-Value Adjustment. Wiley, New York.

See Also

summary.minPtest, plot.minPtest

Examples

# generate a simulated data set as in the example of the function generateSNPs
# consisting of 100 subjects and 200 SNPs on 5 genes.

SNP <- c(6,26,54,135,156,186)
BETA <- c(0.9,0.7,1.5,0.5,0.6,0.8)
SNPtoBETA <- matrix(c(SNP,BETA),ncol=2,nrow=6)
colnames(SNPtoBETA) <- c("SNP.item","SNP.beta")

set.seed(191)
sim1 <- generateSNPs(n=100,gene.no=5,block.no=4,block.size=10,p.same=0.9,
p.different=0.75,p.minor=c(0.1,0.4,0.1,0.4),n.sample=80,SNPtoBETA=SNPtoBETA)

# Cochran Armitage Trend Test without covariates and default permutations.
# Example: Run R sequential

### Seed
set.seed(10)
seed1 <- sample(1:1e7,size=1000)
###
minPtest.object <- minPtest(y=sim1$y, x=sim1$x, SNPtoGene=sim1$SNPtoGene,
	                    seed=seed1)

minPtest documentation built on May 29, 2017, 1:06 p.m.