GxEScanR: Performs GWAS/GWEIS scans using BinaryDosage files

GxEScanR

GxEScanR is designed to efficiently run genome-wide association study (GWAS) and genome-wide by environment interaction study (GWEIS) scans using imputed genotypes stored in the BinaryDosage format. The phenotype to be analyzed can either be a continuous or binary trait. The GWEIS scan performs multiple tests that can be used in two-step methods.

Installation

install.packages("GxEScanR")

To install the development version from GitHub:

# install.packages("devtools")
devtools::install_github("USCbiostats/GxEScanR@develop")

Example

Notation and models

There are five models that can be fit using GxEScanR when running a GWEIS and one that has to be fit before running the GWEIS.

$Y$ represents the outcome or phenotype

$X$ represents the confounding covariates

$E$ represents the covariate that will have its genetic interaction tested

$G$ represents the genetic data

$\beta$ represents the effect estimate that will be fitted in the model

Model 0 - Environment-only Model

This model contains the phenotype, confounding covariates, and the interaction covariate. This must be fit prior to running the GWEIS using the glm() routine.

$Y = \beta_X X + \beta_E E$

Model 1 - Gene-only Model

This model contains all the values as the environment-only model and adds the genetic data.

$Y = \beta_X X + \beta_E E + \beta_G G$

$\beta_G$ is the only term tested in this model and is represented by bg_ge in the output.

Model 2 - GxE Interaction Model

This model contains all the values as the gene-only model and adds the gene-environment interaction term.

$Y = \beta_X X + \beta_E E + \beta_G G + \beta_{GxE} GE$

$\beta_G$ and $\beta_{GxE}$ are both tested for this model as well as a joint test for both terms. They are represented by bg_gxe, bgxe, and joint, respectively, in the output.

Model 3 - EG model

This model tests for a relationship between the environmental value of interest and the gene. If $E$ is coded 0,1, a logistic model is fitted, otherwise a linear model is fitted.

$E = \beta_X X + \beta_G G$

$\beta_G$ is tested in this model and is represented by bg_eg in the output.

Model 4 - Case-only Model

This model is the same as model 3 except it is run on the cases only.

$\beta_G$ is tested and is represented by bg_case in the output.

This model can only be run on a dichotomous phenotype.

Model 5 - Control-only Model

This model is the same as model 3 except it is run on the controls only.

$\beta_G$ is tested and is represented by bg_ctrl in the output.

This model can only be run on a dichotomous phenotype.

Steps needed to run a GWEIS

There are 5 steps to running a GWEIS using GxEScanR

1. Preparing the Data
   • Genetic Data
   • Subject Data
2. Fitting the Base Model
3. Allocating Memory for the GWEIS
4. Running the GWEIS
5. Processing the results

Preparing the Data

Genetic Data

GxEScanR uses data stored in the BinaryDosage format. The BinaryDosage package converts files from the VCF format to the BinaryDosage format.

This example uses a VCF file included with the GxEScanR package. The complete file name is obtained using the system.file() routine. The file is then converted into a BinaryDosage formatted file using the vcftobd() routine from the BinaryDosage package. The getbdinfo() routine is then used to load the information about the BinaryDosage file.

vcffile <- system.file("extdata", "gendata.vcf.gz", package = "GxEScanR")

exampledir <- tempdir()
bdosefile <- file.path(exampledir, "gendata.bdose")

BinaryDosage::vcftobd(vcffile = vcffile, bdose_file = bdosefile)
bdinfo <- BinaryDosage::getbdinfo(bdosefile)

Subject Data

The subject data consists of the subject IDs that will be used to link the genetic data, and the phenotype and the covariates that will be used in the GWEIS. The data included with the GxEScanR package has two phenotypes, one linear and one dichotomous. The data also contains two covariates, x1 and x2.

Important All values for phenotypes and covariates must be numeric; factors are not allowed.

Important All subject IDs must be character values.

Important When using a dichotomous phenotype, it must be coded 0,1.

subjectfile <- system.file("extdata", "subdata.rds", package = "GxEScanR")
subjectdata <- readRDS(subjectfile)
head(subjectdata)
#>      subid y_linear y_logistic x1 x2
#> 1 subject1 7.939120          1  0  0
#> 2 subject2 5.965786          0  1  0
#> 3 subject3 3.047969          0  1  0
#> 4 subject4 1.173026          0  0  0
#> 5 subject5 7.691592          1  1  0
#> 6 subject6 3.931234          0  1  0

Fitting the Base Model

The base model is fit using the glm() routine using data that contains only the subject IDs, phenotype, and covariates of interest. No genetic data will be in the model.

It is important to remove subjects that have missing data and subjects that don’t have genetic data. The subjects with genetic data are listed in the BinaryDosage information file. It is also important to keep the subject IDs in the data used by the glm() routine because they are used to match the genetic data to the correct phenotype and covariates.

The formula used in the glm() routine must have the covariate whose interaction with the genetic data is being tested listed last. In this example we will be testing for a genetic interaction with x1, so the formula is:

y ~ x2 + x1

Note: It is recommended to look at the results from the glm() routine to verify the model was fitted correctly.

Continuous Phenotype

# Remove y_logistic from the subject data
lineardata <- subjectdata[, c(1, 2, 4, 5)]

# Keep only subjects with complete data
lineardata <- lineardata[complete.cases(lineardata), ]

# Keep only subjects with genetic data
lineardata <- lineardata[lineardata$subid %in% bdinfo$samples$sid, ]

# Fit the linear model using the glm routine
linearmodel <- glm(y_linear ~ x2 + x1, data = lineardata)

Dichotomous Phenotype

# Remove y_linear from the subject data
logisticdata <- subjectdata[, c(1, 3, 4, 5)]

# Keep only subjects with complete data
logisticdata <- logisticdata[complete.cases(logisticdata), ]

# Keep only subjects with genetic data
logisticdata <- logisticdata[logisticdata$subid %in% bdinfo$samples$sid, ]

# Fit the logistic model using the glm routine
logisticmodel <- glm(y_logistic ~ x2 + x1, family = binomial, data = logisticdata)

Allocating Memory for the GWEIS

Memory needs to be allocated for the GWEIS. This makes the GWEIS run much quicker. Allocating memory for the GWEIS is done using gweis.mem(), passing it the model fitted without the genetic data, the subject IDs, and the list of GWEIS tests to perform. The subject IDs should come from the dataset that was used in the glm() routine.

The following tests are the ones normally performed in a GWEIS with a continuous phenotype: “bg_ge”, “bg_gxe”, “bgxe”, and “joint”.

The following tests are the ones normally performed in a GWEIS with a dichotomous phenotype: “bg_ge”, “bg_gxe”, “bgxe”, “joint”, “bg_eg”, “bg_case”, and “bg_ctrl”.

# GWEIS with continuous outcome
lineartests <- c("bg_ge", "bg_gxe", "bgxe", "joint")
linearmem <- gweis.mem(gemdl = linearmodel,
                       subids = lineardata$subid,
                       tests = lineartests)

# GWEIS with dichotomous outcome
logistictests <- c("bg_ge", "bg_gxe", "bgxe", "joint", "bg_eg", "bg_case", "bg_ctrl")
logisticmem <- gweis.mem(gemdl = logisticmodel,
                         subids = logisticdata$subid,
                         tests = logistictests)

Running the GWEIS

To run the GWEIS you will need the BinaryDosage information file, the list returned from gweis.mem(), a list of SNPs to test, and the name of the output file.

The list of SNPs to perform the GWEIS on can be either the SNP IDs or the indices in the BinaryDosage information. In this example all the SNPs in the genetic data will be tested and they will be identified by index.

snpindex <- 1:nrow(bdinfo$snps)

# GWEIS with continuous outcome
linearresults <- file.path(exampledir, "linear.txt")
rungweis(gweismem = linearmem,
         bdinfo = bdinfo,
         snps = snpindex,
         outfilename = linearresults)

# GWEIS with dichotomous outcome
logisticresults <- file.path(exampledir, "logistic.txt")
rungweis(gweismem = logisticmem,
         bdinfo = bdinfo,
         snps = snpindex,
         outfilename = logisticresults)

Processing the results

Processing the results is done mostly with other tools. The following describes the output columns. Only the first 3 lines of the results are displayed.

lineardf <- read.table(linearresults, header = TRUE, sep = "\t")
logisticdf <- read.table(logisticresults, header = TRUE, sep = "\t")

For a continuous outcome the SNP data consists of SNP ID (snpid), chromosome number (chr), location (loc), reference allele (ref), alternate allele (alt), and alternate allele frequency (aaf).

knitr::kable(lineardf[1:3, 1:6])

snpid	chr	loc	ref	alt	aaf
chr22:10527916:T:G	chr22	10527916	T	G	0.6494005
chr22:10648779:G:T	chr22	10648779	G	T	0.0854975
chr22:10648794:G:A	chr22	10648794	G	A	0.0536727

For a dichotomous outcome, two additional columns are added: the alternate allele frequency in cases (aaf_case) and in controls (aaf_ctrl).

knitr::kable(logisticdf[1:3, 1:8])

snpid	chr	loc	ref	alt	aaf	aaf_case	aaf_ctrl
chr22:10527916:T:G	chr22	10527916	T	G	0.6494005	0.6264272	0.6540861
chr22:10648779:G:T	chr22	10648779	G	T	0.0854975	0.0882864	0.0849287
chr22:10648794:G:A	chr22	10648794	G	A	0.0536727	0.0457573	0.0552871

For model 1 there are two values output: bg_ge and bg_ge_lrt. The first value is the $\beta_G$ estimate in model 1. The second value is the likelihood ratio test (LRT) 1 df chi-squared value for $\beta_G = 0$. These values are output when “bg_ge” is included in the tests value passed to gweis.mem().

knitr::kable(lineardf[1:3, 7:8])

bg_ge	bg_ge_lrt
-0.8900621	5.2558824
-0.4766408	0.6737040
-0.6481754	0.8070534

For model 2 there are five possible values that can be output.

bg_gxe and bg_gxe_lrt are the estimate of $\beta_G$ and the LRT 1 df test that the value of the estimate is zero. These values are output when “bg_gxe” is included in the tests value passed to gweis.mem().

bgxe and bgxe_lrt are the estimate of $\beta_{GxE}$ and the LRT 1 df test that the value of the estimate is zero. These values are output when “bgxe” is included in the tests value passed to gweis.mem().

joint_lrt is the LRT 2 df chi-squared value for the test that $\beta_G = 0$ and $\beta_{GxE} = 0$. This value is output when “joint” is included in the tests value passed to gweis.mem().

knitr::kable(lineardf[1:3, 9:13])

bg_gxe	bg_gxe_lrt	bgxe	bgxe_lrt	joint_lrt
-1.4864649	8.4363261	1.4016489	3.2015268	8.4574092
-0.4591966	0.3719000	-0.0430555	0.0013243	0.6750284
-0.6509083	0.5512644	0.0084781	0.0000301	0.8070835

For models 3 through 5 the estimate of $\beta_G$ and the LRT 1 df chi-squared value for $\beta_G = 0$ are output. The estimates are labeled bg_eg, bg_case, and bg_ctrl for the three tests, respectively. The test statistics are labeled bg_eg_lrt, bg_case_lrt, and bg_ctrl_lrt, respectively.

knitr::kable(logisticdf[1:3, 16:21])

bg_eg	bg_eg_lrt	bg_case	bg_case_lrt	bg_ctrl	bg_ctrl_lrt
-0.3460221	0.7862570	0.8264053	0.8541469	-0.4673483	1.1215720
-0.0867130	0.0219140	-0.7545895	0.4064202	0.0815513	0.0144966
-0.2255839	0.0949399	-0.4952261	0.0361789	-0.0220965	0.0008210