The lit package implements a flexible kernel-based
multivariate testing procedure, called Latent Interaction Testing (LIT),
to detect latent genetic interactions in a genome-wide association
study. In a standard GWAS analysis, one typically attempts to determine
which SNPs are associated with one (or many) traits. Another important
question is
This question has been very difficult to answer because effect sizes of interactions are likely small, interactive variables are unknown, and there’s often a large multiple testing burden from testing many candidate interactive variables.
One way to help address some of these issues is to use a variance-based testing procedure which does not require the interactive variable(s) to be specified or observed. These procedures can detect any unequal residual trait variation among genotype categories at a specific SNP (i.e., heteroskedasticity), which could suggest an unobserved (or latent) genetic interaction. However, researchers apply such procedures on a trait-by-trait basis and ignore any biological pleiotropy among traits. In fact, it is simple to show that a latent genetic interaction not only induces a variance effect but also a covariance effect between traits, and these covariance patterns can be harnessed to improve the statistical power.
The lit package addresses this gap by leveraging both
the differential variance and differential covariance patterns
to substantially increase power to detect latent genetic interactions in
a GWAS. In particular, LIT assesses whether the trait
variances/covariances vary as a function of genotype using a
kernel-based distance covariance (KDC) framework. LIT often provides
substantial increases in power compared to trait-by-trait univariate
approaches, in part because LIT uses shared information (i.e.,
pleiotropy) across tests and does not require a multiple testing
correction which negatively impacts power.
Note that this package contains the core functionality for the methods described in
Bass AJ, Bian S, Wingo AP, Wingo TS, Culter DJ, Epstein MP. Identifying latent genetic interactions in genome-wide association studies using multiple traits. Submitted; 2023.
Additional software features will be added in the future.
We provide two ways to use the lit package. For small
GWAS datasets where the genotypes can be loaded in R, the
lit() function can be used:
library(lit)
# set seed
set.seed(123)
# generate SNPs and traits
X <- matrix(rbinom(10 * 10, size = 2, prob = 0.25), ncol = 10)
Y <- matrix(rnorm(10 * 4), ncol = 4)
# test for latent genetic interactions
out <- lit(Y, X)
head(out)
#> wlit ulit alit
#> 1 0.2681410 0.3504852 0.3056363
#> 2 0.7773637 0.3504852 0.6044655
#> 3 0.4034423 0.3504852 0.3760632
#> 4 0.7874949 0.3504852 0.6157108
#> 5 0.8701189 0.3504852 0.7337565
#> 6 0.2352616 0.3504852 0.2847600The output is a data frame of p-values where the rows are SNPs and
the columns are different implementations of LIT to test for latent
genetic interactions: the first column (wlit) uses a linear
kernel, the second column (ulit) uses a projection kernel,
and the third column (alit) maximizes the number of
discoveries by combining the p-values of the linear and
projection kernels.
For large GWAS datasets (e.g., biobank-sized), the lit()
function is not computationally feasible. Instead, the
lit_plink() function can be applied directly to plink
files. To demonstrate how to use the function, we use the example plink
files from the genio package:
# load genio package
library(genio)
# path to plink files
file <- system.file("extdata", 'sample.bed', package = "genio", mustWork = TRUE)
# generate trait expression
Y <- matrix(rnorm(10 * 4), ncol = 4)
# apply lit to plink file
out <- lit_plink(Y, file = file, verbose = FALSE)
head(out)
#> chr id pos alt ref maf wlit ulit alit
#> 1 1 rs3094315 752566 G A 0.3888889 0.7908763 0.3422960 0.6150572
#> 2 1 rs7419119 842013 T G 0.3888889 0.1552580 0.3422960 0.2194972
#> 3 1 rs13302957 891021 G A 0.2500000 0.4088937 0.3325939 0.3687589
#> 4 1 rs6696609 903426 C T 0.3125000 0.5857829 0.3325939 0.4519475
#> 5 1 rs8997 949654 A G 0.4375000 0.6628300 0.3325939 0.4969663
#> 6 1 rs9442372 1018704 A G 0.2500000 0.3192430 0.3325939 0.3258332See ?lit and ?lit_plink for additional
details and input arguments.
Note that a marginal testing procedure for latent genetic
interactions based on the squared residuals and cross products (Marginal
(SQ/CP)) can also be implemented using the marginal and
marginal_plink functions: