FEMA-GWAS: mixed-effects algorithms for discovery of genome-wide non-linear SNP-by-age interactions
Poster No:
869
Submission Type:
Abstract Submission
Authors:
Pravesh Parekh1, Nadine Parker1, Diana Smith2, Diliana Pecheva2, Carolina Makowski2, Evgeniia Frei1, Gleda Kutrolli1, Hao Wang2, Dennis van der Meer1,3, Alexey Shadrin1, Thomas Nichols4, Oleksandr Frei1, Anders Dale2, Ole Andreassen1
Institutions:
1University of Oslo, Oslo, Norway, 2University of California San Diego, San Diego, CA, 3Maastricht University, Maastricht, Netherlands, 4University of Oxford, Oxford, United Kingdom
First Author:
Co-Author(s):
Introduction:
Coupling longitudinal neuroimaging and genetics data can be useful in discovering the genetic associations underlying neurodevelopment, brain maturation, and neurodegeneration. These discoveries can then be linked to disease but are computationally intensive and almost impractical at the voxel- or element-level. In this work, we present a novel mixed-effects framework for performing longitudinal genome-wide association studies (GWAS), as well as discovery of non-linear interactions of single nucleotide polymorphisms (SNPs) or other genetic variants with age/time.
Methods:
The longitudinal nature of the data and family structure (in studies like the ABCD Study) requires a mixed effect model, so we use a mixed modeling framework where the SNPs are the fixed effects, modeled individually, in addition to other nuisance effects. Considering the dimensionality of whole-brain neuroimaging data as well as the number of SNPs typically modeled in the GWAS framework (>1 million), it is computationally extremely laborious to use currently available mixed modeling frameworks. Therefore, we leverage our recently developed fast and efficient mixed-effects algorithm (FEMA) framework, which is a computationally efficient solution for performing whole-brain mixed modeling.
For FEMA-GWAS with non-linear interactions, our approach is as follows (see Figure 1 for an overview): 1) using FEMA, estimate the effect of the fixed covariates on the imaging variables (phenotypes) as well as the variance parameters for the random effects. 2) Use a two-stage estimation approach that is equivalent to the conventional approach (Frisch-Waugh theorem) where we pre-residualize the genotype as well as the imaging variables for the effects of these covariates. This simplifies the estimation problem to estimating the effect of the residualized SNPs on the residualized phenotype. 3) For non-linear interactions, use natural cubic splines with user-defined knots – these basis functions are used for expanding the (main) effect of SNP. 4) Finally, we use generalized least squares (re-using the estimated variance parameters from the first step) to estimate the parameters for this expanded set of predictors. As an additional feature, we allow these expanded set of predictors to interact with dummy-coded sex variable, thereby separately estimating the effects of the SNP-by-age interactions for males and females.
For FEMA-GWAS with non-linear interactions, our approach is as follows (see Figure 1 for an overview): 1) using FEMA, estimate the effect of the fixed covariates on the imaging variables (phenotypes) as well as the variance parameters for the random effects. 2) Use a two-stage estimation approach that is equivalent to the conventional approach (Frisch-Waugh theorem) where we pre-residualize the genotype as well as the imaging variables for the effects of these covariates. This simplifies the estimation problem to estimating the effect of the residualized SNPs on the residualized phenotype. 3) For non-linear interactions, use natural cubic splines with user-defined knots – these basis functions are used for expanding the (main) effect of SNP. 4) Finally, we use generalized least squares (re-using the estimated variance parameters from the first step) to estimate the parameters for this expanded set of predictors. As an additional feature, we allow these expanded set of predictors to interact with dummy-coded sex variable, thereby separately estimating the effects of the SNP-by-age interactions for males and females.
Results:
The main result is the development of FEMA-GWAS, a novel MATLAB-based software that allows computationally efficient discovery of SNP-by-age interaction across the entire genome for a large number of imaging variables. Using simulations, we show that FEMA-GWAS has i) well-controlled type I error; and provides: ii) equivalent estimates to standard mixed effects implementation, and iii) accurate parameter recovery for various interaction terms. As a practical use-case, we present results for non-linear interaction of SNP and age for cortical thickness using samples from the longitudinal ABCD Study.
Conclusions:
We have developed a software that enables users to perform longitudinal GWAS and discover the non-linear interactions of SNPs with age/time. FEMA-GWAS is specifically designed with the needs of the neuroimaging community in mind, easily scaling to a large number of outcome variables. Our work will enable users to perform these analyses at multiple levels of granularity – region of interest, voxel-wise, vertex-wise, and connectome-wide. We expect FEMA-GWAS to have wide-ranging impact – from applications in discovery of SNPs which drive the rate of change of imaging variables to establishing genetically adjusted normative models of brain phenotypes. FEMA is available to the public at: https://github.com/cmig-research-group/cmig_tools and FEMA-GWAS will be available as part of the same package soon.
Genetics:
Genetic Association Studies
Genetic Modeling and Analysis Methods 1
Modeling and Analysis Methods:
Methods Development 2
Univariate Modeling
Keywords:
Modeling
Statistical Methods
Univariate
Other - imaging-genetics; GWAS; longitudinal modeling
1|2Indicates the priority used for review