A key goal in imaging neuroscience is to discover specific genetic variants that influence brain structure and function (Glahn et al., 2007a
; Glahn et al., 2007b
). The dynamic trajectory of brain development and aging throughout life is strongly influenced by genetic factors, and genetic variants have been discovered that increase risk the for Alzheimer's disease (Corder et al., 1993
), other mental illness, (Gottesman and Gould, 2003
; Meyer-Lindenberg and Weinberger, 2006
; Purcell et al., 2009
) and even obesity (Frayling et al., 2007
; Ho et al., submitted for publication
). The goals are both scientific and practical: by selecting those at genetic risk for early treatment, drug trials will be better powered to detect treatment effects (Frisoni et al., in press
). A more mechanistic understanding of mental illness will be achieved if gene variants over-represented in patients are studied both at the molecular level and in terms of their effects on brain structure.
Early neuroimaging studies of twins found that several aspects of brain structure are under strong genetic control (Thompson et al., 2001
; Posthuma et al., 2002
) and that common sets of genes may influence brain structure and cognition (Posthuma et al., 2002
). These “first-generation” studies estimated the relative influence of genetic contributions from relatives or family members, based on the expected genetic similarity among different types of relatives. Studies of identical and fraternal twins, and their siblings, have consistently identified heritable aspects of brain structure (Thompson et al., 2001
; Styner et al., 2005
; Hulshoff Pol et al., 2006
; Peper et al., 2007
; Schmitt et al., 2008
; Brun et al., 2009
; Chou et al., 2009
). Except for the genotyping necessary to confirm the zygosity of twins in these studies, specific variations at the DNA level are not used in these analyses.
Early studies that use more detailed genotype information focus on specific candidate gene effects on brain structure. Several studies of candidate genes such as APOE
, and BDNF
have divided populations into carriers and non-carriers of risk polymorphisms within these genes, and detected systematic differences in brain structure using a standard statistical comparison of two groups (Egan et al., 2001
; Pezawas et al., 2004
; Hua et al., 2008
; Chiang et al., 2009
More recently, the second generation of studies has used genome-wide scans to search the entire genome for genetic polymorphisms that influence brain structure. In Stein et al. (submitted for publication)
, a common variant in the GRIN2B
glutamate receptor gene was found to be over-represented in Alzheimer's disease and was associated with ~1.5% lower temporal lobe volume per risk allele in the elderly (N
=742 subjects; P
). Genome-wide searches have not generally been the most efficient or feasible approach in imaging genetics, as they require large samples of subjects to discover gene effects that survive stringent multiple comparisons corrections for searching over the entire genome. However, several international efforts are now underway to scan genotyped healthy and diseased subjects with the goal of discovering which genetic variants contribute to brain architecture (Thompson and Martin, 2010
Perhaps surprisingly, no genome-wide study of brain images has used the armory of statistical methods that are now standard in human brain mapping, such as statistical parametric mapping (Friston et al., 1994
; Frackowiak, 2004
). One study has looked at statistical power for statistical parametric mapping with simulated genome-wide data (Hayasaka, 2009
), but no experimental whole-brain whole-genome approach has been implemented to our knowledge. Most twin morphometric studies still break up the brain into subvolumes (Schmitt et al., 2007
) and run genetic analysis on the numerical summaries (subvolumes).
By contrast, voxel-based morphometric approaches can make detailed 3D images of volume differences throughout the brain, without the need to specify a priori
regions of interest or time consuming manual tracing of anatomy in brain images. These maps of individual differences in brain morphometry make it possible to create detailed maps of gene and environmental effects on the brain, identifying spatially-varying patterns of genetic control that may not be evident if the images were summarized using a few summary indices. Maps of genetic influences on cortical anatomy reveal strong genetic control of frontal anatomy (Thompson et al., 2001
), and regionally-varying gene effects (Panizzon et al., 2009
). Genetic maps based on tensor-based morphometry suggest that there may be some gradients in the degree of genetic influence, with earlier developing occipital lobe structures showing stronger genetic control than frontal brain regions that mature over a more protracted developmental time-course (Brun et al., 2009
; Lee et al., submitted for publication).
Here we extend the notion of statistical parametric mapping, using voxel-based methods, to include genome-wide association (GWAS) data in large populations. The result may be termed voxelwise GWAS (or vGWAS). GWAS is usually applied to study a single trait, such as IQ or the diagnosis of a specific disease, but here it is applied at each location in a brain image. The result is a 3D map of the specific genetic variants that have the greatest statistical effect in accounting for volume variations in each part of the brain, and a method to assess their statistical significance.
Recent advances in neuroimaging and genetics have made it possible, and financially feasible, to scan populations with multi-modality brain imaging and collect genome-wide data (Toga, 2002
; McCarthy et al., 2008
). The Alzheimer's Disease Neuroimaging Initiative (ADNI) has recently acquired genome-wide genotype data as well as structural MRI scans of 740 subjects (Mueller et al., 2005
). This wealth of data is a blessing and a burden: 448,293 genotypes and 31,622 voxels in the brain in each of 740 subjects present powerful and previously unknown spatial and genetic resolution to detect specific variants that influence the brain. However, this vast amount of data requires new ways to deal with the computational load and account statistically for multiple comparisons. A genetic association is usually conducted by performing a linear regression of a phenotype on each genotype of interest, controlling for other confounding variables of no interest. Generally, a genome-wide association study examines only a few phenotypes of interest (Wellcome Trust Case Control Consortium, 2007
; Sabatti et al., 2009
). When conducting a voxelwise genome-wide association study, each voxel represents a phenotype, so a regression must be run at each voxel and at each SNP (~1.4×1010
tests), which requires large amounts of computation time (years) if run serially on one computer. Parallelizing this process across a computing cluster can ease the computational burden, giving results in a reasonable amount of time (days). Additionally, by conducting many statistical tests (in this case ~1.4×1010
) on the same dataset, we are highly prone to false-positive findings (Curran-Everett, 2000
). Finding a method to determine only those genetic hits that are interesting to pursue without overlooking those with potentially important effects is a difficult question explored further here.
For the first time, we conducted a voxelwise genome-wide association study (vGWAS) in 740 subjects to discover genes influencing brain structure across the entire brain. Each genetic variant identified is a potential candidate with the ability to effect brain structure. If these brain traits lie on the path from genes to disorders that involve the brain (Gottesman and Gould, 2003
), they could represent candidates for further study in neurological and psychiatric diseases.