General discussion
METHODOLOGICAL BARRIERS
Several chapters of the work presented in this dissertation (4.1, 5.2, and 5.3) became pos- sible as a result of the original idea of voxel-wise genome-wide analysis. At the beginning of my PhD (March 2014), there were no options to perform such analyses due to computational requirements of thousand years for individual sites. Additionally, the sheer size of intermediate results was not suitable for multicenter settings where summary statistics would need to be exchanged for meta-analysis. I therefore developed algorithms and a framework to overcome both issues. In chapter 4.1 I showed that the co-called HASE framework allows for reducing the computational time to just hours and limiting intermediate summary data from TBs to GBs. The framework was developed to be user-friendly for multicenter studies, and to employ it as part of GWAS pipelines in different analysis settings. In chapter 5.2 the framework was employed in a cortical brain regions GWAS study in combination with classical meta-analysis workflow, just to speed up a single side analysis. In chapter 5.3 the full functionality was used in a voxel-wise GWAS, and all findings in this chapter could not have been obtained without HASE. Having started as a tool to address a quite specific research problem, our HASE framework has now became a broadly used instrument for analysis genetics, gene expression and methylation data in many ongoing projects. We have published the framework online as open source software, and expect that many projects are currently using it and that the number of applications for which it will be used will grow significantly. We are planning to maintain this framework and add additional functionality to support the large scale explora- tion of omics data in both single- and multicenter studies.
In chapter 4.2 I adressed another fundamental problem in the genetics field, pleiotropy, which relates to the fact that genes may associate with more than one phenotype. Currently, around 43% of genes have been reported in GWAS catalogs to associate with more than
one phenotype1, suggesting that the pleiotropy might be rather a common phenomenon,
than something rare. There are several approaches to investigate the genetic correlation between phenotypes. However, identifying a single pleiotropy variant has always proven to
be challenging2. Existing methods to study pleiotropy are lacking or suffer from a bias toward the most statistically powered phenotype. In this thesis I developed a solution to address this issue, taking a combinatorial approach, based on sum ranks of SNPs statistics. This approach directly incorporates information from GWAS summaries and tests the pleiotropy hypothesis. Compared to others tests our approach is robust to extreme p-values and only finds significant pleiotropy when a signal of association is present for all traits. Moreover, the method is able to identify single loci pleiotropy, even when there is no significant genetic correlation. This property made it possible to find important genetic variants which affect several traits and disorders (chapter 4.2, 5.1, 5.3). An important finding in this thesis, and for the imaging genetics field, is that using this method we found significant pleiotropic loci between subcor-
tical structures and schizophrenia (chapter 4.2). In a previous proof-of-concept study3, using
other techniques, no evidence was found for such a relationship, neither at a high level nor for single genetic markers. This raised considerable debate in the research community about the importance of imaging GWAS discoveries for understanding disease etiology since no
genetic overlap was found. Importantly, the latest subcortical GWAS4 with larger sample size
also identified significant SNPs which associated with schizophrenia. Our method was able to obtain this result based on summary statistics on a smaller dataset. Additionally, we found variants not yet significant for each trait independently. Finally, when applying this method for new GWAS results of the commissure tract (chapter 5.1), we found that regardless of the underlying mechanism through which genes affect the anterior commissure, there seems to be pleiotropy with various neurodegenerative diseases (Alzheimer ’s, Parkinson, frontotem- poral dementia). The results reported in this thesis provide only a few examples, but most probably pleiotropy between neuroimaging biomarkers and neurodegenerative diseases is more common. Taken together this demonstrates the potential of the imaging genetics field to significantly contribute to a better understanding of complex disease genetic architecture. An additional advantage of our method is that it is not limited to study pleiotropy between two phenotypes, but it is also possible to apply it to several traits. We demonstrated this by investigating pleiotropy between five psychiatric disorders in chapter 4.2. Clinical observa-
tions show varying degrees of symptoms overlap between these diseases5, suggesting the ex-
istence of a (partially) shared genetic architecture. Additionally, the strong genetic correlation
between pairs of these disorders has been shown using GWAS summary statistics6. However,
to discover specific loci remained challenging. Using our method we were able to find pleio- tropic genes, not only between pairs, like RERE, CACNA1C for bipolar/schizophrenia but also between three or four diseases, like ITIH3, SFMBT1 genes. Our findings can help with better understanding of complex etiology of psychiatric diseases and provide candidate genes for drug development.
gests that multi-phenotype pleiotropy is a common phenomenon and as a result the method we developed has the potential to be applied widely.