A test to asess differential genetic bases amongst disease subtypes
The wide phenotypic variation encountered in many common diseases may be analysed by genetic analysis of disease subtypes. In situations where a phenotype may arise from distinct mechanisms - for example, ischaemic and haemorrhagic strokes - identification and understanding of subtypes is essential to a therapeutic understanding of disease pathology.
An important initial problem is determining whether phenotypically or clinically-defined subgroups of a case group represent different genetic pathophysiologies or subtypes. We present a statistical methodology for assessing this, based on the assertion that under this scenario, some SNPs associated with the phenotype as a whole should also have different effect sizes between subtypes.
Rather than initially attempting to find SNPs responsible for the difference, a task for which GWAS are typically underpowered, our method uses mixture Gaussians to model overall effect size distributions, allowing insight into the genetic architecture of the disease and subtypes. Given whole-genome evidence for different pathologies in subtypes, we can then use our fitted models to determine the SNPs most likely to be responsible for the difference. Our method makes allowance for linkage disequilibrium between SNPs, and we demonstrate some asymptotic properties of our test statistic which enable computationally efficient estimation of p-values with greatly reduced need for permutation testing.
We demonstrate our approach by showing evidence for differential genetic architecture in two major clinical subtypes of autoimmune thyroid disease, but no such difference between subtypes defined by geographic location. We also analyse type 1 diabetes GWAS data subgrouped by level of relevant autoantibodies.
Overall, we demonstrate that modelling GWAS summary statistics from related diseases using mixtures of multivariate Gaussians enables discovery of subtype structures without the need to directly identify associated SNPs.
