Using correlated phenotypes to functionally classify GWAS loci. N. Eriksson, J. Y. Tung, D. A. Hinds 23andMe, Mountain View, CA.
While there have been thousands of genetic loci convincingly associated with hundreds of different phenotypes through GWAS, the function of only a few of these associations has been uncovered. New methods are needed in order to uncover the pathways through which these associations function. Luckily, there are a large number of pleiotropic relationships among these loci (i.e., loci that affect multiple different phenotypes). This may occur with highly correlated traits (e.g., eye color and hair color), modestly correlated traits (e.g., different autoimmune diseases) or even seemingly unrelated traits (e.g., Parkinson's disease and baldness). Here, we take advantage of these pleiotropic effects across the wide range of phenotypes collected from 23andMe customers in order to cluster GWAS SNPs by functional category. We jointly model the correlation structure between a set of phenotypes using generalized estimating equations and look across all SNPs associated with at least one of the phenotypes, allowing us to classify them according to their effect across all the phenotypes. We apply this method to three examples from the 23andMe dataset, showing how correlated phenotypes can provide insight into the function of disease associated loci. First, we look at breast cancer SNPs using breast size as a proxy. There is at most a weak correlation between breast size and breast cancer risk, however we find that many of the loci associated with breast cancer also show associations with breast size. Of these, most have the cancer risk allele leading to larger size, however, several lead to smaller size, leading to clues about the functions of these regions. Next, we study BMI and food choice. It is suspected that several BMI SNPs, such as those in FTO, are related to BMI through neurological pathways. Here, we show that indeed, several of the 32 SNPs associated with BMI show a consistent signal across several food choice phenotypes, whereas others don't, suggesting a division of BMI loci into those involved in BMI through food choice and those through other pathways. Finally, we jointly investigate hair color and basal cell carcinoma. Our method leads to the discovery of several novel associations (including TGM3, TERT, and GATA3 for skin cancer). These SNPs show a firm division of skin cancer loci into those related to pigmentation and those shared with other cancers. We believe this method will be of increasing importance as the number of phenotypes studied grows.
You may contact the first author (during and after the meeting) at