MultiPhen: Joint Model of Multiple Phenotypes Can Increase Discovery in GWAS

Abstract

The genome-wide association study (GWAS) approach has discovered hundreds of genetic variants associated with diseases and quantitative traits. However, despite clinical overlap and statistical correlation between many phenotypes, GWAS are generally performed one-phenotype-at-a-time. Here we compare the performance of modelling multiple phenotypes jointly with that of the standard univariate approach. We introduce a new method and software, MultiPhen, that models multiple phenotypes simultaneously in a fast and interpretable way. By performing ordinal regression, MultiPhen tests the linear combination of phenotypes most associated with the genotypes at each SNP, and thus potentially captures effects hidden to single phenotype GWAS. We demonstrate via simulation that this approach provides a dramatic increase in power in many scenarios. There is a boost in power for variants that affect multiple phenotypes and for those that affect only one phenotype. While other multivariate methods have similar power gains, we describe several benefits of MultiPhen over these. In particular, we demonstrate that other multivariate methods that assume the genotypes are normally distributed, such as canonical correlation analysis (CCA) and MANOVA, can have highly inflated type-1 error rates when testing case-control or non-normal continuous phenotypes, while MultiPhen produces no such inflation. To test the performance of MultiPhen on real data we applied it to lipid traits in the Northern Finland Birth Cohort 1966 (NFBC1966). In these data MultiPhen discovers 21% more independent SNPs with known associations than the standard univariate GWAS approach, while applying MultiPhen in addition to the standard approach provides 37% increased discovery. The most associated linear combinations of the lipids estimated by MultiPhen at the leading SNPs accurately reflect the Friedewald Formula, suggesting that MultiPhen could be used to refine the definition of existing phenotypes or uncover novel heritable phenotypes.

Evidence for Extensive Pleiotropy Among Pharmacogenes

“…When we performed the same multivariate analysis on other variants that approached phenome-wide significance in the single-trait analysis (SLC22A2 rs316019, DYPD rs1801265, NAT2 rs1799931, CYP2B6 rs8192709, CYP2D6 rs5030656, CYP2D6 rs35742686 and CYP2D6 rs5030655), we also did not observe a similar inflation of p-values (Supplementary Figure 4). Given the expected low type-1 error rate of MultiPhen [36], we can conclude that the inflation we observed for the SLC15A2 paired-trait analysis was likely driven by true pleiotropic effects across a spectrum of phecodes, specifically those associated with high morbidity, such as kidney transplant (Supplementary Figure 5).…”

“…We declared phenome-wide significance at a Bonferronicorrected phenome-wide significance in EuropeanAmericans (p < 6.18 × 10 -5 ) and African-Americans at (p < 1.50 × 10 -4 ). The paired trait analysis was performed with the R package MultiPhen [36]. In a MultiPhen analysis, the genetic variant is held as the dependent variable and a joint test of association is performed with a linear combination of phenotypes as predictor variables.…”

“…Among the novel results, we identified a significant association in European-Americans between a nonsynonymous SNP in the gene SLC15A2 rs1143672 and renal osteodystrophy, a disorder of mineral and bone metabolism (p = 2.67 × 10 -6 ; odds ratio [OR]: 0.61; 95% CI: 0.49-0.75). Follow-up tests of association for rs1143672 with all pairs of phenotypes using the multivariate platform MultiPhen [36] revealed significant associations (p < 1.88 × 10 -6 ) with either renal osteodystrophy or end-stage renal disease (ESRD) and abnormal heart sounds, skin neoplasm (of un certain behavior), joint pain and back pain. These results indicate that 'unmeasured phenotypes', which are not well represented in single phenotype analyses, likely contribute to an array of diseases recorded by International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) codes.…”

Joint association analysis of a binary and a quantitative trait in family samples

“…, the variance formula of U (2) will reduce to U Ã 22 À U Ã 21 U ÃÀ1 11 U Ã 12 (the subscript 1 and 2 corresponds to θ (1) and θ (2) , respectively). The test statistic with this restriction is termed as 'BivarEGEE R '.…”

“…2 However, none of these approaches are as powerful or efficient as a joint multivariate test with each trait treated as a dependent variable in discovering genetic loci associated with all traits under study. 1,3,4 For example, in the case of two continuous traits assumed to be normally distributed, a joint test can be derived as a simple extension of a univariate normal test. However, if one of the two traits is a discrete trait, for example, a binary trait, deriving such a test becomes challenging, and it further complicates in family samples.…”