Replication of large-scale epistasis studies: an example on ankylosing spondylitis

Ankylosing spondylitis (AS) is a common form of inflammatory arthritis occurring in approximately 5 out of 1,000 adults of European descent. Recently, the Australo-Anglo-American Spondyloarthritis Consortium and the WTCCC2 showed that polymorphisms of ERAP1 only affect AS risk in HLA-B27-positive individuals, hereby establishing an interaction between ERAP1 and HLA in the TASC, WTCCC2 and replication datasets [2,5]. We were able to confirm this interaction although using other SNPs.

In this study, we use the aforementioned data from WTTCC2 on AS to address unresolved issues when performing large-scale SNP-SNP interaction studies, so as to better guarantee “stable” and “truly replicable” results. These issues are 1) the choice of variable selection method (e.g., of known loci mapping to genes part of know pathways), 2) the choice of SNPs representing a genomic region (e.g., SNPs with modest versus negligible LD between them), 3) the choice of analysis method (e.g., regression-based versus data-reduction (non-parametric) based), 4) different adjustment schemes for lower-order effects (using additive/co-dominant genetic models). We show that even modest changes in 1)-4) may give rise to quite varying epistasis findings for AS, and motivate some “optimal” choices via extensive simulation studies.

In this work we rely on a minimal GWAI protocol for genome-wide epistasis detection using SNPs, as developed in our lab [6][9], using the advanced non-parametric Model-Based Multifactor Dimensionality Reduction (MB-MDR) method [1] and an adapted [*] BOolean Operation-based Screening and Testing (BOOST) algorithm [4].

[*] A BOOST [4] like implementation based on the original BOOST algorithm which accounts for missing genotypes