Case(Control)-Free Multi-SNP Combinations in Case-Control Studies

Several genome-wide searches for disease-associated gene variations have been recently reported (Spinola et al, 2006; Herbert et al, 2006). However, complex diseases can be caused by combinations of several unlinked gene variations. Unfortunately, an exhaustive search among all possible corresponding multi-SNPs can be unfeasible even for small number of SNPs let alone the complete genome. Disease association analysis searches for a SNP with frequency among diseased individuals (cases) considerably higher than among non-diseased individuals (controls). In this work we first explore the problem of searching for the most disease-associated and the most disease-resistant multi-gene interactions for a given case-control study. In the previously published work [3] we proposed fast complimentary greedy search (CGS) which finds multi-SNP combinations with non-trivially high association on real data. Here we present two new approaches which are slightly slower than CGS but aimed to find more MSCs associated with a disease. One is k-level CGS (k-CGS) which is a modification of CGS with fixed k SNPs. Second one is k-level Alternating Closure Search (k-ACS) which starts with a set defined by k-length MSC found by the exhaustive search and minimizes its p-value by reducing the subset of controls and keeping the subset of cases as large as possible. We compare proposed methods with CGS on case/control datasets of Crohn’s disease (Daly et al, 2001), autoimmune disorder (Ueda et al, 2003), tick-borne encephalitis (Barkash et al, 2006), and rheumatoid arthritis (GAW15 NARAC 18q, 2006). When dealing with common diseases, it is necessary to search and analyze multiple independent causes each resulted from interaction of multiple SNPs scattered over the entire genome. Exploiting k-CGS and k-ACS methods for searching associated risk and resistance factors, we address the disease susceptibility prediction problem. We use greedy version of proposed in [3] optimum clustering formulation to extract independent causes. Then we use model-fitting algorithm that transforms clustering algorithm into susceptibility prediction algorithm. We compare the accuracy of the proposed prediction methods (k-CGSP and k-ACSP) with previously known methods using leave-one-out (LOO) test.