DEPTH: A Novel Algorithm for Feature Ranking with Application to Genome-Wide Association Studies

Analysis of genome-wide association studies (GWAS) involves selecting variables from a large set of variables, an issue common to science. The sample could contain people affected by a disease (cases) and people that are not affected by the disease (controls), though other designs are possible and the approach below applies to non-human samples. The ultimate aim of a GWAS is to identify susceptibility genes, and the dominant paradigm to date has been to do this by identifying single nucleotide polymorphism (SNPs) associated with risk of disease. We propose a new algorithm, called DEPendence of association on the number of Top Hits (DEPTH), for variable selection based on permutation statistics and stability selection. DEPTH is: (i) applicable to any parametric regression model, (ii) designed to be run in a parallel computing environment, and (iii) exploits information from the correlation structure of the predictors. In the context of a GWAS, the algorithm can be used for all SNPs in the genome, or for any subset of SNPs (e.g. in a region, in a pathway, functional, etc.). We have found that the algorithm shows good performance when compared to several established procedures using simulated data. We applied DEPTH to a breast cancer GWAS of only 204 cases (about 50:50 ER-ve and ER+ve) and 287 controls. DEPTH found evidence that variants in a particular gene are more associated with ER-ve breast cancer, and more so than ER+ve disease. This finding has been replicated using a large international breast cancer case-control GWAS data set (55,540 cases, 51,168 controls), and would have been unlikely to be identified by conventional analyses. DEPTH is currently running on an IBM BlueGene/Q supercomputer, though versions can also be run on a lap top for targeted analyses. A summary of findings and insights into aetiology and method performance will be presented.