Greedy Approach to Reliable Disease Susceptibility Prediction

One of the main problems in genetic epidemiology is to robustly predict genetic susceptibility to complex diseases based on the data from case/control studies. This becomes computationally challenging in presence of interactions between multiple genes. In order to efficiently search through enormous amount of possible combinations, it is necessary to apply heuristics and the greedy approach has been successfully validated on many real data. In this paper we modify the genotype covering phase of the model-fitting susceptibility prediction algorithm from [1]. We improve reliability of the previously known prediction method based on greedy approach by replacing clean genotype coverage with dirty coverage, i.e., clusters participating in coverage do not overlap or overlap respectively. We have leave-one/many-out cross-validated existed and proposed prediction methods on real case/control studies of four diseases (Chron’s disease, autoimmune disorder, tick-born encephalitis, and lung cancer). Our results show that relaxation of the clean coverage significantly improves reliability of the greedy based susceptibility prediction approach.