A Multi-objective Ant Colony Optimization Algorithm For Genome-wide Association Studies

The goal of genome-wide association studies(GWAS) is to explore strong associations between phenotypes and genetic variations in the whole genetics. Nowadays, we can see an increasing number of methods for genetic interaction detecting in the genome-wide association studies(GWAS). Although great success has been made by these methods, they suffer from many problems.On the one hand, various approaches perform significant differently on different disease models. One of the important reasons is that the existing approaches were constructed on only one correlation model while the different disease models vary a lot. Thus the one correlation model-based methods can’t fit the GWAS datasets well, resulting in low power and a high false-positive rate. On the other hand, compared with the high dimensional real datasets in GWAS, many existing approaches are limited to low dimensional datasets, which also restrict their applications in real datasets. To solve the aforementioned problems, the paper presents a multi-objective ant colony optimization algorithm for SNP epistasis detection. The work and novelties in this dissertation include:To solve the first problem, we for the first time implemented a multi-objective optimization framework based on swarm intelligence optimization to the GWAS field in this paper. In the algorithm, we combine both the standard logistical regression and the Bayesian network methods, which are from the opposing schools of statistics, and adopt the corresponding AIC score and K2 score to evaluate the constructed models. In the experiments,the combination of these two evaluation objectives is proved to be complementary to each other resulting in a performance of higher power and lower false positives.To solve the space and time complexity for large dimension problems in real GWAS datasets, we designed a wrapped feature selection protocol based on ant colony optimization, which is proved to can be applied in large datasets directly. In the experiments, we used a series of the simulated datasets and the real LOAD dataset to prove its effectiveness with high accuracy.