A comparative analysis of machine learning algorithms for genome wide association studies

Variations present in human genome play a vital role in the emergence of genetic disorders and abnormal traits. Single Nucleotide Polymorphism (SNP) is considered as the most common source of genetic variations. Genome Wide Association Studies (GWAS) probe these variations present in human population and find their association with complex genetic disorders. Now these days, recent advances in technology and drastic reduction in costs of Genome Wide Association Studies provide the opportunity to have a plethora of genomic data that delivers huge information of these variations to analyze. In fact, there is significant difference in pace of data generation and analysis, which led to new statistical, computational and biological challenges. Scientists are using numerous approaches to solve the current problems in Genome Wide Association Studies. In this thesis, a comparative analysis of three Machine learning algorithms is done on simulated GWAS datasets. The methods used for analysis are Recursive Partitioning, Logistic Regression and Naïve Bayes Classifier. The classification accuracy of these algorithms is calculated in terms of area under the receiver operating characteristic curve (AUC). Conclusively, the logistic regression model with binary classification seems to be the most promising one among the other four algorithms, as it outperformed the other tools in the AUC value