Multiobjective evolutionary computation for high-order genetic interactions

Abstract

A number of research works support the relationship between Single Nucleotide Polymorphisms (SNPs) and neurodegenerative diseases (e.g. Alzheimer’s or Parkinson’s Disease). It has been proven that these neurodegenerative diseases are mainly caused by the interaction of different SNPs. The complexity of identifying genetic interactions increases exponentially with two factors: (i) the number of SNPs contained in the biological dataset under study and (ii) the number of SNPs involved in the interaction. Therefore, this paper proposes the application of two of the most successful multiobjective evolutionary algorithms to solve this problem: a Reference-point based Many-objective Fast Non-dominated Sorting Genetic Algorithm (NSGA-III) and a Multiobjective Evolutionary Algorithm based on Decomposition with Dynamical Resource Allocation (MOEA/D-DRA). These algorithms have been tested with four datasets (including a real dataset about Bipolar Disorder with 425,574 SNPs) and three interaction sizes: 2, 5, and 8 loci. In addition, they have been compared against well-known and relevant approaches published in the literature, in both multiobjective and biological terms. The results clearly show the advantages of the approach based on NSGA-III. Particularly, NSGA-III improves the results obtained by other algorithms in multiobjective terms (by means of Hypervolume and Set Coverage indicators) and in biological terms (by means of Power, Recall, Precision, and F-measure metrics). Moreover, it reveals new 2 and 5 loci interactions over a Bipolar Disorder real dataset. Therefore, NSGA-III represents a relevant approach to detect high-order genetic interactions.