Uma metodologia para detecção de interações epistáticas em estudos de associação

Abstract

Genomic Wide Association Studies (GWAS) aims to identify SNPs that in uence a particular phenotype, such as speci c characteristics or diseases. SNPs are responsible for allele formation and this markers are used to identify a lócus that may represent a close correlation to a gene or the mutation itself. To determine the genetic mechanisms that in uence the phenotype are used thousands or even hundreds of thousands of SNPs that are genotyped from two groups of individuals: case and control. The reason behind some complex diseases such as cervical cancer, leukemia, type I and II diabetes involve multiple genes combining to express this phenotype. This interaction is known as epistasis. In epistasis a gene can inhibit or potentiate the e ects of the other. From the statistical point of view, the objective is to nd a non-linear mapping between the genotype and the phenotype. In addition to the complexity of nding causal SNPs involved in epistatic interactions, most complex diseases have low heritability. Thus, the lower the heritability, the greater the in uence of environmental factors and the less the genotype explanation of the phenotype. Most of GWAS use statistical tests of p-value hypotheses of each SNP present in the data set. These studies are often based on regression models. However, this type of model is e cient to capture only marginal e ects or e ects of additive gene actions. Cases where the genotype-phenotype relationship is linear. This work proposes a methodology capable of nding non-linear relationships between genotype-phenotype in data sets with large amounts of SNPs. The approach was also developed to deal with di erent levels of heritability. The proposed model is composed of three distinct steps. The rst step is responsible for identifying and selecting subgroups of signi cant SNPs. The dataset is partitioned into small xed-size groups and all possible permutations of each group are sorted by cross-validation by a ensemble method of boosting classi ers. The best groups are selected for the later stage. In the second step a process of ranking the selected SNPs from the subgroups de ned in the rst step is performed. A random forest model with potential to capture the relevance of the SNPs evaluated is applied at this stage of the process. The ranking of markers serves as a reference for the generation of a population markers, which will serve as the basis for the application of an evolutionary method of genetic programming that aims to determine possible associations between the SNPs ranked. At the end of the process, the genotypic relations that express the phenotype of interest from epistatic interactions are presented, based on the interpretability of the generated rules. The proposed methodology was compared with other models in the literature, including the reference method known as MDR, a variation of the MDR with an initialization method known as ReliefF and GPAS. Several experiments were carried out with simulated datasets, including data sets composed of 100, 1000 and 10000 SNPs with di erent levels of heritability varying from 0:4 to 0:1 and MAF of 0:4 and 0:2. The method was analyzed in data presenting epistasis without main e ect in 70 models known in the literature with heritability ranging from 0:4 to 0:01. We also performed experiments with interactions between more than two SNPs and experiments with data sets that present an expressive number of SNPs. The results denote the use of the methodology is promising compared to other models in GWAS literature.