Application of Self-Organizing Maps to Detect Population Stratification

Genetic epidemiology has become a major field of interest at the border of traditional epidemiology and genetics. Genetic data call for specific epidemiological and statistical methods that have to account on the one hand for the family structure within a data set and on the other hand for the paired information of two alleles at each gene locus (geno-type). In addition to these obvious dependencies genetic data can be very complex. Often enormous numbers of hypotheses are investigated simultaneously and various levels of data can be thought of, e.g. gene expression data from different pathways as well as data on the protein level. Complexity arises from possible interactions between genes and the environment and from the lack of methods to model biological interactions by statistical interaction terms. This paper aims to provide an alternative approach to the modelbased cluster analysis that is carried out within the structured association approach. The dependence on distributional assumptions (see Pritchard et al. (2000a)) when identifying the number of underlying subpopulations will be avoided by the application of an exploratory method, namely Self-Organizing Maps (SOMs). The paper organizes as follows. Section 2 will give a brief introduction to the idea of the Self- Organizing Maps. Section 3 investigates the limitations of this proposed method as clustering tool by means of a simulation study. The focus is on the identification of different population structure models and thus on the identification of the number of underlying subpopulations. Some concluding remarks will be given in Section 4 of the paper.