Several machine learning methods used in data mining may be more appropriate to discover and describe the genetic base of complex traits [119]. Classification and regression trees (CART) [120], random forests [121], and Bayesian networks [122] have been proposed for modeling complex gene-environment interactions when the phenotype is a well defined variable [123,124]. CART is a multivariate statistical technique that creates a set of if-then rules linking combinations of genotypes and environmental exposures to the phenotypes. The if-then rules are created with a recursive procedure that groups data into sets to maximize the overall information. Random forests are an expansion of CART that have shown particular promise for the analysis of genotype–phenotype correlations, taking into account gene–gene interactions. The intuition is to use permutation methods and bootstrapping techniques to create thousands of CART models. From the analysis of these models, one can produce an importance measure for each SNP that takes into account interactions with other SNPs that affect the phenotype [125]. It has been shown that when unknown interactions among SNPs and genetic heterogeneity exist, random forest analysis can be substantially more powerful than standard univariate screening methods [126].
