Background: it is increasingly recognized that gene-gene interaction is an important genetic component for complex diseases.Although several metrics for detecting gene-gene interactions have been developed, there is dearth of knowledge regarding their statistical power under different scenarios of interactions.This study thus aimed at systematic evaluation of the merits of three data mining metrics for detecting gene-gene interactions under various genetic models.Methods: In this article, we compared three different metrics of non-linear dependence between two genetic loci for detecting gene-gene interactions.The three metrics were two-loci Fst, mutual information and linkage disequilibrium, respectively.We generated case-control data by using simulation software genomeSIMLA based on 17 genetic models (including both epistatic and multiplicative interaction models).For each model, two different sets of disease allele frequencies and five different interaction effects were explored.The statistical power was computed from 500 simulated datasets.Results: In eight epistatic models, except for the model of recessive versus recessive, two-loci Fst and mutual information had higher statistical power than the linkagedise-quilibrium based metric.In nine multiplicative models, two-loci Fst and mutual information also had higher statistical power.Conclusions: The evolution-based metric, Fst, and information-based metric, mutual information, are more powerful and robust statistic for detecting both epistatic and multiplicative gene-gene interactions .