Interactions between unlinked loci, or epistatic effects, have important roles in shaping human phenotypic variability and disease susceptibility. With the resources and technology for conducting large-scale association studies now imminent, there is a growing need to develop new analytical strategies for detecting and characterizing such interactions on a genome-wide scale.

Detecting interactions between loci requires statistical corrections to account for the extensive multiple testing inherent to multi-locus search strategies, and it is generally assumed that such corrections will weaken the power of the approach. To test this assumption, Jonathan Marchini and colleagues carried out a series of simulation studies to evaluate the performance of analytical strategies that look for interactions between loci. They considered several different two- and three-locus inheritance models using three different gene-detection strategies: a locus-by-locus search, a search over all combinations of loci and a combination two-stage strategy. Surprisingly, they found that, for a large number of epistatic configurations having equal single-locus effects, the interaction-based search strategies were more powerful than the locus-by-locus search strategy, even when a conservative (Bonferroni) correction for multiple testing was used. They concluded that, when analysing genome-wide association data, it might often be advantageous to explore models that explicitly allow for interactions between loci. In particular, they recommend a two-step approach in which single-locus effects are first detected using liberal statistical criteria, followed by a search for all possible interactions among the detected loci under rigorous criteria, corrected for multiple-testing.

The simulation model also reveals that, when interacting loci have different allele frequencies across study populations, the differences in power to detect the marginal effects of each locus can hinder reproducibility. This finding supports recent assertions that the use of locus-by-locus strategies to analyse interacting loci might be a major contributor to the lack of replication in association studies.

The importance of considering genetic epistasis in the analysis of complex traits is becoming increasingly recognized. By showing here that analytical strategies that explicitly consider interactions between loci are computationally feasible and often yield increased power, Marchini and colleagues offer clear guidance to the community about the design and analysis of future genome-wide association studies.