A novel evolution-based method for detecting gene-gene interactions

Abstract

Background: The rapid advance in large-scale SNP-chip technologies offers us great opportunities in elucidating the genetic basis of complex diseases. Methods for large-scale interactions analysis have been under development from several sources. Due to several difficult issues (e.g., sparseness of data in high dimensions and low replication or validation rate), development of fast, powerful and robust methods for detecting various forms of gene-gene interactions continues to be a challenging task.

Methodology/principal findings: In this article, we have developed an evolution-based method to search for genome-wide epistasis in a case-control design. From an evolutionary perspective, we view that human diseases originate from ancient mutations and consider that the underlying genetic variants play a role in differentiating human population into the healthy and the diseased. Based on this concept, traditional evolutionary measure, fixation index (Fst) for two unlinked loci, which measures the genetic distance between populations, should be able to reveal the responsible genetic interplays for disease traits. To validate our proposal, we first investigated the theoretical distribution of Fst by using extensive simulations. Then, we explored its power for detecting gene-gene interactions via SNP markers, and compared it with the conventional Pearson Chi-square test, mutual information based test and linkage disequilibrium based test under several disease models. The proposed evolution-based method outperformed these compared methods in dominant and additive models, no matter what the disease allele frequencies were. However, its performance was relatively poor in a recessive model. Finally, we applied the proposed evolution-based method to analysis of a published dataset. Our results showed that the P value of the Fst -based statistic is smaller than those obtained by the LD-based statistic or Poisson regression models.

Conclusions/significance: With rapidly growing large-scale genetic association studies, the proposed evolution-based method can be a promising tool in the identification of epistatic effects.

Figure 1. Frequency histogram of the Fst based statistic based on 10,000 simulations, compared with F(2, 3994).

The gray bar denotes the frequency histogram of the Fst based statistic, corresponding to the null hypothesis that two SNPs are of no interaction. The red line is the density curve of the theoretical one.

Figure 2. Frequency histogram of 2000×MI based on 10,000 simulations, compared with χ²(8).

The gray bar denotes the frequency histogram of 2000×MI, corresponding to the null hypothesis that two SNPs are of no interaction. The red line is the density curve of χ²(8).

Figure 3. Frequency histogram of the LD based statistic based on 10,000 simulations, compared with χ²(1).

The gray bar denotes the frequency histogram of the LD based statistic, corresponding to the null hypothesis that two SNPs are of no interaction. The red line is the density curve of the theoretical one.

Figure 4. Power of four statistics under three different models when the disease allele frequencies at the two loci are high.

The disease prevalence is assumed to be 1%. The disease allele frequencies at the two loci (g1 and g2) are 0.3 and 0.8, respectively. The power, at significance level α of 0.05, is obtained based on simulations of 500 cases and 500 controls. The green, red, blue, and cyan lines are the power of Fst based statistic, Pearson's Chi-square statistic, MI based statistic, and LD based statistic, respectively. Three plots (A, B, C) correspond to the dominant model, the additive model and the recessive model, respectively.

Figure 5. Power of four statistics under three different model when the disease allele frequencies at the two loci are low.

The disease prevalence is assumed to be 1%. The disease allele frequencies at the two loci are 0.2 and 0.4, respectively. The power, at significance level α of 0.05, is obtained based on simulations of 500 cases and 500 controls. The green, red, blue, and cyan lines are the power of Fst based statistic, Pearson's Chi-square statistic, MI based statistic, and LD based statistic, respectively. Three plots (A, B, C) correspond to the dominant model, the additive model and the recessive model, respectively.

Figure 6. Power of Fst under different parameter settings.

The disease prevalence is assumed to be 1%. The power, at significance level α of 0.05, is obtained based on simulations of 500 cases and 500 controls. Three solid colorful lines (red, green and blue) correspond to the power curves of the Fst-based statistic under three genetic models (dominant, additive and recessive), when the disease allele frequencies at the two loci (g1 and g2) are 0.3 and 0.8, respectively. The dotted lines are the power curves under the assumption that the disease allele frequencies at the two loci are 0.2 and 0.4, respectively.