Corrigendum of ‘High throughput analysis of epistasis in genome-wide association studies with BiForce’

Abstract

Contact: Wenhua.Wei@igmm.ed.ac.uk

Following the publication of our article, describing the use of BiForce (http://bioinfo.utu.fi/biforcetoolbox) for the analysis of epistasis (Gyenesei et al., 2012), we observed that inflated evidence for epistasis may arise under exceptional circumstances when analyzing quantitative traits. This may occur when two neighboring (e.g. <200 kb apart) single nucleotide polymorphisms (SNPs) in an epistatic pair are in linkage disequilibrium (LD) and at least one of them carries strong marginal effects. Similar inflation was discovered recently in other LD- or haplotype-based methods for the analysis of epistasis in disease traits (Ueki and Cordell, 2012). This issue does not affect the analysis of disease traits in our case because BiForce uses logistic regression as the final step to generate the results for such traits (Wan et al., 2010). Thus, Table 3 of the original paper (Gyenesei et al., 2012) is correct. However for quantitative traits, BiForce uses contingency table-based F ratio tests for interactions without the fitting step applied in linear regression. It is known that such tests are not orthogonal, but they are robust when LD between two SNPs is low, allowing the fast screening achieved by BiForce. When LD is high, however, the test for interaction between two correlated SNPs is inflated by the marginal effects of the pair of SNPs, and therefore the inflation is critical when marginal effects are strong but not when marginal effects are weak. Nevertheless, because BiForce uses stringent Bonferroni-adjusted thresholds by default, the chance of inflated epistatic pairs being genome-wide significant should be low in general.

This issue affected the results in Table 2 of the original article (Gyenesei et al., 2012). The correct interaction P-values (P_int) of each SNP pair are listed in the updated Table 2 later in the text, suggesting that none remained genome-wide significant in C-reactive protein (CRP), glucose (GLU), high-density lipoprotein (HDL), low-density lipoprotein (LDL) and triglycerides (TRI). Results from the analyses of simulated data on quantitative and disease traits are unaffected because in simulation SNPs were randomly drawn from a chromosome assuming they were in Hardy–Weinberg equilibrium, i.e. the chance of high LD coming together with strong marginal effects is very low, which is supported by the results of false positive rate in Figure 3 of the original article. In summary, the issue only affects a small part of the results (i.e. Table 2) concerning analyses of quantitative traits in real data. The main biological results and overall conclusions are unaffected. A script to address this issue is available from the dedicated website and will be incorporated in the next BiForce release.

Table 2.

Previous genome-wide significant epistatic pairs identified from the NFBC199 cohort^a (update)

Trait	SNP1	SNP2	Pint	distance	LD (r2)	correct Pint
CRP	rs1811472b (1q23.2; 0.41)	rs2592887b (1q23.2; 0.40)	3.0E-12	10 590	0.86	2.1E-01
CRP	rs1811472b (1q23.2; 0.41)	rs2794520b (1q23.2; 0.36)	3.5E-11	36 467	0.62	1.1E-01
CRP	rs2592887b (1q23.2; 0.40)	rs2794520b (1q23.2; 0.36)	2.9E-12	25 877	0.70	1.6E-01
CRP	rs2650000b (12q24.31; 0.45)	rs7953249b (12q24.31; 0.48)	2.6E-09	14 762	0.76	9.1E-01
CRP	rs1169300b (12q24.31; 0.32)	rs2464196b (12q24.31; 0.32)	3.4E-10	4202	0.99	4.1E-01
GLU	rs560887b (2q31.1; 0.30)	rs563694b (2q31.1; 0.34)	1.3E-08	10 923	0.81	5.2E-01
HDL	rs3764261b (16q13; 0.28)	rs1532624b (16q13; 0.41)	2.0E-14	12 155	0.53	6.8E-01
LDL	rs157580b (19q13.32; 0.29)	rs405509 (19q13.32; 0.46)	6.9E-10	13 570	0.35	1.6E-04
TRI	rs1260326b (2p23.3; 0.36)	rs780094 (2p23.3; 0.36)	5.8E-08	10 297	0.95	9.6E-01

^aAll SNP pairs listed detected as marginal-SNP interactions, with the threshold of 1.5E-08 for CRP, 2.2E-08 for HDL, 3.9E-08 for GLU and LDL, 7.7E-07 for TRI; SNP₁ (SNP₂) – name, genomic location and minor allele frequency (the latter two in bracket) of the first (second) SNP; P_int – P-value of the interaction test; distance – the distance in base pairs between two SNPs; LD – linkage disequilibrium (in r²) between a pair of SNPs; the SNP pair in HDL was also detected via the pair-wise genome scan (P < 9.54E-13); correct P_int – the corrected P-value of the interaction test. ^bThe marginal-SNP.

Conflict of Interest: none declared.