Venous Thrombosis (VT) is a common complex disease affecting ~0.2% of individuals a year. VT includes deep vein thrombosis and pulmonary embolism, the latter being characterized by a one year mortality rate of ~10% excluding patients with malignancies [1]. As a complex trait, VT is considered as resulting from the interplay between environmental and genetic factors, that could interact with each other, to modulate VT risk [2, 3]. The recent Genome Wide Association Studies (GWAS) strategy brought great hopes to identify novel susceptibility loci to human diseases and some true successes were obtained in the field of VT genetics. Novel genes recently identified to harbor common susceptibility alleles (i.e. with allele frequency > 0.05) for VT include GP6, HIVEP1, KNG1, STAB2, STXBP5 and VWF (reviewed in [4]). However, none of the identified risk alleles demonstrated genetic effects stronger than those of the established VT-associated genes known before the GWAS era, ABO, F2, F5 and FGG[5]. As for most multifactorial diseases, risk alleles for VT identified so far only explain a small proportion of the familial risk of disease [6]. Alternative strategies are needed to identify the army sources that could contribute to the unexplained heritability and these include gene-gene and gene-environment interactions, deep sequencing, transcriptomic analyses and epigenomics [7–10].

In this work, we were interested in assessing whether interaction between common polymorphisms could contribute to VT risk. To our knowledge, studies that have investigated this hypothesis were mainly dedicated to known candidate genes [11, 12] and no attempt has been made to address it without any a priori hypothesis. This is why, we here take advantage of the large amount of genetic information we have collected through two French GWAS on VT [6, 13] to conduct the first genome-wide search for SNP x SNP interaction with respect to VT risk.

This work was based on two French GWAS on VT, the Early-Onset Venous Thrombosis (EOVT) and the Marseille Thrombosis Association (MARTHA) studies. These two studies have already been extensively described in [5, 6, 14] for EOVT and in [6, 15–17] for MARTHA.

We first applied a pairwise tagging approach to discard redundant SNPs using a r² threshold of 0.90, that led to the final selection of 243,189 SNPs from the EOVT study.

2.96 10¹⁰ pairwise SNPs interactions were then tested in EOVT, but none of them reached the Bonferroni corrected p-value of 1.69 10^-12. Nevertheless, all interactions with p-value less than 10^-4 (n = 2,126,084) were further assessed in MARTHA. The smallest observed p-value was 6.73 10^-7, but it did not pass the Bonferroni correction (p < 2.35 10^-8) for the number of interactions tested at this second step.

To our knowledge, this work is the first attempt in the field of VT genetics to investigate, at the genome-wide scale, the presence of interactive effects derived from common SNPs. This study did not detect interactions that reached the Bonferroni correction for the number of investigated interactions. The absence of such interaction could of course be due to low power. According to the distributions of the minor allele frequencies and the marginal allelic effects observed in the EOVT study, we computed the minimum OR for interaction that could be detectable with a 80% power [24, 25]. These calculations suggest that our discovery cohort was only well powered to detect interactive ORs greater than 2.8 at the genome-wide statistical level of 1.69 10^-12 and ORs greater than 1.8 at the p <10^-4 threshold [Additional file 2]. The power to detect in our second sample the most significant observed interactions was about 50% [24, 25]. As a consequence, despite the use of two large GWAS datasets on VT, this study was not powerful enough to detect interactions between common SNPS characterized by interactive ORs smaller than ~2.

There is still no consensus about the most efficiency way to perform a genome-wide search for SNP x SNP interaction. A plethora of statistical methods are applicable to the detection of such interactions eg [8, 26–29] and none of them could be considered as the panacea. Comparing the performances of different methodologies is of great importance but out of the scope of this manuscript. We rather focused in the present work on the application of a standard methodology, the logistic regression model, that has been shown to be a valid methodology for detecting interaction between SNPs [8]. Different strategies can still be adopted within the logistic regression framework. Some people advocate to restrict the search for interaction to the set of most “significant” SNPs observed in single locus analysis. However, in that case, which statistical threshold should be used for selecting SNPs with significant marginal associations? Nevertheless, we further confined our search for interaction to SNPs with statistical evidence for association in univariate analysis as low as p < 10^-3 or p < 0.05. We did not identify pair-wise significant interaction that were homogeneous between EOVT and MARTHA, and that satisfied the relevant Bonferroni correction (data not shown). Others suggest to use external biological information to refine the research strategy. Pathway-based analysis focusing only on the pairwise interactions between candidate gene SNPs could be such a strategy. By focusing only on SNPs mapping the VT candidate genes listing in the Supplementary Table 1 in [6], we did not detect any Bonferroni-corrected significant interaction that replicate in the EOVT and MARTHA study (data not shown). Another possibility could consist in assessing whether the most promising interactive effects could also be observed on quantitative traits known to be associated with the disease. Doing so, we observed that the rs9804128 and rs4784379 could interact to modulate both the risk of VT and the variability of FVIII levels. The rs9804128 lies in the proximal promoter of the IGFS21 gene and, according to the SNAP database [21], it is not in strong LD (r² > 0.8) with any other SNP. Conversely, the rs4784379 is in strong LD with several SNPs, all located at least 100 kb away from the IRX3 locus. However, the observed interaction could be considered as counterintuitive since the allele combination associated with increased FVIII levels was found less frequent in cases than in controls. This phenomenon could nevertheless be observed in presence of a mortality bias when patients with high levels of FVIII levels are at a higher risk of VT-associated mortality (eg. pulmonary embolism) and then under-represented in the cases sample. Further investigations are needed to replicate this association that involved SNPs at genes on which very little is known with respect to VT.

In conclusion, our work suggests that strong interactive (~OR >2) phenomena between common SNPs are unlikely to contribute much to the risk of the VT.