Volume 14 Supplement 3
SNPSIG 2012: Identification and annotation of SNPs in the context of structure, function, and disease
GWIS  modelfree, fast and exhaustive search for epistatic interactions in casecontrol GWAS
 Benjamin Goudey^{1, 2}Email author,
 David Rawlinson^{1, 3},
 Qiao Wang^{1, 2},
 Fan Shi^{1, 2},
 Herman Ferra^{1},
 Richard M Campbell^{1, 3},
 Linda Stern^{2},
 Michael T Inouye^{4, 5},
 Cheng Soon Ong^{1, 2} and
 Adam Kowalczyk^{1, 2}Email author
DOI: 10.1186/1471216414S3S10
© Goudey et al.; licensee BioMed Central Ltd. 2013
Published: 28 May 2013
Abstract
Background
It has been hypothesized that multivariate analysis and systematic detection of epistatic interactions between explanatory genotyping variables may help resolve the problem of "missing heritability" currently observed in genomewide association studies (GWAS). However, even the simplest bivariate analysis is still held back by significant statistical and computational challenges that are often addressed by reducing the set of analysed markers. Theoretically, it has been shown that combinations of loci may exist that show weak or no effects individually, but show significant (even complete) explanatory power over phenotype when combined. Reducing the set of analysed SNPs before bivariate analysis could easily omit such critical loci.
Results
We have developed an exhaustive bivariate GWAS analysis methodology that yields a manageable subset of candidate marker pairs for subsequent analysis using other, often more computationally expensive techniques. Our modelfree filtering approach is based on classification using ROC curve analysis, an alternative to much slower regressionbased modelling techniques. Exhaustive analysis of studies containing approximately 450,000 SNPs and 5,000 samples requires only 2 hours using a desktop CPU or 13 minutes using a GPU (Graphics Processing Unit). We validate our methodology with analysis of simulated datasets as well as the seven Wellcome Trust CaseControl Consortium datasets that represent a wide range of real life GWAS challenges. We have identified SNP pairs that have considerably stronger association with disease than their individual component SNPs that often show negligible effect univariately. When compared against previously reported results in the literature, our methods redetect most significant SNPpairs and additionally detect many pairs absent from the literature that show strong association with disease. The high overlap suggests that our fast analysis could substitute for some slower alternatives.
Conclusions
We demonstrate that the proposed methodology is robust, fast and capable of exhaustive search for epistatic interactions using a standard desktop computer. First, our implementation is significantly faster than timings for comparable algorithms reported in the literature, especially as our method allows simultaneous use of multiple statistical filters with low computing time overhead. Second, for some diseases, we have identified hundreds of SNP pairs that pass formal multiple test (Bonferroni) correction and could form a rich source of hypotheses for followup analysis.
Availability
A webbased version of the software used for this analysis is available at http://bioinformatics.research.nicta.com.au/gwis.
Background
Genomewide association studies (GWAS) have discovered many underlying genetic causes of disease, but have also raised important questions about standard approaches to modelling complex traits [1]. While commonlyused univariate analysis techniques have been able to detect a number of significantly associated loci, for many conditions these discovered variants do not account for a majority of the theoretical estimates of genetic heritability. Multivariate approaches may help to alleviate this issue of "missing heritability" [2]. Theoretically, it has been shown that 2way and 3way single nucleotide polymorphism (SNP) interactions can explain up to ~ 50% and ~ 100% of trait variance while each SNP involved explains none [3], indicating that critical SNP pairs may be ignored by univariate analysis predominantly applied to GWAS so far. It is hypothesised that systematic detection methods may assist discovery of such potentially epistatic interactions between DNA loci.
Motivation
To date there exists little experimentallyvalidated evidence of SNP interactions in humans, largely due to the complexity of multivariate GWAS analysis. Even in only bivariate analysis, the number of possible SNP interactions that need to be searched is extremely large, as there are 125 billion possible SNP pairs in a GWAS of 500,000 SNPs. The scale of the problem produces significant computational and statistical challenges. Numerous approaches proposed to address these challenges are unable to scale to this large number of tests, due to both performance and accuracy (a large number of falsepositive results are expected from so many tests, generating concerns about the effectiveness of multipletest correction). This has led to claims that finding epistatic interactions via exhaustive search is infeasible [4, 5]. While these pessimistic claims have recently been proven wrong (e.g. [6–10]), techniques that do scale to exhaustive search currently require weeks or months to process GWAS of 5 million SNPs, which are becoming increasingly common. As GWAS studies continue to grow in size, faster analysis techniques will be needed. This paper aims to offer solutions that meet these everincreasing requirements.
Epistatic interactions
Our goal is to present a system capable of exhaustive search through all SNP pairs in an entire GWAS, detecting all significant epistatic interactions. As discussed in [11], both the terms "significant" and "epistatic interaction" have diverse definitions when used by biologists, epidemiologists, statisticians and geneticists and are often not made explicit. We specify the precise meanings of these terms as used in this paper, presenting a verbal description now and a more specific elaboration in the Methods section. We say that two SNPs have an epistatic interaction if using both of them allows discrimination between Cases and Controls with significantly higher sensitivity and specificity than is possible using any one of them individually. The significance is quantified as a pvalue for rejection of a well specified null hypothesis (see Methods for details). This rejection implies in particular, that the improvement cannot be explained by biased sampling from a population preclassified by any one of the SNPs in the pair. In the Discussion section we argue that our generic formal definition of epistasis captures some biological aspects of epistasis that Fisher's popular definition of interaction [12] misses.
GWIS approach
The definitions given above can be directly converted into computational methods, suitable for scanning trillions of SNP pairs in a modern GWAS and providing an alternative to widelyused regression based approaches. In this work, we present a platform called Genome Wide Interaction Search (GWIS), that is based on classification, and novel rigorous statistical tests based on receiver operating characteristic (ROC) curve analysis [13]. Our proposed method is genuinely "model free", since we do not assume any interaction model between SNP genotypes. In this sense we are close to other model free approaches, in particular Multifactor Dimensionality Reduction (MDR) [14–16], although we rely on analytical solutions to hypothesisbased testing rather than slower, computationallycostly crossvalidation and permutation testing.
We demonstrate that exhaustive search of all possible pairs in standard GWAS is feasible and fast on a desktop computer and that our proposed technique is faster than currently available exhaustive techniques. Aside from the computational challenges mentioned above, there are a number of statistical challenges that also need to be resolved. Principled methods are needed that allow for significancecorrection of the billions of SNPpair and genotype combinations, and that are able to cope with characteristics of realworld data, e.g. confounding factors due to strong univariate signals, examining significance in the far tail of distributions where the central limit approximation does not hold, and SNPs with low minor allele frequency giving rise to very low genotype counts.
We introduce a novel and theoretically wellfounded, modelfree hypothesis test specifically designed for multivariate GWAS analysis. It is based on relating the sensitivity and specificity observed in the sampled data to the sensitivity and specificity that could be achieved in the 'true' population. The test, named gain in sensitivity and specificity (GSS), is designed to detect epistatic SNP interactions, and computes exact pvalues, without using large sample normal approximations. Each application of the GSS test to a pair of SNPs involves solving a number of minmax optimisations, which are pair specific and are therefore impractical for scanning trillions of putative SNPpairs. Hence, we introduce two extra tests, referred to as sensitivity and specificity (SS) and difference in sensitivity and specificity (DSS), that act as practical fast proxies for the GSS test.
Validation
Algorithms for detecting epistatic SNP interactions are typically evaluated using simulated data, for reasons of both scalability and interpretation [17–19]. However, the creation of realistic structure in simulated data is problematic as much is unknown about the nature and existence of epistasis in humans [20, 21]. Therefore, we primarily focus on seven GWAS datasets from the Wellcome Trust CaseControl Consortium (WTCCC) [22]. These data include various real GWAS challenges that are not always represented in simulated data. Although the set of true SNP interactions is not yet known for WTCCC data, analysis of this data using multiple types of analysis provides evidence on the properties of the epistatic interactions that can be observed, reveals confounding factors not generally modelled in synthetic data, and demonstrates the advantages and limitations of different statistical filtering approaches. The efficiency of our methods is demonstrated by comparing timings of our methods on various size datasets to those reported in several recent publications. The proposed statistical filters are further benchmarked by confirming their theoretically advantageous properties and validation of their power and false positive rates over an extensive collection of synthetic datasets available from [23]. We show the importance of exhaustive search without which heuristics may miss significant SNP pairs. We demonstrate that our GSS test is able to identify a number of interesting SNP pairs that show significant epistatic effects. Detected results are compared to those from existing literature, showing that GWIS repeats many known results, as well as suggesting many novel interactions.
Contributions
This paper makes several contributions. First, we use an operational definition of epistasis based on classification of individuals into Cases or Controls to develop a set of robust, principled methods for explicitly detecting significant epistatic interactions in GWAS data. Second, we demonstrate that our proposed methods scale well and are fast enough to permit exhaustive analysis of current and nearfuture GWAS data. Third, we have applied GWIS to a diverse range of both simulated and real life benchmark data, and detected many significant associations in addition to confirming many associations previously reported. Finally, our analysis of real data indicates the limitations of conventional statistical methods such as Pearson's ${\chi}^{2}$ test for detecting epistatic interactions in the presence of strong main effects.
Results
An exhaustive evaluation of all possible SNP pairings is the most powerful strategy to detect epistatic interactions [24] but to date remains a computationally challenging task. Most methods have been unable to scale exhaustive methods to entire GWAS without performing some reduction in the number of pairs to be evaluated [5], or requiring special hardware such as a compute cluster [25–27].
Comparison of computation time
Runtime required for GWIS compared to recent CPU and GPU methods.
Method  Time for n_{ SNP }× n_{ samples }  Exhaustive Search  

0.45M × 5K  5M × 10K  
CPU Implementation:  
(4 cores, 64 bits, 3GHz Intel)  
GWIS (1 filter)  2.7 hours*  28 days  Yes 
GWIS (3 filters)  10.9 hours*  113 days  Yes 
BOOST [31]  23 hours  8 months  Yes 
PLINK [53]  89 days  60 years  Yes 
RAPID [32]  15 mins  NA  No 
SIXPAC [24]  8.0 hours  NA  No 
GPU Implementation:  
(448 CUDA cores, 1.215 GHz, NVIDIA GTX 470)  
GWIS (1 filter)  13 mins*  2.2 days  Yes 
GWIS (3 Filters)  22 mins*  3.8 days  Yes 
GBOOST [7]  1.4 hours  15 days  Yes 
EpiGPU [8]  17 hours  6 months  Yes 
SHEsisEPI [9]  28 hours  10 months  Yes 
EPIBLASTER [6]  8.9 days  6 years  Yes 
Table 1 demonstrates that exhaustive evaluation of all possible SNP pairs is feasible on a standard desktop machine with GWIS taking 2.7 hours for CPU and 13 minutes for GPU implementations. This represents an approximate 9× and 6× speed up over other alternative CPU and GPU exhaustivesearch methods respectively, and is faster than many methods that use heuristic search strategies. The only faster method reported here is a nonexhaustive search algorithm RAPID, whose timing reported here excludes parameter tuning that increases the actual time dramatically and has profound impact on performance (see the following Section).
For GWIS, we report runtime using one filter and three filters, namely ${\chi}^{2}$ alone or in combination with DSS and SS tests. The latter two tests are more computationally intensive than most existing statistical filters such as ${\chi}^{2}$, Difference of Odds (DoO) and the Fisher Exact test (FE). Approximately 60% of the runtime for ${\chi}^{2}$ alone is spent computing contingency tables, that are subsequently used by all statistical tests. On the reference machine used for CPU results in Table 1, ${\chi}^{2}$ alone runs in 2.7 hours. ${\chi}^{2}$, DoO and FE can be completed in 4.6 hours. ${\chi}^{2}$, DSS and SS requires 10.9 hours.
If we consider arrays of 5M SNPs, the estimated difference in times shows the necessity of faster exhaustive methods. Many algorithms that had acceptable runtime on current size GWAS will take weeks or months to compute on the larger number of SNPs as the total number of pairs to be evaluated grows quadratically. While the CPU implementation of GWIS would require about 3 months, the GPU implementation requires 3 days, a feasible wait for research results. Both CPU and GPU implementations could be deployed on a computing cluster to easily reduce this runtime down to a few minutes.
We expected the runtime of our methods to increase linearly with the number of samples and quadratically with increasing SNPs (i.e. linear in terms of SNPpairs). To verify this, we examined program runtime on simulated datasets varying both the number of samples and the number of SNPs. These datasets contained between 125K and 1M SNPs and between 1250 and 10K samples. Due to the independence of computations on each SNPpair, both CPU and GPU implementations show the expected relationships between samples, SNPs and runtime. Note that actual timings will be affected by machine architecture; in addition to obvious factors such as clock speed, we exploit lowlevel functions that are found in most modern CPUs. Older CPUs without high performance functions will not execute GWIS as quickly.
Summary and analysis of interactions detected using different statistical filters
Abbreviations and number of Case samples for each WTCCC dataset.
Abbreviation  Num. of Cases  Disease 

BD  1868  Bipolar Disorder 
CAD  1926  Coronary Artery Disease 
CD  1748  Crohn's Disease 
HT  1952  Hypertension 
RA  1860  Rheumatoid Arthritis 
T1D  1963  Type 1 Diabetes 
T2D  1924  Type 2 Diabetes 
Summary of the number of SNP pairs detected by different filtering methods.
Univariate  Bivariate Filter  Bivariate Filter + GSS  

Dataset  log_{10}${P}_{{\chi}^{2}}$  ${\chi}^{2}$  DSS  GBOOST  ${\chi}^{2}$  DSS  GBOOST  
HT  9.8  128  429  51  41  107  24  
BD  10.9  2445  556  34  44  179  27  
CAD  13.1  210147  7807  43  42  116  39  
T2D  13.3  56592  3105  52  79  134  41  
CD  34.3  > 500000^{∗}  5591  25  29  57  22  
RA  37.7  > 500000^{∗}  823  99  59  312  95  
T1D  133.6  > 500000^{∗}  4993  37  2  107  33 
We found that the evaluated methods varied greatly in the number of interactions detected. ${\chi}^{2}$ reported many interactions that passed Bonferroni correction, totalling many hundreds of thousands of SNP pairs in some datasets. This suggests additional filtering is required. GBOOST was also able to detect a number of SNP pairs with significant association in all datasets, though this is reduced compared with previously reported results and is less than we report using our novel DSS test. We also attempted to run RAPID [32], which is based on a geometric approximation to ${\chi}^{2}$, but despite a lengthy parameter tuning stage, requiring multiple iterations over the WTCCC data, we were unable to detect any significant SNPs in real data. These differences in results with previous reports for GBOOST and RAPID may be caused by varying quality control measures, or parameter settings.
The vast number of positive results that a conventional ${\chi}^{2}$ statistic generates for some datasets appears to be associated with the strength of univariate SNP association seen in the data. We hypothesise that SNPs showing strong univariate association may have a possible confounding effect. If a strongly associated SNP is paired with a SNP showing no association, the resulting pair is likely to have at least the same level of association according to ${\chi}^{2}$ as the strongest of the two. Given the vast number of pairs being examined, it is likely that such "univariatelydriven" pairs overwhelm the results and reduce the ranking of SNP pairs with "genuine" epistatic effects enough that they are impossible to recover using postprocessing techniques.
In Figure 1(b) we plot the number of times that each SNP occurs in the list of top pairs reported by ${\chi}^{2}$. While most SNPs occur in fewer than 10 pairs, the two outliers correspond to the two SNPs with strongest univariate significance indicating they occur in 99% of the 500,000 topranked SNP pairs reported by ${\chi}^{2}$. The majority of these SNP pairs are therefore unlikely to be evidence of epistatic interactions as their perceived association is due to univariate effects only. When used for the detection of epistatic SNP pairs, the ${\chi}^{2}$ statistic tests only for an association with phenotype but, unfortunately, fails to adequately take into account whether this association is due to univariate effects only. In the search for epistatic interactions, such pairs represent a source of noise that can cause practical problems for many standard tests of association.
Novel statistics to account for strong univariate effects
The confounding by strong univariate signals similar to the results of ${\chi}^{2}$ filtering in Figure 1(a) has been seen elsewhere [26, 27, 33], but previously proposed methods of accounting for these effects are either heuristic (difficult to interpret and lacking in statistical rigour), or are based on regression (requiring slower iterative solutions and assumptions about the way in which SNPs interact). Here we present the results of our novel GSS test as an alternative solution for dealing with these effects.
In Table 3, we indicate the number of pairs detected by ${\chi}^{2}$ that are significant according to the GSS test given the conservative Bonferroni threshold of significance. The number of significant pairs falls dramatically for diseases with strong univariate signals, from hundreds of thousands down to tens. These reductions support our hypothesis that most of the SNP pairs detected by the ${\chi}^{2}$ filter show very weak or no epistatic effect.
Interestingly, repeating the same approach over the pairs detected by GBOOST removes very few pairs for most datasets. This is likely because GBOOST looks for significant interactions by examining the improvement of fit in loglinear regression models with and without an interaction term, in essence searching for SNP pairs with no strong univariate effects. The downside of such a technique is that a number of assumptions must be made, in this case requiring that the epistatic SNP pair must fall under an additive model. Such assumptions are not made by the GSS test.
Current implementations of the GSS statistic are too computationally expensive to use on all possible SNP pairs but can easily be run over a few million candidate pairs (our MATLAB implementation requires ≈ 90 minutes for evaluation of 1 million SNP pairs, see Additional File 1 Section 1.6). We therefore take a twostage filtering approach similar to many other methods [6, 10, 29, 32, 34], running a fast but lenient primary filter exhaustively over all pairs, followed by the slower but more accurate GSS test on the smaller subset of pairs selected by the initial filter. As a primary filter, we could use ${\chi}^{2}$, though the proliferation of strong univariate SNPs is often so large that it is not feasible to store all significant pairs within a ranked list. As an alternative primary filter, we introduce the DSS, based on similar concepts to the GSS statistic. The DSS test measures the logpvalue difference between a pair of SNPs and the strongest individual SNP in the pair. This approach is similar to that used in [26, 27], and is well correlated with the GSS test (see Additional File Figures 3 and 4) but is much faster to compute.
To demonstrate the effectiveness of the DSS heuristic followed by the GSS filter, we repeat the same analysis as performed for ${\chi}^{2}$ and GBOOST. Table 3 shows that DSS detects hundreds or thousands of SNP pairs in all datasets and after filtering using the GSS statistic, there are more SNP pairs remaining than for either ${\chi}^{2}$ or GBOOST, in every dataset, respectively.
In Figure 1(c), we plot the significance of SNP pairs chosen by the DSS filter. The figure demonstrates that the DSS heuristic largely addresses the proliferation of SNP pairs caused by strong univariate SNPs, with the chosen SNP pairs no longer showing the similar banding effect seen in the corresponding plot for ${\chi}^{2}$ shown in Figure 1(a). The frequency plot (Figure 1(d)) further demonstrates this, indicating that while some SNPs appear more frequently than others, no single SNP dominates the entire list. The SNP pairs with high DSS show an improved concordance with GSS compared to the concordance seen for ${\chi}^{2}$ in Figure 1(a).
Comparison to previously reported interactions
The WTCCC datasets have been thoroughly examined by a number of epistasis detection methods many of which have reported significantly interacting SNP pairs, including some with evidence of replication of association in other datasets. We have conducted a comparison of these previous results [26, 35–37] with the SNP pairs reported by GWIS using a combination of DSS and GSS filters.
Each study reported in the literature uses its own statistics for determining a pair's significance and while direct comparison between pvalues from these statistics is not meaningful, we can instead evaluate the usefulness of a SNP or SNP pair directly. Namely, we would like to find a pair of SNPs which segregate a significant subset of Cases with no or very few Controls, or conversely a significant fraction of Controls with few Cases.
Odds Ratios (ORs) are commonly used to measure effect size [28] and have the advantage that they can also show whether the effect is protective or contributory. It is well known that the OR can be meaningless if the "odds" are close to zero. For contributory (deleterious) alleles this occurs when the critical parameter sensitivity ≈ 0 while for protective alleles, this is reversed and the odds ratio becomes uninformative when specificity ≈ 0. As we are only interested in either sensitivity or specificity depending on direction of association, we use the term "critical sens/spec" to refer to sensitivity and specificity depending on whether a given pair is contributory or protective. By examining the OR and the critical sens/spec we are able to summarise information on effect size, association direction and the proportion of correctly classified samples.
Although a substantial number of pairs were detected by both GWIS and literature methods, there are some discrepancies. Many pairs detected by GWIS alone often have greater odds ratios and critical sens/spec than pairs detected by the literature only. This suggests that GWIS can detect many potentially interacting pairs that are missed by methods in the literature.
Overlap between SNP pairs found by GWIS using the DSS filter and previous studies after GSS filtering.
Disease  Lit. total  Lit. after GSS  Overlap  GWIS only  GWIS total 

BD  134  57 (43%)  56  123 (69%)  179 
CAD  117  37 (32%)  37  79 (68%)  116 
CD  234  21 (9%)  21  36 (63%)  57 
HT  93  53 (57%)  49  58 (54%)  107 
RA  293  191 (65%)  190  122 (39%)  312 
T1D  801  35 (4%)  35  72 (67%)  107 
T2D  230  59 (26%)  56  78 (58%)  134 
The second category, marked by cyan diamonds, predominantly consist of literature pairs where the level of improvement of the pair over its individual SNPs is insufficient to be deemed epistatic according to our stringent requirement of improvement above Bonferroni threshold, i.e. $\mathsf{\text{fl}}{\mathsf{\text{t}}}_{\mathsf{\text{GSS}}}\ge \mathsf{\text{lo}}{\mathsf{\text{g}}}_{10}\left(\begin{array}{c}\hfill 459,012\hfill \\ \hfill 2\hfill \end{array}\right)\approx 11$. In essence, these pair are deemed to be driven by main effects alone. This category also includes a few literature pairs that had insufficient critical sens/spec to be considered by GWIS (< 2%, see "Minimal sensitivity and specificity" in Methods section). The supplementary Figures 2224 show how this category changes, and the overlap with literature improves, once the Bonferroni threshold requirement is relaxed.
Recall that GWIS is intended to detect potential epistatic interactions. It is very encouraging that although GWIS' epistasis definition does not explicitly maximize oddsratios or critical sens/spec, literature pairs with high oddsratios and critical sens/spec are reliably detected by GWIS.
The analysis across datasets shows the expected trend of lower critical sens/spec having increased log_{2}OR and SNP pairs with low critical sens/spec (≤ 7%) often having very large log_{2}OR. These pairs are often closely located (≤ 1Mb) and a number have been detected by previous studies. Some exceptions do exist to these trends with T1D showing a SNP pair detected only by GWIS that has OR above 4 and critical sens/spec above 30%.
As discussed earlier, GBOOST results are largely significant according to the GSS filter. There were a few points detected by GBOOST but not by GSS filtering. These pairs tended to have relatively small OR and were only just under the strict Bonferroni threshold being used.
We can also use the previous literature to provide evidence that the DSS statistic is acting as a reasonable proxy for the GSS filter. If pairs from previous literature that are significant according to GSS but were not detected by DSS, then the DSS filter has failed to detected some relevant pairs. In Table 4, we show the number of interactions reported by GWIS using the combination of DSS and GSS filters, the number of pairs reported only by previous literature after GSS filter and the number of pairs that appear in both sets of results.
The results indicate that the number of previously reported pairs that remain significant under GSS varies dramatically ranging from ≈ 4% to ≈ 66%. This large variance is likely related to the fact that different methods chose to focus on one or two WTCCC datasets rather than all seven. Datasets that show a large reduction in the number of reported pairs after GSS filtering have tended to be driven by methods searching for strong phenotype association rather than purely epistatic effects.
The "Overlap" column shows that aside from nine pairs in four diseases, all previously reported pairs that are significant under GSS were also detected by GWIS using the combination of DSS and filters. This provides additional evidence that the DSS primary filter is sensitive enough to detect pairs that are likely to be significant under GSS.
We also note that many 'novel' SNP pairs were also detected by GWIS. While reiterating that further quality control and inspection would need to be performed to validate such pairs, it is indicative that exhaustive search combined with the statistics we propose here is likely able to detect a greater quantity of novel epistatic interactions. Such further analysis may also involve readjustment of the cutoff threshold to values below Bonferroni threshold used in Figure 2.
Further validation over simulated data
To further validate our proposed statistic GSS and heuristic DSS, we evaluate their power and false positive rates over a set of synthetic benchmark datasets. The datasets chosen were generated for [23] and simulate 5 models of SNP interaction. The data shows association with phenotype only when the "true" SNPs are considered as a pair, with no association univariately. For each combination of heritability, minor allele frequency and samplesize, 500 datasets were generated, creating a total of 70 penetrance functions and 42,000 datasets. These datasets have been used to evaluate the results of several previous methods [10, 23, 38].
Over all the parameter combinations, DSS provided higher power than ${\chi}^{2}$, albeit with a slightly higher falsepositive rate. This matches our expectations for DSS as a heuristic fast filter for epistasis (i.e. a manageable number of falsepositives are expected). The number of falsepositives from ${\chi}^{2}$ was extremely low (0 or 1 per 500,000 SNPpairs) suggesting that the Bonferronicorrected significance threshold was too strict for the ${\chi}^{2}$ test on this data. With a different threshold, ${\chi}^{2}$ might have recovered some false negative errors.
The number of falsepositives from DSS was also very low, and appeared to grow linearly with increases in the number of samples. The maximum falsepositive rate observed for DSS on any dataset was 0.003 and the average false positive rate over all parameter combinations was 0.001.
Although with the Bonferronicorrected pvalue threshold DSS performed better than ${\chi}^{2}$, these results should be viewed with caution. Both methods could have performed better with a different significance threshold. Many of the DSS falsepositives could have been filtered with a stricter threshold and likewise, many of the ${\chi}^{2}$ falsenegatives could have been detected with a weaker threshold. However, generation of pvalues is intrinsic to the tests being evaluated, and in real datasets the set of true interactions is unknown making it impossible to tune the significance threshold. Our results on the WTCCC datasets show that SNPpair pvalue assignment by the DSS heuristic is of practical use for quickly finding SNPpairs with characteristics suggestive of phenotype association. Although we could have adjusted the pvalue threshold to suit either algorithm, we felt the strict Bonferroni level is the only meaningful threshold that could be applied to real world data and therefore the only threshold that is justifiable on simulated data.
While these figures validate our proposed DSS filter, it is worth noting that the simple scenario of a single epistatic interaction is unlikely to emulate that of real datasets, and as such, the conclusions that can be drawn from current synthetic benchmarks, including that used here, are limited. For instance, the QQ plots in Additional File Figure 2 clearly indicate that in the reallife WTCCC data used in previous sections the DSS filters yield systematically fewer false positives than ${\chi}^{2}$ filters, contrary to the observations for simulated data above. We elaborate on this in the Discussion section.
Discussion
Improved efficiency allows analysis on current and future datasets
In recent years, there have been several proposals that exploit the inherently parallelisable structure of GWAS data to provide reasonably fast solutions capable of processing a WTCCC dataset in several hours. However, SNP arrays currently being used in GWAS studies are an order of magnitude larger [39], resulting in two orders of magnitude increase in the number of pairs and a pressing need for ever more efficient processing of GWAS. Moreover, datasets are often processed repeatedly as data and parameters are altered, quality control measures applied or to correct for population and batch effects, meaning that effective research demands rapid processing. The analysis of higherorder interactions will also dramatically increase the computational burden of epistasis detection. Combined, these points indicate that multivariate GWAS analysis is still a computational challenge.
Our method provides faster discovery of epistatic interactions, which enables more effective, interactive usage. The tool provides an efficient and fast screening capabilities that can be run locally on researchers' desktop computers rather than expensive computing clusters. The reported results can then be refined with more computationally expensive methods such as logistic regression or permutation testing, or in combination with additional biological reference material.
Feasibility of exhaustive search removes the need for adhoc constraints
As indicated by several previous publications [21, 24, 40], there is a need for exhaustive search over all bivariate associations in CaseControl studies. While there are several established heuristics that aim to reduce the number of pairs considered, they all have corresponding weaknesses.
An alternative strategy is the use of known biological data. Here, the number of SNPs examined is reduced to those with prior evidence of possible epistatic effects [42] or that can be mapped to known biological networks [35]. These strategies are likely to be hindered by a lack of epistasis understanding in complex organisms.
Distance constraints, in which SNP pairs are discarded if they are too close together [10, 24, 32], are commonly used with some evidence [26, 43] indicating that such pairs may be linked to genotyping errors. However, it is not always clear that all closely located SNP pairs are due to genotyping errors [26]. Moreover, some recent methods [44] have been designed specifically to look for pairs that were closely located, in order to find associations caused by nontyped SNPs.
The feasibility of exhaustive search as demonstrated in this work removes the need for such constraints. Exhaustive search can examine all possible SNP pairings and, if a robust statistical filter is used, will greatly reduce the set of epistatic interactions requiring followup analysis. Further filtering can then be applied to remove those SNP pairs that are not relevant for a given experiment.
Comments on the definition of epistasis
Our prime goal in this paper is to present a practical system capable of exhaustive search through all SNP pairs in real, full scale GWAS, detecting all pairs evidencing significant epistatic effects. This requires a robust definition of epistasis which can be translated into an actionable mathematical algorithm [11].
 (i)
that there exists a scoring function of genotype calls for the pair of SNPs and a decision threshold such that a substantial subset of subjects scoring above the threshold is significantly enriched (biased) in either Cases or Controls, and the split of the sample according to this threshold results in OR significantly different from 1;
 (ii)
for any scoring function depending on a single SNP of the pair only, such an enrichment is highly unlikely to be achievable by resampling data from the population.
In particular, our definition captures three examples of penetrance tables for "nonstandard" epistatically interacting loci discussed by Cordell [11, Tables 1, 2, 3], and moreover, this can be done with a suitable choice of "purely additive" scoring functions and appropriate decision thresholds (no need for any crossterms). In that respect our generic formal definition of epistasis is closer to its biological counterpart than Fisher's definition of interaction [12], which focuses on fitted models' deviation from additivity. Note that even the original review of Fisher's paper pointed out that his definition does not capture a number of biologically plausible aspects of epistatic interaction, see [11]. However, Fisher's definition is mathematically sound and thus widely used in analysis of contingency tables in statistical literature [28], in quantitative genetics [21] and has been applied in a number of GWAS analysis papers using model based regression approaches [10, 26, 35, 37, 45–47].
Analysis of real datasets may improve simulated data
Despite advances in speed, the most common benchmark for epistasis remains simulated data, where a single epistatic interaction embedded in a small number of SNPs is used to judge a method's power and false positive rate under various parameter settings. In this work, we evaluate the power of a standard ${\chi}^{2}$ and our proposed DSS filter over many such datasets. In conjunction, we also extensively and exhaustively examined multiple real life GWAS, revealing complexities such as confounding signals generated by highly associated univariate SNPs and multiple epistatic signals of varying strength. Such complexities are rarely modelled together in a single epistasis simulation and indicates limitations in the ability of simulated data to be indicative of true power or false positive rates. We believe that further analysis of real data may help better characterise the complexities of GWAS which can be used to create more realistic simulated data. Broader scenarios with multiple epistatic, nonepistatic and univariate signals may better emulate the complexities which we believe are still hidden in real datasets.
Univariate associations can have a confounding effect on standard tests for association
Using the ${\chi}^{2}$ statistic as a filter to detect epistatic SNP pairs, we discovered that topranked SNP pairs were almost always driven by univariately strong SNPs. If a dataset contains a SNP with strong univariate association its pairings with random SNPs will cause the ${\chi}^{2}$ filter to report many thousands of SNP pairs that show an association with phenotype but do not show epistaticlike effects according to our definition.
Studying pairwise associations in GWAS data is necessarily a filtering process, reducing the billions of possible interactions by several (5 or 6) orders of magnitude down to a small number that can be analysed in detail. In order to have any chance of discovering epistatic interactions, the majority of pairs of SNPs that show little improvement over their univariate associations must be explicitly discarded; in other words, we must specifically look for pairs of SNPs that together show improved association with phenotype.
Empirical evidence showing the impact this confounding has on the ${\chi}^{2}$ statistic provided in this paper is intrinsically interesting. Indeed, ${\chi}^{2}$ filtering has been used in bivariate analysis of GWAS in the past using the standard ${\chi}^{2}$ test directly [25, 29, 30] or some variant of it [27, 32, 48, 49]. It is also likely that the same confounding will affect other standard tests for association. Such confounding has been previously observed but has rarely been dealt with in a rigorous manner that is not based on regression. Our GSS/DSS test, explicitly searching for gains in specificity and sensitivity, is a new, efficient alternative in this regard.
Multivariate analysis increases the need for stringent quality control and followup analysis
GWIS is a modelfree method for detecting epistatic SNPs, designed to be sensitive to any associations in the data that separate Cases and Controls. However, this separation may be due to signals other than that caused by phenotype. It has been noted that pairwise SNP analysis may be more susceptible to noise caused by genotyping errors, population structure or batch effects [26, 43] compared to univariate analysis and reported interactions may be a product of these sources of noise. Given that these will vary between experiments followup analysis of reported interactions, especially quality control of genotype calls, remains critical for determining their validity.
Methods
In this section we outline various filtering procedures used in this paper for detection of putative epistasis loci. We shall focus particularly on the receiver operating characteristic (ROC) analysis method, which is part of the novelty of this paper. More details and formal descriptions have been shifted to the Additional File 2 Materials and forthcoming papers will contain the full details and formal proofs.
ROC analysis for GWAS
Here we outline three particular "model free" statistical filtering methods implemented in GWIS and explicitly used in this presentation.
Our filtering approach quantifies the ability of a pair of SNPprobes to segregate Cases from Controls in available data sample compared to the segregation ability of the two SNPprobes taken individually. There are a number of methods in the literature that attempt to measure this type of improvement for epistasis detection, e.g. BOOST uses the decrease in residual error between additive and full interaction regression models [10] while the “random chemistry” approach of Eppstein et. al. [50] uses Euclidean distance between ROC curves.
The key distinct features of our method can be summarised follows:
• It is based on ROC curve analysis, focussing on classification rather than regression;
• The filters use an exact quantification of underlying probability distributions rather than relying on asymptotic normality;
• The approach permits a natural interpretation that links the properties of the sample data back to population data.
With each SNPprobe or pair of SNPprobes, we associate a sample prevalence mapping, allocating to each individual the ratio of the number of Cases to the sum of Cases and Controls in the dataset which carry exactly the same genotype as this individual. For any pair of SNPprobes we have three such prevalence mappings, one for the pair and two for the individual probes. Each mapping can be used to construct a ROC curve: the plot of the true positive rate (TPR) versus the false positive rate (FPR). These are piecewise linear curves. Specifically, a 9piece ROC curve for the pair, ROC (g_{1}, g_{2}), which dominates both 3piece curves ROC(g_{ i } ), i = 1, 2, for the individual SNPprobes. This domination results from the increased number of genotype calls for a pair of SNPs which allows for finer stratification of the data. For most probe pairs, this stratification will have little effect on the ability to separate Cases from Controls but for some the difference will be significant.
This will be referred to as the score or output of the GSSfilter.
while the "max" is over the smallest convex region ${\mathcal{H}}_{0}$ of the unit square I^{2} := [0, 1]^{2} containing ROC(g_{1}) and ROC(g_{2}), see the shaded region in Figure 5; and t_{0}, t_{1} are the total numbers of Controls and Cases in the sample dataset, respectively. In this case π_{0} and π_{1} represent the (unknown) population proportion of deleterious alleles in Controls and Cases respectively. For a given point on the ROC curve (defined by x_{0} and x_{1}), maximizing over the unknown population probabilities corresponds to a worst case scenario for rejection of the null hypothesis $\left({\pi}_{0},{\pi}_{1}\right)\in {\mathcal{H}}_{0}$, with the pvalues quantifying the largest probability of observing a sensitivity greater than x_{1}/t_{1} and a specificity greater than 1  x_{0}/t_{0} by biased sampling. The true pvalue, for the actual (π_{0}, π_{1}) for the population, must obviously be less than this. Minimizing over the pairs of points of ROC(g_{1}, g_{2}) curve gives the set of alleles with the "best" capability to discriminate Cases from Controls.
The optimisation itself is relatively easily computable on modern hardware with carefully crafted algorithms. More details are given in the Supplement and [51].
In this case π_{0} and π_{1} again represent the (unknown) population proportions of deleterious alleles in Controls and Cases, respectively, but since the null hypothesis is in fact restricted to the main diagonal, the optimisation over the population parameters reduces to maximisation over a single variable π_{0}. The interpretation is as before with the "max" part corresponding to an upper limit on the true pvalue and minimisation over the pairs of points corresponding to selection of the smallest such upper limit, thereby giving the most significant improvement of the pair is a classification of individuals into Controls and Cases using the pair's genotype calls over bias sampling from hypothetically inseparable population.
The crucial, "max" part of this statistic can be easily tabulated (as a function of counts (x_{1}, x_{2})), and therefore P_{SS} is relatively easy to implement in practice for exhaustive scanning of probepairs as a primary filter.
We shall call flt_{SS} and flt_{DSS} the filters for SS and DSS, respectively. The fit_{DSS} quantifies an improvement of a pair over its individual constituents allowing it to act as a computationally inexpensive proxy for flt_{GSS} which is suitable for scanning massive numbers of pairs (g_{1}, g_{2}); see Additional File Figures 3 and 4.
Odds ratio
GWAS studies aim in particular at identification of genomic rare variants in the population which are associated with increased or decreased risk of developing a disease. At data filtering stage, the main focus in this paper, we would like to identify SNPpairs and sets of genotyping calls which allow us to identify subsets of the dataset with an OR for developing disease significantly different from 1. There are two possibilities illustrated by example in Figure 4. The first of them, the contributing or high risk scenario, (cntr ≡ OR ≫ 1), is illustrated in subplot (c). Here the red star corresponds to set of genotype calls with the highest flt_{GSS} for contributing gain, which happen to be a singleton set {(1, 0)}. The carriers of this genotype constitute ξ_{1}:= x_{1}/t_{1} = 5.79% of Cases and ξ_{0}:= x_{0}/t_{0} = 0.24% Controls, resulting in extremely high odds ratio OR = 25.73 and significant flt_{GSS} = 33.83. The opposite, protective scenario, (prtv ≡ OR ≪ 1), illustrated in Figure 4(b). Here we find that for the set two genotype calls, {(2, 1), (1, 1)}, we have very low number of Cases carrying these genotypes, ξ_{1} = 0.42% and relatively high fraction of Controls ξ_{0} = 7.39% resulting in OR ≈ 0.05. In the contributing scenario we would like to increase ξ_{1} = SEN and decrease ξ_{0} to ≈ 0; in the protective situation, we would like to maximize ξ_{0} = SPE with simultaneous reduction of ξ_{1} to ≈ 0.
Implementation of tests
The above section introduces principles on which our custom filtering algorithms are built. In this subsection we describe some additional enhancements and heuristics which were added to practical implementations used.
Protective and contributing capabilities
Minimal sensitivity and specificity
Both SS and GSS tests are capable of identification of genotype probes which allow for strong separation in relatively small fractions of the population. This is a desired property for detection of rare variants. However, in practice the limited sample size imposes practical limitations on minimal size which could be of practical interest and is immune to noise or numerical instability of the optimisation procedures used. In our analysis we demanded that in computation of the outer minimum in either (2) or (4) we disregarded all contributions from x_{0}, x_{1} such that min(1  x_{0}/t_{0}, x_{1}/t_{1}) < 0.02.
Limited precision implementation
The solution of this optimisation is not straightforward, for an average GWAS, t_{0} and t_{1} have sizes measured in thousands. This means in practice that in evaluating (2) and (4) we need to deal with multiplications, divisions and summations of thousands of numbers either so small or so large that they cannot be represented directly in computer hardware. For the description of the specific procedures developed to deal with this task and presentation of related formal proofs of their correctness we refer to a dedicated methods paper [51]. Here we only outline the main steps of those derivations:

First, we prove that the functions under "max" in (2) and (4) have no local maxima;

For (2) the maximum is achieved on the boundary of ${\mathcal{H}}_{0}$.

Due to that uniqueness, we can efficiently use any iterative procedure for finding the maximum. In particular we have used the bisection method, which converges to the solution along the boundary.

Finally, for numerical efficiency we have developed specific numerical simplifications which effectively reduce computation of the sums in (2) and (4) down to additions of small numbers of terms of order of one, with provably negligible penalty errors.
With the simplifications outlined above, the computation of values for individual probes and probepairs becomes a tractable numerical task. However when it comes to an exhaustive tabulation of the whole 2dimensional distribution underpinning computation of P_{ SS } for tens of thousands of possible values of counts x_{0} and x_{1}, hence for the multiple millions of pairs (x_{0}, x_{1}), the computing burden could become significant, warranting additional simplifications and reductions. In the case of GSS the computational burden is even harder, direct scan with this statistical filter becomes impractical (see Additional File Section 4), and so arises the need for developing more efficient proxies such as DSS (5).
Other filters used
We have used a number of other techniques than those described above for filtering putative interactions in GWAS. We outline them here for completeness.
${\chi}^{2}$ for independence
This is one of the most popular methods for interactions detection in GWAS. It has two distinct components:

Computation of${\chi}^{2}$statistics. This is a well defined statistic which could be used directly for ranking of hits;

Computation of pvalue for determination of significance. This part is is more complex and the usual solution is to apply a formula which is rigorously derived for sampling from a normal distribution [28].
We have used such formulae with 8 and 2 degrees of freedom when dealing with bivariate or univariate analysis, respectively. Additionally, we have applied the ${\chi}^{2}$ distribution with 4 degrees of freedom to scores derived by the BOOST algorithm, following the original recommendation of the authors of that method (see [7]). In all those cases we have serious reservations regarding allocation of such pvalues (see Discussion for an elaboration of this point).
2 × Vcontingency table.
Genotype Frequencies  

Phenotype  1  2  ...  V  Row Counts 
Controls ≡ 0  n _{01}  n _{02}  ...  n _{0V}  t _{0} 
Cases ≡ 1  n _{11}  n _{12}  ...  n _{1V}  t _{1} 
Col. Counts=  n _{:1}  n _{:2}  ...  n _{:V}  n 
This statistic is known to have approximately ${\chi}^{2}$ distribution with V  1 degrees of freedom [28, 52], which is used to allocate the pvalues. Note, if the null hypothesis H_{0} : N_{ iυ } = E_{ iυ } for all i, υ holds, then X^{2} = 0.
Fisher Exact test
Fisher Exact test is often used for evaluation of 2×2 contingency tables [28] and as such can be applied for allocation of pvalues to observed cumulated count (x_{0}, x_{1}). Such pvalues turn out to be in fact very close numerically to the P_{ss} test, see Additional File Figure 21. For that reason we did not scan data with Fisher Exact test based filters, but the SS filter is a good indicator of its performance.
BOOST
We have used BOOST and GPU version GBOOST algorithms which we have downloaded from the web, and for details we refer to [7, 10]. These algorithms perform exhaustive search though all pairs of probes, but they use different methodology: they use loglinear regression rather than classification and asymptotically justified approximation for allocation of pvalues to derived scores, the 4degree of freedom ${\chi}^{2}$ test (see [31]).
List of abbreviations
 DoO:

Difference of Odds
 DSS:

Difference in Sensitivity and Specificity
 FE:

Fishers Exact
 FPR:

False positive rate
 GPU:

Graphics Processing Unit
 GWAS:

Genome wide association studies
 GWIS:

Genome Wide Interaction Search
 GSS:

Gain in Sensitivity and Specificity
 OR:

Odds Ratio
 ROC:

Receiver operating characteristic
 SNP:

Single nucleotide polymorphism
 SS:

Sensitivity and Specificity
 TPR:

True positive rate
 WTCCC:

Wellcome Trust CaseControl Consortium.
WTCCC Datasets
 BD:

Bipoloar Disorder
 CAD:

Coronary Artery Disease
 CD:

Crohn's Disease
 HT:

Hypertension
 RA:

Rheumatoid Arthritis
 T1D:

Type I Diabetes
 T2D:

Type II Diabetes.
Declarations
Acknowledgements
We would like to thank our colleagues for generous assistance in the preparation of this paper:
Gad Abraham, Leon Gor, Izi Haviv, Eder Kikianty, Andrew Kowalczyk, Geoff Macintye, John Markham and Armita Zarnegar.
This study was supported by National ICT Australia (NICTA). NICTA is funded by the Australian Government's Department of Communications, Information Technology and the Arts, the Australian Research Council through Backing Australia's Ability, and the ICT Centre of Excellence programs.
This study has also used resources provided by the Multimodal Australian Sciences Imaging and Visualisation Environment (MASSIVE) through the National Computational Merit Allocation Scheme supported by the Australian Government.
Sections of the data used here were generated by the Wellcome Trust CaseControl Consortium. A full list of the investigators who contributed to the generation of the data is available from http://www.wtccc.org.uk. Funding for the project was provided by the Wellcome Trust under award 076113 and 085475.
AK, MI and BG were partially supported by NHMRC grant 1033452.
This article has been published as part of BMC Genomics Volume 14 Supplement 3, 2013: SNPSIG 2012: Identification and annotation of SNPs in the context of structure, function, and disease. The full contents of the supplement are available online at http://www.biomedcentral.com/bmcgenomics/supplements/14/S3
Declarations
The publication costs for this article were funded by the above grants.
Authors’ Affiliations
References
 Makowsky R, Pajewski NM, Klimentidis YC, Vazquez AI, Duarte CW, Allison DB, de los Campos G: Beyond missing heritability: Prediction of complex traits. PLoS Genet. 2011, 7 (4):
 Zuk O, Hechter E, Sunyaev SR, Lander ES: The mystery of missing heritability: Genetic interactions create phantom heritability. Proc Natl Acad Sci USA. 2012, 109 (4): 11931198. 10.1073/pnas.1119675109.PubMed CentralView ArticlePubMedGoogle Scholar
 Culverhouse R, Suarez BK, Lin J, Reich T: A perspective on epistasis: limits of models displaying no main effect. Am J Hum Genet. 2002, 70 (2): 461471. 10.1086/338759.PubMed CentralView ArticlePubMedGoogle Scholar
 Greene CS, SinnottArmstrong NA, Himmelstein DS, Park PJ, Moore JH, Harris BT: Multifactor dimensionality reduction for graphics processing units enables genomewide testing of epistasis in sporadic ALS. Bioinformatics. 2010, 26 (5): 694695. 10.1093/bioinformatics/btq009.PubMed CentralView ArticlePubMedGoogle Scholar
 Ritchie MD, Hahn LW, Roodi N, Bailey LR, Dupont WD, Parl FF, Moore JH: Multifactordimensionality reduction reveals highorder interactions among estrogenmetabolism genes in sporadic breast cancer. Am J Hum Genet. 2001, 69: 138147. 10.1086/321276.PubMed CentralView ArticlePubMedGoogle Scholar
 KamThong T, Pütz B, Karbalai N, MüllerMyhsok B, Borgwardt K: Epistasis detection on quantitative phenotypes by exhaustive enumeration using GPUs. Bioinformatics. 2011, 27 (13): i214i221. 10.1093/bioinformatics/btr218.PubMed CentralView ArticlePubMedGoogle Scholar
 Yung LS, Yang C, Wan X, Yu W: GBOOST: a GPUbased tool for detecting genegene interactions in genomewide case control studies. Bioinformatics. 2011, 27 (9): 13091310. 10.1093/bioinformatics/btr114.PubMed CentralView ArticlePubMedGoogle Scholar
 Hemani G, Theocharidis A, Wei W, Haley C: EpiGPU: exhaustive pairwise epistasis scans parallelized on consumer level graphics cards. Bioinformatics. 2011, 27 (11): 14621465. 10.1093/bioinformatics/btr172.View ArticlePubMedGoogle Scholar
 Hu X, Liu Q, Zhang Z, Li Z, Wang S, He L, Shi Y: SHEsisEpi, a GPUenhanced genomewide SNPSNP interaction scanning algorithm, efficiently reveals the risk genetic epistasis in bipolar disorder. Cell Res. 2010, 20 (7): 854857. 10.1038/cr.2010.68.View ArticlePubMedGoogle Scholar
 Wan X, Yang C, Yang Q, Xue H, Tang NLS, Yu W: Detecting twolocus associations allowing for interactions in genomewide association studies. Bioinformatics. 2010, 26 (20): 25172525. 10.1093/bioinformatics/btq486.View ArticlePubMedGoogle Scholar
 Cordell HJ: Epistasis: what it means, what it doesn't mean, and statistical methods to detect it in humans. Hum Mol Genet. 2002, 11 (20): 24632468. 10.1093/hmg/11.20.2463.View ArticlePubMedGoogle Scholar
 Fisher RA: The correlation between relatives on the supposition of Mendelian inheritance. Trans R Soc Edin. 1918, 52: 399433.View ArticleGoogle Scholar
 Krzanowski WJ, Hand DJ: ROC Curves for Continuous Data, Volume 111 of CRC Monographs on Statistics and Applied Probability. 2009, Chapman & Hall/CRCView ArticleGoogle Scholar
 Hahn LW, Ritchie MD, Moore JH: Multifactor dimensionality reduction software for detecting genegene and geneenvironment interactions. Bioinformatics. 2003, 19 (3): 376382. 10.1093/bioinformatics/btf869.View ArticlePubMedGoogle Scholar
 Moore JH, Gilbert JC, Tsai CT, Chiang FT, Holden T, Barney N, White BC: A exible computational framework for detecting, characterizing, and interpreting statistical patterns of epistasis in genetic studies of human disease susceptibility. J Theor Biol. 2006, 241 (2): 252261. 10.1016/j.jtbi.2005.11.036.View ArticlePubMedGoogle Scholar
 Moore JH, Barney N, Tsai CT, Chiang FT, Gui J, White BC: Symbolic modeling of epistasis. Hum Hered. 2007, 63 (2): 120133. 10.1159/000099184.View ArticlePubMedGoogle Scholar
 Chen L, Yu G, Langefeld CD, Miller DJ, Guy RT, Raghuram J, Yuan X, Herrington DM, Wang Y: Comparative analysis of methods for detecting interacting loci. BMC Genomics. 2011, 12: 34410.1186/1471216412344.PubMed CentralView ArticlePubMedGoogle Scholar
 Wang Y, Liu G, Feng M, Wong L: An empirical comparison of several recent epistatic interaction detection methods. Bioinformatics. 2011, 27 (21): 29362943. 10.1093/bioinformatics/btr512.View ArticlePubMedGoogle Scholar
 Shang J, Zhang J, Sun Y, Liu D, Ye D, Yin Y: Performance analysis of novel methods for detecting epistasis. BMC Bioinformatics. 2011, 12: 47510.1186/1471210512475.PubMed CentralView ArticlePubMedGoogle Scholar
 Ritchie MD: Using biological knowledge to uncover the mystery in the search for epistasis in genomewide association studies. Ann Human Genet. 2011, 75: 172182. 10.1111/j.14691809.2010.00630.x.View ArticleGoogle Scholar
 Cordell HJ: Detecting genegene interactions that underlie human diseases. Nat Rev Genet. 2009, 10 (6): 392404.PubMed CentralView ArticlePubMedGoogle Scholar
 The Wellcome Trust CaseControl Consortium: Genomewide association study of 14,000 cases of seven common diseases and 3,000 shared controls. Nature. 2007, 447 (7145): 661678. 10.1038/nature05911.PubMed CentralView ArticleGoogle Scholar
 Velez DR, White B, Motsinger A, Bush WS, Ritchie MD, Williams SM, Moore JH: A Balanced Accuracy Function for Epistasis Modeling in Imbalanced Datasets using Multifactor Dimensionality Reduction. Genet Epidemiol. 2007, 315 (4): 306315.View ArticleGoogle Scholar
 Prabhu S, Pe'er I: Ultrafast genomewide scan for SNPSNP interactions in common complex disease. Genome Res. 2012, 22: 22302240. 10.1101/gr.137885.112.PubMed CentralView ArticlePubMedGoogle Scholar
 Wang Z, Wang Y, Tan KLKLL, Wong L, Agrawal D: eCEO: an efficient Cloud Epistasis cOmputing model in genomewide association study. Bioinformatics. 2011, 27 (8): 10451051. 10.1093/bioinformatics/btr091.View ArticlePubMedGoogle Scholar
 Liu Y, Xu H, Chen S, Chen X, Zhang Z, Zhu Z, Qin X, Hu L, Zhu J, Zhao GP, Kong X: GenomeWide InteractionBased Association Analysis Identified Multiple New Susceptibility Loci for Common Diseases. PLoS Genet. 2011, 7 (3): e100133810.1371/journal.pgen.1001338.PubMed CentralView ArticlePubMedGoogle Scholar
 Fang G, Haznadar M, Wang W, Yu H, Steinbach M, Church TR, Oetting WS, Van Ness B, Kumar V: Highorder SNP combinations associated with complex diseases: efficient discovery, statistical power and functional interactions. PloS ONE. 2012, 7 (4): e3353110.1371/journal.pone.0033531.PubMed CentralView ArticlePubMedGoogle Scholar
 Agresti A: Categorical Data Analysis. 2002, WileyView ArticleGoogle Scholar
 Zhang X, Zou F, Wang W: FastChi: an efficient algorithm for analyzing genegene interactions. Pacific Symposium on Biocomputing. 2009, 14: 528539.Google Scholar
 Zhang X, Huang S, Zou F, Wang W: TEAM: efficient twolocus epistasis tests in human genomewide association study. Bioinformatics. 2010, 26 (12): i217i227. 10.1093/bioinformatics/btq186.PubMed CentralView ArticlePubMedGoogle Scholar
 Wan X, Yang C, Yang Q, Xue H, Fan X, Tang NLS, Yu W: BOOST: A fast approach to detecting genegene interactions in genomewide casecontrol studies. Am J Hum Genet. 2010, 87 (3): 14.View ArticleGoogle Scholar
 Brinza D, Schultz M, Tesler G, Bafna V: RAPID detection of genegene interactions in genomewide association studies. Bioinformatics. 2010, 26 (22): 28562862. 10.1093/bioinformatics/btq529.PubMed CentralView ArticlePubMedGoogle Scholar
 Bell JT, Timpson NJ, Rayner NW, Zeggini E, Frayling TM, Hattersley AT, Morris AP, McCarthy MI: Genomewide association scan allowing for epistasis in type 2 diabetes. Ann Human Genet. 2011, 75: 1019. 10.1111/j.14691809.2010.00629.x.View ArticleGoogle Scholar
 Zhang X, Zou F, Wang W: Fastanova: an efficient algorithm for genomewide association study. Proceedings of the 14th ACM SIGKDD international conference on Knowledge discovery and data mining. 2008, KDD '08, New York, NY, USA: ACM, 821829.View ArticleGoogle Scholar
 Emily M, Mailund T, Hein J, Schauser L, Schierup MH: Using biological networks to search for interacting loci in genomewide association studies. Eur J Hum Genet. 2009, 17 (10): 12311240. 10.1038/ejhg.2009.15.PubMed CentralView ArticlePubMedGoogle Scholar
 Wan X, Yang C, Yang Q, Xue H, Tang NLS, Yu W: Predictive rule inference for epistatic interaction detection in genomewide association studies. Bioinformatics. 2010, 26: 3037. 10.1093/bioinformatics/btp622.View ArticlePubMedGoogle Scholar
 Gyenesei A, Moody J, Semple CAM, Haley CS, Wei WH: Highthroughput analysis of epistasis in genomewide association studies with BiForce. Bioinformatics. 2012, 28 (15): 19571964. 10.1093/bioinformatics/bts304.PubMed CentralView ArticlePubMedGoogle Scholar
 Jiang X, Barmada MM, Cooper GF, Becich MJ: A bayesian method for evaluating and discovering disease loci associations. PloS ONE. 2011, 6 (8): e2207510.1371/journal.pone.0022075.PubMed CentralView ArticlePubMedGoogle Scholar
 Illumina Inc: GenomeWide DNA Analysis BeadChips. 2011, Data sheet, Illumni Inc, [http://www.illumina.com/documents/products/datasheets/datasheet_omni_wholegenome_arrays.pdf]Google Scholar
 Marchini J, Donnelly P, Cardon LR: Genomewide strategies for detecting multiple loci that influence complex diseases. Nat Genet. 2005, 37 (4): 413417. 10.1038/ng1537.View ArticlePubMedGoogle Scholar
 Jiang R, Tang W, Wu X, Fu W: A random forest approach to the detection of epistatic interactions in casecontrol studies. BMC Bioinformatics. 2009, 10 (Suppl 1): S6510.1186/1471210510S1S65.PubMed CentralView ArticlePubMedGoogle Scholar
 Bochdanovits Z, Sondervan D, Perillous S, van Beijsterveldt T, Boomsma D, Heutink P: Genomewide prediction of functional genegene interactions inferred from patterns of genetic differentiation in mice and men. PloS ONE. 2008, 3 (2): e159310.1371/journal.pone.0001593.PubMed CentralView ArticlePubMedGoogle Scholar
 Lee SH, Nyholt DR, Macgregor S, Henders AK, Zondervan KT, Montgomery GW, Visscher PM: A simple and fast twolocus quality control test to detect false positives due to batch effects in genomewide association studies. Genet Epidemiol. 2010, 34 (8): 854862. 10.1002/gepi.20541.PubMed CentralView ArticlePubMedGoogle Scholar
 Slavin TP, Feng T, Schnell A, Zhu X, Elston RC: Twomarker association tests yield new disease associations for coronary artery disease and hypertension. Hum Genet. 2011, 130 (6): 725733. 10.1007/s0043901110096.PubMed CentralView ArticlePubMedGoogle Scholar
 Marchini J, Donnelly P, Cardon LR: Genomewide strategies for detecting multiple loci that influence complex diseases. Nat Genet. 2005, 37 (4): 413417. 10.1038/ng1537.View ArticlePubMedGoogle Scholar
 Evans DM, Marchini J, Morris AP, Cardon LR: Twostage twolocus models in genomewide association. PLoS Genet. 2006, 2 (9): e15710.1371/journal.pgen.0020157.PubMed CentralView ArticlePubMedGoogle Scholar
 Schüpbach T, Xenarios I, Bergmann S, Kapur K: FastEpistasis: a high performance computing solution for quantitative trait epistasis. Bioinformatics. 2010, 26 (11): 14681469. 10.1093/bioinformatics/btq147.PubMed CentralView ArticlePubMedGoogle Scholar
 Ackermann M, Beyer A: Systematic detection of epistatic interactions based on allele pair frequencies. PLoS Genet. 2012, 8 (2): e100246310.1371/journal.pgen.1002463.PubMed CentralView ArticlePubMedGoogle Scholar
 Gayán J, GonzálezPérez A, Bermudo F, Sáez ME, Royo JL, Quintas A, Galan JJ, Morón FJ, RamirezLorca R, Real LM, Ruiz A: A method for detecting epistasis in genomewide studies using casecontrol multilocus association analysis. BMC Genomics. 2008, 9: 36010.1186/147121649360.PubMed CentralView ArticlePubMedGoogle Scholar
 Eppstein MJ, Payne JL, White BC, Moore JH: Genomic mining for complex disease traits with\random chemistry". Genet Program Evolvable Mach. 2007, 8 (4): 395411. 10.1007/s1071000790395.View ArticleGoogle Scholar
 Kowalczyk A, Shi F, Kikianty E: Accuracy test for genome wide selection of biomarkers. 2011, [http://videolectures.net/nipsworkshops2011_sierranevada]Google Scholar
 Hogg RV, Tanis EA: Probability and Statistical Inference. 2010, Prentice Hall, 2009, 7Google Scholar
 Purcell S, Neale B, ToddBrown K, Thomas L, Ferreira M, Bender D, Maller J, Sklar P, De Bakker P, Daly M et al: PLINK: a tool set for wholegenome association and populationbased linkage analyses. The American Journal of Human Genetics. 2007, 81 (3): 559575. 10.1086/519795.View ArticlePubMedGoogle Scholar
Copyright
This article is published under license to BioMed Central Ltd. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.