Data mining and machine learning approaches for the integration of genome-wide association and methylation data: methodology and main conclusions from GAW20

Background

GAW20 working group 5 brought together researchers who contributed 7 papers with the aim of evaluating methods to detect genetic by epigenetic interactions. GAW20 distributed real data from the Genetics of Lipid Lowering Drugs and Diet Network (GOLDN) study, including single-nucleotide polymorphism (SNP) markers, methylation (cytosine-phosphate-guanine [CpG]) markers, and phenotype information on up to 995 individuals. In addition, a simulated data set based on the real data was provided.

Results

The 7 contributed papers analyzed these data sets with a number of different statistical methods, including generalized linear mixed models, mediation analysis, machine learning, W-test, and sparsity-inducing regularized regression. These methods generally appeared to perform well. Several papers confirmed a number of causative SNPs in either the large number of simulation sets or the real data on chromosome 11. Findings were also reported for different SNPs, CpG sites, and SNP–CpG site interaction pairs.

Conclusions

In the simulation (200 replications), power appeared generally good for large interaction effects, but smaller effects will require larger studies or consortium collaboration for realizing a sufficient power.

     

Indirect effect inference and application to GAW20 data

Background

Association studies using a single type of omics data have been successful in identifying disease-associated genetic markers, but the underlying mechanisms are unaddressed. To provide a possible explanation of how these genetic factors affect the disease phenotype, integration of multiple omics data is needed.

Results

We propose a novel method, LIPID (likelihood inference proposal for indirect estimation), that uses both single nucleotide polymorphism (SNP) and DNA methylation data jointly to analyze the association between a trait and SNPs. The total effect of SNPs is decomposed into direct and indirect effects, where the indirect effects are the focus of our investigation. Simulation studies show that LIPID performs better in various scenarios than existing methods. Application to the GAW20 data also leads to encouraging results, as the genes identified appear to be biologically relevant to the phenotype studied.

Conclusions

The proposed LIPID method is shown to be meritorious in extensive simulations and in real-data analyses.

     

Epigenetics,heritability and longitudinal analysis

Background

Longitudinal data and repeated measurements in epigenome-wide association studies (EWAS) provide a rich resource for understanding epigenetics. We summarize 7 analytical approaches to the GAW20 data sets that addressed challenges and potential applications of phenotypic and epigenetic data. All contributions used the GAW20 real data set and employed either linear mixed effect (LME) models or marginal models through generalized estimating equations (GEE). These contributions were subdivided into 3 categories: (a) quality control (QC) methods for DNA methylation data; (b) heritability estimates pretreatment and posttreatment with fenofibrate; and (c) impact of drug response pretreatment and posttreatment with fenofibrate on DNA methylation and blood lipids.

Results

Two contributions addressed QC and identified large statistical differences with pretreatment and posttreatment DNA methylation, possibly a result of batch effects. Two contributions compared epigenome-wide heritability estimates pretreatment and posttreatment, with one employing a Bayesian LME and the other using a variance-component LME. Density curves comparing these studies indicated these heritability estimates were similar. Another contribution used a variance-component LME to depict the proportion of heritability resulting from a genetic and shared environment. By including environmental exposures as random effects, the authors found heritability estimates became more stable but not significantly different. Two contributions investigated treatment response. One estimated drug-associated methylation effects on triglyceride levels as the response, and identified 11 significant cytosine-phosphate-guanine (CpG) sites with or without adjusting for high-density lipoprotein. The second contribution performed weighted gene coexpression network analysis and identified 6 significant modules of at least 30 CpG sites, including 3 modules with topological differences pretreatment and posttreatment.

Conclusions

Four conclusions from this GAW20 working group are: (a) QC measures are an important consideration for EWAS studies that are investigating multiple time points or repeated measurements; (b) application of heritability estimates between time points for individual CpG sites is a useful QC measure for DNA methylation studies; (c) drug intervention demonstrated strong epigenome-wide DNA methylation patterns across the 2 time points; and (d) new statistical methods are required to account for the environmental contributions of DNA methylation across time. These contributions demonstrate numerous opportunities exist for the analysis of longitudinal data in future epigenetic studies.

     

Genome scan linkage analysis comparing microsatellites and single-nucleotide polymorphisms markers for two measures of alcoholism in chromosomes 1, 4, and 7

Background

We analyzed 143 pedigrees (364 nuclear families) in the Collaborative Study on the Genetics of Alcoholism (COGA) data provided to the participants in the Genetic Analysis Workshop 14 (GAW14) with the goal of comparing results obtained from genome linkage analysis using microsatellite and with results obtained using SNP markers for two measures of alcoholism (maximum number of drinks -MAXDRINK and an electrophysiological measure from EEG -TTTH1). First, we constructed haplotype blocks by using the entire set of single-nucleotide polymorphisms (SNP) in chromosomes 1, 4, and 7. These chromosomes have shown linkage signals for MAXDRINK or EEG-TTTH1 in previous reports. Second, we randomly selected one, two, three, four, and five SNPs from each block (referred to as Rep1 – Rep5, respectively) to conduct linkage analysis using variance component approach. Finally, results of all SNP analyses were compared with those obtained using microsatellite markers.

Results

The LOD scores obtained from SNPs were slightly higher but the curves were not radically different from those obtained from microsatellite analyses. The peaks of linkage regions from SNP sets were slightly shifted to the left when compared to those from microsatellite markers. The reduced sets of SNPs provide signals in the same linkage regions but with a smaller LOD score suggesting a significant impact of the decrease in information content on linkage results. The widths of 1 LOD support interval of linkage regions from SNP sets were smaller when compared to those of microsatellite markers. However, two linkage regions obtained from the microsatellite linkage analysis on chromosome 7 for LOG of TTTH1 were not detected in the SNP based analyses.

Conclusion

The linkage results from SNPs showed narrower linkage regions and slightly higher LOD scores when compared to those of microsatellite markers. The different builds of the genetic maps used in microsatellite and SNPs markers or/and errors in genotyping may account for the microsatellite linkage signals on chromosome 7 that were not identified using SNPs. Also, unresolved map issues between SNPs and microsatellite markers may be partly responsible for the shifted linkage peaks when comparing the two types of markers.

     

Investigation of parent-of-origin effects induced by fenofibrate treatment on triglycerides levels

Background

Genome-wide association studies performed on triglycerides (TGs) have not accounted for epigenetic mechanisms that may partially explain trait heritability.

Results

Parent-of-origin (POO) effect association analyses using an agnostic approach or a candidate approach were performed for pretreatment TG levels, posttreatment TG levels, and pre- and posttreatment TG-level differences in the real GAW20 family data set. We detected 22 genetic variants with suggestive POO effects with at least 1 phenotype (P ≤ 10^− 5). We evaluated the association of these 22 significant genetic variants showing POO effects with close DNA methylation probes associated with TGs. A total of 18 DNA methylation probes located in the vicinity of the 22 SNPs were associated with at least 1 phenotype and 6 SNP-probe pairs were associated with DNA methylation probes at the nominal level of P < 0.05, among which 1 pair presented evidence of POO effect. Our analyses identified a paternal effect of SNP rs301621 on the difference between pre- and posttreatment TG levels (P = 1.2 × 10^− 5). This same SNP showed evidence for a maternal effect on methylation levels of a nearby probe (cg10206250; P = 0.01). Using a causal inference test we established that the observed POO effect of rs301621 was not mediated by DNA methylation at cg10206250.

Conclusions

We performed POO effect association analyses of SNPs with TGs, as well as association analyses of SNPs with DNA methylation probes. These analyses, which were followed by a causal inference test, established that the paternal effect at the SNP rs301621 is induced by treatment and is not mediated by methylation level at cg10206250.

     

Association analyses of repeated measures on triglyceride and high-density lipoprotein levels: insights from GAW20

Background

The GAW20 group formed on the theme of methods for association analyses of repeated measures comprised 4sets of investigators. The provided “real” data set included genotypes obtained from a human whole-genome association study based on longitudinal measurements of triglycerides (TGs) and high-density lipoprotein in addition to methylation levels before and after administration of fenofibrate. The simulated data set contained 200 replications of methylation levels and posttreatment TGs, mimicking the real data set.

Results

The different investigators in the group focused on the statistical challenges unique to family-based association analyses of phenotypes measured longitudinally and applied a wide spectrum of statistical methods such as linear mixed models, generalized estimating equations, and quasi-likelihood–based regression models. This article discusses the varying strategies explored by the group’s investigators with the common goal of improving the power to detect association with repeated measures of a phenotype.

Conclusions

Although it is difficult to identify a common message emanating from the different contributions because of the diversity in the issues addressed, the unifying theme of the contributions lie in the search for novel analytic strategies to circumvent the limitations of existing methodologies to detect genetic association.

     

A double classification tree search algorithm for index SNP selection

Background  

In population-based studies, it is generally recognized that single nucleotide polymorphism (SNP) markers are not independent. Rather, they are carried by haplotypes, groups of SNPs that tend to be coinherited. It is thus possible to choose a much smaller number of SNPs to use as indices for identifying haplotypes or haplotype blocks in genetic association studies. We refer to these characteristic SNPs as index SNPs. In order to reduce costs and work, a minimum number of index SNPs that can distinguish all SNP and haplotype patterns should be chosen. Unfortunately, this is an NP-complete problem, requiring brute force algorithms that are not feasible for large data sets.     

A Modified Entropy-Based Approach for Identifying Gene-Gene Interactions in Case-Control Study

Gene-gene interactions may play an important role in the genetics of a complex disease. Detection and characterization of gene-gene interactions is a challenging issue that has stimulated the development of various statistical methods to address it. In this study, we introduce a method to measure gene interactions using entropy-based statistics from a contingency table of trait and genotype combinations. We also developed an exploration procedure by using graphs. We propose a standardized relative information gain (RIG) measure to evaluate the interactions between single nucleotide polymorphism (SNP) combinations. To identify the k ^th order interactions, contingency tables of trait and genotype combinations of k SNPs are constructed, with which RIGs are calculated. The RIGs are standardized using the mean and standard deviation from the permuted datasets. SNP combinations yielding high standardized RIG are chosen for gene-gene interactions. Detection of high-order interactions and comparison of interaction strengths between different orders are made possible by using standardized RIG. We have applied the proposed standardized entropy-based method to two types of data sets from a simulation study and a real genetic association study. We have compared our method and the multifactor dimensionality reduction (MDR) method through power analysis of eight different genetic models with varying penetrance rates, number of SNPs, and sample sizes. Our method shows successful identification of genetic associations and gene-gene interactions both in simulation and real genetic data. Simulation results suggest that the proposed entropy-based method is better able to detect high-order interactions and is superior to the MDR method in most cases. The proposed method is well suited for detecting interactions without main effects as well as for models including main effects.     

Assessing statistical power of SNPs for population structure and conservation studies

Single nucleotide polymorphisms (SNPs) have been proposed by some as the new frontier for population studies, and several papers have presented theoretical and empirical evidence reporting the advantages and limitations of SNPs. As a practical matter, however, it remains unclear how many SNP markers will be required or what the optimal characteristics of those markers should be in order to obtain sufficient statistical power to detect different levels of population differentiation. We use a hypothetical case to illustrate the process of designing a population genetics project, and present results from simulations that address several issues for maximizing statistical power to detect differentiation while minimizing the amount of effort in developing SNPs. Results indicate that (i) while ~30 SNPs should be sufficient to detect moderate (F_ST = 0.01) levels of differentiation, studies aimed at detecting demographic independence (e.g. F_ST < 0.005) may require 80 or more SNPs and large sample sizes; (ii) different SNP allele frequencies have little affect on power, and thus, selection of SNPs can be relatively unbiased; (iii) increasing the sample size has a strong effect on power, so that the number of loci can be minimized when sample number is known, and increasing sample size is almost always beneficial; and (iv) power is increased by including multiple SNPs within loci and inferring haplotypes, rather than trying to use only unlinked SNPs. This also has the practical benefit of reducing the SNP ascertainment effort, and may influence the decision of whether to seek SNPs in coding or noncoding regions.     

Powerful testing via hierarchical linkage disequilibrium in haplotype association studies

Marginal tests based on individual SNPs are routinely used in genetic association studies. Studies have shown that haplotype‐based methods may provide more power in disease mapping than methods based on single markers when, for example, multiple disease‐susceptibility variants occur within the same gene. A limitation of haplotype‐based methods is that the number of parameters increases exponentially with the number of SNPs, inducing a commensurate increase in the degrees of freedom and weakening the power to detect associations. To address this limitation, we introduce a hierarchical linkage disequilibrium model for disease mapping, based on a reparametrization of the multinomial haplotype distribution, where every parameter corresponds to the cumulant of each possible subset of a set of loci. This hierarchy present in the parameters enables us to employ flexible testing strategies over a range of parameter sets: from standard single SNP analyses through the full haplotype distribution tests, reducing degrees of freedom and increasing the power to detect associations. We show via extensive simulations that our approach maintains the type I error at nominal level and has increased power under many realistic scenarios, as compared to single SNP and standard haplotype‐based studies. To evaluate the performance of our proposed methodology in real data, we analyze genome‐wide data from the Wellcome Trust Case‐Control Consortium.     

The impact of missing and erroneous genotypes on tagging SNP selection and power of subsequent association tests

OBJECTIVE: Single nucleotide polymorphisms (SNPs) serve as effective markers for localizing disease susceptibility genes, but current genotyping technologies are inadequate for genotyping all available SNP markers in a typical linkage/association study. Much attention has recently been paid to methods for selecting the minimal informative subset of SNPs in identifying haplotypes, but there has been little investigation of the effect of missing or erroneous genotypes on the performance of these SNP selection algorithms and subsequent association tests using the selected tagging SNPs. The purpose of this study is to explore the effect of missing genotype or genotyping error on tagging SNP selection and subsequent single marker and haplotype association tests using the selected tagging SNPs. METHODS: Through two sets of simulations, we evaluated the performance of three tagging SNP selection programs in the presence of missing or erroneous genotypes: Clayton's diversity based program htstep, Carlson's linkage disequilibrium (LD) based program ldSelect, and Stram's coefficient of determination based program tagsnp.exe. RESULTS: When randomly selected known loci were relabeled as 'missing', we found that the average number of tagging SNPs selected by all three algorithms changed very little and the power of subsequent single marker and haplotype association tests using the selected tagging SNPs remained close to the power of these tests in the absence of missing genotype. When random genotyping errors were introduced, we found that the average number of tagging SNPs selected by all three algorithms increased. In data sets simulated according to the haplotype frequecies in the CYP19 region, Stram's program had larger increase than Carlson's and Clayton's programs. In data sets simulated under the coalescent model, Carlson's program had the largest increase and Clayton's program had the smallest increase. In both sets of simulations, with the presence of genotyping errors, the power of the haplotype tests from all three programs decreased quickly, but there was not much reduction in power of the single marker tests. CONCLUSIONS: Missing genotypes do not seem to have much impact on tagging SNP selection and subsequent single marker and haplotype association tests. In contrast, genotyping errors could have severe impact on tagging SNP selection and haplotype tests, but not on single marker tests.     

Comparison of SNPs and microsatellites for fine-scale application of genetic stock identification of Chinook salmon in the Columbia River Basin

Genetic stock identification (GSI) is an important tool in fisheries management. Microsatellites (μSATs) have been the dominant genetic marker for GSI; however, increasing availability and numerous advantages of single-nucleotide polymorphism (SNP) markers make them an appealing alternative. We tested performance of 13 μSAT vs. 92 SNP loci in a fine-scale application of GSI, using a new baseline for Chinook salmon consisting of 49 collections (n = 4014) distributed across the Columbia River Basin. In GSI, baseline genotypes for both marker sets were used independently to analyse a real fishery mixture (n = 2731) representing the total run of Chinook salmon passing Bonneville Dam in the Columbia River. Marker sets were evaluated using three criteria: (i) ability to differentiate reporting groups, (ii) proportion of correct assignment in mixture simulation tests and baseline leave-one-out analyses and (iii) individual assignment and confidence intervals around estimated stock proportions of a real fishery mixture. The μSATs outperformed the SNPs in resolving fine-scale relationships, but all 105 markers combined provided greatest power for GSI. SNPs were ranked by relative information content based on both an iterative procedure that optimized correct assignment to the baseline and ranking by minor allele frequency. For both methods, we identified a subset of the top 50 ranked loci, which were similar in assignment accuracy, and both reached maximum available power of the total 92 SNP loci (correct assignment = 73%). Our estimates indicate that between 100 and 200 highly informative SNP loci are required to meet management standards (correct assignment > 90%) for resolving stocks in finer-scale GSI applications.     

Effects of ascertainment bias and marker number on estimations of barley diversity from high-throughput SNP genotype data

The capability of molecular markers to provide information of genetic structure is influenced by their number and the way they are chosen. This study evaluates the effects of single nucleotide polymorphism (SNP) number and selection strategy on estimates of germplasm diversity and population structure for different types of barley germplasm, namely cultivar and landrace. One hundred and sixty-nine barley landraces from Syria and Jordan and 171 European barley cultivars were genotyped with 1536 SNPs. Different subsets of 384 and 96 SNPs were selected from the 1536 set, based on their ability to detect diversity in landraces or cultivated barley in addition to corresponding randomly chosen subsets. All SNP sets except the landrace-optimised subsets underestimated the diversity present in the landrace germplasm, and all subsets of SNP gave similar estimates for cultivar germplasm. All marker subsets gave qualitatively similar estimates of the population structure in both germplasm sets, but the 96 SNP sets showed much lower data resolution values than the larger SNP sets. From these data we deduce that pre-selecting markers for their diversity in a germplasm set is very worthwhile in terms of the quality of data obtained. Second, we suggest that a properly chosen 384 SNP subset gives a good combination of power and economy for germplasm characterization, whereas the rather modest gain from using 1536 SNPs does not justify the increased cost and 96 markers give unacceptably low performance. Lastly, we propose a specific 384 SNP subset as a standard genotyping tool for middle-eastern landrace barley.     

Identification of favorable SNP alleles and candidate genes responsible for inflorescence-related traits via GWAS in chrysanthemum

Key message

81 SNPs were identified for three inflorescence-related traits, in which 15 were highly favorable. Two dCAPS markers were developed for future MAS breeding, and six candidate genes were predicted.

Abstract

Chrysanthemum is a leading ornamental species worldwide and demonstrates a wealth of morphological variation. Knowledge about the genetic basis of its phenotypic variation for key horticultural traits can contribute to its effective management and genetic improvement. In this study, we conducted a genome-wide association study (GWAS) based on two years of phenotype data and a set of 92,617 single nucleotide polymorphisms (SNPs) using a panel of 107 diverse cut chrysanthemums to dissect the genetic control of three inflorescence-related traits. A total of 81 SNPs were significantly associated with the three inflorescence-related traits (capitulum diameter, number of ray florets and flowering time) in at least one environment, with an individual allele explaining 22.72–38.67% of the phenotypic variation. Fifteen highly favorable alleles were identified for the three target traits by computing the phenotypic effect values for the stable associations detected in 2 year-long trials at each locus. Dosage pyramiding effects of the highly favorable SNP alleles and significant linear correlations between highly favorable allele numbers and corresponding phenotypic performance were observed. Two highly favorable SNP alleles correlating to flowering time and capitulum diameter were converted to derived cleaved amplified polymorphic sequence (dCAPS) markers to facilitate future breeding. Finally, six putative candidate genes were identified that contribute to flowering time and capitulum diameter. These results serve as a foundation for analyzing the genetic mechanisms underlying important horticultural traits and provide valuable insights into molecular marker-assisted selection (MAS) in chrysanthemum breeding programs.

     

Detecting responses to treatment with fenofibrate in pedigrees

Background

Fenofibrate (Fb) is a known treatment for elevated triglyceride (TG) levels. The Genetics of Lipid Lowering Drugs and Diet Network (GOLDN) study was designed to investigate potential contributors to the effects of Fb on TG levels. Here, we summarize the analyses of 8 papers whose authors had access to the GOLDN data and were grouped together because they pursued investigations into Fb treatment responses as part of GAW20. These papers report explorations of a variety of genetics, epigenetics, and study design questions. Data regarding treatment with 160 mg of micronized Fb per day for 3 weeks included pretreatment and posttreatment TG and methylation levels (ML) at approximately 450,000 epigenetic markers (cytosine-phosphate-guanine [CpG] sites). In addition, approximately 1 million single-nucleotide polymorphisms (SNPs) were genotyped or imputed in each of the study participants, drawn from 188 pedigrees.

Results

The analyses of a variety of subsets of the GOLDN data used a number of analytic approaches such as linear mixed models, a kernel score test, penalized regression, and artificial neural networks.

Conclusions

Results indicate that (a) CpG ML are responsive to Fb; (b) CpG ML should be included in models predicting the TG level responses to Fb; (c) common and rare variants are associated with TG responses to Fb; (d) the interactions of common variants and CpG ML should be included in models predicting the TG response; and (e) sample size is a critical factor in the successful construction of predictive models representing the response to Fb.

     

Comparing single-nucleotide polymorphism marker-based and microsatellite marker-based linkage analyses

We compared linkage analysis results for an alcoholism trait, ALDX1 (DSM-III-R and Feigner criteria) using a nonparametric linkage analysis method, which takes into account allele sharing among several affected persons, for both microsatellite and single-nucleotide polymorphism (SNP) markers (Affymetrix and Illumina) in the Collaborative Study on the Genetics of Alcoholism (COGA) dataset provided to participants at the Genetic Analysis Workshop 14 (GAW14). The two sets of linkage results from the dense Affymetrix SNP markers and less densely spaced Illumina SNP markers are very similar. The linkage analysis results from microsatellite and SNP markers are generally similar, but the match is not perfect. Strong linkage peaks were found on chromosome 7 in three sets of linkage analyses using both SNP and microsatellite marker data. We also observed that for SNP markers, using the given genetic map and using the map by converting 1 megabase pair (1 Mb) to 1 centimorgan (cM), did not change the linkage results. We recommend the use of the 1 Mb-to-1 cM converted map in a first round of linkage analysis with SNP markers in which map integration is an issue.     

Predictive ability of selected subsets of single nucleotide polymorphisms (SNPs) in a moderately sized dairy cattle population

Several studies have shown that computation of genomic estimated breeding values (GEBV) with accuracies significantly greater than parent average (PA) estimated breeding values (EBVs) requires genotyping of at least several thousand progeny-tested bulls. For all published analyses, GEBV computed from the selected samples of markers have lower or equal accuracy than GEBV derived on the basis of all valid single nucleotide polymorphisms (SNPs). In the current study, we report on four new methods for selection of markers. Milk, fat, protein, somatic cell score, fertility, persistency, herd life and the Israeli selection index were analyzed. The 972 Israeli Holstein bulls genotyped with EBV for milk production traits computed from daughter records in 2012 were assigned into a training set of 844 bulls with progeny test EBV in 2008, and a validation set of 128 young bulls. Numbers of bulls in the two sets varied slightly among the nonproduction traits. In EFF₁₂, SNPs were first selected for each trait based on the effects of each marker on the bulls’ 2012 EBV corrected for effective relationships, as determined by the SNP matrix. EFF₀₈ was the same as EFF₁₂, except that the SNPs were selected on the basis of the 2008 EBV. In DIF_max, the SNPs with the greatest differences in allelic frequency between the bulls in the training and validation sets were selected, whereas in DIF_min the SNPs with the smallest differences were selected. For all methods, the numbers of SNPs retained varied over the range of 300 to 6000. For each trait, except fertility, an optimum number of markers between 800 and 5000 was obtained for EFF₁₂, based on the correlation between the GEBV and current EBV of the validation bulls. For all traits, the difference between the correlation of GEBV and current EBV and the correlation of the PA and current EBV was >0.25. EFF₀₈ was inferior to EFF₁₂, and was generally no better than PA EBV. DIF_max always outperformed DIF_min and generally outperformed EFF₀₈ and PA. Furthermore, GEBV based on DIF_max were generally less biased than PA. It is likely that other methods of SNP selection could improve upon these results.     

Application of novel and existing methods to identify genes with evidence of epigenetic association: results from GAW20

Background

The rise in popularity and accessibility of DNA methylation data to evaluate epigenetic associations with disease has led to numerous methodological questions. As part of GAW20, our working group of 8 research groups focused on gene searching methods.

Results

Although the methods were varied, we identified 3 main themes within our group. First, many groups tackled the question of how best to use pedigree information in downstream analyses, finding that (a) the use of kinship matrices is common practice, (b) ascertainment corrections may be necessary, and (c) pedigree information may be useful for identifying parent-of-origin effects. Second, many groups also considered multimarker versus single-marker tests. Multimarker tests had modestly improved power versus single-marker methods on simulated data, and on real data identified additional associations that were not identified with single-marker methods, including identification of a gene with a strong biological interpretation. Finally, some of the groups explored methods to combine single-nucleotide polymorphism (SNP) and DNA methylation into a single association analysis.

Conclusions

A causal inference method showed promise at discovering new mechanisms of SNP activity; gene-based methods of summarizing SNP and DNA methylation data also showed promise. Even though numerous questions still remain in the analysis of DNA methylation data, our discussions at GAW20 suggest some emerging best practices.