Bayesian model search and multilevel inference for SNP association studies

Technological advances in genotyping have given rise to hypothesis-based association studies of increasing scope. As a result, the scientific hypotheses addressed by these studies have become more complex and more difficult to address using existing analytic methodologies. Obstacles to analysis include inference in the face of multiple comparisons, complications arising from correlations among the SNPs (single nucleotide polymorphisms), choice of their genetic parametrization and missing data. In this paper we present an efficient Bayesian model search strategy that searches over the space of genetic markers and their genetic parametrization. The resulting method for Multilevel Inference of SNP Associations, MISA, allows computation of multilevel posterior probabilities and Bayes factors at the global, gene and SNP level, with the prior distribution on SNP inclusion in the model providing an intrinsic multiplicity correction. We use simulated data sets to characterize MISA’s statistical power, and show that MISA has higher power to detect association than standard procedures. Using data from the North Carolina Ovarian Cancer Study (NCOCS), MISA identifies variants that were not identified by standard methods and have been externally ``validated’’ in independent studies. We examine sensitivity of the NCOCS results to prior choice and method for imputing missing data. MISA is available in an R package on CRAN.

💡 Research Summary

The paper introduces a novel Bayesian framework called MISA (Multilevel Inference of SNP Associations) designed to address several persistent challenges in modern single‑nucleotide‑polymorphism (SNP) association studies. Traditional genome‑wide association analyses often rely on single‑marker tests followed by multiple‑testing corrections (e.g., Bonferroni, FDR) or on set‑based tests that ignore the uncertainty about the optimal genetic parametrization (additive, dominant, recessive) for each variant. Moreover, linkage disequilibrium (LD) creates strong correlations among markers, inflating the effective number of tests and reducing statistical power. Missing genotype data further complicates inference.

MISA tackles these issues by embedding variable selection and model choice within a single Bayesian hierarchical model. For each SNP the method defines two discrete latent variables: (i) an inclusion indicator that decides whether the marker participates in the disease model, and (ii) a parametrization indicator that selects among a small set of genetic models (e.g., additive, dominant, recessive). The prior on inclusion is a Bernoulli distribution with a low success probability, which serves as an intrinsic multiplicity correction: the more stringent the prior, the fewer spurious markers are admitted. The parametrization prior is typically uniform, reflecting no a priori preference for a particular genetic mode.

Because the total model space grows as 2^p × K (p = number of SNPs, K = number of parametrizations), exhaustive enumeration is impossible. The authors therefore develop an efficient stochastic search algorithm based on reversible‑jump Markov chain Monte Carlo (RJMCMC). At each iteration the chain proposes one of three moves: add a SNP, delete a SNP, or switch its parametrization. Acceptance probabilities are computed from the ratio of posterior probabilities, which combine the likelihood of the logistic disease model with the specified priors. The algorithm is equipped with adaptive tuning of proposal probabilities and a “sweeping” strategy that periodically revisits low‑probability regions to avoid local traps, ensuring good mixing even in high‑dimensional settings.

A key innovation of MISA is its multilevel inference capability. After the MCMC run, posterior samples are used to compute:

1. Global evidence – the posterior probability that at least one SNP is associated, providing a single number that quantifies overall genetic contribution.
2. Gene‑level evidence – the sum of posterior probabilities of all models that contain any SNP from a given gene, yielding a Bayes factor for each gene.
3. SNP‑level evidence – the marginal inclusion probability (PIP) for each variant and a Bayes factor comparing models that include the SNP versus those that do not.

These quantities allow researchers to rank genes and individual variants while accounting for model uncertainty and LD structure.

The authors evaluate MISA through extensive simulations that vary LD patterns, effect sizes, and numbers of causal SNPs. Compared with standard logistic regression with stepwise selection and with multiple‑testing corrections, MISA consistently shows higher true‑positive rates and comparable or lower false‑positive rates. Its advantage is most pronounced when causal variants are in high LD with many non‑causal neighbors, a scenario where conventional single‑marker tests lose power.

To demonstrate real‑world utility, the method is applied to data from the North Carolina Ovarian Cancer Study (NCOCS). Using the same genotype panel that previous analyses examined, MISA identifies several SNPs with high posterior inclusion probabilities that were not flagged by the original study. Independent replication studies have subsequently reported associations for many of these variants, providing external validation of MISA’s ability to uncover signals missed by traditional pipelines.

Missing genotype data are handled in two ways. First, a multiple‑imputation (MI) approach integrates over the posterior distribution of missing genotypes within the Bayesian model, effectively treating them as additional latent variables. Second, a simpler mean‑imputation scheme is examined for comparison. The MI strategy yields more stable posterior probabilities and is less sensitive to the proportion of missingness, confirming the benefit of fully Bayesian treatment of missing data.

MISA has been released as an R package on CRAN, complete with functions for data preprocessing, model‑search execution, posterior summarization, and graphical diagnostics. The package supports parallel computation, user‑defined priors, and can be scaled to genome‑wide data sets by employing chromosome‑wise or region‑wise analyses.

In conclusion, the paper presents a comprehensive Bayesian solution that simultaneously searches over marker inclusion, genetic parametrization, and handles missing data, while providing coherent multilevel posterior inference. Simulation studies and the NCOCS application illustrate that MISA achieves higher power and better interpretability than existing methods, making it a valuable addition to the toolkit of genetic epidemiologists and statistical geneticists.