Ultra-high-dimensional genomic studies require simultaneous identification of sparse signals, control of spurious discoveries, and computationally feasible inference when the number of predictors can reach hundreds of thousands or more. We propose Bayesian Univariate-Guided Sparse Regression (BUGS), a global--local shrinkage framework that incorporates marginal association information directly into the prior through continuous modulation of shrinkage. By embedding univariate guidance within the nonlinear variance structure of a regularized horseshoe prior, BUGS adaptively differentiates predictors with strong and weak marginal evidence without relying on hard screening. We establish theoretical results on prior concentration, posterior contraction, and guidance-induced shrinkage separation, and characterize the behavior of the prior under uninformative guidance. For ultra-high-dimensional problems, we develop BUGS-Active, an active-set MCMC approximation that reduces the cost of local-scale updates from $O(p)$ to $O(|A_n|)$ while retaining global coefficient updates. Simulations under independent and correlated designs show strong signal recovery and substantially lower empirical false discovery rates relative to competing methods. We apply BUGS-Active to a DNA methylation study of developmental age with $n=1051$ samples and approximately $850{,}000$ CpG sites, jointly analyzing all probes without pre-screening. The analysis yields strong out-of-sample prediction and a sparse set of age-associated CpG sites with stable posterior support. These results illustrate the practical value of marginally guided Bayesian shrinkage for large-scale genomic regression.

翻译：暂无翻译