Abstract: Variable selection and large-scale hypothesis testing are techniques commonly used to analyze high-dimensional genomic data. Despite recent advances in theory and methodology, variable selection and inference with highly collinear features remain challenging. For instance, collinearity poses a great challenge in Genome-Wide Association Studies (GWAS) involving millions of variants, many of which may be in high linkage disequilibrium. In such settings, collinearity can significantly reduce the power of variable selection methods to identify individual variants associated with an outcome. To address such challenges, we developed a Bayesian Hierarchical Hypothesis Testing (BHHT)–a novel multi-resolution testing procedure that offers high power with adequate error control and fine-mapping resolution. We demonstrate through simulations that the proposed methodology has a power-FDR performance that is competitive with (and in many scenarios better than) state-of-the-art methods. Finally, we demonstrate the feasibility of using the proposed methodology with big data to map risk variants for serum urate using data (n∼300,000) on phenotype and ultra-high-dimensional genotypes (∼15 million SNPs) from the UK-Biobank. Our results show that the proposed methodology leads to many more discoveries than those obtained using traditional feature-centered inference procedures. The article is accompanied by open-source software that implements the methods described in this study using algorithms that scale to biobank-size ultra-high-dimensional data.
Introduction
In modern genomics, identifying Single Nucleotide Polymorphisms (SNPs) associated with complex traits is a primary goal. However, Genome-Wide Association Studies (GWAS) are often hampered by Linkage Disequilibrium (LD), where SNPs are inherited together in blocks. This correlation, or collinearity, causes a "signal-spreading" problem: a standard Bayesian model might distribute the posterior probability of a single causal variant across multiple correlated SNPs. Consequently, no individual marker reaches the significance threshold, leading to a loss of discovery power. This paper introduces Bayesian Hierarchical Hypothesis Testing (BHHT), a multi-resolution framework designed to overcome this "collinearity trap."
Methodology: Multi-Resolution Testing
BHHT shifts the focus from individual variants to hierarchical clusters. The process begins with a Bayesian Variable Selection (BVS) model, such as a "Spike-and-Slab" prior, to estimate effects across all variants. SNPs are then organized into a tree structure based on physical location and correlation.
The core of the methodology is a multi-resolution testing procedure. Instead of testing every individual SNP, the algorithm starts at the root of the tree (representing large genomic regions). If a region shows a high Posterior Inclusion Probability (PIP), the algorithm moves down the branches to test finer resolutions. To maintain statistical rigor, we introduce the Discovery Set Bayesian False Discovery Rate (DS-BFDR), which controls the error rate for the entire set of discovered regions rather than at each node, maximizing power while ensuring reliability.
Experimental Validation
The BHHT framework was validated using three distinct approaches:
- Simulated Data: 1,000 replicates were used to benchmark BHHT against state-of-the-art tools like SuSiE under varying correlation levels ($\rho = 0.45$ and $0.90$).
- UK Biobank (Simulated Phenotypes): Real human genotypes from ~300,000 individuals were used to test performance under realistic genomic complexity.
- UK Biobank (Serum Urate): A real-world application mapping risk variants for serum urate using ~15 million SNPs, testing the algorithm’s scalability to ultra-high-dimensional data.
Results and Discussion
Simulations revealed that BHHT achieves higher power than SuSiE in scenarios with multiple causal variants in a single region, as it avoids the local optima issues sometimes encountered by optimization-based methods. A key observation was the "U-shaped" distribution of probabilities across clusters, where values tend toward 0 or 1, providing stable FDR control.
In the serum urate analysis, traditional marginal testing identified 98 significant SNPs. BHHT identified 180 discoveries—an 83% increase. These discoveries include both single variants and "credible sets" of variants, offering a more complete picture of the genetic factors influencing the trait.
Conclusion
BHHT provides a powerful, scalable solution for GWAS in the presence of high collinearity. By leveraging a multi-resolution approach and a novel FDR control algorithm, it significantly increases discovery power compared to traditional feature-centered methods. Future research will focus on extending this framework to pleiotropic traits and other high-dimensional data types, such as neuroimaging.
Software Availability: The authors have released the Bayes_HHT R package on GitHub to allow other researchers to apply these multi-resolution tests to their own high-dimensional datasets.
Summary updated