Set-based rare variant association tests for biobank scale sequencing data sets

Wei Zhou, Wenjian Bi, Zhangchen Zhao, Kushal K. Dey, Karthik A. Jagadeesh, Konrad J. Karczewski, Mark J. Daly, Benjamin M. Neale, Seunggeun LeemedRxiv 2021 · DOI

Abstract

As large-scale whole-exome sequencing (WES) data becomes available through resources like the UK Biobank (UKBB), identifying rare variant associations remains a significant challenge. Existing methods often suffer from inflated Type I error rates or prohibitive computational costs. This paper introduces SAIGE-GENE+, a scalable statistical framework that optimizes rare variant testing through ultra-rare variant collapsing and multi-set aggregation. In an analysis of UKBB data covering 30 quantitative and 141 binary traits, SAIGE-GENE+ identified 551 gene-phenotype associations while achieving up to a 1,400x speedup over previous iterations.

Introduction

The search for "rare variants"—genetic changes occurring in less than 1% of the population—is central to understanding the genetic architecture of complex diseases. However, the extreme sparsity of these variants in biobank-scale datasets creates two primary hurdles. First, standard statistical tests often break down when cases are rare, leading to false positives (Type I error inflation). Second, the computational burden of testing tens of thousands of genes across hundreds of thousands of individuals, while accounting for complex population structures, is immense.

Methodology

SAIGE-GENE+ addresses these challenges through several key innovations:

  • Ultra-rare Variant Collapsing: To stabilize statistics in sparse data, variants with a Minor Allele Count (MAC) of 10 or less are collapsed into a single "absence/presence" marker. This reduces the number of zero-heavy variables, making mathematical approximations (like the saddlepoint approximation) far more accurate.
  • Multi-set Aggregation: Rather than relying on a single frequency threshold, the tool performs tests across multiple Minor Allele Frequency (MAF) cutoffs (1%, 0.1%, 0.01%) and functional categories (e.g., Loss-of-Function vs. missense). These results are then combined using the Cauchy combination test to produce a single, robust p-value.
  • Computational Optimization: The framework utilizes a sparse Genetic Relationship Matrix (GRM) to account for participant relatedness and optimizes data handling to minimize memory usage, allowing for the analysis of massive datasets on standard high-performance computing clusters.

Key Results

Performance and Accuracy

SAIGE-GENE+ maintains well-controlled Type I error rates even in highly unbalanced scenarios (case-control ratios as extreme as 1:267). This reliability is crucial for studying rare diseases where the number of affected individuals is small relative to the total population.

Quantile-Quantile (QQ) plots for four binary traits (eczema, type 2 diabetes, multiple sclerosis, and thyroid problems) across three MAF cutoffs (<=0.01%, <=0.1%, <=1%). The plots demonstrate the method's ability to control Type I error and identify associations across varying frequency thresholds.

Biological Discovery

The team identified 551 gene-phenotype associations in the UKBB WES data. The results highlighted the importance of testing multiple MAF thresholds; some signals only became apparent when restricting the analysis to ultra-rare variants (MAF ≤ 0.01%), as broader thresholds can sometimes introduce genetic noise that dilutes the signal.

Detailed QQ plot for eczema/dermatitis highlighting the significant association with the FLG gene. The green and blue points represent SKAT and SKAT-O tests, respectively, which are aggregated by the SAIGE-GENE+ framework.

Computational Efficiency

The efficiency gains were dramatic. For the TTN gene—the largest in the human genome—computation time was reduced from nearly 10,000 minutes to just 7 minutes. This 1,400x improvement, paired with a reduction in memory usage from 65GB to 2.1GB, makes biobank-wide rare variant screening feasible.

Conclusion

SAIGE-GENE+ provides a robust and efficient solution for rare variant association studies in the era of massive sequencing datasets. By fixing the mathematical issues associated with data sparsity and optimizing the computational workflow, it enables researchers to uncover novel genetic drivers of disease that were previously inaccessible due to statistical noise or technical limitations.

Summary updated

Abstract

Abstract UK Biobank has released the large-scale whole-exome sequencing (WES) data, but the best practices remain unclear for rare variant tests, and an existing approach, SAIGE-GENE, can have inflated type I error rates with high computation costs. Here, we propose SAIGE-GENE+ with greatly improved type I error control and computational efficiency compared to SAIGE-GENE. In the analysis of UKBB WES data of 30 quantitative and 141 binary traits, SAIGE-GENE+ identified 551 gene-phenotype associations. In addition, we showed that incorporating multiple MAF cutoffs and functional annotations can help identify novel gene-phenotype associations.
BibTeX
@misc{zhou2021set,
  author = {Wei Zhou and Wenjian Bi and Zhangchen Zhao and Kushal K. Dey and Karthik A. Jagadeesh and Konrad J. Karczewski and Mark J. Daly and Benjamin M. Neale and Seunggeun Lee},
  title = {{Set-based rare variant association tests for biobank scale sequencing data sets}},
  year = {2021},
  doi = {10.1101/2021.07.12.21260400},
  howpublished = {medRxiv},
}
Figure