A powerful subset-based gene-set analysis method identifies novel associations and improves interpretation in UK Biobank

Diptavo Dutta, Peter VandeHaar, Lars G. Fritsche, Sebastian Zöllner, Michael Boehnke, Laura J. Scott, Seunggeun LeebioRxiv 2019 · DOI

TL;DR

This paper introduces GAUSS, a novel gene-set association method that significantly improves power and interpretability in genetic studies. GAUSS not only identifies associations between sets of genes and phenotypes but also pinpoints the specific "core subset" of genes driving these associations, offering deeper biological insights than existing methods.

What Problem Does This Paper Address?

Genome-wide association studies (GWAS) have identified thousands of genetic variants linked to complex diseases. However, these variants often explain only a small fraction of heritability. Gene-set analysis (GSA) aims to overcome this by examining groups of genes with shared biological functions, potentially revealing associations missed by single-variant analyses. The challenge is that many existing GSA methods lack the power to detect associations when only a small subset of genes within a set are truly involved. Furthermore, these methods don't identify which genes within a set are the most important drivers of the association, hindering downstream interpretation and the identification of therapeutic targets.

Methods

The core innovation is the Gene-set analysis Association Using Sparse Signals (GAUSS) method. Unlike traditional GSA that might average signals across all genes in a set, GAUSS employs a subset-based statistic. It identifies the subset of genes within a gene-set that exhibits the maximal evidence of association. This "core subset" (CS) is key for interpretation.

GAUSS works in two main steps:

  1. Test Statistic Calculation: Gene-based p-values (obtained from GWAS summary statistics or individual-level data) are converted to z-scores. The GAUSS statistic is then defined as the maximum association score among all possible non-empty subsets of genes within a gene-set. This is computationally optimized to O(mlogm).
  2. P-value Calculation: Due to linkage disequilibrium (LD) among genes, z-scores are dependent. GAUSS uses a fast simulation approach based on a normal-Copula to estimate the null distribution. It leverages pre-computed correlation structures from reference panels and employs adaptive resampling and a generalized Pareto distribution (GPD) for estimating very small p-values, making it computationally efficient.

Datasets

The study utilizes:

  • UK Biobank: GWAS summary statistics for 1,403 binary phenotypes and 10,679 gene-sets (from MSigDB v6.2, including curated gene-sets and GO terms).
  • 1000 Genomes Project: European ancestry data used as a reference panel for estimating LD and correlation structures.
  • Michigan Genomics Initiative (MGI): An independent dataset of ~38,000 European samples used for validating findings from the UK Biobank.

Key Insights and Ideas

  • Focusing on the "Core Subset" is Powerful: By identifying the subset of genes with the strongest signal, GAUSS can aggregate weaker associations that would otherwise be missed, leading to increased power.
  • Improved Interpretability: Pinpointing the CS genes provides direct biological context, helping researchers understand why a gene-set is associated with a phenotype and suggesting potential targets for further investigation.
  • Computational Efficiency: GAUSS's simulation-based approach, combined with adaptive resampling and GPD estimation, makes it significantly faster than traditional permutation-based methods, enabling large-scale analyses.
  • Robustness to Sparsity: GAUSS excels when only a small fraction of genes in a set are truly associated with the phenotype, a common scenario in complex trait genetics.

Key Results

  • Increased Power: Simulations showed GAUSS consistently outperformed existing methods (aSPUpath, SKAT-Pathway, MAGMA), especially in sparse scenarios where few genes drive the association.
  • Accurate Type I Error Control: GAUSS maintained well-calibrated type I error rates across various gene-set sizes and significance thresholds.
  • Large-Scale UK Biobank Analysis: GAUSS identified 13,466 significant phenotype-gene-set associations from 1,403 phenotypes and 10,679 gene-sets.
  • Identification of Driving Genes: For significant associations, GAUSS identified an average of 17.2 CS genes, with some cases identifying up to 405 genes. Notably, many of these CS genes were not individually significant at a stringent gene-based cutoff.
  • Validation: Findings from UK Biobank were partially validated in the smaller MGI dataset, supporting the robustness of the identified associations.
  • Computational Speed: GAUSS was comparable in speed to MAGMA and significantly faster than aSPUpath for large-scale analyses.

Limitations and Future Work

  • Self-Contained Null Hypothesis Only: GAUSS is designed to test the self-contained null hypothesis (no genes in the set are associated with the phenotype) and does not directly provide information on gene-set enrichment.
  • GPD Method Needs Further Research: The authors suggest that the GPD method for estimating very small p-values requires additional research and exploration.
  • Reference Panel Choice: While the choice of reference panel (1000 Genomes vs. UK Biobank) did not substantially impact results, it's important for the reference and study samples to share similar ancestry to reflect accurate LD patterns.

Future work could explore extending GAUSS to test other null hypotheses or integrate its findings with other types of genetic data.

Summary updated

Abstract

Abstract Tests of association between a phenotype and a set of genes in a biological pathway can provide insights into the genetic architecture of complex phenotypes beyond those obtained from single variant or single gene association analysis. However, most existing gene set tests have limited power to detect gene set-phenotype association when a small fraction of the genes are associated with the phenotype, and no method exists which identifies the potentially “active” genes that might drive a gene-set-based association. To address these issues, we have developed Gene-set analysis Association Using Sparse Signals (GAUSS), a method for gene-set association analysis that requires only GWAS summary statistics. For each significantly associated gene set, GAUSS identifies the subset of genes that have the maximal evidence of association and can best account for the gene set association. Using pre-computed correlation structure among test statistics from a reference panel, our p-value calculation is substantially faster than other permutation or simulation-based approaches. In simulations with varying proportions of causal genes, we find that GAUSS effectively controls type 1 error rate and has greater power than several existing methods, particularly when a small proportion of genes account for the gene set signal. Using GAUSS, we analyzed UK Biobank GWAS summary statistics for 10,679 gene-sets and 1,403 binary phenotypes. We found that GAUSS is scalable and identified 13,466 phenotype and gene-set association pairs. Within these genes sets, we identify an average of 17.2 (max=405) genes that underlie these gene set associations.
BibTeX
@misc{dutta2019powerful,
  author = {Diptavo Dutta and Peter VandeHaar and Lars G. Fritsche and Sebastian Zöllner and Michael Boehnke and Laura J. Scott and Seunggeun Lee},
  title = {{A powerful subset-based gene-set analysis method identifies novel associations and improves interpretation in UK Biobank}},
  year = {2019},
  doi = {10.1101/799791},
  howpublished = {bioRxiv},
}
Figure