LDAK-GBAT: Fast and powerful gene-based association testing using summary statistics

Abstract

We present LDAK-GBAT, a tool for gene-based association testing using summary statistics from genome-wide association studies that is computationally efficient, produces well-calibrated p values, and is significantly more powerful than existing tools. LDAK-GBAT takes approximately 30 min to analyze imputed data (2.9M common, genic SNPs), requiring less than 10 Gb memory. It shows good control of type 1 error given an appropriate reference panel. Across 109 phenotypes (82 from the UK Biobank, 18 from the Million Veteran Program, and nine from the Psychiatric Genetics Consortium), LDAK-GBAT finds on average 19% (SE: 1%) more significant genes than the existing tool sumFREGAT-ACAT, with even greater gains in comparison with MAGMA, GCTA-fastBAT, sumFREGAT-SKAT-O, and sumFREGAT-PCA.

Keywords: UK Biobank; complex traits; gene-based association testing; genome-wide association study; statistical genetics.

Figure 1

Number of significant genes We analyze each of the 109 phenotypes by using LDAK-GBAT and five existing tools. Bars report the relative number of genome-wide significant genes (p ≤ 2.5 × 10⁻⁶) from each tool, with LDAK-GBAT as the reference. Segments mark 95% confidence intervals for the ratios. The total number of genome-wide significant genes from each tool is reported above each bar.

Figure 2

Clumping results from LDAK-GBAT Segments provide p values from LDAK-GBAT when analyzing the MVP phenotype type 2 diabetes; red segments indicate significant genes that are selected by clumping, gold segments indicate significant genes that are discarded by clumping, while gray segments indicate non-significant genes. Blue points report p values for genic SNPs from single-SNP analysis. The gold and blue horizontal lines mark p = 2.5 × 10⁻⁶ and p = 5 × 10⁻⁸, respectively (the significance thresholds used with LDAK-GBAT and single-SNP analyses).

Figure 3

Varying the heritability model We analyze each of the 109 phenotypes by using LDAK-GBAT with seven heritability models, defined by ${\rm E}\left[h_j^2\right]\propto {\left[p_j\left(1-p_j\right)\right]}^{1+\alpha }$, where $p_j$ is the MAF of $SN{P}_j$ and α is −1.25, −1, −0.75, −0.5, −0.25, 0, or 0.25. Bars report the total number of genome-wide significant genes (p ≤ 2.5 × 10⁻⁶) for each heritability model. For comparison, the red horizontal line reports the total number of genome-wide significant genes from sumFREGAT-ACAT (the best existing tool).

Figure 4

Comparing LDAK-GBAT and single-SNP analysis for the first ten UK Biobank phenotypes The gold bar reports the total number of genome-wide significant genes (p ≤ 2.5 × 10⁻⁶) from LDAK-GBAT (using 50,000 individuals). The blue bars report how many of these genes contain a SNP with Bonferroni-corrected p ≤ 2.5 × 10⁻⁶ from single-SNP analysis as the sample size is increased from 50,000 to 200,000.