Mixed-model association for biobank-scale datasets

Abstract

Biobank-based genome-wide association studies are enabling exciting insights in complex trait genetics, but much uncertainty remains over best practices for optimizing statistical power and computational efficiency in GWAS while controlling confounders. Here, we introduce a much faster version of our BOLT-LMM Bayesian mixed model association method—capable of running analyses of the full UK Biobank cohort in a few days on a single compute node—and show that it produces highly powered, robust test statistics when run on all 459K European samples (retaining related individuals). When used to conduct a GWAS for height in UK Biobank, BOLT-LMM achieved power equivalent to linear regression on 650K samples—a 93% increase in effective sample size versus the common practice of analyzing unrelated British samples using linear regression (UK Biobank documentation; Bycroft et al. bioRxiv). Across a broader set of 23 highly heritable traits, the total number of independent GWAS loci detected increased from 5,839 to 10,759, an 84% increase. We recommend the use of BOLT-LMM (retaining related individuals) for biobank-scale analyses, and we have publicly released BOLT-LMM summary association statistics for the 23 traits analyzed as a resource for all researchers.

Figure 1:. Power, calibration, and speed of BOLT-LMM v2.3 in UK Biobank analyses.

(a) Numbers of independent genome-wide significant associations (p<5×10–9) identified by BOLT-LMM analyses of all European-ancestry individuals (N=459,327) versus linear regression analyses of unrelated British individuals (N=337,539, following common practice [5]). Results for 23 phenotypes are plotted, with 8 representative phenotypes highlighted. (b) Variance explained by genome-wide SNPs on which BOLT-LMM implicitly conditions to increase power. Conditioning on BOLT-LMM’s polygenic predictions—which attain accuracy (r2BOLT-LMM) approaching SNP-heritability (hg2) for some traits—achieves effective sample sizes as high as ~700K. (We measured effective sample size by comparing χ² statistics at associated SNPs; Supplementary Note.) (c) Test statistic calibration of BOLT-LMM on all European individuals versus linear regression on unrelated British individuals (using 20 principal component covariates). Attenuation ratios from LD score regression [7, 8] match closely between the two methods, indicating that BOLT-LMM properly controls false positives (Supplementary Fig. 2). Error bars, jackknife s.e. (d) Computational cost of association analysis using BOLT-LMM v2.3, the previous version of BOLT-LMM [3], and linear regression (implemented efficiently within the BOLT-LMM software) on the UK Biobank N=150K and N=500K data releases. Analyses were run on 8 threads on a 2.10 GHz Intel Xeon E5–2683 v4 processor. Additional details and numerical data are provided in the Supplementary Note, Supplementary Fig. 1, and Supplementary Tables 1–7.