N-GWAMA / MA-GWAMA
Summary¶
N-GWAMA and MA-GWAMA are two methods for multivariate genome-wide-association meta-analysis (GWAMA) that pool summary statistics across genetically correlated traits, using only published GWAS/GWAMA summary statistics (no individual-level genotypes required) and LD score regression (LDSC) to estimate and correct for unknown sample overlap between the input studies. Baselmans, Jansen, Ip, et al. (2019) introduce both methods and apply them to a "well-being spectrum" meta-analysis of life satisfaction, positive affect, neuroticism, and depressive symptoms (N=2,370,390), finding 304 independent loci — a 26% increase over the four univariate GWAMAs — and a ~57% average gain in polygenic score predictive power [1].
Method¶
Both methods define a multivariate model as one in which a single SNP's effect is considered jointly across multiple traits, differing in how strictly they enforce a shared effect:
- N-GWAMA: assumes a unitary (shared) SNP effect across all input traits, weighted by sample size.
- MA-GWAMA: relaxes the unitary-effect assumption, generating trait-specific effect estimates per SNP (comparable to the "Sₕₑₜ" statistic) and allowing heterogeneous SNP effects across traits.
Both methods estimate the error correlation between univariate GWAMA effect sizes — induced by unknown sample overlap — from LDSC applied to the univariate summary statistics, and use the univariate LDSC intercept to correct for population stratification and cryptic relatedness. Relative to prior multivariate methods, N-GWAMA/MA-GWAMA require only summary statistics (unlike MultiPhen, mv-PLINK/CCA, Combined-PC, and mv-BIMBAM, which need individual-level genotypes), estimate error correlation more precisely than Sₕₒₘ, generate signed effect-size estimates usable for polygenic scoring (unlike TATES, which produces only a P-value), and — for MA-GWAMA — allow per-SNP heterogeneity across traits (unlike MTAG, which assumes a shared effect).
Simulation Validation¶
Across simulated scenarios (four traits, h²SNP = 30%, genetic correlations 0.1–0.9, sample overlap 0–25,000 of 40,000 individuals per trait), both methods outperformed univariate GWAMA when genetic correlations were ≥ 0.5; the advantage disappeared at rg ≤ 0.4. When a causal SNP influenced 3 of 4 simulated traits, N-GWAMA and MA-GWAMA performed equally; when a SNP influenced only 1 or 2 of 4 traits, MA-GWAMA outperformed N-GWAMA (and, in scenarios where SNPs affect fewer than half the traits considered, univariate GWAMA can outperform both multivariate methods).
Application: The Well-Being Spectrum¶
Combining published univariate GWAMAs of life satisfaction (N=80,852), positive affect (N=410,603), neuroticism (N=582,989), and depressive symptoms (N=1,295,946) — mean cross-trait genetic correlation rg=0.7 — into a joint "well-being spectrum" analysis (N=2,370,390):
- N-GWAMA identified 231 independent loci; MA-GWAMA identified 148/191/263/239 loci for life satisfaction/positive affect/neuroticism/depressive symptoms respectively (289 independent signals after merging, 78.3% within 50 kb of an N-GWAMA signal). Combined, the two methods yielded 304 independent genome-wide signals, a 26% increase over the four univariate GWAMAs.
- Polygenic scores built from N-GWAMA/MA-GWAMA outperformed univariate-GWAMA-derived scores for all four traits (e.g., depressive symptoms incremental R² rose from 1.22% univariate to 1.64%/1.63%), with an average 59% (N-GWAMA) and 55% (MA-GWAMA) gain in predictive power.
- Follow-on TWAS uncovered 97 transcript-trait associations (45 loci, 17 not previously significant in the N-GWAMA locus set); MWAS uncovered 913 CpG methylation-trait associations (141 loci, 75 not previously significant).
- Stratified LDSC (histone-mark annotations, regional brain gene-expression atlases, single-nucleus RNA-seq) linked the well-being spectrum signal to central-nervous-system tissue, specifically the subiculum (hippocampal formation; Z=3.47, P<0.001) and GABAergic interneurons in hippocampus/prefrontal cortex (Z=3.42–3.70), with negative controls (educational attainment, height) showing no or weaker enrichment in the same regions.
Availability¶
Software: github.com/baselmans/multivariate_GWAMA.
See Also¶
- GWIS (Genome-Wide Inferred Statistics) — a related but distinct multivariate summary-statistics method from an overlapping research group (shared senior author Michel Nivard): GWIS infers the GWAS of a single phenotype defined as a function of other phenotypes, whereas N-GWAMA/MA-GWAMA meta-analyze several already-measured, genetically correlated phenotypes into a joint signal.