Best practices for multi-ancestry, meta-analytic transcriptome-wide association studies: Lessons from the Global Biobank Meta-analysis Initiative

Abstract

The Global Biobank Meta-analysis Initiative (GBMI), through its diversity, provides a valuable opportunity to study population-wide and ancestry-specific genetic associations. However, with multiple ascertainment strategies and multi-ancestry study populations across biobanks, GBMI presents unique challenges in implementing statistical genetics methods. Transcriptome-wide association studies (TWASs) boost detection power for and provide biological context to genetic associations by integrating genetic variant-to-trait associations from genome-wide association studies (GWASs) with predictive models of gene expression. TWASs present unique challenges beyond GWASs, especially in a multi-biobank, meta-analytic setting. Here, we present the GBMI TWAS pipeline, outlining practical considerations for ancestry and tissue specificity, meta-analytic strategies, and open challenges at every step of the framework. We advise conducting ancestry-stratified TWASs using ancestry-specific expression models and meta-analyzing results using inverse-variance weighting, showing the least test statistic inflation. Our work provides a foundation for adding transcriptomic context to biobank-linked GWASs, allowing for ancestry-aware discovery to accelerate genomic medicine.

Graphical abstract

Figure 1

Challenges in multi-ancestry, meta-analytic TWASs Each level of data in a TWAS introduces a set of challenges: (1) genetics data include confounding from genetic ancestry, population structure and relatedness, and complex linkage disequilibrium patterns, (2) gene expression data introduces context-specific factors, such as tissue-, cell-type-, or cell-state-specific expression, and (3) phenotypic data involve challenges in acquiring and aggregating phenotypes, properly defining controls for phenotypes, and ascertainment and selection bias from non-random sampling.

Figure 2

Comparison of predictive performance of expression prediction models across ancestry (A) Adjusted R² difference (y axis) when predicting expression in the AFR imputation sample between models trained in EUR and AFR training samples across tissue (x axis). Proportion of models with improved R² using ancestry-aligned models versus ancestry-mismatched models is labeled. (B) Adjusted R² difference between ancestry-specific and ancestry-unaware models imputing into EUR (left) and AFR (right) samples.

Figure 3

Comparison of meta-analytic strategies for multi-biobank, multi-ancestry TWASs (A) Per-ancestry meta-analyzed TWAS scores across EUR (x axis) versus AFR ancestry (y axis). The dotted lines indicate p < 2.5×10⁻⁶ with a 45-degree line for reference. Points are colored by which ancestry population the TWAS association meets p < 2.5×10⁻⁶. (B) QQ-plot of TWAS Z scores , colored by meta-analytic strategies. Per ancestry refers to TWAS meta-analysis across meta-analyzed ancestry-specific GWAS summary statistics. Per bank/per ancestry refers to TWAS meta-analysis using all biobank- and ancestry-specific GWAS summary statistics. (C) Effect sizes and Bonferroni-corrected confidence intervals (CIs) for TWAS associations across 17 individual biobanks (EUR in teal, AFR in red) and two IVW meta-analysis strategies (in yellow) for five representative genes. The Higgins-Thompson I² statistic is provided.

Figure 4

GReX-PheWAS for categorizing phenome-wide associations for TAF7 genetically regulated expression in UKBB (A) –log₁₀ Benjamini-Hochberg FDR-adjusted p values of GTAs (y axis) across nine phenotype groups (x axis). Dotted line shows FDR-adjusted p = 0.05. (B) Miami plot of TWAS Z scores (y axis) across phenotypes (x axis), colored by phecode group. Dotted line shows Benjamini-Hochberg FDR-corrected significance, and phenotypes are labeled if the association passes Bonferroni correction.