Genome-wide Association Studies in Ancestrally Diverse Populations (2019)

1. Problem & Tasks

• Review methodological pitfalls and recommendations for conducting GWAS beyond European cohorts.
• Critical for how we handle ancestry PCs and covariates when integrating genetics with MRI.

2. Datasets

• Summarizes lessons from PAGE, HCHS/SOL, UKB, biobanks in East Asia and Africa.
• Emphasizes differences in sample sizes and LD structure across populations.

3. Model / Method Details

3.1 Key Recommendations

• Use linear mixed models or ancestry-informed stratification to control population structure.
• Include local ancestry estimates when available; otherwise rely on global PCs.
• Validate PRS / embeddings separately per ancestry group; avoid pooling unless harmonized.

3.2 Confound Handling & Evaluation Discipline

• Check for residual stratification using QQ plots and genomic control.
• Report ancestry composition and site distribution with each GWAS/PRS result.
• Use jackknife/bootstraps for uncertainty when sample sizes are small per subgroup.

4. Results & Tables

• Quantifies performance drop (R² decrease up to 5×) when PRS trained in Europeans applied to African ancestry cohorts.
• Provides tables linking LD differences to effect estimation bias.

5. Limitations & Cautions

• 2019 state of the field; newer multi-ancestry methods exist but core cautions remain.
• Does not cover foundation models directly; we must adapt recommendations to embeddings.

---

6. Hooks into Neuro-Omics KB

Relevant KB assets

• kb/paper_cards/gwas_diverse_populations.yaml
• kb/datasets/ukb_manifest_stub.yaml (records ancestry composition, PCs).

Configs / recipes informed

• Covariate list (age/sex/site/PCs) in configs/experiments/01_cca_gene_smri.yaml and 02_prediction_baselines.yaml.
• Future documentation for PRS/embedding fairness analyses.

Concrete guidance for our project

• Always log ancestry distribution and include PCs in dataset cards; cite this paper when explaining why.
• If/when we add non-European cohorts, run stratified evaluations instead of assuming global models transfer.
• Keep hooks for local ancestry or site-specific PCs should we extend beyond UKB Europeans.