Ancestry mismatches in discovery-GWAS and LD-reference compromise African-centric polygenic risk scores

A 2 × 2 factorial comparison of how the ancestry of the discovery GWAS and the ancestry of the LD reference panel shape the composition of a Type 2 Diabetes (T2D) polygenic risk score (PRS) built by P-value thresholding and LD clumping.

Scope: this project studies PRS construction and composition (which variants and weights end up in the score). It does not evaluate individual-level predictive accuracy.

---

Research question

Do the ancestry of the discovery GWAS and the ancestry of the LD reference independently or jointly influence the genetic composition of a T2D PRS?

Design

Pipeline	Discovery GWAS	LD reference
AFR-AFR	African ancestry	1000 Genomes AFR
AFR-EUR	African ancestry	1000 Genomes EUR
EUR-AFR	European ancestry	1000 Genomes AFR
EUR-EUR	European ancestry	1000 Genomes EUR

• Phenotype: Type 2 Diabetes (PheCode 250.2), Million Veteran Program (MVP) GWAS summary statistics for African and European, and 1000 Genomes Project Phase 3 for African and European LD-reference populations
• P-value thresholds (all pipelines): 5×10⁻⁸, 1×10⁻⁵, 1×10⁻⁴, 1×10⁻³, 0.01, 0.05.
• Changing one factor while holding the other fixed separates the GWAS effect from the LD effect and allows an interaction check.

Workflow

Notebooks

Part 1: Pipelines (one per ancestry combination, GWAS through PRS weights)

Notebook	Pipeline	Output
PRS_T2D_A-Gwas+A-LD.ipynb	AFR GWAS + AFR LD	candidate SNPs, clumped index SNPs, weights
PRS_T2D_A-Gwas+E-LD.ipynb	AFR GWAS + EUR LD	candidate SNPs, clumped index SNPs, weights
PRS_T2D_E-Gwas+A-LD.ipynb	EUR GWAS + AFR LD	candidate SNPs, clumped index SNPs, weights
PRS_T2D_E-Gwas+E-LD.ipynb	EUR GWAS + EUR LD	candidate SNPs, clumped index SNPs, weights

Part 2: Comparisons

Notebook	Question	Main factor
PRS_Candidate_SNP_Comparison.ipynb	Does GWAS ancestry change the initial variant pool?	GWAS ancestry
PRS_LD_Clumping_Comparison.ipynb	Does LD-reference ancestry change which independent variants are kept?	LD ancestry
PRS_PRS_Weight_Comparison.ipynb	How does the final weighted variant set differ (BETA, EAF, P, overlap)?	GWAS + LD
PRS_Integrated_Comparison.ipynb	What are the separate and joint effects?	Both

Key results

1. Candidate SNPs

• The same GWAS gave identical candidate counts under both LD references (pipeline consistency control passed).
• The EUR GWAS produced about 2.0-2.5x more candidates (66,996 vs 27,292 at 5×10⁻⁸; 3,066,477 vs 1,500,187 at 0.05).
• 92% of AFR candidates at 5×10⁻⁸ were also in the EUR set; only 37% of EUR candidates were in the AFR set.

2. LD clumping

Index SNPs per pipeline:

Threshold	AFR-AFR	AFR-EUR	EUR-AFR	EUR-EUR
5×10⁻⁸	1,710	633	4,944	2,063
1×10⁻⁵	4,305	1,636	16,558	9,216
1×10⁻⁴	7,666	3,302	37,511	25,465
1×10⁻³	19,691	11,288	107,335	83,925
0.01	90,187	74,325	364,262	319,256
0.05	357,804	358,128	920,240	842,495

• AFR LD retained 2.7x (AFR GWAS) and 2.4x (EUR GWAS) more index SNPs than EUR LD at 5×10⁻⁸.
• Equal counts do not mean equal content: at P ≤ 0.05 the AFR-GWAS counts were nearly identical, yet 224,817 variants were shared (about 63% of each set; Jaccard 0.46).
• Jaccard overlap between LD references was lowest at strict thresholds (0.28 for AFR GWAS, 0.32 for EUR GWAS at 5×10⁻⁸).

3. PRS weights

• Index counts equal weight counts in all 24 files.
• EUR weight sets have larger median |BETA| and much higher median EAF (0.78-0.999) than AFR sets (0.01-0.46). No MAF filter was applied, so low-frequency variants are included.
• These are differences in composition, not in score quality.

4. Integrated comparison

Share of index SNPs lost when switching AFR LD to EUR LD:

Threshold	AFR GWAS	EUR GWAS
5×10⁻⁸	63%	58%
1×10⁻⁵	62%	44%
1×10⁻⁴	57%	32%
1×10⁻³	43%	22%
0.01	18%	12%
0.05	~0%	8%

The LD-reference effect is larger for the AFR GWAS at most thresholds, so the two factors interact (descriptive, no statistical test).

Tools

Python Google colab PLINK 2 NumPy Pandas

Data quality note

A weight-file audit found that at multi-allelic sites some EUR weight rows carried the wrong allele's BETA, EAF and P (16-1,911 rows per file, 0.2-0.8%). Clumping and index SNP selection were not affected. Weights were rebuilt from the candidate files (lowest-P allele retained, assumed to match the clumping selection) and saved as *_FIXED.tsv. The corrected files contain the same SNPs as the originals, and all retained variants meet their thresholds. No issues were found in the AFR files. Audit summary: weight_fix_report.tsv.

Conclusions

• Discovery-GWAS ancestry sets the size and identity of the starting variant pool.
• LD-reference ancestry changes which independent variants survive clumping, even when counts converge.
• Both factors shape T2D PRS composition, and their effects depend on the P-value threshold and on each other.
• For T2D, a condition with a large burden across populations, this suggests that ancestry-specific choices in score construction can change which variants a score is built from, which may contribute to poor portability of scores across ancestries. Predictive performance was not tested.

Limitations

• No individual-level validation; a follow-up should test the scores in independent AFR and EUR cohorts.
• No MAF filter; rare variants likely contribute to large EUR effect sizes (untested).
• Variant overlap is by SNP ID, not allele-harmonized or locus-based.
• The 1000 Genomes AFR and EUR panels differ in size and composition.
• Mechanistic explanations (e.g., shorter LD blocks in AFR) are interpretations, not tested here.

Reproducibility

• Runs in Google Colab with data on Google Drive.
• Python: pandas, numpy, matplotlib; PLINK for clumping.
• TODO: clumping parameters (r², window), GWAS sample sizes, 1000 Genomes panel sizes, genome build.