ukb

Real data analysis

We are not able to share individual-level data so we describe procedures to select individuals and pre-process the dataset.

Admixed individual selection

We run SCOPE, a scalable program to infer admixture proportion in biobank-scale data. We used four population EUR, AFR, EAS, SAS, as ancestral population. We select individuals based on (EUR > 0.05) & (AFR > 0.05) & (EAS < 0.05) & (SAS < 0.05). This yields 4327 individuals as admixed population in this study. See 00_data.ipynb notebook for details.

Purple points were selected as admixed individuals with European and African ancestries

Genotype processing

With the subset of admixed individuals, we filter for hwe < 1e-6, MAF > 0.01, genotype missing rate < 0.05 to select SNPs, using the following comamand.

plink --bfile ${bfile} \
    --keep-fam ${admix_id} \
    --keep-allele-order \
    --make-bed \
    --hwe 1e-6 \
    --maf 0.01 \
    --geno 0.05 \
    --out ${out}

Then we perform phasing and imputation using TOPMed Imputation Server. We perform post-imputation QC to filter for imputation R2 > 0.8 and MAF > 0.5%.

bcftools filter -i 'INFO/R2>0.8 & INFO/MAF > 0.005' ${vcf_input} -Oz -o ${vcf_output}
tabix -p vcf ${vcf_output}

Local ancestry inference

We follow Tractor paper: we use AFR and EUR in 1000G reference panel and RFmix to infer local ancestry. See s03_rfmix.sh for details. The inferred local ancestry will be used as input to SNP1 and Tractor.

Association testing on known risk regions to lipid traits

We compared ATT, SNP1, Tractor on four well-known risk regions (LDLR, APOE, PCSK9, SORT1) of lipid traits (HDL, TC). For all three methods, we include age, sex, dilution factor, and top 10 PCs as covariates.

No inflation of p-values are observed for the three methods. We take +- 50kb window around the transcribed regions of each gene, and compare the association strength for each of the method.

See s04_assoc.ipynb for more details. See results/risk_regions.xlsx for numerical results.