This repository details the quality control (QC) pipeline of the MGB Biobank samples genotyped on Illumina’s GSA array, which largely follows the recommendations described the reference below and leverages the scripts (https://github.com/Annefeng/PBK-QC-pipeline) used to QC ealier MGBB releases genotyped on Illumina’s Multi-Ethnic Global array:
Peterson, R. E., Kuchenbaecker, K., Walters, R. K., Chen, C.-Y., Popejoy, A. B., Periyasamy, S., et al. (2019). Genome-wide Association Studies in Ancestrally Diverse Populations: Opportunities, Methods, Pitfalls, and Recommendations. Cell, 179(3), 589–603. http://doi.org/10.1016/j.cell.2019.08.051
The dataset includes 47,321 individuals genotyped on Illumina’s GSA array in hg19 coordinates.
NOTE: There is substantial overlap between the samples genotyped on the MEG array and GSA array.
-
QC for each genotyping batch (metrics not affected by ancestry;
scripts 01-04)- SNP-level call rate >0.95
- Sample-level call rate >0.98
- SNP-level call rate >0.98
- Maximum SNP-level missing rate difference between two batches < an empirical threshold (0.75%)
-
Merge genotyping batches (
scripts 05-06)- Remove duplicated SNPs
- Remove monomorphic SNPs
- Remove SNPs not confidently mapped (chr=0)
-
Population assignment (
scripts 07-08)- Select common, high-quality SNPs for population inference
- SNP-level call rate >0.98
- MAF >5%
- Remove strand ambiguous SNPs and long-range LD regions (chr6:25-35Mb; chr8:7-13Mb inversion)
- Prune to <100K independent SNPs
- Identify individuals with European ancestry using selected SNPs
- Run PCA combining study samples + 1KG data
- Use Random Forest to classify genetic ancesty with a prediciton prob. >0.8
- Remove PCA outliers (6SD away from the population mean for each of the top 10 PCs)
- Select common, high-quality SNPs for population inference
-
QC within European samples (metrics may be affected by ancestry;
scripts 09-13)- Remove samples that fail sex check (--check-sex)
- Absolute value of autosomal heterozygosity rate deviating from the mean (5SD; --het)
- Remove SNPs that show batch associations
- Regress each batch indicator on SNPs, adjusting for sex (association P < 1e-4)
-
Calculate PCs within European samples using common, high-quality SNPs (
script 14) -
Final SNP-level QC within European samples (
script 15)- SNP-level call rate >0.98
- HWE >1e-10
- Retain only autosomal SNPs, excluding indels and monomorphic SNPs (for imputation; HRC: SNP only)
-
Prepare data for HRC/1KG imputation using Michigan server (
scripts 16-17)- Harmonize study data with HRC/1KG data
- Convert plink to vcf by chromosome
-
Send European samples to Michigan server for imputation (using HRC as the reference panel)
-
Post-imputation QC (converting vcf dosages to plink hard-call genotypes;
scripts 18-20)- INFO score/Imputation R2 >0.6
- MAF (based on imputed dosages) >0.5%
- HWE >1e-10
- SNP-level call rate >0.98
- Random Forest prediction probability > 0.8 based on top 6 PCs, with 1KG samples as the training data
- "Unclassified" consists of mostly admixed individuals
| 1KG superpop | EUR | AMR | AFR | EAS | SAS | Unclassified | Total |
|---|---|---|---|---|---|---|---|
| # Samples | 33,614 | 2,599 | 2,036 | 908 | 480 | 7,682 | 47,319 |
| % Total | 71.1% | 5.5% | 4.3% | 1.9% | 1.0% | 16.2% | 100% |
- Samples are genotyped on 3 batches. Batch 0110: 13,140; batch 0111: 11,649; batch 0113: 22,532
| Sample QC metric | # Samples | % Total |
|---|---|---|
| Initial sample size | 47,321 | 100% |
| Batch QC: | ||
| Sample-level call rate < 0.98 | 2 | 0.0% |
| Merged QC: | ||
| Non-European | 13,705 | 29.0% |
| EUR (pop-specific) QC: | ||
| Failing sex check (reported != imputed sex, using F < 0.2 for female & >0.8 for male) |
332 | 0.87% |
| Outlying heterozygosity rate (>5SD from the mean) | 134 | 0.35% |
| Any of the above two | 451 | 0.95% |
| PCA outliers (6SD away from the mean in top 10 PCs) | 96 | 0.20% |
| Post-QC | 33,067 | 69.9% |
| Variant QC metric | # Variants | % Total |
|---|---|---|
| Initial variant count (avg. across batches) | 0.68M | - |
| Batch QC: | ||
| SNP call rate < 0.95 and then < 0.98 (avg. across batches) | 10K | - |
| Common across batches | 656,598 | - |
| Missing rate diff > 0.0075 between any two batches | 8,295 | - |
| Merged QC: | ||
| Total | 648,303 | 100% |
| Monomorphic SNPs | 14,753 | 2.28% |
| Duplicated SNPs | 427 | 0.07% |
| Not confidently mapped (chr & pos=0) | 0 | 0% |
| Any of the above three | 15,184 | 2.34% |
| EUR (pop-specific) QC: | 633,119 | 97.66% |
| Showing batch association (p < 1e-04) | 4,112 | 0.63% |
| Final QC: | ||
| SNP-level call rate < 0.98 | 14 | 2.16e-3% |
| pHWE < 1e-10 | 7,541 | 1.16% |
| Non-autosomal, indel, or monomorphic | 36,868 | 5.69% |
| Post-QC | 584,584 | 90.17% |
| HRC QC (for imputation): | ||
| Not in HRC or mismatched info | 41,944 | 6.47% |
| Send to Michigan imputation server | 542,640 | 83.7% |
| Variant filter | # Variants | % Total |
|---|---|---|
| Total imputed | 34,921,029 | 100% |
| MAF (dosage) < 0.005 or Imputation R2 < 0.6 | 26,240,369 | 75.1% |
| Call rate < 0.98 or pHWE < 1e-10 | 1,395 | 4.00e-3% |
| Post-QC | 8,679,265 | 24.9% |