- We ran PHESANT on continuous (including biomarker) and categorical phenotypes. This performed inverse-rank normal transformation (IRNT) of all continuous traits within each ancestry group.
- We did not run PHESANT on ICD-10 codes or prescription data.
- We constructed Phecodes from ICD-9 and ICD-10 codes using our custom script based on the interim outputs of the createUKBphenome scripts. ICD-9 codes include primary codes (data field: 41203) and secondary codes (41205), and ICD-10 codes include primary codes (41202), secondary codes (41204), external codes (41201), and cause of death codes (40001). The mappings between phecodes and ICD codes are summarized here.
- For binary traits, we required at least 50 cases in each population except EUR given the larger sample size. In EUR, we required at least 100 cases.
- Prescription data was extracted from UKB data field 42039 and processed through a custom pipeline to harmonize data.
- Special treatment of some phenotypes was applied, including for those related to COVID-19 as well as custom phenotype combinations such as waist-hip ratio (phenotype computed from UKB code 48 / 49, then IRNT).
Pre-QC phenotype counts are broken down by ancestries as follows:
- We removed participants flagged with
putative.sex.chromosome.aneuploidy == 1
We first combined reference data from the 1000 Genomes Project and Human Genome Diversity Panel (HGDP). Briefly, we used 1000 Genomes phase 3 sequenced data and publicly available genotyping data from HGDP. We combined these reference datasets into continental ancestries according to their corresponding meta-data. Details of meta-data grouping combining these two datasets are available here.
We then used a two-stage approach to:
- assign continental ancestries, and then
- prune ancestry outliers within continental groups.
First, we ran PCA on unrelated individuals from the 1000 Genomes + HGDP combined reference dataset. To partition individuals in the UK Biobank based on their continental ancestry, we used the PC loadings from the reference dataset to project UK Biobank individuals into the same PC space. We trained a random forest classifier given continental ancestry meta-data based on top 6 PCs from the reference training data. (We also assessed whether including more PCs -- up to 20 -- changed the ancestry assignments but identified no change, so we continued with the assignments from the top 6 PCs). We applied this random forest to the projected UK Biobank PCA data and assigned initial ancestries that were subsequently refined if the random forest probability was >50%. Other individuals with a probability < 50% for any given ancestry group were dropped from further analysis.
Second, we refined initial ancestry assignments by pruning outliers within each continental assignment. We started by rerunning PCA among UK Biobank individuals within each assigned continental ancestry group (i.e. excluding reference panel data). We calculated the total distance from population centroids across 10 PCs. Using the PC scores, we computed centroid distances across 3-5 centroids spanning these PCs depending on the degree of heterogeneity within each continental ancestry as follows:
For each individual and for each ellipse across 10 PCs, subtract the population PC mean from the individual's PC, square this value, and divide by the variance of the PC.
We identified ancestry outliers based by plotting histograms of centroid distances and removing those individuals from the extreme high end of the distribution.
Within each population, we ran PC-Relate with
min_individual_maf=0.05. To get the maximal set of unrelateds, we then ran
hl.maximal_independent_set using Hail. We used these unrelated individuals to define the ancestry-specific PC space, then projected related individuals into the same PC space for use as covariates in SAIGE.
- We used imputed variants from the UK Biobank (97,059,328 in version 3).
- We retained variants with INFO scores > 0.8 (29,865,259 variants on the autosomes and X chromosome). - For each population, we retained variants with an allele count of at least 20, computing using the genotypes where available or imputed dosages.
- We created a low-confidence filter for those variants with a minor allele count less than 20.
We ran GWAS for all phenotypes and ancestry groups using the Scalable and Accurate Implementation of GEneralized mixed model package, SAIGE, which runs a linear or logistic mixed model including a kinship matrix as a random effect and covariates as fixed effects. This approach was chosen due to its suitability for a wide range of GWAS scenarios, its ability to account for more heterogeneity in sample makeup, and its computational scalability. Specifically, it provides accurate p values even when case-control ratios are extremely imbalanced by using the saddlepoint approximation to calibrate the distribution of score test statistics.
For each GWAS conducted for each phenotype and ancestry group, we included the following covariates:
- Age * Sex
- Age2 * Sex
- The first 10 PCs
We have recently developed a best practice set of quality control steps for analysis of summary statistics from a wide range of phenotypes using the following sequential filters:
GWAS_run: if the GWAS was performed for the ancestry-trait pair
ancestry_reasonable_n: if the ancestry has a reasonable sample size; has the effect of removing AMR
defined_h2: if the heritability estimate is non-missing
significant_z: if the ancestry-trait pair shows z-score
in_bounds_h2: if, for all ancestries for a given trait, observed-scale heritability estimates
normal_lambda: if, for the top three best powered ancestry groups (EUR, CSA, AFR),
normal_ratio: if, for the top three best powered ancestry groups (EUR, CSA, AFR), the S-LDSC ratio, given by , or the ratio z-score
EUR_plus_1: if the trait passes all above filters in EUR and at least 1 other ancestry group
More detailed documentation on these filters is forthcoming. The number of traits passing all QC per-ancestry are:
As with any analysis of large-scale datasets, we urge caution in interpretation especially of outliers (i.e. genetic variants, samples, and/or whole phenotypes) that can signal technical artifacts or noise.
- We have recently released new heritability estimates using summary statistics with LD-score regression and stratified LD-score regression as well as genotype-level data using a randomized Haseman-Elston regression estimator (RHEmc). More information can be found here.
- We note that LDSC and S-LDSC are produce very noisy estimates in non-EUR ancestry groups likely due to small sample sizes.
- RHEmc produces heritability estimates much more significantly different from 0 than LDSC/S-LDSC in non-EUR groups (see post), however we caution that the impact of stratification has yet to be fully explored on these estimates (as PCs are included only as fixed effects) and these may have a differential impact on different ancestry groups. However, these estimates have shown promise in summary statistics QC (see Quality control of summary statistics).
- Due to computational costs we only produce RHEmc EUR estimates for a subset of traits as a pilot, finding concordance with S-LDSC run on summary statistics from the same traits.
- The heritability computed by SAIGE is known to be downwardly biased: see the supplementary information from the SAIGE publication.