BridgePRS is a trans-ancestry PRS software which improves polygenic risk score analysis in under-represented target populations by combining summary statistics from under-represented target populations (e.g. Africans) with those from a powerful GWAS (e.g. Europeans).
Background
PRS methods require as input summary statistics from genome wide association studies. If unfamiliar with GWAS or in need of a refresher consider reading this paper.
- Genome-wide association studies (GWAS) involve analyzing the genomes of a large group of individuals.
- For a binary trait (like blue or brown eyes) this involves comparing the frequency of genetic variations in each group to produce an odds-ratio (measure of association) and a p-value to measure the signficance of the realtionship at each SNP.
- For continuous measures (like height) this results in an effect size (measure of continuous association) and p-value that measure the significance of our relationship.
- GWAS results can be summarized in a sumstats file which looks like this:
| CHR | ID | REF | A1 | A1_FREQ | OBS_CT | SE | BETA | P |
|---|---|---|---|---|---|---|---|---|
| 1 | rs100 | A | T | 0.1 | 100 | 0.01 | 1.10 | 0.01 |
| 2 | rs200 | C | G | 0.2 | 100 | 0.02 | -1.10 | 0.05 |
| 3 | rs300 | G | A | 0.3 | 100 | 0.03 | 1.02 | 0.10 |
PRS basics
- Polygenic scores combine genetic variant associations genome-wide by summing their effects to produce individual scores that predict the phenotype.
- Different PRS methods use different statistical methods to select variants and estimate their effect sizes for use in the PRS.
- BridgePRS, like many other PRS methods, requires genotype and phenotype data from individuals (several hundred samples) in addition to GWAS summary statistics to estimate the PRS.
- The predictive power of PRS is assessed in genotype and phenotype data from unseen samples. BridgePRS reports PRS accuracy by the residual variance explained \(R^2\) (accounting for the variance explained by the non-genetic covariates included in the model). For binary traits Nagelkerke \(R^2\) is used.
Cross Population Analysis
The PRS Portability Problem
- Often when a PRS trained using data from one population, e.g. European, is applied to a population of a different ancestry, e.g. African, the PRS model is less predictive.
- Often there is insufficient data in the non-European population to
estimate good ancestry specific PRS using this data alone. Using
single ancestry PRS methods researchers are left with two choices:
- An underpowered PRS model estimated using only data the target (non-European) population.
- A PRS estimated using a well powered GWAS from Europeans
The BridgePRS Solution
- BridgePRS combines GWAS summary statistics from two populations, typically a large well powered GWAS from a "base" population (e.g. European) and a smaller GWAS from the target population (e.g. African).
- Figure 1 of our Nature Genetics Paper gives an overview of the modelling implemented by BridgePRS.
- Figure 1 shows the three models, M1-3, estimated by BridgePRS which are subsequently combined to give the final PRS. Below we outline these three models
- Model M1: BridgePRS uses two prior distributions to estimate PRS which we describe as Stage 1 and Stage 2. Stage 1 analysis is a single population analysis and applies zero centred Gaussian prior distributions to SNP effect sizes within clumps. Stage 2 uses the output from a Stage 1 analysis of the base population as the prior for the target population.
- Model M2: This is a single population analysis using Stage 1 analyis applied to the target populationto produce two sets of SNP weights
- Model M3: M1 and M2 both produce PRS estimated under different prior parameter settings. These PRS are combined in a ridge regression fit to the target data to give models M1 and M2. Model M3 is estimated by combining both sets of PRS in a ridge regression fit.
- Model M1 reflects the belief that the target population GWAS is only informative in conjugtion with the base population GWAS.
- Model M2 reflects the belief that the target population GWAS is informative and the base population GWAS gives no addition information.
- Models M3 reflects the belief both the base and target population GWAS contribute independent information.
- Since we do not know apriori which of the three scenarios (M1, M2, or M3) are true, BridgePRS weights and sums all three models to produce and evaluate a single target population PRS. BridgePRS also reports SNP weights and PRS R2 values in the target population for each of the three other models (M1, M2, or M3), enabling the user to use any of the four models.
- Further detals outlining this modelling is shown in Figure 1 of our Nature Genetics Paper