SDPRX
Summary¶
SDPRX (Summary statistics-based Discovery and Prediction of Risk with Cross-population information) is a hierarchical Bayesian method designed for cross-population polygenic risk score (PRS) prediction. By integrating genome-wide association study (GWAS) summary statistics and linkage disequilibrium (LD) reference panels from two populations (typically European and a non-European cohort), SDPRX accounts for differences in LD structures and models shared versus population-specific genetic architectures to improve predictive accuracy in underrepresented populations.
Methodology¶
SDPRX is structured as a hierarchical Bayesian regression model consisting of two core components: 1. Likelihood Function: Models the relationship between marginal effect sizes from GWAS summary statistics and true underlying effect sizes. It uses a multivariate normal distribution to incorporate the LD covariance structure. To handle potential mismatches between GWAS summary cohorts and external reference panels, SDPRX applies a shrinkage factor to the off-diagonal covariance entries. 2. Prior Distribution: Employs a Bayesian nonparametric prior based on a Dirichlet process mixture model. Specifically, SNP effect sizes across the two populations are modeled using a four-component prior: - Null: Non-causal variants in both populations (Dirichlet delta distribution at zero). - Population 1 Specific: Non-zero effect only in population 1. - Population 2 Specific: Non-zero effect only in population 2. - Shared: Non-zero effects in both populations, allowing for cross-population correlation.
Parameter estimation is performed using a Markov Chain Monte Carlo (MCMC) algorithm with a truncated stick-breaking process, allowing SDPRX to adaptively learn the genetic architecture without needing independent tuning datasets.
Relationship to Other Methods¶
- SDPR: SDPRX is a direct cross-population extension of SDPR. While SDPR is restricted to single-population analysis, SDPRX models joint effect distributions across two populations.
- PleioSDPR: A subsequent expansion that applies similar Dirichlet process mixture priors to model shared genetic architectures between correlated traits (pleiotropy) rather than populations.