A statistical review of polygenic risk scores: from heuristics to model-based inference

Polygenic risk scores (PRS) were initially developed as pragmatic tools to aggregate genome-wide association study (GWAS) signals for individual-level prediction, relying on heuristic strategies such as clumping and thresholding to approximate independence among variants. Although computationally efficient and widely accessible, these early approaches were sensitive to tuning parameters and limited in their ability to capture the diffuse signal characteristic of highly polygenic traits. As GWAS sample sizes expanded and biobank-scale resources emerged, methodological priorities shifted toward statistically principled models that explicitly represent linkage disequilibrium, effect-size heterogeneity, and population structure. In this review, we examine the methodological evolution of PRS construction from threshold-based aggregation to fully model-based inference frameworks, including linear mixed models, LD-aware Bayesian shrinkage approaches, machine learning, and recent multi-ancestry extensions, and summarize practical considerations for method selection under different data-access, LD-reference, tuning, and ancestry settings. Collectively, these developments mark a transition from heuristic scoring algorithms to a mature, statistically grounded paradigm for genomic risk prediction.