The problem. A polygenic risk score sums the effect of many variants to predict a person’s liability for a trait. The simplest version takes the GWAS effect sizes, thresholds by significance, and adds them up. That mishandles linkage disequilibrium: nearby variants are correlated, so their effects get counted several times, and the arbitrary significance cutoff throws away real signal sitting just below the line.

The idea. LDpred puts a Bayesian model over the true effect sizes and corrects the observed GWAS effects using a reference panel’s linkage structure. It assumes a prior in which some fraction of variants are causal, then updates each effect given its neighbors’ correlations, shrinking and reweighting so correlated variants are not double-counted. The result is a set of adjusted weights that make a more accurate score, and it uses summary statistics plus a reference panel rather than raw genotypes.

Why it matters. Polygenic scores are how GWAS results reach an individual prediction, and their accuracy hinges on handling linkage and the many sub-threshold effects that GCTA showed carry most of the heritability. The reference-panel dependence is the same reliance on population resources like 1000 Genomes that runs through this whole area.

Verdict. A key step in making polygenic scores usable. Read it for why modeling linkage, not just thresholding, is what raises accuracy.