Genetic Architecture

The total count of independent genetic loci that exert a non-zero effect on a complex trait. Polygenicity quantifies how many distinct regions of the genome contribute to phenotypic variance.

• High polygenicity: Traits influenced by thousands to tens of thousands of variants (e.g., height, schizophrenia)
• Low polygenicity: Traits dominated by a handful of moderate-to-large effect loci (e.g., age-related macular degeneration)
• Directly determines whether sparse methods like LASSO or dense methods like LDpred2 are more appropriate
• Estimated from GWAS using techniques like LD Score regression or BayesS

The statistical distribution characterizing the magnitude of each causal variant's contribution to the phenotype. This distribution fundamentally shapes which PRS method will achieve optimal prediction.

• Infinitesimal model: All variants have tiny, non-zero effects — best captured by Bayesian shrinkage methods like PRS-CS
• Non-infinitesimal (sparse): A mixture of null and larger-effect variants — suited for clumping and thresholding or LDpred2 with a point-normal prior
• Negative selection: Variants with larger effects tend to be rarer due to purifying selection, creating an inverse relationship between minor allele frequency (MAF) and effect size
• The SNP heritability attributable to each MAF bin reveals the trait's allelic architecture

The distribution of causal variant frequencies across the allele frequency spectrum, from rare (MAF < 1%) to common (MAF > 5%). This spectrum determines the type of genotyping or sequencing data required for adequate PRS performance.

• Common variant architecture: Most heritability concentrated in variants with MAF > 5%, well-tagged by standard GWAS arrays and imputation panels
• Rare variant architecture: Substantial heritability from low-frequency (0.5% < MAF < 5%) or rare (MAF < 0.5%) variants, requiring whole-exome or whole-genome sequencing
• Allelic heterogeneity: Multiple rare variants at the same gene can collectively mimic a common variant signal in GWAS
• The MAF-effect size relationship informs whether imputation accuracy at rare variants will limit PRS transferability

The complex correlation structure between genetic variants arising from shared evolutionary history. LD patterns dictate how GWAS signals are distributed across correlated markers and how PRS methods must model or prune these dependencies.

• LD blocks: Regions of high correlation where a single GWAS hit may tag multiple causal variants, complicating fine-mapping
• LD decay: The rate at which correlation diminishes with physical distance, varying dramatically across ancestral populations
• LD reference panels (e.g., 1000 Genomes, UK Biobank) are essential for Bayesian PRS methods to disentangle true effect sizes from tagging effects
• Cross-ancestry LD differences are a primary driver of PRS portability loss when applying European-derived scores to non-European populations

The decomposition of total SNP heritability into functional genomic categories, revealing which biological annotations are enriched for causal variants. This informs biologically-informed PRS methods that upweight variants in relevant annotations.

• Functional annotations: Coding regions, promoters, enhancers, conserved elements, and cell-type-specific regulatory elements
• Stratified LD Score regression (S-LDSC) estimates enrichment of heritability in each annotation category
• Partitioned heritability reveals that disease-relevant cell types and tissues harbor disproportionately high concentrations of causal variants
• Methods like AnnoPred and LDpred-funct incorporate these functional priors to improve PRS accuracy by leveraging biological knowledge

The extent to which the same causal variants influence multiple traits, quantified by the genetic correlation (r_g). Pleiotropic architectures have profound implications for PRS specificity and multi-trait prediction strategies.

• Positive genetic correlation: Shared genetic basis between traits (e.g., LDL cholesterol and coronary artery disease, r_g ≈ 0.8)
• Negative genetic correlation: Variants that increase risk for one trait decrease risk for another (e.g., anorexia and obesity)
• Multi-trait PRS methods like MTAG and wMT-SBLUP leverage genetic correlations to improve prediction accuracy for each trait
• Pleiotropy can inflate PRS associations with non-target outcomes, requiring careful interpretation in clinical settings
• Estimated from GWAS summary statistics using LD Score regression or Genomic SEM