SNP Heritability

SNP heritability specifically estimates the additive component of genetic variance—the sum of independent effects from individual variants. This excludes dominance effects (interactions between alleles at the same locus) and epistatic effects (interactions between different loci). The additive assumption underlies most PRS methods because additive effects are directly transmissible from parents to offspring and are the primary drivers of prediction accuracy in linear models.

SNP heritability is not a fixed biological constant but a population-specific estimate that depends on:

• Allele frequencies in the study cohort
• Linkage disequilibrium patterns between genotyped and causal variants
• Environmental variance in the sampled population

A trait may show h²_SNP = 0.30 in a European cohort but 0.15 in an African-ancestry cohort due to differing LD structures and allele frequency spectra, even if the underlying causal variants are identical.

Narrow-sense heritability (h²) represents the total additive genetic variance, including effects from all causal variants—both genotyped and ungenotyped. SNP heritability (h²_SNP) captures only the variance attributable to variants directly measured or well-tagged on genotyping arrays. The gap between h² and h²_SNP is the missing heritability, arising from:

• Rare variants not captured by standard arrays
• Structural variants poorly tagged by SNPs
• Incomplete LD between genotyped markers and causal variants

LD Score Regression (LDSC) is the dominant method for estimating SNP heritability from GWAS summary statistics. It exploits the relationship: the expected χ² statistic for a variant equals 1 + (N × h²_SNP × l_j / M), where l_j is the variant's LD Score (sum of squared correlations with neighboring variants). This approach:

• Corrects for population stratification by distinguishing polygenic signal from confounding
• Requires only summary statistics, not individual-level data
• Provides intercept estimates that quantify residual confounding

SNP heritability defines the theoretical maximum R² achievable by a polygenic risk score constructed from the same marker set. In practice, PRS accuracy falls below this ceiling due to:

• Finite GWAS sample sizes producing noisy effect size estimates
• Imperfect LD tagging between markers and causal variants
• Winner's curse inflating effect estimates in discovery cohorts

The ratio of observed PRS R² to h²_SNP quantifies how efficiently a method extracts the available genetic signal.

Stratified LD Score Regression extends heritability estimation by partitioning h²_SNP across functional genomic annotations. This reveals enrichment patterns:

• Coding regions often show disproportionate heritability relative to their genomic proportion
• Conserved non-coding elements harbor significant trait-associated variance
• Cell-type-specific regulatory elements (e.g., DNase hypersensitivity sites) can localize heritability to relevant tissues

These partitions inform fine-mapping efforts and validate the biological relevance of GWAS loci.

The foundational discovery method that scans millions of single nucleotide polymorphisms across the genome to identify genotype-phenotype associations. GWAS summary statistics—including effect sizes, standard errors, and p-values for each variant—serve as the primary input for estimating SNP heritability via methods like LD Score Regression. A well-powered GWAS with adequate sample size is essential for producing reliable heritability estimates.

A technique that leverages the correlation structure between genetic variants to estimate SNP heritability from GWAS summary statistics alone, without requiring individual-level data. LD Score Regression partitions the inflation in test statistics into polygenic signal versus population stratification confounding. Key outputs include the genomic inflation factor and estimates of heritability attributable to all measured SNPs.

Systematic differences in allele frequencies and disease prevalence between subpopulations due to ancestry, which can inflate SNP heritability estimates if not properly corrected. Methods like Principal Component Analysis (PCA) model continuous axes of genetic ancestry variation and are routinely included as covariates in GWAS to mitigate this confounding. Residual stratification remains a concern in cross-ancestry analyses.

The comprehensive characterization of the number, frequency, and effect size distribution of causal variants underlying a complex trait. Genetic architecture directly determines SNP heritability: traits influenced by many common variants of small effect yield different estimation challenges than those shaped by rare variants of large effect. Understanding this architecture informs the choice between methods like LDpred2 and PRS-CS for downstream prediction.

The proportion of phenotypic variance in a target dataset accounted for by all measured SNPs, quantifying the model's overall predictive power. In SNP heritability contexts, R² is estimated on the liability scale for binary disease traits using the liability threshold model, which assumes a continuous, unobserved liability distribution underlying case-control status. This transformation is critical for comparing heritability estimates across traits with different population prevalences.

SNP heritability estimates derived from European-ancestry cohorts often fail to replicate in diverse global populations due to differences in LD patterns, allele frequencies, and gene-environment interactions. Cross-ancestry heritability analysis requires GWAS from multiple populations and methods that account for heterogeneous genetic architectures. Poor transferability highlights the urgent need for inclusive biobank recruitment to ensure equitable polygenic prediction.