Variant Allele Frequency

VAF is calculated as the ratio of alternate allele reads to total reads at a genomic locus. For germline variants in a diploid genome, expected VAFs are approximately 0% (homozygous reference), 50% (heterozygous), or 100% (homozygous alternate). Somatic variants in tumor samples exhibit continuous VAF distributions due to tumor purity and clonal heterogeneity.

• Formula: VAF = (Alternate Allele Depth) / (Total Depth)
• Germline heterozygous expectation: ~0.5
• Somatic VAF range: 0.01 to 1.0 depending on clonality

Synthetic genomic data generators must preserve realistic VAF distributions to ensure downstream analyses remain valid. Generative models like VAEs and GANs must learn the empirical allele frequency spectrum from training populations and reproduce it without introducing allele frequency bias.

• Synthetic VCF files must match expected site frequency spectra
• Failure to model VAF accurately leads to spurious association signals
• TSTR evaluation validates that synthetic VAF distributions mirror real cohorts

Hardy-Weinberg Equilibrium (HWE) serves as a null model for validating synthetic population genomics data. Under HWE, genotype frequencies are predictable from allele frequencies: p² + 2pq + q² = 1. Synthetic data generators must produce variant calls where observed heterozygosity matches expected heterozygosity.

• HWE deviation indicates genotyping errors or population structure
• Synthetic cohorts should pass HWE tests for randomly mating populations
• Chi-squared goodness-of-fit tests quantify HWE adherence

The Site Frequency Spectrum (SFS) is the distribution of allele frequencies across all variant sites in a population. It captures the proportion of variants present at 1, 2, 3, ..., n copies in a sample of n chromosomes. Synthetic data must reproduce the expected SFS shape, which is influenced by demographic history, selection, and mutation rates.

• Rare variants (low VAF) dominate the SFS in human populations
• SFS distortion indicates model failure or privacy leakage
• Watterson's estimator uses SFS to estimate population mutation rate

Linkage Disequilibrium (LD) describes the non-random association of alleles at different loci. VAF at one site is often correlated with VAF at nearby sites due to shared haplotype backgrounds. Synthetic data generators must preserve LD decay patterns across genomic distances to maintain realistic population structure.

• LD is measured by metrics like r² and D'
• LD blocks define haplotype structures inherited together
• Synthetic genomes with broken LD produce inflated false-positive associations

Differential privacy mechanisms add calibrated noise to VAF estimates to prevent membership inference attacks. The privacy budget (epsilon) controls the trade-off between VAF accuracy and individual privacy guarantees. Lower epsilon values provide stronger privacy but may distort rare variant frequencies.

• Laplace noise is commonly added to allele counts
• Rare variants (low VAF) are most vulnerable to re-identification
• Privacy-preserving synthetic data must balance utility and protection

A foundational population genetics principle stating that allele and genotype frequencies remain constant across generations in the absence of evolutionary forces. For synthetic data generation, Hardy-Weinberg Equilibrium serves as a null model: a synthetic cohort's observed VAF distribution should conform to expected genotype proportions (p², 2pq, q²) unless specific selection or drift effects are intentionally simulated. Deviations from equilibrium in synthetic data often indicate bugs in the generative model's allele sampling logic.

The non-random association of alleles at different loci, measured as the difference between observed and expected haplotype frequencies. Synthetic genomic data generators must preserve linkage disequilibrium patterns to produce realistic VAF correlations across adjacent variants. Failure to model LD results in synthetic cohorts where allele frequencies are statistically independent, breaking the haplotype block structures critical for genome-wide association studies and population stratification analyses.

An artificially generated Variant Call Format file containing simulated SNPs, insertions, and deletions with realistic VAF distributions. A high-quality Synthetic VCF must accurately reflect:

• Allele frequencies matching the source population's minor allele frequency spectrum
• Genotype likelihoods consistent with sequencing depth and error profiles
• Site-level quality metrics indistinguishable from real variant calls These files are essential for benchmarking variant callers without accessing restricted clinical data.

A quality assessment technique where a classifier is trained to distinguish real from synthetic genomic data. For VAF evaluation, adversarial validation tests whether the distribution of allele frequencies in synthetic data contains detectable artifacts. A generator passes this test when the discriminator achieves near-chance accuracy (AUC ≈ 0.5), indicating that the synthetic VAF spectrum is statistically indistinguishable from the real population distribution.

A mathematical framework that injects calibrated noise into generative model training to prevent individual re-identification. When applied to VAF preservation, differential privacy introduces a privacy budget (epsilon) that bounds how much any single individual's variants can influence the synthetic allele frequency distribution. Lower epsilon values provide stronger privacy guarantees but may distort rare variant frequencies, creating a direct trade-off between VAF fidelity and participant protection.

An evaluation paradigm that measures synthetic data utility by training a predictive model on artificial data and testing on real data. For VAF-dependent tasks, TSTR assesses whether a model trained on synthetic allele frequencies can accurately predict phenotypes, disease risk, or population structure in genuine cohorts. High TSTR performance indicates that the synthetic VAF distribution captures the biological signal necessary for downstream machine learning applications.