Linkage Disequilibrium

Linkage disequilibrium occurs when alleles at two or more genomic loci are inherited together more frequently than expected by chance under independent assortment. This non-random association arises from physical proximity on the same chromosome, population bottlenecks, natural selection, or recent admixture. LD is quantified using metrics such as D' (Lewontin's D-prime) and r² (the squared correlation coefficient), which measure the statistical deviation from linkage equilibrium. In synthetic genomic data generation, failing to model LD results in sequences that lack realistic haplotype structure, producing artificial recombination breakpoints that degrade the utility of the data for downstream analyses like genome-wide association studies.

The rate at which LD decays with increasing genomic distance is a signature of a population's demographic history. In populations with large effective population sizes and ancient origins, LD decays rapidly over short distances (e.g., African populations). In populations that experienced recent bottlenecks or founder effects, LD extends over much longer genomic intervals (e.g., European and East Asian populations). Synthetic genomic data generators must accurately model this population-specific LD decay curve to produce realistic sequences. A generator that produces uniform LD patterns across populations will fail validation tests and introduce biases in simulated cohort studies.

The genome is organized into discrete haplotype blocks—segments of high LD separated by recombination hotspots. Within each block, a limited number of common haplotypes account for most of the genetic variation in a population. This block-like structure is a direct consequence of the non-uniform distribution of meiotic recombination events. Synthetic data generators must preserve these block boundaries and the haplotype diversity within each block. Failure to do so produces sequences with unrealistic combinations of alleles that would never co-occur in nature, a critical flaw detectable by tools like Haploview or PLINK.

Preserving LD is a primary quality metric for synthetic genomic data. Validation approaches include:

• LD decay curve comparison: Overlaying the r² vs. distance curves of real and synthetic data to ensure they match across all genomic scales.
• Haplotype block concordance: Verifying that synthetic data reproduces the same block boundaries and haplotype frequencies as the training population.
• Principal component analysis (PCA): Projecting synthetic samples onto real population PCA plots to confirm they cluster appropriately without creating artificial subpopulations.
• Adversarial validation: Training a classifier to distinguish real from synthetic haplotypes; high discriminability indicates poor LD preservation.

When synthetic genomic data generators fail to preserve LD, several downstream artifacts emerge:

• Inflated false-positive rates in GWAS due to spurious associations between unlinked causal variants and tag SNPs.
• Artificially deflated p-values in fine-mapping studies because the correlation structure among variants is broken.
• Unrealistic polygenic risk scores that overestimate or underestimate genetic liability when applied to real cohorts.
• Broken epistatic interactions where synthetic sequences fail to capture the joint effects of variants at interacting loci. These failures render synthetic data unsuitable for its primary use case: serving as a realistic proxy for real genomic data in method development and benchmarking.

Different generative model classes handle LD with varying fidelity:

• VAEs with structured latent spaces can capture global LD patterns but often smooth over fine-scale haplotype structure due to the Gaussian prior.
• GANs, particularly WGAN-GP variants, can learn sharp LD patterns but are prone to mode collapse that eliminates rare haplotypes.
• Autoregressive models (e.g., DNA language models) that generate sequences base-by-base conditioned on preceding context can implicitly learn LD through attention mechanisms.
• Haplotype-aware architectures that explicitly model phased diploid genomes as paired sequences with recombination-aware loss functions provide the most faithful LD preservation.

The computational process of determining which genetic variants are inherited together on the same parental chromosome. Synthetic data generators must accurately model haplotype structure to preserve linkage disequilibrium patterns.

• Resolves diploid genotypes into maternal and paternal haplotypes
• Critical for imputing missing genotypes in genome-wide association studies
• Statistical phasing uses Hidden Markov Models and reference panels
• Experimental phasing uses long-read sequencing or Hi-C data

A fundamental population genetics principle stating that allele and genotype frequencies remain constant across generations in the absence of evolutionary forces. Synthetic population data must satisfy HWE as a validation checkpoint.

• Null model for detecting genotyping errors
• Deviations indicate selection, drift, or population stratification
• Tested using chi-square goodness-of-fit on observed vs expected genotypes
• Serves as a quality control metric for synthetic cohort generation

The proportion of sequencing reads or alleles in a population that carry a specific genetic variant at a given locus. Synthetic genomic data must accurately reflect VAF distributions to enable realistic cohort simulations.

• Calculated as alternate allele count divided by total allele count
• Low-frequency variants (<1%) are critical for rare disease modeling
• VAF spectra differ across populations and genomic ancestries
• Generative models must preserve the site frequency spectrum

The occurrence count of all possible subsequences of length k within a genome. Preserving k-mer distributions is essential for synthetic generators to maintain biological plausibility and avoid artifacts.

• k=3 captures codon usage patterns in coding regions
• k=6-8 captures transcription factor binding site motifs
• Deviations in k-mer spectra indicate mode collapse in GAN training
• Used as a lightweight evaluation metric for synthetic sequence quality

A quantitative metric for evaluating synthetic genomic data quality by comparing the distribution of generated sequences to real sequences in a learned feature space. Analogous to FID in computer vision.

• Lower FGD scores indicate higher fidelity synthetic data
• Computed using embeddings from a pretrained genomic model
• Captures both diversity and realism in a single scalar value
• Sensitive to mode collapse and distributional shifts

An evaluation paradigm where a predictive model is trained exclusively on synthetic genomic data and tested on real held-out data. Measures the utility of synthetic data by its ability to substitute for real data in downstream tasks.

• If TSTR performance ≈ Train-Real-Test-Real, synthetic data is high-quality
• Commonly used for variant calling and gene expression prediction benchmarks
• Complements distributional metrics with task-specific utility assessment
• Gold standard for validating generative model outputs