Haplotype Phasing

The specific assignment of alleles to their maternal and paternal chromosomes of origin. Phase resolves whether two heterozygous variants at different loci are located on the same chromosome (cis configuration) or on opposite homologous chromosomes (trans configuration). This information is lost in standard genotyping arrays and short-read sequencing, which report unphased genotypes. Computational phasing algorithms reconstruct this relationship using linkage disequilibrium patterns from reference panels or pedigree information.

Contiguous chromosomal segments inherited as a single unit due to low historical recombination rates. These blocks are defined by strong linkage disequilibrium between variants within the block and recombination hotspots at the boundaries. Key characteristics:

• Block lengths range from a few kilobases to over 100 kb
• Within a block, a small number of common haplotypes account for most of the population variation
• The HapMap Project and 1000 Genomes Project catalogued these structures across diverse populations
• Synthetic data generators must preserve block boundaries to maintain realistic recombination patterns

A quantitative metric measuring the accuracy of computational phasing by counting the frequency of incorrect transitions between parental haplotypes. A switch error occurs when the phasing algorithm incorrectly reassigns a segment from the maternal to the paternal chromosome. Long-range phasing accuracy is critical for:

• Identity-by-descent detection
• Accurate imputation of untyped variants
• Detecting compound heterozygous mutations in clinical genomics State-of-the-art methods achieve switch error rates below 0.5% when using large reference panels.

Large, pre-phased population datasets used as templates to infer missing genotypes and resolve phase in study samples. These panels leverage the observation that haplotypes are shared among individuals due to common ancestry. Critical reference resources include:

• 1000 Genomes Project: 2,504 individuals from 26 populations
• Haplotype Reference Consortium: 64,976 haplotypes with whole-genome sequencing
• TOPMed: Over 100,000 deeply sequenced genomes Synthetic data generators must replicate the haplotype diversity represented in these panels to produce realistic imputation benchmarks.

A direct phasing method that uses paired-end sequencing reads or linked-read technology to physically connect variants on the same DNA molecule. Unlike statistical phasing, which relies on population reference panels, read-backed phasing provides molecular evidence of cis relationships. Technologies enabling this include:

• 10x Genomics Linked-Reads: Barcoded short reads from high molecular weight DNA
• Hi-C sequencing: Chromatin conformation capture that links distant loci
• Long-read sequencing: PacBio HiFi and Oxford Nanopore reads spanning tens of kilobases Synthetic read generators must simulate these physical connections to benchmark phasing tools realistically.

Generative models must explicitly model the joint distribution of alleles along chromosomes rather than generating variants independently. Failure to preserve phase results in:

• Unrealistic haplotype structures that break linkage disequilibrium patterns
• Artificially inflated recombination rates in downstream analyses
• Erroneous compound heterozygote calls in clinical simulations Advanced approaches use haplotype-aware VAEs or conditional GANs that condition generation on phased reference panels to maintain chromosomal coherence in synthetic VCF outputs.

The non-random association of alleles at two or more loci. LD is the fundamental population-genetic signal that phasing algorithms exploit. It arises from mutation, genetic drift, and selection, and decays over generations due to recombination. High LD between a pair of SNPs means they are almost always inherited together, allowing one to serve as a tag for the other.

• Measured by statistics like D' and r².
• LD patterns form the basis of haplotype blocks.
• Phasing accuracy is directly proportional to the strength of LD in a population.

A null model stating that allele and genotype frequencies remain constant across generations in the absence of evolutionary forces. For phasing, HWE is a critical statistical prior. Algorithms like Expectation-Maximization (EM) use HWE assumptions to estimate haplotype frequencies from unphased genotype data. A deviation from HWE in a dataset can indicate genotyping errors or population structure that complicates phasing.

An iterative statistical method for finding maximum likelihood estimates of haplotype frequencies from unphased genotype data. The E-step computes the probability of each possible haplotype pair for an individual given current frequency estimates. The M-step re-estimates population haplotype frequencies by counting these expected pairings. EM is foundational but computationally intensive for large datasets.

• Handles phase ambiguity probabilistically.
• Assumes Hardy-Weinberg Equilibrium.
• Largely superseded by faster heuristic methods for genome-wide data.

A contiguous genomic region of high Linkage Disequilibrium and low recombination, typically spanning tens to hundreds of kilobases. Within a block, only a few common haplotypes exist. The HapMap Project cataloged these blocks, showing the genome is structured as a mosaic of discrete segments. Phasing algorithms leverage this block structure to simplify computation by solving the phase within each block independently.

The primary metric for evaluating phasing accuracy. A switch error occurs when the inferred haplotype phase incorrectly jumps between the maternal and paternal chromosomes at a recombination point. It is measured as the percentage of heterozygous positions where the phase assignment is flipped relative to the true, known phase. Long-range phasing methods aim to minimize switch errors to accurately reconstruct whole-chromosome haplotypes.

A genomic segment shared between two individuals because it was inherited from a common ancestor without recombination. IBD segments are the ultimate ground truth for phasing in population cohorts. If two individuals share a long IBD segment, they must share the same phased haplotype on that segment. Statistical phasing methods like Beagle and SHAPEIT use a Li-Stephens model to implicitly detect and leverage IBD for highly accurate inference.