Haplotype Copying In Imputation

Why Copying Haplotypes Can Fill in Untyped Markers

A study may have genotype calls at only some of the variants used in a reference panel. Imputation uses the observed calls and the panel's haplotype patterns to infer alleles at markers that were not typed. The central idea in haplotype-copying imputation is that a sample chromosome often resembles a stretch of one or more reference chromosomes. If a reference haplotype matches the sample at nearby typed markers, its allele at an untyped marker is a plausible candidate for the sample.

This is a segment-matching idea, not a claim that the sample chromosome is an exact copy of one complete reference chromosome. A chromosome can follow one reference haplotype in one region and another nearby. This accommodates the fact that ancestral chromosomes have recombined over generations. Earlier material on haplotypes and phase ambiguity and statistical phasing matters here: the copying procedure needs an ordering of alleles along each chromosome, and uncertain phase can change which strings are compared.

For clarity, represent each biallelic marker by 0 for one allele and 1 for the other. For example, the string \(0,1,1,0\) is one haplotype's alleles at four ordered markers. The coding is arbitrary, but the sample and reference must use the same allele coding and marker order. A diploid individual has two haplotypes; the discussion below follows one at a time. The two inferred haplotypes can then be combined to describe the individual's genotype, but the focus here is how copying suggests the chromosome-level alleles.

Matching and Reading an Untyped Allele

Suppose the sample's alleles are observed at markers 1, 2, and 4, while marker 3 is untyped. A reference haplotype matching the sample at the typed markers provides its allele at marker 3 as a candidate. A longer, locally distinctive pattern is generally more informative than a short match shared by many panel haplotypes. Matching is usually approximate: observed alleles can disagree with a candidate because of genotyping error, mutation, or the fact that the sample's best local template is not identical to any one panel haplotype.

Real methods use more than a raw count of matches. A model assigns a probability to observing each typed allele given a proposed copying state, and allows copying states to change along the chromosome. A likely path tends to use templates that explain the observed sequence while avoiding implausibly frequent switches. The method's result depends both on the observed matches and on the model's assumptions about error and switching.

A Simple Likelihood Comparison

A small likelihood calculation makes the matching principle precise. Consider two candidate copying states, A and B, and eight observed typed markers. Suppose, for this deliberately simplified example, that observations are conditionally independent given the candidate state. At four markers the probability of a match is 0.99 and the probability of a mismatch is 0.01. At the other four markers both candidates match, with match probability 0.98 at each. Assign prior probability \(1/2\) to each candidate before using these observations. The likelihood score below includes that prior weight; it is a weighted score for comparing candidates, not a genotype posterior.

Candidate A matches at all eight markers. Candidate B matches at two of the first four and mismatches at the other two, while both candidates match at the remaining four. Multiplying the prior weight and the marker-specific probabilities gives:

For candidate B, two of the first four observations are matches and two are mismatches:

The ratio of these weighted scores is \(L_B/L_A\approx 0.0001020\). Under these stated assumptions, the observed sequence is far more compatible with A than with B. This comparison does not establish that A is the true ancestral source. It only shows how a model can use the pattern of agreement, rather than the allele at any single marker, to favor one candidate copying state.

Copying Paths and Switches

A single reference haplotype may not match well across an entire region. A copying model can instead assign one state to a stretch of markers and another state to a later stretch. The point at which the state changes is a copying-state switch. Switches provide a statistical way to represent mosaic ancestry; they are not direct observations of a particular historical recombination event.

In a hidden Markov model formulation, the copying state is a hidden state that can persist between nearby markers or switch to another reference template. The observed typed alleles contribute emission probabilities: a match is generally more probable than a mismatch, but mismatches can be allowed. At an untyped marker there is no observed allele to score. Its absence contributes no match or mismatch evidence; the copying path instead provides the candidate allele from the reference state at that position. A target lying near an uncertain switch can therefore be less well resolved than one inside a long, consistent copied segment.

Assumptions, Checks, and Failure Modes

The copying interpretation depends on several conditions. The reference panel should contain haplotypes relevant to the study population, and the sample and panel markers must refer to the same genomic locations and allele coding. The sample's marker order and phase should be sufficiently reliable for meaningful haplotype comparison. The model must also represent genotyping errors and the tendency of copying states to persist or switch in a reasonable way. These are modeling assumptions, not guarantees supplied by the act of matching.

Useful diagnostics focus on whether the data support coherent local copying. Many mismatches across a region can indicate poor marker alignment, allele-coding errors, low-quality calls, an unrepresentative panel, or an unusual sample haplotype. A path that switches repeatedly over a short distance may reflect weak evidence, an overly permissive switching model, or problematic input rather than a sequence of biologically meaningful events. A long match shared by many reference haplotypes may also be uninformative: agreement is not strong evidence if the same pattern is common and does not distinguish candidate alleles.

A common mistake is to say, “The closest reference matches three of four typed alleles, so its allele at the missing marker must be correct.” The correction is that a best match is a model-based candidate, not a certainty. The comparison may change when mismatch probabilities, switch costs, other nearby markers, phase uncertainty, or panel composition are considered. A further mistake is to treat one whole reference haplotype as the source of the entire chromosome. Copying permits local changes of template, and a state switch can make the allele at a target depend on where that switch is placed.

Haplotype copying differs from exact-match lookup, which requires a complete matching string and can fail when there is even one discrepancy. It also differs from selecting a single nearest reference across the whole region, which ignores evidence that different segments may match different templates. A probabilistic copying model accommodates imperfect matches and local changes, but its flexibility does not remove the need for well-aligned data, adequate reference diversity, and sensible model assumptions. A reference panel cannot supply reliable information for an allele or haplotype pattern it does not represent.