Imputation Of Rare Variants

Why Rare Variants Need Special Attention

An imputed genotype is an inference, not a direct observation. For a biallelic variant, let \(G\) denote the number of copies of the allele being studied, so \(G\in\{0,1,2\}\). An imputed dosage is the conditional expected genotype given the available data and reference information: \(D=E(G\mid\text{observed data, reference})\). A dosage near zero can reflect a high probability of no copies, while a dosage between zero and two can reflect uncertainty across possible genotypes. Dosages and their distinction from hard calls are covered in the earlier material on dosage as the expected genotype and dosages versus hard calls.

Rare variants are challenging for two related but distinct reasons. First, a reference panel may contain very few copies of the allele, or none at all. The panel then provides little evidence about the haplotypes carrying it or its relationship to genotyped markers. Second, even when the variant is represented in the panel, the available markers may not predict its presence reliably in a target individual. These are different failures: sparse representation concerns the information available in the reference; low-quality prediction concerns how uncertain or inaccurate the resulting dosages are.

Imputation relies on reference haplotypes and the relationship between observed markers and the unobserved allele. Its performance therefore depends on assumptions that the target and reference data are correctly aligned, that the reference is relevant to the target population, and that the observed markers provide useful information about the allele. Earlier material on allele harmonization and genome build matching addresses two important prerequisites: the alleles and positions must refer to the same variant. Correct harmonization alone, however, cannot create information about an allele that the panel barely represents.

How Low Counts Limit Reference Information

A simple sampling calculation shows why frequency matters. Suppose a reference panel contains \(M\) sampled chromosomes from a population in which the allele frequency is \(q\). Under random sampling and a common allele frequency across those chromosomes, the count \(K\) of allele copies has a binomial distribution. This approximation treats sampled chromosomes as independent; relatedness or population structure can make the effective information smaller than the nominal chromosome count.

The expected count can be small even in a large panel when \(q\) is very small. Moreover, a frequency estimated from only a few copies is imprecise. Under the binomial model, the estimated frequency \(\hat{q}=K/M\) has standard error approximately \(\sqrt{q(1-q)/M}\), with the true \(q\) in this expression. If \(K=0\), the panel gives no observed copy of the allele; it does not prove the allele is absent from the target population. How a particular imputation method handles a variant absent from its reference depends on that method and its input data.

For a separate target-sample calculation, assume Hardy–Weinberg proportions at the variant and allele frequency \(q\). The probability that a person carries at least one copy is \(1-(1-q)^2=2q-q^2\). Thus a sample of \(n\) unrelated individuals has expected carrier count \(n(2q-q^2)\). For small \(q\), this is approximately \(2nq\). The approximation concerns the expected number of carriers; it does not guarantee that any particular sample will contain that many.

Diagnosing Frequency-Dependent Errors

Validation should compare imputed dosages with genotypes observed in a suitable validation set, such as a subset with direct, reliable genotyping or sequencing. The comparison must use the same variant and allele coding on both sides. Masking an observed genotype and then imputing it can assess performance under that masking design; it does not automatically guarantee the same performance for every population or for variants whose frequency differs from those in the validation set.

Summaries should be examined across frequency groups, not only averaged over all variants. Useful checks include allele frequency and allele count, dosage-based measures such as dosage \(R^2\), genotype or carrier-status sensitivity when a threshold is defined, and concordance measures interpreted according to exactly what was compared. The earlier material on imputation quality metrics, dosage \(R^2\), INFO scores, and validation by masking establishes these measures and their general use. For rare variants, inspect the number of validation allele copies and carriers underlying each estimate: a metric based on one or two observed copies can be highly variable. If there is no variation in true genotypes, a correlation-based dosage \(R^2\) is undefined, not zero evidence of good or bad performance.

A high overall agreement can also conceal missed carriers when noncarriers dominate. Carrier-status agreement is the proportion classified correctly as carrier or noncarrier under a stated dosage threshold. It is not the same as hard-genotype agreement, which requires comparing predicted and observed calls across genotype categories \(0\), \(1\), and \(2\). The threshold and classification target must be stated before interpreting carrier metrics.

Separate Sparse Representation From Uncertain Dosage

Reference allele count and prediction uncertainty answer different questions. A low count indicates that few reference examples carry the allele. A dosage is a prediction for a target individual or sample and reflects evidence available for that prediction. A variant can be sparse in the reference yet predict well if the few copies occur on a recognizable haplotype shared with the target. Conversely, a variant with many reference copies can predict poorly if its surrounding markers do not distinguish carriers from noncarriers in the target population.

For an individual with posterior genotype probabilities \(P(G=0\mid\text{data})\), \(P(G=1\mid\text{data})\), and \(P(G=2\mid\text{data})\), the dosage is their probability-weighted mean. The spread of those probabilities indicates uncertainty for that individual. A dosage is not itself a hard call, and an individual dosage alone cannot establish that the imputation method is well calibrated across a population.

Practical Comparisons, Diagnostics, and Limitations

When performance is poor, first ask whether the allele is represented and then whether the available data predict it. Low reference allele count, or no copies, points toward sparse representation. If the allele is present but masked validation shows low dosage agreement, low dosage \(R^2\), or many missed carriers, the predictions are not sufficiently informative in that validation setting. These observations can coexist. A large reference panel may improve the chance of observing rare alleles, but size alone does not guarantee that its ancestry, haplotypes, variant coverage, or marker relationships suit the target.

Frequency-stratified validation is generally more informative than a single summary over common and rare variants. A useful report gives the frequency bins, variant counts, allele or carrier counts where available, the metric and its definition, and the validation sample’s relationship to the target population. For a very rare variant, per-variant estimates may be too unstable to support a precise ranking. Aggregating results across a frequency bin can describe typical performance in that bin, but it cannot certify every variant in it.

A common alternative to dosage-based evaluation is hard-call or carrier-status evaluation. Hard calls are easier to interpret as discrete genotypes but discard uncertainty and depend on a calling rule; carrier classification at a chosen threshold answers a narrower question. Dosage-based metrics retain graded predictions, but can be unstable when observed genotypes have little variation and are not interchangeable with sensitivity or concordance. No single metric replaces inspection of variant counts, frequency, and the exact comparison being made.

A key limitation is that observed frequency and predictive performance depend on the population sampled. A variant rare in one group may be more common in another; a reference panel that poorly represents the target can produce frequency-dependent errors even when its total size is large. In addition, an apparent validation failure may arise from mismatched alleles or positions rather than from the imputation model itself. Check variant identity and validation design before attributing an error to rarity alone.

The practical lesson is to separate three claims: whether an allele is represented in the reference, whether its dosage predictions are informative, and how those predictions perform in a specified validation sample. State the evidence for each claim rather than treating “rare” as a complete diagnosis.