Reference Databases for Rare-Variant Interpretation

Why a Reference Database Needs Context

A reference database can help assess how often a genetic variant has been observed in sampled people. But its value is not determined by the database’s total number of participants alone. The relevant questions include how many chromosomes were actually assessed at the site, whether the sampled people represent the population of interest, and whether the data are sufficiently reliable for the type of variant being considered.

In particular, “not listed” can have several meanings. A variant may have been assessed and observed zero times; the site may not have been adequately assessed; or the database may not report variants that fail its inclusion or quality rules. Those situations provide different evidence. An absent record is not, by itself, proof that a variant is exceptionally rare.

For a diploid autosomal site, a fully assessed individual can contribute two chromosome copies, but missing or unusable data can reduce the number assessed. Thus, participant count and AN are not interchangeable. Frequency estimates should be considered within the relevant ancestry group when possible, not only in a pooled sample whose composition may differ from the population of interest.

Read the Denominator Before the Frequency

The observed frequency is meaningful only alongside its denominator. A count of one allele copy among 100 assessed chromosomes gives a different estimate, and much more sampling uncertainty, than a count of one among 100,000. The key quantity is the number of relevant chromosome copies successfully assessed, not just the headline number of database participants.

Population-specific frequency matters for the same reason. If an allele occurs in one ancestry group but is uncommon in a much larger group, pooling those groups can make the pooled frequency a poor summary for either group. Ancestry categories in databases are approximations, and they may not capture all relevant population structure. A frequency estimate should therefore be matched as closely as possible to the population about which an interpretation is being made.

What a Zero Count Can Establish

A reported zero allele count can still be informative if the database provides a credible denominator for the relevant site and population. Under a simple sampling model, suppose \(N\) independent chromosome copies are assessed and each has probability \(p\) of carrying the allele. The number of observed copies follows a binomial model. The probability of seeing zero copies is \( (1-p)^N \). A zero count is more surprising when \(N\) is large than when \(N\) is small, if the true frequency is held fixed.

A useful way to express the limit of a zero count is a one-sided 95% upper bound. It is the frequency \(p_U\) for which the probability of observing zero copies is 0.05:

For large \(N\), this is approximately \(3/N\). The bound is not an estimate that the frequency equals \(3/N\); it indicates that, under the model, frequencies substantially above that value would make a zero count unlikely. This calculation assumes independent sampled chromosome copies and a common frequency \(p\) among them. It does not fix a mismatch between the database sample and the target population, or compensate for a site that was not adequately assessed.

Ancestry Representation and Pooled Summaries

A large total database can still offer limited evidence for a particular ancestry group if relatively few people from that group were included or if relatively few chromosome copies were assessable. A pooled frequency can conceal this imbalance. For interpretation, examine both the overall summary and the ancestry-stratified counts and denominators, while recognizing that ancestry labels are imperfect proxies for genetic and population history.

Assess Data Quality and Database Scope

A database frequency is useful only when the allele and its denominator are credible. Before treating a reported count or absence as evidence, consider whether the database’s scope and data-generation procedures fit the question. Quality is not a single yes-or-no property: performance can differ by genomic region, variant type, sequencing methods, and filtering rules.

• Site assessment: Determine whether the site was assessed in the relevant samples. An AN of zero or a missing denominator is not evidence for a zero frequency. A reported zero count is informative only when an adequate number of copies was actually assessed.
• Variant type and region: Some variant classes or difficult genomic regions may be less reliably represented by a resource’s methods. A database designed primarily for one data type may not provide equally strong evidence for every other type.
• Quality and inclusion rules: Resources may exclude low-confidence observations or omit variants that do not meet reporting thresholds. A missing entry can therefore reflect filtering or reporting policy, not necessarily a verified absence in all sampled individuals.
• Consistent variant representation: Confirm that the database record refers to the same genomic position and alternate allele as the variant being interpreted. Differences in genome build, allele alignment, or representation of complex variants can make matching records difficult.
• Sample composition: Consider how participants were selected and whether their ancestry and other characteristics are relevant. A database sample is not automatically a random, representative sample of every population.

These considerations do not require treating every database result as unusable. Rather, they determine how much weight the result can bear. A well-supported count in a relevant group can inform frequency assessment; a missing record without a reliable site-specific denominator usually supports a much weaker conclusion.

Putting Database Evidence Together

A careful interpretation treats database information as one component of evidence, not as a verdict. The practical sequence is to identify the exact variant and relevant population, check the count and denominator, judge the precision of any zero count, and evaluate whether data quality and database scope support the comparison. If sources disagree, differences in ancestry composition, sample size, site assessment, or representation may explain some of the discrepancy; the largest database total does not automatically settle the question.

Common Mistake: Treating “Not Listed” as “Exceptionally Rare”

The incorrect reasoning is: “The database has no record of this variant, so the variant must be exceptionally rare.” This conclusion skips two questions: was the relevant site actually assessed, and how many relevant chromosome copies were assessed? Even a genuine zero count can be weak evidence when the denominator is small or the target ancestry is underrepresented. And when there is no site-specific denominator, a missing entry is not a measured zero.

The correction is to separate absence of a record from zero observed copies among assessed chromosomes. For the latter, report the ancestry group, AC, AN, and the resulting frequency or uncertainty. For the former, do not infer a frequency unless the database documentation supports that inference. In either case, describe the limits of the database evidence and avoid converting a sample observation into a claim of population-wide absence.