A Worked Audit of Rare-Variant Data

Why Audit Before Analysis?

A rare-variant analysis can be affected by errors that begin well before a statistical test is run. A frequency may use the wrong denominator, a sequencing artifact may look like a genuine low-frequency allele, or a transcript-dependent annotation may not match the intended functional mask. An audit makes these decisions visible and checks them in a defined order.

The goal here is not to test whether a gene is associated with a trait. It is to decide which variant records are adequately supported and meet the study’s stated frequency and functional inclusion rules. The resulting list is an input to downstream gene-level analysis, not evidence that any included variant has an effect.

Set the Audit Rules Before Reviewing Candidates

Consider a hypothetical study of 100 cases and 100 controls, each with a binary trait. All candidate records are autosomal and diploid. For each group, the maximum possible number of called chromosomes at a fully observed site is \(2(100)=200\). In practice, missing genotypes reduce that number. The audit uses a study-specific rule set declared in advance: call rate must be at least 95% in each group, the absolute difference in call rate between groups must be no more than 2 percentage points, and listed site-level diagnostics must pass the project’s calibrated quality criteria. The external population allele-frequency threshold is 0.01. The functional mask includes high-confidence protein-truncating and canonical splice variants, plus missense variants meeting the project’s pre-specified annotation rule.

These cutoffs are illustrative, not universal recommendations. Thresholds should be chosen for the sequencing platform, study design, ancestry context, and intended analysis, and then applied consistently. The audit also assumes that variant records have been harmonized to the same reference build and allele representation, so that frequency and annotation records refer to the same alternate allele.

Worked Example: Check Frequency Denominators

Audit Call Completeness and Quality

Frequency summaries are only interpretable alongside call completeness. A variant can appear to differ between groups because one group has more missing genotypes, or because a low-quality call has been counted as an allele. Review call rate separately in cases and controls, and examine the diagnostics selected for the sequencing assay, such as depth, mapping quality, strand balance, and evidence for alignment artifacts. The relevant metrics and thresholds should be defined for the project; a single number is not a substitute for checking the evidence that supports a call.

The call-rate comparison below is a basic screen, not a proof that missingness is harmless. Even similar overall call rates can conceal sample-specific or batch-related problems. Earlier material on coverage-driven missingness and case-control batch effects motivates checking these patterns in addition to a site-level summary.

Apply Functional Annotations With Transcript Context

A consequence label describes a variant relative to a transcript; it is not an intrinsic, transcript-free property of the genomic position. The same allele may receive different consequence labels for different transcripts. For that reason, record the genome build, annotation source and version, and the transcript-selection rule used in the audit. When the project specifies a preferred transcript, such as a selected canonical or MANE transcript, apply that choice consistently rather than selecting whichever annotation makes a variant appear most damaging.

The functional mask defines which annotation classes can proceed to analysis. In this example it includes high-confidence protein-truncating variants, canonical splice variants, and qualifying missense variants under a pre-specified rule. In silico predictions are uncertain evidence, so a missense variant’s eligibility depends on the stated mask, not on treating a prediction as a confirmed functional effect. A synonymous variant does not qualify for this particular mask, though another scientific question could define a different mask.

Worked Example: Combine Frequency, Quality, and Annotation Rules

A Worked Audit From Summary Table to Variant List

The following table consolidates the hypothetical records. “Call rate” is shown as cases/controls; \(AC/AN\) is shown separately by group. The quality column summarizes the project’s pre-specified site-level diagnostics, not a universal quality score.

For V4, \(AN=198\) in controls because one of the 100 control genotypes is missing; its control frequency is \(2/198=0.01010\). This is a useful reminder that a frequency calculation and a call-rate check answer different questions. For V6, complete call rates do not override a failed mapping-quality diagnostic. A high call rate alone cannot establish that the observed alleles are genuine.

Worked Example: Produce and Document the Final List

Why the Audit Order Matters

The checks form a dependency chain. A mismatched allele representation can make a correct frequency summary or annotation appear to describe the wrong record. Missingness changes the observed \(AN\), and poor call evidence can make even a correctly computed frequency misleading. Annotation is then interpreted under an explicit transcript and mask, and only after these checks are inclusion rules applied. The final record should preserve both the decision and its reason.

Common Mistake

Wrong reasoning: “The variant is very rare and annotated as damaging, so it should be included even though one group has many missing calls or the mapping metric fails.” This combines evidence from different checks as if a strong annotation could compensate for unreliable genotype data.

Correction: Treat required criteria as separate gates. First determine whether the allele calls and denominators are trustworthy; then apply the frequency and annotation rules as specified. A failed essential quality check is a reason to exclude or investigate the record, not a reason to reinterpret it as a convincing rare variant. Likewise, passing the audit means only that the record qualifies for the defined downstream analysis.