Why Define a Mask Explicitly?
A gene can contain many variants, but a burden test does not automatically treat every variant in that gene as relevant to the same biological question. A mask specifies which variants contribute to the gene-level analysis. For example, one analysis might select qualifying loss-of-function annotations, while another might select qualifying loss-of-function and missense annotations. The distinction is meaningful only in relation to the variants that actually meet all the rules: adding a category changes the selected set if it contributes at least one eligible variant, but not otherwise.
A mask is therefore both a scientific choice and an operational rule. Its functional categories describe which consequences are of interest; its inclusion and exclusion criteria determine how annotation and quality information translate into membership. The set of selected variants then determines which genetic information is summarized for each participant and associated with the trait.
Writing the Membership Rule
Begin by defining the candidate variant universe \(V_g\): the variants considered for gene \(g\) after the analysis has established the relevant genomic region and data-quality procedures. Each candidate variant \(v\) has annotation information, which may depend on transcript choice, and may also have eligibility information such as call quality. Let \(C\) be the set of functional consequence categories selected by the investigator. A general membership rule is:
Here \(A(v)\) is the set of consequence categories assigned to variant \(v\) under the specified annotation procedure, \(E(v)=1\) means that the variant passes all required inclusion or eligibility criteria, and \(X(v)=1\) means that it meets a stated exclusion criterion. The realized mask \(M_g(C)\) contains only variants that pass eligibility, have at least one selected consequence category, and are not excluded.
This expression makes the logic explicit. A variant must satisfy the required conditions together; a selected consequence label alone does not guarantee membership. If a variant has more than one annotation, the rule should say whether any selected annotation is sufficient or whether another annotation takes precedence. If exclusions override inclusions, say so directly. Also specify the annotation source and transcript rule: a consequence label can differ across transcripts, so membership can change when the transcript-selection procedure changes.
The categories and exclusions should correspond to the intended scientific question. A mask restricted to predicted loss-of-function consequences, for example, targets a different functional definition from one that also admits missense consequences when qualifying missense variants are present. Neither rule establishes that every selected variant changes the trait, or that all selected variants have effects in the same direction. A mask defines the variants grouped for analysis; it does not prove their biological effects.
From Mask Membership to the Tested Association
A burden test associates a participant-level summary of variants in a specified mask with a phenotype. Write \(B_i(M_g)\) for the summary calculated for participant \(i\) using variants in the realized mask. The precise summary and genotype coding are separate analysis choices; for understanding the role of the mask, the key point is that the mask determines which variant data are supplied to that summary.
For a continuous outcome, a schematic regression model is:
Here \(Y_i\) is the outcome, \(Z_i\) is a vector of covariates, and \(\beta_{M_g}\) describes the conditional association between the specified genetic summary and the outcome in this model. The corresponding null hypothesis is \(H_0:\beta_{M_g}=0\). For another outcome type, the regression model and link may differ, but the mask still determines the variants entering the genetic summary. An association is not, by itself, evidence that the variants cause the outcome.
Thus, interpreting a burden result requires more than naming the gene. The mask defines what counts as contributing variation in that gene. If the mask includes an additional eligible category, the genetic predictor may represent a broader functional set. If no eligible variants belong to that added category, the realized mask is unchanged. Even when the realized variant sets differ, the observed participant summaries can coincide if the added variants contribute no observed genotype information under the chosen coding. These distinctions matter when describing what the test actually examined.
Worked Example: Selecting Predicted Loss-of-Function Variants
Hypothetical data. A team considers five candidate variants in one gene. Its specification selects stop-gain and frameshift consequences, requires a call-quality eligibility flag \(E(v)=1\), and has no additional exclusion category in this example.
| Variant | Consequence category | \(E(v)\) | Selected? |
|---|---|---|---|
| v1 | Stop-gain | 1 | Yes |
| v2 | Missense | 1 | No |
| v3 | Frameshift | 0 | No |
| v4 | Frameshift | 1 | Yes |
| v5 | Synonymous | 1 | No |
Let \(C=\{\text{stop-gain},\text{frameshift}\}\). Applying the membership rule gives \(v1\) because it is in a selected category and \(E(v1)=1\); \(v2\) and \(v5\) fail the category requirement; and \(v3\) is in a selected category but fails eligibility. Therefore,
The mask specification is not simply “all loss-of-function variants”: it also records the named categories and the eligibility rule used. In the association analysis, the participant-level genetic summary is formed from \(v1\) and \(v4\), not from every candidate in the gene.
Inclusions, Exclusions, and Annotation Ambiguity
A reliable mask has rules that can be applied consistently to every candidate variant. Inclusion rules state what qualifies; exclusions state what must be left out even if an inclusion rule appears to be met. For example, a team might select specified missense annotations but exclude variants whose annotation is marked uncertain under its prespecified procedure. The exclusion should be explicit rather than left to ad hoc decisions after seeing which variants are associated with the outcome.
When a variant has multiple consequence labels, the rule must also make the treatment of multiple labels reproducible. If the mask includes a variant whenever any selected category applies, state that. If only the consequence on a particular transcript is used, specify how that transcript is chosen. Otherwise, analysts can make different membership decisions from the same underlying sequence data and arrive at different realized masks.
A mask specification should be recorded in enough detail to reconstruct the realized set. At minimum, report the gene or gene boundary definition used by the analysis, the selected consequence categories, the annotation and transcript procedure, all relevant eligibility rules, and any exclusions. The aim is not to list every general data-processing step again, but to make the criteria that determine mask membership transparent and auditable.
Worked Example: Applying an Exclusion Rule
Hypothetical data. A second analysis selects missense variants with \(E(v)=1\), but excludes variants flagged as uncertain by a prespecified annotation review. Let \(C=\{\text{missense}\}\), and let \(U(v)=1\) indicate that a variant has the uncertainty flag. The rule is \(E(v)=1\), \(A(v)\cap C\ne\varnothing\), and \(U(v)=0\).
| Variant | Consequence | \(E(v)\) | \(U(v)\) | Selected? |
|---|---|---|---|---|
| v1 | Missense | 1 | 0 | Yes |
| v2 | Missense | 1 | 1 | No |
| v3 | Missense | 0 | 0 | No |
| v4 | Synonymous | 1 | 0 | No |
Variant \(v1\) passes all three conditions. Variant \(v2\) passes category and eligibility requirements but is excluded because \(U(v2)=1\). Variant \(v3\) fails eligibility, and \(v4\) fails the category requirement. Thus the realized set is \(M_g(C)=\{v1\}\), with \(1\) selected variant. If the exclusion were omitted, the realized set would instead be \(\{v1,v2\}\). The written exclusion therefore changes membership here because an eligible missense variant meets it.
Comparing Specifications and Realized Sets
A comparison of masks should distinguish three things: the written specifications, the resulting variant membership, and the participant-level summaries used in the test. A difference at one level does not automatically imply a difference at the next. Suppose specification A selects eligible loss-of-function variants, while specification B selects eligible loss-of-function or missense variants. The specifications differ. Their realized sets differ only if at least one eligible missense variant is present and not excluded. If no such variant is present, both specifications select the same realized set.
Worked Example: A Broader Specification That Selects the Same Set
Hypothetical data. Four variants in a gene are evaluated under two specifications. Both require eligibility and exclude variants marked uncertain. Specification A selects stop-gain and frameshift consequences. Specification B selects those same categories plus missense.
| Variant | Consequence | Eligible? | Uncertain exclusion? |
|---|---|---|---|
| v1 | Stop-gain | Yes | No |
| v2 | Frameshift | Yes | No |
| v3 | Missense | No | No |
| v4 | Synonymous | Yes | No |
For specification A, the eligible selected variants are \(v1\) and \(v2\), so \(M_A=\{v1,v2\}\) and \(|M_A|=2\). Specification B adds the missense category, but \(v3\) fails eligibility, while \(v4\) is not in any selected category. Therefore \(M_B=\{v1,v2\}\) and \(|M_B|=2\). The specifications differ, but the realized sets do not. In this data, the tests use the same variant membership; describing them as tests of different realized variant sets would be inaccurate.
If \(v3\) instead passed eligibility and was not excluded, the result would be \(M_B=\{v1,v2,v3\}\), while \(M_A=\{v1,v2\}\). Then the realized sets would differ. Whether the observed participant-level summaries also differ depends on the genotypes and summary rule: for example, if no participant has a called contributing genotype at \(v3\), the summaries could still coincide in the analyzed sample. A precise report distinguishes these possibilities rather than inferring them from category names alone.
Why the Composition Matters
Different functional categories can encode different biological hypotheses. A mask restricted to consequences expected to disrupt a transcript asks about variation meeting that narrower definition. Adding missense annotations can broaden the set toward variants that alter protein sequence but may have uncertain or heterogeneous functional consequences. The broader mask does not necessarily select any additional variants, and even if it does, it does not establish that the additional variants have the same effect direction or magnitude as those already selected.
The mask also affects the interpretation of a null result. A non-significant association is evidence about the tested predictor under the specified model and data, not a general conclusion that the gene has no role in the trait. The test may have used a narrow category, a broader category, or a realized set containing only a few variants. Clear membership rules make it possible to understand the scope of that conclusion and to compare analyses meaningfully.
A useful audit follows the logic from rule to result: list the candidate variants, apply the selected categories, apply eligibility criteria, apply exclusions, and compare the final set with the intended specification. Then verify that the association analysis used that realized set to construct its genetic summary. This sequence helps reveal both accidental omissions and claims that overstate the difference between two mask analyses.
Common Mistake
Wrong reasoning: “The second mask includes missense variants, so it tests a different set of variants.” That conclusion does not follow from the specifications alone. If there are no eligible missense variants, or all such variants are excluded, the realized sets can be identical. Conversely, distinct realized sets do not guarantee different observed participant summaries if the additional variants contribute no observed genotype information under the coding used.
Correction: First describe the mask specifications as different. Then compare their realized membership. Say that they test different realized variant sets only when at least one variant is selected by one specification and not the other. If relevant, separately state whether the resulting participant-level genetic summaries differ in the analyzed data.
Check Your Understanding
Use the distinction between a mask specification, its realized set, and the association predictor to answer each question.
- A mask selects eligible stop-gain and frameshift variants. One frameshift variant fails the eligibility rule. Is it part of the realized mask? Explain which conditions it satisfies and which it fails.
- Specification B adds missense consequences to specification A, but no eligible missense variants are present. How should the specifications and realized sets be described?
- Why should a mask state how transcript choice is handled when consequence labels can vary across transcripts?
- Two specifications select different realized variant sets, but the extra variant has no observed contributing genotype in any participant. Must the participant-level summaries differ? Explain.
- What association does a burden test examine when its mask selects a stated set of variants?