Why Weight and Mask Structural-Variant Events?
A gene-level burden test combines information across selected events in a gene, rather than testing each event separately. For structural variants, the selection and weighting decisions are consequential: a coding deletion, an intronic duplication, and a low-frequency event may not be treated alike. A mask specifies which events enter the burden; weights specify how much each included event contributes.
These are analysis choices, not neutral data-cleaning steps. A functional annotation rule determines which consequences count as relevant, and a frequency threshold determines which events count as rare enough. Changing either can change each participant’s burden and therefore the association hypothesis. The choices should be stated clearly and set before interpreting the association result.
This coding assumes the event representation and gene assignment have already been harmonized, and that an event is not duplicated in the event list. It is a carrier-burden definition, not a general encoding for every copy-number state. If the scientific question requires distinguishing copy-number states or counting multiple copies, the dosage coding must be specified instead; the binary definition above should not be silently substituted for it.
Defining Masks and Weights
A functional mask is a rule based on annotations, such as retaining events that overlap a specified coding consequence. A frequency mask retains events whose estimated frequency is no greater than a chosen cutoff. For this tutorial, let \(q_j\) be the estimated carrier frequency of event \(j\) in a prespecified reference population, and let \(q_{\max}\) be the frequency cutoff. The functional rule can be written \(F_j=1\) when the annotation meets the rule and 0 otherwise. The combined mask is
where \(I(\cdot)\) equals 1 when its condition is true and 0 otherwise. The reference population, frequency estimator, annotation source, gene-overlap rule, and treatment of uncertain annotations all need to be specified. Otherwise, the same nominal cutoff can produce different event sets.
There is no single required weighting formula for all studies. One illustrative scheme separates a functional multiplier \(a_j\) from a frequency component:
where \(q_0\) is a reference frequency chosen in advance and \(q_j>0\). For example, \(a_j\) might be larger for a prespecified high-impact consequence than for another qualifying consequence. This formula is a design choice, not a universal standard. With \(a_j\) held fixed, the frequency component increases as \(q_j\) decreases. The overall weight need not increase as frequency decreases across events with different functional multipliers.
Worked Example: Constructing a Burden From a Mask
Consider a hypothetical gene with three harmonized events. Use \(q_{\max}=0.01\), \(q_0=0.01\), and the weighting formula above. The functional rule retains coding exon events, and the frequency estimate is the carrier frequency in the prespecified reference population.
Event A is an exon deletion with \(q_A=0.002\) and \(a_A=2\), so it passes both masks. Event B is an intronic duplication with \(q_B=0.008\) and \(a_B=1\); it passes the frequency rule but not the functional rule. Event C is an exon duplication with \(q_C=0.01\) and \(a_C=1.5\), so it passes both masks, including the frequency boundary.
The included weights are
Event B contributes nothing because \(M_B=0\), even though it is frequent enough to pass the frequency rule. Suppose participant 1 carries A only, participant 2 carries C only, participant 3 carries both A and C, and participant 4 carries neither. Their burdens are \(B_1=4.472\), \(B_2=1.500\), \(B_3=4.472+1.500=5.972\), and \(B_4=0\). These numbers describe the chosen mask and weights; they are not copy-number dosages.
What the Burden Test Assumes
For a binary phenotype \(Y_i\), a common association model is logistic regression. Let \(C_i\) denote the prespecified covariates for participant \(i\). The model is
where \(\operatorname{logit}(p)=\log\{p/(1-p)\}\). The burden association null hypothesis is \(H_0:\beta=0\); a two-sided alternative is \(H_a:\beta\ne0\). Under this model, \(\beta\) is the change in log odds per one-unit increase in the specified weighted burden, conditional on the covariates. Thus \(\exp(\beta)\) is the conditional odds ratio for a one-unit increase in \(B_i\). If the weights are rescaled, the burden unit changes, and so does the numerical coefficient.
The procedure assumes that the chosen mask and weights represent the intended scientific burden, that the phenotype model is adequate, and that relevant confounding is handled by the covariates or study design. It also assumes observations are independent under the fitted model, unless the analysis uses an appropriate method for dependence. Event calls and annotations must be sufficiently reliable for the chosen event set. Standard errors and p-values must be calibrated for the data and design; sparse carrier counts can make ordinary large-sample approximations unreliable.
A burden model also constrains how event contributions combine: each event enters through \(w_jx_{ij}\) and a single coefficient \(\beta\). It does not estimate an independent effect for every event. The coefficient is therefore an association with the aggregate specified by the mask and weights, not a claim that every event has an identical biological effect.
Why the Burden Coefficient Is Tested
The basic likelihood calculation shows what the test asks. For simplicity, first consider a model with no covariates. For independent binary outcomes, the log likelihood is
Differentiating with respect to \(\beta\) gives the score, the rate at which the log likelihood changes as the burden coefficient changes:
Under \(H_0\), the model has no burden term, so the fitted probability does not depend on \(B_i\). The score compares observed outcomes with those null-model probabilities, weighted by each participant’s burden. A score test asks whether this contrast is too large to attribute to sampling variation under the null. With covariates, the nuisance parameters must also be accounted for; a fitted logistic regression can instead report a Wald statistic, \(\hat{\beta}/SE(\hat{\beta})\), when its approximation is suitable.
Worked Example: Testing a Unit-Weighted Carrier Burden
In a hypothetical cohort of 100 unrelated participants, a prespecified mask identifies six carriers of at least one qualifying event. For this illustration, each qualifying event has weight 1, and participants carry at most one qualifying event, so \(B_i\) is 1 for those six participants and 0 for everyone else. There are 3 cases and 3 controls among carriers; among the 94 noncarriers, there are 20 cases and 74 controls. Assume observations are independent, there are no covariates, and the logistic model is used.
With a binary burden, the fitted burden coefficient is the log odds ratio comparing carriers with noncarriers. The estimated odds ratio is
The usual large-sample standard error for this log odds ratio is
Thus \(z=1.3083328/0.854506\approx1.531\), giving a two-sided normal-approximation p-value of about 0.126. The 95% Wald interval for the log odds ratio is \(1.3083328\pm1.96(0.854506)=[-0.3665,2.9832]\); exponentiating gives an odds-ratio interval of approximately \([0.693,19.75]\). At \(\alpha=0.05\), this example does not provide statistically significant evidence of association. The interval is wide, and the ordinary approximation is questionable with only six carriers; the calculation illustrates the test, not a recommendation to ignore sparse-data diagnostics.
Annotation and Thresholds Define the Hypothesis
A mask is part of the definition of the predictor \(B_i\), not just a technical filter applied after the hypothesis is set. If the annotation rule changes, different events may contribute. If \(q_{\max}\) changes, the event set may change as well. In either case, participant burdens can change, and the coefficient tested is no longer the coefficient for the original burden. Testing several masks and reporting only the most favorable result also changes the inferential problem: the reported p-value does not account for the search unless the analysis addresses that selection.
Worked Example: A Frequency Cutoff Changes the Tested Burden
Consider two hypothetical coding events in one gene. Event D is a moderate-impact deletion with \(q_D=0.008\) and \(a_D=1\). Event E is a high-impact duplication with \(q_E=0.020\) and \(a_E=2\). Set \(q_0=0.01\), and let the functional rule include both events. Compare a narrow frequency mask with \(q_{\max}=0.01\) to a broader one with \(q_{\max}=0.025\).
Event D has weight \(w_D=\sqrt{0.01/0.008}\approx1.118\). Event E has weight \(w_E=2\sqrt{0.01/0.020}\approx1.414\). Under the narrow mask, E is excluded, so a participant carrying D only has burden 1.118 and a participant carrying E only has burden 0. Under the broader mask, those burdens are 1.118 and 1.414, respectively; a participant carrying both has burden \(1.118+1.414=2.532\).
The narrow-mask test asks whether the burden of qualifying events with frequency at most 0.01 is associated with the phenotype. The broader-mask test asks about a different aggregate, which also includes E. Neither test is simply a more accurate version of the other: each refers to the burden defined by its own annotation and frequency rules.
Worked Example: Frequency and Functional Multipliers Together
Suppose two hypothetical events pass the same functional mask, with \(q_0=0.01\) and the illustrative weight formula. A non-high-impact event has \(q=0.005\) and \(a=1\). Its weight is \(\sqrt{0.01/0.005}=\sqrt{2}\approx1.414\). A high-impact event has \(q=0.01\) and \(a=2\). Its weight is \(2\sqrt{0.01/0.01}=2.000\).
The first event is rarer, but its total weight is lower because the second event has a larger functional multiplier. Holding \(a_j\) fixed, the frequency component increases as \(q_j\) decreases; comparing total weights across events with different multipliers does not establish that the rarer event must receive the larger weight.
Common Mistakes and Practical Interpretation
- Calling a mask an objective biological truth. A mask is an operational definition based on annotations, frequency estimates, and thresholds. State those choices and interpret the test relative to them.
- Assuming rarer always means a larger total weight. The frequency component may increase as frequency falls when the functional multiplier is fixed, but the full weight also depends on \(a_j\).
- Interpreting \(\beta\) without its burden scale. The coefficient is per one unit of the specified weighted burden. If weights are changed, the unit and coefficient change.
- Treating a nonsignificant result as proof of no association. A p-value above \(\alpha\) is not evidence that \(H_0\) is true; consider the estimate, interval, carrier count, and test calibration.
- Changing masks after seeing results without accounting for the search. The resulting p-value does not automatically reflect how many annotations, thresholds, or weights were tried.
Check Your Understanding
Answer each question by referring to the definitions and model above.
- An event passes the frequency cutoff but fails the functional rule. What is its contribution to the burden, and why?
- Using \(w_j=a_j\sqrt{q_0/q_j}\), explain why a rarer event is not guaranteed to have a larger total weight than a more common event.
- In the logistic burden model, interpret \(\exp(\beta)\) and state what happens to its interpretation if the burden weights are rescaled.
- How does raising \(q_{\max}\) change the set of events considered by the test?
- Why is selecting and reporting only the most significant result from many masks a problem for interpreting its p-value?