Why Weight Variants by Functional Annotation?
A variant set can contain alleles with very different predicted consequences. A stop-gain variant and a missense variant predicted to be damaging may plausibly matter more than a synonymous variant, but a simple count gives each observed allele the same contribution. Functional annotation weights offer one way to express that distinction in a burden score.
A weight is a rule for combining genotype data with annotation information; it is not a measured effect size. Choosing a larger weight means that, under the score’s working assumptions, the corresponding variant contributes more to the individual’s burden. It does not establish that the variant has a larger causal effect, nor that the annotation prediction is correct.
The genotype coding must be consistent: \(G_{ij}\) counts the same allele at every variant, typically the allele designated as alternate or minor according to the analysis definition. A larger \(w_j\) increases the contribution of each counted copy at variant \(j\). If weights are nonnegative, the score is also nonnegative; the score’s units are weighted allele copies, not a clinical measurement.
What a Weighting Scheme Assumes
A burden score becomes especially interpretable when connected to a model for the phenotype. Suppose a model’s linear predictor contains the variant contributions \(\sum_j \beta_jG_{ij}\), where \(\beta_j\) denotes the per-allele effect at variant \(j\). A weighted burden construction imposes a particular pattern on those effects:
This is the key justification: replacing the separate variant contributions with one burden score corresponds to assuming that variant effects follow the relative pattern specified by the weights, up to a common multiplier \(\theta\). With nonnegative weights, the pattern does not encode opposing directions of effect. The common multiplier may be positive or negative, but the effects are still constrained to share the direction implied by that multiplier. The annotation weights specify relative contributions, not the value of \(\theta\).
This interpretation requires care. Annotation categories or prediction scores must be assigned to the same variants represented in the genotype data; dosage coding and allele orientation must be consistent; and the weights should be chosen independently of the phenotype association results being tested. If weights are selected after inspecting which choices produce the strongest association, nominal evidence can be overstated unless that selection is handled in the analysis. The working effect pattern is also an assumption to examine, not a biological fact guaranteed by the annotation.
Turning Annotations Into Weights
There is no single universal conversion from annotation to weight. A simple scheme can assign fixed values to consequence classes. A continuous prediction can be used directly, transformed, capped, or grouped into categories. Whatever the rule, state it explicitly so another analyst can reproduce which variants receive more influence.
| Annotation information | Possible weight rule | What larger weights express |
|---|---|---|
| Predicted consequence class | For example, 1.0 for loss-of-function, 0.6 for damaging missense, and 0.1 for synonymous | A prespecified ranking of expected functional relevance |
| Continuous deleteriousness prediction \(a_j\) | Use \(w_j=a_j\), or use \(w_j=a_j^2\) | Direct proportional weighting, or greater emphasis on high predictions |
| Prediction with extreme values | Use a capped weight such as \(w_j=\min(a_j,c)\) | High values still receive more weight, but cannot dominate without limit |
These are examples of possible rules, not universal biological calibrations. A prediction may be on a scale for which larger values indicate greater deleteriousness, or it may use the reverse direction; the analyst must check the definition before constructing weights. When annotations are categorical, the numerical spacing between assigned weights is a choice. Assigning 1.0 and 0.5 does not prove that one consequence is twice as harmful.
Worked Example: From Consequence Classes to a Burden Score
Consider a hypothetical set of three variants. The chosen rule assigns weights 1.0 to a loss-of-function prediction, 0.6 to a damaging missense prediction, and 0.1 to a synonymous prediction. Individual A has dosages \((1,0,2)\) at those variants, respectively. Calculate the weighted burden.
The score is $$ B_A=(1.0)(1)+(0.6)(0)+(0.1)(2)=1.0+0+0.2=1.20. $$
Individual A therefore has a burden of 1.20 weighted allele copies under this rule. The two synonymous allele copies contribute 0.20 in total, while the single loss-of-function allele contributes 1.0. This result describes the arithmetic of the chosen score; it does not show that the loss-of-function allele has a true effect five times the effect of a synonymous allele.
For comparison, if individual B has dosages \((0,1,1)\), then \(B_B=(1.0)(0)+(0.6)(1)+(0.1)(1)=0.70\). Under these weights, A has the larger score, 1.20 versus 0.70.
How Transformations Change Contributions
A nonlinear transformation can change the relative influence of variants, even when it preserves their ranking. Squaring prediction values, for example, makes high values larger relative to low values. The score can consequently change both in magnitude and in how much each variant contributes. A numerical comparison is more informative than describing the transformation only as “up-weighting” damaging variants.
Worked Example: Equal Versus Squared Prediction Weights
Suppose three hypothetical variants have prediction values 0.2, 0.6, and 0.9. Compare equal weights, \(w_j=1\), with squared weights, \(w_j=a_j^2\). Individual A carries one allele at each of the first two variants and none at the third, so the dosage vector is \((1,1,0)\). Individual B carries one allele only at the third variant, giving \((0,0,1)\).
Under equal weights, A’s score is \((1)(1)+(1)(1)+(1)(0)=2\), and B’s is \((1)(0)+(1)(0)+(1)(1)=1\). Under squared weights, the weights are \(0.2^2=0.04\), \(0.6^2=0.36\), and \(0.9^2=0.81\). Thus A’s score is \((0.04)(1)+(0.36)(1)+(0.81)(0)=0.40\), while B’s score is \((0.04)(0)+(0.36)(0)+(0.81)(1)=0.81\).
The first individual’s burden is 0.40 under the squared scheme, compared with 2 under equal weights. The second individual’s squared-weight burden is 0.81, concentrated at the highest-prediction variant. In this hypothetical comparison, the ranking changes: A has the larger score under equal weights, but B has the larger score under squared weights. The change reflects the selected weighting rule, not new evidence about either individual’s phenotype.
A common transformation is a positive global rescaling, \(w_j^\star=cw_j\) for \(c>0\). It changes the numerical units of the score but not the relative contributions:
For example, multiplying every weight by 5 multiplies every individual’s score by 5 and leaves their ordering unchanged. This differs from squaring prediction values, which generally changes the relative weights and can change that ordering. When comparing weight schemes, separate pure rescaling from a change in relative contributions.
Worked Example: Capping an Extreme Prediction
Suppose a prediction scale has values 0.2, 0.6, and 1.4, where larger values are intended to indicate greater predicted deleteriousness. Use the raw prediction values as weights, then compare them with capped weights \(w_j=\min(a_j,1.0)\). Individual C carries two alleles at the third variant and none at the others, with dosage vector \((0,0,2)\). Individual D carries one allele at the first variant and one at the second, with vector \((1,1,0)\).
With raw weights \((0.2,0.6,1.4)\), C’s score is \((0.2)(0)+(0.6)(0)+(1.4)(2)=2.8\). D’s score is \((0.2)(1)+(0.6)(1)+(1.4)(0)=0.8\). With capped weights \((0.2,0.6,1.0)\), C’s score is \((0.2)(0)+(0.6)(0)+(1.0)(2)=2.0\), while D’s remains 0.8.
C still has the larger score under the capped scheme, but the gap falls from 2.0 to 1.2 weighted allele copies. C’s score falls because both counted alleles are at the variant whose weight was reduced. Capping can limit the leverage of an extreme annotation value; it does not establish that the capped value is more accurate.
Checking a Weighting Choice
Before interpreting or testing a weighted burden, check that the annotation and score construction behave as intended. These checks focus on whether the score represents the prespecified rule, rather than on a particular downstream association procedure.
Confirm what larger and smaller values mean, how missing annotations are treated, and whether the prediction corresponds to the consequence definition used for the variant set.
Tabulate weights by consequence class and examine their range. Check for unintended zeros, negative values, extreme values, or ties that erase distinctions the rule was meant to preserve.
For representative individuals, calculate each term \(w_jG_{ij}\), not only the total. Confirm that the dominant terms are plausible under the stated annotation rule and dosage coding.
Compare prespecified alternatives such as equal, direct, squared, or capped weights. Large changes in scores or rankings mean conclusions may depend strongly on the annotation-to-weight mapping.
A mismatch between allele orientation and dosage coding can reverse the intended interpretation of a genotype contribution. A reversed prediction scale can give the largest weights to variants intended to receive the smallest ones. An annotation pipeline can also label variants differently across transcripts, so a class assignment should follow a stated transcript or consequence policy. These are construction failures, not subtle biological disagreements.
Alternatives, Limitations, and a Common Pitfall
Equal-weight allele burden is a useful comparator: it treats every counted allele alike and avoids claiming that the annotation scale gives reliable relative importance. A hard-filter approach instead includes only variants meeting a specified annotation criterion, leaving excluded variants out of the score. Functional weighting retains variants but changes their contributions. Frequency-weighted scores use allele frequency information to shape contributions; functional and frequency information answer different prioritization questions and can be combined only with an explicit rule.
None of these choices removes the central limitation of a single burden score: it compresses multiple variant genotypes into one number. As the derivation shows, its simplest interpretation relies on effects following a common weighted pattern. If some variants increase risk and others decrease it, or if their effects depart substantially from the weight ratios, contributions can cancel or the score can be poorly aligned with the phenotype. Annotation weighting does not by itself solve that problem. Methods that allow more heterogeneous effects are distinct alternatives and require their own assumptions and calibration.
Weights may also be noisy, correlated with one another, or poorly transferable across genes, transcripts, or populations. A weight is not a probability that a variant is causal. Nor is a prediction value necessarily calibrated to the size of a phenotype effect. Sensitivity analyses can reveal dependence on arbitrary choices, but they do not identify which weighting scheme is biologically correct.
Functional annotation weights are most useful when the biological rationale is explicit, the mapping from annotation to weight is reproducible, and conclusions are not presented as if the weights were observed effects. Report the annotation source and version, the variant-consequence policy, the exact weight formula, and how missing or extreme values were handled. That information makes the score interpretable and allows its sensitivity to alternative rules to be assessed.
Check Your Understanding
Answer each question, then compare with the worked solutions below.
- Two variants have weights 0.3 and 0.8. An individual has dosages 2 and 1, respectively. What is the weighted burden?
- What effect pattern is implied by setting \(\beta_j=\theta w_j\) for all variants in a burden score?
- If every weight is multiplied by 4, what happens to the burden values and their ordering?
- Why does squaring nonnegative prediction values generally do more than change the units of a burden score?
- Name one diagnostic check that could identify an annotation-weighting error.
Solutions. (1) \(B=(0.3)(2)+(0.8)(1)=0.6+0.8=1.4\) weighted allele copies. (2) Variant effects are constrained to be proportional to their weights through one common multiplier; with nonnegative weights, the pattern does not represent arbitrary opposing directions. (3) Every burden is multiplied by 4, since \(B_i^\star=\sum_j4w_jG_{ij}=4B_i\); the ordering is unchanged. (4) Squaring generally changes ratios between weights, so variants can contribute differently relative to one another and individual score rankings may change. (5) For example, verify that the prediction scale’s direction matches the rule and tabulate weights by consequence class; either check can reveal a reversed mapping.