Allele Harmonization Across Association Results

Why Allele Harmonization Matters

An association result is not fully interpretable from its effect estimate alone. The estimate is tied to the allele whose copies were counted: a positive coefficient means the modeled outcome or trait increases with each additional copy of the effect allele, on the model’s reported scale. If another cohort counts the opposite allele, the same association can have the opposite coefficient sign. Comparing those numbers without checking allele coding can make agreement look like disagreement, or vice versa.

Harmonization means putting association results for the same genetic variant into a consistent allele orientation before comparing their estimates. It is not a method for making different variants equivalent, nor does it resolve differences in phenotype definition or effect scale. Those conditions must be assessed separately before estimates are compared.

Definition: For a biallelic variant, the effect allele (EA) is the allele whose dosage is used to define the reported effect estimate. The other allele (OA) is the second allele. In an additive dosage analysis, \(G\) is the number of EA copies an individual carries, typically 0, 1, or 2, and \(\beta\) is the estimated change in the model’s outcome scale per one-copy increase in \(G\). The reported standard error \(SE\) quantifies uncertainty in that estimate.

Two Different Operations: Complementing and Swapping

A DNA strand complement changes the letter used to represent an allele: A complements T, and C complements G. Thus, a variant written as C/T on one strand may be written as G/A on the other. If the effect allele on the second record is G, it corresponds to C on the first record. This change of strand notation does not, by itself, change the biological allele being counted or reverse the coefficient sign.

A different operation is an effect-allele swap: changing the counted allele from EA to OA. For a diploid biallelic variant, if \(G\) counts copies of the original EA, then the dosage \(H\) of the other allele is \(2-G\). In a model whose predictor is linear in dosage, the coefficient changes sign when the dosage is recoded this way.

\eta=\alpha+\beta G,\qquad G=2-H \quad\Longrightarrow\quad \eta=(\alpha+2\beta)-\beta H.

The new intercept is \(\alpha+2\beta\), and the coefficient for \(H\) is \(-\beta\). The same algebra applies to a standard additive linear regression and to a logistic regression with a linear dosage term in the linear predictor. The standard error is unchanged by multiplying the coefficient by \(-1\), so a two-sided test statistic changes sign but has the same absolute value and two-sided p-value. This conclusion assumes the same individuals, model, and test; it does not assert that every analysis or one-sided test is invariant.

The simple sign rule depends on the predictor being linear in dosage. For a model with a quadratic dosage term, the transformed linear coefficient is not generally just the negative of the original linear coefficient.

\alpha+\beta G+\gamma G^2 =\alpha+2\beta+4\gamma+(-\beta-4\gamma)H+\gamma H^2.

Here the coefficient of \(H\) is \(-\beta-4\gamma\), while the coefficient of \(H^2\) remains \(\gamma\). Therefore, do not apply the simple sign-reversal rule to a nonlinear dosage model without transforming the full predictor.

Worked Example: Complementing the Strand and Swapping the Effect Allele

Two hypothetical cohorts report the same biallelic variant at the same verified genome-build coordinate. Cohort A reports C/T, with C as EA and \(\beta_A=0.120\), \(SE_A=0.040\). Cohort B reports the complementary letters G/A, with A as EA and \(\beta_B=-0.110\), \(SE_B=0.035\). Both estimates are per EA copy on the same model scale.

Step 1: Translate the strand letters. G complements C and A complements T. Therefore, B’s G/A pair corresponds to A’s C/T pair, and B’s effect allele A corresponds to T in A’s representation.

Step 2: Identify the dosage orientation. A counts C copies; B counts copies of the allele corresponding to T. These are opposite alleles at the same biallelic site, so harmonizing B to C requires an effect-allele swap. The needed operation is not a sign change merely because the strand letters differ.

Step 3: Transform the estimate. Since the model is linear in dosage, the coefficient per C copy is \(-\beta_B=-(-0.110)=+0.110\). Its standard error remains \(0.035\). The aligned estimates are therefore \(+0.120\) and \(+0.110\), close in magnitude and direction on the stated scale.

As a contrast, if B had reported G as its effect allele with \(\beta_B=+0.110\), G would complement C. That would be the same biological effect allele, so the aligned estimate would remain \(+0.110\); no sign reversal would be made.

Palindromic Variants Require Extra Evidence

A/T and C/G variants are called palindromic because complementing their allele letters produces the same unordered pair. For example, A/T complements to T/A. If strand information is missing, the letters alone cannot distinguish the two possible strand orientations. A/T and C/G variants therefore need particular care in cross-cohort comparisons.

An effect-allele frequency can sometimes help. If two records are known to describe the same variant and their cohorts have comparable allele frequencies, the frequency of an allele provides evidence about whether the reported effect alleles refer to the same allele or to opposite alleles. This is evidence, not proof: ancestry differences, sampling variation, allele-frequency estimation, and data errors can all affect the comparison. Frequency matching is especially uninformative when the allele frequency is near 0.5, because the allele and its counterpart have similar frequencies.

Worked Example: Assessing a Palindromic A/T Variant

Suppose two hypothetical cohorts report the same verified A/T variant. Cohort A reports EA A, effect-allele frequency \(0.12\), and \(\beta_A=+0.100\). Cohort B reports EA T, effect-allele frequency \(0.87\), and \(\beta_B=-0.090\). Assume, for this frequency check, that the cohorts have sufficiently comparable allele frequencies.

The complement of A is T, so the letters cannot resolve the strand orientation. Compare the reported frequencies: if B’s T allele were the same allele as A’s A, its frequency would be expected to be near \(0.12\), not \(0.87\). The frequency of the allele opposite B’s T is \(1-0.87=0.13\), which is close to A’s \(0.12\). This supports the interpretation that B’s T corresponds to the allele opposite A’s A.

Under that interpretation, B must be swapped to A as the effect allele. The aligned estimate is \(-\beta_B=-(-0.090)=+0.090\), with \(SE\) unchanged. The two estimates, \(+0.100\) and \(+0.090\), are directionally consistent. The frequency check supports this alignment but does not establish it with certainty; if the frequency evidence were weak or the cohorts poorly comparable, the orientation should remain unresolved rather than being guessed.

Variant Identity Comes Before Allele Alignment

Allele harmonization is meaningful only after establishing that records refer to the same variant. A shared rsID can be a useful clue, but identifiers, genomic coordinates, and allele labels should be checked together. Coordinates depend on the reference genome build, so records at the same numeric position on different builds are not automatically the same variant. For indels, equivalent sequence changes can also have different representations; comparison may require consistent trimming and left alignment against the same reference sequence. Multiallelic sites require allele-specific matching: an estimate for one alternate allele is not automatically interchangeable with an estimate for another.

If the variant identity or allele mapping cannot be established, do not repair the record by changing a sign or by matching on the nearest coordinate. Mark it as unresolved or incompatible and exclude it from comparisons that require a verified match. A failed harmonization is safer than a confident but incorrect alignment.

Worked Example: Rejecting an Unverified Variant Match

A hypothetical result file lists a variant at chromosome 5, position 400,000, with A/G alleles on build 38 and an effect estimate per G copy. A second file lists chromosome 5, position 400,000, with C/T alleles on build 37 and an effect estimate per T copy. The records have a similar identifier, but no build conversion or sequence-based match has yet been performed.

The letters C/T complement to G/A, which might appear to suggest a match. However, the coordinate systems differ, and the apparent allele relationship does not establish that the records refer to the same genomic site. In particular, a build difference cannot be resolved merely by complementing letters. Without verifying the build conversion and the variant identity, there is no justified dosage transformation and no justified coefficient sign change.

The correct decision is to flag this pair as unresolved pending variant matching. If subsequent reference-based checking establishes that the records are the same biallelic site and that T corresponds to G, the estimates can then be aligned without a sign change. If it establishes that T corresponds to the opposite allele, the effect estimate must be sign-reversed for a linear dosage model. Until that evidence exists, neither conclusion is warranted.

A Practical Harmonization Sequence

A consistent workflow separates variant matching, allele orientation, and estimate transformation. The following sequence makes each decision explicit and leaves an audit trail.

Verify the variant.
Check chromosome, position, reference genome build, and allele pair. For indels or multiallelic records, establish a consistent representation and an allele-specific match before proceeding.

Record the reported coding.
For every result, record EA, OA, the effect-allele dosage definition, effect estimate, and \(SE\). Do not infer the effect allele from the sign or size of the estimate.

Translate strand notation.
Apply A↔T and C↔G where relevant. Determine whether the translated EA matches the chosen target allele or its opposite. Complementation alone does not imply a sign change.

Resolve ambiguity or stop.
For palindromic pairs, use reliable strand information or carefully assessed frequency evidence. If neither supports an orientation, retain an unresolved status rather than selecting one arbitrarily.

Transform and document.
For a verified effect-allele swap in a linear dosage model, multiply the coefficient by \(-1\), retain its \(SE\), and record the target EA. Use the full reparameterization for nonlinear dosage models.

Worked Estimate Transformation

Once a pair is verified as the same variant, the dosage transformation is simple for a linear model. It is still important to show the calculation and retain the original result, so a later reviewer can see both the reported and harmonized orientations.

Worked Example: Harmonizing a Direct Allele Swap

Two hypothetical results describe the same verified A/G variant and report effects from additive logistic models with linear dosage terms. Study 1 reports EA G, \(\beta_1=+0.150\), \(SE_1=0.050\). Study 2 reports EA A, \(\beta_2=-0.130\), \(SE_2=0.040\). The target orientation is G.

Study 1 is already oriented to G, so its aligned estimate is \(+0.150\) with \(SE=0.050\). Study 2 counts A, the opposite allele, so its aligned coefficient is:

\beta_{2,\text{aligned}}=-\beta_2=-(-0.130)=+0.130, \qquad SE_{2,\text{aligned}}=0.040.

For a check, Study 2’s original coefficient divided by its standard error is \(-0.130/0.040=-3.25\). After harmonization it is \(+0.130/0.040=+3.25\): the sign changes and the magnitude does not. Its 95% Wald interval before harmonization is \(-0.130\pm1.96(0.040)=[-0.2084,-0.0516]\). Reversing the allele orientation gives \([+0.0516,+0.2084]\). On the stated model scale, the aligned estimates are both positive per G copy. This establishes directional agreement, not equality of the estimates or proof that their underlying effects are identical.

Common Mistakes and Interpretation

Flipping the sign whenever allele letters differ. Different letters can be strand complements for the same allele. Translate the strand first, then decide whether the effect allele was actually swapped.
Assuming a negative coefficient means the other allele is the effect allele. The coefficient’s sign does not identify the dosage coding. Use the reported EA and dosage definition.
Forcing an orientation for a palindromic variant. If strand and frequency evidence do not resolve it, label it unresolved. A guessed sign can manufacture apparent replication or apparent disagreement.
Applying the sign rule to every model. Sign reversal follows from replacing \(G\) with \(2-G\) in a predictor linear in dosage. With nonlinear dosage terms, transform the complete model expression.
Matching on an identifier alone. Verify coordinate build and alleles; do not treat a plausible-looking identifier or coordinate as proof that the variants match.

Key takeaway: First establish that the records refer to the same variant; then translate strand notation; then determine whether the effect allele itself has been swapped. Only a verified swap calls for a sign reversal under a linear-in-dosage model. Ambiguous cases should remain flagged, not silently forced into alignment.

Check Your Understanding

Answer each question using the stated allele coding and model assumptions.

A result reports C/T with EA C. Another reports G/A with EA G at the verified same site. After complementing the second record, should its coefficient change sign? Explain.
For a linear dosage model, show how the coefficient changes when the dosage \(G\) is replaced by \(H=2-G\). What happens to its standard error?
Why can allele frequencies help with an A/T variant, and why might they still be insufficient to determine orientation?
What evidence is needed before treating two records at the same numeric coordinate but on different genome builds as the same variant?
For a predictor \(\alpha+\beta G+\gamma G^2\), what is the coefficient of \(H\) after substituting \(G=2-H\)?