Allele Harmonization for Cross-Population Summary Statistics

Why Allele Harmonization Matters

A reported association estimate is tied to a particular allele coding. One study may report the change in outcome per copy of G, while another reports the change per copy of A at the same biallelic site. If these estimates are compared without accounting for which allele is counted, their signs can appear to disagree even when the underlying association is consistent. A strand mismatch or genome-build mismatch can also make records that look similar refer to different alleles or positions.

Harmonization is the process of putting variant records and effect estimates into a common, verified orientation before comparing them. It is a data-integrity step, not a test of whether effects are equal and not evidence that the same causal mechanism operates across populations. The earlier material on association effect scales across populations is relevant here: estimates should be compared on compatible outcome and effect scales, but this tutorial focuses on identifying and aligning the alleles to which those estimates refer.

Allele Recoding and the Sign of an Effect

Suppose a diploid biallelic genotype is coded as a dosage \(G\), the number of copies of allele A, so \(G\) is 0, 1, or 2. If the other allele is B, the dosage of B is \(G'=2-G\). Under an additive model for a quantitative outcome, with covariates \(C\), write

Substituting \(G=2-G'\) gives the same model in terms of B dosage. This establishes the sign change directly:

Thus, changing the counted allele reverses the coefficient under this equivalent additive coding. The intercept changes, while the fitted values do not. In a linear model, the standard error is unchanged by this sign reversal; the test statistic changes sign and a two-sided p-value is unchanged. The same recoding logic applies to an additive logistic regression model on the log-odds scale: the coefficient changes sign and the intercept changes, while fitted probabilities remain the same. This reasoning assumes the same individuals, outcome, covariates, and model, and a dosage range of 0 to 2. Do not apply \(2-G\) without adjustment to a variant whose ploidy or dosage scale differs, such as some sex-chromosome analyses.

Strand Orientation and Variant Identity

DNA alleles can be reported on the forward or reverse strand. The base complements are A with T and C with G. For example, a forward-strand C/T record corresponds to G/A on the reverse strand: C maps to G, and T maps to A. Complementing the allele labels is distinct from changing which of the two alleles is counted. First translate a reverse-strand record to the chosen strand; then, if necessary, recode the effect estimate to the chosen effect allele. Keeping these as separate steps prevents sign errors.

A variant identifier alone is not a sufficient harmonization check. Records should be checked using the chromosome, position, genome assembly, and alleles, with a consistent representation of the variant. For insertions and deletions, equivalent records can have different textual representations unless they are normalized, commonly by left alignment and trimming redundant sequence against the reference. A genome-build conversion can change coordinates and may fail or map ambiguously. After conversion, verify that the alleles agree with the reference sequence for the target build; do not assume a successful coordinate conversion guarantees a valid allele match.

Genome Builds and Frequency Checks

A coordinate is meaningful only with its genome assembly. Before combining records, specify a target build and confirm that each record has been mapped to it. Check that the target position and alleles are valid for that assembly, and inspect any conversion failures, ambiguous mappings, or reference-allele mismatches. A record that maps to a nearby coordinate is not automatically the same variant. For indels and other complex variants, compare normalized allele representations rather than relying on position alone.

Allele frequencies are useful as cross-checks, but they must be interpreted in population context. If one record reports the frequency of an allele and another reports the frequency of the other allele, their frequencies should sum to one within the same sample, apart from rounding or differences in how frequencies were estimated. Across ancestry groups, however, the frequency of the same correctly aligned allele can differ substantially. Harmonization means aligning which allele a frequency and effect describe; it does not mean forcing ancestry-specific frequencies to agree.

Palindromic Variants Require Special Care

A palindromic single-nucleotide variant has allele pairs A/T or C/G. The pair looks the same after complementing and reversing the order of the alleles: for example, A/T becomes T/A. Consequently, allele letters alone cannot reveal whether a palindromic record was reported on the forward or reverse strand. This is different from a non-palindromic pair such as C/T, whose reverse-strand complement G/A gives a visible clue about strand orientation.

For a palindromic variant, effect-allele frequency may help only under suitable conditions. If the allele frequency is reliable, well away from 0.5, and can be compared to a frequency from an appropriate, sufficiently similar population, the frequency can distinguish candidate orientations. If the frequency is close to 0.5, the two orientations are difficult to distinguish because the complementary frequencies are also close to 0.5. If ancestry groups differ, comparing their frequencies as though they should match can create a false orientation decision. When orientation cannot be resolved confidently, flag the variant as ambiguous and exclude it from analyses requiring a known common effect allele, or handle it with a prespecified method that accounts for the uncertainty.

A Practical Harmonization Sequence

Diagnostics, Alternatives, and Limitations

Useful diagnostics include counting how many records required a sign flip, checking whether aligned frequencies are complements when expected, and examining the distribution of frequency discrepancies within ancestry groups. A concentration of apparent sign reversals, unexpected allele mismatches, or implausible frequency patterns can indicate a pipeline problem. Frequency discrepancies should be investigated, not automatically “corrected”: they may reflect ancestry, sampling variation, different sample inclusion, imputation quality, or a coding error.

Matching only by variant identifier is quicker but can miss build, allele, and representation errors. Inferring orientation from the sign of the association estimate is not a valid alternative: it makes the desired direction determine the coding and can conceal real differences. Dropping every palindromic variant is conservative and often appropriate when orientation information is poor, but it can discard usable variants when reliable frequency information and strand metadata resolve them. Conversely, frequency-based resolution is unsafe when frequencies are imprecise, near 0.5, or drawn from populations that are not comparable.

Harmonization cannot repair a wrongly specified model, an incompatible outcome scale, an incorrect sample label, or a variant that was misidentified before the summary statistics were produced. Nor does successful alignment imply that an allele has the same association effect in each population. It only ensures that the reported estimates are expressed for a verified common allele at a common variant representation, as far as the available data permit.