Palindromic Variants And Strand Ambiguity

Why Palindromic Variants Need Special Care

In summary-data Mendelian randomization (MR), each variant’s association with an exposure and an outcome must refer to the same effect allele before the associations are combined. For an ordinary biallelic variant, the allele labels often reveal whether one dataset reports the complementary DNA strand: for example, complementing A/G gives T/C, a different pair. A/T and C/G variants are different. Their complements are still A/T and C/G, respectively, so the allele pair alone cannot show whether one dataset uses the opposite strand.

This is strand ambiguity. It is not a claim that the variant has no biological identity or that its association is inherently uninterpretable. It means that the allele labels do not, by themselves, determine how to align the reported associations. A wrong alignment can reverse an outcome association’s sign, and thus change the direction of an MR estimate.

Two separate operations must be kept distinct. First, a strand mapping translates alleles reported on one strand to the other strand. Second, effect-allele alignment puts both association estimates on the same allele. After translating the strand, if the outcome effect allele is the exposure effect allele, retain the outcome coefficient. If it is the other allele, negate the outcome coefficient. This follows because changing the reference allele reverses the direction of the measured allele contrast.

Using Frequencies to Compare Strand Mappings

An effect-allele frequency (EAF) is the proportion of sampled chromosomes carrying the allele for which the association coefficient is reported. For a biallelic variant, the other allele has frequency one minus the EAF. Consequently, frequencies can help discriminate between candidate mappings even when the allele pair cannot.

Suppose the exposure dataset reports effect allele \(E_X\), EAF \(p_X\), and coefficient \(\hat\beta_X\). The outcome dataset reports effect allele \(E_Y\), EAF \(p_Y\), and coefficient \(\hat\beta_Y\). For a palindromic variant, assess both candidate mappings of the outcome alleles: leave the outcome labels as reported, or complement them. Under each mapping, identify which outcome allele corresponds to \(E_X\). Compare the frequency of that allele with \(p_X\). If the outcome effect allele maps to \(E_X\), its frequency is \(p_Y\); if it maps to the other allele, its frequency is \(1-p_Y\). Compare that frequency with \(p_X\). The mapping with the more concordant frequencies is supported, provided the frequency estimates are comparable.

Once a mapping is selected, align the effect alleles. Keep \(\hat\beta_Y\) if the mapped outcome effect allele is \(E_X\); otherwise use \(-\hat\beta_Y\). Allele frequencies help choose a mapping, but they do not replace checking allele labels, variant identity, or the meaning of the reported coefficients.

A useful diagnostic is whether the two candidate mappings lead to clearly different discrepancies. If the exposure EAF is near \(0.5\), its two allele frequencies, \(p_X\) and \(1-p_X\), are nearly equal. In that case, frequency evidence may not distinguish the mappings. The same problem occurs if the outcome EAF is near \(0.5\). There is no universal frequency cutoff that guarantees safe resolution: uncertainty, ancestry, and the quality and comparability of the frequency estimates all matter.

Assumptions Behind Frequency-Based Resolution

Frequency-based harmonization is useful only if the data support its assumptions. Before relying on it, check that the exposure and outcome records refer to the same biallelic variant, with consistent genomic position and genome build, and that each EAF is for the stated effect allele. The datasets should represent comparable ancestry groups, because allele frequencies can differ across populations. Differences in imputation, quality control, or sampling variation can also make frequencies disagree even when the orientation is correct.

In practice, compare the frequencies under both candidate mappings and consider whether the observed difference is small relative to plausible sampling and population differences. Do not treat the smaller discrepancy as proof when both candidates fit poorly. A disagreement may indicate an allele-label error, a variant-matching problem, population differences, or another data-quality issue. If a reliable forward-strand reference or trusted allele annotation is available, it may provide stronger orientation information than comparing frequencies alone.

A Short Justification for the Sign Change

The sign rule follows directly from the definition of an allele contrast. If the reported coefficient is the difference in outcome per copy of allele \(A\), relative to allele \(T\), then reversing the contrast compares \(T\) with \(A\). The difference changes sign. With allele-frequency notation, a biallelic variant has frequency \(p\) for one allele and \(1-p\) for the other. Thus, when the candidate mapping says the outcome effect allele is the exposure’s other allele, both the frequency used for comparison and the sign-aligned coefficient must be interpreted for that other allele.

The sign reversal is an exact recoding of the same two-allele contrast; it is not an additional statistical adjustment. The frequency complement is likewise an identity for a biallelic variant, not an approximation. What is uncertain is which candidate strand mapping is correct when the allele labels are palindromic.

When Frequency Evidence Is Not Enough

Near one half, the two allele frequencies are almost indistinguishable. For example, if the exposure EAF is \(0.490\), the frequency of the other allele is \(0.510\). A modest difference in estimated frequencies, or ordinary variation between samples, can change which candidate mapping appears closer. In this setting, frequency alone is weak evidence. A similar problem arises if both mappings have similar discrepancies, if the samples have different ancestry, or if the EAF is imprecise.

When orientation cannot be resolved confidently, the conservative choice is generally to exclude the ambiguous variant from the harmonized analysis, or to use independently verified strand information if it is available. Exclusion should be recorded and reported, because it changes the set of variants contributing to the analysis. Retaining a variant with an unresolved sign is not a neutral choice: it can move a combined estimate in either direction.

Diagnostics, Alternatives, and Common Pitfalls

A reliable harmonization check makes the choice visible rather than silently applying a sign change. For every palindromic variant, record the allele pair, effect alleles, EAFs, candidate mappings, frequency discrepancies, chosen orientation, and whether the outcome coefficient was reversed. Investigate variants with poor frequency agreement under both mappings. Also check for inconsistent genome builds, duplicate or mismatched variant identifiers, and records whose stated alleles or EAFs do not agree with the underlying data. These checks help detect errors that frequency comparison cannot repair.

Common mistake: “The allele pair is A/T in both datasets, so the strands must match and the outcome coefficient can be kept.” This does not follow. A/T remains A/T after strand complementation, so the matching pair provides no strand evidence. Instead, compare candidate mappings using reliable frequency information or independent strand annotation, align the effect alleles under the supported mapping, and exclude the variant if the evidence remains ambiguous.

A second mistake is to treat whichever frequency discrepancy is smaller as conclusive, even when both are large or nearly equal. The comparison is informative only when its assumptions are credible and the candidate mappings are meaningfully separated. Frequency matching is a diagnostic aid, not a guarantee of correct harmonization.