Breakpoint Harmonization Across Structural-Variant Calls

Why Breakpoint Harmonization Matters

A structural variant (SV) may be reported differently by two sequencing pipelines even when both calls describe the same underlying event. One caller may place a deletion boundary a few bases earlier, another may represent the same event with a slightly different interval, and a third may fail to call it in a particular sample. Conversely, two calls can overlap substantially while representing different events or different copy-number changes. Harmonization is the process of applying explicit rules to decide which calls should be treated as the same event for comparison across samples or studies.

The decision matters before comparing frequencies or constructing an association data set. If one event is split into several labels, its carriers may be divided across analyses. If distinct events are merged, carriers of different changes may be treated as if they carried the same variant. A reliable analysis therefore defines matching criteria in advance, documents them, and preserves the original calls so that a harmonized grouping can be checked.

Define the Matching Rule Before Comparing Calls

For interval-based calls, use a consistent reference genome assembly and coordinate convention. In this tutorial, intervals are 1-based and closed: an interval \([a,b]\) includes both endpoint bases and has length \(b-a+1\). Two calls cannot be compared reliably if their coordinates refer to different assemblies or if one file's coordinate convention has not been converted to the other's.

At minimum, compare chromosome and event type. A deletion should not match a duplication merely because their intervals coincide. For copy-number events, the direction of dosage change is relevant: a loss and a gain are not the same event. Inversions and insertions also have distinct representations and should not be matched to deletions or duplications solely by interval similarity. If a study distinguishes subtypes or orientations that matter to the scientific question, the rule should require those to agree too.

For two calls on the same chromosome with intervals \(A=[a_1,a_2]\) and \(B=[b_1,b_2]\), define their overlap length \(L_{\mathrm{ov}}\) and reciprocal overlap fractions as follows. The fractions express how much of each call is covered by the other.

A common interval criterion requires both \(R_A\) and \(R_B\) to be at least a prespecified threshold \(\rho\). Requiring both fractions guards against a short interval lying entirely inside a much longer one: the shorter interval could be fully covered even though only a small fraction of the longer call overlaps. A complementary breakpoint criterion may require the absolute differences between the left endpoints and between the right endpoints to be no greater than a tolerance \(\delta\) bases.

These are rule components, not universal biological constants. The threshold \(\rho\) and tolerance \(\delta\) should be chosen for the event class, sequencing resolution, and study purpose, then applied consistently. Some analyses use reciprocal overlap without endpoint constraints; others use endpoint tolerance or confidence intervals around uncertain breakpoints. The chosen policy must say which criteria are required and how uncertainty is handled.

Breakpoint Tolerance, Uncertainty, and Event Representation

Breakpoint tolerance acknowledges that estimated boundaries can vary between callers or samples. It is not a license to ignore location. A large \(\delta\) can cause nearby but distinct events to be grouped, while a very small \(\delta\) can split calls for one event when the data do not support base-level precision. Interval overlap and endpoint tolerance provide related but not identical information: a long interval can have substantial reciprocal overlap despite a boundary difference that exceeds a chosen tolerance.

Some callers report breakpoint confidence intervals rather than one precise endpoint. A confidence interval describes uncertainty in the breakpoint estimate; it is not the same thing as the called variant's affected interval. A harmonization protocol can compare point estimates with a fixed tolerance, or can define a rule for overlapping breakpoint uncertainty intervals. These approaches should not be mixed silently. For example, calling two breakpoints compatible whenever their uncertainty intervals overlap can be permissive when those intervals are wide, so the uncertainty widths and the policy should be retained for review.

Interval overlap is not appropriate for every event representation. Insertions may be represented by an insertion site rather than a span of affected bases. Such calls need a site-distance rule, and, when available and relevant, agreement in inserted sequence or insertion subtype. A zero-length or one-base placeholder should not be treated as an ordinary interval and evaluated with reciprocal overlap as though it had meaningful span. The same principle applies to complex events: the representation and matching criteria must reflect the event structure being compared.

Resolve Candidate Matches Without Losing Distinctions

A call can have more than one plausible match, especially in regions with recurrent events or dense call sets. Pairwise similarity is not necessarily transitive. Call A may satisfy the rule against B, and B may satisfy it against C, while A does not satisfy it against C. Consequently, it is unsafe to assume that every chain of pairwise matches defines one unambiguous event group.

A representative-based rule makes the decision reproducible but can depend on which call is chosen as representative. A clustering method can also have choices that affect group membership. Whichever policy is used, retain the pairwise comparisons and flag groups whose members are not mutually compatible. Avoid silently taking connected components of a match graph: that procedure joins any calls linked by a path, including the A-to-C mismatch in the example.

The following sequence makes the decision process explicit.

Consequences for Association Grouping

Harmonization determines which calls can contribute to the same event-level genotype or carrier indicator. Suppose two studies report different breakpoint estimates for one deletion in a sample. If those calls are harmonized as one event, the sample should not be counted as carrying two independent events merely because it has two records describing the same call. Conversely, if two calls fail the declared criteria, merging them can assign a sample to the wrong event group.

This decision is upstream of association testing. It defines the variant labels and the mapping from call-level records to analysis-level observations; it does not itself establish an association or a causal effect. When converting to a genotype or dosage representation, follow the event's copy-number coding and the analysis specification. Do not add dosages simply because multiple records were matched to one event: first determine whether they are duplicate descriptions of one call or distinct alleles, and apply a consistent sample-level rule.

For cross-study comparisons, apply the same harmonization policy to all studies or explain any justified differences. Report the assembly, event-type requirements, overlap threshold, breakpoint tolerance, uncertainty policy, and ambiguity resolution procedure. Sensitivity checks using reasonable alternative thresholds can reveal whether an association grouping depends heavily on an arbitrary boundary, but they do not replace a primary rule defined before examining the results.

Common Mistake and Key Takeaway

• Wrong reasoning: “These calls overlap, so they must be the same variant.” Overlap alone ignores event type, the fraction of each interval covered, boundary differences, coordinate conventions, and competing matches.
• Correction: Require compatible event classifications and apply a declared location rule, such as reciprocal overlap plus endpoint tolerance. Treat ambiguous matches explicitly and preserve the original calls.
• Wrong reasoning: “If A matches B and B matches C, all three are automatically one event.” Pairwise matching need not be transitive.
• Correction: Specify the grouping policy and check group-wide compatibility rather than assuming that a chain of matches defines one event.