Variant-Level Quality Filters For Structural Variants

Why Filter Structural-Variant Calls?

A structural-variant call is an inference from sequencing evidence, not a direct observation of a biological event. Repetitive sequence, alignment ambiguity, local coverage variation, and limitations of a caller can all produce convincing-looking but incorrect records. Conversely, a real event may have weak or incomplete read support. Variant-level quality filters make this uncertainty explicit: they define which candidate records have enough evidence for the analysis at hand.

The goal is not to find a threshold that eliminates every questionable call. A stricter threshold can remove artifacts, but it can also discard real variants, including variants whose evidence is weaker because of event size, genomic context, or sequencing characteristics. Those two errors have different consequences for association analysis: false calls can create apparent signals, while missed true calls can reduce or distort them.

Evidence Dimensions and Their Limits

Quality is not a single property. Several partly independent evidence dimensions can support or challenge a call. The relevant evidence and its interpretation depend on the structural-variant class and on the calling method.

• Call confidence. A caller may report a score or a probability-like quantity reflecting how well the data fit its calling model. A score is not automatically a calibrated probability that the event is real. Its scale and meaning are caller-specific, so a threshold should be justified for the particular software and data.
• Breakpoint support. Reads that span a junction, split across the breakpoint, or have paired ends consistent with the proposed event can support its location and structure. A high count is not conclusive if the region is difficult to map; a low count may be expected for an event or library with limited informative reads.
• Read depth. For a copy-number event, normalized depth can provide evidence that the number of copies differs from the reference state. For example, a deletion should generally have lower depth across its interval than nearby expected sequence, while a duplication should generally have higher depth. Local coverage variation can obscure this pattern, and read depth alone may not establish a breakpoint.
• Call consistency. Evidence may agree across multiple samples carrying the call, supporting reads, or independent call-generation procedures. Consistency is informative only when the sources being compared are sufficiently independent and comparable. Repeated use of the same underlying evidence does not amount to independent confirmation.

For these reasons, a single confidence score should rarely be treated as a complete quality assessment. A more defensible rule specifies which evidence is required and why. For instance, a deletion filter might require adequate confidence and breakpoint support, and also require depth to be compatible with a deletion. A record with one very strong metric but contradictory evidence may deserve review rather than automatic retention.

Defining a Filter Before Association Analysis

Write the decision rule in terms of quantities that are available and interpretable for the data being analyzed. One simple rule is a conjunction of thresholds: a record passes only if every required criterion is satisfied. If \(C\) is a confidence measure, \(B\) is breakpoint-support count, and \(R\) is a normalized depth ratio, an illustrative deletion rule might be

Here \(c_0\), \(b_0\), and \(r_0\) are thresholds chosen for the caller, event class, and sequencing data; \(\mathbb{1}\{\cdot\}\) equals 1 when all conditions inside the braces hold and 0 otherwise. The ratio \(R\) must be defined consistently, such as normalized depth in the proposed interval relative to an expected reference level. This example is a form for a rule, not a universal prescription. A duplication would generally need a different depth criterion, and different evidence might be needed for other event classes.

Thresholds should be chosen without using the association result they are meant to assess. If a threshold is adjusted until a desired association appears or disappears, the resulting estimate and its uncertainty no longer reflect a prespecified analysis. When feasible, evaluate candidate rules against an independent validation set in which event status has been assessed by a method not used to make the original calls.

Quantifying the Trade-Off With Validation Data

A filter can be evaluated by comparing its decisions with an independent reference assessment. Let \(T\) denote that a candidate is truly present and \(F\) denote that it passes the filter. Sensitivity is \(P(F\mid T)\), the fraction of true events retained. Specificity is \(P(\text{fail}\mid \text{not }T)\), the fraction of non-events rejected. Positive predictive value (PPV) is \(P(T\mid F)\), the fraction of retained candidates that are true events. Sensitivity and specificity describe performance conditional on truth status; PPV also depends on how common true events are among candidates.

The dependence of PPV on event prevalence follows directly from the definition of conditional probability. If \(\pi=P(T)\), sensitivity is \(P(F\mid T)\), and specificity is \(P(\text{fail}\mid \text{not }T)\), then the probability that a retained candidate is true is

The denominator adds the probability of retaining a true event to the probability of retaining an artifact. For example, if sensitivity is 0.90, specificity is 0.80, and only 10% of candidate records are true events, then

Thus, even with 90% sensitivity, only about 33.3% of retained candidates are expected to be true under these assumptions. A filter's performance in a validation set should not be summarized by sensitivity alone.

How Filtering Can Change an Association Signal

A false-positive variant can create an association signal if its calls differ between phenotype groups. Such a difference can arise, for example, when an artifact is more likely to be called in one group than another. In contrast, false-negative calls change the observed carrier counts and can reduce power or bias an effect estimate. The direction and size of bias depend on the error pattern; it is not correct to assume every filter simply moves an estimate toward zero.

The following examples use odds ratios to show how call errors can alter an association estimate. An odds ratio here compares the odds of a recorded variant call in cases with the odds in controls. These calculations illustrate the effect of call classification; they do not by themselves establish statistical significance or biological causation.

Practical Choices and Common Mistakes

A defensible filtering strategy connects each threshold to the event type, evidence-generation process, and intended analysis. Before applying a rule, specify how its metrics are computed, which records it can exclude, and whether a failed variant is discarded entirely or retained for review. Use independent validation where possible, and report how many records pass each criterion. Where the data allow it, examine whether the filter's decisions are associated with phenotype or study design features; an unexpected imbalance can indicate that technical error is entangled with the comparison of interest.

A common mistake is to assume that a stricter threshold must make the association more trustworthy. The wrong reasoning is: “The strict rule retains fewer calls, so the remaining association must be real.” Fewer calls may mean fewer artifacts, but it may also mean that real events with weak evidence were preferentially removed. The correction is to assess both false-positive and false-negative behavior using validation evidence, then examine how plausible filtering choices affect the association estimate. The filter should be selected on technical grounds and applied consistently, not tuned to deliver a preferred result.

• Do not treat scores as interchangeable. A numeric score from one caller cannot be assigned another caller's threshold without checking its definition and calibration.
• Do not rely on one metric when evidence conflicts. Strong confidence paired with incompatible depth or weak breakpoint support calls for scrutiny, not automatic acceptance.
• Do not assume one cutoff fits every event class. Deletions, duplications, and other structural events can produce different expected evidence patterns.
• Do not interpret a passing call as proof. Passing means only that the record meets the stated rule under the available data.
• Do not choose thresholds using the desired association result. That practice makes the reported signal depend on outcome-driven analysis choices.