Sequencing Errors That Mimic Rare Variants

Why a Few Alternate Reads Need an Explanation

At a rare-variant site, even a small number of reads carrying an alternate base can look important. But a small alternate-read count can arise when the true DNA carries the reference base: base-calling mistakes may substitute the alternate base, contaminating DNA may contribute it, or PCR may amplify one error or one molecule into many reads. These mechanisms matter because many rare-variant calls are supported by only a few observations, and technical artifacts can therefore resemble genuine low-frequency alleles.

The key is to ask what the observations represent. A read is a sequenced output, not necessarily an independent DNA molecule. Several reads can descend from the same original molecule, especially after PCR. Evidence is stronger when alternate observations occur on multiple independent molecules and persist in separately prepared material, and weaker when they show a pattern expected from a specific technical mechanism.

Base-Calling Errors as a Probability Model

Suppose the true sequence at a site is the reference base. Let \(q\) be the probability that one read from a distinct, independently sampled original DNA molecule is called as the particular alternate base. If there are \(n\) such reads, and each has the same error probability \(q\) independently of the others, the number of alternate calls \(X\) follows a binomial model:

This is an error-only model: it assumes the DNA is truly reference at the site and that the specified error probability applies to each observation. In practice, \(q\) should represent the chance of a call to this particular alternate base, not automatically the total probability of any mismatched base. It may depend on the base substitution, sequence context, read quality, and sequencing process. If reads are PCR duplicates, errors are correlated, or error rates differ across reads, the simple binomial model may not be adequate.

Read-level patterns help assess whether the assumed model is plausible. For example, alternate calls concentrated in low-quality reads, at a particular read position, or on only one sequencing orientation can be consistent with technical error. Support distributed across high-quality reads and independent molecules is more reassuring, though it does not by itself prove that the allele is real. A striking tail probability from an overly optimistic error model should not be mistaken for strong evidence if that model fails to represent the data-generation process.

Contamination Can Introduce Genuine Alternate Bases

Contamination differs from a base-calling error: the alternate base may be correctly read, but it comes from DNA that is not the intended sample. A small contribution from another individual can produce a low alternate fraction at sites where that source carries an allele absent from the intended sample. A useful first calculation relates the contaminating fraction to the expected alternate fraction.

For example, suppose a fraction \(c\) of DNA molecules comes from a contaminating donor who is heterozygous at a site, while the intended sample is reference homozygous. If the DNA molecules contribute equally and there is no allele-specific amplification, the donor contributes half alternate-bearing molecules. The expected alternate fraction is therefore \(c/2\). This simple relationship is conditional on those assumptions; other donor genotypes or unequal amplification give different expectations.

Contamination can be introduced at different stages, so comparisons among samples and controls processed together can be informative. A low-frequency allele in a sample that also appears in a negative control, or at a similar fraction in several samples sharing a likely source, raises concern. Conversely, not seeing an allele in a control does not guarantee that the sample is uncontaminated: the contaminant may be sparse, and its alleles may not be observed at every site.

PCR Can Inflate the Apparent Number of Supporting Reads

PCR produces many copies from a limited set of starting molecules. If an error occurs during amplification, copies descended from the error-containing product can make the erroneous base appear repeatedly. Even without a new error, many reads can arise from one original molecule. Consequently, the raw alternate-read fraction can differ substantially from the fraction among distinct starting molecules.

Molecular identifiers, when used and reliably assigned before amplification, can help group reads by their likely original molecule. They do not make every family infallible: an error early in amplification can dominate a family, and errors during identifier assignment or grouping can also mislead. The useful question is not simply whether many reads agree, but whether independent molecules support the allele and whether the family-level evidence is coherent.

Replicates: What They Can and Cannot Establish

A replicate is most useful when it is genuinely independent for the technical question being asked. Resequencing the same amplified library can check some sequencing-stage behavior, but it does not independently test an error introduced before or during that library’s amplification. Preparing a new library from a separate aliquot tests more of the process; an independent extraction can test still earlier steps. The design should therefore be stated before interpreting agreement or disagreement.

Replicate comparisons also depend on the biological model. A germline heterozygote, a low-level mosaic allele, and a contaminating allele imply different expected fractions. A variant absent from one replicate may be inconsistent with a well-sampled germline heterozygote yet plausible when the expected fraction is low and the number of independent molecules is small.

A Practical Diagnostic Sequence

A disciplined review separates three questions: are the reads technically credible, are the observations independent, and does the signal recur in material that was prepared independently? The sequence below keeps those questions distinct.

Common Mistake and Why This Matters

Common mistake: “Four alternate reads have a very small probability under the sequencing-error model, so the variant must be real.” The reasoning overlooks the model’s conditions. The tail probability only describes alternate calls under the particular assumed error rate, independence, and true-reference null. It does not account for contamination or PCR dependence, and it is not the probability that the call is false.

Correction: Treat the error-model calculation as one diagnostic component. Ask whether its assumptions fit the observations, determine how many independent molecules support the allele, and compare controls and independent preparations. A low-frequency call can be technically suspicious even when a simple error-only probability is small; conversely, one negative replicate does not automatically disprove an allele when few molecules were sampled.

Reliable distinction among artifacts matters because false rare-variant calls can distort interpretation and subsequent analyses. The goal is not to declare every low-count observation false, but to match the strength of the conclusion to the evidence: read quality, molecular independence, plausible contaminant patterns, and reproducibility all contribute. These checks address base-calling, contamination, and PCR; other technical explanations require their own diagnostics.