Question 9 of 9 Advanced Practical
Design a variant-calling validation strategy for a low-frequency mutation assay.
Short answer
Validate low-frequency variant calling with materials and replicates that probe the full workflow at its intended VAF, DNA input and sequence-context limits, not only software benchmark scores.
Why
At low variant allele fractions, polymerase errors, oxidative damage, mapping artifacts and stochastic molecule sampling compete with true signal. Raw coverage is not equivalent to independent original molecules. A 0.5% variant can be invisible in a low-complexity sample.
How to work it through
- Define clinical or research use, target regions and detection threshold.
- Build truth materials using characterized variants diluted across multiple VAFs and genomic contexts, plus negative and difficult-matrix samples.
- Evaluate extraction, library prep, UMI consensus, sequencing and calling end-to-end across runs and operators.
- Estimate sensitivity, false-positive rate, precision and LOD with confidence intervals.
- Separately measure reproducibility near the boundary and confirm with orthogonal methods when feasible.
An example
An assay reports 10,000 overlapping reads, yet only 200 distinct molecules. A variant seen in two read families has much weaker support than 100 raw variant reads suggest. Validation must account for library complexity.
Where the answer stops holding
Orthogonal technology may have a worse LOD than the NGS assay, so disagreement is not automatically disproof. FFPE damage and homopolymer regions need dedicated challenges. Germline GATK thresholds are not universal somatic cutoffs.
A strong answer covers
- Validate low-frequency variant calling with materials and replicates that probe the full workflow at its intended VAF, DNA input and sequence-context limits, not only software benchmark scores
- Define clinical or research use, target regions and detection threshold
Common mistakes
- Publishing only overall accuracy from high-VAF controls
- Treating different callers as fully independent validation
- Ignoring no-call regions
The interviewer might ask next
How would you calculate confidence in a 0.2% allele call with only 500 original molecules?