The Normal You Never Sequenced

Zetobit

Bioinformatics Insight Series  ·  No. 16

The Normal You Never Sequenced

Tumor-only panels lean on population databases to subtract the germline. The subtraction is incomplete — and how incomplete it is depends on whose genome you are filtering.

By Kanna Nandakumar, PhD  ·  Founder & CEO, Zetobit

The matched normal is the control you wish you had. Sequence the tumor, sequence the patient's blood, subtract one from the other, and what remains is somatic by construction. It is clean, it is principled, and in a great deal of real clinical work it is simply not available. The block is archival and the patient is long since lost to follow-up. The consent covered the tumor and nothing else. The budget covered one library, not two. So the sample goes through tumor-only, and somewhere in the pipeline a filter stands in for the control that was never collected.

That filter is almost always a population database. Call the variants, drop anything that appears in gnomAD above some allele frequency threshold, and treat the remainder as somatic. The logic is sound as far as it goes: germline variants are, by definition, variants that people carry, and a database of 800,000 people should know about the variants people carry.

The trouble is the gap between should and does. And the size of that gap is not the same for every patient.

What the filter actually removes

A population database can only remove a germline variant it has seen before. Every individual carries some number of variants that are, in the database's view, unprecedented — not rare, but absent. Within a single individual's exome, gnomAD analyses found a mean of 27 ± 13 novel coding variants absent from all other individuals in the database, with the count running higher in East Asian and South Asian participants.¹ On average an individual also carries roughly 200 very rare coding variants below 0.1% allele frequency, a number that shifts with ancestry depending on which populations the database represents and on the heterozygosity rate.¹

Those private variants are the ones the filter cannot touch. They survive into the somatic call set wearing exactly the same clothes as a real somatic mutation: present in the tumor, absent from the database, no matched normal to say otherwise. They are not a rounding error. They are the dominant failure mode of the entire approach.

A private germline variant and a somatic mutation look identical to a filter. The only thing that distinguishes them is a normal sample you don't have.

The number that should worry you

The critical work here comes from a study that put the filtering strategy on a scale and read off the result by ancestry. For samples of European American ancestry, the positive predictive value of the filtering approach ranged from 35% to 62%. For samples of Hispanic, African American, or African ancestry, the same approach ranged from 20% to 40%.²

Read that again with the clinical meaning restored. A positive predictive value of 35% means that roughly two out of every three variants your pipeline calls somatic are not somatic. At 20%, four out of five are wrong. This is not a subtle degradation at the margins. It is a call set in which the errors outnumber the truth, and it is worse for exactly the patients that genomic medicine has historically served worst.

The same analysis quantified the trade the filter is making. The filtering approach achieved better sensitivity — mean true positive rate 87%, range 78–96% — than a dedicated tumor-only caller at 52% mean. But its precision was far worse: mean PPV 35% against 75% for the tumor-only caller, which leveraged allele frequency differences rather than database membership alone.² Filtering catches nearly everything real. It also catches an enormous amount that isn't.

Figure 1. Reported positive predictive value ranges for somatic calls made without a matched normal. Database filtering alone loses roughly half its precision for European-ancestry samples and considerably more for others; a caller that models variant allele frequency rather than relying on database membership recovers much of it, at a cost in sensitivity. Ranges from Halperin et al. (2017).²

Why ancestry drives the failure

The mechanism is not subtle, and it is not biological in any interesting sense. It is a sampling problem.

A variant's presence in a population database is a function of how many people like the patient were sequenced into it. The number of private variants an individual carries depends on ancestry, driven by undersampling of some populations within databases together with population-specific characteristics such as admixture and recent rapid expansion.² More private variants means more variants the filter cannot see, which means more germline leaking into the somatic call set.

gnomAD has moved substantially in the right direction. The v4 release covers 807,162 individuals — 730,947 exomes and 76,215 genomes on GRCh38 — and adds roughly threefold additional global diversity through the inclusion of about 169,000 individuals of non-European genetic ancestry.^3,4 That is a genuinely large improvement over v2 and v3 combined. But scale alone does not close the gap: even at v4's size, the residual private-variant burden remains unevenly distributed, and there is evidence from an adjacent problem that representation matters more than raw count. In work on gene intolerance metrics, a score trained on 43,000 multi-ancestry exomes outperformed the same score trained on a nearly tenfold larger set of 440,000 non-Finnish European exomes.⁵

The practical read

If your tumor-only pipeline reports a single validation figure — one sensitivity, one PPV, one FDR — that number was measured on some cohort, and it does not transfer to a patient whose ancestry is underrepresented in that cohort or in your filtering database. The pipeline's performance is not a property of the pipeline. It is a property of the pipeline and the patient jointly.

The other direction: variants the filter should not have removed

Everything above concerns germline that survives filtering. The reciprocal error gets less attention and is arguably more dangerous, because it is silent.

During filtration, true somatic variants that happen to match germline variants in population databases can be inadvertently removed, producing false negatives.⁶ Recurrent hotspots are the obvious exposure — a somatic mutation at a position that also segregates as a rare germline variant in some population will be filtered out on database membership alone, no matter that in this patient it arose in the tumor. The standard mitigation is to rescue variants reported in COSMIC, an approach many groups have adopted specifically to avoid discarding cancer-related somatic mutations this way.⁶

A rescue list helps. It also means your filter's behavior now depends on the completeness of a second database, with its own ascertainment biases, layered on top of the first.

Germline findings you did not go looking for

There is a third problem, and it is the one with a consent form attached. Tumor-only sequencing does not only fail to remove germline variants — it detects them, including the ones that matter to the patient's family.

In an MD Anderson study of 1,000 cancer patients sequenced on a 202-gene panel with matched germline samples available, focused analysis of 19 cancer-predisposing genes found that about 5% of pathogenic mutations identified in the tumor were in fact germline.⁷ The exposure scales with panel size. Reported rates of incidental germline pathogenic mutations rose from 6.4% on a 26-gene panel to 12.6% on a 93-gene panel and 15.7% on a 187-gene panel.⁷

Running the inference in reverse is no more reliable. Compared against clinical genetic testing in a cohort of more than 21,000 cancer patients, tumor-only sequencing missed 10.5% of clinically actionable pathogenic germline variants in cancer susceptibility genes — 18.8% in mismatch repair genes, 12.8% in DNA damage response genes, and 7.3% in homologous recombination deficiency genes — for an overall sensitivity of 89.5%. Exonic SNVs and small indels were mostly detected; germline copy number variants, intronic variants, and repetitive element insertions largely were not.⁸

And the intuitive VAF heuristic — call it germline if it sits near 50% — is weaker than it looks. Variants with allele fractions close to 50% are assumed to be germline heterozygous, but tumor purity and somatic deletion or amplification events shift tumor allele fractions and complicate that assumption.⁹ A germline variant in a region of copy-neutral LOH can present near 100%. A germline variant in an amplified region can present anywhere. The heuristic works on a purity-corrected, copy-number-aware view of the genome, and not otherwise.

What to do instead

None of this argues against tumor-only sequencing. It argues against tumor-only sequencing that pretends a database is a control.

Model allele frequency, don't just filter on membership

The single largest available gain is moving from "is it in the database?" to "given tumor purity, local copy number, and observed VAF, what is the posterior probability this is somatic?" That is the design principle behind VAF-aware tumor-only callers, and it is what buys back the precision that filtering gives away.² Combining a filtering approach with a tumor-only caller improves positive predictive value further than either alone² — the two make different errors, which is exactly when combining helps.

Stratify your validation by ancestry

A single headline PPV is a summary statistic over a cohort composition you did not choose and your patients do not match. Report performance by ancestry group, and if the cohort was too homogeneous to support that, report that instead of a number that implies more than it knows.

Set the AF threshold deliberately

A global cutoff — drop everything above 0.1% — treats a variant common in one population and absent from another as though the database's aggregate frequency were the relevant quantity. Population-specific maximum allele frequencies exist for this reason. Use them, and be explicit about what a variant absent from the database actually licenses you to conclude, which is less than it feels like.

Plan the germline referral path before you need it

Tumor-only sequencing will surface probable germline pathogenic variants whether or not your workflow has anywhere to put them. ESMO's Precision Medicine Working Group noted wide disparity in both the extent of systematic germline-focused analysis on tumor sequencing data and in which variants trigger follow-up germline testing.¹⁰ There is real evidence that a defined process changes outcomes: after one center implemented a formal tumor-only data review and genetic counseling referral process, the share of patients with a detected pathogenic germline variant rose from 1.4% to 7.5%.⁹ That is a sixfold change from process, not from chemistry.

Confirm before it becomes a clinical claim

An orthogonal germline test on a normal sample resolves the question the pipeline structurally cannot. When a probable germline finding is actionable, the confirmatory test is not a nice-to-have.

The honest framing

Tumor-only variant calling is a reasonable answer to a real constraint. Archival tissue is what exists. Budgets are what they are. The failure is not in doing it — the failure is in reporting its output as though a matched normal had been involved.

The population database is not a control. It is a prior, and a prior with a known and quantified demographic skew. Treating it as a control produces a call set whose error rate you have not measured, varying with a patient characteristic you did not adjust for, in a direction that compounds the disparities the field already has.

The database can tell you a variant is common. It cannot tell you a variant is not yours.

Design the pipeline around that sentence and most of the rest follows: model the allele frequencies, stratify the validation, rescue the hotspots, route the germline findings, confirm what matters. None of it makes the missing normal appear. All of it makes its absence something you have accounted for rather than something the report quietly assumes away.

Zetobit builds CAP/CLIA-compliant NGS pipelines and variant interpretation systems for biotech, pharma, and academic groups. If you are running tumor-only workflows and want the error model characterized rather than assumed — get in touch.

References

1. Gudmundsson S, et al. Variant interpretation using population databases: Lessons from gnomAD. Human Mutation, 2022.
2. Halperin RF, et al. A method to reduce ancestry related germline false positives in tumor only somatic variant calling. BMC Medical Genomics, 2017;10:61.
3. Chen S, et al. gnomAD v4.0 release notes. Broad Institute, November 2023.
4. Gross AM, et al. Analysis of >800,000 diverse sequenced humans in gnomAD improves clinical interpretation and provides insight into gene function. Genetics in Medicine Open, 2024.
5. Ren Z, et al. Diverse ancestral representation improves genetic intolerance metrics. 2025.
6. Jones W, et al. Somatic Tumor Variant Filtration Strategies to Optimize Tumor-Only Molecular Profiling Using Targeted Next-Generation Sequencing Panels. The Journal of Molecular Diagnostics, 2018.
7. Wong M, et al. BRCA sequencing of tumors: understanding its implications in the oncology community. Chinese Clinical Oncology, 2020.
8. Mandelker D, et al. Comprehensive assessment of germline pathogenic variant detection in tumor-only sequencing. Annals of Oncology, 2022.
9. Bombard Y, et al. A Clinical Approach to Detecting Germline Pathogenic Variants From Tumor-Only Sequencing. JNCI Cancer Spectrum, 2020.
10. Mandelker D, et al. Germline-focussed analysis of tumour-only sequencing: recommendations from the ESMO Precision Medicine Working Group. Annals of Oncology, 2019;30(10):1221–1231.