FFPE Artifacts as a Pre-Analytic Axis: Why a Low-VAF C>T Call Cannot Be Read as Biology

Zetobit · Bioinformatics Insight Series Clinical Genomics

Bioinformatics Insight Series

FFPE Artifacts as a Pre-Analytic Axis: Why a Low-VAF C>T Call Cannot Be Read as Biology

Formalin fixation is a chemical reaction performed on the specimen before any sequencing decision is made. It deaminates cytosine, and the resulting C>T changes enter the pipeline as ordinary reads carrying ordinary quality scores. Nothing downstream is malfunctioning when they are called. The consequence reaches further than a few spurious variants: because the artifact concentrates in one substitution class at low allele fraction, it lands directly on top of the measurements that depend on counting low-fraction substitutions — tumor mutational burden, mutational signatures, and microsatellite instability.

Kanna Nandakumar, PhD · Zetobit, LLC

A somatic pipeline validated on high-quality DNA can be applied without modification to an FFPE block and will run to completion. Coverage will be adequate, duplication rates ordinary, mapping rates normal. The variant list will be longer than expected, and the excess will sit at low variant allele fraction. There is no flag in the BAM for a base that was chemically altered in a cassette three years before sequencing.

This is a different class of problem from the ones a pipeline is built to catch. Sequencing errors are a property of the instrument, and alignment errors are a property of the reference. A deamination artifact is a property of the specimen — introduced by a decision made in a histology lab, usually by someone who was not thinking about sequencing at all, and frequently before the assay existed. By the time the DNA reaches library prep, the artifact is a real chemical feature of a real molecule. It amplifies faithfully. It sequences faithfully. It is not noise in the signal-processing sense; it is a mutation that happened in a cassette rather than in a patient.

The series has treated pre-analytic effects before, most directly in the piece on distinguishing mosaicism from sequencing artifact, where FFPE damage appears as one of several sources of low-VAF noise. This piece takes the other direction: not how to keep FFPE artifacts out of a variant call, but what their presence does to the aggregate measurements a report is built on.

What formalin actually does to a cytosine

Buffered formalin cross-links intracellular macromolecules — DNA to DNA, DNA to RNA, DNA to protein — and those cross-links stall polymerases during library amplification, which is why FFPE libraries have depleted template diversity relative to fresh material.¹ The cross-linking is the damage most people associate with FFPE, and it mostly costs yield.

The damage that costs accuracy is hydrolytic deamination. Formalin deaminates cytosine to uracil; during amplification a polymerase incorporates adenine opposite that uracil, and the resulting read carries a T where the genome has a C.¹ Because the complementary strand produces the mirror event, the artifact is conventionally written as C:G>T:A.

Two properties make this consequential rather than merely annoying.

It is dose-dependent on fixation time. In samples fixed for 2, 15, 24 and 48 hours, FFPE-only C>T counts rose with exposure — and rose far more steeply in unrepaired than in repaired material.¹ Fixation duration is a pre-analytic variable that no bioinformatics step observes and that is rarely recorded in a form the pipeline can read. In a separate series of fixation conditions, C:G>T:A changes accounted for between roughly 21% and 54% of detected substitutions depending on condition, with a pronounced increase past three days of fixation, and acidic formalin produced more variant calls than neutral or basic.²

It is stochastic and sparse. Only a small fraction of cytosines are deaminated, and which ones is essentially random, so the artifacts appear as low-frequency, unpredictable single-nucleotide calls scattered across the target.³ Any given artifact is supported by few molecules. That is precisely what makes the class hard: low frequency and non-recurrent are also the properties of genuine subclonal mutation.

The artifact is not distinguishable from the finding by any property the variant record contains. It is distinguishable only by properties of the class — substitution type, allele fraction, read orientation, and how many independent molecules support it.

The repair step, and the half of the problem it does not solve

The standard mitigation is enzymatic: uracil-DNA glycosylase excises uracil bases before amplification, so the lesion never becomes a miscalled base. Treating FFPE DNA with UDG before PCR produces a marked reduction in C:G>T:A artifacts.⁴

The limitation is chemical, and it is worth being precise about, because it determines what a pipeline still has to handle after the wet-lab has done everything right. Formalin-induced deamination of 5-methylcytosine yields thymine directly rather than uracil.¹ There is no uracil for the glycosylase to excise. Since 5mC in the human genome occurs essentially at CpG dinucleotides, UDG treatment removes most of the artifact burden but leaves behind a residue that is concentrated at CpG sites.

This shows up in the data as a shift in the artifact's shape, not merely its size. In one paired analysis, C>T made up about 72% of FFPE-only discordant mutations in unrepaired samples and about 39% in repaired ones — and the artifacts that survived repair were disproportionately in CpG contexts.¹ The important consequence is that repair changes which biological process the leftover artifact resembles, and that turns out to matter a great deal.

Where it stops being a variant problem: mutational signatures

Mutational signature analysis classifies substitutions by trinucleotide context and decomposes the resulting spectrum into the processes that produced it. It is an aggregate measurement — the signal is the shape of the whole catalogue, so anything that adds mutations with a consistent shape enters as a process.

Formalin's footprint has now been characterized directly. Working from 110 FFPE samples across two studies, the unrepaired-FFPE artifact signature was found to closely resemble COSMIC SBS30 (cosine similarity 0.92), and the repaired-FFPE signature to closely resemble SBS1 (0.90).¹

Read that pairing carefully, because the coincidence is not arbitrary. SBS1 is the clock-like ageing signature, whose accepted aetiology is spontaneous deamination of 5-methylcytosine in vivo. The repaired-FFPE artifact is deamination of 5-methylcytosine in vitro. The same chemistry runs in a cell over decades and in a cassette over days, and it leaves the same fingerprint. SBS30, which the unrepaired artifact mimics, is the footprint of base-excision-repair deficiency from biallelic NTHL1 inactivation — a finding with hereditary cancer implications.¹

The measured effect on decomposition is substantial. When signature activities were refit on uncorrected profiles, the present/absent classification error rate was roughly 41% for unrepaired and 37% for repaired samples; on corrected profiles the false discovery rate fell to 8–10%.¹ The direction of the errors is what makes this clinical rather than academic: SBS3, the homologous-recombination-deficiency signature that informs PARP inhibitor candidacy, was substantially underestimated in both unrepaired and repaired FFPE.¹ Artifactual C>T mutations crowd the catalogue, and the relative contribution of the signature someone is actually looking for falls.

Tumor mutational burden: the threshold moves with the specimen

TMB is a count of somatic mutations per megabase, compared against a threshold. Every component of that sentence is sensitive to a process that adds low-fraction substitutions.

The practical consequence is that the VAF filter — the parameter that decides which variants are counted — cannot be held constant across specimen types. In a 685-sample clinical series spanning 390 FFPE and 295 frozen specimens, an optimal VAF cut-off of 10% was established for FFPE against 5% for frozen material, with the authors attributing the difference to deamination artifacts inflating TMB in low-quality FFPE DNA.⁵ Both sample type and DNA quality affected the resulting scores.

A doubled VAF threshold is not a cosmetic adjustment. It is a different assay. Raising the floor to 10% to suppress artifacts also discards genuine subclonal mutations between 5% and 10%, and discards real variants in low-purity specimens where clonal mutations sit below 10% simply because tumor content is low. The artifact is controlled by trading away sensitivity — a reasonable trade, but one that changes what the number means and is rarely visible beside it.

This compounds a point the series has made about tumor-only calling: the sources of TMB inflation are additive. Database-based germline filtering already inflates TMB relative to matched-normal filtering, and does so unevenly across ancestry groups. Deamination artifacts add a second inflation with a different mechanism. A TMB near a decision threshold, derived from an FFPE block without a matched normal, has two independent upward pressures on it.

Figure 1Raising the VAF floor from 5% to 10% suppresses the artifact mass but forfeits the true variants that lived between the two thresholds. The distributions here are schematic — they illustrate the overlap that forces the trade-off, and are not drawn from a specific dataset. The 5%-frozen and 10%-FFPE thresholds are the operating points reported for one clinical panel;⁵ the correct thresholds for any assay must be established on that assay.

Microsatellite instability: an indel measurement disturbed by a substitution artifact

MSI is the case where the consequence is most surprising, because MSI is not a substitution measurement at all. NGS-based MSI callers evaluate length variability at homopolymer and microsatellite loci and compare the unstable fraction against a threshold.

The reported result is stark. Applying MSIsensor to a colorectal FFPE sample, 8.3% of microsatellite sites showed somatic changes in the unrepaired sample against 0.23% in the repaired one — and 8.3% exceeds the 3.5% threshold used to call MSI, meaning the unrepaired specimen would have been miscalled as MSI, while the corrected profile classified it correctly as MSS.¹

Two things are worth separating here. The comparison is repaired versus unrepaired aliquots of the same tumor, so the entire difference is attributable to the pre-analytic and repair pathway rather than to biology. And it is a single case, so the magnitude should not be generalized — what generalizes is the mechanism: formalin damage in a low-complexity locus disturbs the very quantity an MSI caller measures, and the disturbance crosses a clinical threshold in at least one documented instance.

Independent evidence points the same direction from a different angle. In a DNA-input titration, the frequency of LOH and MSI artifacts related significantly to template input, and the proportion of LOH artifacts was significantly higher for paraffin-embedded material than for fresh-frozen at comparable input.⁶ Low template diversity — the direct consequence of cross-linking — means each locus is genotyped from few independent molecules, and stochastic events are not averaged away.

Since MSI status gates immune checkpoint inhibitor eligibility, this belongs in the same risk category as a TMB near threshold.

Three aggregate measurements, one pre-analytic cause

Measurement	How the artifact enters	Documented effect	Direction
Tumor mutational burden	Artifactual C>T calls counted as somatic mutations at low VAF	VAF cut-off of 10% required for FFPE vs 5% for frozen on the same panel5	Inflates TMB; controlling it forfeits genuine subclonal variants
Mutational signatures	Artifact spectrum mimics real processes (SBS30 unrepaired; SBS1 repaired)1	~41% / ~37% present-absent error on uncorrected profiles; SBS3 underestimated1	Overstates ageing/BER activity; understates HR deficiency
Microsatellite instability	Damage and low template diversity at homopolymer loci	8.3% unstable sites unrepaired vs 0.23% repaired, against a 3.5% threshold1	Can cross the MSI threshold in the false-positive direction

Effects are drawn from the cited studies and reflect those specimens, panels, and pipelines. The mechanism generalizes; the magnitudes do not.

Why the standard filter is not sufficient on its own

The established computational defence is orientation bias. Deamination affects one strand of a duplex, so the artifact appears preferentially in read 1 versus read 2 in paired-end data — a statistical asymmetry that genuine heterozygous or clonal variants do not show.⁷ GATK's orientation-bias filter was designed for exactly this, alongside the analogous oxidative G>T artifact.

It works, and it should be in every FFPE pipeline. It is also partial. In one benchmarking analysis, the MuTect orientation-bias filter retained true variants at a sensitivity of 0.969 but removed only 40.7% of FFPE artifacts (11,204 of 27,510).⁸ A filter that leaves the majority of the artifact class in place is a component of a strategy, not the strategy.

The deeper structural problem is that the most widely used alternative — a blanket low-VAF cut-off — cannot separate the two populations because it does not distinguish them. Existing FFPE filtering tools tend to label all low-frequency variants as artifacts,⁸ which is why methods work has moved toward classifiers that combine allele fraction with orientation and quality features rather than thresholding on frequency alone.

And it is worth stating plainly what a blanket cut-off costs, because it is the reason this cannot be dismissed as solved: low-frequency C>T mutations also occur genuinely in cancer and can be clinically important.³ The substitution class the artifact occupies is not an empty class.

The counterpoint, which is real

It would be easy to read the above as an argument that FFPE data should not be trusted, and that would be both wrong and unhelpful — clinical archives consist near-exclusively of FFPE blocks, and abandoning them would mean abandoning most of translational oncology.¹

The evidence supports a narrower claim. In a controlled fixation study, mutation calls did not differ significantly between conditions until 48 hours of fixation, at which point C/G>T/A calls rose significantly in untreated DNA — but the allele frequencies of those events remained below the assay's reportable limit of detection, leading the authors to conclude that formalin fixation increases the deamination signature without producing false positives in clinical practice for that assay.⁹ A well-designed assay with an appropriate reporting floor absorbs the artifact.

That finding and the ones above are not in conflict; they describe different regimes. A hotspot assay reporting at 5% VAF on well-fixed tissue is genuinely robust. The failure modes appear when the reporting floor drops toward the artifact band (subclonal calling, MRD-adjacent applications), when fixation was prolonged or acidic, when input DNA is scarce, when blocks are old — and, critically, whenever the output is an aggregate count rather than a single hotspot call. TMB, signature activity, and MSI fraction all sum over exactly the low-fraction territory a hotspot assay excludes by design. The artifact does not need to reach reportable VAF to move a count that includes everything above the noise floor.

Similarly, the recent large-scale evidence deserves acknowledgement: whole-genome sequencing of FFPE specimens at scale has been shown to preserve clinical utility.¹⁰ The conclusion is not that FFPE data are unusable. It is that FFPE status is a parameter of the measurement, and it belongs where parameters belong — in the validation, in the thresholds, and in the report.

What this means for a pipeline

Treat fixation metadata as an analysis input

Fixation time, formalin pH, block age, and whether the extraction included UDG repair are all determinants of the artifact burden,^1,2,4 and none is recoverable from the FASTQ. Repair status in particular changes which COSMIC signature the residual artifact impersonates, which means a signature analysis cannot be performed correctly without knowing it.¹ Capture these fields at accessioning. Where a dataset arrives without them, the extraction kit is often the fastest route to inferring repair status.

Quantify the artifact burden per sample, and report it

The C>T fraction at low VAF is measurable in every FFPE run, and it is a per-sample quantity, not a per-batch constant — fixation was done one specimen at a time. A deamination metric alongside coverage and duplication turns an invisible confounder into a QC axis with a threshold. It also makes the failure mode legible: a sample with an anomalous C>T fraction is a sample whose TMB and MSI results deserve a second look before they leave the lab.

Layer the filters; do not rely on the VAF floor

Orientation bias removes a substantial minority of the artifact class,⁸ and enzymatic repair removes most of the non-CpG burden.⁴ Combining wet-lab repair with orientation-aware and context-aware filtering outperforms either alone, and comparative work indicates the enzymatic approach carries the most weight of the available options. Reserve the blanket low-VAF cut-off as a last resort, and recognize that when you use it you have traded subclonal sensitivity for specificity.

Correct the catalogue before fitting signatures

Signature refitting on an uncorrected FFPE profile mis-assigns activities at a rate that makes the output unreliable,¹ and correction tools exist. They have documented limits — correction performs poorly when the true mutation load is low relative to the artifact burden, and it fails when the tumor's genuine profile closely resembles the artifact signature, which is the case for repaired FFPE in MSS colorectal cancer.¹ Knowing when correction will not work is as useful as running it.

Do not compare TMB or MSI across specimen types

This is the operational bottom line. A TMB from an FFPE block and a TMB from frozen tissue are measurements on different scales, produced with different thresholds, carrying different artifact burdens. Comparing them across a cohort — or against a threshold established on the other material — is a category error. Validate specimen types separately, state which was used, and treat the specimen type as part of the result.

The shape of the problem

Everything above has the same structure. A chemical process alters bases in a specimen; the alteration is faithfully preserved through extraction, amplification, sequencing, and alignment; and it arrives at the variant caller indistinguishable, at the level of an individual record, from the biology it mimics. No stage is broken. The information required to separate artifact from finding was never in the data at that resolution — it lives in the substitution class, in the strand asymmetry, in the molecular support, and in the specimen metadata that the pipeline usually never sees.

The field has characterized this thoroughly. The signatures are published, the filters are implemented, the correction tools are open-source, and the mechanism has been understood for over two decades. What is not solved is transmission. A TMB of 11 leaves the pipeline as a number, and the fixation duration, the repair status, the VAF threshold that number required, and the specimen type it was measured on do not travel with it. By the time it reaches a treatment decision, it is a scalar compared against a threshold, and every pre-analytic condition that set its value has become invisible.

Related in this series: the mosaicism piece covers separating true low-VAF variants from noise at the level of an individual call; the tumor-only piece covers the other major source of TMB inflation; the piece on tumor purity and ploidy covers a second set of parameters every somatic number is computed downstream of; and the homopolymer and low-complexity piece covers why indels in exactly the loci MSI callers interrogate are systematically miscalled.

References

1. Guo Q, Lakatos E, Al Bakir I, Curtius K, Graham TA, Mustonen V. The mutational signatures of formalin fixation on the human genome. Nature Communications. 2022;13:4487. https://www.nature.com/articles/s41467-022-32041-5
2. Kim S, et al. Deamination effects in formalin-fixed, paraffin-embedded tissue samples in the era of precision medicine. Journal of Molecular Diagnostics. 2017;19(1):137–146. https://www.sciencedirect.com/science/article/pii/S152515781630188X
3. Mathieson W, Thomas GA. Why formalin-fixed, paraffin-embedded biospecimens must be used in genomic medicine: an evidence-based review and conclusion. Journal of Histochemistry & Cytochemistry. 2020;68(8):543–552. https://journals.sagepub.com/doi/10.1369/0022155420945050
4. Do H, Dobrovic A. Dramatic reduction of sequence artefacts from DNA isolated from formalin-fixed cancer biopsies by treatment with uracil-DNA glycosylase. Oncotarget. 2012;3(5):546–558. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3388184/
5. Tumor mutational burden assessment and standardized bioinformatics approach using custom NGS panels in clinical routine. BMC Biology. 2024;22:29. https://link.springer.com/article/10.1186/s12915-024-01839-8
6. Microsatellite instability (MSI) detection in DNA from FFPE tissues: relation of template input to LOH and MSI artifacts. Journal of Molecular Diagnostics / methods literature. https://www.researchgate.net/publication/301190673_Microsatellite_Instability_MSI_Detection_in_DNA_from_FFPE_Tissues
7. Khan MS, Shih DJH. Filtering sequencing artifacts from FFPE tissues: an orientation-bias-based statistical tool. Briefings in Bioinformatics. 2025. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12699713/
8. Heo D-h, Kim I, Seo H, et al. DEEPOMICS FFPE, a deep neural network model, identifies DNA sequencing artifacts from formalin fixed paraffin embedded tissue with high accuracy. Scientific Reports. 2024;14:2559. https://www.nature.com/articles/s41598-024-53167-0
9. Prentice LM, et al. Formalin fixation increases deamination mutation signature but should not lead to false positive mutations in clinical practice. PLoS ONE. 2018;13(4):e0196434. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5919577/
10. Large-scale analysis of whole genome sequencing data from formalin-fixed paraffin-embedded cancer specimens demonstrates preservation of clinical utility. Nature Communications. 2024;15:7561. https://www.nature.com/articles/s41467-024-51577-2