Disclosure note. Public SEQC2 data are the reproducible benchmark. Translation results from a matched FF/FFPE cohort are rounded and aggregate-only. Patient/sample identifiers, exact variants, sequencing-run identifiers, filenames, and storage or infrastructure details have been removed. Proxy labels are not molecular or clinical truth.

When Sequence Is Not Enough

Formalin-fixed, paraffin-embedded (FFPE) tissue makes archival tumor material available for genomic analysis, but fixation can damage DNA. Some damage survives variant calling and appears as a plausible somatic variant.

The tempting AI question is: “Can a large DNA model recognize the artifact from sequence context?”

Our experiments showed why that framing is incomplete. A real mutation and a fixation artifact at the same locus with the same reference and alternate alleles produce the same alternate DNA sequence. The information that separates them lives primarily in the reads: strand and orientation imbalance, base quality, read position, allele fraction, and caller evidence.

I built and evaluated an FFPE-risk modeling workflow across a public multi-caller benchmark and a de-identified matched fresh-frozen/FFPE cohort. The work produced a practical classifier, an operating-policy analysis, and an important negative result for frozen DNA foundation-model embeddings.

Table of contents

1. Problem
2. Label design
3. Public benchmark
4. Modeling approach
5. Results
6. Why the DNA foundation model did not help
7. Operating policy
8. Product impact
9. Limitations
10. Take-home
11. References

Problem

FFPE artifact filtering is asymmetric:

• missing an artifact can create a false biological story;
• removing a real variant can erase clinically important evidence.

Accuracy at a default threshold is therefore not enough. A useful system must answer:

At a required retention of likely-real variants, what fraction of artifact-enriched calls can be flagged for review?

Label design

Public benchmark labels

The public anchor was the SEQC2 tumor-normal reference system, which provides deeply characterized somatic truth calls across assay and computational conditions.

Calls in high-confidence regions were labeled:

y = 0  if supported by the reference truth set
y = 1  if an FFPE call was not supported by the truth set

This is stronger than labels from unmatched clinical samples, but still has a noise floor near detection limits.

Matched-cohort proxy labels

For translation, matched fresh-frozen (FF) and FFPE samples from the same tumors enabled:

FF-concordant call → likely-real proxy
FFPE-only call      → artifact-enriched proxy

These are technical proxies. Tumor heterogeneity, sampling, coverage, and subclonality can make a real variant appear in only one preparation.

Public benchmark

I unified:

• 587 variant-call files;
• five callers;
• six assay conditions;
• three aligners;
• 54.9 million raw rows;
• 18.6 million labeled rows after removing uncertain labels.

Evaluation was group-held-out by physical sample token. No sample could contribute rows to both training and test sets.

This is essential: row-random splitting would let the model learn sample- or workflow-specific fingerprints and overstate generalization.

Modeling approach

Feature groups

The model used evidence available at or near variant calling:

• alternate and reference read counts;
• depth and variant allele fraction;
• caller confidence;
• strand or read-pair orientation imbalance;
• base and mapping quality;
• local substitution context;
• assay, caller, and aligner metadata when available.

Models

I compared:

• histogram gradient-boosted trees;
• random forests;
• a small sequence-context transformer fused with numeric features;
• frozen Evo2 likelihood and embeddings;
• published FFPE-filtering approaches where the available inputs allowed a meaningful comparison.

Evaluation

The primary metrics were:

ROC AUC  = ranking quality across thresholds
AUPRC    = precision–recall quality for the artifact class
F1       = harmonic mean of artifact precision and recall
Retention = TN / (TN + FP) for likely-real proxies

Results

Public multi-caller benchmark

The performance drop in the reduced model quantifies a deployment trade-off. Caller metadata improves discrimination, but caller-agnostic features are easier to transfer across pipelines.

Mechanistic ablation

In a narrower public experiment, removing the dominant caller-confidence score reduced F1 by only about 0.003. The apparent importance of that field was largely redundant with mechanistic read-level evidence.

This was a useful model-review lesson:

High permutation importance does not prove a feature is load-bearing; correlated evidence must be removed and the model retrained.

De-identified matched-cohort translation

In the full, pre-filter aggregate of more than 85,000 calls, the tree ensemble reached ROC AUC 0.923, while the sequence-context transformer reached 0.919.

After current production-style filters, the remaining deployment population was approximately:

• 16,700 calls;
• 12,200 FF-concordant, likely-real proxies;
• 4,500 FFPE-only, artifact-enriched proxies.

The harder post-filter model reached leave-one-patient-out ROC AUC 0.790. This lower number is the more relevant deployment estimate: easy artifacts had already been removed upstream.

Why the DNA foundation model did not help

On a stratified pilot, frozen Evo2 features achieved:

Evo2 alone was indistinguishable from random ranking.

The reason is information-theoretic, not a failure of scale:

Hreal:      locus L, REF→ALT, observed from a true mutation
Hartifact:  locus L, REF→ALT, observed from fixation damage
 
Sequence input under both hypotheses = identical ALT sequence

A model receiving only the resulting reference/alternate sequence cannot infer how the observed base arose. Adding a larger context window or a more expressive embedding head does not restore evidence that was never present in the input.

Useful signals instead include:

P(artifact | reads, orientation, quality, VAF, caller, context)

The negative result prevented further investment in a category-mismatched foundation-model approach and redirected effort to the evidence that can actually solve the problem.

Operating policy

A classifier becomes useful only when tied to a retention policy.

For a threshold τ:

Flag as FFPE-risk if P(artifact | evidence) ≥ τ

In the post-filter matched cohort:

• at 90% likely-real retention, the model flagged 55.8% of artifact-enriched proxies;
• at 95% retention, it flagged 24.3%;
• at 99% retention, it flagged 3.3%.

There is no universally correct threshold. A review-support tool may tolerate more flags than an automatic filter.

Product impact

The recommended deployment pattern is additive:

Variant caller
    → existing hard filters
    → calibrated FFPE-risk score
    → visible evidence and reason codes
    → expert review

The score should:

• annotate, never silently delete;
• expose its operating threshold;
• retain caller and assay provenance;
• report calibration drift by batch and workflow;
• preserve the original call and evidence;
• be prospectively validated before clinical use.

This creates a practical bridge between ML research and diagnostic-quality engineering: the model optimizes review effort while preserving human and pipeline controls.

Limitations

• FF-concordant and FFPE-only classes are proxy labels.
• The matched-cohort result is retrospective and rounded for publication safety.
• Performance varies by caller, assay, aligner, and filtering stage.
• Cross-caller transfer should not be assumed when feature definitions differ.
• Published comparator tools require different input evidence; VCF-only comparisons can be incomplete.
• The operating point requires prospective validation and clinical/QC governance.
• This system estimates technical artifact risk, not pathogenicity or actionability.

Take-home

• The public multi-caller model reached ROC AUC 0.959 and AUPRC 0.920 with group-held-out evaluation across 18.6 million labeled rows.
• The post-filter matched-cohort task was harder but operationally relevant: AUC 0.790.
• At 90% likely-real retention, the score flagged 55.8% of residual artifact-enriched proxies.
• A frozen genomic foundation model performed at chance because sequence alone does not contain variant-provenance evidence.
• The responsible product is a calibrated risk annotation with provenance and expert review—not an autonomous deletion rule.

References

1. Fang, L. T. et al. Establishing community reference samples, data and call sets for benchmarking cancer mutation detection using whole-genome sequencing. Nature Biotechnology (2021).
2. Sahraeian, S. M. E. et al. Achieving robust somatic mutation detection with deep learning models derived from reference data sets of a cancer sample. Genome Biology (2022).
3. Heo, D. et al. DEEPOMICS FFPE identifies DNA-sequencing artifacts from FFPE tissue with deep learning. Scientific Reports (2024).
4. Cebeci, E. et al. FFPolish: a machine-learning-based method for refining FFPE somatic variant calls. Frontiers in Genetics (2022).
5. Frankell, A. M. et al. Enabling whole-genome sequencing analysis from FFPE specimens in clinical oncology. Nature Communications (2025).
6. Brixi, G. et al. Genome modelling and design across all domains of life with Evo 2. Nature (2026).