MutFormer

Recurrent somatic mutations are the classic signal that a genomic position is under positive selection in cancer, but recurrence alone is a treacherous statistic. Some positions are mutated repeatedly across tumors simply because local DNA sequence makes them fragile substrates for a particular mutational process — APOBEC deamination, UV photodimer formation, or polymerase-epsilon proofreading failure. The problem is acute outside protein-coding genes, where whole-genome studies routinely surface hotspots that turn out to be high-mutability passengers rather than drivers. The standard correction, a trinucleotide context model inherited from the COSMIC mutational signature framework, captures only one base on either side of the mutated position and is too coarse to explain much of the observed variation.

MutFormer is a hybrid transformer/convolutional neural network that predicts base-pair-resolution somatic mutation risk from DNA sequence alone, producing separate predictions for individual COSMIC signatures. It was developed by Iván Galván-Femenía, Marcell Veiner, Daniel Naro, and Fran Supek in the Genome Data Science lab at IRB Barcelona, and posted to bioRxiv in July 2026 under a CC BY license with support from two European Research Council grants. The work is a preprint and has not yet completed peer review.

The name collides with an unrelated 2021 model, also called MutFormer, which is a BERT-style protein language model for missense pathogenicity prediction. That earlier system operates on amino acid sequence and scores germline variant deleteriousness; this one operates on DNA and estimates how often a given base is expected to mutate somatically.

#Key Features

• Per-signature mutation risk: Rather than a single genome-wide mutation rate, the model emits a separate base-pair-resolution risk track for each COSMIC signature, matching the fact that different tumors are dominated by different mutagenic processes.
• Context beyond the trinucleotide: MutFormer learns extended sequence determinants that often span up to roughly 20 nucleotides, well past the ±1 base window that conventional signature models use.
• Recovers known mutagenic preferences: Without being told what to look for, the model reproduces the sequence preferences of APOBEC, UV, POLE and SBS17 mutagenesis, along with additional mutation-risk-prone motifs.
• Neutral recurrence modeling: Sequence-derived risk is combined with tumor mutation burden, per-sample signature exposures, and epigenetic covariates to compute how often a specific position should recur under neutrality across more than 18,000 tumor whole genomes.
• Calibrated false discovery: Coding-region analyses benchmark the framework against known driver genes and AlphaMissense scores, supporting conservative false-discovery estimates before the method is applied to non-coding space.

#Technical Details

The architecture pairs convolutional layers with transformer blocks over raw DNA sequence and was trained on more than 90 million high-confidence, signature-assigned single nucleotide variants drawn from cancer genomes. Because each training mutation carries a probabilistic signature assignment, the network learns signature-specific sequence determinants rather than an averaged mutability landscape. On top of the neural predictions sits a statistical layer that converts expected per-base risk into an expected recurrence count for each position in a cohort, using mutation burden, signature exposures, and epigenetic covariates. Applied genome-wide, this recurrence model explains the large majority of recurrent non-coding hotspots as passenger mutability. The residual outliers are enriched near known cancer genes and were corroborated with orthogonal evidence from SpliceAI, PromoterAI, AlphaGenome, and expression data. Prioritized candidates include splice-region or deep-intronic hotspots in BCL6, PTEN, TCF7L2, PBRM1, PTPRT and VHL, and promoter hotspots in SHKBP1, PRSS3 and BCL2.

#Applications

The primary use case is cancer driver discovery in whole-genome sequencing cohorts, where MutFormer functions as the null model that a candidate hotspot must beat. Groups analyzing pan-cancer or tumor-type-specific genome collections can use it to filter recurrent positions that are explained by local mutability, concentrating experimental follow-up on the small set of positions with genuine excess recurrence. The signature-resolved output is also directly useful for studying mutagenesis itself — characterizing which sequence contexts a given endogenous or exogenous process targets, at a resolution finer than signature catalogs provide.

#Impact

The result that reframes the field is a negative one: across more than 18,000 tumor whole genomes, most recurrent non-coding hotspots are best explained by passenger mutability rather than selection, which sets a high evidentiary bar for claims of non-coding drivers. The candidate list the method does produce — splice-region, deep-intronic, and promoter hotspots near established cancer genes — is a concrete set of testable hypotheses supported by convergent computational evidence, though the variants themselves await functional validation. As a preprint, the work has not been peer reviewed, and no public code repository or released model weights have been made available, which currently limits independent reuse of the predictions.