Neural network reads cancer’s mutational fingerprints from tiny data scraps

Cancer genomes carry the scars of the mutational processes that shaped them. Each process, from faulty DNA repair to environmental exposures, leaves a characteristic pattern of mutations called a signature, and measuring how active each process was in a given tumor, its so-called exposure, can help stratify patients and predict drug response. The trouble is that the most widely used tools for reading these fingerprints were built for whole-genome sequencing data, where thousands of mutations provide plenty of statistical signal. In targeted panel sequencing, the workhorse of clinical oncology, a sample may harbor only a handful of mutations, and the standard refitting methods begin to falter.

In a study published in Algorithms for Molecular Biology, Gal Gilad, Teresa Przytycka and Roded Sharan introduce SuRe, short for Supervised Refitting, a neural network trained to predict signature exposures directly from sparse mutation counts. Across synthetic benchmarks, real breast cancer genomes and a pan-cancer collection spanning more than 4,500 patients, SuRe outperformed established unsupervised methods, and its advantage grew wider as the data became sparser.

Signature refitting is the task of inferring, from the mutations a tumor carries, how many of them were caused by each of dozens of known mutational processes cataloged in the COSMIC database. The most popular tool for this job is non-negative least squares, which finds exposures that minimize the Kullback-Leibler divergence between the observed mutation counts and the product of signatures and exposures. Alternatives include deconstructSigs, a heuristic based on iterative regression; SigLASSO, which adds L1 regularization to produce sparse, high-confidence solutions; SignatureEstimation, which probes the stability of exposure estimates by perturbing the input counts; SigProfilerAssignment, which iteratively adds and removes signatures guided by reconstruction similarity; and MuSiCal, a likelihood-based sparse NNLS variant. A method called Mix was designed specifically for sparse panel data, clustering samples and learning exposures per cluster rather than per sample.

SuRe departs from all of these by learning from labeled data. The model takes as input a vector of 96 counts, one for each standard single-base substitution category defined by the mutated base, the substituted base and the flanking nucleotides. Optionally, it also receives a one-hot encoded tissue-type vector. The combined representation passes through a Mixture-of-Experts architecture: several independent expert modules, each a stack of fully connected layers with ReLU activations and dropout regularization, produce probability distributions over the signatures, while a gating network learns how much to weight each expert for a given input. The final prediction is a weighted average of the expert outputs. The idea is that individual experts can specialize in particular tissues, cancer types or exposure patterns, with the gate deciding which specialists to consult.

Two design choices stand out. First, SuRe processes raw absolute mutation counts rather than converting them to compositional probabilities. The authors argue this is biologically motivated: a surge in overall mutation burden typically reflects an additive increase in a specific process, such as mismatch repair deficiency, rather than a uniform scaling of all background signatures, and preserving absolute counts keeps that additive signal visible. It also lets the network’s internal logit magnitudes scale with the volume of evidence, producing more conservative predictions on very sparse inputs and sharper ones when the mutational signal is strong. Second, the loss function is the Kullback-Leibler divergence between predicted and actual relative exposures, rather than a multinomial likelihood on the observed counts, because the COSMIC database provides aggregate exposures per sample rather than labels for individual mutations, and fitting a likelihood to sparse observed counts would push the model to overfit sampling noise.

Training data came from previously published analyses of COSMIC mutation catalogs, including whole-genome sequencing data from thousands of cancer patients along with their estimated signature exposures. The researchers assembled a breast cancer set of 679 patients, a pan-cancer set of 4,568 patients and a synthetic set of 2,700 simulated cancer genomes generated from 21 COSMIC reference signatures. To mimic panel sequencing, they downsampled the whole-genome data to as few as 3, 6, 10 or 30 mutations per sample. During training, the model was exposed to subsampled mutations drawn from a power-law distribution, forcing it to learn robust representations across the full range of mutation burdens, while the original whole-genome samples were also fed in.

The benchmark results tell a consistent story. On the synthetic data set, where ground-truth exposures are known, MuSiCal was the top performer on full genome-wide mutation counts, but SuRe beat all competing methods on both reconstruction error and correlation whenever fewer than 300 mutations per sample were available. In the sparse regimes that are the study’s central focus, at 30 mutations or fewer, a Wilcoxon signed-rank test confirmed a statistically significant advantage over the next-best method across all downsampling repetitions. When the data were refitted against the entire COSMIC catalog rather than only signatures known to be active, NNLS, MuSiCal and SigProfiler degraded sharply, over-assigning exposures to inactive signatures, while SuRe remained comparatively resilient. Similar patterns held on real breast cancer genomes, where SuRe’s predictions agreed better with COSMIC exposures across all sampling sizes, and on the pan-cancer data, where the model leveraged tissue-type input to refit against all 58 active signatures and still outperformed methods that had been given the easier, tissue-restricted signature sets.

The clinical demonstrations are where the approach may matter most. Using triple-negative breast cancer samples with homologous recombination deficiency status predicted by HRDetect, the team downsampled mutation data to the genomic regions covered by the MSK-IMPACT panel and asked whether inferred exposure to SBS3, a signature known to correlate strongly with HRD, could classify samples. SuRe, fine-tuned to predict SBS3 exposure directly, and Mix were the only methods that clearly improved on a baseline classifier based on total mutation burden alone, and SuRe significantly outperformed Mix. A bias analysis showed why: competing methods tended to underestimate exposure to SBS3 and to SBS5, the other flat signature common in breast cancer, while SuRe showed no such tendency for these or for the spiky signatures common in the disease.

The team then turned to native panel data rather than downsampled genomes, analyzing 13,076 samples from the MSK-IMPACT and MSK-MET cohorts across lung, breast and prostate cancers. After normalizing predicted exposures and clustering patients into two groups, SuRe produced clusters with extremely strong associations to high tumor mutational burden, with p-values below 10 to the minus 40 in lung and breast cancer. Most strikingly, in the prostate cohort, a tissue with typically low mutation counts, SuRe was the only method to identify a significant TMB-associated subgroup, while NNLS and SigProfiler failed to reach significance. SuRe also recovered metastatic-status signals in prostate and breast cohorts that the competing methods missed or matched only weakly, and in lung cancer it separated smokers from never-smokers with a 17-percentage-point gap between clusters, compared with 10 percent for SigProfiler. Associations with HR pathway alterations in prostate cancer were insignificant for all methods, and no method stratified lung cancer samples by metastatic status.

The authors are candid about a central limitation: SuRe is trained and evaluated against COSMIC exposures that were themselves computationally estimated, which could restrict and bias the learning process. They chose these labels because they are a field standard and have been shown to correlate with clinical attributes including age, smoking history, HRD status and mismatch repair deficiency. They also note that the model’s ability to reconstruct, from very few mutations, exposures estimated from rich data suggests it genuinely captures correlations between co-active signatures. With hyperparameters tuned by grid search, four experts and 100 hidden units per layer for the breast cancer model, eight experts and 500 units for the pan-cancer version, and an ablation study confirming the Mixture-of-Experts design outperforms a comparable standard network, SuRe offers a template for supervised refitting that could open the door to reliably detecting rare signatures in the sparse panel data that dominates clinical sequencing.

Subject of Research: A supervised neural network method for inferring mutational signature exposures from sparse pan-cancer sequencing data

Article Title: Mutational signature refitting on sparse pan-cancer data

Article References: Gilad, G., Przytycka, T. M., & Sharan, R. (2026). Mutational signature refitting on sparse pan-cancer data. Algorithms for Molecular Biology, 21(1), Article 16. https://doi.org/10.1186/s13015-026-00305-0

Image Credits: AI Generated

DOI: 10.1186/s13015-026-00305-0

Keywords: mutational signatures, cancer genomics, signature refitting, neural network, sparse data, panel sequencing, COSMIC, homologous recombination deficiency, MSK-IMPACT, tumor mutational burden, machine learning, pan-cancer