MutFormer: a multimodal deep learning framework for predicting somatic mutation hotspots from sequence and chromatin features
- 1School of Software Engineering, Quanzhou University of Information Engineering, Quanzhou, 362000 China
- 2Department of Molecular Medicine, UT Health San Antonio, San Antonio, TX 78229 USA
- 3Department of Mathematics and Statistics, University of Wisconsin-La Crosse, La Crosse, WI 54601 USA
Abstract
Background: Somatic mutations are unevenly distributed across the genome, with certain loci recurrently mutated across independent tumors. Both local DNA sequence context and chromatin organization have been associated with mutation rate variation, yet their combined effects at single-nucleotide resolution remain incompletely characterized.
Results: We developed MutFormer, a multimodal framework that integrates sequence context with chromatin-associated features to estimate mutation hotspot susceptibility at single-base resolution. Model performance was evaluated using a chromosome-partitioned strategy to reduce potential bias from local sequence similarity. The model achieved an AUROC of 0.979 and an AUPRC of 0.955 on TCGA data, and showed consistent performance on an independent METABRIC cohort.
In addition to overall predictive accuracy, the integration of chromatin features altered the relative ranking of genomic loci. A subset of hotspot positions showed increased predicted susceptibility compared with sequence-only models, suggesting that chromatin context contributes information not captured by sequence composition alone. Sequence-derived features, particularly CpG-related metrics, accounted for the majority of predictive signal, whereas chromatin features provided a secondary, context-dependent contribution. The absolute AUROC gain from chromatin integration over the sequence-only model was modest (ΔAUROC = 0.004). Chromatin features instead reordered the prediction ranks of a specific subset (∼14%) of hotspot loci, with most loci unaffected.
Conclusions: These results support a model in which intrinsic sequence composition defines baseline mutation susceptibility, while chromatin context modulates the relative prioritization of specific loci. This framework provides a complementary approach for identifying mutation-prone regions beyond sequence-based models. MutFormer runs in a sequence-only mode by default and incorporates chromatin features as an optional enhancement when matched data are available.
Supplementary Information: The online version contains supplementary material available at https://doi.org/10.1186/s12864-026-13246-0.
Related
- Intraoperative indocyanine green fluorescence imaging and multimodal deep learning for predicting non-salvage outcomes in pediatric testicular torsion
- Multi-Scale MRI, Radiology Reports, and Blood Biomarker–Guided Multimodal Deep Learning for Predicting Postoperative Recurrence in Cervical Cancer: A Multicenter Study
- Multimodal graph-based deep learning for predicting heat of combustion as a key hazard indicator under GHS/CLP standards
- Systematic Modality Ablation of Multimodal Machine Learning for Predicting 24-Month Progression from Mild Cognitive Impairment to Alzheimer's Disease
- An integrated evolutionary optimization and machine learning framework for predicting pharmacokinetics of nanocarrier-encapsulated drugs
BibTeX
@article{li2026mutformer,
title = {{MutFormer: a multimodal deep learning framework for predicting somatic mutation hotspots from sequence and chromatin features}},
author = {Tianbao Li and Yan Jiang and Junjie Wu and Junru Lin and Reisa Widjaja and Ailan Wang and Qi Liu and Hongbo Zhou},
journal = {BMC Genomics},
year = {2026},
month = aug,
doi = {10.1186/s12864-026-13246-0},
url = {https://doi.org/10.1186/s12864-026-13246-0},
note = {CC BY-NC-ND 4.0},
}
RIS
TY - JOUR
TI - MutFormer: a multimodal deep learning framework for predicting somatic mutation hotspots from sequence and chromatin features
AU - Tianbao Li
AU - Yan Jiang
AU - Junjie Wu
AU - Junru Lin
AU - Reisa Widjaja
AU - Ailan Wang
AU - Qi Liu
AU - Hongbo Zhou
PY - 2026
DA - 2026/08/11
JO - BMC Genomics
DO - 10.1186/s12864-026-13246-0
UR - https://doi.org/10.1186/s12864-026-13246-0
AB - Background: Somatic mutations are unevenly distributed across the genome, with certain loci recurrently mutated across independent tumors. Both local DNA sequence context and chromatin organization have been associated with mutation rate variation, yet their combined effects at single-nucleotide resolution remain incompletely characterized. Results: We developed MutFormer, a multimodal framework that integrates sequence context with chromatin-associated features to estimate mutation hotspot susceptibility at single-base resolution. Model performance was evaluated using a chromosome-partitioned strategy to reduce potential bias from local sequence similarity. The model achieved an AUROC of 0.979 and an AUPRC of 0.955 on TCGA data, and showed consistent performance on an independent METABRIC cohort. In addition to overall predictive accuracy, the integration of chromatin features altered the relative ranking of genomic loci. A subset of hotspot positions showed increased predicted susceptibility compared with sequence-only models, suggesting that chromatin context contributes information not captured by sequence composition alone. Sequence-derived features, particularly CpG-related metrics, accounted for the majority of predictive signal, whereas chromatin features provided a secondary, context-dependent contribution. The absolute AUROC gain from chromatin integration over the sequence-only model was modest (ΔAUROC = 0.004). Chromatin features instead reordered the prediction ranks of a specific subset (∼14%) of hotspot loci, with most loci unaffected. Conclusions: These results support a model in which intrinsic sequence composition defines baseline mutation susceptibility, while chromatin context modulates the relative prioritization of specific loci. This framework provides a complementary approach for identifying mutation-prone regions beyond sequence-based models. MutFormer runs in a sequence-only mode by default and incorporates chromatin features as an optional enhancement when matched data are available. Supplementary Information: The online version contains supplementary material available at https://doi.org/10.1186/s12864-026-13246-0.
SN - 1471-2164
ER -