SigFormer: an Attention-Based Framework for Robust Single-Sample Mutational Signature Decomposition.

Area of Science:

• Genomics
• Computational Biology
• Cancer Research

Background:

• Somatic mutational signatures reveal genomic history from exposures and endogenous processes.
• Accurate single-sample signature decomposition is challenging due to low mutation burden, high noise, and complex catalogs.

Purpose of the Study:

• To develop a robust computational framework, SigFormer, for accurate somatic mutation analysis, particularly in challenging low-burden and high-noise scenarios.
• To improve the decomposition and detection of mutational signatures at the single-sample level.

Main Methods:

• Development of SigFormer, a set-conditioned transformer framework utilizing a cross-attention mechanism.
• Comparison of SigFormer against likelihood-driven refitting methods like MuSiCal.
• Application to PCAWG genomes and low-burden normal-tissue datasets.

Main Results:

• SigFormer demonstrates improved exposure recovery and detection accuracy over existing methods, especially in high-noise and overcomplete settings.
• Analysis of PCAWG genomes shows SigFormer preserves tissue structure and captures co-occurring low-abundance signatures without tumor-type specificity.
• SigFormer accurately identifies stable tissue-dependent mutational patterns in normal tissues and quantifies unattributable residuals.

Conclusions:

• SigFormer offers a robust and accurate method for somatic mutation signature analysis, outperforming traditional approaches.
• The framework is effective across diverse datasets, including low-burden normal tissues, revealing mutagenic heterogeneity.
• SigFormer provides a valuable tool for understanding genomic damage, disease risk, and for downstream analyses by explicitly handling incomplete signature catalogs.