Somatic mutation probabilities learned from the DNA sequence
Somatic mutations distribute heterogeneously over genomes, dependent on immediate sequence context, function, and non-genetic factors. Mutations present the foundation for somatic genome evolution towards age-related diseases. Here, we present EAGLE-MUT, a group of DNA language models trained on sparse single base substitution data from esophageal adenocarcinoma (EAC) and melanoma (MELA) samples. In both cancer types, predicted mutation probabilities vary across several orders of magnitude, depending on sequence context. For example, T>G hotspots in EAC localize to A/T-rich ~200 base pair DNA motifs, associated with nucleosome positioning. Coldspots localize to G/C-rich DNA, nucleosome linker DNA, CTCF binding sites, potential origins of replication, and intron-exon boundaries with a transcription strand bias. Fine-tuning EAGLE-MUT for individual patient samples leads to varying mutation distributions relative to genome elements and potential cancer driver mutations. EAGLE-MUT provides new links between DNA sequence, genome function, and stability in individual samples, a foundation for understanding somatic genome evolution, and more personalized strategies for risk-assessment, prevention, and treatment.