Chromatin accessibility of primary cancers informs regional mutagenesis in metastases through multi-scale deep learning

Abstract

Tissue-specific chromatin states shape regional mutation density in primary tumors, but whether this relationship persists in metastases is unclear. We integrated whole-genome sequencing data from 2,507 metastatic tumors across six cancer types with 892 chromatin accessibility and replication timing profiles, and developed CAMM, a multi-scale deep learning model, to jointly predict SNV and indel density across multiple genomic resolutions. The model explained regional variance in mutation density across metastatic cancers and preserved mutagenesis patterns in an independent primary tumor cohort. Chromatin accessibility profiles from tissue-matched primary cancers were informative predictors of regional mutagenesis in metastases, supporting a contribution of lineage-linked chromatin context, as indicated by feature attribution analyses. Genomic windows with mutation burdens exceeding epigenome-based expectations were enriched for known and candidate cancer loci. These results link metastatic regional mutagenesis to tissue-of-origin chromatin accessibility and provide a framework for interpreting mutational processes and prioritizing mutation-enriched loci.