Abstract 1500: Modeling base-pair level mutation rate in metastatic breast cancer using a sequence-based deep learning model

Abstract By leveraging a sequence-based deep learning framework, we seek to uncover mechanisms by which mutations arise in noncoding regulatory regions, potentially leading to the discovery of novel targets in metastatic breast cancer. Research around metastatic breast cancer has largely focused on analyzing coding mutations to characterize progression. Though noncoding mutations are known to affect transcription factor binding and regulation of gene expression, few noncoding mutation drivers have been identified. Previously published work has shown that metastatic mutation rate correlates with open chromatin in the cells-of-origin. However, this work has mostly been done at a coarse, region-level scale, identifying mutational hotspot regions. In this work, we aim to uncover genomic positions in regulatory regions of metastatic breast cancer with elevated mutation rates, and identify their potential mutation mechanism. We propose a novel deep learning model that uncovers the sequence context-based covariates of per-base mutation rate in regulatory regions of metastatic breast cancer. With access to over 500,000 mutations from the Hartwig Medical Foundation cohort, the neural network is trained on sequences from regulatory regions in normal breast epithelium, and predicts per-base mutation rate profiles for the region. As a result, the model learns how sequence features change mutation likelihood at particular genomic positions. Analysis of the saliency map of the model allows for identification of specific sites with higher-than-expected mutation rates, which is potentially indicative of increased transcription factor binding that extends beyond selection by pro-metastatic regulatory programs. These findings provide a method to connect noncoding mutation patterns to mutation mechanism and regulatory effects in metastatic genomes. Future work aims at expanding this model to a pan-cancer level, revealing shared and cancer-specific noncoding mutations that have potential to reveal patterns of metastasis. This scalable sequence model framework provides an advantage over existing methods particularly due to its base-pair level resolution modeling of the metastatic cancer genome. Thus, its implications for modeling and uncovering regulatory mechanisms of cancer is key. Citation Format: Ariaki Dandawate, Christina Leslie, Ekta Khurana. Modeling base-pair level mutation rate in metastatic breast cancer using a sequence-based deep learning model [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 1500.