Mechanism-Driven Radiogenomic Analysis Reveals Potential Molecular Drivers of Pretreatment MRI Phenotypes in Breast Cancer

Abstract Purpose To develop a pretreatment MRI model for predicting residual cancer burden (RCB) after neoadjuvant chemotherapy and a mechanism-driven radiogenomic framework to identify molecular programs underlying response-associated MRI phenotypes. Methods A multimodal deep learning model was developed using pretreatment MRI and clinical data from 492 patients across two centers. Radiomic phenotypes from 61 patients were integrated with public transcriptomic cohorts using differential expression, pathway enrichment, WGCNA, survival modeling, feature selection, and in silico perturbation analyses. Results The model achieved AUCs of 0.919 in training and 0.852 in testing. Of 195 response-associated differentially expressed genes, 109 showed concordant changes in an independent cohort. Response was linked to proliferation, DNA repair, hypoxia, immune-inflammatory, and estrogen-response programs. A 12-gene prognostic signature achieved a C-index of 0.798. Multilayer prioritization identified ESR1, TBC1D9, GREB1 , and related genes, while RARRES1 uniquely showed strong directional concordance between predicted and observed radiomic changes. Conclusion Pretreatment MRI heterogeneity reflects coordinated biological programs relevant to neoadjuvant treatment response. This framework supports noninvasive RCB stratification and generates testable mechanistic hypotheses for functional validation.