Predicting MammaPrint Recurrence Risk from Breast Cancer Pathological Images Using a Weakly Supervised Transformer

Abstract

Recurrence related to poor prognosis is a leading cause of mortality in patients with breast cancer (BC). The MammaPrint (MP) genomic assay is designed to stratify recurrence risk and evaluate chemotherapy benefits for early-stage HR+/HER2- BC patients. However, MP fails to reveal spatial tumor morphology and is limited by high costs. In this study, we establish a BC MP cohort and develop CPMP, a weakly supervised agent-attention transformer model, to predict MP recurrence risk from annotation-free BC histopathological slides. CPMP achieves an AUROC of 0.824 {+/-} 0.03 in predicting MP risk groups. We further leveraged CPMP for spatial and morphological analyses to explore histological patterns associated with MP risk groups. The model reveals tumor spatial localization at the whole-slide level and highlights distinct intercellular interaction patterns of MP groups. It also characterizes the diversity in tumor morphology and uncovers MP High-specific, Low-specific, and colocalized morphological phenotypes that differ in quantitative cellular composition. Prognostic evaluation in the external cohort exhibits significant stratification of distant metastasis risk (HR: 3.14, p-value = 0.0014), underscoring the prognostic power of CPMP. These findings demonstrate the capability of CPMP in MP risk prediction, offering a flexible supplement to genomic risk assessment in early-stage BC.