Integrative multiomics-histopathology analysis for breast cancer classification
Yasha Ektefaie, William Yuan, Deborah A. Dillon, Nancy U. Lin, Jeffrey A. Golden, Isaac S. Kohane and Kun-Hsing Yu
npj. 2021. DOI: 10.1038/s41591-018-0300-7. DOI:10.1038/s41523-021-00357-y
Could deep learning reveal connections between histology and multi-omics for breast cancer?
Histopathologic evaluation of biopsy slides is a pivotal procedure in diagnosing and subtyping breast cancers, an indispensable step for the diagnosis and treatment of breast cancer. However, the connections between visual morphology and genetic statuses have never been systematically explored or interpreted.
- The authors associate deep learning image models from digitized histopathology slides with transcriptomic analyses to expose molecular and morphological profiles associated with hormone receptor status and genomic subtypes of breast cancer.
- Ektefaie et al. developed image classifiers for several classification tasks including hormone receptor (HR) status prediction in breast cancer patients using only hematoxylin-and-eosin slides. They then trained a regression model on RNA-seq profiles and identified important genes for the task of HR classification. The image classifiers did learn lymphocyte-specific morphological signals for the HR classification task.
- This work confirms the utility of deep learning-based image models in both clinical and research settings, through its capability to uncover connections between visual morphology and genetic statuses. This pipeline can be immediately used to examine the underlying biology of other breast cancer markers.