Background: The ability to predict future risk of cancer development in non-malignant biopsies is poor. Cellular senescence has been associated with cancer as either a barrier mechanism restricting autonomous cell proliferation or a tumor-promoting microenvironmental mechanism that secretes pro-inflammatory paracrine factors. With most work done in non-human models and the heterogenous nature of senescence the precise role of senescent cells in the development of cancer in humans is not well understood. Further, more than one million non-malignant breast biopsies are taken every year that could be a major source of risk-stratification for women.
Methods: We applied single cell deep learning senescence predictors based on nuclear morphology to histological images of 4,411 H&E-stained breast biopsies from healthy female donors. Senescence was predicted in the epithelial, stromal, and adipocyte compartments using predictor models trained on cells induced to senescence by ionizing radiation (IR), replicative exhaustion (RS), or antimycin A, Atv/R and doxorubicin (AAD) exposures. To benchmark our senescence-based prediction results we generated 5-year Gail scores, the current clinical gold standard for breast cancer risk prediction.
Findings: We found significant differences in adipocyte-specific IR and AAD senescence prediction for the 86 out of 4,411 healthy women who developed breast cancer an average 4.8 years after study entry. Risk models demonstrated that individuals in the upper median of scores for the adipocyte IR model had a higher risk (OR=1.71 [1.10-2.68], p=0.019), while the adipocyte AAD model revealed a reduced risk (OR=0.57 [0.36-0.88], p=0.013). Individuals with both adipocyte risk factors had an OR of 3.32 ([1.68-7.03], p<0.001). Alone, 5-year Gail scores yielded an OR of 2.70 ([1.22-6.54], p=0.019). When combining Gail scores with our adipocyte AAD risk model, we found that individuals with both of these risk predictors had an OR of 4.70 ([2.29-10.90], p<0.001).
Interpretation: Assessment of senescence with deep learning allows considerable prediction of future cancer risk from non-malignant breast biopsies, something that was previously impossible to do. Furthermore, our results suggest an important role for microscope image-based deep learning models in predicting future cancer development. Such models could be incorporated into current breast cancer risk assessment and screening protocols.
Funding: This study was funded by the Novo Nordisk Foundation (#NNF17OC0027812), and by the National Institutes of Health (NIH) Common Fund SenNet program (U54AG075932).