Editors' ChoiceCancer

Predicting pathology provides prostate prognostication

See allHide authors and affiliations

Science Translational Medicine  10 Oct 2018:
Vol. 10, Issue 462, eaav3886
DOI: 10.1126/scitranslmed.aav3886


A live primary cell phenotypic assay with single-cell resolution can predict postsurgical adverse pathology with >80% accuracy for patients with prostate or breast cancer.

Clinicians have long sought ways to identify physiologically relevant biomarkers to improve cancer diagnosis and stratify patients to guide therapeutic intervention and avoid overtreatment; however, efforts to do so have largely been impeded by patient-to-patient and disease heterogeneity.

Manak et al. described a risk stratification assay, Stratification of Adverse Pathology (STRAT-AP), for the analysis of single cells derived from prostate and breast tumor biopsies. To overcome the challenge of tumor heterogeneity—a thorn in the side of most clinicians and pathologists—the authors utilized machine-vision software for image analysis with spatiotemporal resolution of single cells. For each patient in the study, roughly 5000 primary live biopsy cells were analyzed using sequential live-cell imaging across 26 time points to quantify individual cell morphology and motility. The cells were then fixed for immunohistochemical protein quantification, with each cell assigned a unique identification number in order to co-register the data from individual cells, totaling 42 million measurements per biopsy sample.

These techniques generated over 300 features from each cell, focusing on focal adhesion and cytoskeletal markers. A random forest machine learning algorithm was then trained using features from 70% of the cells as well as the formal surgical pathology outcome for the associated tumor. The authors predicted surgical disease severity using the remaining 30% of the cells and successfully identified both local and metastatic disease severity with greater than 80% accuracy.

This work provides a key step forward in stratification of risk to guide cancer prognosis and could provide opportunities for personalized therapy selection and for improving the process of lead-compound discovery for patients with prostate and breast cancer, reducing unnecessary surgical interventions. However, one limitation of the study is that each cell was treated as an independent entity, and the model was trained and validated on cells from the same patients. As cells from the same tumor are likely to have overlapping characteristics, data leak between the training and validation datasets could have occurred, which may have inflated the predictive accuracy of the system. Evaluation in tumors from new patients and other tumors and further validation on risk stratification with a larger patient sample size are key next steps.

Highlighted Article

Stay Connected to Science Translational Medicine

Navigate This Article