A deep learning program using live cell images to classify melanoma cell populations.
Live cell histology for classification of melanoma cell population based on single cell actions
It has long been speculated that metastatic cells adopt distinct appearance and dynamics (morphodynamics) that are predictive of their potential to disseminate and survive in remote tissues. However, this hypothesis has not been systematically tested, due to the challenges in associating clinically-documented outcomes of metastatic development in patients to behavior of individual cells. By combining a unique cell system, high-content imaging pipeline and a novel analytical approach we demonstrate that morphodynamic behaviors are characteristic traits of the metastatic efficiency of individual cells.
We built on a transplantation xenograft model of melanoma that maintains, with high reproducibility, the clinical signature of tumors extracted from patients. We implemented an unsupervised deep neural network (Generative adversarial networks) followed by supervised machine learning to capture the subtle details of cell appearance and actions that discriminate between cells from high- versus low-metastatic potential tumors. The same methodology could distinguish between established melanoma cell lines versus primary-derived cell systems, between transformed and untransformed cell lines, and even between two different cell lines or primary tumors. Altogether, we demonstrate that the dynamics of unstructured textural data extracted from live melanoma imaging encapsulates information predictive of the cell origin, including the metastatic efficiency.
We coin this new assay quantitative live cell histology. Beyond the potential of becoming an invaluable assay for the discovery of metastasis-promoting pathways, quantitative live cell histology may become the backbone assay to investigate various aspects in cellular heterogeneity, plasticity, and identification of molecular players governing particular cell functions.