Disseminating cells in human oral tumours possess an EMT cancer stem cell marker profile that is predictive of metastasis in image-based machine learning.

When oral cancers metastasise ­ that is, when tumour cells invade other parts of the body ­ they typically do so by first colonizing the lymph nodes present in the neck. As this event significantly reduces chances of survival, oral cancer patients often have their neck lymph nodes removed to prevent the spread of the disease. However, this surgery carries risks and leads to longer hospital stays, stressing the need for better ways to predict which oral tumours will metastasise. Evidence from lab-grown cells and mice studies suggest that, in oral cancer, metastasis occurs when some cells in the original tumour go through a process called the epithelial-mesenchymal transition (EMT for short). This transformation allows the cells to detach from the tumour and become invasive. However, it has so far been difficult to observe this process in actual human tumours; this is partly because cells undergoing EMT stop producing the proteins that scientists rely on to distinguish cancer and healthy cells. To address this knowledge gap, Youssef et al. focused on three proteins: two tumour markers, EpCAM and CD24; and Vimentin, which is produced in greater quantities in the invasive mesenchymal state. Previous work had shown that a specific population of oral tumour cells can continue to express all three proteins even when adopting a mesenchymal identity through EMT. Based on this knowledge, Youssef et al. hypothesised that tracking Vimentin, EpCAM and CD24 using fluorescence microscopy would allow them to identify metastasising cells in human samples. An analysis of over 12,000 images from 74 tumours obtained from surgeries revealed that, in the metastatic samples, the cells detaching from primary tumours were more likely to express these three proteins. Finally, Youssef et al. used these images to train a machine learning algorithm. When applied to data from new oral cancer patients, the programme was able to predict whether their tumours were likely to spread with 89% accuracy. If confirmed by further work, and in particular on larger samples, these findings could in the future help clinicians decide which patients with oral cancer would benefit the most from surgery to remove neck lymph nodes.