Professor Kwonmoo Lee along with a diverse team consisting of biomedical engineers, computer scientists and cell biologists have taken upon themselves a project in which they are trying to develop a computational methodology that would combine machine learning and live cell imaging that could solve complex problems in cell biology. While machine learning is involved in cell biology data, particularly immunofluorescence images, it has not been widely applied to live cell movies. The reason for this discrepancy is that with fixed cells it is easier to obtain a colorful and clear image, as well as analyze the image. However, living cell imaging is more difficult and challenging. Part of the difficulty lies in the necessity of maintaining the cell viability and making sure it is not extremely stressed, while still receiving the necessary data. Further difficulty is found in the fact that the images that are obtained contain a substantial amount of noise, and cannot be manipulated the same way fixed cell images can be. The exceptional benefit of living cell imaging however is that temporal changes can be observed, dynamics of the cell can be recorded, cell division can be studied and the development and movement of individual cells can be tracked. Professor Lee and his team are specifically looking to integrating live cell imagining and machine learning to study the cell protrusion process, which initiates cell migration. Applying machine learning to the study of cells would provide the advantage of studying the complex and heterogeneous processes of the cells and would allow to track complicated changes in the individual cells, as well as cell populations. The use of machine learning enables them to identify small individual activities that are different from the larger group, isolate that activity from large data sets and allow researchers to study the diverse behaviors of cells as well as heterogeneous drug susceptibility.
The team currently already has the machine learning pipeline set up and is moving to the actual cellular processes of making their machine learning to work by identifying subcellular phenotypes of cell protrusion. Since the beginning of this project the team’s hard work and dedication has already identified an interesting protrusion pattern, called ‘accelerating cell protrusion’ which makes up only a fraction of the whole data and for this reason has been overlooked in the past. The heterogeneous protrusion has been correlated with specific actin regulator molecules. Furthermore, the machine learning pipeline could pinpoint specific subcellular responses under pharmacological perturbations. Further research with the use of live cell imagining will be looking into the specifics of this protrusion, whether it has a long-term migratory consequence in cancer metastasis. The pipeline will be also augmented by deep learning which would provide the machine with raw pixel images that it would then analyze and identify the important features automatically. The team expect that this technology will help cell biologists to discover rare and potent cellular activities particularly involved in cancer, and characterize the complex drug responses.