Recent advances in fluorescence microscopy allow researchers to acquire an unprecedented amount of live cell image data at high spatial and temporal resolutions. However, these images pose a significant challenge for data analyses due to massive subcellular heterogeneity. Although conventional computer vision algorithms have facilitated automatic image analysis, traditional ensemble-averaging of subcellular heterogeneity could lead to the loss of critical mechanistic details. Given the current rapid growth of cell biological data from new technological development, it is nearly impossible to keep up with the data generation if we solely rely on human intelligence for algorithm development and data analysis. Recently, machine learning (ML) is making tremendous progress and has shown that computers can outperform humans in the analysis of complex high dimensional datasets. Conventional ML application in cell biology, however, is usually limited to fixed cells or low spatial resolution setting (single cell resolution), which is limited in analyzing dynamic subcellular information. To fill this voids, we have been developing an ML framework for fluorescence live cell image analyses at the subcellular level. In our previous study, we established the method to deconvolve the subcellular heterogeneity of lamellipodial protrusion from live cell imaging, which identified distinct subcellular protrusion phenotypes with differential drug susceptibility. Thus, our goal is to advance this ML framework and address technical and cell biological challenges in the live cell analysis. The overall goal of our research is two- fold: i) advancing a new ML framework for cell biological research (technological development) and ii) applying our ML framework to integrate mechanobiology and metabolism in cell protrusion (targeted cell biological study). First, we will advance our ML framework for the deconvolution of subcellular heterogeneity of protrusion and molecular coordination in live cells. This method will integrate time-series modeling and ML to deconvolve subcellular molecular coordination. Second, we will develop deep learning based high-throughput fluorescence live cell imaging. This will include microscope automation, resolution enhancement, and data synthesis, which will build up the massive dataset for ML. Third, we will apply our ML framework to study the mechanosensitivity of subcellular bioenergetic status in cell protrusion. We will evaluate how AMPK reacts to mechanical forces and controls the subcellular organization of actin assembly and mitochondria to promote energy-demanding protrusion phenotypes. Our ML framework will bring unprecedented analytical power to cell biology by analyzing a large numbers of individual cells at the high spatial resolution and automatically extracting a multitude of subcellular phenotypes. This framework can be applied to various areas of cell biology such as cytoskeleton, membrane remodeling, and membrane-bound organelles.
We propose to develop a novel machine learning framework for automated, large-scale analyses of single cells at the subcellular level. By integrating time-series modeling and machine learning, this new system will enable us to identify hidden phenotypes and molecular coordination from live cell movies. We will apply the developed technology to study the interplay between mechanical forces and metabolism in cell protrusion.