Machine Learning for Lifespan Inference from Time-Lapse Microfluidic Images of Dividing Yeast Cells

High-throughput microfluidics-based assays can potentially increase the speed and quality of yeast replicative lifespan measurements that are related to aging. One major challenge is to efficiently convert large volumes of time-lapse images into quantitative measurements of yeast cell lifespan measurements. To address these issues, we developed several deep learning methods to analyze a large number of images collected from microfluidic experiments. First, we compared three deep learning architectures to classify microfluidic time-lapse images of dividing yeast cells into categories that represent different stages in the yeast replicative aging process. Second, we evaluated convolutional neural networks for detecting cells from microfluidic images. The YOLO and Mask R-CNN are trained with yeast microfluidic images and tested for object detection, and features extraction. The results indicate that YOLO had better performance in terms of object detection and accuracy. In contrast, the Mask R-CNN had better performance in terms of cell area and better detection when the number of cells inside the trap is less than 3 cells. Third, prototyping an algorithm that can evaluate cell division events through family trees of cells. We generated a null distribution using single cells inside microfluidic traps. Based on this null distribution, we prototyped a likelihood algorithm for cell tracking between images at different time-points. We inferred cell family trees through a trace-back method. The replicative lifespan of a mother cell can be counted as the number of bifurcating branches of its family tree. Linear regression showed that predictions of our prototype correlated with experimental observations. Forth, since it is challenging to visualize and interpret the time-series data gathered through time-lapse microscopy images, we have developed a circular plotting software tool, mPolar, to visualize the trends and patterns of the cell movements, and cell division events in a time-series. Overall, our methods have the potential to accelerate the efficiency and expand the range of quantitative measurement of yeast replicative aging experiments. This work lays a useful framework for sophisticated deep-learning processing of microfluidic-based assays of yeast replicative aging.