Computational analysis of single-cell dynamics: protein localisation, cell cycle, and metabolic adaptation

Abstract

Cells need to be able to adapt quickly to changes in nutrient availability in their environment in order to survive. Budding yeasts constitute a convenient model to study how eukaryotic cells respond to sudden environmental change because of their fast growth and relative simplicity. Many of the intracellular changes needed for adaptation are spatial and transient; they can be captured experimentally using fluorescence time-lapse microscopy. These data are limited when only used for observation, and become most powerful when they can be used to extract quantitative, dynamic, single-cell information.

In this thesis we describe an analysis framework heavily based on deep learning methods that allows us to quantitatively describe different aspects of cells’ response to a new environment from microscopy data. chapter 2 describes a start-to-finish pipeline for data access and preprocessing, cell segmentation, volume and growth rate estimation, and lineage extraction. We provide benchmarks of run time and describe how to speed up analysis using parallelisation. We then show how this pipeline can be extended with custom processing functions, and how it can be used for real-time analysis of microscopy experiments.

In chapter 3 we develop a method for predicting the location of the vacuole and nucleus from bright field images. We combine this method with cell segmentation to quantify the timing of three aspects of the cells’ response to a sudden nutrient shift: a transient change in transcription factor nuclear localisation, a change in instantaneous growth rate, and the reorganisation of the plasma membrane through the endocytosis of certain membrane proteins. In particular, we quantify the relative timing of these processes and show that there is a consistent lag between the perception of the stress at the level of gene expression and the reorganisation of the cell membrane.

In chapter 4 we evaluate several methods to obtain cell cycle phase information in a label-free manner. We begin by using the outputs of cell segmentation to predict cytokinesis with high accuracy. We then predict cell cycle phase at a higher granularity directly from bright field images. We show that bright field images contain information about the cell cycle which is not visible by eye. We use these methods to quantify the relationship between cell cycle phase length and growth rate.

Finally, in chapter 5 we look beyond microscopy to the bigger picture. We sketch an abstract description of how, at a genome-scale, cells might choose a strategy for adapting to a nutrient shift based on limited, noisy, and local information. Starting from a constraint-based model of metabolism, we propose an algorithm to navigate through metabolic space using only a lossy encoding of the full metabolic network. We show how this navigation can be used to adapt to a changing environment, and how its results differ from the global optimisation usually applied to metabolic models.