Ulicna, Kristina; (2023) Machine Learning for Single-Cell Trajectory Analysis. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text Kristina_Ulicna_PhD_Thesis.pdf - Accepted Version Download (143MB) | Preview

Abstract

Single-cell methods are beginning to reveal the intrinsic heterogeneity in cell populations. However, it remains challenging to quantify single-cell behaviours from time-lapse microscopy data, owing to the difficulty of extracting reliable cell trajectories and lineage information over long time-scales and across several generations. To address this challenge, I used a hybrid deep learning and Bayesian cell tracking approach to reconstruct lineage trees from live-cell microscopy data. The cell detection step consisted of a residual U-Net model coupled with a cell state CNN classifier to allow accurate instance segmentation of the cell nuclei with classification of its mitotic stage. To track the cells over time and through cell divisions, I utilised a Bayesian cell tracking methodology that uses input features from the images to enable the retrieval of multi-generational lineage information from a corpus of thousands of hours of live-cell imaging data. Using this approach, I extracted 20,000+ fully annotated single-cell trajectories from over 3,500 hours of video footage, organised into multi-generational lineage trees spanning up to 8 generations and fourth cousin distances. To demonstrate the robustness of this minimally supervised cell tracking methodology, I retrieve cell cycle durations and their extended inter- and intra-generational family relationships in 5,000+ fully annotated cell lineages without any manual curation. My analysis expands the depth and breadth of investigated cell lineage relationships in approximately two orders of magnitude more data than in previous studies of cell cycle heritability. To further dissect the origins of the cycling duration at the single-cell level, an additional proliferating cell reporter was introduced to the imaging. I developed a self-supervised deep neural network to learn a low-dimensional representation of the underlying cell state. Using this approach, I successfully annotated the phases and the continuous transitions within the cell cycle as well as identified regions of high self-similarities within and across single cell clones. Moreover, this approach enabled detailed trajectory comparison directly from the image data without incorporation of prior knowledge about the cell cycle phases into the model. I envisage that this dynamic and time-dependent similarity scoring could be used to reveal resemblances between cells at scale, and applied across more heterogeneous cell systems, with putative applications in high-throughput screening.

Download activity - last month

Export as XMLCSVJSON

Download activity - last 12 months

Export as JSONXMLCSV

Downloads by country - last 12 months

Export as CSVXMLJSON

Archive Staff Only

View Item