Learning cell representations in temporal and 3D contexts

Thesis

Cell morphology and its changes under different circumstances is one of the primary ways by which we can understand biology. Computational tools for characterization and analysis, therefore, play a critical role in advancing studies involving cell morphology.

In this thesis, I explored the use of representation learning and self-supervised methods to analyze nuclear texture in fluorescence imaging across different contexts and scales. To analyze the cell cycle using 2D temporal imaging data, as well as DNA damage in 3D imaging data, I employed a simple model based on the VAE-GAN architecture. Through the VAE-GAN model, I constructed manifolds in which the latent representations of the data can be grouped and clustered based on textural similarities without the need for exhaustive training annotations. I used these representations, as well as manually engineered features, to perform various analyses both at the single cell and tissue levels.

The application on the cell cycle data revealed that common tasks such as cell cycle staging and cell cycle time estimation can be done even with minimal fluorescence information and user annotation. On the other hand, the texture classes derived to characterize DNA damage in 3D histology images unveiled differences between control and treated tissue regions. Lastly, by aggregating cell-level information to characterize local cell neighborhoods, interactions between DNA-damaged cells and immune cells can be quantified and some tissue microstructures can be identified.

The results presented in this thesis demonstrated the utility of the representations learned through my approach in supporting biological inquiries involving temporal and 3D spatial data. The quantitative measurements computed using the presented methods have the potential to aid not only similar experiments on the cell cycle and DNA damage but also in exploratory studies in 3D histology.

Authors: delas Penas, K

• DOI: 10.5287/ora-mzndbaezm
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: machine learning; representation learning; cell cycle; histology