Preview

PDF, English - main document Download (54MB) | Terms of use

Abstract

A cell is the basic functional unit of life. Most ulticellular organisms, including animals, are composed of a variety of different cell types that fulfil distinct roles. Within an organism, all cells share the same genome, however, their diverse genetic programs lead them to acquire different molecular and anatomical characteristics. Describing these characteristics is essential for understanding how cellular diversity emerged and how it contributes to the organism function. Probing cellular appearance by microscopy methods is the original way of describing cell types and the main approach to characterise cellular morphology and position in the organism. Present cutting-edge microscopy techniques generate immense amounts of data, requiring efficient automated unbiased methods of analysis. Not only can such methods accelerate the process of scientific discovery, they should also facilitate large-scale systematic reproducible analysis. The necessity of processing big datasets has led to development of intricate image analysis pipelines, however, they are mostly tailored to a particular dataset and a specific research question. In this thesis I aimed to address the problem of creating more general fully-automated ways of describing cells in different imaging modalities, with a specific focus on deep neural networks as a promising solution for extracting rich general-purpose features from the analysed data. I further target the problem of integrating multiple data modalities to generate a detailed description of cells on the whole-organism level. First, on two examples of cell analysis projects, I show how using automated image analysis pipelines and neural networks in particular, can assist characterising cells in microscopy data. In the first project I analyse a movie of drosophila embryo development to elucidate the difference in myosin patterns between two populations of cells with different shape fate. In the second project I develop a pipeline for automatic cell classification in a new imaging modality to show that the quality of the data is sufficient to tell apart cell types in a volume of mouse brain cortex. Next, I present an extensive collaborative effort aimed at generating a whole-body multimodal cell atlas of a three-segmented Platynereis dumerilii worm, combining high resolution morphology and gene expression. To generate a multi-sided description of cells in the atlas I create a pipeline for assigning coherent denoised gene expression profiles, obtained from spatial gene expression maps, to cells segmented in the EM volume. Finally, as the main project of this thesis, I focus on extracting comprehensive unbiased cell morphology features from an EM volume of Platynereis dumerilii. I design a fully unsupervised neural network pipeline for extracting rich morphological representations that enable grouping cells into morphological cell classes with characteristic gene expression. I further show how such descriptors could be used to explore the morphological diversity of cells, tissues and organs in the dataset.