Abstract

Microscopes have been an invaluable tool throughout the history of the life sciences, as they allow researchers to observe the miniscule details of living systems in space and time. However, modern biology studies complex and non-obvious phenotypes and their distributions in populations and thus requires that microscopes evolve from visual aids for anecdotal observation into instruments for objective and quantitative measurements. To this end, many cutting-edge developments in microscopy are fuelled by innovations in the computational processing of the generated images. Computational tools can be applied in the early stages of an experiment, where they allow for reconstruction of images with higher resolution and contrast or more colors compared to raw data. In the final analysis stage, state-of-the-art image analysis pipelines seek to extract interpretable and humanly tractable information from the high-dimensional space of images. In the work presented in this thesis, I performed super-resolution microscopy and wrote image analysis pipelines to derive quantitative information about multiple biological processes. I contributed to studies on the regulation of DNMT1 by implementing machine learning-based segmentation of replication sites in images and performed quantitative statistical analysis of the recruitment of multiple DNMT1 mutants. To study the spatiotemporal distribution of DNA damage response I performed STED microscopy and could provide a lower bound on the size of the elementary spatial units of DNA repair. In this project, I also wrote image analysis pipelines and performed statistical analysis to show a decoupling of DNA density and heterochromatin marks during repair. More on the experimental side, I helped in the establishment of a protocol for many-fold color multiplexing by iterative labelling of diverse structures via DNA hybridization. Turning from small scale details to the distribution of phenotypes in a population, I wrote a reusable pipeline for fitting models of cell cycle stage distribution and inhibition curves to high-throughput measurements to quickly quantify the effects of innovative antiproliferative antibody-drug-conjugates. The main focus of the thesis is BigStitcher, a tool for the management and alignment of terabyte-sized image datasets. Such enormous datasets are nowadays generated routinely with light-sheet microscopy and sample preparation techniques such as clearing or expansion. Their sheer size, high dimensionality and unique optical properties poses a serious bottleneck for researchers and requires specialized processing tools, as the images often do not fit into the main memory of most computers. BigStitcher primarily allows for fast registration of such many-dimensional datasets on conventional hardware using optimized multi-resolution alignment algorithms. The software can also correct a variety of aberrations such as fixed-pattern noise, chromatic shifts and even complex sample-induced distortions. A defining feature of BigStitcher, as well as the various image analysis scripts developed in this work is their interactivity. A central goal was to leverage the user's expertise at key moments and bring innovations from the big data world to the lab with its smaller and much more diverse datasets without replacing scientists with automated black-box pipelines. To this end, BigStitcher was implemented as a user-friendly plug-in for the open source image processing platform Fiji and provides the users with a nearly instantaneous preview of the aligned images and opportunities for manual control of all processing steps. With its powerful features and ease-of-use, BigStitcher paves the way to the routine application of light-sheet microscopy and other methods producing equally large datasets.