Logo Logo
Hilfe
Kontakt
Switch language to English
Integrating deep learning and genetic approaches to uncover molecular mechanisms of cellular organization
Integrating deep learning and genetic approaches to uncover molecular mechanisms of cellular organization
Organisation is a fundamental principle of life. Matter needs to be arranged in space in such a way that it enables reproduction. In biology, this organisation occurs at different scales: from whole ecosystems to multicellular organisms and their tissues, to individual cells and subcellular compartments with defined functions. The blueprint for the spatial organisation of individual organisms is outlined in their genomes. Understanding these blueprints to define what differentiates individual organisms from one another is a fundamental task in biology. Since an organism's identity is defined by its genome, relating its structural composition directly to its genome provides deeper biological insights into how life is organised. In cell biology, we focus on understanding these relationships from the level of individual cells to tissues. One biochemical method to analyse the spatial composition of individual cells builds on subcellular fractionation. In this approach cells are split into their distinct compartments, for example by sequential centrifugation steps. The composition of each compartment can then be investigated separately. By coupling cellular fractionation to mass spectrometry (MS), this in principle allows for the unbiased identification of the subcellular localisation of all components in a cell. This technique can further be combined with perturbing a cell’s genome to directly link specific genes to their effect on subcellular composition. I demonstrated the strengths of this approach in my characterisation of the molecular mechanisms underlying activation of the immune sensor NLRP3. Using subcellular fractionation coupled to mass spectrometry, we identified the recruitment of NLRP3 to the trans-Golgi network as a key mechanism governing inflammasome activation. While this approach can generate deep biological insights, it is restricted to a comparatively low number of genes that can be investigated and provides limited spatial resolution. Light microscopy delivers much higher spatial resolution while also allowing for high-throughput analysis of composition and architecture of millions of cells. However, gaining biological insights from microscopy images is not trivial. In recent years, a new method has emerged from computer vision research that uses machine learning powered by deep neural networks to identify and compress complex patterns into a representative feature space. This approach, called deep learning, shows promise for extracting meaningful biological information from microscopy data. Another technology that allows for the investigation of spatial composition at the level of tissues is deep visual proteomics (DVP). In DVP, we use microscopy images to identify cells within the larger spatial context of tissues and analyse them further using mass spectrometry. This allows us to collect unbiased information on the molecular composition of these cells while preserving spatial information. By increasing MS sensitivity, we can even break this down to investigate the molecular composition of single cells. Using this approach, I was able to delineate key markers defining hepatocyte zonation in the liver. Taking it a step further, I used deep learning models to unbiasedly phenotype hepatocytes on the basis of their subcellular distribution of alpha-1 antitrypsin (AAT) in the fibrogenic liver disease AAT deficiency (AATD), which is characterised by the misfolding and accumulation of AAT. Combining this deep learning-driven phenotyping of cellular morphology with DVP, resulted in the identification of a terminal hepatocyte state marked by globular protein aggregates with a distinct proteomic signature, that holds promise for understanding and ultimately counteracting the molecular mechanisms underlying AATD disease progression. The above-described approaches are observational, linking distinct cellular compositions assayed using microscopy and MS to their functional implications. However, the high throughput facilitated by modern microscopes allows for the assessment of various aspects of cellular composition over millions of cells, which is compatible with a perturbational approach that looks at the effect of all coding genes on specific subcellular phenotypes. To enable this type of analysis, I developed spatially resolved CRISPR screening (SPARCS). SPARCS uses automated high-speed laser microdissection to physically isolate phenotypic variants in situ for subsequent genotyping. This enables robust, genome-wide, high-throughput screening for spatial cellular phenotypes. Using SPARCS, I was able to identify most known regulators of the cellular process of macroautophagy in a single experiment, and even identified a gene with a previously undescribed cellular phenotype. SPARCS opens up a new paradigm for investigating the genetic basis of subcellular phenotypes that can be applied to a variety of biological contexts. Finally, to facilitate the types of spatial analysis performed throughout this thesis, I developed a software platform called scPortrait that generates single-cell images from raw microscopy data. These single-cell images can be used for deep learning-based cell phenotyping, as demonstrated throughout this thesis, but also for the development of new deep learning models that generate even deeper biological insights. Completely open source and building on available open data formats, scPortrait is maximally compatible and provides a framework for the routine implementation of deep learning-based investigation of cellular composition across various areas of biology.
CRISPR genome editing, deep learning, image-based phenotyping, single-cell analysis, SPARCS, scPortrait, computational biology
Mädler, Sophia Clara
2024
Englisch
Universitätsbibliothek der Ludwig-Maximilians-Universität München
Mädler, Sophia Clara (2024): Integrating deep learning and genetic approaches to uncover molecular mechanisms of cellular organization. Dissertation, LMU München: Fakultät für Chemie und Pharmazie
[thumbnail of Maedler_Sophia_Clara.pdf]
Vorschau
PDF
Maedler_Sophia_Clara.pdf

78MB

Abstract

Organisation is a fundamental principle of life. Matter needs to be arranged in space in such a way that it enables reproduction. In biology, this organisation occurs at different scales: from whole ecosystems to multicellular organisms and their tissues, to individual cells and subcellular compartments with defined functions. The blueprint for the spatial organisation of individual organisms is outlined in their genomes. Understanding these blueprints to define what differentiates individual organisms from one another is a fundamental task in biology. Since an organism's identity is defined by its genome, relating its structural composition directly to its genome provides deeper biological insights into how life is organised. In cell biology, we focus on understanding these relationships from the level of individual cells to tissues. One biochemical method to analyse the spatial composition of individual cells builds on subcellular fractionation. In this approach cells are split into their distinct compartments, for example by sequential centrifugation steps. The composition of each compartment can then be investigated separately. By coupling cellular fractionation to mass spectrometry (MS), this in principle allows for the unbiased identification of the subcellular localisation of all components in a cell. This technique can further be combined with perturbing a cell’s genome to directly link specific genes to their effect on subcellular composition. I demonstrated the strengths of this approach in my characterisation of the molecular mechanisms underlying activation of the immune sensor NLRP3. Using subcellular fractionation coupled to mass spectrometry, we identified the recruitment of NLRP3 to the trans-Golgi network as a key mechanism governing inflammasome activation. While this approach can generate deep biological insights, it is restricted to a comparatively low number of genes that can be investigated and provides limited spatial resolution. Light microscopy delivers much higher spatial resolution while also allowing for high-throughput analysis of composition and architecture of millions of cells. However, gaining biological insights from microscopy images is not trivial. In recent years, a new method has emerged from computer vision research that uses machine learning powered by deep neural networks to identify and compress complex patterns into a representative feature space. This approach, called deep learning, shows promise for extracting meaningful biological information from microscopy data. Another technology that allows for the investigation of spatial composition at the level of tissues is deep visual proteomics (DVP). In DVP, we use microscopy images to identify cells within the larger spatial context of tissues and analyse them further using mass spectrometry. This allows us to collect unbiased information on the molecular composition of these cells while preserving spatial information. By increasing MS sensitivity, we can even break this down to investigate the molecular composition of single cells. Using this approach, I was able to delineate key markers defining hepatocyte zonation in the liver. Taking it a step further, I used deep learning models to unbiasedly phenotype hepatocytes on the basis of their subcellular distribution of alpha-1 antitrypsin (AAT) in the fibrogenic liver disease AAT deficiency (AATD), which is characterised by the misfolding and accumulation of AAT. Combining this deep learning-driven phenotyping of cellular morphology with DVP, resulted in the identification of a terminal hepatocyte state marked by globular protein aggregates with a distinct proteomic signature, that holds promise for understanding and ultimately counteracting the molecular mechanisms underlying AATD disease progression. The above-described approaches are observational, linking distinct cellular compositions assayed using microscopy and MS to their functional implications. However, the high throughput facilitated by modern microscopes allows for the assessment of various aspects of cellular composition over millions of cells, which is compatible with a perturbational approach that looks at the effect of all coding genes on specific subcellular phenotypes. To enable this type of analysis, I developed spatially resolved CRISPR screening (SPARCS). SPARCS uses automated high-speed laser microdissection to physically isolate phenotypic variants in situ for subsequent genotyping. This enables robust, genome-wide, high-throughput screening for spatial cellular phenotypes. Using SPARCS, I was able to identify most known regulators of the cellular process of macroautophagy in a single experiment, and even identified a gene with a previously undescribed cellular phenotype. SPARCS opens up a new paradigm for investigating the genetic basis of subcellular phenotypes that can be applied to a variety of biological contexts. Finally, to facilitate the types of spatial analysis performed throughout this thesis, I developed a software platform called scPortrait that generates single-cell images from raw microscopy data. These single-cell images can be used for deep learning-based cell phenotyping, as demonstrated throughout this thesis, but also for the development of new deep learning models that generate even deeper biological insights. Completely open source and building on available open data formats, scPortrait is maximally compatible and provides a framework for the routine implementation of deep learning-based investigation of cellular composition across various areas of biology.