22455 research outputs found
Sort by
GARP derived TGFβ promotes acetylation-mediated Foxp3 protein stabilization and Treg functionality
Get PDFDiseases of immune dysregulation are frequently caused by single gene mutations in central pathways of immune tolerance. They are characterised by heterogeneous clinical manifestations with autoimmune symptoms based on defects in mechanisms regulating self-tolerance. Identification of the genetic cause of such diseases has critical implications for the treatment of patients. To identify underlying molecular etiologies, patients suffering from immune dysregulation were subjected to whole-exome sequencing. Two patients bearing previously undescribed mutations of LRRC32 encoding glycoprotein A repetitions predominant (GARP) were identified. GARP has been recently described to be specifically expressed on regulatory T cells (Tregs) and is important for their suppressive capacity. GARP is involved in TGFβ1 activation by binding latent TGFβ1 in the cytoplasm and translocating it to the cell surface. The importance of fully functional Tregs is well described for maintaining immunological self-tolerance and homeostasis of the immune system. Tregs control destructive immune responses against pathogens and limit reactions towards self antigens. However, the function of this TGFβ1 source and the underlying mechanisms are not yet completely understood.
My PhD thesis presents data that reinforces the importance of GARP as a regulator of Treg function and stability. It describes two patients with LRRC32 mutations, characterises the function of GARP in cellular and molecular detail and demonstrates the importance of functional Tregs for physiological immune homeostasis in men and mice. The study identifies a novel link between GARP dependent TGFβ signalling in Tregs and expression of the Treg specific histone deacetylase (HDAC)9 that promotes Foxp3 deacetylation, which contributes to an instable dysfunctional Treg phenotype.
The research shows that Tregs from patients with LRRC32 mutations have only minimal GARP expression on the cell surface and reduced TGFβ signalling. Tregs from these patients further show a strongly diminished Treg suppressor function and significant reduction in Treg numbers and frequency. GARP functions are characterised in a novel molecular detail using a model of conditional Garp-deficiency in mice. Here the study confirms increased susceptibility to inflammatory diseases once GARP expression is decreased on Tregs. Consistent with the effects observed in patients, Garp-deficiency in mice leads to absence of latent TGFβ on the cell surface of Tregs, reduced TGFβ signalling and diminished suppressor function. Further, Treg from Garp-deficient mice exhibit an unstable phenotype due to diminished Foxp3 protein acetylation and stability. In sum, the PhD thesis reinforces the understanding of immunological mechanisms of immune dysregulation and expands the knowledge of immunological function of GARP as an important regulator of Treg stability
Vergleich von Tacrolimus zweimal täglich mit retardiertem Tacrolimus einmal täglich bei einem pädiatrischen Patientenkollektiv im Verlauf nach Herztransplantation
Get PDFEndovascular treatment of renal bleeding: Technical success, clinical tolerability and overall success
Get PDFGestational diabetes mellitus in Tajikistan
Get PDFBackground: Gestational diabetes mellitus (GDM) is currently the most common medical complication of pregnancy worldwide. The prevalence of undiagnosed hyperglycemia and even overt diabetes in women of reproductive age is increasing. No data are available on the prevalence of GDM in Tajikistan. This study evaluated the prevalence of GDM and the obstetric and neonatal outcomes of pregnancies in an urban and a rural setting of Tajikistan.
Methods: Cross-sectional study conducted among pregnant women presented to the the Reproductive Health Centers in Dushanbe and Qurghonteppa between December 2015 and May 2018. Pregnant women were included in the study during the 1st trimester of pregnancy. The collection of data was carried out according to a specially structured questionnaire, where data on visits were recorded. Between weeks 24 and weeks 28 of gestation an oral glucose tolerance test (OGTT) with 75 g of glucose was performed. GDM was diagnosed if any one of the venous plasma glucose values was met or exceeded (fasting ≥ 5.1 mmol/L; at 60 min ≥ 10.0 mmol/L and at 120 min ≥ 8.5-11.0 mmol/L). Obstetric and neonatal otcomes were recorded after delivery. H-Kruskal-Wallis, Mann-Whitney U and Chi-square tests were used.
Results: Of 2643 women (age 25.3±5.3 years, BMI 23.8±4.3 kg/m2), 92.2% underwent an OGTT and of these 29.7% had elevated fasting plasma glucose values (mostly minimally elevat-ed), while 2.8% had elevated 60 min and/or 120 min values. The overall prevalence of GDM was 32.4%. Age (p=0.001), weight (p=0.001), BMI (p=0.002) and parity (p=0.012) were associated with GDM. The obstetric and neonatal outcome of women with only elevated fasting glucose levels was not different from women with normal glucose levels. Women with abnormal blood glucose concentration after 60 min and/or 120 min had a significantly higher rate of complica-tions, threatening miscarriage, infection of urinary tract and emergency Cesarean section while affected newborns had lower birth weight, lower APGAR and lower 30 min glucose levels.
Discussion: The study determined for the first time the prevalence of GDM in Tajikistan both in urban and rural areas. Most cases of GDM were diagnosed on the basis of slightly elevated fast-ing glucose level, which was not associated with adverse obstetric or neonatal outcomes, while women and neonates from women with elevated 60 min or 120 min values had significantly more complications. These findings are in agreement with recent studies from Denmark and the USA, showing that very mild forms of GDM (identified by slightly elevated fasting glucose levels) are not associated with an adverse outcome.
Conclusion: Although the formal prevalence of GDM is high in Tajikistan, the applicability of the one step OGTT for the screening and diagnosis of GDM must be questioned, as most of the identified women have a normal pregnancy outcome. At the same time this strategy puts the burden of receiving a diagnosis of GDM on individual women and the burden of treating many more women on the health care system. A two-step screening or a one step screening in women with risk factors for GDM maybe a better strategy in a setting were the prevalence of « severe » GDM is low
Democratizing machine learning
Get PDFModelle des maschinellen Lernens sind zunehmend in der Gesellschaft verankert, oft in Form von automatisierten Entscheidungsprozessen. Ein wesentlicher Grund dafür ist die verbesserte Zugänglichkeit von Daten, aber auch von Toolkits für maschinelles Lernen, die den Zugang zu Methoden des maschinellen Lernens für Nicht-Experten ermöglichen.
Diese Arbeit umfasst mehrere Beiträge zur Demokratisierung des Zugangs zum maschinellem Lernen, mit dem Ziel, einem breiterem Publikum Zugang zu diesen Technologien zu er- möglichen. Die Beiträge in diesem Manuskript stammen aus mehreren Bereichen innerhalb dieses weiten Gebiets. Ein großer Teil ist dem Bereich des automatisierten maschinellen Lernens (AutoML) und der Hyperparameter-Optimierung gewidmet, mit dem Ziel, die oft mühsame Aufgabe, ein optimales Vorhersagemodell für einen gegebenen Datensatz zu finden, zu vereinfachen. Dieser Prozess besteht meist darin ein für vom Benutzer vorgegebene Leistungsmetrik(en) optimales Modell zu finden. Oft kann dieser Prozess durch Lernen aus vorhergehenden Experimenten verbessert oder beschleunigt werden.
In dieser Arbeit werden drei solcher Methoden vorgestellt, die entweder darauf abzielen, eine feste Menge möglicher Hyperparameterkonfigurationen zu erhalten, die wahrscheinlich gute Lösungen für jeden neuen Datensatz enthalten, oder Eigenschaften der Datensätze zu nutzen, um neue Konfigurationen vorzuschlagen.
Darüber hinaus wird eine Sammlung solcher erforderlichen Metadaten zu den Experimenten vorgestellt, und es wird gezeigt, wie solche Metadaten für die Entwicklung und als Testumgebung für neue Hyperparameter- Optimierungsmethoden verwendet werden können. Die weite Verbreitung von ML-Modellen in vielen Bereichen der Gesellschaft erfordert gleichzeitig eine genauere Untersuchung der Art und Weise, wie aus Modellen abgeleitete automatisierte Entscheidungen die Gesellschaft formen, und ob sie möglicherweise Individuen oder einzelne Bevölkerungsgruppen benachteiligen. In dieser Arbeit wird daher ein AutoML-Tool vorgestellt, das es ermöglicht, solche Überlegungen in die Suche nach einem optimalen Modell miteinzubeziehen. Diese Forderung nach Fairness wirft gleichzeitig die Frage auf, ob die Fairness eines Modells zuverlässig geschätzt werden kann, was in einem weiteren Beitrag in dieser Arbeit untersucht wird. Da der Zugang zu Methoden des maschinellen Lernens auch stark vom Zugang zu Software und Toolboxen abhängt, sind mehrere Beiträge in Form von Software Teil dieser Arbeit. Das R-Paket mlr3pipelines ermöglicht die Einbettung von Modellen in sogenan- nte Machine Learning Pipelines, die Vor- und Nachverarbeitungsschritte enthalten, die im maschinellen Lernen und AutoML häufig benötigt werden. Das mlr3fairness R-Paket hingegen ermöglicht es dem Benutzer, Modelle auf potentielle Benachteiligung hin zu über- prüfen und diese durch verschiedene Techniken zu reduzieren. Eine dieser Techniken, multi-calibration wurde darüberhinaus als seperate Software veröffentlicht.Machine learning artifacts are increasingly embedded in society, often in the form of automated decision-making processes. One major reason for this, along with methodological improvements, is the increasing accessibility of data but also machine learning toolkits that enable access to machine learning methodology for non-experts. The core focus of this thesis is exactly this – democratizing access to machine learning in order to enable a wider audience to benefit from its potential.
Contributions in this manuscript stem from several different areas within this broader area. A major section is dedicated to the field of automated machine learning (AutoML) with the goal to abstract away the tedious task of obtaining an optimal predictive model for a given dataset. This process mostly consists of finding said optimal model, often through hyperparameter optimization, while the user in turn only selects the appropriate performance metric(s) and validates the resulting models. This process can be improved or sped up by learning from previous experiments.
Three such methods one with the goal to obtain a fixed set of possible hyperparameter configurations that likely contain good solutions for any new dataset and two using dataset characteristics to propose new configurations are presented in this thesis.
It furthermore presents a collection of required experiment metadata and how such meta-data can be used for the development and as a test bed for new hyperparameter optimization methods. The pervasion of models derived from ML in many aspects of society simultaneously calls for increased scrutiny with respect to how such models shape society and the eventual biases they exhibit. Therefore, this thesis presents an AutoML tool that allows incorporating fairness considerations into the search for an optimal model. This requirement for fairness simultaneously poses the question of whether we can reliably estimate a model’s fairness, which is studied in a further contribution in this thesis. Since access to machine learning methods also heavily depends on access to software and toolboxes, several contributions in the form of software are part of this thesis. The mlr3pipelines R package allows for embedding models in so-called machine learning pipelines that include pre- and postprocessing steps often required in machine learning and AutoML. The mlr3fairness R package on the other hand enables users to audit models for potential biases as well as reduce those biases through different debiasing techniques. One such technique, multi-calibration is published as a separate software package, mcboost
Multilingual representations and models for improved low-resource language processing
Get PDFWord representations are the cornerstone of modern NLP. Representing words or characters using real-valued vectors as static representations that can capture the Semantics and encode the meaning has been popular among researchers. In more recent years, Pretrained Language Models using large amounts of data and creating contextualized representations achieved great performance in various tasks such as Semantic Role Labeling. These large pretrained language models are capable of storing and generalizing information and can be used as knowledge bases.
Language models can produce multilingual representations while only using monolingual data during training. These multilingual representations can be beneficial in many tasks such as Machine Translation. Further, knowledge extraction models that only relied on information extracted from English resources, can now benefit from extra resources in other languages.
Although these results were achieved for high-resource languages, there are thousands of languages that do not have large corpora. Moreover, for other tasks such as machine translation, if large monolingual data is not available, the models need parallel data, which is scarce for most languages. Further, many languages lack tokenization models, and splitting the text into meaningful segments such as words is not trivial. Although using subwords helps the models to have better coverage over unseen data and new words in the vocabulary, generalizing over low-resource languages with different alphabets and grammars is still a challenge.
This thesis investigates methods to overcome these issues for low-resource languages. In the first publication, we explore the degree of multilinguality in multilingual pretrained language models. We demonstrate that these language models can produce high-quality word alignments without using parallel training data, which is not available for many languages. In the second paper, we extract word alignments for all available language pairs in the public bible corpus (PBC). Further, we created a tool for exploring these alignments which are especially helpful in studying low-resource languages. The third paper investigates word alignment in multiparallel corpora and exploits graph algorithms for extracting new alignment edges. In the fourth publication, we propose a new model to iteratively generate cross-lingual word embeddings and extract word alignments when only small parallel corpora are available. Lastly, the fifth paper finds that aggregation of different granularities of text can improve word alignment quality. We propose using subword sampling to produce such granularities
Epigenetische Veränderungen bei der Geweberegeneration und der Onkogen-vermittelten Tumorentstehung im Pankreas und in anderen Tumoren
Get PDFMolecular dissection of pericyte-to-neuron reprogramming reveals cellular identity safeguarding mechanisms
Get PDFNeurodegenerative diseases, strokes, and injuries affect millions of people worldwide and current treatment options are insufficient. Since death of neurons in the brain is a common feature of all these disorders, a potential therapy could replace the lost neurons by newly generated ones to restore brain function. Natural adult neurogenesis in humans has been proven inadequate to deal with a major loss of brain cells. Therefore, for many years, transplantation of fetal tissue or stem cell-derived neural progenitors have been the focus of investigations regarding new treatments. More recently, new methods and insights have rendered brain-resident cells a promising means of an alternative therapeutic approach. While cellular identity was believed to be irreversible once differentiated for a long time, this view has changed gradually over the last decades. Among other cells, it has been shown for human brain pericytes that retroviral expression of the transcription factors (TFs) Ascl1 and Sox2 (AS) is sufficient to generate functional induced neurons (iNs) by direct reprogramming, and that this process is accompanied by a neural stem cell (NSC)-like state.
While it is clear now that even a terminal cellular identity can be changed, the exact mechanisms remain elusive. Therefore, in this study we aimed at (i) identifying barriers and molecular mechanisms involved in cellular identity conversion from somatic cells into induced neurons, (ii) improving the efficiency of pericyte-to-neuron reprogramming, and (iii) directing the reprogramming process towards the desired cell types.
By single cell RNA sequencing, we generated a high-resolution dataset of cells during pericyte-to-iN conversion. Using RNA velocity analysis, we were able to predict the progression of cells towards the neuronal fate and could identify blocker and facilitator genes that obstruct or enable cells to pass past a designated decision point. Among the facilitator genes, we identified several chromatin remodelers and cytoskeleton genes, and revealed a temporal heterogeneity regarding their expression pattern. Interestingly, we show that the blocker genes are part of a cellular identity safeguarding mechanism triggered by AS reprogramming. We demonstrate that the metabolic transition from glycolysis to oxidative phosphorylation is an essential barrier cells must overcome to transit from a pericyte towards a neuronal identity. Our findings suggest that any failure to meet metabolic requirements results in cells being either unable to change their identity or adopting a confused fate.
To impact on the NSC-like state, we used either modulation of NOTCH signaling or TGF-β signaling by inhibition of the γ-secretase or dual SMAD inhibition, respectively, via small molecules. Strikingly, both treatments counteracted pericyte identity safeguarding mechanisms and significantly lowered reprogramming barriers. Consequently, our results show a strong increase in the number of generated iNs. Interestingly, we demonstrate that TGF-β signaling inhibition is more potent in lowering these metabolic barriers than NOTCH signaling inhibition, re-routing cells onto an entirely different route towards neurons. Additionally, TGF-β signaling inhibition almost completely suppresses the generation of undesired off-target cells without a clear identity, likely due to antioxidant regulon activity, which supports the metabolic transition.
Remarkably, we illustrate that despite different treatments, iNs are transcriptionally similar and that both neuronal subtypes can be mapped to developing human brain regions. Finally, we used a different approach and reprogrammed pericytes into TUBB3+ cells using Neurog2/Sox2 (NS). We show that NS generated cells have a distinct transcriptomic identity from AS generated ones: While they are more likely to lose their original identity, the NS-generated iNs exhibit more progenitor-like properties, pointing at the different reprogramming capacities of proneural TFs.
Altogether, this thesis emphasizes not only that cellular identity even in terminally differentiated cells can still be altered without returning to a pluripotent state. It further illustrates several previously unknown mechanisms during direct pericyte-to-iN reprogramming and opens new ways to improve its efficiency. Every new insight into cross-lineage cellular identity conversion paves the way for future neuronal replacement therapies