22455 research outputs found
Sort by
Studying stress-related epigenetic and metabolic changes in fruit flies and mice through monogenic and polygenic models
Get PDFEpigenetic factors and metabolic intermediates have been at the forefront of chromatin research in the past few years. The communication between metabolic process and epigenetic factors with the genetic component is especially important to translate environmental changes to the organisms to either improve its survival in short term or provide the ability to adapt in the longer run. We studied these aspects using a poikilothermic species such as Drosophila melanogaster (Publication-1) and a higher organism, Mus musculus, which are more closely related to humans (Publication-2).
In the first study, Venkatasubramani et al., 2023, we aimed to understand the evolutionary importance of chameau (chm), a MYST-domain acetyltransferase. We have previously shown that absence of this protein improves both physical activity and longevity in Drosophila melanogaster (Peleg, Feller, Forne, et al., 2016). We employed UAS-GAL4 system to knockdown the gene both ubiquitously and tissue-specifically. Upon observing massive physiological changes with ubiquitous, neuronal and fatbody loss of chm, we also confirmed the importance of MYST-domain, i.e. acetyltransferase activity, using a heterozygous mutant with a deletion of MYST genomic region that also showed similar observation. Following this, we used a multi-omics strategy to assess the effect of mutation (or knockdown) on transcriptome, histone acetylome, proteome and non-histone acetylome. Observations from these approaches pointed towards misregulated metabolism. We therefore tested the ability of organisms lacking the protein to respond to metabolic stress by administering wet-starvation and as expected, these flies showed reduced starvation resilience. Interestingly, this was independent of chm’s role in development as assessed using GeneSwitch system, where chm knockdown is induced in an adult-specific manner. Further, we also validated the perceived role of chm in metabolism and stress response by over-expressing the protein ubiquitously and in different tissues via the UAS-GAL4 system. All these experiments resulted in an improved survival upon metabolic stress and specifically upon over-expression in tissues such as neurons and fatbody. In addition, we have evidence that support this putative evolutionary role of chm in stress response is limited only to certain temperature ranges, which could be considered non-ideal for the organism. This is also supported by evidence of sequence variation and balancing selection in different climatic populations (Levine & Begun, 2008; Croze et al., 2017). This is being addressed in a follow-up study that is in preparation.
The second study was a collaborative effort with Peleg lab. In this study, (Müller-Eigner et al., 2022), we characterized a unique and novel mouse model named Titan that has been selected for over 180 generations based on their high body mass. After assessing their genetic, biochemical and physiological status, we perceived that these mouse show similarities to metabolic syndrome in humans. We further compared Titan mice to unselected mice and characterized histone PTMs and subsequently proteome and transcriptome, all of which pointed towards metabolic differences. All these were assessed in two ages in unselected and Titan mouse as the latter also showed a steep decline in lifespan. Intriguingly, genetic analysis showed variation in acetyltransferase genes and accordingly, acetylation of H4 was altered between the mouse models. Finally, we employed dietary intervention, an intervention that may result in extended life span to test if these mice can display improved survival and rescue associated molecular changes. This study characterized and provided an additional animal model for studying metabolic diseases like obesity and showed the importance of having a non-inbred model that displays genetic variability. The study also provided an interesting take on reversing the possible effects of genetic differences via changes in lifestyle.
Altogether, these studies provide an understanding of metabolic dysregulation in organisms, either upon mutation or phenotypic selection. In particular, the evolutionary importance and a previously unknown role of chm in metabolism of Drosophila melanogaster and the characterization of a long-term phenotypically selected Mus musculus model, Titan mice, as a novel and unique resource for studying metabolic disorders that closely reflects conditions in human
Untersuchungen zur prognostischen Bedeutung des Rezeptors TAAR1 und zum Proliferationsverhalten von TAAR1 Liganden in Mammakarzinomen in vitro
Get PDF
Artificial Intelligence for automated decision-making in error-pattern recognition
Get PDFThe demand for integrating Artificial Intelligence (AI) into diverse systems continues to expand rapidly. With growing reliance on AI, there is a constant need to deliver increasingly more dependable and robust solutions. This thesis aims to address common Machine Learning (ML) challenges and offers solutions in the field of automated decision-making and pattern recognition. It explores, presents, and analyzes novel methods and algorithms for Representation Learning, focusing on "Neural Architecture & Data Representation", "Manifold & Embedding Learning", and "Data Integration & Analysis".
The contributions of this thesis include: (1) a novel neural architecture based on complex-valued neural networks, (2) a framework for encoding hierarchical one-to-many relationship databases into contextualized numeric data representations, (3) an adaptive and robust feature normalization and pre-processing technique, (4) a method for synthetic data augmentation on hierarchical databases, (5) a contrastive Representation Learning framework for tree structures, (6) extending and generalizing hierarchical Embedding Learning to multiple data views, (7) methods and diverse loss functions for Manifold Representation Learning and Clustering, (8) approaches for automated textual description generation for cluster groups, (9) a quality evaluation metric for clustering under coarse label uncertainty, (10) methods for determining representative textual labels for clustering accurate sensor data with inexact annotations, (11) a Transfer Learning approach to distilling insights from ML models trained on different data sets, (12) a novel approach to Out-of-Distribution Detection and Novelty Detection by leveraging mismatches in model calibration alignment, and (13) a method for representative sampling within quantile windows in data streams of unknown final length
Towards a better understanding of chicken macrophage function in interkingdom signalling with postbiotic SCFAs
Get PDF
Verarbeitung von Sequenzen mit Neuronalen Netzen
Get PDFDie Verarbeitung von geordneten Daten (Sequenzen) hat seit Längerem in vielen Bereichen der Informatik und damit auch in unserem alltäglichen Leben Einzug gehalten. Sei es bei offensichtlichen Anwendungsfällen, wie der Verarbeitung von Messreihen und der Vorhersage von Trends, z. B. in meteorologischen Systemen, als auch bei nicht offensichtlichen Anwendungen, wie der Interpretation nicht sequenzieller Datentypen. Wie in vielen anderen Teilbereichen der Informatik hat die Anwendung selbstlernender Algorithmen auch hier aufgrund ihrer ausgezeichneten Generalisierungsfähigkeiten jüngst zugenommen. Dabei ist das Lernen an viele Einflüsse gebunden, so eignen sich viele bekannte Algorithmen z. B. für einen konkreten Anwendungsfall, verfehlen aber ihr Leistungsziel bei artverwandten Datensätzen. Hier müssen passende Hyperparameter-Konfigurationen oder andere, geeignetere Architekturen gefunden werden. Die vorliegende Arbeit bearbeitet Fragestellungen aus diesem Teilbereich der Anwendung tiefer neuronaler Netzwerke im Kontext sequenzieller Daten (Deep Neural Network) in fünf Kapiteln.
In der vorliegenden Arbeit wird zunächst die generelle Funktionsweise des Informationstransports im sog. Hidden Vector untersucht. Dies spielt bei der Verarbeitung von Sequenzen typischen Architektur sog. Recurrent Neural Network eine besondere Rolle. Dabei lässt sich eine harte Grenze der sog. Memory Horizon feststellen, die mit der Anzahl an verwendeten Neurons korreliert. Auch verwandte Zelle-Architekturen werden hierbei untersucht. Anschließend beschäftigt sich diese Arbeit mit alternativen Architekturen in der Sequenzverarbeitung im Bereich der Audio-Klassifizierung. So werden zwei Architekturen für die Verarbeitung von Mel-Spektrogrammen von Audio-Signalen aus Real-World-Datasets eingeführt und im Kontext bestehender Ansätze, unter Beachtung verschiedener Data Augmentation (DA)-Methoden, evaluiert. Dann werden etablierte Data Augmentation-Methoden in einem neuen Verfahren („Transport“) zusammengefasst und evaluiert. Erste Ergebnisse lassen die Hoffnung zu, dass dieser neue Ansatz nicht nur die Varianz steigern, sondern auch einen Leistungsgewinn im Bereich der Audio-Klassifizierung erzielen kann. Schließlich werden Self-Replicating Neural Network betrachtet, deren Lerndaten sich fortlaufend (sequenziell) verändern, bis diese auf sich selbst abgebildet werden können. Durch die Erweiterung dieses Forschungsfeldes auf Aufgaben aus dem Bereich des Deep Learning ist das Phänomen des natürlichen Dropout (Pruning), durch einen aus solchen Particle Network zusammengesetzten Organismus (Organism Network) zu beobachten.The processing of ordered data (sequences) has been used for a long time in many areas of computer science, and thus also in our everyday life. Be it in obvious use cases, such as the processing of series of measurements and the prediction of trends, meteorological systems e.g., as well as in non-obvious applications, such as the interpretation of non-sequential data types. As in many other subfields of computer science, the use of self-learning algorithms has recently increased due to their excellent generalization capabilities. Thereby, learning is bound to many influences, e.g., many known algorithms are suitable for a concrete use case, but fail to achieve their performance goal on related or slightly altered data sets. Then, only suitable hyperparameter configurations or entirely other more appropriate architectures must be found. This thesis addresses issues from this subfield of applying deep neural networks in the context of sequential data in five chapters.
In this thesis, we first examine the general operation of information transport in hidden vectors. A typical problem when dealing with architectures (Recurrent Neural Networks), specialized in processing sequences. A hard limit, the so called memory horizon, can be found, which correlates with the number of neurons used. Related cell types are also examined here. Next, this thesis deals with alternative architectures in sequence processing in the field of audio classification. Thus, two candidates for processing mel-spectrograms of audio signals from real-world datasets are introduced and evaluated in the context of existing approaches, considering different data augmentation techniques. Then, established data augmentation methods are combined and evaluated in a newly proposed method („transport“). Initial results give hope that this new approach can not only increase variance, but also gain performance in audio classification. Finally, we consider self-replicating neural networks, which learn on continuously changing data sources (sequentially) until it can be mapped onto itself. By extending this field of research to tasks from the imag
