Dokument: TopSuite: A meta-suite for protein structure prediction using deep neural networks

Titel:	TopSuite: A meta-suite for protein structure prediction using deep neural networks
URL für Lesezeichen:	https://docserv.uni-duesseldorf.de/servlets/DocumentServlet?id=54861
URN (NBN):	urn:nbn:de:hbz:061-20201125-114030-3
Kollektion:	Dissertationen
Sprache:	Englisch
Dokumententyp:	Wissenschaftliche Abschlussarbeiten » Dissertation
Medientyp:	Text
Autor:	Mulnaes, Daniel [Autor]
Dateien:	[Dateien anzeigen] Adobe PDF [Details] 40,03 MB in einer Datei [ZIP-Datei erzeugen] Dateien vom 24.11.2020 / geändert 24.11.2020
Beitragende:	Prof. Dr. Gohlke, Holger [Gutachter] Jun.-Prof. Dr. Schröder, Gunnar [Gutachter]
Stichwörter:	protein structure prediction neural network TopSuite topsuite
Dewey Dezimal-Klassifikation:	600 Technik, Medizin, angewandte Wissenschaften » 610 Medizin und Gesundheit
Beschreibung:	All known life to date depends on proteins. Proteins are essential molecular machines that participate in every process within cells and play vital roles in, for example, cell structure, cell signalling, cell division, motor function, metabolism, and immune responses. Proteins owe their diverse functions to their vast array of different structures. Knowing the 3D structure of a protein is therefore a critical step towards understanding its function. That knowledge can, in turn, help researchers to figure out how to modulate the protein function, leading to new and/or improved drugs or cleaner and more environmentally sustainable industrial processes. At present, resolving a protein structure experimentally is laborious, time consuming and cost intensive. Therefore, being able to predict a protein structure accurately using computational methods is of high interest in biochemical, biomedical, and biotechnological research. Many computational structure prediction methods have been developed in the last two decades, but no single method is consistently the best for every protein. Since different methods use different ideas, databases, algorithms and machine learning techniques, they provide different answers to the same types of problems. Consequently, integrating multiple so-called primary methods into a single meta-method harnesses their strengths and counteracts their weaknesses. This results in more robust and accurate structure predictions. However, when the majority of primary methods consent on the wrong prediction, out-numbering those that make the right one, meta-methods that rely on majority consensus make wrong structure predictions. The goal of this thesis is to provide a toolbox and fully automated protein structure prediction workflow called TopSuite. This workflow consists of multiple meta-methods, each of which solve different tasks for protein structure prediction. Rather than using consensus though, these meta-tools make use of deep neural networks that have been trained on large datasets to learn when, and how much, to trust each primary method. As such, the TopSuite meta-methods are able to go against the majority when needed, and yield predictions that are significantly better than any of their respective primary methods. Furthermore, the utility of TopSuite, in particular the template-based structure prediction workflow TopModel, is demonstrated through the application to target proteins of high biological, medical, and industrial importance.
Lizenz:	Urheberrechtsschutz
Fachbereich / Einrichtung:	Mathematisch- Naturwissenschaftliche Fakultät » WE Pharmazie » Pharmazeutische und Medizinische Chemie
Dokument erstellt am:	25.11.2020
Dateien geändert am:	25.11.2020
Promotionsantrag am:	11.02.2020
Datum der Promotion:	23.10.2020