Dokument: Bi-Level optimization algorithms for improving genome-scale metabolic models

Titel:	Bi-Level optimization algorithms for improving genome-scale metabolic models
URL für Lesezeichen:	https://docserv.uni-duesseldorf.de/servlets/DocumentServlet?id=57499
URN (NBN):	urn:nbn:de:hbz:061-20210927-111114-3
Kollektion:	Dissertationen
Sprache:	Englisch
Dokumententyp:	Wissenschaftliche Abschlussarbeiten » Dissertation
Medientyp:	Text
Autor:	Hartleb, Daniel [Autor]
Dateien:	[Dateien anzeigen] Adobe PDF [Details] 40,42 MB in einer Datei [ZIP-Datei erzeugen] Dateien vom 21.09.2021 / geändert 21.09.2021
Beitragende:	Prof. Dr. Lercher, Martin [Betreuer/Doktorvater] Prof. Dr. Ebenhöh, Oliver [Gutachter]
Dewey Dezimal-Klassifikation:	000 Informatik, Informationswissenschaft, allgemeine Werke » 004 Datenverarbeitung; Informatik
Beschreibung:	Genome-scale metabolic models are reconstructed for many organisms. They are routinely used to predict metabolic behavior, simulate evolutionary adaptation, and help to design organisms of bioengineering interest. However, the quality of metabolic models is highly variable. Typically, metabolic models are refined by comparing in silico predictions to in vivo experiments (e.g., viability of gene knock-outs or growth in different nutritional environments). Several different algorithms have been developed to resolve the resulting inconsistencies between prediction and experiment. However, these tools can iteratively correct only one inconsistency at a time. Thus, the total number of network changes may not be globally optimal, a modification introduced earlier might prevent the resolution of other inconsistencies, or a potential solution might not be found because the combination of different types of network changes is not supported. In Manuscript 1, a novel bi-level optimization algorithm – GlobalFit – is introduced. GlobalFit is the first algorithm that can simultaneously solve multiple inconsistencies at a time and allows the combination of different network modifications. We applied the algorithm to the genome-scale metabolic model for Mycoplasma genitalium,improving the overall accuracy for viability predictions from 87.3% to 97.3%. Interestingly, solving all inconsistencies at a time resulted in the same network changes as iteratively solving each erroneous prediction together with a corresponding counter-case, while the overall time for solving decreased dramatically. Applying this subset strategy to the much better curated genome-scale metabolic model for Escherichia coli, we could again substantially improve the accuracy, from 90.8% to 95.4%. Reconstructing metabolic models is still a laborious and time-consuming task. To accelerate this process, automatic reconstruction algorithms have been developed. However, predictions by automatically reconstructed networks generally have low accuracy and still need to be refined manually. In Manuscript 2, a novel pipeline is introduced, which gathers information from metabolic networks from closely related organisms and metabolic databases (i.e., KBase, TransportDB). At each step, the metabolic information of each gene is replaced by newer information. Finally, the draft metabolic network is refined with GlobalFit based on genome-wide gene knock-out data. We demonstrate the applicability of this pipeline by reconstructing genomescale metabolic models for three different Streptococci genomes. The predictive power of the resulting metabolic models was of the same quality as for manually curated models (e.g., E. coli iJO1366). In addition to the low predictive power of automatically reconstructed metabolic models, they often contain internal energy generating cycles. These cycles can charge energy-rich metabolites such as ATP without the uptake of any nutrients. Thus, they can severely affect the energy metabolism of the model and can unrealistically inflate the maximal biomass production. However, no systematic method to eliminate those cycles had been developed previously. In Manuscript 3, a variant of FBA is described to identify energy generating cycles, and a modified version of GlobalFit is subsequently used to eliminate the detected cycles. We could identify energy generating cycles in 65% of metabolic networks from three different databases, and ModelSEED). In the following step, GlobalFit could fully eliminate energy generating cycles in 94% of the affected metabolic models.
Lizenz:	Urheberrechtsschutz
Fachbereich / Einrichtung:	Mathematisch- Naturwissenschaftliche Fakultät » WE Informatik » Bioinformatik
Dokument erstellt am:	27.09.2021
Dateien geändert am:	27.09.2021
Promotionsantrag am:	10.07.2019
Datum der Promotion:	20.09.2021