Dokument: Prediction of enzyme kinetic parameters and substrate scopes using artificial intelligence
| Titel: | Prediction of enzyme kinetic parameters and substrate scopes using artificial intelligence | |||||||
| URL für Lesezeichen: | https://docserv.uni-duesseldorf.de/servlets/DocumentServlet?id=64883 | |||||||
| URN (NBN): | urn:nbn:de:hbz:061-20240209-090801-6 | |||||||
| Kollektion: | Dissertationen | |||||||
| Sprache: | Englisch | |||||||
| Dokumententyp: | Wissenschaftliche Abschlussarbeiten » Dissertation | |||||||
| Medientyp: | Text | |||||||
| Autor: | Kroll, Alexander [Autor] | |||||||
| Dateien: |
| |||||||
| Beitragende: | Prof. Dr. Lercher, Martin [Gutachter] Prof. Dr. Kollmann, Markus [Gutachter] | |||||||
| Dewey Dezimal-Klassifikation: | 000 Informatik, Informationswissenschaft, allgemeine Werke » 004 Datenverarbeitung; Informatik | |||||||
| Beschreibung: | Cells are the building blocks for all living organisms on earth. In each cell, a complex network of biochemical reactions facilitates cellular metabolism, which is crucial for many biological functions. Metabolic network models are powerful tools that allow the simulation of cellular metabolism and can thus provide fundamental mechanistic insights. For example, metabolic network models can be used to predict environment-dependent growth rates, various phenotypic states under different cultural conditions, and the flow of metabolites through the metabolic reaction network. Predicting more detailed quantities like optimal metabolite and enzyme concentrations or substrate-level regulatory mechanisms requires to incorporate information about enzyme kinetic parameters. Unfortunately, even for model organisms, experimentally measured kinetic parameters are not available for the vast majority of enzymatic reactions. Prediction methods for kinetic parameters could help to overcome this issue, but previously developed methods can either be only applied to a small subset of enzymatic reactions, lead to unrealistic values that are largely uncoupled from the true kinetic parameters, or they provide inaccurate predictions for enzymes that are not highly similar to proteins with measured kinetic parameters.
However, not only missing kinetic parameters but also missing functional information for enzyme-encoding genes lead to knowledge gaps in metabolic networks. Even in model organisms, large fractions of genes do not have high-quality functional annotations. One the one hand, this can lead to important reactions missing in metabolic networks, and on the other hand, many reactions cannot be associated with the catalyzing enzyme. In this thesis, I aim to overcome the issues of missing enzyme kinetic parameters and of not yet annotated enzymes with the use of machine and deep learning models. I developed the first general prediction model for the Michaelis-Menten constant KM. The resulting model is applicable to any natural enzyme-substrate pair and achieves a coefficient of determination R2=0.53 on a test set. Moreover, I developed a general model for predicting enzyme turnover numbers kcat for natural reactions of wild-type enzymes. The model outperforms previously developed prediction models and achieves a coefficient of determination R2=0.40 on a test set. To predict the function of not yet fully annotated enzymes, I developed a general model for predicting the substrate scopes of enzymes. The resulting model generalizes well even to enzymes that are not highly similar to enzymes in the training set, and it achieves an accuracy of over 91% on a test set. To develop these general prediction models, it was necessary to create maximally informative numerical representations of the proteins and the small molecules relevant for the downstream prediction tasks. This was achieved by using and modifying state-of-the-art deep learning methods for converting the linear protein amino acid sequences and the structures of the small molecules into numerical vectors. For all of the developed prediction models, we only used input information that is easily accessible, which makes the prediction models broadly applicable. | |||||||
| Lizenz: |  Dieses Werk ist lizenziert unter einer Creative Commons Namensnennung 4.0 International Lizenz | |||||||
| Fachbereich / Einrichtung: | Mathematisch- Naturwissenschaftliche Fakultät » WE Informatik | |||||||
| Dokument erstellt am: | 09.02.2024 | |||||||
| Dateien geändert am: | 09.02.2024 | |||||||
| Promotionsantrag am: | 21.12.2022 | |||||||
| Datum der Promotion: | 24.08.2023 |
