Dokument: Modeling Biological Systems - from Mechanistics to Machine Learning
| Titel: | Modeling Biological Systems - from Mechanistics to Machine Learning | |||||||
| URL für Lesezeichen: | https://docserv.uni-duesseldorf.de/servlets/DocumentServlet?id=51640 | |||||||
| URN (NBN): | urn:nbn:de:hbz:061-20191127-084821-8 | |||||||
| Kollektion: | Dissertationen | |||||||
| Sprache: | Englisch | |||||||
| Dokumententyp: | Wissenschaftliche Abschlussarbeiten » Dissertation | |||||||
| Medientyp: | Text | |||||||
| Autor: | Zhao, Linlin [Autor] | |||||||
| Dateien: |
| |||||||
| Beitragende: | Kollmann, Markus [Gutachter] Prof. Dr. Ebenhöh, Oliver [Gutachter] | |||||||
| Dewey Dezimal-Klassifikation: | 500 Naturwissenschaften und Mathematik » 570 Biowissenschaften; Biologie | |||||||
| Beschreibungen: | The dissertation is devoted to studying various biological systems through different modeling strategies, from mechanistic modeling to machine learning. The complex nature of biological systems imposes difficulties on both experimental data collection and
theoretical modeling. The main contribution is twofold. First, with emphasis on the system components and their interconnections, it was shown that mechanistic modeling led to significant understandings of investigated systems. Second, with emphasis on learning directly from data, machine learning was deployed to study biological problems from mRNA translation to automated annotations of high throughput biological data. The investigated systems include: • A network system of interconnected oscillators inspired by quorum sensing, a communication mechanism in bacterial culture. The network was mechanistically modeled and its synchronization conditions were analyzed. It was shown that synchronization could be achieved when the subsystems were stable oscillators and their interconnections influenced their input-output properties. • Flowering decision making in plants. Mechanistic modeling and machine learning were both employed to investigate how plants use temperatures and day length signals to make flowering decisions. Climate data from different regions were used to reconstruct the regulatory gene expression patterns by regarding plants as information processing units. It was shown that temperate plants can make use of cold winter memory and short term of day lengths to robustly determine flowering time. • The translational mechanism of mRNA sequences. Biological understandings of the mechanism were used to convert the sequences in the form of series of “AUCG” letters to numerical features. The learning results using gradient boosting trees suggested that the mRNA structures at the starting part of the sequences was the dominant feature. As machine learning models specially emphasize predictions, the trained model correctly predicted nine out of ten new sequences, which were verified by in-house experiments. • Computational identification of potential unconventional protein secretions (UPS) which do not require a guide signal at the starting of protein sequences (N-terminus) to enter the extracellular space. The method named OutCyte was built on inhouse experimental data. Based on a convolutional neural network model, OutCyte first only allows sequences without N-terminal signals to be fed into the cascaded XGBoost model for identifying UPS. OutCyte predicted 14 out of 18 experimentally verified UPS. • A bee behavior annotating system. It consisted of a high speed camera for monitoring tagged bees, a tracking system for the moving trajectories of bees, and an interactive machine learning framework for learning the general patterns of bee behaviors. The system correctly annotated 93% of the encounter behaviors in a group of bees.The dissertation is devoted to studying various biological systems through different modeling strategies, from mechanistic modeling to machine learning. The complex nature of biological systems imposes difficulties on both experimental data collection and theoretical modeling. The main contribution is twofold. First, with emphasis on the system components and their interconnections, it was shown that mechanistic modeling led to significant understandings of investigated systems. Second, with emphasis on learning directly from data, machine learning was deployed to study biological problems from mRNA translation to automated annotations of high throughput biological data. The investigated systems include: • A network system of interconnected oscillators inspired by quorum sensing, a communication mechanism in bacterial culture. The network was mechanistically modeled and its synchronization conditions were analyzed. It was shown that synchronization could be achieved when the subsystems were stable oscillators and their interconnections influenced their input-output properties. • Flowering decision making in plants. Mechanistic modeling and machine learning were both employed to investigate how plants use temperatures and day length signals to make flowering decisions. Climate data from different regions were used to reconstruct the regulatory gene expression patterns by regarding plants as information processing units. It was shown that temperate plants can make use of cold winter memory and short term of day lengths to robustly determine flowering time. • The translational mechanism of mRNA sequences. Biological understandings of the mechanism were used to convert the sequences in the form of series of “AUCG” letters to numerical features. The learning results using gradient boosting trees suggested that the mRNA structures at the starting part of the sequences was the dominant feature. As machine learning models specially emphasize predictions, the trained model correctly predicted nine out of ten new sequences, which were verified by in-house experiments. • Computational identification of potential unconventional protein secretions (UPS) which do not require a guide signal at the starting of protein sequences (N-terminus) to enter the extracellular space. The method named OutCyte was built on inhouse experimental data. Based on a convolutional neural network model, OutCyte first only allows sequences without N-terminal signals to be fed into the cascaded XGBoost model for identifying UPS. OutCyte predicted 14 out of 18 experimentally verified UPS. • A bee behavior annotating system. It consisted of a high speed camera for monitoring tagged bees, a tracking system for the moving trajectories of bees, and an interactive machine learning framework for learning the general patterns of bee behaviors. The system correctly annotated 93% of the encounter behaviors in a group of bees. | |||||||
| Lizenz: |  Urheberrechtsschutz | |||||||
| Fachbereich / Einrichtung: | Mathematisch- Naturwissenschaftliche Fakultät | |||||||
| Dokument erstellt am: | 27.11.2019 | |||||||
| Dateien geändert am: | 27.11.2019 | |||||||
| Promotionsantrag am: | 03.06.2019 | |||||||
| Datum der Promotion: | 11.10.2019 |
