Publicação
Network inference for logic-based ordinary differential equations
| Resumo: | Signaling is a highly dynamic and context specific process. When cells fail to interpret external stimuli from the environment or emitted by other cells the consequences can be disastrous. Mechanistic signaling models with predictive value have the potential to help developing new therapheutical strategies targeting molecules involved in signal transduction. However, the complexity of signaling networks, the nonlinear nature of these systems and several technological limitations regarding the ability to manipulate cells in vitro and measure post translational modifications experimentally, make the task of building quantitative models for signaling very difficult. Many interactions in signaling pathways are known but, because they are not well characterized from the biochemical point of view, it is not straightforward to turn this information into a model. In this thesis, we present methods for reverse engineering mechanistic models combining data from cell-line perturbation experiments. Here, the model dynamics is described by means of logic-based ordinary differential equations, a recent formalism that through a set of reasonable assumptions describes regulatory mechanisms in a relatively simple, yet, dynamic and continuous manner. We formulate model selection and network inference as dynamic optimization problems, which are nonlinear non-convex and, thus very hard to solve. Here, we formulate model selection as a mixed-integer dynamic optimization problem and solve it recurring to state of the art meta-heuristics for optimization and numerical methods for simulation. We apply the methods to several signaling case-studies and concluded the method scales up well. In addition, we develop a relaxation tailored for this problem that improves convergence in large problems. The network inference problem is tackled with the help of mutual information and an ensemble approach. To compensate for the lack of prior knowledge, we build data-driven networks based in mutual information. With the ensemble approach, we explore the landscape of possible models, providing more reliable predictions for trajectories and network inference. The method was applied to several in silico and experimental case studies including data from the HPN-DREAM Breast Cancer Network Inference challenge. We were able to generate predictions that were in some cases significantly better than those provided by the best performers. To facilitate the implementation and redistribution of dynamic optimization problems in systems biology, such as those described above, we also develop a C library. This library is open-source and platform independent. The implementation and some applications of the library are discussed. Building dynamic models of signaling with predictive power is possible despite of a number of well known pitfalls and limitations. The heavy computational cost of simulating ordinary differential equations models can be palliated by combining state of the art numerical methods with meta-heuristics and the power of cluster computing. |
|---|---|
| Autores principais: | Henriques, David Saque |
| Assunto: | Engenharia e Tecnologia::Engenharia Eletrotécnica, Eletrónica e Informática |
| Ano: | 2016 |
| País: | Portugal |
| Tipo de documento: | tese de doutoramento |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade do Minho |
| Idioma: | inglês |
| Origem: | RepositóriUM - Universidade do Minho |
| Resumo: | Signaling is a highly dynamic and context specific process. When cells fail to interpret external stimuli from the environment or emitted by other cells the consequences can be disastrous. Mechanistic signaling models with predictive value have the potential to help developing new therapheutical strategies targeting molecules involved in signal transduction. However, the complexity of signaling networks, the nonlinear nature of these systems and several technological limitations regarding the ability to manipulate cells in vitro and measure post translational modifications experimentally, make the task of building quantitative models for signaling very difficult. Many interactions in signaling pathways are known but, because they are not well characterized from the biochemical point of view, it is not straightforward to turn this information into a model. In this thesis, we present methods for reverse engineering mechanistic models combining data from cell-line perturbation experiments. Here, the model dynamics is described by means of logic-based ordinary differential equations, a recent formalism that through a set of reasonable assumptions describes regulatory mechanisms in a relatively simple, yet, dynamic and continuous manner. We formulate model selection and network inference as dynamic optimization problems, which are nonlinear non-convex and, thus very hard to solve. Here, we formulate model selection as a mixed-integer dynamic optimization problem and solve it recurring to state of the art meta-heuristics for optimization and numerical methods for simulation. We apply the methods to several signaling case-studies and concluded the method scales up well. In addition, we develop a relaxation tailored for this problem that improves convergence in large problems. The network inference problem is tackled with the help of mutual information and an ensemble approach. To compensate for the lack of prior knowledge, we build data-driven networks based in mutual information. With the ensemble approach, we explore the landscape of possible models, providing more reliable predictions for trajectories and network inference. The method was applied to several in silico and experimental case studies including data from the HPN-DREAM Breast Cancer Network Inference challenge. We were able to generate predictions that were in some cases significantly better than those provided by the best performers. To facilitate the implementation and redistribution of dynamic optimization problems in systems biology, such as those described above, we also develop a C library. This library is open-source and platform independent. The implementation and some applications of the library are discussed. Building dynamic models of signaling with predictive power is possible despite of a number of well known pitfalls and limitations. The heavy computational cost of simulating ordinary differential equations models can be palliated by combining state of the art numerical methods with meta-heuristics and the power of cluster computing. |
|---|