16 May 2024
08:00 Master's Defense Room 53 of IC2
Theme
Modeling Cellular Signaling Pathways via Universal Differential Equations and Joint Inference of First-Principles Model Parameters and Neural Network Weights
Student
Cristiano Gabriel de Souza Campos
Advisor / Teacher
Marcelo da Silva Reis
Brief summary
The orchestration of cellular processes occurs through sequences of chemical reactions known as cell signaling pathways. These pathways, fundamental for the regulation of cellular behavior, face the challenge of the "lack of isolation problem", which consists of the lack of communication between the reactions contained in the model and the reactions in the rest of the cell. This problem hampers the modeling of cell signaling pathways using ordinary differential equations (ODE)-based approaches, as it can lead to the loss of critical contextual information, thus hampering prediction accuracy. To address this problem, one possibility is to use a hybrid modeling, in which a model based on first principles and ODE is combined with a model based on neural networks and data. A mathematical framework that implements such a solution is the Universal Differential Equation (UDE). However, in real UDE-based cellular signaling pathway modeling settings, one must infer not only neural network weights but also unknown first-principles parameters (e.g., missing rate constants); To the best of our knowledge, no inference method is available in the literature. Therefore, here we propose an approach for modeling cellular signaling pathways that leverages the UDE framework and also jointly infers the missing parameters from the first-principles model and the neural network weights. To evaluate this approach, we performed computational experiments using four different cell signaling models and also the UDE implementation available in the SciML ecosystem. Our findings demonstrate marked improvements in both prediction accuracy and interpretability compared to the ODE-based approach, thus highlighting the effectiveness of UDE-based hybrid models for cell signaling pathway studies. In conclusion, this research has provided some promising tools for exploring the complex dynamics of biological systems.
Examination Board
Headlines:
| Marcelo da Silva Reis | IC / UNICAMP |
| Yayoi Natsume-Kitatani | NIBIOHN/Japan |
| João Meidanis | IC / UNICAMP |
Substitutes:
| Andre Santanche | IC / UNICAMP |
| André Fujita | IME / USP |