Exploring Hybrid Modeling of Cell Signaling Pathways: Combining Neural Networks and ODEs to Tackle the Lack of Isolation Problem

• Ano de defesa: 2025
• Autor(a) principal: Sousa, Ronaldo Nogueira de
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Biologia de sistemas; Cell signaling pathways; Equações diferenciais universais; Falta de isolamento; Hybrid modeling; Modelagem híbrida; Neural networks; Redes neurais; Systems biology; Universal Differential Equations. Lack of isolation; Vias de sinalização celular
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/95/95131/tde-10022026-122908/

Cell signaling pathways regulate key biological processes through complex biochemical interactions. Traditional mechanistic models based on ordinary differential equations (ODEs) attempt to capture these dynamics using established biochemical knowledge. However, these models often rely on the assumption that the modeled subsystem behaves in isolation. In practice, this assumption rarely holds: signaling components frequently interact with external species and parallel pathways, giving rise to incomplete models that fail to reproduce experimental observations, a challenge referred to as the lack of isolation problem. This thesis explores hybrid modeling as a strategy to address this limitation. Specifically, we combine mechanistic ODE-based models with data-driven neural networks to approximate latent or unmodeled contributions. The hybrid formulation leverages Universal Differential Equations (UDEs), enabling the incorporation of prior mechanistic knowledge while providing flexibility to infer unknown interactions directly from data. The methodology was evaluated in three phases: (i) controlled toy models, (ii) a realistic cell-cycle model with synthetic data, and (iii) real-world proteomic time-series. Results show that hybrid models consistently outperform purely mechanistic ODE models, particularly when perturbations or partial observability are present. Across all experimental scenarios, hybrid models exhibited improved predictive accuracy, achieving significantly lower MAE and SMAPE values, and demonstrated strong generalization to unseen initial conditions. Importantly, the neural component successfully captured missing signaling contributions, improving trajectory reconstruction without compromising biological interpretability. These findings highlight the potential of hybrid modeling to overcome the lack of isolation inherent in cell signaling models. By integrating mechanistic and learning-based components, this work provides a computational framework capable of better reflecting biological complexity, extending the applicability of dynamic modeling in systems biology.