Data-gedreven modelidentificatie van dynamische biologische systemen

This project addresses the challenge of understanding biological dynamical systems through data-driven methods, with a focus on interpretable machine learning approaches. After framing the concept of "understanding" in terms of differential equations and tracing the evolution of dynamical systems theory, the work systematically reviews existing machine learning techniques for model discovery, identifying key limitations. A central emphasis is on symbolic approaches, particularly Sparse Identification of Nonlinear Dynamics (SINDy). Through a comprehensive evaluation, the thesis delineates obstacles to applying SINDy to biological data, such as measurement noise, data scarcity, nonlinearities, and multiple time scales. To mitigate these issues, practical guidelines are proposed for experimentalists. Furthermore, novel strategies are introduced to improve equation recovery, including the incorporation of system constraints (e.g., fixed points, nullclines). As part of the project, CLINE (Computational Learning and Identification of Nullclines) method, a neural network-based framework for model-free inference of nullcline structures, is introduced. Unlike SINDy, CLINE does not rely on predefined candidate libraries and thus avoids bias in model selection. Results demonstrate how CLINE complements symbolic approaches, providing robust support for model discovery in complex biological systems characterized by strong nonlinearities and scale separation.