Presentation Information

[C04-05]Data-driven Identification of Biological Systems Using
Multi-scale Analysis

*ISMAILA MUHAMMED1, DIMITRIS M MANIAS1, DIMITRIS A. GOUSSIS1, HARALAMPOS HATZIKIROU1 (1. KHALIFA UNIVERSITY (United Arab Emirates))

Keywords:

Multiscale,Dynamics,Quasi-Steady-State,System Identification,Neural Network.

Most biological systems evolve across multiple spatiotemporal scales, making it challenging to capture their dynamics within biologically relevant regimes. This complexity necessitates the integration of system identification and multiscale analysis
techniques. Traditional approaches, such as Computational Singular Perturbation (CSP), rely on explicit governing equations, which are often unavailable in biomedical
contexts where only observational data are provided. To address this limitation, we
propose a data-driven CSP framework that combines Sparse Identification of Nonlinear Dynamics (SINDy) and neural networks (NNs) to identify fast and slow dynamics from
data, enabling region-wise time-scale decomposition. The methodology is validated on the Michaelis-Menten enzyme kinetics model -a well-established multiscale system- by
identifying regions where valid quasi-steady-state approximations (sQSSA and rQSSA) can be constructed. In scenarios where SINDy fails due to noise, data sparsity, or model complexity, neural networks are employed to estimate the Jacobian matrix, enabling CSP to reveal dominant dynamical structures and detect regions consistent with reduced dynamics. We further examine a case where the system spans multiple
dynamical regimes, necessitating dataset partitioning. Within each partition, SINDy is used to recover region-specific reduced models. The results demonstrate that while SINDy alone is sensitive to noise and often fails in globally multiscale settings, its performance significantly improves when guided by CSP-based time scale decomposition and region identification. This work introduces a robust, equation-free framework for
analyzing biological systems, extending the applicability of CSP to data-driven settings where explicit governing equations are not available.