Speaker
Description
Mechanistic models based on ordinary differential equations typically involve many parameters that must be estimated from limited experimental data. This process is hindered by multimodal objective landscapes, expensive model simulations, sparse and noisy measurements, plus structural and/or practical identifiability issues, all of which make calibration computationally demanding.
Here, we propose a strategy to reduce the computational burden of model calibration by focusing only on the most relevant parameters. Instead of estimating all parameters simultaneously, we define a reduced calibration problem in which non-influential (non-identifiable) parameters are excluded prior to optimization. In contrast to traditional sensitivity-based approaches relying on local analyses, our method uses global sampling and a decision tree–based strategy to identify and discard parameters that cannot be reliably inferred from the available data, thereby simplifying the estimation problem while preserving model fidelity.
We evaluate the methodology on a collection of representative ODE-based systems biology models. The results show that the proposed approach achieves statistical fits comparable to full-scale calibration while substantially reducing computation time. Overall, the framework provides a practical and promising route to accelerate parameter estimation in complex mechanistic models.
