Parameter estimation in models of biological oscillators: an automated regularised estimation approach

Area of Science:

• Systems Biology
• Computational Biology
• Mathematical Biology

Background:

• Dynamic modeling is crucial for understanding complex biological systems.
• Parameter estimation in nonlinear differential equation models of biological oscillators presents significant challenges.
• Common pitfalls include large parameter spaces, local optima, overfitting, and parameter non-identifiability.

Purpose of the Study:

• To present a novel automated approach to address parameter estimation challenges in biological oscillator models.
• To improve the efficiency, accuracy, and generalizability of model calibration.
• To overcome limitations of standard estimation methods.

Main Methods:

• A novel automated workflow involving two sequential optimization steps.
• Utilizes sampling strategies to reduce parameter search space.
• Employs efficient global optimization to avoid local optima.
• Incorporates advanced regularization techniques to prevent overfitting.
• Includes tests for structural and practical parameter identifiability.

Main Results:

• Successfully evaluated the approach on four challenging biological oscillator models (Goodwin, FitzHugh-Nagumo, Repressilator, metabolic oscillator).
• Demonstrated the inability of local gradient-based methods, even with multi-start, to avoid common pitfalls.
• The novel approach yields more efficient estimations and avoids local optima.

Conclusions:

• The proposed method leads to more efficient parameter estimations through bounding strategies.
• Global optimization successfully prevents convergence to local optima.
• Regularization effectively combats overfitting, producing models with enhanced predictive power and generalizability.