Robust parameter estimation for dynamical systems from outlier-corrupted data.

Area of Science:

• Systems Biology
• Computational Biology
• Mathematical Modeling

Background:

• Mechanistic mathematical models are crucial for understanding cellular processes.
• Parameter estimation from experimental data often assumes normally distributed noise.
• Outliers in datasets can significantly distort parameter estimates and lead to inaccurate model predictions.

Purpose of the Study:

• To develop and evaluate robust parameter estimation methods for ordinary differential equation models using outlier-corrupted data.
• To assess the impact of alternative noise distribution assumptions on parameter estimation accuracy and robustness.

Main Methods:

• Proposed and evaluated parameter estimation methods for ordinary differential equation models.
• Considered Laplace, Huber, Cauchy, and Student's t distributions as alternatives to the normal distribution for noise modeling.
• Assessed accuracy, robustness, and computational efficiency using artificial and experimental data (Epo-induced JAK/STAT signaling).

Main Results:

• Alternative noise distribution assumptions significantly improve the robustness of parameter estimates in the presence of outliers.
• Methods demonstrated the ability to compensate for and identify artificially introduced outliers.
• Evaluation included comparisons of accuracy, robustness, and computational efficiency across different distribution assumptions.

Conclusions:

• Employing alternative noise distributions enhances the reliability of parameter estimation for mechanistic models with corrupted data.
• The proposed methods offer a more robust approach to analyzing biological system dynamics from experimental measurements.
• MATLAB implementation of likelihood functions is available for broader application.