Novel and flexible parameter estimation methods for data-consistent inversion in mechanistic modelling.

Area of Science:

• Computational modeling
• Physical sciences
• Biological sciences

Background:

• Mechanistic models (MMs) are crucial for analyzing ensembles in physical and biological systems.
• Current parameter estimation methods, like Bayesian inference, can introduce bias via uninformative priors.
• Population-based methods also face limitations in parameter estimation.

Purpose of the Study:

• To propose a novel parameter inference framework using stochastic inverse problems (SIPs) for mechanistic models.
• To address and mitigate bias introduced by uninformative priors in parameter estimation.
• To develop new computational methods for solving SIPs and overcoming their limitations.

Main Methods:

• Introduced stochastic inverse problems (SIPs), or data-consistent inversion, for parameter inference.
• Developed new SIP solution methods including rejection sampling, Markov chain Monte Carlo, and generative adversarial networks (GANs).
• Reformulated SIPs using constrained optimization and presented a novel GAN for this problem.

Main Results:

• The proposed SIP framework effectively infers parameters by targeting uncertainties from model non-invertibility.
• New computational methods, including GANs, provide efficient solutions for SIPs.
• Constrained optimization reformulation and its GAN solution overcome existing SIP limitations.

Conclusions:

• Stochastic inverse problems offer a robust framework for parameter inference in mechanistic models.
• The developed methods, particularly GAN-based approaches, enhance the accuracy and efficiency of parameter estimation.
• This work advances the analysis of complex systems by improving mechanistic model parameterization.