Predictive power of non-identifiable models.

Area of Science:

• Computational modeling
• Bayesian inference
• Systems biology

Background:

• Computational models often face non-identifiability issues, requiring complex solutions like data augmentation or model reduction.
• Model reduction can lead to parameters lacking direct interpretability, hindering practical application.
• Alternative methods are needed to leverage the predictive capabilities of non-identifiable models.

Purpose of the Study:

• To develop and evaluate a Bayesian approach for quantifying the predictive power of non-identifiable computational models.
• To demonstrate that specific measurements can reduce parameter space dimensionality, enabling predictions.
• To assess the utility of iterative measurements in enhancing model predictive capabilities.

Main Methods:

• Exploration of a Bayesian framework to assess model predictive power.
• Application to a biochemical signaling cascade model and its mechanical analog.
• Utilizing targeted measurements and stimulation protocols to reduce parameter space dimensionality.

Main Results:

• Demonstrated that measuring a single variable under specific stimulation reduces parameter space dimensionality.
• Enabled prediction of variable trajectories and their transformation under parameter changes.
• Showcased how successive measurements further reduce dimensionality and enable new predictions.

Conclusions:

• The proposed Bayesian approach effectively quantifies and utilizes the predictive power of non-identifiable models.
• Iterative measurements enhance model predictability and allow for assessment at each step.
• The method offers a practical alternative to model reduction for handling non-identifiable computational models.