Integrated / Uncertainty Quantification Method for Protein Crystallization Models.

The complexity of protein crystallization and modeling challenges process intensification and the wider adoption of crystallization in biomanufacturing. For computational models to support and replace extensive experiments, they must accurately reflect experiments. However, parameter estimation can be ineffective due to the highly nonlinear model structure and inaccurate online process analytical technology, which must be addressed. In this work, an experimentally validated and model-driven parametrization methodology is presented, developed for an antisolvent batch protein crystallization system with limited offline measurements. Global sensitivity analysis is performed to assess parameter identifiability during batch operations and inform optimal measurement points. Experiments at three different initial lysozyme concentrations ( = 15, 18, 19 mg/mL) are used for estimation. Parameter uncertainty distributions are recovered through an Approximate Bayesian Computation algorithm and propagated to model outputs through Monte Carlo simulations, avoiding linearization or unnecessary assumptions on the parametric and output uncertainty distributions. The methodology was successfully validated under two new experimental conditions. The shapes of the recovered parametric and output uncertainties highlight the need for parameter estimation methodologies specifically tailored to nonlinear models.