Mechanistic chromatography models digitalize biopharmaceutical process development and offer numerous advantages from early stage process development to root-cause investigation during manufacturing. Digital twins based on mechanistic models enable an in-depth understanding of even very complex separation problems and turn data into even deeper process knowledge.
The initial hurdle to take: model calibration
The industrial adoption of mechanistic modeling is however slowed down by cumbersome model calibration approaches. Multicomponent feed stocks lead to multidimensional parameter estimation problems with many unknown protein properties. Standard model calibration techniques may result in unreasonable correlations and unphysical parameter estimates. The demand for a more straightforward and robust model calibration strategy, ensuring adequate model certainty in the fastest time possible is therefore high. But how does such a strategy look like?
Straightforward model calibration for mechanistic chromatography models
In their study on the “Straightforward method for calibration of mechanistic cation exchange chromatography models for industrial applications”, Boehringer Ingelheim (BI) addressed this question. David Saleh and co-workers implemented a robust workflow to calibrate a mechanistic model for a pH-dependent cation-exchange (CEX) operation to polish a monoclonal antibody (mAb).
Model calibration by Yamamoto and inverse peak fitting
Besides low effort experiments for system and column characterization, BI performed linear gradient elutions (LGE) at low loading density as well as one LGE and one step elution at a process relevant loading density. These experiments were sufficient to determine the charge and equilibrium parameters using the Yamamoto approach. This technique is conveniently available in the Peak Finder of GoSilico’s simulation software ChromX. The remaining parameters (binding kinetic and ligand shielding) where determined by inverse peak fitting, again using ChromX. To include a pH-dependency in the calibration space some of the LGEs were performed at pH values varying from the standard pH. Again, BI used the Yamamoto method to derive the pH dependency of the charge and equilibrium parameters.
Enriched feedstock experiments to increase parameter sensitivity
As the feedstock contained a very low concentration of HMW (<0.4%) all LGEs at low load density were again performed with an aggregate enriched loading material to increase the parameter sensitivity for this species. This resulted in a total of 20 experiments with 18 of these experiments being very low effort LGEs at low loading density (1 g/l, 0.05 g/L) with low material consumption.
Peak shapes above 3000 mAU were insignificant in this study
As a short cut, peak shapes above the 3000 mAU measurement limit were neglected during inverse parameter estimation. BI considered the model’s predictive power for yield, elution volume and purity as the key requirement for performing in silico process characterization while the individual peak shapes are less important for the project at hand. The goal could be achieved with a simple SMA adsorption model that is able to accurately describe the frustum of the peaks. A high-fidelity model for the full peak shape might have required a more complex adsorption model that would have complicated the parameter estimation.
Model uncertainty and model validation
All obtained parameters were in physically reasonable ranges and consistent with values found in literature. Moreover, a thorough model validation including 12 validation experiments was performed, covering LGEs and step elutions at all pH values under low, medium and high load conditions. The validation runs met the demanded accuracy for future application of the model during in silico process characterization studies.
The straightforward approach by BI combined well-established theories like the Yamamoto method with inverse peak fitting. The study allowed to systematically reduce the number of unknown model parameters from 32 to 7. While the experimental effort was decreased compared to previous mechanistic modeling projects, the risk of overfitting was mitigated und model uncertainty was reduced. The model was further also successfully validated beyond the operating ranges of the final unit operation, enabling its application to late‐stage downstream process development. In the publication of BI, a systematic model calibration strategy is provided, ensuring adequate model certainty in the fastest time and with the least experimental effort possible.
Outlook and conclusions
GoSilico’s Tobias Hahn summarizes: “What I like about this publication is that it dares to strike a balance between speed and accuracy. The chosen procedure ensures the physical correctness of the main model parameters by using Yamamoto’s method. On the other hand, the model and the calibration effort are deliberately reduced by modeling only cut off peaks with SMA. There is a risk in this decision, but it is remedied by careful validation of the model. I am convinced that the procedure can be used for many other mAb processes. Many thanks to the colleagues from BI for taking the time to make their methodology accessible to other modelers.”
The full paper is available in Biotechnology Progress.