The traditional approach to design space identification typically comprises an initial mechanistic modelling step, where either a new model is created for the system under investigation or an off-the-shelf existing model is reparametrized. Such an approach enables the application of model-based design space identification strategies that provide high accuracy [4, 5]. The main challenge of this approach is related to the mechanistic modelling stage, which might require extensive experts’ time and effort. In fact, it is not always the case that such experts are readily available in-house.
To address these issues, a valid alternative route is to employ a data-driven approach directly harnessing the available data. Naturally, this approach relies heavily on the quality and quantity of data itself. If the data is sufficiently good and enough, then the data-driven DSpI strategy enables a rapid and efficient way of applying the QbD framework.
In this work, we propose a data-driven DSpI methodology based on machine learning that utilizes artificial neural network classifiers to distinguish between feasibility and infeasibility regions. The methodology is systematically applied to different experimental scenarios with varying data quality and quantity to determine its effectiveness. Thus, the role of data quality and quantity is investigated to assess how robust and accurate the machine learning-based approach is in predicting the design space.
The experimental setup analysed is that of an industrial Protein A chromatography column for monoclonal antibodies separation using the MabSelect™ Sure™ resin. Additionally, transfer learning is implemented to support the data-driven approach in data-scarce scenarios to boost the reliability and accuracy of the design space predictions. Its application involves an initial training of the artificial neural network on the mechanistic model of Grom et al. [6], developed and validated for Protein A chromatography, and a final specialization on the target system under investigation. Transfer learning is able to extract intrinsic knowledge from the mechanistic model that has not been parametrized or adapted to the current system investigated and to enhance the predictive accuracy and reliability of the DSpI in data-scarce scenarios.
Acknowledgements
Funding from the UK Engineering & Physical Sciences Research Council (EPSRC) for the i-PREDICT:Integrated adaPtive pRocEss DesIgn and ConTrol (Grant reference: EP/W035006/1) is gratefully acknowledged.
References
[1] Nunavath, R. S., Singh, M. T., Jain, A., Chakma, M., Arivuselvam, R., & Azeeze, M. S. T. A. (2023). Quality by design in pharmaceuticals: A review of its impact on regulatory compliance and product quality. Drug Research, 74(1), 18–23. https://doi.org/10.1055/a-2185-4916
[2] Mehta, A. (2019). Downstream processing for biopharmaceuticals recovery. https://doi.org/10.1007/978-3-030-01881-8_6
[3] Simões, A., Veiga, F., & Vitorino, C. (2024). Question-based review for pharmaceutical development: An enhanced quality approach. European Journal of Pharmaceutics and Biopharmaceutics, 195. https://doi.org/10.1016/j.ejpb.2023.114174
[4] Sachio, S., Kontoravdi, C., & Papathanasiou, M. M. (2023). A model-based approach towards accelerated process development: A case study on chromatography. Chemical Engineering Research and Design, 197, 800–820. https://doi.org/10.1016/j.cherd.2023.08.016
[5] Geremia, M., Bezzo, F., & Ierapetritou, M. G. (2024). Design space determination of pharmaceutical processes: Effects of control strategies and uncertainty. European Journal of Pharmaceutics and Biopharmaceutics, 194, 159–169. https://doi.org/10.1016/j.ejpb.2023.12.008
[6] Grom, M., Kozorog, M., Caserman, S., Pohar, A., & Likozar, B. (2018). Protein A affinity chromatography of Chinese hamster ovary (CHO) cell culture broths containing biopharmaceutical monoclonal antibody (mAb): Experiments and mechanistic transport, binding and equilibrium modeling. Journal of Chromatography B, 1083, 44–56. https://doi.org/10.1016/j.jchromb.2018.02.032