(504c) Enhancing Resin Screening with Machine Learning in Protein a Chromatography

The concept of Quality by Design (QbD) is rapidly reshaping the workflow of how pharmaceutical processes are being developed [1, 2, 3, 4, 5]. The integration of quality restrictions in process design provides a more holistic approach that allows for a significant reduction in research and development time and costs compared to traditional Quality by Testing (QbT) [6, 1]. Quality by Digital Design (QbDD) further improves this by using computer-aided tools for design and optimization [7, 8], leveraging mathematical models for streamlined process development, and faster identification of design spaces. This innovative approach promises significant advancements in sustainable, efficient, and compliant biopharmaceutical manufacturing [9].

A key challenge in QbDD is its reliance on a high-fidelity mechanistic model, i.e. a digital twin [10]. Even though digital twins of the physical assets are of paramount importance, especially from an Industry 4.0 perspective [11], their development can be lengthy and complex. In the case of design variables such as resin type in separations, each alternative would possibly require a tailored digital twin. This can challenge and delay the early stages of process design where multiple variables must be screened and the plethora of candidate designs must be narrowed down. In this context, computer modeling tools can aid and accelerate process development, particularly in cases where the design is multifactorial.

Transferring existing knowledge, particularly in the form of rigorous kinetic modeling, from one system to a different one is a challenging task. Transferring modeling knowledge across systems could result in reducing wet-lab experimentation required and costs, and shaping an evidence-based process development strategy. Nevertheless, digital twins are naturally rather stiff in regards to transferability given that they are customized for a particular unit operation. However, the intrinsic physical knowledge contained in digital twins is invaluable for enabling transfer learning across multiple systems through means of machine learning.

The challenges mentioned above necessitate innovative approaches in process development. In response, this work demonstrates the effectiveness of a data-driven approach, coupled with transfer learning, for design space identification using experimental data. The case study analyzed in this work is a lab-scale Protein A chromatography setup validated on five different resins [12]. In particular, the data generated in the array of experiments has been used for the development of a machine learning algorithm comprising a binary classification neural network to directly estimate the feasible design space. To evaluate the performance of the data-driven model, an accurate digital twin of the unit operation is used and sampled for validation [12].

Transfer learning expands the capabilities of the above-mentioned approach by effectively reducing the need for extensive and costly experimental data while improving predictive accuracy. This is achieved by developing a machine learning model that inherently understands the system's physics through training on diverse rigorous models, capturing their intrinsic characteristics. In practice, the neural network is pre-trained on data generated by digital twins to understand underlying patterns for various resins. The final training is performed solely on experimental data, resulting in an algorithm tailored to the specific experimental campaign.

The presented study highlights the effectiveness of a purely data-driven approach for design space identification and emphasizes the role of transfer learning in improving model performance. Integrating these methods can significantly decrease the need for extensive experimental datasets, optimizing the balance between rigor and computational efficiency. This advancement can be beneficial for biopharmaceutical process development, providing more sustainable, flexible, and cost-effective manufacturing practices.

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] A. Simoes, F. Veiga, and C. Vitorino. “Question-Based Review for Pharmaceutical Development: An Enhanced Quality Approach”. In: European Journal of Pharmaceutics and Biopharmaceutics 195 (2024).
[2] T.T. Khuat, R. Bassett, E. Otte, A. Grevis-James, and B. Gabrys. “Applications of Machine Learning in Antibody Discovery, Process Development, Manufacturing and Formulation: Current Trends, Challenges, and Opportunities”. In: Computers and Chemical Engineering 182 (2024).
[3] R.V. Barenji, R.E. Hariry, D. Demirkol, and T.U. Daim. “Research Landscape Analysis for Quality in Pharma 4.0 Era”. In: Technology in Society 76 (2024).0160-791X.
[4] R.S. Nunavath, M.T. Singh, A. Jain, M. Chakma, R. Arivuselvam, and M.S.T.A. Azeeze. “Quality by Design in Pharmaceuticals: A Review of Its Impact on Regulatory Compliance and Product Quality”. In: Drug Research 74.1 (2023), pp. 18–23.
[5] Z. Kis, K. Tak, D. Ibrahim, S. Daniel, D. van de Berg, M.M. Papathanasiou, B. Chachuat, C. Kontoravdi, and N. Shah. “Quality by Design and Techno-Economic Modelling of RNA Vaccine Production for Pandemic-Response”. In: Computer Aided Chemical Engineering 49 (2022), pp. 2167–2172.50361-4.
[6] G. V. Reklaitis, J. Khinast, and F. Muzzio. “Pharmaceutical Engineering Science—New Approaches to Pharmaceutical Development and Manufacturing”. In: Chemical Engineering Science. Pharmaceutical Engineering Science, A Key for Tomorrow’s Drugs 65.21 (Nov. 2010), pp. iv–vii.
[7] H. Narayanan, F. Dingfelder, I. Condado Morales, B. Patel, K.E. Heding, J.R. Bjelke, T. Egebjerg, A. Buttè, M. Sokolov, N. Lorenzen, and P. Arosio. “Design of Biopharmaceutical Formulations Accelerated by Machine Learning”. In: Molecular Pharmaceutics 18.10 (2021), pp. 3843–3853.1c00469.
[8] Steven Sachio, Cleo Kontoravdi, and Maria M. Papathanasiou. “A Model-Based Approach towards Accelerated Process Development: A Case Study on Chromatography”. In: Chemical Engineering Research and Design 197 (Sept. 2023), pp. 800–820.
[9] H. Narayanan, M.F. Luna, M. von Stosch, M.N. Cruz Bournazou, G. Polotti, M. Morbidelli, A. Buttè, and M. Sokolov. “Bioprocessing in the Digital Age: The Role of Process Models”. In: Biotechnology Journal 15.1 (2020).
[10] Chaoying Ding and Marianthi Ierapetritou. “A Novel Framework of Surrogate-Based Feasibility Analysis for Establishing Design Space of Twin-Column Continuous Chromatography”. In: International Journal of Pharmaceutics 609 (Nov. 2021), p. 121161.
[11] Tsega Y. Melesse, Valentina Di Pasquale, and Stefano Riemma. “Digital Twin Models in Industrial Operations: A Systematic Literature Review”. In: Procedia Manufacturing. International Conference on Industry 4.0 and Smart Manufacturing (ISM 2019) 42 (Jan. 2020), pp. 267–272.084.
[12] Matic Grom, Mirijam Kozorog, Simon Caserman, Andrej Pohar, and Bla ˇz Likozar. “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”. In: Journal of Chromatography B 1083 (Apr. 2018), pp. 44–56.