Data-driven or hybrid models can potentially be used to predict the behaviour of complex bioprocesses. In this study, we consider the whole cell biotransformation of oripavine to nororipavine, an N-demethylation step in the production of buprenorphine, where cell growth occurs simultaneously with enzymatic transformation and mechanistic understanding is not yet complete. Data-driven machine learning approaches, including artificial neural networks and recurrent neural networks, were assessed for their ability to predict cell growth, as well as reaction progress from simple online process measurements. Hybrid models were also developed and compared, together with simpler multivariable linear regression. A mass balance was used to pre-process training data for the hybrid model, unknown parameters were then predicted and used to calculate nororipavine concentration. A second rates-based hybrid model predicted the unknown rates of production and consumption of nororipavine, which were then used in a mass balance to predict nororipavine concentration. The hybrid model incorporating a rates-based approach was able to predict process metrics, including the transient peak in nororipavine concentration. The simpler hybrid model, or the data-driven approaches with no mechanistic detail, could not accurately predict nororipavine concentration, only oripavine concentration and cell growth, with the simpler multivariable linear regression only able to predict cell growth. The successful rates-based hybrid model demonstrates a proof-of-concept that could potentially be used for process prediction and monitoring, to supplement offline analytical methods, to provide new data on by-product formation or to simulate process data and improve process control. The study demonstrates the utility of hybrid models, where data-driven components describe complex process behaviour, while mechanistic components with known process knowledge increase model accuracy.
Autologous Chimeric Antigen Receptor (CAR) T cell manufacturing remains costly and labour-intensive, with workflows reliant on manual interventions despite advances in automation. Here, we present a compact microfluidic cell culture device (µCCD) with a simple, modular architecture for intensified CAR T cell expansion. Over four days, the µCCD yielded 10.5 × 10⁶ viable primary T cells compared with 49.8 × 10⁶ viable cells in the G-Rex®, while achieving a 6.7-fold higher viable cell concentration. For CAR T cells, the µCCD achieved a 14-fold expansion, yielding 14.1 × 10⁶ viable cells, compared with 25.1 × 10⁶ viable cells in the G-Rex®, and reaching 56.5 × 106 viable cells·mL−1 – a concentration seldom achieved in conventional culture systems and 18-fold higher than a G-Rex® well-based static culture control. Continuous perfusion maintained glucose availability, while integrated oxygen sensors enabled real-time monitoring. Expanded CAR T cells retained cytotoxic activity comparable to G-Rex®-expanded cells under the tested conditions, and microfluidic culture yielded a higher proportion of central memory CAR T cells, which are associated with long-term therapeutic persistence. The platform's modular architecture and low seeding requirement provide a foundation for future clinical-dose manufacturing and rapid assessment of patient-to-patient variability. Together, these results demonstrate the feasibility of the µCCD as a compact and adaptable platform for the high-density expansion of functionally robust CAR T cells. By demonstrating that a modular microfluidic architecture can support the expansion of non-adherent therapeutic cell cultures, this work represents an important step towards more standardised microfluidic platforms for next-generation cell therapy manufacturing.
