(198u) A Data-Driven Approach to Harmonising Experimental Design and Mechanistic Modelling: Application to mRNA Production | AIChE

(198u) A Data-Driven Approach to Harmonising Experimental Design and Mechanistic Modelling: Application to mRNA Production


Kalospyrou, I. - Presenter, University of Sheffield
Milton, R., University of Sheffield
Kis, Z., Imperial College London
Brown, S. F., University of Sheffield
In the initial experimental stages of process development, the determination of ideal ranges for process inputs and conditions to achieve optimisation of production is necessary. Experimentation is typically followed by the development of mechanistic models to predict the outputs of processes. The optimisation of these models is done on the basis of the experimental outputs derived first.

The implementation of suitable methods to determine optimal input and condition ranges as well as to facilitate the optimisation of mechanistic models using experimental data is critical but more often than not, challenging. Especially in cases when the experimental data obtained is limited or the process examined involves highly complex and not fully understood physicochemical mechanisms and interactions.

In this study, a novel data-driven method for the simultaneous optimisation of mechanistic models and experimental practice is developed, through the consecutive application of Latin Hypercube sampling (LHS), Gaussian Process Regression (GPR) and Global Sensitivity Analysis (GSA) on a series of data. The specific application where this method is examined is the batch in vitro transcription (IVT) bioreactor where DNA is transcribed to mRNA in mRNA vaccine production lines. The optimisation was conducted on the basis of mRNA yield and it was achieved through the separate investigation of the mechanistic model and experimental data obtained, followed by the development of a surrogate model of error between the mechanistic model and experimentally derived mRNA yields. The bioreactor inputs whose effects were examined were the nucleoside triphosphate (NTP), magnesium acetate (MgOAc), T7-RNA polymerase, template DNA, and spermidine concentrations. Incubation time was also accounted for.

The application of this method to the mRNA production example showed the statistical significance of different inputs and conditions on mRNA production as well as their contribution to the variation between experimental data and mechanistic model predictions. In this manner, it facilitated the determination of the optimal input ranges for mRNA yield maximisation. Moreover, this method pointed out chemical reaction instabilities unaccounted for previously during experimental practice. Finally, it clearly pinpointed technical issues related to the implementation of the mechanistic model as well as weaknesses of said model in describing the biochemical process.