Generation of therapeutic mAbs in a bioreactor using cells is a complex biochemical process. Design elements and process variables, such as media and feed strategy, initial seeding, environmental conditions for pH, pO2, pCO2, among many other parameters, all collectively shape the cellular environment. Traditional mechanistic models often struggle to adequately capture the intricate interplay between these parameters, especially their effects on the multiple quality attributes. To address this challenge, we present GARNET (Generalized Architecture for Reaction NETwork), a hybrid modeling tool developed in-house at Merck. GARNET captures the complex reaction kinetics in bioreactors by combining neural ordinary differential equations (neural-ODEs) with biological mechanistic knowledge and machine learning. Existing experimental data serves as the training set for GARNET’s artificial neural networks, which learns the cellular kinetics, viz., growth rates, metabolic behaviors, etc. as a function of the local environmental conditions. The trained model is then capable of accurately predicting profiles for protein titer, cell counts, and metabolite levels, as well as critical quality attributes such as charge variants and glycan profiles. As the model is trained to learn cellular kinetics, which is scale independent, it is particularly useful for process development and scaling-up to commercial scale operations.
The advancement with GARNET is the capability to model industrially-relevant, complex bioreactor systems with remarkably good accuracy, hitherto unachieved in our opinion. We are able to incorporate as many process parameters and quality attributes of interest in the modeling framework (including scale effects), and combine with the available experimental data to build the trained model. GARNET can, in effect, be used as a "digital twin" of the bioreactor, enabling accurate virtual simulations of various bioreactor strategies to guide and/or replace traditional process development experiments. For example, the glucose concentration that maximizes cell viability at harvest can be virtually optimized using GARNET. Similarly, GARNET can be used for broad exploration of the process parameters, helping define ranges for the critical process parameters at the larger scales that yield acceptable protein quality attributes. Finally, this work sets the foundation for future use of GARNET as a true digital twin integrated real-time in a commercial manufacturing facility.