(261b) Hybrid Modeling Employing Neural Differential Algebraic System of Equations (DAEs) for Enhanced Monoclonal Antibody Production in Fed-Batch Reactors

Monoclonal antibodies (mAb) are an essential asset for treating a wide range of diseases including certain types of cancers, autoimmune disorders, and infectious diseases, with an expected market of US$ 300 billion in 2025 [1] and to surpass US$ 700 billion by 2033 [2]. The manufacture of mAbs is mainly performed using Chinese hamster ovary (CHO) cell lines with fed-batch processes, corresponding to around 70% of all mAb production [3]. This process has complex dynamics associated with cell growth, metabolism, and production of mAbs. Nutrients (e.g., glucose, amino acids) are fed regularly in boluses, and undesired byproducts such as ammonia and lactate are also produced, which inhibit cell growth and decrease mAb productivity. Such complex dynamics make it challenging to model this process and require a differential algebraic system of equations (DAE) with nontrivial solution techniques. While modeling of CHO-based mAb production from a mechanistic perspective is present in the literature [4], [5], [6], there is a critical need to embed machine learning and model predictions beyond the capabilities of purely mechanistic approaches.

To achieve this goal, we developed a hybrid model that embeds both mechanistic model equations, which incorporates domain knowledge, with neural network components trained based on experimental data. We aim to generate a model that is both interpretable and able to capture the complexity of mAb production processes. Inspired by recent hybrid modeling approaches such as neural ordinary differential equations (ODEs) and universal differential equations (UDEs) [7], [8] we generated a novel system of so-called neural DAEs [9]. This hybrid formulation is built leveraging the Pyomo.DAE framework [10], which allows for the algebraic modeling and optimization of DAE systems. The training and simulation of the CHO-based mAb production hybrid model is performed using state-of-the art nonlinear programming algorithms, such as Ipopt [11] and dynamic optimization techniques, namely the simultaneous approach [12]. The Pyomo [13] framework can accurately model a non-smooth process that is interrupted by feeding and sampling events based on experimental data available. We train the neural DAE system using data from experiments at small and large scales and evaluate the hybrid model performance across scales.

The hybrid model offers enhanced prediction capabilities while still maintaining domain knowledge. This approach can mitigate extrapolation issues often observed in purely data-driven models and reduce data needed to train the embedded neural networks. In this presentation, we highlight the potential of hybrid modeling enabled by nonlinear programming as a new venue for enhancing biopharmaceutical manufacturing.