(374b) Hybrid Modeling of CHO Cell Culture Using Physics Informed Neural Networks
AIChE Annual Meeting
2023
2023 AIChE Annual Meeting
Computing and Systems Technology Division
Data-driven and hybrid modeling for decision making II
Tuesday, November 7, 2023 - 12:55pm to 1:20pm
The development of accurate bioreactor models is hindered by two primary factors: the complexity of the underlying mechanism, and the limited availability of data. From a first principle modeling (FPM) perspective, bioreactors involve multiscale biochemical and physical processes that encompass growth, metabolism, and transport phenomenon, which can be difficult or even infeasible to capture. As a standalone FPM is also susceptible to process perturbation and characteristics that are specific to the cell line or equipment. From a data-driven modeling (DDM) perspective, due to the limited number of runs, the data quantity is usually insufficient to construct a high-precision model and the resulting model often generalizes poorly. Hybrid modeling is a promising approach to address the challenge of system complexity and data scarcity, which marry the strengths of both components 5.
Physics informed neural networks (PINNs) is an emerging hybrid modeling technique that combines the adaptability of deep learning with the rigor of mechanistic laws 6. Compared to traditional hybrid modeling techniques where DDMs and FPMs are arranged in a structural manner, PINN completely integrates the FPM component into a neural network by enforcing the mechanistic laws (usually in the form of ordinary or partial differential equations) during the model training process. This ensures model predictions fall within realistic bounds. The PINN hybrid modeling framework has shown itself to be useful in describing a growing number of diverse and complex systems while remaining computationally scalable 5.
This study discusses using PINNs for the macroscopic modeling of CHO cell bioreactor. Our strategy uses PINN to combine mechanistic cell culture model (such as cell growth and metabolite consumption) with process data (such as pH and dissolved oxygen). The proposed modeling approach will be illustrated using real-world data from Pfizerâs bioreactors across multiple scales. Predictive performance of the proposed PINN model is compared with pure FPM and DDM, and improved modeling accuracy and extrapolation performance is shown.
References:
(1) Dumont, J.; Euwart, D.; Mei, B.; Estes, S.; Kshirsagar, R. Human Cell Lines for Biopharmaceutical Manufacturing: History, Status, and Future Perspectives. Critical Reviews in Biotechnology 2016, 36 (6), 1110â1122. https://doi.org/10.3109/07388551.2015.1084266.
(2) Kontoravdi, C.; Pistikopoulos, E. N.; Mantalaris, A. Systematic Development of Predictive Mathematical Models for Animal Cell Cultures. Computers & Chemical Engineering 2010, 34 (8), 1192â1198. https://doi.org/10.1016/j.compchemeng.2010.03.012.
(3) Karra, S.; Sager, B.; Karim, M. N. Multi-Scale Modeling of Heterogeneities in Mammalian Cell Culture Processes. Ind. Eng. Chem. Res. 2010, 49 (17), 7990â8006. https://doi.org/10.1021/ie100125a.
(4) Hong, M. S.; Braatz, R. D. Mechanistic Modeling and Parameter-Adaptive Nonlinear Model Predictive Control of a Microbioreactor. Computers & Chemical Engineering 2021, 147, 107255. https://doi.org/10.1016/j.compchemeng.2021.107255.
(5) Bradley, W.; Kim, J.; Kilwein, Z.; Blakely, L.; Eydenberg, M.; Jalvin, J.; Laird, C.; Boukouvala, F. Perspectives on the Integration between First-Principles and Data-Driven Modeling. Computers & Chemical Engineering 2022, 166, 107898. https://doi.org/10.1016/j.compchemeng.2022.107898.
(6) Raissi, M.; Perdikaris, P.; Karniadakis, G. E. Physics-Informed Neural Networks: A Deep Learning Framework for Solving Forward and Inverse Problems Involving Nonlinear Partial Differential Equations. Journal of Computational Physics 2019, 378, 686â707. https://doi.org/10.1016/j.jcp.2018.10.045.