Cells are extremely complex, with numerous reaction pathways, intermediates, products, and regulatory networks. Precise models of cells are a necessity for optimizing performance and controlling bioprocesses. Current bioprocess phenomenological models, similar to reactor kinetic models, bury information regarding biomass heterogeneity and intercellular reactions within empirical parameters based on biological intuition.[3] On the other hand, purely machine learning (ML) models have shown to capture the nonlinear complexity of bioprocess but struggle with extrapolation and physical interpretability.[3] Experimental datasets are often sparse and noisy resulting in poor development and calibration of both these models. To leverage the physical constraints of phenomenological models, and the flexibility and practicality of ML models, hybrid models have been proposed.[3] In this work, an embedded hybrid modelling structure is developed, wherein parameters like growth and consumption rates are modelled using ML models. Our hypothesis is that hybrid models can capture the complex relationships between external metabolites and bacteria physiology.
A time variant parameter estimation (TV-PE) technique is employed to train the ML component of the hybrid model. A parameter estimation strategy is explored wherein the errors in the state space and derivative space are minimized to reinforce the learning of the underlying physics behavior. This method is compared to the traditional method of minimizing the error in the state space. Another variation in the hybridization structure is explored: (i) a sequential method where the TV-PE and ML model training is performed sequentially and (ii) an integrated method where the two steps occur simultaneously. A comparison of all methods is performed to determine which can best capture the physics of the bioprocess under interpolating and extrapolating scenarios. Results show that the hybrid model consistently outperforms purely phenomenological and black box models across various data availability and noise levels.
Finally, we do not stop at the development of a “first-generation” hybrid model. In fact, hybrid models lend a framework of interpretability by enabling sensitivity analyses on the ML components, providing qualitative insight for missing phenomena. We employ interpretable AI techniques to obtain insight from the ML components, which when combined with biologically-guided intuition, lead to improved hybrid models, with less dependence on black-box components. The sensitivity analysis gives the opportunity to present a train-build-improve model construction cycle. Accurate models will aid in process design and optimization and will allow for implementation of more sound and sophisticated control strategies. These are necessary steps for expediting the widespread scale-up and commercialization of bio-manufacturing processes.
References:
[1] J. Wesseler, et al., “Measuring the Bioeconomy: Economics and Policies,” Annu. Rev. Resour. Econ., 2017, pp. 275–298, Oct. 2017.
[2] N.-Z. Xie, et al., “Biotechnological production of muconic acid: current status and future prospects,” Biotechnol. Adv., vol. 32, no. 3, pp. 615–622, May 2014.
[3] A. Tsopanoglou, et al., “Moving towards an era of hybrid modelling: advantages and challenges of coupling mechanistic and data-driven models for upstream pharmaceutical bioprocesses,” Curr. Opin. Chem. Eng., vol. 32, p. 100691, Jun. 2021.