Adaptive Nonlinear Model Predictive Control of Monoclonal Antibody Glycosylation in CHO Cell Culture

N-glycosylation is a critical quality attribute of monoclonal antibodies (mAbs), the dominant class of biopharmaceuticals. Controlling glycosylation remains difficult due to intrinsic pathway complexity, limited online measurements, and a lack of tailored control strategies. This work applies an adaptive nonlinear model predictive control (ANMPC) framework to a fed-batch mAb production process, using a multiscale model that links extracellular conditions to intracellular Golgi reactions to predict glycan profiles. Model parameters are updated online as new measurements arrive, after which a shrinking-horizon optimization computes the control inputs; only the first control move is implemented each cycle. Case studies show that, with a minimal day-1 galactose excitation, ANMPC mitigates model-plant mismatch and achieves up to 130% and 96% higher performance than open-loop optimization and state NMPC, respectively. Under more realistic conditions (partial measurement availability and longer preparation time), ANMPC maintains comparable performance, indicating robustness to practical limitations. Overall, the results demonstrate that ANMPC can actively shape glycan distributions in silico and offers a viable path toward closed-loop control of mAb glycosylation.

💡 Research Summary

Monoclonal antibodies (mAbs) are the most valuable class of biopharmaceuticals, and their therapeutic efficacy is strongly influenced by the N‑linked glycosylation pattern attached to the Fc region. Achieving a consistent glycan distribution in CHO‑based fed‑batch processes is notoriously difficult because the glycosylation pathway is a highly complex, non‑template enzymatic network, online glycan analytics are scarce, and existing control strategies focus on productivity rather than quality attributes. This paper addresses these challenges by developing an adaptive nonlinear model predictive control (ANMPC) scheme that directly manipulates the glycan profile in silico.

The authors employ a recently published multiscale mechanistic model that couples three hierarchical layers: (1) a reactor‑scale cell culture model describing viable cell density, extracellular metabolite concentrations, specific productivity, and mAb secretion; (2) an intracellular nucleotide‑sugar donor (NSD) synthesis model that converts extracellular nutrients into seven key NSDs (GDP‑Man, GDP‑Fuc, UDP‑Gal, UDP‑Glc, UDP‑GalNAc, UDP‑GlcNAc, CMP‑Neu5Ac) via Michaelis‑Menten kinetics; and (3) a Golgi‑level glycosylation model formulated as a plug‑flow reactor (PFR) with partial differential‑algebraic equations (PDAEs) that predicts the fractional distribution of glycoforms as the protein transits the Golgi apparatus. The overall system comprises 30 ordinary differential equations, 34 partial differential equations, and a large set of nonlinear algebraic relations, making it computationally demanding.

To render real‑time control feasible, the ANMPC framework integrates two key innovations. First, at each sampling instant (every few hours), all available measurements (cell density, extracellular metabolites, mAb concentration, and limited glycan fractions) are used to update the model parameters via a nonlinear least‑squares estimator. This online adaptation compensates for model‑plant mismatch caused by biological variability or parameter drift. Second, the control problem is solved using control‑vector parameterization (CVP) combined with a parallel quasi‑steady‑state (QSS) simulation technique previously shown to accelerate the solution of the underlying dynamic optimization. The resulting shrinking‑horizon optimization yields a sequence of control moves (e.g., feed rates of galactose, manganese, and other nutrients); only the first move is implemented before the next estimation‑optimization cycle.

Three case studies evaluate the approach. In the first scenario, the initial model parameters are deliberately perturbed to mimic severe uncertainty. Compared with a conventional state‑feedback NMPC that uses fixed parameters, ANMPC restores the desired glycan distribution and achieves a 96 % higher penalized merit. In the second scenario, a minimal day‑1 galactose spike (≈0.5 g L⁻¹) is used as the sole excitation. Despite this modest intervention, ANMPC outperforms an open‑loop optimal feed trajectory by 130 % in the same merit metric, demonstrating that active feedback can extract far more value from limited inputs. The third scenario introduces practical constraints: NSD concentrations are not measured online, and a realistic 4‑hour delay between sampling, analysis, and actuation is imposed. Even under these conditions, ANMPC maintains performance comparable to the ideal case, confirming robustness to measurement sparsity and latency.

The study concludes that adaptive NMPC, when coupled with a high‑fidelity multiscale model, can actively steer the glycosylation state space of mAbs in fed‑batch CHO cultures. The combination of online parameter adaptation and accelerated optimization mitigates the computational barrier that previously limited model‑based control of such complex systems. The authors suggest that future work should focus on integrating real‑time glycan analytics, scaling the approach to pilot‑scale bioreactors, and extending the methodology to other critical quality attributes such as fucosylation or sialylation. Overall, the paper provides a compelling proof‑of‑concept that closed‑loop control of mAb glycosylation is within reach, opening a new avenue for quality‑by‑design biomanufacturing.