Alex Vrbancic
Thermo Fisher
14:15
Abstract
Upstream process development remains resource-intensive, particularly when scaling from high-throughput systems to pilot and manufacturing scales. Limited large-scale data often constrains purely data-driven approaches, while mechanistic models alone may lack flexibility across varying operating conditions.
We evaluated a hybrid modeling framework that integrates mechanistic understanding with machine learning using a 28-run multi-scale dataset spanning 15mL to 500L bioreactors. The objective was to assess cross-scale predictive capability and reduce experimental burden during process optimization.
Hybrid models were trained to predict key process variables and final titer. Robust large-scale predictions were achieved using predominantly small-scale training data, enabling a >30% reduction in calibration experiments. Compared to standard multivariate linear regression, the hybrid approach maintained predictive performance across scales.
The trained hybrid model was subsequently used to explore alternative feeding strategies in silico, allowing rapid evaluation of scenarios that would otherwise require multiple additional wet-lab studies. Simulation-guided optimization identified a front-loaded feeding strategy predicted to improve titer. This recommendation was experimentally verified in a 5 L glass bioreactor system with glucose feedback control, where the front-loaded strategy resulted in improved titer while maintaining stable growth.
These findings demonstrate that hybrid in silico modeling can support data-lean, scale-relevant process optimization and reduce experimental effort without compromising predictive reliability.
