Digital Twins And Hybrid Models In Biologics Production: A Model Comparison

By Thomas de Marchin

Predicting how cells behave in a bioreactor is one of the more demanding challenges in biologics manufacturing. Classical data-driven models offer a practical starting point, but they rely on fitting mathematical functions to experimental data without accounting for the biological mechanisms behind what's actually happening. This creates real limitations when conditions shift or data is scarce.

Hybrid models take a different approach by combining mechanistic knowledge — nutrient consumption rates, metabolic pathways, physical laws — with machine learning components that handle the complex or poorly understood dynamics. In a head-to-head comparison using a perfusion bioreactor case study, a hybrid model consistently outperformed a classical model on prediction accuracy, both within and beyond the conditions it was trained on. The difference was particularly pronounced when extrapolating to new scenarios, which is where real-world manufacturing decisions are made.

For organizations working with limited experimental data, strategies like Intensified Design of Experiments and federated pre-training offer promising paths to making hybrid modeling more accessible. If process optimization in biologics manufacturing is on your agenda, these approaches are worth understanding in depth.