Bioprocessing Blog

The complexity of biopharmaceutical manufacturing demands precision, efficiency, and a deep understanding of process dynamics. Over the years, Design of Experiments (DoE) models have served as a critical tool in the process development toolbox, providing a systematic approach to exploring the effects of process variables on outcomes.

However, as the calls for increased speed and efficiency grow louder, the limitations of DoE have become more evident, adding pressure on development organizations. To meet this new challenge, new analytical technologies and methods are emerging.

Hybrid models represent the next frontier in process modeling, offering enhanced predictive capabilities, reduced experimental burdens, and deeper process insight by combining the structured knowledge of mechanistic models with the adaptability of data-driven techniques. They are proving to be superior to traditional DoE models and are transforming biopharmaceutical process development.

Hybrid models combine the strengths of mechanistic and data-driven approaches. By incorporating first principles into data-driven frameworks, they deliver superior predictive capabilities, efficiency, and process insights.

Key Features of Hybrid Models

1. Mechanistic Foundations: Hybrid models use well-established scientific knowledge (e.g., reaction kinetics, mass balances) to describe known physical and biochemical phenomena. This ensures that predictions are grounded in robust, universal principles.
2. Data-Driven Flexibility: Machine learning and statistical techniques complement the mechanistic framework by capturing complex, nonlinear, or poorly understood relationships. These data-driven components fill gaps where mechanistic knowledge is incomplete.
3. Dynamic and Adaptive Modeling: Unlike static DoE models, hybrid models can adapt dynamically as new data becomes available, continuously refining predictions and improving accuracy over time.

While statistical DoE models have been a valuable tool for process development, hybrid modeling goes beyond what can be acheived with DoE models offering several advantages:

Statistical DoE relies on carefully designed experiments to generate robust models that answer the process development question at hand. As the number of process variables increases, the number of experiments required grows exponentially due to factorial combinations. This “curse of dimensionality” becomes particularly problematic in bioprocess development, where experiments are resource-intensive, time-consuming, and costly.

For example, optimizing a fed-batch cell culture process with five factors might require dozens of experimental runs to capture all possible interactions and achieve statistical significance. This level of experimentation is often impractical in an industrial setting, forcing developers to accept degrees of uncertainty and risk.

How is this different with hybrid models? Due to the mechanistic backbone certain variable interactions are already accounted for in grounded scientific knowledge, wherefore experimental evidence is not required. In addition, hybrid models can make use of data from previous development phases to either answer the question at hand directly or aid in the design of a limited number of process runs to gather the insight.

The integration of mechanistic process knowledge and the ability to leverage insights from previous phases significantly reduces the need for extensive new experimental data and process runs, enabling a more efficient and reliable understanding of process objectives.

DoE models perform well within the design space of their training data but struggle to extrapolate beyond it. Predictions for conditions outside the tested range—such as during scale up —are often unreliable, forcing developers to rely on additional experiments or assumptions.

Hybrid models excel at both interpolation (predicting outcomes within the design space) and extrapolation (predicting outcomes outside it). By combining mechanistic and data-driven components, they can better predict system behavior under untested conditions.

• Scenario: Scaling up a process from 5 L to 500 L bioreactors often introduces variability due to changes in mixing, oxygen transfer, and heat dissipation. A hybrid model can incorporate mechanistic principles for scale effects while adapting to real-world data, ensuring accurate predictions for the larger system.

DoE models focus on correlating inputs and outputs but provide limited insights into the underlying mechanisms driving those relationships. Moreover, they largely disregard process dynamics as evolutions of cell or metabolite concentrations are not captured. This limitation of process understanding can hinder troubleshooting and limit the ability to generalize findings to new systems or conditions.

Hybrid models provide insights into the underlying mechanisms driving process behaviour and are uniquely equipped to capture the nonlinear and time-dependent relationships that characterize biopharmaceutical processes.

By blending mechanistic equations with machine learning, they can model dynamic systems with high fidelity and with relative ease. The mechanistic backbone which incorporates first principles, also offers a transparent framework that explains hy specific inputs lead to certain outputs.  They further capture the dynamic nature of the process explicitly.

• Hybrid vs DoE: While a DoE model might indicate that increasing agitation speed improves productivity, a hybrid model could explain that the improvement is due to enhanced oxygen transfer and reduced shear stress during the process evolution, offering actionable insights for optimization.

As biopharmaceutical processes grow more complex, hybrid models are emerging as the gold standard for process development and optimization. Their ability to integrate mechanistic understanding with data-driven adaptability addresses the limitations of traditional DoE statistical models, offering unparalleled predictive accuracy, efficiency, and insight.

Adopting hybrid models is not just an incremental improvement, it’s a transformative step forward. By embracing this technology, biopharmaceutical companies can achieve faster development timelines, more robust processes, and greater scalability while maintaining the highest standards of quality and compliance. They are also believed to be the key to high fidelity digital twins.

Read a case study about the impact of Hybrid Models vs Statistical “Black Box” DoE Models