Most digital twins used in biopharma are built with process data. But, while this fits with ideas like Quality by Design (QbD), it is more labor intensive than it needs to be, according to the authors of a new study who suggest hybrid models are a better option.
Digital twins—computer simulations used to design, test, and fine-tune physical processes—are made up of three components: a physical system, a digital infrastructure and, most importantly, a virtual system or model.
The models on which twins are based are either mechanistic—leveraging theoretical understanding of the underlying mechanisms of a process—or empirical, based on experimental data derived from experimentation.
Most twins employed in biopharma use empirical models created from observed relationships between process inputs and outputs. These models utilize statistical methods such as regression analysis for development.
Production process insights
The main advantage is that they provide insights into the behavior of a production process without the need for in-depth understanding, in theory making them more straightforward to build. It is also in keeping with QbD principles.
However, the approach is not well suited to highly complex production processes say researchers at the University of Cork and Netherlands-based engineering firm DPS.
“The data-centric method has notable disadvantages. It often necessitates an increased number of experiments, particularly with an increasing number of process parameters, leading to a ‘curse of dimensionality.’”
“Additionally,” the researchers continue, “insights from this method are generally limited to the specific data set analyzed, which restricts the broader application of the knowledge gained.”
Instead, the authors suggest that for experienced biopharmaceutical process developers, twins based on first principles are a better option.
“The knowledge-centric approach offers a deeper comprehension of systems and is built on equations refined over time. This method allows for knowledge transfer across projects as physical laws remain consistent. However, its application can be challenging due to the extensive understanding and expertise needed, especially as the number of process parameters and their potential interactions increase.”
Another alternative would be to combine both modeling methods, according to the authors. In a hybrid approach, a mechanistic framework could be used to describe the biological system, while cell-specific characteristics—like growth rate—could be expressed as statistical expressions using input variables and parameters.
“Hybrid models leverage the depth of the knowledge-centric method and the simplicity of the data-centric approach. They provide a cost-effective and modular way to understand key system behaviors, enhancing the effectiveness of digital twins,” they write.