Digital twins—predictive computer simulations of drug production processes—are only as good as the software from which they are constructed. The key to building them, say researchers, is knowing which algorithms will yield the most accurate models.
Industry interest in twins has increased markedly in recent years. Patient demand for safer, higher quality medicines combined with manufacturers’ desire to make production more efficient has fueled demand for process optimization tools.
Regulations are also a factor. For example, ICH Q8 (2) guidelines on pharmaceutical development to mathematical models are used to gain enhanced process understanding and meet Quality-by-Design (QbD) goals.
Selection
But developing a digital twin is not straightforward according to Dong-Yup Lee, PhD, associate professor at the School of Chemical Engineering at Sungkyunkwan University in Korea, who says software choice is critical.
“It is essential to select best artificial intelligence or machine learning (AI/ML) algorithms for best prediction. Simply relying on some methods which you may be familiar with is risky since basically data-driven models (DDM) are a black-box approach and condition specific. The model may not be applicable when conditional variation occurs. In addition, we may not explain why we have such results.”
To try and address this, Lee and colleagues developed an evaluation framework for real-time data-driven predictive models.
“The framework uses historical data and real-time measurements to select the best prediction model structures, including forecasting strategies, model inputs, and machine and or deep learning algorithms for predicting near-future profiles or identifying fault patterns upon receiving new monitoring data.
“We can use the reliable AI/ML models to find optimal process conditions, such as basal or feed media, and use them to forecast culture behavior in real time. In theory, faulty runs could be predicted and process inputs can be adjusted to correct them by control.”
And the approach can be used to determine the predictive accuracy of a wide range of different types of model algorithms, Lee says.
“We can train a large number of culture profiles generated from many bioreactor runs under various or similar operating conditions in real time, allowing us to build effective AI/ML models to relate inputs—like glucose, lactate, ammonia, DO, pH—and output culture performance, for example VCD, viability, and IgG titer.”
The new method can even be used to test the accuracy of process models that combine both statistical and mechanistic elements.
“We are currently working on a hybrid model where the DDM and mechanistic models can be combined or linked to better explain and comprehensively predict culture behaviors described by cell growth, viability, productivity, and metabolite concentration as well as process parameters in real time.”