A multi-step machine learning approach for accelerating QbD-based process development of protein spray drying

This study proposes a new material-efficient multi-step machine learning (ML) approach for the development of a design space (DS) for spray drying proteins. Typically, a DS is developed by performing a design of experiments (DoE) with the spray dryer and the protein of interest, followed by deriving the DoE models via multi-variate regression. This approach was followed as a benchmark to the ML approach. The more complex the process and required accuracy of the final model is, the more experiments are necessary. However, most biologics are expensive and thus experiments should be kept to a minimum. Therefore, the suitability of using a surrogate material and ML for the development of a DS was investigated. To this end, a DoE was performed with the surrogate and the data used for training the ML approach. The ML and DoE model predictions were compared to measurements of three protein-based validation runs. The suitability of using lactose as surrogate was investigated and advantages of the proposed approach were demonstrated. Limitations were identified at protein concentrations >35 mg/ml and particle sizes of x₅₀ > 6 µm. Within the investigated DS protein secondary structure was preserved, and most process settings, resulted in yields >75% and residual moisture <10 wt%.