Machine learning algorithms to predict spray dried protein/peptide formulations

Proteins and peptides hold immense promises for treating fatal and rare diseases. However, their complex structures and intrinsic instability pose significant challenges in drug development. The spray drying process offers a continuous and rapid method to stabilize these biomacromolecules, converting them from liquid to solid formulations. However, the spray drying process remains trial-and-error based, requiring extensive resources. This study employs machine learning (ML) algorithms to predict key properties of spray-dried protein and peptide powders, including yield, particle size, residual solvent content, solid states properties, and aggregation, with the aim to help accelerate formulation development and optimize process parameters. In total, 321 yield data points, 288 particle size data points, 357 residual solvent content data points, 205 solid states of the dry powder data points, and 305 aggregation data points were collected and described using various molecular descriptors for model building. Seven ML algorithms were tested to identify the best model to predict the key properties. Light Gradient Boosting Machine (LightGBM) exhibited the best performance for regression tasks, particularly for residual solvent content (mean absolute error = 0.841), and logistic regression excelling in predicting solid state characteristics and aggregation. Feature importance analysis identified protein, excipient, processing parameters, and environmental conditions as critical factors influencing various properties of spray-dried proteins/peptides formulation. The generalizability of the models was experimentally validated using alpha-lactalbumin formulations, the mean absolute error (MAE) in the models of yield, particle size, and residual solvent content were 0.755, 1.591, and 14.492 respectively, and the accuracies in predicting solid states of the dry powder and aggregation were 78.0% and 100.0% respectively. This study demonstrates the potential of machine learning to streamline the development of spray-dried protein formulations, providing a material- and time-saving solution as a reference.