Data-driven prediction of polydispersity in microfluidic liposome production for pharmaceutical formulation optimization

A data-driven approach was developed to model microfluidic liposome formulation and predict two important quality attributes: particle size (SIZE) and polydispersity index (PDI). The data set contained 306 formulations with process factors (e.g., flow rate ratio and total flow rate) and formulation factors, supported by initial exploratory analysis (feature engineering) and interpretability analysis. Eleven machine learning (ML) and deep learning (DL) models were tested. Predictions of PDI were highly accurate, with the best validation performance achieved with gradient boosting (R 2 = 0.765; RMSE = 0.106; MAE = 0.059), which exceeded a previously reported benchmark. However, predictions of SIZE were much less accurate than those of PDI, with the best validation performance achieved with a deep multilayer perceptron (R 2 = 0.460), which is consistent with the presence of target skewness and outliers in the size data. Both feature importance and SHAP analyses indicated that engineered ratios of the formulation composition, particularly the main lipid-to-PEG ratio, along with process effects, such as the flow rate ratio (FRR), were primary contributors to the prediction of PDI. Overall, the findings provide evidence supporting the application of explainable ML to direct the development of microfluidic liposomes, specifically for improving formulation uniformity, while indicating the need for larger and more robust data sets and strategies to improve the accuracy of size predictions.