Machine Learning–Driven QSAR Modeling for pKa Prediction of Ionizable Lipids in Lipid Nanoparticles for Hepatic Gene Silencing

Liver cancer remains a significant global health burden, requiring the development of precise nucleic acid delivery systems. Lipid nanoparticles (LNPs) are leading candidates; however, their efficiency is governed by the pKa of ionizable lipids, which dictates nanoparticle stability and endosomal escape. In this study, we employed a machine learning–driven quantitative structure–activity relationship framework to predict the pKa of ionizable lipids derived from the DLin–KC2–DMA scaffold. Utilizing a dataset of 56 compounds, we compared Random Forest, Artificial Neural Network, and Extreme Gradient Boosting (XGB) models integrated with Permutation Importance (PI) for feature selection. The optimized PI–XGB model exhibited exceptional predictive accuracy (R2 = 0.970, R2CV = 0.901, RMSEtest = 0.115) and robust generalization confirmed via external validation (RMSEext. = 0.313). Mechanistic insights derived from SHapley Additive exPlanation analysis identified charge distribution, molecular topology, and polarity as critical determinants of lipid ionization. These results demonstrate the power of interpretable machine learning in elucidating molecular structure–property relationships, offering a robust computational strategy for the rational design of next–generation ionizable lipids to optimize LNP–mediated gene therapy for liver cancer.