The adsorption of biomolecules onto nanoparticle (NP) surfaces, forming the so-called biocorona (BC), plays a crucial role in defining the biological identity and fate of nanomaterials. While extensive research has focused on protein coronas, lipid adsorption remains underexplored despite its critical importance in physiological environments. In this study, we introduce a systematic machine learning framework to predict lipid corona formation on iron oxide (Fe3O4) nanoparticles using experimentally obtained lipidomic data comprising 1740 distinct lipids. Each lipid molecule was described by eight interpretable physicochemical descriptors, and a Sparse Gaussian Process (SGP) was trained to predict binary adsorption outcomes across 20 biological conditions defined by NP size, serum concentration, and donor sex. Analysis of individual groups revealed the highest performance for males with 75% serum and 50 nm particles (accuracy = 0.77, recall = 0.79, precision = 0.74, F1 score = 0.76, and AUROC = 0.83) underscoring the importance of class balance for model accuracy. SHapley Additive exPlanations (SHAP) analysis indicated that features such as hydrogen bond acceptor (HBA) count, van der Waals volume, and log P played key roles in NP-lipid interactions. Despite challenges posed by class imbalance in low-serum groups, the SGP framework offered generalizable predictions and mechanistically interpretable insights. This work lays a data-driven foundation for predictive modeling of lipid coronas, opening avenues for multiomics integration in nanoparticle design and evaluation.
Basu et al. (Mon,) studied this question.