Abstract
<jats:p>However, serum proteins bind to nanoparticle surfaces, changing how they interact with on-and off-target cells. Nanoparticles have a vast design space spanning continuous and categorical design features, including size, material, charge, and surface ligands. In this work, we study representations of nanoparticle physicochemical properties to enable Gaussian Processes and Prior-Data Fitted Networks to predict the serum protein binding affinity of nanoparticles. These models represent the first data-driven approaches to predicting serum protein-nanoparticle affinities. Despite scarce data (≈ 250 data points), the models generalize to unseen nanoparticle designs with moderate predictive accuracy (R2 ≈ 0.50) and strong ranking performance (Spearman’s ρ ≈ 0.73). Because experimental design depends on correctly prioritizing candidates rather than on exact prediction, the latter is a key measure of practical utility. Our work demonstrates that predictive models trained on small but systematically measured datasets may enable model-guided exploration of nanoparticle design properties, reducing the experimental burden of navigating a large design space.</jats:p>