Cross-Scale Discovery of Surface Ligand-Modified Nanoparticles via Combined Transfer-Cascade Machine Learning

Surface ligand-modified nanoparticles (SMNPs), which immobilize surface ligands on nanoscale scaffolds, represent important nanomaterials for diverse applications. While surface ligands critically dictate the functions of SMNPs, their systematic discovery remains elusive. The major challenge in using artificial intelligence to predict and screen effective surface ligands and SMNPs across extensive chemical space is the scarcity of high-quality, standardized data sets. This impedes SMNP nanotechnology and leaves an enormous functional landscape largely unmapped. For antibacterial gold nanoparticles, available surface-ligand data sets are extremely sparse (<102), whereas public small-molecule libraries exceed 1010 entries. Here, we introduce a cross-scale (from small molecules to nanoparticles) transfer-cascade learning framework, called TRADE, that transforms the structural–activity patterns of molecular antibiotics into predictive principles for the surface ligands of nanoparticle antibiotics. Within TRADE, we chemically deconstruct molecular antibiotics into conserved scaffolds and variable substituents for transfer learning and combine a multitier, coarse-to-fine cascade to narrow the search space. We identify a new class of 24 surface ligands with high predicted antibacterial efficacy for decorating gold nanoparticles across over 9 million molecules. The leading auricidin inhibits bacterial peptidoglycan biosynthesis and demonstrates potent antimicrobial activity against multidrug-resistant pathogens, including methicillin-resistant Staphylococcus aureus and vancomycin-resistant Enterococcus faecium, with minimum inhibitory concentrations as low as 0.5–1 μg/mL. Across multiple animal infection models, auricidin exhibits exceptional therapeutic efficacy and robustness, superior to vancomycin, minimal toxicity, and low susceptibility to resistance evolution. This work proposes an intelligent strategy for discovering new functional nanoscale entities by integrating cross-domain transfer and cascade learning.