Abstract
Abstract
When nanoparticles enter biological environments, their surfaces are rapidly coated by proteins, forming a protein corona that governs their biological identity and downstream responses. Existing computational methods treat each protein as an independent prediction task, preventing the learning of shared adsorption principles and limiting generalization beyond experimentally characterized proteins. Here we introduce NanoPCT, a universal protein-conditioned learning framework that jointly models nanoparticle physicochemical properties and protein sequence representations. Using frozen ESM-2 embeddings and a joint interaction transformer, NanoPCT learns transferable representations of nano–bio interaction patterns across hundreds of proteins and maintains robust performance as protein diversity increases. The learned protein representation further enables zero-shot prediction for previously unseen proteins and biologically distinct protein groups through classifier reconstruction in embedding space. Beyond prediction, NanoPCT supports large-scale in silico inverse design and reveals consistent adsorption preferences between observed and unseen protein populations, suggesting that it captures transferable interaction patterns rather than protein-specific memorization. NanoPCT provides a unified framework for protein corona prediction, zero-shot generalization, and rational nanomaterial design.