Geometric-aware deep learning enables discovery of bifunctional ligand-based liposomes for tumor targeting therapy

Abstract

Limited tumor targeting capacity of conventional liposomes compromises their clinical outcomes in tumor therapy. Although ligand-based liposomes show promise for improved tumor targeting efficiency, their transition to clinical use is impeded by the complexity of necessary ligand modifications on liposomal membranes. Certain bifunctional natural products, offering both liposomal membrane-regulating and tumor-targeting ligands properties, have shown tumor targeting potential after prepared into liposomes without the need for ligands synthesis, but their discovery has been hindered by the constraints of conventional screening methods. Here, we propose combining deep learning with wet experimentation for rapid discovery of new bifunctional ligands. Utilizing pre-trained geometric-aware neural networks, we simultaneously modeled predictions for membrane-regulating and glucose transporter 1-ligand functions. The trained models identified nine top candidates from > 300,000 natural products, six of which demonstrated the anticipated dual functionality upon experimental validation. The lead liposome, Ilexgenin A (Ile)-based liposome, demonstrated superior tumor-targeting and anti-tumor effect compared to the existing bifunctional ligand-based liposome. Further analysis elucidated Ile's mechanisms in immunoregulation and chemotherapy sensitization. This approach signifies the potential of deep learning in design of intelligent and targeting drug delivery systems.