<i>De novo</i> Design of Biocompatible Nanomaterials\nUsing Quasi-SMILES and Recurrent Neural Networks

Screening nanomaterials (NMs) with desired properties\nfrom the\nextensive chemical space presents significant challenges. The potential\ntoxicity of NMs further limits their applications in biological systems.\nTraditional methods struggle with these complexities, but generative\nmodels offer a possible solution to producing new molecules without\nprior knowledge. However, converting complex 3D nanostructures into\ncomputer-readable formats remains a critical prerequisite. To overcome\nthese challenges, we proposed an innovative deep-learning framework\nfor the <i>de novo</i> design of biocompatible NMs. This\nframework comprises two predictive models and a generative model,\nutilizing a Quasi-SMILES representation to encode three-dimensional\nstructural information on NMs. Our generative model successfully created\n289 new NMs not previously seen in the training set. The predictive\nmodels identified a particularly promising NM characterized by high\ncellular uptake and low toxicity. This NM was successfully synthesized,\nand its predicted properties were experimentally validated. Our approach\nadvances the application of artificial intelligence in NM design and\nprovides a practical solution for balancing functionality and toxicity\nin NMs.