A modular artificial intelligence framework to facilitate fluorophore design

Abstract

Fluorescence imaging, indispensable for fundamental research and clinical practice, has been driven by advances in fluorophores. Despite fast growth over the years, many available fluorophores suffer from insufficient performances, and their development is highly dependent on trial-and-error experiments due to subtle structure-property effects and complicated solvent effects. Herein, FLAME (FLuorophore design Acceleration ModulE), an artificial intelligence framework with a modular architecture, is built by integrating open-source databases, multiple prediction models, and the latest molecule generators to facilitate fluorophore design. First, we constructed the largest open-source fluorophore database to date (FluoDB), containing 55,169 fluorophore-solvent pairs. Then FLSF (FLuorescence prediction with fluoroScaFfold-driven model) with a domain-knowledge-derived fingerprint for characterizing fluorescent scaffolds (called fluoroscaffold) was designed and demonstrated to predict optical properties quickly and accurately, whose reliability and potential have been verified via molecular and atomistic interpretability analysis. Further, a molecule generator was incorporated to provide new compounds with desired fluorescence. Representative 3,4-oxazole-fused coumarins were synthesized and evaluated, creating an unreported compound with bright fluorescence.