Dual-Module Near-Infrared Fluorophores Discovery System via Knowledge Transfer

In vivo near-infrared (NIR) imaging is an emerging technique in biomedical research. It is particularly important for examining living tissue owing to its deep tissue penetration and low autofluorescence within the NIR optical window. Here, we present a deep learning discovery system for suggesting potential NIR fluorophores. Our approach employs a dual-module framework, incorporating a predictive module with transfer learning to estimate properties and a generative module for constructing synthetically accessible NIR fluorophore candidates. This system predicts key optical properties, addressing limitations of labor-intensive experimentation, with transfer learning incorporated to handle data scarcity. Through this system, three molecules (NTDT-TPA, NPA-BTD, DPP-TPA) are synthesized for experimental validation. Among them, NTDT-TPA is formulated into nanoparticles and evaluated in vitro and in vivo, showing its potential for fluorescence bioimaging.