Small molecule near-infrared (NIR) fluorophores play a critical role in disease diagnosis and early detection of various markers in living organisms. To accelerate their development and design, a deep learning platform, NIRFluor, was established to rapidly screen small molecule NIR fluorophores with the desired optical properties. The core component of NIRFluor is a state-of-the-art deep learning model trained on 5179 experimental big data. First, novel hybrid fingerprints including Morgan fingerprints, physicochemical properties, and solvent properties were proposed. Then, a powerful deep learning model, multitask fingerprint-enhanced graph convolutional network (MT-FinGCN), was designed, which combines fingerprint information and molecule graph structure information to achieve accurate prediction of six properties (absorption wavelength, emission wavelength, Stokes shift, extinction coefficient, photoluminescence quantum yield, and lifetime) of different small molecule NIR fluorophores in different solvents. Furthermore, the "black-box" of the GCN model was opened through interpretability studies. Finally, the well-trained models were placed on the web platform NIRFluor for free use (https://nirfluor.aicbsc.com).
Read full abstract