Deep learning-based prediction of highly efficient optical probes for two-photon fluorescence imaging and photodynamic therapy of tumor

Abstract

This study reports an innovative intelligent screening strategy based on a directed message-passing neural network (DMPNN) to predict two key physical properties of corrole molecule probes involving the Q_band absorption and singlet–triplet intersystem crossing rate (kISC). Feature engineering and heterogeneous model fusion are performed, including the concatenation of molecular fingerprints, integration with additional algorithms, incorporation of physical constraints into molecular descriptors, and introduction of attention mechanisms. The DMPNN model alone achieves excellent predictive performance for Q_band absorption. The combination of the DMPNN with specific Morgan fingerprints yields an R2 of up to 0.926 for kISC prediction. Interpretable analyses of node- and bond-level features reveal the primary structural information related to the metal center and its N4 coordination environment, the conjugated macrocyclic framework, and the presence of lone pair electrons. The model screens gallium-centered corrole derivatives with peripheral pentafluorophenyl substitutions to identify promising candidates. According to two-photon fluorescence lifetime imaging, the probes enable effective discrimination between normal and cancer cells. Under red-light irradiation, the probe efficiently generates reactive oxygen species and effectively eliminates cancer cells both in vitro and in vivo. The intelligent screening strategy provides a new paradigm for developing high-performance optical theranostic probes with strong potential for clinical translation.