Machine learning-driven discovery of novel photosensitizer for cancer therapy

Photodynamic therapy enables spatially controlled cancer treatment but remains limited by the scarcity of effective photosensitizers. A main bottleneck limiting the discovery of new photosensitizers is the high computational cost of predicting excited-state properties critical for therapeutic activity. Here, we introduce X-MACE, a neural network that predicts key excited-state and redox properties with near quantum chemical accuracy at a fraction of the original computational costs. Screening thousands of drug-like molecules within minutes, we identified candidates active in the therapeutic window. Experimental validation of the top-ranked compound confirms visible light absorption and strong light enhanced cytotoxicity across four cancer cell lines, with up to 302-fold IC₅₀ improvement upon irradiation. This framework transforms photosensitizer discovery from empirical screening to physics-informed rational design.