Machine Learning Guided by Physicochemical Principles Enables Generalized Prediction of Small-Molecule Subcellular Localization and Discovery of Targeted Molecules

Precise subcellular localization is crucial for the design of molecular probes and targeted therapeutics, yet selectively distinguishing organelles with similar physicochemical properties, such as lipid droplets, mitochondria, and the cell membrane, remains a formidable challenge. Traditional empirical methods struggle to capture complex structure-activity relationships and exhibit limited generalizability. Here, we report an innovative machine learning strategy that achieves precise prediction of small-molecule localization among these organelles by learning the fundamental physicochemical forces governing molecular partitioning. Notably, this approach yields exceptional predictive accuracy (cross-validation accuracy >94%) using only a modest data set of 355 samples. We then prospectively employed the resulting model to guide the screening of a probe library based on a diphenylamine-coumarin-thiophene scaffold, which was absent from the training set. This led to the design and synthesis of three aggregation-induced emission probes with their predicted organelle-targeting specificities. Subsequent experiments confirmed their precise localization, demonstrating the model's generalization ability. Furthermore, the model rationally explains the anomalous targeting behaviors of previously reported probes. This work showcases the potential of ML guided by physicochemical principles to empower chemical biology research and drug discovery, offering a highly generalizable paradigm for the efficient development of chemical tools with precise biological functions.