Machine Learning Guided by Physicochemical Principles Enables Generalized Prediction of Small-Molecule Subcellular...

Area of Science:

• Chemical Biology
• Computational Chemistry
• Molecular Imaging

Background:

• Precise subcellular localization is essential for designing molecular probes and targeted therapeutics.
• Distinguishing organelles with similar physicochemical properties (e.g., lipid droplets, mitochondria, cell membrane) is challenging.
• Traditional methods lack generalizability and struggle with complex structure-activity relationships.

Purpose of the Study:

• To develop an innovative machine learning (ML) strategy for precise prediction of small-molecule organelle localization.
• To overcome limitations of traditional empirical methods in predicting molecular partitioning.
• To guide the design and synthesis of novel chemical probes with specific organelle targeting.

Main Methods:

• Developed an ML model learning fundamental physicochemical forces governing molecular partitioning.
• Utilized a modest dataset of 355 samples for training.
• Prospectively employed the model to screen a probe library and guide synthesis.

Main Results:

• Achieved exceptional predictive accuracy (>94% cross-validation) for small-molecule localization.
• Successfully designed and synthesized three aggregation-induced emission probes with predicted organelle-targeting specificities.
• Demonstrated model generalization by confirming probe localization and explaining anomalous targeting behaviors.

Conclusions:

• ML, guided by physicochemical principles, can accurately predict organelle localization.
• This approach empowers chemical biology research and drug discovery by enabling efficient development of functional chemical tools.
• The developed model offers a generalizable paradigm for creating targeted therapeutics and molecular probes.