AI-based Docking Methods Fail to Predict Olfactory Receptor Activation

The combinatorial mapping from odorants to olfactory receptors (ORs) remains a central mystery for the chemical senses. We evaluate physics-based docking methods, structure prediction models, and OR-specific frameworks on a ligand-generalization benchmark. Across model families, performance approaches chance when predicting functional activation for unseen odorants across a panel of human ORs. Further analysis reveals that all models consistently underperform simple ligand marginals and fail to demonstrate robust ligand-receptor specificity.