Mapping the combinatorial coding between olfactory receptors and perception with deep learning.

The sense of smell remains poorly understood compared with vision and audition. At its core is an information flow in which odorant molecules activate subsets of olfactory receptors (ORs) and combinations of receptor activations encode distinct percepts. However, predicting molecule-OR interactions, and linking them to perception, remains difficult. Here, we develop MolOR, an approach that maps odorants to their OR-activation profiles and then predicts their odor percepts. Using cross-attention between a graph neural network over molecules and protein-language-model embeddings of receptors, we predict OR activation and-despite no molecular overlap between binding and percept datasets-improve percept prediction by using predicted OR profiles as auxiliary features. Structurally diverse molecules sharing a percept show similar predicted OR profiles, and the model distinguishes protein-coding ORs from pseudogenes across the human subgenome. This may aid the discovery of ligands for orphan ORs and the design of odorants with desired perceptual qualities.