Interpretable Latent Representations of Odor Perception by Integrating Molecular Structure and Olfactory Receptor Responses

<title>Abstract</title> <p>Odor perception arises from many-to-many relationships among molecular structures, olfactory receptors, and perceptual descriptors, making it difficult to define a compact and mechanistically interpretable odor space. Here we present a receptor-informed sparse representation framework for odor perception by integrating molecular structure, olfactory receptor (OR) docking profiles, and human odor-descriptor data. A curated set of 136 ATLAS odorants was represented by docking scores against 409 human olfactory receptor models and by 146 ATLAS perceptual descriptors. We applied sparse autoencoder (SAE)-based representation learning to extract low-dimensional latent odor components from receptor-only and joint receptor–perceptual input matrices. The Joint SAE identified interpretable latent odor concepts, including green–vegetable, musk–perfumery, citrus–fruity, sulfur–garlic/gas, and burnt–smoky/rubber. These components were characterized by coherent ATLAS descriptor profiles, chemically interpretable RDKit feature associations, representative odorant molecules, and distributed olfactory receptor ensembles. Sequence-based receptor-tree analysis and OR-subfamily aggregation showed that the receptor signatures of these concepts are not confined to single receptor families but are distributed across multiple receptor subfamilies, consistent with combinatorial odor coding. Comparison with M2OR records provided partial experimental support and suggested candidate receptor–odorant pairs for future deorphanization studies. Cross-dataset validation using the Keller-Vosshall dataset further showed that several latent odor dimensions were concordant with independent human perceptual ratings. These results suggest that odor space can be represented by sparse, receptor-informed latent variables linking molecular structure, receptor interaction patterns, and perceptual organization.</p>