Decoding Smell from Receptor Structure

Abstract Olfaction enables animals to detect and discriminate an immense diversity of volatile chemicals, yet how individual odorant receptors determine ligand selectivity remains poorly understood (Buck and Axel 1991, Bushdid, Magnasco et al. 2014, Niimura, Matsui and Touhara 2014, Tsukahara, Brann et al. 2021). Recent cryo-electron microscopy structures of mammalian odorant receptors(Billesbølle, de March et al. 2023, Choi, Bae et al. 2023, de March, Ma et al. 2024, Ge, Cheng et al. 2026, Gil, Choi et al. 2026, Han, Zhang et al. 2026, Wang, Wu et al. 2026), together with advances in protein structure prediction(Jumper, Evans et al. 2021, Lin, Akin et al. 2023, Abramson, Adler et al. 2024), now provide an opportunity to test whether receptor structure can explain odor recognition at scale. Here, we combine AlphaFold3-predicted odorant receptor structures with ESM2-derived protein sequence embeddings and large-scale sequencing linked to in vivo olfactory sensory neuron activation across a chemically diverse odorant panel. Using these data, we develop a structure-informed deep learning framework that predicts odorant receptor-ligand interactions and maps relationships between receptor space and chemical space. The learned receptor representation organizes odorant receptors by functional similarity rather than primary sequence alone, while feature attribution identifies binding-cavity subregions that contribute to chemical recognition. Together, these results show that odorant receptor ligand selectivity is encoded in three-dimensional receptor features and establish a predictive framework for interpreting odorant receptor coding.