Data-driven Olfactory Cartography

Quantitative maps connecting sensory phenomena with each other and with physical stimuli are essential to understanding the sensory world and the neural computations that underlie it.  Maps for normal color vision have existed in vision for >100 years (e.g. the color wheel), but have remained elusive in olfaction (there is no smell wheel).  Obstacles have included the difficulty of expressing odor stimuli (i.e. molecules) as fixed-length vectors, collecting sufficient sensory data, and finding a modeling framework for linking stimulus and percept.  I describe a novel graph neural network approach to solving this problem, linked with new datasets of unprecedented size in olfaction, leading to a performant model whose embedding space has several powerful features: (1) unification of previously disparate psychophysical measurements under a single geometry; (2) human-level performance in olfactory prediction tasks; (3) an understanding of olfactory perception in terms of nature's metabolic graph; and (4) computational molecular screening for novel odorants and insect repellents.