Molecular Nearest Neighbors Determine Generalization of Odorant Mixtures
Most natural odors arise from mixtures of multiple odorants. Some such mixtures are perceived \"elementally\", with each odorant component clearly identifiable, while others are perceived \"configurally\", with the mixture adopting a perceptual quality distinct from any of the components. While the perceptual similarity of two mixtures is presumably related in some way to the similarity of the corresponding components, given the elemental/configural dichotomy it is unclear if any formal principle can be used to predict mixture similarity. To investigate this problem, we trained mice to respond to a binary reference mixture of structurally similar odorants (S+) and then tested generalization of this response to other structurally related binary test mixtures. Across 5 experiments, we parametrically varied these mixtures in distinct ways to test candidate models for the perceptual similarity of mixtures. The best-performing model predicted behavioral responses by considering, for each component of the S+, only the similarity of the most structurally similar (\"nearest neighbor\") component of each test mixture. We conclude that for mixture generalization tasks the olfactory system may deemphasize or discard information about mixture components not perceptually \"near\" enough to any of those in the S+, consistent with a sparse and elemental rule for perception of structurally-related binary mixtures.