While vision and hearing are already digitized, the sense of smell has proven to be more complex and perplexing. Researchers at Monell Chemical Senses Center and start-up Osmo, which was spun out of machine learning research at Google Research, are studying how airborne chemicals connect to odor perception in the brain. They discovered a machine-learning model achieving human-level proficiency at describing, in words, how chemicals might smell. Their research is published in Science.
The research brings us closer to digitizing odors for recording and reproduction and could potentially identify new odors for the fragrance and flavor industry. This could decrease dependence on naturally sourced endangered plants and identify new functional scents such as mosquito repellent or malodor masking.
The team’s model learned how to match the prose descriptions of a molecule’s odor with the odor’s molecular structure. The study proposes and validates a novel data-driven map of human olfaction and matches chemical structure to odor perception.
The model was trained using a dataset including the molecular structures and qualities of 5,000 known odorants. Researchers conducted a blind validation procedure in which a panel of trained research participants described new molecules and compared their answers with the model’s description. Fifteen panelists were each given 400 odorants and trained to use a set of 55 words to describe each molecule. Training sessions and lab-designed odor reference kits taught each panelist the odor quality associated with each descriptive term.
The panelists were asked to select which of the 55 descriptors applied and to rate the extent to which the term best pertained to the odor on a 1-to-5 scale for each of the 400 scents. For example, one panelist rated the smell of the previously uncharacterized odorant 2,3-dihydrobenzofuran-5-carboxaldehyde as very powdery (5) and somewhat sweet (3).
The model’s performance was compared to that of individual panelists, and the model achieved better predictions of the average of the group’s odor ratings than any single panelist in the study, performing better than the average panelist for 53% of the molecules tested.
Surprisingly, the model succeeded at olfactory tasks it was not trained to do, such as predicting odor strength. It identified dozens of pairs of structurally dissimilar molecules with counter-intuitively similar smells while characterizing a wide variety of odor properties, such as odor strength, for 500,000 potential scent molecules.