One key focus in neuroscience is unraveling how our senses interpret stimuli: light becomes sight, sound becomes hearing, and so on. The sense of smell, however, presents a unique challenge.
A collaborative research team, led by the Monell Chemical Senses Center and startup Osmo, spun out of Google Research’s machine learning efforts, is delving into the intricate relationship between airborne chemicals and our perception of odors.
Their breakthrough involves a machine-learning model that matches human proficiency in describing chemical scents in words. It has identified structurally dissimilar molecules with surprisingly similar smells and characterized odor properties for an astounding 500,000 potential scent molecules. This research, published in Science, fills gaps in our understanding of smell and moves us closer to digitizing scents.
Humans possess around 400 olfactory receptors, far more than those used for color vision or taste. Understanding how molecular structure correlates with odor perception has long eluded researchers. Osmo’s model bridges this gap, proposing a data-driven map of human olfaction.
Training the model involved a dataset with molecular structures and odor characteristics of 5,000 known odorants. Validation included human assessments, and the model remarkably outperformed individual panelists for 53% of molecules tested, even in tasks it wasn’t explicitly trained for.
This model promises to be a valuable tool for chemistry, olfactory neuroscience, and psychophysics research, potentially reshaping our understanding of how our brains organize odors based on their nutritional origins.