A scientific group, in collaboration with the Monell Chemical Senses Center and the OSMO startup based in Cambridge, is investigating the relationship between airborne chemicals and the brain’s perception of odor. Recent studies published in Science magazine have revealed that a machine learning-based model can better describe the smells of various chemicals compared to a human being.
“The model resolves long-standing gaps in our understanding of the sense of smell,” commented Joel Mainland, the chief co-author of the study.
The primary objective of the study was to determine how molecular structures are correlated with the perception of smell. The team developed a model that learned to connect descriptions of smells with the molecular structure of the corresponding substances. The model was trained using a dataset comprising 5,000 known substances with distinct odors.
To assess the model’s effectiveness, researchers from Monell conducted a “blind” verification process. Participants in the experiment described new molecules, and their responses were then compared to the model’s descriptions. The model outperformed the predictions of individual participants in the panel.
Remarkably, the model demonstrated the ability to perform tasks for which it was not originally trained, such as determining the intensity of a smell.
In the future, this map of smell perception could prove valuable for researchers in the fields of chemistry, neurobiology of smell, and psychophysics, offering a novel tool for studying olfactory sensations. Additionally, it opens up possibilities for creating new aromas and flavors in the fragrance and food industries.