A significant challenge in neuroscience is unraveling the complexities of sensory perception, particularly when it comes to the sense of smell. While it’s relatively straightforward to understand how our senses translate light into sight or sound into hearing, decoding how chemicals in the air become odor perceptions in the brain is a more intricate puzzle. However, a recent breakthrough in machine learning and olfaction research, led by the Monell Chemical Senses Center and start-up Osmo, could pave the way for digitizing odors and expanding our understanding of the sense of smell.
The research team’s primary focus was to investigate how airborne chemicals correlate with odor perception in the human brain. Their efforts culminated in the development of a machine-learning model that achieved human-level proficiency in describing odors using words. The results of their study were published in the September 1 issue of Science.
Joel Mainland, a senior co-author and Monell Center Member, emphasized that the model addresses long-standing gaps in scientific knowledge about the sense of smell. This breakthrough not only brings us closer to digitizing odors for recording and reproduction but also holds potential implications for industries like fragrance and flavor, where it could lead to the discovery of new scents for various applications, including mosquito repellents and malodor masking.
Intriguingly, the human sense of smell is a complex interplay of approximately 400 functional olfactory receptors, which are specialized proteins located at the end of olfactory nerves. These receptors interact with airborne molecules, transmitting electrical signals to the olfactory bulb in response. This number of olfactory receptors significantly exceeds those dedicated to other senses, such as the four used for color vision or the roughly 40 for taste.
The fundamental question that has confounded olfaction researchers for years is how the physical properties of an airborne molecule determine the way it smells to the human brain. Machine learning offers a unique opportunity to explore this relationship by discerning patterns between molecular shapes and odor perceptions.
To tackle this challenge, Osmo’s CEO, 20877, and his team developed a machine-learning model. This model learned how to associate written descriptions of a molecule’s odor with its molecular structure, effectively creating groupings of similarly smelling odors. While computers have successfully digitized vision and hearing, this study represents a significant step forward in digitizing the sense of smell.
The model was trained using an industry dataset containing molecular structures and odor qualities of 5,000 known odorants. It learned to predict which odor words best described a molecule’s smell based on its shape. To validate the model’s effectiveness, Monell researchers conducted a blind validation procedure, where a panel of trained participants described new molecules, and their descriptions were compared to the model’s.
The results were remarkable, with the machine-learning model outperforming individual human panelists for 53% of the molecules tested. The model’s success extended to olfactory tasks it wasn’t explicitly trained for, such as predicting odor strength. This unexpected ability to make accurate predictions suggests the model’s potential in advancing our understanding of olfactory sensations.
Beyond its predictive capabilities, the model could offer a novel way to organize odors based on metabolism, a shift from the traditional chemical categorization. In this new framework, perceptually similar odors would also be metabolically related, potentially revealing fascinating insights into the connections between smell and nutrition.
In summary, this research brings us one step closer to demystifying the complex world of olfaction and offers exciting possibilities for digitizing and understanding the sense of smell in unprecedented ways.