Scientists from the Monell Chemical Senses Center at the University of Pennsylvania, in collaboration with Osmo (a division of Google DeepMind), have created a computer model of smell that outperforms humans in odor recognition. This innovative artificial intelligence (AI) system has the potential to revolutionize the field of olfaction research and deepen our understanding of how the brain and nose work together.
The challenge of developing a computer-aided sense of smell
While machines have made significant progress in mimicking human senses such as sight and taste, developing computer-assisted olfaction has proven more challenging. Unlike vision, which relies on four receptors, and taste, which uses about 40 receptors, our sense of smell is mediated by 400 olfactory receptors. This complexity has hindered the development of artificial intelligence systems capable of accurately analyzing and describing smells.
Creating a master odor map (POM)
To address this gap, the research team set out to develop a neural network system capable of analyzing odor molecules and describing their smell in human language. This artificial intelligence system enabled the creation of the Principal Odor Map (POM), which provides insight into the relationship between the physical properties of air molecules and their perception by the brain.
Joel Mainland, senior co-author of the study from Monell, explains, “If a computer can determine the relationship between how molecules are shaped and how we perceive their odors, scientists can use this knowledge to further our understanding of how the brain and nose work together.”
Applications and implications
The implications of this research are far-reaching. With a better understanding of the relationship between molecular structure and odor perception, scientists will be able to develop more effective mosquito repellents, deodorizers, and perhaps even new fragrances. POM could become a valuable tool for chemists, olfactory neuroscientists and psychophysicists, allowing them to explore the nature of olfactory sensation in unprecedented ways.
Artificial intelligence system training and surprising results
To train the artificial intelligence system, the researchers provided it with the molecular structure of 5,000 odorants and descriptions of odors such as “minty” or “musty.” In addition, 15 experts had to identify 400 odors and describe them using a set of 55 words. In the tests, the artificial intelligence system performed slightly better than the participants in the odor identification experiment. However, the most remarkable thing was that the model successfully coped with olfactory tasks for which it was not specially trained, such as predicting the strength of an odor.
Mapping unsynthesized odor molecules
Based on these promising results, the researchers used the artificial intelligence system to map 500,000 odor molecules that have never been synthesized. This would have required 70 years of traditional olfactory methods for humans. The resulting data-driven map of human olfaction is a valuable resource for future research in chemistry, olfactory neuroscience and psychophysics.
The study, published in the journal EurekAlert, is a milestone in the field of olfactory research and paves the way for further progress in understanding how our sense of smell functions.
