Floral-fresh or more sweet-rosy? Odor molecules in the air activate a variety of different receptors in our nose and thus create complex odor perceptions. With the help of machine learning, researchers have now succeeded in connecting the molecular structure of thousands of odorants with their perceptual properties. An artificial intelligence trained by them was able to describe unknown smells more precisely than trained human test sniffers based solely on the molecular structure of the fragrance.
For many of our senses, it is relatively easy to predict which external stimulus will lead to which perception. Light of a certain wavelength has a certain color and sound with a certain frequency is perceived as a certain pitch. Things are more complicated when it comes to the sense of smell: over 400 different olfactory receptors in our nose react to an incredible variety of chemical molecules. Even molecules that are structurally very similar can differ greatly in terms of smell. It was therefore not previously possible to reliably predict whether and how new substances smell.
From molecular structure to smell
A team led by Brian Lee from the Google Research Brain Team in Cambridge has now changed that. Using a training data set with 5,000 odorants described in detail, the researchers trained an AI model to associate the molecular structure with the odor properties and to describe them with words. In doing so, they created an olfactory map that makes the most complex of our senses predictable and machine-capable.
“The model fills in age-old gaps in the scientific understanding of olfaction,” says co-author Joel Mainland of the Monell Chemical Senses Center in Philadelphia. “In olfactory research, there has long been a question as to which physical properties make a molecule smell the way the brain perceives it. But if a computer can now see the relationship between the shape of molecules and the way we smell them, then science can use that knowledge to understand how our brain and nose work together.”
Better than human test smellers
To validate the model, the research team trained 15 human subjects to describe smells using a selection of 55 words, including floral, musty, musky, smoky, meaty and earthy. In addition, they were asked to indicate on a scale from one to five how pronounced the respective odor nuance was. The team then had both the AI and the human testers rate 400 new smells. The artificial intelligence received the molecular structure and the human sniffers received a purified sample of the respective substance.
The result: “The odor profile generated by the AI matched the mean of the human descriptions better than any human tester individually,” the authors report. For more than half of the molecules tested, the machine model performed better than the human sample sniffers. “The most surprising result, however, is that the model was also successful at olfactory tasks that it was not trained for,” says Mainland. “The trick was that we had never trained it to learn the strength of the odor, but it was still able to make accurate predictions.” smell similar and which, despite having an almost identical structure, differ greatly from one another in terms of smell.
Digitization of smells
“We hope this map will be useful to researchers in chemistry, olfactory neuroscience, and psychophysics as a new tool to explore the nature of olfaction,” said Mainland. Among other things, the predictions of the model could help to create new scents for the perfume industry, to create functional smells, for example to repel insects, or to produce new flavors. “Our approach enables odor predictions on a broad basis and paves the way for the digitization of odors,” write the researchers.
Source: Brian Lee (Google Research, Brain Team, Cambridge) et al., Science, doi: 10.1126/science.ade4401