x
Neural network can provide descriptions, such as ‘grassy’, for a wide variety of molecules, including some that don’t exist in nature.
Humans who were taught to describe specific odours often did so less precisely than a newly developed artificial-intelligence tool.Credit: Andia/ Universal Images Group via Getty
An artificial-intelligence system can describe how compounds smell simply by analysing their molecular structures — and its descriptions are often similar to those of trained human sniffers.
The researchers who designed the system used it to list odours, such as ‘fruity’ or ‘grassy’, that correspond to hundreds of chemical structures. This odorous guidebook could help researchers to design new synthetic scents and might provide insights into how the human brain interprets smell.
The research is reported today in Science1.
Smells are the only type of sensory information that goes directly from the sensory organ — the nose, in this case — to the brain’s memory and emotional centres; the other kinds of sensory input first pass through other brain regions. This direct route explains why scents can evoke specific, intense memories.
“There’s something special about smell,” says neurobiologist Alexander Wiltschko. His start-up company, Osmo in Cambridge, Massachusetts, is a spin-off from Google Research that is trying to design new smelly molecules, or odorants.
‘The entire protein universe’: AI predicts shape of nearly every known protein
To explore the association between a chemical’s structure and its odour, Wiltschko and his team at Osmo designed a type of artificial intelligence (AI) system called a neural network that can assign one or more of 55 descriptive words, such as fishy or winey, to an odorant. The team directed the AI to describe the aroma of roughly 5,000 odorants. The AI also analysed each odorant’s chemical structure to determine the relationship between structure and aroma.
The system identified around 250 correlations between specific patterns in a chemical’s structure with a particular smell. The researchers combined these correlations into a principal odour map (POM) that the AI could consult when asked to predict a new molecule’s scent.
To test the POM against human noses, the researchers trained 15 volunteers to associate specific smells with the same set of descriptive words used by the AI. Next, the authors collected hundreds of odorants that don’t exist in nature but are familiar enough for people to describe. They asked the human volunteers to describe 323 of them and asked the AI to predict each new molecule’s scent on the basis of its chemical structure. The AI’s guess tended to be very close to the average response given by the humans — often closer than any individual’s guess.
“It’s a nice advance using machine learning,” says Stuart Firestein, a neuroscientist at Columbia University in New York City. He says that the POM could be a useful reference tool in the food and cleaning-product industries, for example.
But Firestein points out that the POM doesn’t reveal much about the biology behind the human sense of smell — how different molecules interact with the approximately 350 odour receptors in the human nose, for instance. “They’ve got the chemical side and the brain side, but we don’t know anything about the middle yet,” he says.
How do we smell? First 3D structure of human odour receptor offers clues
Pablo Meyer, a systems biologist at the IBM Center for Computational Health in Yorktown Heights, New York, praises the paper’s use of language to link structures with subjective smells. But he disagrees that the average of the humans’ answers is the “correct” way to describe a smell. “Smell is something personal,” he says. “I don’t think there’s a correct perception of something.”
The next step, Wiltschko says, is to find out how odorants combine and compete with one another to create what the human brain interprets as a smell that is entirely different from those of each of the individual odorants. Meyer and Firestein say this will be very difficult: mixing just 100 molecules in different combinations of 10 produces 17 trillion variations, and the number of possible combinations quickly becomes far too many for a computer to analyse.
