When Jonathan Deutsch agreed to sniff 400 vials of unlabeled liquid for science, he didn’t know he would be competing with a computer. A research chef who helps with food product development at Drexel University, he simply welcomed the chance to hone his sense of smell. But odor profiles generated by Deutsch and 13 other volunteers served as a test for a computer program that had been trained to produce these same types of descriptions—such as fruity, cooling, fishy, piny—using chemical structure alone.
The results, reported today in Science, show that the program, a so-called graph neural network, is excellent at imitating human sniffers, at least when it comes to simple odors. It reliably predicted what the volunteers smelled, a feat sensory biologists have been working toward for decades. It also predicted the smells of 500,000 other molecules, with no need to make or sniff them.
The result is a boon for the study of olfaction, a field that has “floundered around for years looking for this information,” says Stuart Firestein, a neuroscientist at Columbia University who was not involved in the work. “The approach offers great potential” to speed up the search for better smelling consumer products, adds Andreas Grasskamp, a neurobiologist who studies perception at the Fraunhofer Institute for Process Engineering and Packaging.
The findings may also help establish olfactory research as a field on par with sight or vision. Smell, which in humans involves a smaller proportion of the brain and fewer types of receptor cells than in other mammals, was long considered “a primitive sense that wasn’t worth studying by neurobiologists,” says Pablo Meyer Rojas, a physicist at IBM Research. It has also defied systematic study. Whereas what we see and hear reflects quantifiable properties such as wavelength and frequency, smell doesn’t neatly correspond with the shape of a molecule. Similarly structured molecules can smell different, whereas dissimilar molecules can produce the same odor.
Ten years ago, a crowdsourced competition challenged researchers to use artificial intelligence (AI) to predict an odor’s smell from its structure. Asked to “smell” 69 chemicals, the winning algorithms could pick out eight of 19 possible odors that people had attributed to these samples. But it couldn’t group the samples according to how similar they smelled, says Joel Mainland, a neuroscientist at the Monell Chemical Senses Center who helped organize the challenge.
Rick Gerkin, a neuroscientist at the AI company Osmo, was one of the winners. Because the number of well-characterized odor molecules for training AI “sniffers” has greatly expanded since the competition, he thought he could do better. With Alexander Wiltschko, first at Google Research and now at Osmo, Gerkin and colleagues input the structures and odor descriptions of 5000 molecules into a more sophisticated AI. It learned to recognize patterns in the training data, correlating a molecule’s smell with features of its constituent atoms—their identities, sizes, and connecting bonds.
Then came the testing, led by Mainland. Deutsch and his fellow volunteers signed onto Zoom sessions with Monell sensory biologist Emily Mayhew to sniff the 400 mystery vials and report which of 55 odors they detected in each. Mayhew, now at Michigan State University, notes that perceived smells vary greatly between people. So the team calculated average human ratings to compare with the network’s predictions. In more than half of cases, the neural network got even closer to this average than any individual in the group did. “This is a difficult accomplishment,” Grasskamp says.
Next the team generated 500,000 hypothetical chemical structures. The network quickly inferred how they should smell, providing a database that should help in the search for odors for new foods, perfumes, cleaners, and other products.
Although the neural network shows it’s possible to map a chemical’s structure to an odor, it sheds little light on smell’s underlying biology. From a basic research perspective, “it’s not clear whether this is an important development rather than an incremental one,” says Linda Buck, a neuroscientist at the Fred Hutchinson Cancer Center. She adds that the neural network still hasn’t proved it can evaluate mixtures of molecules—the kinds of complex odors we encounter in the real world.
“The frontier is now about mixtures,” Gerkin agrees. That’s what he hopes to teach his algorithms to do next.
