Computer programs’ odor predictions are on the nose.

Image Source/Alamy Stock Photo

Artificial intelligence grows a nose

Predicting color is easy: Shine a light with a wavelength of 510 nanometers, and most people will say it looks green. Yet figuring out exactly how a particular molecule will smell is much tougher. Now, 22 teams of computer scientists have unveiled a set of algorithms able to predict the odor of different molecules based on their chemical structure. It remains to be seen how broadly useful such programs will be, but one hope is that such algorithms may help fragrancemakers and food producers design new odorants with precisely tailored scents.

This latest smell prediction effort began with a recent study by olfactory researcher Leslie Vosshall and colleagues at The Rockefeller University in New York City, in which 49 volunteers rated the smell of 476 vials of pure odorants. For each one, the volunteers labeled the smell with one of 19 descriptors, including “fish,” “garlic,” “sweet,” or “burnt.” They also rated each odor’s pleasantness and intensity, creating a massive database of more than 1 million data points for all the odorant molecules in their study.

When computational biologist Pablo Meyer learned of the Rockefeller study 2 years ago, he saw an opportunity to test whether computer scientists could use it to predict how people would assess smells. Besides working at IBM’s Thomas J. Watson Research Center in Yorktown Heights, New York, Meyer heads something called the DREAM challenges, contests that ask teams of computer scientists to solve outstanding biomedical problems, such as predicting the outcome of prostate cancer treatment based on clinical variables or detecting breast cancer from mammogram data. “I knew from graduate school that olfaction was still one of the big unknowns,” Meyer says. Even though researchers have discovered some 400 separate odor receptors in humans, he adds, just how they work together to distinguish different smells remains largely a mystery.

In 2015, Meyer and his colleagues set up the DREAM Olfaction Prediction Challenge. They divided the Rockefeller group’s data set into three parts. Participants were given the volunteer ratings for two-thirds of the odors, along with the chemical structure of the molecules that produced them. They were also given more than 4800 descriptors for each molecule, such as the atoms included, their arrangement, and geometry, which constituted a separate set of more than 2 million data points. These data were then used to train their computer models in predicting smells from chemical structural information. The remaining groups of data—two sets of 69 ratings and their corresponding chemical information—were used to test how well the models predicted both how an average person would rate an odor and how each of the 49 individuals would rate them. 

Twenty-two teams from around the globe took up the challenge. Many did well, but two stood out. A team led by Yuanfang Guan, a computer scientist at the University of Michigan in Ann Arbor, scored best at predicting how individual subjects rate smells. Another team led by Richard Gerkin at Arizona State University in Tempe best predicted how all the participants on average would rate smells, Meyer and his colleagues report today in Science.

“We learned that we can very specifically assign structural features to descriptions of the odor,” Meyer says. For example, molecules with sulfur groups tend to produce a “garlicky” smell, and molecules with a similar chemical structure to vanillin, from vanilla beans, predicts whether subjects will perceive a “bakery” smell.

Meyer suggests such models may help fragrance and flavor companies come up with new molecules tuned to trigger particular smells, such as sandalwood or citrus. But Avery Gilbert, a biological psychologist at Synesthetics in Fort Collins, Colorado, and a longtime veteran of the fragrance and flavor industry, says he’s not so sure. Gilbert says the new work is useful in that it provides such a large data set. But the 19 different verbal descriptors of different scents, he says, is too limited. “That’s really a slim number of attributes,” he says. Alternative studies have had volunteers use 80 or more categories to rate different smells.

The upshot is that even though the current study showed computers can predict which of 19 words people will use to describe this set of odors, it’s not clear whether the same artificial intelligence programs would rise to the challenge if there were more categories. “If you had different descriptors, you might have had different models predict them best. So I’m not sure where that leaves us,” Gilbert says. Perhaps it serves mostly as a reminder that odor perception remains a challenge both for human scientists and artificial intelligence.