Researchers have long known that the chemical makeup of the molecules we inhale affects what we smell. But in most cases, no one can know exactly how. Scientists have deciphered some of the specific rules that govern how the nose and brain perceive an airborne molecule based on its properties. It’s becoming clear that we quickly identify some sulfur-containing compounds as the smell of garlic, for example, and some amines derived from ammonia as fishy odors. But these are exceptions.
It turns out that molecules that are not structurally related can have similar odors. For example, the larger ring-shaped hydrogen cyanide and benzaldehyde both smell like almond. Meanwhile, small structural changes – even changing the location of a single double bond – can dramatically alter aroma.
To understand this perplexing chemistry, researchers have turned to the computational power of artificial intelligence. Now one team has trained a type of AI known as a graph neural network to predict the scent of a compound in relation to humans — rose, medicinal, earthy, and so on — based on the chemical features of the scent molecules. computer model Evaluate new scents as reliably as humansthe researchers report in a draft of a new paper published in the preprint repository bioRxiv.
“I actually learned something fundamental about how the world smells and how smell works, which is amazing to me,” says Alex Welchko, now at venture capital firm GV subsidiary Google, who led the digital olfactory team while working at Google Research.
The average human does not have about 350 types of olfactory receptors, which can bind to a huge number of airborne particles. These receptors then initiate nerve signals that the brain later interprets as the smell of coffee, gasoline, or perfume. Although scientists know how this process works in a broad sense, many details — such as the exact shape of the odor receptors or how the system encodes these complex signals — are still far from them.
Stuart Ferstein, an olfactory neuroscientist at Columbia University, calls the model “a powerful work in computational biology.” But as is typical of a lot of machine learning-based studies, “it never, in my opinion, gives you much of a deep sense of how things work,” says Fierstein, who was not involved in the paper. His critique stems from an advantage inherent in technology: such neural networks are generally uninterpretable, which means that human researchers do not have access to the logic that the model uses to solve a problem.
Furthermore, this model bypasses the ambiguous workings of the nervous system, and instead makes direct connections between molecules and odors. However, Firestein et al. describe it as a useful tool with which to study the sense of smell and its charged relationship with chemistry. For the researchers involved, the model also represents a move toward a more accurate, numbers-based way of describing the world of smell, which they hope will eventually bring that meaning to the digital world.
“I strongly believe in a future in which you can hear, the same way computers can see, and they can smell,” says Wilchko, who is now exploring commercialization of this technology.
For a while, researchers have been using computer modeling to investigate the sense of smell. In a paper published in 2017, a crowdsourcing competition has produced a model that is capable of this Matching molecular structures with some labels— including “sweet,” “burnt,” and “flower” — which describe their odors in terms of what humans experience. In the new follow-up effort, Wilczko’s team trained their model with data from about 5,000 well-studied molecules, including the attributes of their atoms and the bonds between them. As a result, the model created a very complex “map” of smells. Unlike a traditional paper map, which plots locations in two dimensions, the model placed odor molecules into “locations” based on 256 dimensions—features the algorithm determined it could use to distinguish between molecules.
To see if this map corresponds to actual human perception, Wilchko’s team turned to Joel Mainland, an olfactory neuroscientist at the Monell Chemical Senses Center. “Defining success here is a bit tricky in that.” How do you define the smell of something? The mainland says. “Kyo [the fragrance] The industry — and what we’re doing here — is basically you get a group of people together, and they describe their scent.”
First, Mainland et al. identified a group of molecules with an odor that has not been documented. At least 15 trained study participants inhaled. Because perceptions of smell can vary greatly from person to person, thanks to genetic differences and personal experience and preferences, the researchers averaged participants’ ratings and compared this average to the model’s predictions. They found that for 53 percent of the molecules, the model came close to the plate average compared to the average individual team member — a performance they say surpassed the previous label-based model.
While the new model has demonstrated its ability to simulate a human’s perception of smell when given single molecules, it wouldn’t do as well in the real world. From roses to cigarette smoke, most scents are a mixture. Also, the team trained the new model using perfume data, which tends toward unpleasant odors and away from scent.
Even with these limitations, the model can still help those interested in smell chemistry, for example, by directing researchers who want to identify unstudied odors or test how alterations in the molecular structure alter perception. Perfumers can refer to it when refining fragrance formulations or identifying potential new ingredients.
Wiltschko’s team has already used the model to test a theory about the relationship between the structure of a chemical and how people and other creatures perceive its smell. In another paper published on bioRxiv in August, researchers suggested that an animal’s metabolism — the chemical processes that sustain its life, such as turning food into energy — could hold the explanation. From a database, they chose the metabolic compounds that are expected to excite odors and analyzed the molecules using an odor map model. The team concluded that molecules play closely related roles in metabolic reactions Tend to smell both, even if they differ in structure. Mainland, who was not a co-author of this separate preprint paper but consulted with the team on the project, describes his discovery as “really exciting.” “We don’t just build a model that solves some problems,” he says. “We’re trying to figure out the logic behind all of this.”
The model may also open the door to a new technology that records or produces specific scents on demand. Wilchko describes his team’s work as a step toward a “complete map” of human scent perception. The final version will be similar to “Color Space” Set by the International Commission on Illumination, which defines the visible colors. Unlike the new olfactory map, the color space is not dependent on words, notes Asifa Majid, a professor of cognitive sciences at Oxford University, who was not involved in the studies. Majid questions the use of language as a basis for mapping human perception. “Speakers of different languages have different ways of referring to the world, and the categories don’t always translate exactly,” she says. For example, English speakers often describe the aroma with reference to a possible source such as coffee or cinnamon. But in Jahai, an indigenous language spoken in Malaysia and parts of Thailand, one chooses from a vocabulary of 12 basic words for smell.
“We simply don’t know how this work can be extended to other languages,” Majid says, without empirical research to validate this. In theory, the researchers could identify smells without labels by measuring panel members’ reaction times when asked to compare smells: similar smells are hard to distinguish, so participants need more time to do so. However, according to the mainland, this behavioral approach has proven less realistic. Since the model has learned something fundamental about the organization of the scent world, it expects the map to be applicable elsewhere in the world.
Although it is possible to study human perception of smells without relying on words, researchers still lack the ability to represent these experiences in a crucial universal language: numbers. By developing olfactory equations for color space coordinates or hexagrams (which encode colors in terms of red, green, and blue), researchers aim to describe odors with new accuracy—and, perhaps, eventually, digitize them.
For vision and hearing, researchers learned which features the brain pays attention to, explains Michael Schmucker, who uses chemical informatics to study smell at the University of Hertfordshire in England and was not involved in the studies. As for the sense of smell, “there are many things that need to be resolved at the moment,” he says.
One of the main challenges is identifying primary odors. To create the olfactory equivalent of digital images, in which odors (such as scenes) are efficiently recorded and reconfigured, researchers need to identify a set of odor molecules that will reliably produce a series of odors when mixed – just like the red, green, and blue hues each screen generates.
“It’s science fiction that’s pretty far fetched right now, even though people are working on it,” Schmoker says.
