The smell in the laboratory was new. It was, in the language of the business, tenacious: for more than a week, the odour clung to the paper on which it had been blotted.
To researcher Alex Wiltschko, it was the smell of summertime in Texas: watermelon, but more precisely, the boundary where the red flesh transitions into white rind.
“It was a molecule that nobody had ever seen before,” says Wiltschko, who runs a company called Osmo, based in Cambridge, Massachusetts. His team created the compound, called 533, as part of its mission to understand and digitize smell. His goal — to develop a system that can detect, predict or create odours — is a tall order, as molecule 533 shows. “If you looked at the structure, you would never have guessed that it smelled this way.”
That’s one of the problems with understanding smell: the chemical structure of a molecule tells you almost nothing about its odour. Two chemicals with very similar structures can smell wildly different; and two wildly different chemical structures can produce an almost identical odour. And most smells — coffee, Camembert, ripe tomatoes — are mixtures of many tens or hundreds of aroma molecules, intensifying the challenge of understanding how chemistry gives rise to olfactory experience.
Another problem is working out how smells relate to each other. With vision, the spectrum is a simple colour palette: red, green, blue and all their swirling intermediates. Sounds have a frequency and a volume, but for smell there are no obvious parameters. Where does an odour identifiable as ‘frost’ sit in relation to ‘sauna’? It’s a real challenge to make predictions about smell, says Joel Mainland, a neuroscientist at the Monell Chemical Senses Center, an independent research institute in Philadelphia, Pennsylvania.
Animals, including humans, have evolved a remarkably complex decoding system befitting the enormous repertoire of odour molecules. All sensory information is processed by receptors, and odour is no different — except in its scale. For light, the human eye has two types of receptor cell; for smell, there are 400. How the signals from these receptors combine to trigger a particular perception is unclear. Plus, the receptor proteins themselves are hard to work with, so what they look like and how they function has mostly been guesswork.
Things are beginning to change, however, thanks to improvements in structural biology, data analytics and artificial intelligence (AI). Many scientists hope that cracking the olfactory code will help them to understand how animals use this essential sense to find food or mates, and how it feeds into memory, emotion, stress, appetite and more.
Others are trying to digitize smell to build new technologies: devices that diagnose disease on the basis of odours; better, safer insect repellents; and affordable or more-effective aroma molecules for the US$30-billion flavour and fragrance market. At least 20 start-up firms are trying to make electronic noses for applications in health and public safety.
This all adds up to a surge of research into the biology of olfaction, says Sandeep Robert Datta, a neuroscientist at Harvard Medical School in Boston, Massachusetts. “Smell is having a moment,” he says.
Even for experts, the physical properties of an odour molecule typically offer little insight into how it will actually smell.
Researchers have come up with a few computational models that can relate structure to odour, but early versions tended to be based on quite narrow data sets or could only make predictions when smells had been calibrated to have the same perceived intensity. In 2020, one team reported a model that could predict how similar real-world mixtures were to each other, correctly identifying that rose and violet odorants are more similar to one another than either is to the pungent spice asafoetida, often used in Indian cuisine1.
Previous attempts to use machine learning were good, but not great. For example, when researchers ran a competition to create the best odour-predicting model, algorithms from 22 teams could effectively predict only 8 out of 19 smell descriptors2.
Last year, Wiltschko’s team — then part of Google’s AI research division — collaborated with researchers at Monell, including Mainland, to publish a map for smell3 that made use of AI.
Their program was trained by feeding the model thousands of descriptions of molecular structures from fragrance catalogues, along with smell labels for each — terms such as ‘beefy’ or ‘floral’.
Then, the researchers compared the AI system with human noses. They trained 15 panellists to rate a few hundred aromas using 55 labels, such as ‘smoky’, ‘tropical’ and ‘waxy’.
Humans have a hard time with this task because smell is so subjective. “There’s no universal truth,” says Mainland. Most smell descriptions lack detail, too. For one smell, panellists chose the words ‘sharp, sweet, roasted, buttery’. A master perfumer, asked to describe the same smell, noted ‘ski lodge, fireplace without a fire’. “That shows you the gap,” says Mainland. “Our lexicon is not good enough.” Nonetheless, a human panel is one of the best available tools for coming up with consistent smell descriptors because the average rankings of the group for different smells tend to be stable.
Using the structure of these molecules alone, the AI algorithm did well at predicting the smell of compounds compared with the average group assessments (see ‘Same but different’), and it performed better than the typical individual sniffer. And although the map it produced was very complicated — it has more than 250 dimensions — it was able to group smells by type, such as meaty, alcoholic or woody.
