A new AI can smell as well as a human

The compounds behind your partner’s terrible odours are unknowable to science.

But it turns out that it’s not just them. Mapping any compounds to the scents they make is harder than it looks. A team in the US has set out to solve the problem, by creating a ‘principal odour map’ to decode how our brains perceive scents.

“A fundamental problem in neuroscience is mapping the physical properties of a stimulus to perceptual characteristics,” the researchers write in their new paper, published in Science.

“In vision, wavelength maps to colour; in audition, frequency maps to pitch. By contrast, the mapping from chemical structures to olfactory percepts is poorly understood.”

The researchers set out to create a neural network that created graphs for each molecule. The team had access to a dataset of around 5000 molecules. Information about the molecules and the bonds holding molecules together were recorded down to the atomic level. Each molecule also had multiple odour labels attached to it – everything from beefy to lavender.

To test the model, the team then had to use the current best in smelling technology – humans.

“In olfaction, no reliable instrumental method of measuring odour perception exists, and trained human sensory panels are the gold standard for odour characterisation,” the researchers write.

“We trained a cohort of subjects to describe their perception of odorants using the rate-all-that-apply method (RATA) and a 55-word odour lexicon.”

Humans don’t always agree on what one smell smells like, and so the team had to use the mean and median to try and get an average of what the smell was. But the model did surpass the median panellist for over half of the labels. This did however depend significantly on the scent.

“Our model performs best for labels such as garlic and fishy that have clear structural determinants (sulphur-containing for garlic; amines for fishy) and worst for the label musk, which includes at least five distinct structural classes (macrocyclic, polycyclic, nitro, steroid-type, and straight chain),” the team wrote.

“By contrast, a panellist’s performance for a given label depended on their familiarity with the label in the context of smell; consequently, we observed strong panellist-panel agreement for labels that describe common food smells such as nutty, garlic, and cheesy and weak agreement for labels such as musk and hay.”

Excitingly, the model was found to be better than any other that came before and could describe the smells of the compound, but also the intensity and the similarity between different odours.

Using a model like this could explore new smells – the researchers have compiled a huge list of 500,000 potential compounds that haven’t been created. Using the principal odour map they could understand how they smell just from their atomic structure and molecular bonds.

“We hope this map will be useful to researchers in chemistry, olfactory neuroscience, and psychophysics,” the team conclude. “First, as a drop-in replacement for chemoinformatic descriptors and, more broadly, as a new tool for investigating the nature of olfactory sensation.”

Please login to favourite this article.