Could looking at a molecule’s structure tell us whether it smells floral or grassy? Molecules with similar shapes do not necessarily smell alike, and a small structural change can sometimes produce a very different impression. The relationship between molecules and scent is difficult to capture through a simple classification of shapes.
In a study published in 2023, Lee and colleagues sought to map this relationship using an AI trained on the words people use to describe smells. They called it the Principal Odor Map, or POM. A major feature of the study was that predictions for molecules not used in training were compared with newly collected human smell ratings.
Reading molecular structure in terms of scent
The training data consisted of approximately 5,000 molecules and 138 odor descriptors compiled from existing databases of fragrance and related materials. Each molecule was associated with multiple terms, such as fruity, sweet, or grassy.
The AI received molecular structures represented by atoms as points and bonds as lines. Using a graph neural network, it learned which features of atom types and their connections were useful for predicting scent.
POM is a representation of each molecule as 256 numbers, obtained during training. Rather than the two directions of a map on paper, it has many axes. The aim was to place molecules with similar smells near each other and capture relationships in which specific qualities such as jasmine and lavender sit within the broader floral family.
Compared with conventional representations that, for example, directly count molecular substructures, POM preserved groupings closer to human odor categories. Rather than simply organizing molecular shapes, it created coordinates tailored to their relationship with perceived scent.
Trained panelists smell new molecules
Predicting a hidden subset of existing data is not enough to establish usefulness for unfamiliar smells. The researchers therefore prepared 400 molecules not used in training and collected human evaluations independently of the AI predictions. The molecules were selected to cover a reasonably broad range of structures and odors and to include differences from the training examples.
The evaluators were trained to connect smells with words and met screening criteria in a preliminary test using 20 common odors. Each sample was rated twice by at least 15 people. They selected applicable terms from a set of 55 and rated how well those terms applied.
Samples were diluted to make their perceived intensities similar. Even so, 42 molecules with weak odors, 32 with confirmed or suspected impurity effects, and three duplicates or other exclusions were removed, leaving 323 molecules for prediction evaluation. The model was not retrained on these newly collected ratings before its performance was calculated.
What did human-level performance actually mean?
The reference was the panel’s average odor profile. The researchers compared how closely the AI prediction and each evaluator’s responses matched that average.
For 53% of the 323 molecules, the AI prediction was closer to the panel average than the median evaluator’s response. A conventional comparison model using the same training data achieved 41%. Looking at individual odor descriptors, the AI outperformed the median evaluator on 30 of the 55 terms.
This does not mean that the AI correctly identified 53% of smells. It was a comparison of how closely a profile of ratings across multiple terms corresponded to the panel average. It did not classify individuals’ perceptions as right or wrong, nor establish that AI surpassed humans in every olfactory ability.
Performance also varied across features. It was good for qualities such as garlic-like and fishy odors, whose relationships to structure are relatively easy to capture, but weaker for musk. Musks include several structurally distinct families; the authors suggest that the amount of training data and the chemical range grouped under a descriptor may matter.
The molecule on the bottle’s label may not be the one you smell
The study examined not only AI performance but also what the samples contained. For 50 samples, the researchers combined chemical analysis that separates constituents with a method in which people smell those separated constituents, to determine which molecules produced the odor.
They found cases where the intended molecule was the main source of the smell, as well as cases where trace impurities played a substantial role. In one example, a sample’s buttery impression was attributed to the impurity diacetyl rather than to the nominal molecule that had been purchased.
These 50 samples were not selected at random: their selection took account of patterns of agreement among people and with the AI. The proportion of impurities found in this group therefore cannot be treated as representative of fragrance materials in general. The key issue was to align the molecule supplied to the AI as input with the molecule people were actually smelling.
Can the map do more than predict odor words?
The authors also examined challenging sets of three molecules in which the more structurally similar pair actually smelled less alike. In a test using 41 selected triplets, the AI captured this reversal in half of the cases, compared with 19% for the comparison model. This suggests scope for prediction without assuming that similar shapes must smell alike.
They also applied relatively simple calculations to the map’s coordinates to predict, in other public datasets, how well odor descriptors applied, the concentration thresholds at which odors could be detected, and how similar odors were to one another. Here too, performance was as good as or better than that of comparison models using chemical-structure features alone.
A major aim was to obtain a representation useful across multiple olfactory tasks, rather than merely learn the answers to one task. The researchers also mapped roughly 500,000 candidate molecules likely to have odors. Many had not been synthesized or evaluated, however; people had not verified half a million different smells.
It does not yet tell us how to blend a perfume
POM is a map trained mainly on the odors of individual molecules. It does not provide clear rules for the new impressions that will result from mixing particular ingredients. Nor does its structure directly reproduce the workings of the brain or nose.
Odor quality can change with concentration, but concentration is not explicitly included in this map. Being able to predict a molecule’s detection threshold does not directly tell us how intense people will find it at a particular concentration. Individual differences and structures outside the training domain also remain limitations.
The authors identify the need to update the map with higher-quality evaluation data and eventually develop a representation that can predict perception of mixtures. Even at this stage, they anticipate value in shared coordinates that chemists can use to search for candidates and neuroscientists can use to compare relationships between odors and brain responses.
The research was funded by Google Research and other sources. Some authors disclosed that they worked at Google during the study, moved during peer review to Osmo, which is commercializing related technologies, and held ownership interests in and received salaries from that company. They also disclosed the transfer of related intellectual property.
Summary
An AI trained on molecular structure predicted the odor descriptors people used on average for new molecules at a level close to a typical individual evaluator. Rather than making human evaluation unnecessary, the study used it to build a map organizing relationships among smells.
Importantly, this map was useful not only for odor descriptions but also for similarity and detection thresholds. While mixtures, concentration, and individual differences remain challenges, it offers a common tool for thinking about the molecular world and the world of perceived scent.
References
Lee, B. K., Mayhew, E. J., Sanchez-Lengeling, B., Wei, J. N., Qian, W. W., et al. (2023). A principal odor map unifies diverse tasks in olfactory perception. Science, 381(6661), 999–1006. DOI: 10.1126/science.ade4401 / Full text (accepted author manuscript).
