How AI predicts a smell from the structure of a molecule

Imagine receiving the structure of an unfamiliar molecule and being asked whether it might smell floral, woody or sulphurous. Atoms and bonds are the input; a predicted human description is the output. Molecular AI tries to learn the relationship between them.[5]

A molecule becomes a network

One representation treats atoms as nodes and bonds as connections in a graph. A message-passing neural network repeatedly combines information from neighbouring parts of that graph. It then produces a learned numerical representation of the molecule, which can be used to predict odour labels. The arrangement matters: a list of ingredients without their connections would discard part of the chemistry the model needs.[1]

The 2025 Dense Sense study examined whether electron-density information could enrich this approach. Its preprocessing used quantum-chemical calculations and a partition of electron density before the network learned from those features. The authors reported about forty days for one calculation stage on their computing cluster and another twenty for a subsequent stage in their stated setup. Those are dataset-preparation costs, not the time to describe one smell after training.[1]

A quick prediction can sit on top of a slow, expensive preparation process. The interesting engineering question is whether the representation contains information that improves the task enough to justify obtaining it—not simply whether the final interface responds quickly.

What the human comparison measured

A 2023 Science paper introduced a Principal Odor Map: a learned representation designed to preserve relationships between perceived smells. The researchers prospectively evaluated 400 novel odorants using a trained panel of fifteen people. The model’s predicted profile was closer to the panel’s mean description than the median individual panelist was. That is a meaningful comparison, but it is not “AI smells better than humans” in every sense.[2]

The comparison concerns agreement with a group’s description of odour quality. The paper was published in Science in 2023; a later repository-availability date does not make it a new 2025 discovery.[2]

Why a map helps more than a list of labels

Think of an original, simplified illustration with three imagined samples. Two are described as citrus-like, but one also has a woody character. A third is woody without the citrus impression. A representation retaining several dimensions can express those relationships more usefully than putting each sample into one exclusive box. The illustration is not a set of measured odours or a reconstruction of the paper’s numerical map.

The same reasoning explains why a map can help with a new task. If it preserves relationships relevant to perception, another model may use its coordinates as inputs rather than relearning everything from raw chemistry. Whether that works is an empirical question. It must be tested on the new task, not assumed because a visualisation looks intuitively organised.

A surprising connection with metabolism

A separate eLife analysis published in 2023 asked whether the representation related to biological chemistry. Compounds lying nearer in the odour map were more likely to occur together in a substance and to be closely related through metabolic processes. Sequences of metabolic reactions could trace smooth paths through the learned representation despite larger changes in conventional molecular structure.[3]

The authors interpret this as evidence that natural biological regularities help organise smell. That gives the map scientific interest beyond predicting product descriptions. It is still an interpretation supported by computational relationships, not proof that every aspect of olfaction follows one law or that a model has reconstructed the entire human sensory system.[3]

There is more than one human nose

Human perception also varies at the receptor level. A 2019 PNAS study sequenced olfactory-receptor genes in 332 people and examined perception measures including intensity and pleasantness for 68 odorants at two concentrations. The investigators validated ten cases in which receptor function measured in cells correlated with people’s perception. In eight, reduced receptor function was associated with lower perceived intensity.[4]

This is why a population-average prediction and a personal experience need not match exactly. An individual who finds a sample weaker or more pleasant is not automatically providing a bad label. Biology, the outcome being measured and the testing conditions can all matter. A model aimed at a panel average should be evaluated as that kind of model.[4]

A population-level odour map and individual receptor differences can therefore both be informative. One describes a regularity across people; the other helps explain why a particular person experiences the same molecule differently.

Report an error or suggest a correction