The Percept Is Not in the Molecule

ai blog
July 19, 2026

Joel Kowalewski, PhD

Lucretius and the hooked atoms

Around the middle of the first century before our era, the Roman poet Lucretius set out to explain the whole of nature in verse, and when he came to the senses he gave an account that was, for its time, a triumph of materialism. Following the Greek atomists Democritus and Epicurus, he proposed in De Rerum Natura that everything we perceive is the work of tiny bodies streaming off the surfaces of things and striking the “apertures” of our senses (or what we might now call sensory organs). Sensation, on this view, is a matter of shape. Smooth, round particles slip gently into the passages of the body and produce sweetness and pleasant scents; rough, jagged, hooked particles tear at those passages and produce bitterness and stench. A pleasant smell simply was a stream of smooth atoms. It is the first real theory of olfaction in the Western record, and it did something that we have been doing ever since: assign perceptual qualities entirely to the properties of the things being perceived. The odor was a property of the particle, carried in its shape, waiting in the object to be delivered to the nose. Cause and effect, mechanical and precise in a manner that would ultimately be consistent with the work of Isaac Newton much later, in descriptions of energy of one thing being imparted to another thing.

It is highly intuitive. We say a rose smells sweet as though the sweetness were sitting in the rose alongside its mass, both objective properties, inside, gathered and waiting for an instrument to measure them. Twenty-one centuries after Lucretius we have incomparably better chemistry and the very same instinct: identify the molecule, and you have identified the smell; match structure to percept, and the code is cracked. I have spent a large part of my scientific life inside exactly that project, building machine-learning models that try to predict how a molecule will smell from what the molecule is. And it is from inside the project, not outside it, that I have come to think the ancient intuition is wrong in a precise and interesting way. The smell is not in the molecule. It is performed in a relationship between a molecule and a particular kind of body, and this is why the molecule by itself will never tell you the smell — and why a model that predicts smell, even an excellent one, is a map of that relationship and never the scent itself.

The molecule keeps its secret

The cleanest way to watch the smell slip out of the molecule is to hold two molecules that are, by almost every physical measure, the same. Carvone comes in two forms that are mirror images of each other, related as your left hand is to your right. They share a chemical formula, the same bonds, the same mass, the same boiling point, very nearly every scalar property a chemist could list. And yet one of them, (S)-carvone, smells unmistakably of caraway, and its mirror, (R)-carvone, smells of spearmint. Limonene is additional example: one smells of oranges, the other of lemons and pine. If odor were an essential property that a molecule carried, two molecules identical in every internal measure should be decoded similarly. They do not. What distinguishes them is nothing you can find in either molecule considered individually; it is the way the mirror opposites interact, or fail to interact, with sensory receptors in the nose. The difference lives in the fit — in the relationship — and not in the compound.

This is why predicting odor from chemical structure alone has been a challenging problem. In 2017 a large collaborative contest, the DREAM Olfaction Prediction Challenge, invited teams from around the world to predict how panels of people would describe a molecule’s smell given only its physical structure. The results, published in Science, were genuinely good but also limited: structure predicts perception well above chance and nowhere near perfectly, and some perceptual qualities resist prediction almost entirely. More recent systems have improved on this dramatically — a 2023 paper in Science describing a “principal odor map” showed a neural network learning could predict odor descriptors with something close to human reliability — but the “map” was an embedding in which molecules were placed near or far from one another according to human raters. It was a simulation of olfactory perceptual coding; the relationships could not be inferred, bottom up, from the chemical structure alone.

My own work made the same discovery from the other direction, demonstrating that a molecule’s perceived smell is predicted markedly better from the pattern of receptors it activates than from its chemical structure, its atoms and physicochemical features. While clearly these features are relevant, they are just indirect and noisy measurements of a receptor code. What is captured more directly in an olfactory description is a web of relationships between molecules and odorant receptors. In 1991 Linda Buck and Richard Axel discovered the enormous family of genes that encode odorant receptors, work that won them the Nobel Prize in 2004; humans carry roughly four hundred functional ones. In 1999 Buck and her colleague Bettina Malnic showed how they are used: not one receptor per smell, but a combinatorial code, in which each odorant switches on a particular subset of receptors and each receptor answers to many odorants, so that the identity of a smell is a pattern spread across the whole ensemble. There is no specialized “rose” receptor. “Rose” is a pattern of activation across hundreds of separate relationships given a name later on in the processing stream.

The same molecule, a different world

If the smell were really in the molecule, then olfactory experience or descriptions would be common. Consider androstenone, a steroid found in sweat. In 2007, a team led by researchers at Rockefeller and Duke reported in Nature that a difference in a single odorant-receptor gene, OR7D4, governs how a person experiences it: depending on which version of this receptor you carry, the identical molecule smells like stale urine, or sweet and faintly floral, or of nothing whatsoever. One compound, three different experiences. The same story explains why, for a sizable minority of people, fresh cilantro tastes of soap — a quirk traced in part to a variant in an olfactory-receptor gene. These are not mistakes, errors to be corrected against the true odor printed on the structure. There is no true odor printed on the structure. It is constructed and deeply relational. When the relationships change, the quality changes with it.

I have argued elsewhere that a fact is not an object we discover sitting in the world but a belief that a community of similarly situated observers is compelled to share. Odor qualities behave exactly this way. That coffee “smells like coffee” is a fact in precisely the sense that it is maintained, reliably and in common, by the enormous population of people whose noses are built closely enough alike to have approximately similar experiences. I say “approximately” because complex systems like living organisms and brains in particular tend to reject the term “identical.” Not even identical twins are truly identical. “Similar” is slight deviation around a statistical average. Where the bodies diverge — a different receptor, a different genome — the “fact” of the smell diverges too, not because someone is in error but because the relationship that constitutes the fact has different terms. The objectivity of smell, such as it is, is the agreement of many like instruments, and not the readout of a property.

Chemical mixtures further expose brain processing as relational and computationally irreducible . Smell is a synthetic sense, not an analytic one: combine a few dozen odorants and you do not perceive a list of ingredients but a single thing — “jasmine,” “coffee,” “the sea.” Push this far enough and something remarkable happens. Take many complex mixtures, each made of numerous different odorants, and as the mixtures grow in complexity they begin to smell alike, converging on a single hard-to-place percept that researchers named “olfactory white” in a 2012 study, by analogy with white light and white noise. Two such mixtures can share not one ingredient while converging on a similar perceptual experience. The perception of the mixture is no more contained in the mixture than the smell was contained in the molecule; it is mathematical operations performed by a system that integrates and suppresses and configures its inputs into a whole dynamically. Perception in the dance.

The strongest case for the molecule

First floated by Malcolm Dyson in the 1930s and revived by the biophysicist Luca Turin in 1996, the vibrational theory of odor coding proposed that the nose is a kind of biological spectroscope, reading not the shape of a molecule but the quantum vibrations of its bonds — so that the smell really would be a physical property carried in the molecule, in it vibrational frequencies. It is a beautiful theory, and part of its appeal is precisely that it puts the odor back where intuition wants it, inside the thing. Our brains gravitate toward essentialist descriptions because a property “inside” is simply more intuitive than a web of relationships. But a theory is answerable to evidence, and when Andreas Keller and Leslie Vosshall tested the vibration account in 2004, its central predictions failed: perception did not change with vibrational frequency. The vibrations are real. They are just not essential, nor code-like.

Maybe my own models keep getting better at predicting odor from molecular structure, I might then say, is that not also proof that the smell is in the molecule after all? While structure constrains perception powerfully, the molecule is a genuine and necessary term in the relationship, and the models work at all only because the relationship is lawful rather than random, but the model does not extract the odor from the atoms. The frontier models learn an embedding — a map of how molecules stand in relation to one another, to receptors, and to the words people reach for — and it learns that map from human judgments, or from receptor responses, which is to say it learns because of a relational encoding that is not in the chemical itself. The gains of the last decade are from modeling the relationships with higher fidelity. When a network correctly predicts that (R)-carvone will be called “spearmint,” it has captured a lawful regularity about how one particular molecule meets one particular evolved ensemble of receptors and one particular community of speakers. That is a real and considerable achievement.

To describe the smell is not to smell it

I have argued in other essays that the physical world is relational rather than a collection of discrete objects each holding eternal, essential properties — and to represent a relationship is not to perform or enact it; the properties of things or objects like odorant molecules, although not irrelevant, disclose the potential to interact and form relationships over time. Until the relationships are actual, the possibilities are seemingly endless. Smell is the sense that makes this hardest to deny, because it so plainly refuses to stay put in the molecule. A predictive model of odor — mine, or anyone’s — is a description of the relationship among a chemistry, a set of receptors, and human language. It can be accurate; it can be useful; it can be, in its way, beautiful. But it is a map of the relationships. And it is difficult, if not impossible, to capture everything of relevance in a map. You need the molecule and the receptor it fits and the sensory neuron carried past its threshold and the cortex that gathers the pattern and assigns the name — the entire relation, performed, in a living body, which is constantly changing and rejects simple descriptions admired by the public and pursued by scientists.

None of this is mysticism; the argument is easy to mistake for a retreat into the ineffable. Olfactory perception is physical. The molecule is necessary and it is not sufficient. The body is necessary and it is not sufficient. Perception is the relationship between them, and it exists only while they are in contact. I leave open, as one should, that we might one day simulate the whole relationship finely enough to reconstitute a percept — but to do it, you would have to build the body and the whole organism too, as it is unclear why simple abstractions, simplified sets of relationships that convey essential qualities of the whole, would not be the rule in nature. We must ask why would the simpler configuration not be the goal of evolution, as if it were left to the scientist alone to discover? If evolution is mathematical optimization, the biosphere is the mathematically irreducible result; in general, the simplest, most elegant configuration that still retains the relationships that are necessary for sustaining life. Of course, this does not exclude explanation. The scientist, after all, abstracts or simplifies to understand. Not to reconstruct physical reality in the laboratory. It is only problematic if the very thing that drives our interest requires what the laboratory cannot contain. Life itself or the universe itself.

Lucretius was not being trivial, nor primitive. He was early example, doing as we still do. Twenty-one centuries and one Nobel Prize later, we know that the shape of the molecule matters enormously but for reasons the molecule hides. The odor was, the whole time, a complex web of relationships, which scientists steadily revealed, if only in fragments, over time. When I train an AI model to predict a smell, I am charting these fragments. I know that life is in the encounter, and that science is not well equipped to express or contain in this phenomenon. But quite unlike the popular belief that science is about rational pursuits, the boundaries of science are pushed back because we quite irrationally believe they do not and must not exist. And so, we move forward, accident-by-accident, relationship-to-relationship.