A molecule’s shape does not translate directly into the words people use to describe its smell. MolOR adds an intermediate step: predict the pattern of olfactory receptors it activates, then use that pattern to help predict perception. A Cell Systems paper reports that the approach improved odour-description modelling even though its receptor-binding and percept datasets contained no overlapping molecules.[1]
Seyone Chithrananda, Judith Amores and Kevin K. Yang published the work on 2 September. Their model takes an odorant’s molecular graph and a receptor’s amino-acid sequence as inputs. By estimating interactions across receptors, it creates a computational activation profile that a second model can use alongside chemical features.[1]
From a molecule to a pattern of activated receptors
Smell is combinatorial: a molecule can activate several receptor types, and a receptor can respond to several molecules. MolOR encodes the molecule’s atoms and bonds with a graph neural network and the receptor with protein-language-model representations. Cross-attention connects the two sets of features to predict an interaction. The resulting profile is an estimate of receptor behaviour, rather than a direct measurement from a human nose.[1]
The research challenge is unusually severe because scientists have incomplete receptor–odorant binding measurements and separate databases of human smell descriptors. Simply training one model on an enormous combined table is not an option when the same molecules have not necessarily been observed in both settings. MolOR turns predicted receptor responses into auxiliary features for a percept-prediction system, supplying a biological intermediate where direct paired observations are missing.[1]
From chemical structure to a scent description
| Stage | Input | Output |
|---|---|---|
| Molecular representation | Atoms and bonds | Learned chemical features |
| Receptor interaction | Chemical features and protein sequences | Predicted receptor activation profile |
| Percept modelling | Chemical features and predicted activations | Predicted scent descriptors |
A bridge between datasets
The authors also report that structurally different molecules with similar perceived scents could produce similar predicted receptor profiles. That result gives the intermediate representation a useful role: it can connect molecules that do not look alike chemically but may engage the olfactory system in related ways.[1]
The model further differentiated functional olfactory receptors from pseudogenes in its predictions and could generate estimates for receptors not seen during training. These results suggest possible routes to investigate “orphan” receptors whose activating chemicals remain poorly characterised. They are predictions worth directing experiments towards, not evidence that every proposed chemical partner has been confirmed in a laboratory.[1]
Predictions that can guide receptor experiments
Real-world odours also depend on mixtures, concentration and individual variation. A useful follow-up would select previously unmeasured receptor–molecule pairs from MolOR’s predictions and test them independently. That would show whether an intermediate representation that helps predict descriptors also reliably identifies new biological interactions.[1]
The immediate value is a model that makes the route from chemistry to perception more explicit. Researchers can inspect the predicted receptor pattern, choose a proposed interaction to investigate and compare it with an experiment. That offers a more informative starting point than a scent label alone.[1]
Sources & further reading
- Mapping the combinatorial coding between olfactory receptors and perception with deep learning Cell Systems · 2 Sep 2026 · Accessed 26 Oct 2026