In short
Models can now predict some odor properties of single molecules from their structure, sometimes at about the level of a trained panelist under the conditions tested. That is a real advance. It doesn't mean a program can design a fragrance, since these models learn from limited descriptors, mostly study single molecules at set doses, and can't tell you about safety, cost or how a mixture will develop. This lesson separates what has been shown from what gets claimed.
What you will learn
- Summarize what major prediction studies achieved
- List the limits of the data and the descriptor sets
- Say what a prediction could and couldn't do in practice
Before this lesson: Receptor Genetics and Individual Differences
Main studies
| Study | Data and aim | Result, in brief |
|---|---|---|
| Snitz et al., 2013 | Perceptual similarity from structure, including mixtures | A measure predicting odor similarity |
| Keller et al., 2017 | 49 people, 476 molecules, public prediction challenge | Models predicted some attributes, such as intensity and pleasantness, plus a number of semantic descriptors |
| Lee et al., 2023 | Graph neural network; tested on 400 out-of-sample odorants | The model's profile matched the panel mean more closely than the median panelist did |
| Ravia et al., 2020 | 49,788 similarity estimates from 199 people | A physicochemical measure; pairs of non-overlapping mixtures made to smell alike |
Together, these support a real claim: much of odor perception relates to measurable features of molecules. That's relevant to your earlier study of structure and odor, where rules kept finding exceptions.
Where the limits lie
- The labels. Models learn from people's words. "Woody" and "fresh" are broad, and panels differ by language, culture and training. A model's accuracy is partly its agreement with a panel's vocabulary.
- Training data. Published datasets often come from flavor and fragrance catalogs, so they contain well-known molecules with specific biases. Very new structures may behave differently.
- Purity. Trace impurities affect real samples, as in the enantiomer lesson. Models don't usually know about them.
- Isomers and stereochemistry. Depending on the model and its input, these may be handled poorly.
- Dose. A molecule can change character with concentration, and many datasets have a single dose.
- Individual differences. A panel mean hides the people you met in the last lesson.
- Mixtures and time. Perfume is mixtures changing over hours. Single-molecule prediction doesn't reach that far. See mixture perception.
- Everything else. Stability, synthesis, cost, safety, regulation. A prediction doesn't address these.
Reading a claim well
- What exactly was predicted? A descriptor, a rating, a similarity or something else?
- What was held out? Testing on data not used for training is essential. Were similar molecules in both?
- Compared with what baseline? A correlation means more beside a simple model, and beside human repeatability.
- Who funded or authored it? Commercial interest doesn't invalidate work, but it's something to note.
- Did anybody smell the predictions? Prospective testing means more than retrospective fit.
- What would it change for a perfumer? Be specific.
Where it could help
- Narrowing down candidates for synthesis at a fragrance house.
- Generating hypotheses about why a material has a particular facet.
- Suggesting descriptors as prompts, which you then check by nose.
Any candidate molecule still needs proper chemistry and safety evaluation before anyone smells it, and no independent learner should try to synthesize or smell an unassessed novel molecule on the basis of a prediction.
A note on what this does not change
A perfumer's training stays relevant. Even a very good descriptor prediction doesn't tell you how to bring a material in at the right moment, or whether a fragrance will suit a brief. But do stay open. This field is moving quickly. The best position is a dated, critical reading of the evidence, in place of either dismissal or awe. Your AI uses and limits lesson applies the same care to general tools.
Practical exercise: Evaluate one prediction claim
- Choose one paper from the table, or a commercial tool's public description.
- Answer the six questions above, quoting where you can.
- Choose one molecule you know well. Look up what a public dataset or tool says about it, if one is accessible, or use published descriptors. Compare with your own notes from smelling it.
- Write one page: what the work shows, what it doesn't, and what would have to be true for it to change your work.
You have done this well if you state the test conditions accurately, and you neither dismiss the result nor extend it past them.
Common mistakes
- Reading a molecule model as a perfume model. They address different questions.
- Treating a high correlation as understanding. Look at the test design.
- Smelling novel compounds out of curiosity. Unknown compounds are unknown hazards.
Knowledge check
Written and reviewed by the RUŌOD Lab team. This is general education, not legal, regulatory, medical or safety advice. Example dilutions and exercises are for learning, are not validated products, and have not been safety-assessed. Completing a lesson confers no qualification. How we check these lessons.