Can Molecular Connectivity Predict Odor Words?

Can words such as ‘floral,’ ‘woody,’ and ‘musk’ be predicted simply by showing a molecule’s structure? In 2025, Jinyoung Suh, Yeonju Hong, and Chunho Park used 8,681 molecules and existing odor descriptions to compare ways of representing molecules for machine-learning predictions.

Rather than merely presenting one high-performing model, the study aimed to identify which information about molecular substructures and physicochemical properties helps classify odors. The researchers also explored mapping the relationships between odor words that the models had learned.

A study of collected descriptions, not a new experiment with human volunteers

The researchers combined descriptions from ten sources, including fragrance reference collections and databases. The dataset covered 8,681 molecules; this does not mean that 8,681 people newly smelled and rated them.

Even the same odor may be described using different spellings or expressions in different sources. The team therefore prioritized terminology from the International Fragrance Association (IFRA) and, with help from fragrance experts, standardized typographical errors and variations in wording. There were 200 candidate descriptors, of which 100 had at least 30 records.

Each molecule was not assigned just one word. The task was to predict multiple words at once, as when the same molecule is both ‘floral’ and ‘spicy.’

Three ways to represent a molecule

The first method considered which functional groups a molecule contains. Functional groups are groups of atoms, such as an aldehyde group, that contribute to chemical properties. The presence or absence of predefined substructures was encoded numerically.

The second method used whole-molecule characteristics such as molecular weight, affinity for oily environments, and the size of the surface associated with polarity. This representation summarizes physicochemical properties rather than detailed connectivity.

The third method used Morgan fingerprints, which record the atoms around each atom and how they are connected as small substructures. These are not human fingerprints: the name likens a numerical code of molecular features to a fingerprint.

This code primarily represents two-dimensional bond connectivity. Although preparing the input involves calculating three-dimensional structures, the method does not directly simulate in three dimensions how a molecule fits into an odor receptor.

Nine combinations with three learning methods

The researchers combined each molecular representation with three machine-learning methods, comparing nine combinations in total. All were decision-tree-based methods that combine many decisions splitting the data according to its features.

The data were divided into training and test sets, with fivefold cross-validation also performed on the training side. This involves alternating which subsets are used for learning and evaluation, to check whether good performance depends on one particular split.

In terms of average performance, Morgan fingerprints combined with the learning method XGBoost offered the best balance. The result favored a representation of detailed connectivity around atoms over one limited to preselected functional groups.

Even at 97.8%, many labels were missed

In the test results for this combination, accuracy across both the presence and absence of labels was 97.8%. However, most words do not apply to most molecules. When many of the answers are ‘not present,’ overall accuracy can readily become high.

Other metrics therefore matter. Precision, the proportion of predicted positive labels that were correct, was 41.9%. Recall, the proportion of actual positive labels successfully retrieved, was 16.3%.

The figure of 97.8% therefore cannot be read as ‘almost perfectly identifying the odor of unfamiliar molecules.’ Some predicted odor words were wrong, and the model also failed to retrieve many words present in the existing descriptions.

The paper also compared performance using several metrics as the criterion for accepting a prediction changed. It emphasized the need to examine performance on sparsely recorded odor words rather than rely solely on simple accuracy.

The same method is not necessarily best for every odor

Structural encoding had an advantage in the overall average, but the results differed by word. For citrus and ‘odorless,’ there were cases in which physicochemical descriptors performed well, and this information was also useful for musk.

The researchers also examined the importance of features used by each model in making its decisions. Molecular weight, affinity for oily environments, and the size of the surface associated with polarity helped classification.

The authors discuss how these properties might relate to volatility, passage through nasal mucus, and interactions with receptors. However, the study did not measure receptor binding or persistence on skin. It did not establish that features useful for prediction are themselves the causes of an odor.

Checks with familiar molecules and predictions for new candidates

The researchers also tried combining the three structure-based models and selecting high-scoring word–model pairs. Predictions matched known characteristics, including lavender and floral for linalool and rose for phenethyl alcohol.

However, the three familiar fragrance materials used as examples here were molecules included in the training data. This was a check that the learned information agreed with familiar examples, not an external validation on molecules the models had never encountered.

By contrast, virtually odorless squalane, which was not included in the dataset, received a high prediction for ‘fishy.’ This is a misclassification to treat with caution, not evidence that the correct odor was discovered. An explanation involving a structurally similar molecule or its oxidation products does not make the prediction for squalane itself correct.

The team also computationally screened 400 dipeptides, each made by joining two amino acids, and suggested candidates for odors such as meaty. However, people did not smell the candidates to verify those predictions. These were proposals for subsequent experiments, not results demonstrating their suitability as new fragrance or food ingredients.

Mapping relationships between odor words

Another part of the study compared the molecular features that the models emphasized for different odor words. Words whose predictions rely on similar features can be considered related within the model.

The researchers converted differences in feature importance into distances and compressed them into a small number of axes to produce a two-dimensional map. Fragrance experts interpreted clusters such as words associated with dairy products and fermentation, including buttery and cheesy, and words evoking wood and forests.

This was not a perceptual map created by asking new participants to rate odor similarities. It illustrates the structural cues shared by models trained on existing descriptions.

Even so, it offers a way to consider connections and continuous transitions between words, rather than treating each odor word as an independent category. The authors emphasize its value as an interpretable framework for understanding structure–odor relationships and searching for candidates.

Good predictions also require well-organized terminology

The comparison showed that relatively manageable codes representing atomic connectivity contain useful information, without requiring complex computational representations. However, the best method on average is not necessarily the best for every odor word.

If the odor descriptions used for training are ambiguous, the model inherits that ambiguity. The authors identify expert standardization of terminology as important alongside molecular representation. The targets were existing labels for individual molecules; the model does not directly predict the overall impression of a perfume, including concentration, mixtures, and individual differences.

All three authors work in COSMAX’s research and development division, and the company supported the research, computing equipment, and open-access costs. The paper declares no competing interests.

Summary

In a comparison using existing descriptions of 8,681 molecules, encoding atomic connectivity as substructures was useful on average. Yet the most suitable method varied by odor word, and high overall accuracy alone did not reveal incorrect predictions or missed labels.

The study suggests that both how molecules are represented and how odors are put into words need careful preparation. The authors propose that interpreting the features a model prioritizes and organizing relationships between words can connect prediction-based candidate searches with research into structure–odor relationships.

References

Suh, J., Hong, Y., & Park, C. (2025). A comparative study of machine learning models on molecular fingerprints for odor decoding. Communications Chemistry, 8, 278. DOI: 10.1038/s42004-025-01651-7. Full original article. © The Author(s) 2025. The original article is published under CC BY 4.0.