To train AI to predict how a molecule smells, we need information about the molecule and data on how people perceive it. Yet odor studies differ in the names, concentrations, and recording formats they use. Simply gathering papers does not necessarily make their results ready for joint analysis.
Elizabeth A. Hamel and colleagues described Pyrfume in 2024 as a public resource for organizing olfactory data in a shared, usable format. Their paper does not report a single experiment in which people smelled large numbers of new odors. Instead, it presents a framework connecting existing studies and databases, together with its validation and possible uses.
From people’s impressions to receptor responses
At the time the paper was written, Pyrfume brought together more than 40 datasets containing information on more than 20,000 odorants. These covered more than human ratings of odor intensity, pleasantness, and verbal descriptions: they also included animal behavior, neural activity, and receptor responses to odor molecules.
This breadth is deliberate. For the same molecule, researchers can look for which receptors respond, how neurons respond, and how people describe it, linking information across different levels. However, not every molecule has measurements at all these levels.
The figures of 40 datasets and 20,000 odorants describe the scale reported in the 2024 paper. They are not current counts for an evolving resource, nor numbers of experiments performed under identical conditions.
Keeping track of stimulus conditions, not just molecular names
The organizing principle is to assign identifiers to the stimuli used in experiments. Individual molecules, mixtures, and materials such as cloves whose complete composition is unknown are each treated as stimuli.
Where molecular structure is known, the records link to identifiers in the PubChem chemical database and to textual representations of structure. This helps users find data for the same compound even when different sources use different names.
A compound identifier alone cannot describe experimental conditions, however. The same molecule at a different dilution or mixed with different substances is not necessarily the same stimulus. Pyrfume includes concentrations, mixture ratios, solvents, and other available information in experiment-specific stimulus records.
Nor does it force mixtures of unknown composition into a single molecular identifier. It retains identifiers for the materials themselves, in a format that accommodates the absence of a match to individual molecules.
Linking each value to what was actually measured
Measurement records connect the stimulus, subject, measured variable, and value. For example, a person’s rating of how floral an odor smells and a receptor’s response to the same molecule are recorded as different kinds of measurement.
The paper groups these broadly under “behavior,” but the term does not mean human behavior alone. It provides a place for experimentally observed values, including neural responses and receptor sensitivity. The design avoids placing numbers together without retaining what they measure.
The resource also records original publications, dataset descriptions, and processing programs. Where possible, it includes raw data, allowing users not only to view organized tables but also to trace processing and repeat it if necessary.
Building a basis for comparison before judging AI performance
The authors state that they reviewed and cross-checked the collected material in addition to the evaluations by its original collectors or providers. Standardizing formats does not, however, make the participants, animal species, concentrations, or measurement procedures of different studies identical.
The paper also does not show that using all the data improves AI predictions by a fixed percentage. Its purpose is to enable hypotheses linking molecules and perception to be tested across multiple studies, and prediction models to be compared on shared tasks.
The authors emphasize that researchers and learners can both retrieve existing material and contribute their own experiments in the same format. Shared benchmarks can help reveal not only where models perform well, but also where they struggle and where measurements are missing altogether.
Users must still check the conditions attached to each original dataset. Public availability or access from a common location does not mean that all the data can be used commercially under the same permission. The article’s license and the terms governing individual datasets are separate matters.
Conclusion
Pyrfume connects scattered olfactory findings while preserving stimulus conditions and the meaning of measurements. It offers a shared starting point not only for AI that predicts odors, but also for research into relationships among perception, neurons, and receptors.
The authors hope this organization will let researchers test hypotheses against diverse data, identify blind spots in earlier work, and choose future experiments. Alongside building new models of olfaction, establishing how to test those models is an important part of advancing our understanding of smell.
References
Hamel, E. A., Castro, J. B., Gould, T. J., Pellegrino, R., et al. (2024). Pyrfume: A window to the world’s olfactory data. Scientific Data, 11, 1220. DOI: 10.1038/s41597-024-04051-z / Full text.
The research was supported by the US NIH and NSF. Author affiliations include research institutions and companies; the authors declared no competing interests. Original article: © 2024 The Author(s), CC BY-NC-ND 4.0.
