A measure of smell enables the creation of olfactory metamers

Wavelength is a physical measure of light, and the intricate understanding of its link to perceived colour enables the creation of perceptual entities such as metamers—non-overlapping

spectral compositions that generate identical colour percepts1. By contrast, scientists have been unable to develop a physical measure linked to perceived smell, even one that merely

reflects the extent of perceptual similarity between odorants2. Here, to generate such a measure, we collected perceptual similarity estimates of 49,788 pairwise odorants from 199

participants who smelled 242 different multicomponent odorants and used these data to refine a predictive model that links odorant structure to odorant perception3. The resulting measure

combines 21 physicochemical features of the odorants into a single number—expressed in radians—that accurately predicts the extent of perceptual similarity between multicomponent odorant

pairs. To assess the usefulness of this measure, we investigated whether we could use it to create olfactory metamers. To this end, we first identified a cut-off in the measure: pairs of

multicomponent odorants that were within 0.05 radians of each other or less were very difficult to discriminate. Using this cut-off, we were able to design olfactory metamers—pairs of

non-overlapping molecular compositions that generated identical odour percepts. The accurate predictions of perceptual similarity, and the ensuing creation of olfactory metamers, suggest

that we have obtained a valid olfactory measure, one that may enable the digitization of smell.

All data generated during this study are included in the Article and its Supplementary Information. All the odorants used are included in Supplementary Table 1, all behavioural similarity

results are included in Supplementary Table 2 and all behavioural discrimination results are included in Supplementary Table 3. An additional external dataset used can be found in the

supplementary material of a previously published study15.

The custom code used to process the data collected in this study is available at https://gitlab.com/AharonR/olfaction.

This work was primarily supported by the Horizon 2020 FET Open project NanoSmell (662629). Additional support from grant 1599/14 from the Israel Science Foundation, by a grant from Unilever,

and by the Rob and Cheryl McEwen Fund for Brain Research.

Department of Neurobiology, Weizmann Institute of Science, Rehovot, Israel

Aharon Ravia, Kobi Snitz, Danielle Honigstein, Maya Finkel, Rotem Zirler, Ofer Perl, Lavi Secundo & Noam Sobel

Department of Computer Science and Applied Mathematics, Weizmann Institute of Science, Rehovot, Israel

A.R., K.S., L.S., D. Harel and N.S. developed the concepts. A.R. and N.S. designed experiments. A.R., R.Z. and M.F. ran experiments. A.R., K.S., O.P. and N.S. analysed data. C.L. developed

scent formulas. A.R., D. Honigstein, K.S., O.P. and N.S. constructed the web-tool. A.R., O.P., D. Harel and N.S. wrote the paper.

The Office of Technology Licensing at the Weizmann Institute of Science is filing for patents on the algorithms developed in this study. A small portion of this work was supported by a

research grant from Unilever, a company with interests in the fragrance industry. Unilever had no input or impact on the design of experiments, or on analysis and presentation of the

results. C.L. is the owner of DreamAir LLC, a company with interests in the fragrance industry. DreamAir had no input or impact on the analysis and presentation of the results.

Peer review information Nature thanks Tatyana Sharpee and the other, anonymous, reviewer(s) for their contribution to the peer review of this work.

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

a, As in the main text, the 148 molecules used across experiments overlaid on 4,046 molecules within the first and second principal components of the 21-descriptor physicochemical space. b,

The same molecules within the first and second principal components of perceptual space. Perceptual space data for 470 molecules as background (data from previously published studies4,7),

containing 115 of the 148 molecules that we used. c, Histograms showing the experiment odorant distribution on each principal component (PC) in the range of PC1–PC6. The principal components

were computed as in a, on the 21-descriptor physicochemical space. There is a large decline in the explained variance from the third principal component onward. d, Histograms showing the

distances between all odorant pairs, per experiment. The distances are summed (black line) for the overall distribution. Although monomolecules were not used as a stimulus for

discrimination, this is to show that there was no bias in their selection, because for each experiment the distances of the pairs spanned a range of distances.

Ordered depiction of the tasks across the seven reported experiments.

a, b, Factoring odorant intensity. a, In experiment 1, the overall MC-odorant intensity could have been used to determine similarity, n = 23 participants for intensity ratings and 22

participants for similarity ratings. Correlation coefficient r = −0.61, P