Presentation Details
| Predicting Odor Mixture Character from Chemical Structure Xuebo Song1, Yuanfang Guan2, Matej Hladiš3, 4, Nachman Keren5, 6, Maxence Lalis3, Leonor Saiz7, José Vilar8, 9, Evan Guerra1, 10, Yikun Han10, Ashok Palaniappan11, Maria Diaz12, Gaia Andreoletti12, Verena Chung12, Robert Pellegrino1, Pablo Meyer13, Joel D.Mainland1, 14. 1Monell Chemical Senses Center, Philadelphia, PA, USA.2Department of Computational Medicine and Bioinformatics, University of Michigan, Ann Arbor, MI, USA.3Institut de Chimie de Nice, Université Côte d’Azur, Nice, France.4Department of Computer Science, University of Oxford, Oxford, United Kingdom.5Department of Statistics & Data Science, The Hebrew University of Jerusalem, Jerusalem, Israel.6Food Science and Nutrition, The Hebrew University of Jerusalem, Jerusalem, Israel.7Department of Biomedical Engineering, University of California Davis, Davis, CA, USA.8Biofisika Institute (CSIC, UPV/EHU), University of the Basque Country, Bilbao, Spain.9IKERBASQUE, Basque Foundation for Science, Bilbao, Spain.10School of Information Sciences, University of Illinois Urbana-Champaign, Champaign, IL, USA.11School of Chemical and Biotechnology, SASTRA Deemed University, Thanjavur, India.12Sage Bionetworks, Seattle, WA, USA.13IBM Research, New York, NY, USA.14Department of Neuroscience, University of Pennsylvania School of Medicine, Philadelphia, PA, USA |
Abstract
The mapping between physical properties and perception is well-established in vision and audition, but remains poorly understood in olfaction. Recent advances in predictive modeling and availability of large-scale perceptual datasets have enabled reliable odor prediction from molecular structures. However, current models neglect concentration-dependent perceptual changes and focus on single molecules rather than real-world mixtures. To address these gaps, we organized the third DREAM Olfaction Challenge in 2025, focusing on two prediction tasks: (i) odor quality prediction for single molecules across concentrations, and (ii) prediction of the perceptual qualities of odor mixtures. We curated two complementary datasets: (i) 151 monomolecular odorants measured at two concentrations, and (ii) over 650 odor mixtures composed of 2, 3, 5, or 10 components, all profiled by trained panelists using a standardized 51-word odor lexicon. For single-molecule prediction, the top three teams achieved Pearson correlations ranging from 0.73-0.75, surpassing the measurement error on the provided data for the first task (r = 0.72). For mixtures, the top performing models achieved a Pearson correlation of 0.79, exceeding measurement error on the provided data for the second task (r = 0.73). These results significantly advance the mapping of chemical structure to human olfactory perception, establishing a foundation for digital representation of odor mixtures.
No part of this publication may be reproduced, distributed, or transmitted in any form or by any means, including photocopying, recording, or other electronic or mechanical methods, without the prior written permission of the author.
No part of this publication may be reproduced, distributed, or transmitted in any form or by any means, including photocopying, recording, or other electronic or mechanical methods, without the prior written permission of the author.