September 18, 2013 — Of the five senses, smell is the least understood, but an Oak Ridge National Laboratory researcher is sniffing out answers that could help establish a systematic understanding of how people categorize odors.
The paper, published in PLOS ONE, could ultimately result in more complete explanations of how the brain’s odor processing mechanism represents and categorizes odors, and help in the effort to predict mental impressions of odors from chemicals, said Arvind Ramanathan, co-author and a member of ORNL’s Computational Sciences and Engineering Division.
Ramanathan and co-authors from Bates College and the University of Pittsburgh also envision this work aiding in the design of sensors and other pursuits that have been stymied because of this void.
“While taste has five basic qualities – sweet, sour, bitter, salty and umami, or savory -- smells are notoriously difficult to describe and even more difficult to categorize,” Ramanathan said. “Our results suggest that there are approximately 10 basic odor qualities, which are each associated with distinct chemical features.”
Unlike taste buds, in which the senses are well characterized with clear identities for receptors and how these tastes are recognized, for odor there is a fundamental question about basic qualities of smell. The researchers defined this as “odor space” and sought to answer this question by identifying discrete categories of smell. These categories are intuitive – floral, fruity, sweet, sour – and appeal to the chemical diversity of odor molecules, Ramanathan said.
A key to this research was the use of “non-negative matrix factorization,” which the researchers define as “a dimensionality reduction technique to uncover structure in a panel of odor profiles.” The technique streamlines information and sorts it into coherent categories similar to the way a photo or digital file can be compressed without compromising its inherent usefulness.
This approach has been successfully used in many other domains, including the financial world and for characterizing images and videos. For this study, the research team used a standard set of olfactory perception data, Andrew Dravnieks’ 1985 Atlas of Odor Character Profiles.
“What non-negative matrix factorization is good at,” said co-author Chakra Chennubhotla of the University of Pittsburgh, “is dividing the dataset into its constituent parts. You let the data tell you how many parts it has while building in some constraints to help the technique reveal them.”
The lead author of the paper, titled “Categorical dimensions of human odor descriptor space revealed by non-negative matrix factorization,” is Jason Castro of Bates College. The paper is available at http://dx.plos.org/10.1371/journal.pone.0073289
While no funding was received for this research, Ramanathan sees future opportunities because the work has potential implications in a number of areas, including homeland security.