Scientists Use Machine Learning to Map How Complex Smells Relate

Researchers co-authored by scientists at the Monell Chemical Senses Center have built a machine-learning system that can predict how similar different smell mixtures are to one another, a step toward what the team calls digital olfaction, the ability to digitize scents the way color and sound have long been digitized. The work, published in the Proceedings of the National Academy of Sciences, provides what the researchers describe as a validated metric and benchmark for comparing smells.

Study co-author Joel Mainland of the Monell Center said an earlier 2015 IBM-run DREAM challenge tackled single odor molecules, predicting what one molecule smells like from its chemical structure. But most real-world odors are complex mixtures of dozens or hundreds of molecules, so Mainland and colleagues ran their own DREAM challenge focused on mixtures. They compiled six data sets from three studies into a new set comprising 168 unique single molecules, 731 unique mixtures and 507 mixture-pair measurements, scored on a scale from 0 (indistinguishable) to 1 (most distinct).

Over three months, 26 international teams competed to predict the similarity of paired scents in a hidden test set of 46 mixture pairs, ending in a four-way tie. Mainland's team then built an ensemble model by averaging the four winning teams' predictions with two additional high-performing models, and validated it on an independent set of 50 mixture pairs. The final model achieved a median root mean square error of 0.08 and a Pearson correlation of 0.57, which the researchers said showed good to moderately strong accuracy.

Many sensory scientists had expected predicting mixture similarity to be far harder than predicting single-molecule smells, but that did not turn out to be the case. Mainland said the models were more likely to use everyday descriptive words such as "fruity" and "sweet" than chemistry terms, and removing those semantic labels — tested after a journal reviewer's request — made predictions notably harder. The team is now preparing results from a third DREAM Olfaction challenge, in which predicting a mixture's smell from its individual components' known smells worked out to be roughly an average of the parts.