An AI system can identify the scent profiles of various chemical compounds based on their molecular makeup, and its assessments often align with those of expert human evaluators.
The scientists who developed the AI utilized it to categorize scents like ‘fruity’ or ‘grassy’ linked to many chemical formations. This database of odors could be valuable for experts working on creating new artificial fragrances and could also shed light on the cognitive processes behind how humans perceive smells. This study has been published in the journal Science.
Fast-Track to Memory
Odors have a unique characteristic: they directly move from the olfactory organ—the nose, specifically—to the brain’s centers responsible for memory and emotions. This direct path makes scents particularly potent in triggering vivid and specific memories.
“There’s something special about smell,” notes neurobiologist Alexander Wiltschko. He founded Osmo, based in Cambridge, Massachusetts, which aims to engineer novel odor molecules.
Wiltschko and his team at Osmo employed a kind of AI known as a neural network to assign one or several out of 55 descriptors like ‘fishy’ or ‘winey’ to a smell.
The system was trained to characterize approximately 5,000 scents while studying each chemical’s molecular structure to identify any correlations with its scent.
The AI spotted about 250 relationships between particular molecular patterns and corresponding odors. These findings were integrated into a “Principal Odor Map” (POM), which the AI could refer to when making predictions about the scent of a new molecule.
To verify the POM’s effectiveness, 15 human subjects were trained to associate particular scents with the same descriptors used by the AI. The study then collected various artificial smells and human subjects, and the AI was tasked with describing them based on their chemical structures.
The AI’s evaluations were generally in line with the average human response and often even more accurate than individual human guesses.
An Emerging Tool
“It’s a nice advance using machine learning,” observes Stuart Firestein, a neuroscientist at Columbia University. He suggests that the POM could be a handy reference in industries like food and cleaning.
However, Firestein also remarks that the POM doesn’t really enlighten us about the biological mechanisms behind human olfaction. “They’ve got the chemical side and the brain side, but we don’t know anything about the middle yet,” he states.
Pablo Meyer, a systems biologist at the IBM Center for Computational Health, appreciates the innovative use of language to connect molecular structures to subjective scents.
However, he contests that the average human response should be considered the “right” way to describe a smell. “Smell is something personal,” Meyer points out.
As for future research, Wiltschko aims to explore how different odors interact and compete to produce a new scent as the human brain perceives it. Meyer and Firestein acknowledge that this will be an exceedingly complex task, given the infinite combinations of molecules that could be involved.
“The next frontier,” according to Wiltschko, “is predicting what a mix of scents will smell like.”