TDT | Distributor
Daily Tribune – Bahrain
To the human nose, hydrogen sulfide smells like rotten eggs, and geranyl acetate smells like roses, but the problem of predicting what a new chemical will smell like without letting people smell it has long vexed food scientists, flavorists and neuroscientists.
Now, in a study published in the journal Science, the researchers describe a machine learning model that does just that: Called the “Dominant Odor Map,” the model predicted the smell of 500,000 molecules that had never been synthesized before — a task that would take a human 70 years to accomplish.
“Our bandwidth for molecular profiling is orders of magnitude faster,” says Emily Mayhew, a food scientist at Michigan State University who co-led the study. Although the color of light is defined by its wavelength, there is not such a simple relationship between a molecule's physical properties and its smell.
Small changes to a molecule's structure can dramatically change its smell. Conversely, chemicals can have similar smells even if their molecular structures are different. Previous machine learning models have found associations between the chemical properties of known odorants (known as cheminformatics) and odors, but their predictive performance was limited.
In the new study, the researchers trained a neural network on 5,000 known odorants to emphasize 256 chemical features depending on how much they affect the molecule's smell. Rather than using standard cheminformatics, “they built their own,” says Pablo Mayer-Rojas, a computational biologist at IBM Research who was not involved in the work.
“They directly inferred the properties associated with smell,” he says, but how the models arrive at these predictions is too complex for humans to understand: They created a giant map of smells, with the coordinates of each molecule determined by its chemical properties.
The model also predicts what each molecule smells like to humans, using 55 descriptive labels such as “grassy” or “woody.” Remarkably, odorants with similar smells appeared together on the map, a feature not possible with previous odor maps.
