AI beauty

AI is joining the ranks of superheroes like Daredevil and Wolverine who have a heightened sense of smell.

As reported in the journal Science, an AI program has been trained to produce odor profiles like fruity or fishy based on chemical structure alone. The program, a graph neural network, has proved remarkably accurate. The program reliably predicted the olfactory descriptions of a diverse, 14-member human panel. The program was developed by an AI company called Osmo AI, a spinoff from Google Research. Among the company’s investors is the Bill & Melinda Gates Foundation.

A key development was the creation of a 500,000 molecule database of hypothetical chemical structures and what they would smell like. Basically, Osmo AI converted atoms into digital bits of information. This database will aid in the search for odors in new foods, perfumes, cleaners and other products. The database also will be key in designing insect repellants that smell bad to mosquitoes, for example.

OsmoAISmellMap

OsmoAISmellMap

“We see AI augmenting the role of synthetic chemists and master perfumers, the scent hunters who comb through myriad molecules to create the scents that are an integral part of our lives,” says Osmo AI team leader Alex Wiltschko.

The Osmo AI team input the structures and odor descriptions of 5000 molecules into a sophisticated AI program that learned to recognize patterns in the training data, correlating a molecule’s smell with features of its constituent atoms, like size and connecting bonds. The 5000 molecules were paired with smell labels they evoked such as fruity, floral, cheesy or minty.

Digitizing scent required the creation of the first map for smell, a task daunting in its complexity. While there are three channels of color information in your eyes, there are over 300 channels of odor information in your nose. The Osmo AI team created a Principal Odor Map (POM) using novel datasets and AI. When Osmo AI conducted what it calls the “Odor Turing Test,” its model performed better than the average human panelist on 53 percent of the new molecules tested, predicting odor solely from molecular structure.

Osmo AI says there is still work to be done in refining AI’s nose for smells. Among the challenges are predicting how strong a particular smell might be and how molecules blended together might smell. Osmo AI also would like to predict smell directly from chemical sensors.

Osmo AI’s map of odor—the company calls it an Osmograph– is the foundation for broader ambitions. “If we can build on our insights to develop systems capable of replicating what our nose, or what a dog’s nose can do (smell diseases!), we can spot diseases early, prevent food waste, capture powerful memories and more,” says Wiltschko. “Digitizing scent could catalyze the transformation of scent from something people see as ethereal to enduring.”