This neural net maps aroma molecules

Sensory science has come a long way in explaining how certain physical phenomena—for example, a particular wavelength of light, or a column of air vibrating at a set frequency—are related to common sensory experience. The sense of smell, however, has proven elusive. Until recently, there was no way to take the physical properties of a compound or the structural formula of a molecule and have a sense of what it might smell like.

Using a type of deep-learning algorithm called a graph neural network, the researchers created a model that maps chemical structures to odor descriptors. The model successfully predicted how a panel of humans would describe new smells and could be an important step towards digitizing smells. A study published on August 31 described the work Science.

“This paper is a milestone in predicting aroma from the chemical structure of odors,” said Michael Schmucker, a professor of neural computation at the University of Hertfordshire in England, who was not involved in the study. Although scent maps should be useful and this work “represents a leap forward,” he said, more work is needed to address nose-tingling possibilities, such as sharing smells on the Internet.

This abstract representation of Osmo’s odor map shows the olfactory relationships between molecules.

The model used a special type of graph neural network called a message-passing neural network. It was trained on a combined fragrance industry dataset of over 5,000 molecules and their structure was converted into a graph and tagged with commercial odor notes. Parts of the research group worked at Google when the work began, and some later formed an offshoot company Osmo in January 2023, backed by Google Ventures, the venture capital arm of Alphabet.

“The predictive power of graph neural networks allowed us to do this,” said co-author Alex Wiltsko, CEO of Osmo.

In the long run, Osmo seeks to digitize smell in the same way that images and sounds can be recorded and transmitted. Full-blown scent digitization will help develop new ways of producing or analyzing scents, leading to a wide range of new products and technologies, such as medical tests, treatments or prosthetics.

The model created a spatial representation showing the similarity of odor descriptors applied to different molecules. With more than 250 dimensions, the model is more complex than a similar representation for colors, for example. Given only the chemical structure as a graph of a novel molecule—that is, anything not included in the training set—the model can map it, essentially predicting how the odor will be described. The core odor map, as the team called it, is unprecedented for the sense of smell. “This master odor map is the first step toward giving computers a sense of smell,” Wiltsko said.

Researchers have a good understanding of the individuality and personality of the sense of smell. “The difficult thing about talking about how the model is working is that we don’t have an objective truth,” said co-author Joel Mainland, a neuroscientist at the Monell Chemical Senses Center and the University of Pennsylvania. Mainland is now Osmo’s scientific advisor.

To validate the performance of the model, evaluations of 400 new molecules were compared to ratings of a 15-person panel trained to recognize 55 odor labels. Training reference samples were a mixture of pantry staples, supermarket treats (such as green apple Jolly Rancher for “apple”), and specially acquired odorants. “Animal” labelIt was taught by a vial of horse sweat. “It has a fantastic smell—really complex and interesting,” Mainland said.

Performance was not flawless, but 53 percent of the time the model was closer than the average panelist to the average panel evaluation. In other words, Mainland explains, replacing a panel with a model improves group description.

Schmucker was part of a group of scientists who, independently of the original research team, reproduced the model and master odor map based on a preprinted version of the study. They have made their project openly available.

Areas of further research include odor intensity; mixtures and concentrations of several basic aromatic molecules; digitization of real-world odors when molecular structures are not given; and improving descriptive power.

The model was run on Nvidia’s Tesla P100 GPU. While the GPU is relatively powerful,Compares the current specification of Mainland Vas labeling to 8-bit graphics. Panelists described one molecule as “sharp, sweet, roasted, buttery”. A master perfumer, when consulted on the same scent, wrote: “Ski lodge; A fire without fire.”