Google AI succeeds in developing odor maps of molecules

A challenging process

There are more than 300 scent receptors, as opposed to the human eye's three color sensors (red, green, and blue). Thus, it was thorny to create odor maps.

In 2019, Google AI developed a graph neural network (GNN) model that began to explore thousands of examples of distinct molecules paired with the smell labels that they evoke, e.g., “beefy,” “floral,” or “minty,” to learn the relationship between a molecule’s structure and the probability that such a molecule would have each smell label.

A neural network has now been used by Joel Mainland at the University of Pennsylvania and his colleagues, including researchers at Google, to create a map that connects a molecule's structure with the smell it will emit and measures how near together molecules are in terms of smell.

“The neural network seems to be learning some sort of representation of molecules that is more fundamental than what we expected,” says Joel Mainland, as per New Scientist.

5000 molecules from different scents

In the neural model, over 5000 molecules from two different flavor and scent data sets were used. By asking it to describe how 320 distinct compounds would smell based on their structures and contrasting this with descriptions of smell from a group of 15 individuals, Mainland and his colleagues evaluated the system's performance.

In addition, in a separate study, Mainland's colleagues used the neural network to create a map connecting the molecular structures of compounds that repel mosquitoes to how repellant the insects find the odors.

This allowed the researchers to identify molecules that, according to the model, would be at least as repellent as leading anti-mosquito products and could be tested in future trials.

Joel Mainland also acknowledges this and says future work will focus on producing models that can identify enantiomers and more complex mixtures of molecules rather than just the single molecules the current model works with.

Abstract:

Mapping molecular structure to odor perception is a key challenge in olfaction. Here, we use graph neural networks (GNN) to generate a Principal Odor Map (POM) that preserves perceptual relationships and enables odor quality prediction for novel odorants. The model is as reliable as a human in describing odor quality: on a prospective validation set of 400 novel odorants, the model-generated odor profile more closely matched the trained panel mean (n=15) than did the median panelist. Applying simple, interpretable, theoretically-rooted transformations, the POM outperformed chemoinformatic models on several other odor prediction tasks, indicating that the POM successfully encoded a generalized map of structure-odor relationships. This approach broadly enables odor prediction and paves the way toward digitizing odors.