How AI May Disrupt the Fragrance and Flavor Industries

Source: Engin_Akyurt/Pixabay

Freshly baked bread. Coffee brewing. Buttery popcorn. Can a machine digitally predict with artificial intelligence (AI) how humans will biologically perceive a certain smell? Recently a research duo at the University of California, Riverside, created a machine learning algorithm that can predict how a chemical will smell to a person—accelerating the digital disruption of the fragrance and flavor industry.

The research team of Anandasankar Ray and Joel Kowalewski at the University of California, Riverside, applied machine learning to predict the human perception of chemical smells and published their findings on August 21, 2020, in iScience, an open-access journal from Cell Press. Ray, a Yale alumnus with a Ph.D. in molecular, cellular, and developmental biology, is not only an associate professor at UC Riverside, but also the Founder and President of Sensorygen, Inc., and the Founder and Manager of Sensoromyx.

How does the sense of smell work in people? The human nose can discriminate over a trillion olfactory stimuli, according to a research study published in Science on March 21, 2014. There are special neurons, called the olfactory sensory neurons, which connect directly to the brain, and for every olfactory neuron, there is an odorant receptor (OR), according to the National Institutes of Health (NIH). When an odor molecule binds to olfactory cilia, it stimulates these neurons located in a patch of tissue high inside the nose.

An odor molecule is typically detected when it goes through the nostril; however, that’s not the only area where it may enter. Humans can also smell via the passage that is between the nose and the roof of the throat. Aromas are released when chewing food and travel through this passage. Thus, the sense of smell plays an important role in the enjoyment of food.

Yet, unlike the sense of taste and vision, scientists do not know exactly which odorant receptors are linked to certain smells. Moreover, the quantity of roughly 400 odorant receptors and the complexity of the possible combination of activated multiple receptors make it difficult to identify patterns.

This is where machine learning is a useful tool to make sense of data complexity. To select the important olfactory receptors, the team ran a recursive feature elimination with a random forest (RF) or support vector machine (SVM) algorithm. They decided to use Gaussian kernel or radial basis function (RBF) to enable the learning of non-linear boundaries, a natural fit for biologically active chemicals where physiochemical characteristics differ.

The research duo screened a large database of compounds for new ligands—ions or molecules that bind to a biomolecule to form a complex. As a result, the machine learning algorithm predicted the activity of 34 human odorant receptors from a chemical library of around 450,000 eMolecules. The team then demonstrated that the activity of the odorant receptors could predict 146 different percepts of chemicals and extended the study on fruit flies. Interestingly, the team discovered that not all the odorant receptors were required to predict some of the percepts—only a few olfactory receptor activities could yield behavior predictions.

The worldwide market for flavors and fragrance is projected to reach USD 28.6 billion by 2025, and aroma chemicals are forecasted to hold the largest market share with over 70 percent of the revenue, according to a March 2019 report by Grand View Research, Inc. By using AI to digitize smells and the associated human perception, researchers have a faster way to create novel chemicals with more desirable aroma characteristics for commercial and industrial purposes in the future.

Copyright © 2020 Cami Rosso All rights reserved.