Artificial intelligence got a nose 01.03.2017

Scientists have long argued about exactly how the receptors of the human body allow us to perceive a wide range of odors and give them certain descriptions. In an attempt to solve this problem, teams of engineers from around the world were asked to create an AI that could perceive smells as well as humans.

Predicting color is not that difficult: for example, if a light wave reaches a wavelength of 510 nm, then most people will say that it is green. But figuring out how a particular molecule smells is much more difficult. 22 teams of scientists have created a set of algorithms that can predict the smell of various molecules based on their chemical structure. The full range of practical applications of the program remains to be seen, but the developers hope that, first of all, it will help perfumers, pharmacists and food workers develop new, unique combinations of smells.

The work began with a recent study by Leslie Vosshull and colleagues at the Rockefeller University in New York, in which 49 volunteers had to guess the smell of 467 odorants. For each of them, a comparison system of 19 basic patterns was developed: the subjects said whether the smell was similar to fish or garlic, assessed the intensity and individual pleasantness of the aroma. As a result, a catalog was created with more than a million cells characterizing certain odorous molecules.

When computational biologist Pablo Meyer found out about this, he immediately saw the study as an opportunity to test whether a computer system could predict how people would evaluate smells. Despite the fact that researchers have discovered about 400 odor receptors in the human body, it remains a mystery to scientists how exactly they work together so that a person can distinguish even subtle shades of odors.

In 2015, Meyer and colleagues launched the DREAM Olfaction Prediction Challenge. The participants of the competition received at their disposal the same rating tables of volunteers that describe odors, along with the chemical structure of the molecules that produce them. In addition, the participants were provided with a database of 4800 descriptions for each individual molecule - its atoms, their relative position, general geometry, which eventually amounted to about 2 million data points. Ultimately, the data should be used to train computer programs to recognize odors based on structural information.

Twenty-two teams from all over the world entered the competition, and while many did a good job, two teams stand out in particular. The Michigan State team, led by Yang Fang Guang, was the best at coming up with an algorithm for predicting the smell of individual things. Another team from the University of Arizona, led by Richard Gerkin, was the best at training the program to average odor ratings across the entire sample. Meyer reports this in an article published in the journal Science.

Of course, many scientists are skeptical about the developments, saying that the work done, although it makes a significant contribution to science, is still a rather primitive selection, and 19 descriptive elements for the entire spectrum of smells in nature are clearly very, very few.

Alternative studies with volunteers used 80 or more of these criteria to verbally evaluate different odors. It is not clear whether the existing algorithm will be able to correctly predict the odor score if it has to face such an array of information. So, to date, the perception of smells remains a mystery for both physicians and engineers.