Latest Machine Learning Research Reveals Hidden Order in Scents

Alex Wiltschko is an olfactory neuroscientist for Google Research’s Brain team. In his youth, his fascination with perfumes. He recently used machine learning to analyze their oldest and least known sense of smell. Their discoveries have greatly increased scientists’ ability to determine the smell of a molecule from its structure.

Over 800 chemicals reach your nose when you smell coffee. Our brain creates the general impression of coffee from this chemical image. Scientists are just beginning to understand how many of the 400 receptors in our nose can interact with a specific molecule to detect its chemical composition. Building models capable of deducing the smell of a molecule from its structure was a point of competition for the teams.

Even the finest models of smell can only account for certain things. Sometimes small changes in the chemical composition of a molecule result in a completely different smell. In other cases, substantial structural changes have little or no effect on odor.

The metabolic organization of an odor

Wiltschko and his team investigated what demands evolution might have placed on our senses to explain these unusual events. Over millions of years, each sense has been honed to pick up the most salient variety of stimuli.

“The central metabolic engine inside every living thing is the one thing that has remained constant through evolution, at least for a very long time. The term ‘metabolism’ refers to a set of chemical processes, such as the Krebs cycle, glycolysis, the urea cycle and many others, which are catalyzed by cellular enzymes and transform one molecule into another in living beings.These well-established reaction pathways describe the links between natural substances that enter our nostrils Substances with similar odors are biologically and chemically linked.

His team needed a map of metabolic processes that take place in nature to test the theory. These natural chemical interactions and the enzymes that cause them were fortunately already described in a large database that metabolomics researchers had compiled. Using this information, the researchers were able to determine how many enzymatic processes would be required to transform one odor molecule into another.

They also needed a computer model that could calculate how different odor molecules smell to humans for comparison. The research team worked on improving a neural network model known as the Primary Smell Map based on the results of the DREAM 2015 competition. This map looks like a cloud of 5,000 points, each representing a different smell molecule. Points for molecules with similar smells cluster together, while those with very different smells are discarded. Only advanced computer technologies can understand the structure of the cloud since it is much more than 3D: it has 256 dimensions of information.

In both data sources, the researchers looked for corresponding correlations. They looked at 50 molecule pairings and found that, despite having quite different structures, compounds closer to the metabolic map tended to be more relative to the scent map.

The researchers found that while the predictions were still not accurate, they were still better than any previous model that had been able to predict on chemical structure alone. It was not necessary at all, he said. “Two biologically similar compounds, one enzyme catalysis step apart, they could smell like rotten eggs and roses.” Scientists have also found that molecules with a natural relationship, such as the many chemical components of an orange, tend to resemble each other more than those that don’t.

Chemistry in harmony with nature

“The researchers said that the olfactory system is designed to pick up different [chemical] coincidences. Therefore, metabolism controls the likelihood of coincidences. This suggests that the metabolic process by which a molecule was created in the natural world is important to our nose in addition to its chemical structure.

“The olfactory system is adjusted to the cosmos it perceives, made up of these molecular structures. and a big part of that is how those molecules are created. They hailed the cleverness of the plan to classify odors using more metabolism. Since the metabolic origin of a molecule is already closely related to its structure, the metabolism-based map does little to improve the structural models, but it adds a little more information.

Instead of single molecules, the researcher said mixture odors would represent the next frontier in olfactory neuroscience. Consider the hundreds of chemicals emitted from your coffee mug. In real life, they rarely breathe in just one chemical at a time. There needs to be more information about the odorous mixtures before the researchers can create a model similar to the one used in the current study for the pure compounds. To fully understand our sense of smell, we will have to look at how chemical combinations combine to create complex odors.

Check paper and reference article. All credit for this research goes to the researchers on this project. Also don’t forget to register. our Reddit page and discord channelwhere we share the latest AI research news, cool AI projects, and more.


Ashish Kumar is an intern consultant at MarktechPost. He is currently pursuing his Btech from Indian Institute of Technology (IIT), Kanpur. He is passionate about exploring new technological advances and applying them to real life.


Sherry J. Basler