SENSE OF SMELL PREDICTED BY ARTIFICIAL INTELLIGENCE

We’ve been deluged with news about artificial intelligence (AI) over the past year. We’ve heard how machines are learning to write and speak almost as well as we can.

Computers are even driving cars and composing paintings and music. They can often see and hear better than people when they perform tasks for us.

One area we haven’t been hearing so much about is how machine learning can mimic our sense of smell. It’s been hard to predict how new chemical products will smell, for example.

‘CHEMOINFORMATICS’ OF ODOURS

There’s even a field of study called “chemoinformatics” The chemoinformatics of odours involves using computers to analyze, model and understand how chemical properties stimulate our sense of smell.

In the past, perfumers and chemists simply relied on empirical evidence to associate smells and chemicals. They knew chemical structures influenced odours, but usually, they’d just take a whiff and write down their impressions.

More recently, researchers have worked out how to calculate the Quantitative-Structure-Odour Relationship (QSOR). This establishes mathematical correlations between molecules and their smells.

DR. EMILY MAYHEW STUDIES FOOD SENSORY PROPERTIES

Dr. Emily Mayhew is a food scientist at Michigan State University. For the past six years, she’s been working to understand how the chemistry of a food dictates its sensory properties, like taste and smell.

Professor Mayhew is a co-leader of study that the journal Science published this week. The study describes a machine-learning model that can predict how a new chemical will smell without having anyone smell it.

The model is called the Principal Odor Map, and it’s predicted smells for 500,000 molecules that no one has ever synthesize. If someone tried to do that manually, it would take them over 70 years.

‘ORDERS OF MAGNITUDE FASTER’

“Our bandwidth for profiling molecules is orders of magnitude faster,” Professor Mayhew told Scientific American. The system can also predict an odour’s intensity and whether two molecules have similar smells.

“That was a really cool surprise,” Professor Mayhew explained. The model’s predictions were just as accurate as the judgements of 15 trained scent judges.

The researchers pulled together data on over 5,000 molecules, including their chemical structures and odour labels. This formed what the team calls a “principal odour map” that catalogs 256 dimensions for each compound.

SLIGHT MOLECULAR DIFFERENCES CAUSE RADICAL SMELL CHANGES

Up until now, results from chemoinformatics odour prediction models have been modest. Even slight differences in a molecular structure can cause radical changes in how it smells.

At the same time, very different molecules can have comparable scents. So, understanding our sense of smell is more complex compared to our other senses.

We know that the pitch of a sound or a colour of light is a function of its wavelength. When it comes to our sense of smell, our noses detect molecules in a more complex and subjective way.

WE OFTEN HAVE INCONSISTENT PERCEPTIONS OF SMELLS

That’s why, even though most of us perceive sights and sounds in very similar ways, we often have inconsistent perceptions of smells. That’s where this new study comes in.

The researchers went “back to the drawing board” and created a brand new model, independent of conventional chemoinformatics. They trained a neural network on their database of 5,000 models with 256 dimensions.

That’s too complicated for our human brains to even grasp. The system then generates a massive odour map, plotting each molecule based on its chemistry.

ODOUR MAP DESCRIBES MOLECULES BASED ON 55 CATEGORIES

The principal odour map can describes each molecule based on 55 categories, like floral or woody. Chemicals with similar scents group together on the map, which is a breakthrough for this new model.

Our senses define what it’s like to be human. Other animals process their senses in different ways, which must make what it’s like to be them different from us.

Every dog owner can confirm that dogs can zero in on a smell to which we’re completely oblivious. They also hear sound pitches we don’t realize exist.

AND ANOTHER THING…

Understanding how we perceive the world around us is part of the new story we all need to better grasp the world around us and our place within it. We’re largely visual animals, and we emphasize our perceptions of sounds, especially through music.

We tend to neglect our sense of smell, but it also plays a role in how we perceive the world. In fact, since we pay less attention to smells, our memories of them can be more intense than other senses.

The researchers plan to take their work to the next level by finding ways to predict how we would perceive these molecular odours in context. In daily life, we process a constant blend of smells, so a more practical model would predict our perception of mixtures of molecules.

Professor Mayhew concluded by saying “Mixture perception is the next frontier.”

We always have more to learn if we dare to know.

Learn more:

Machine Learning Creates a Massive Map of Smelly Molecules

A principal odor map unifies diverse tasks in olfactory perception

Nature Walks Benefit Both Mind and Body

Urban Green Space Reduces Property and Violent Crime

Subjective Memories Drive Our Decision Making