The human sense of smell has always been a fascinating realm of neuroscience. While we understand how we translate light into sight or sound into hearing, the realm of olfaction (smell) remains more complex and perplexing.
A groundbreaking collaboration between the Monell Chemical Senses Center and the start-up Osmo - a company emerging from Google's machine learning research - has made significant progress in unravelling this mystery. Their research, published in the acclaimed journal Science, details the creation of a machine-learning model capable of human-level proficiency in predicting how chemicals might smell.
Image Credit: Antonio Guillem via shutterstock / HDR tune by Universal-Sci
Machine Learning Takes on Olfaction
Humans boast roughly 400 functional olfactory receptors. These receptors help identify airborne molecules, a function far more intricate than our vision (which uses four receptors) or even taste (with about 40 receptors). But the relationship between the structure of these airborne molecules and our perception of their scent has long eluded scientists.
Enter Osmo, whose CEO Alex Wiltschko, and his team embarked on a quest to bridge this knowledge gap. By matching prose descriptions of a molecule's odour with its molecular structure, they created a "map" of odours. This innovation brings the world one step closer to digitizing smells, something that's already been achieved with sight and sound.
Training the Machine
To train their model, Osmo utilized a dataset containing the molecular structures and odour characteristics of 5,000 known odorants. The model then predicts which odour words best describe a molecule based on its shape
To ensure the model's accuracy, the Monell team conducted a blind validation procedure. Panellists, trained with specialized odour reference kits, were given 400 odorants and a set of 55 descriptive words to describe each molecule's smell. This method ensured a robust understanding and distinction between similar terms, preventing confusion like the one between "musty" and "musky".
Quality control was imperative, with Prof. Jane Parker of the University of Reading verifying the purity of samples to ensure impurities didn't interfere with the model's predictions.
Image Credit: Studio Romantic via Shutterstock / HDR tune by Universal-Sci
Surpassing Human Perception
Astoundingly, the model not only matched human performance but even surpassed it in some instances. It was better at predicting group average odour ratings than any individual panellist. More surprisingly, it succeeded in tasks it wasn't trained for, like determining odour strength.
The potential implications of this research are vast, ranging from identifying new scents for industries to fundamentally shifting how scientists perceive and categorize odours. Rather than classifying odours based on their chemical structures, it appears that our brains might organize smells based on the nutrients they come from.
As this model map continues to evolve, it may reveal even deeper insights into the intricate dance between our brain, nose, and the world of smells that surround us.
If you are interested in learning more about the underlying research, be sure to check out the paper publisehd in the peer reviewed journal Science, listed below.
Sources and further reading:
Too busy to follow science news during the week? - Consider subscribing to our (free) newsletter - (Universal-Sci Weekly) - and get the 5 most interesting science articles of the week in your inbox
FEATURED ARTICLES:
