The results could have a range of applications, including helping perfume experts sift through billions of different molecules to find exactly the odor they’re searching for.
To begin with, Rockefeller University researchers asked 49 volunteers to assess the odor of 476 different chemicals, based on 21 different descriptors. These ranged from a chemicals’ intensity and pleasantness of smell, to how spicy or fruity it was. Once this data was assembled, the researchers then released it for 407 of the chemicals — together with 4,884 other variables based around chemical structure — for anyone who wanted to take a crack at writing an algorithm to make sense of it all.
The remaining 69 chemicals were held back so that whatever algorithms were created could be tested.
In the end, Professor Richard Gerkin, a neuroscientist at Arizona State University in Tempe, came up with the winning formula. His algorithm proved capable of predicting the scores the volunteers had assigned the chemicals based on their chemical compositions alone. Sure, 21 descriptors isn’t granular enough detail to accurately analyze every different chemical, but it’s an impressive start.
The possibility of using technology such as this to help narrow down the field when searching for a particular smell, or even flavor, is pretty exciting. “Eventually, you can use a database … and say, ‘OK, pick out the top 100 hits out of a billion molecules,'” Gerkin told New Scientist. “A hundred molecules are easier to test than a billion.”
Next up? Flipping the formula to predict which smells arise from mixing certain chemicals. Can it be done? Hey, if there’s one thing we’ve learned about machine learning, it’s that we write it off at our own peril.
