Google AI Can Tell The Smell of Something From Its Molecular Structure

Credit: pexels.com

GOOGLE has used artificial intelligence to produce a map that relates smells to the structures of molecules. it's as reliable as a personality's in describing the aroma of a substance and also the researchers behind the work say it's a step towards digitising odours. 

The hope is that the AI model could in the future be wont to identify new scents for fragrances and flavors for food or come up chemically to repel disease-carrying organisms.

Mapping how our perceptions of smell relate to the physical source of an odour is difficult. Unlike the color sensors within the human eye, which detect just red, green and blue, there are over 300 scent receptors, so there's a far greater diversity of scents that we are able to potentially detect. 

Differences in people’s capacity to smell, subjective opinions on what things smell like and an absence of any obvious “primary” scents, such as what exists for colours, only complicate things further. 

This makes digitising smells – encoding information so any smell is recreated from its digital signature, just like the RGB digital scale for colours – an incredibly hard task. 

Now, Joel Mainland at the University of Pennsylvania and his colleagues, including researchers at Google, have used a neural network to form a map that links a molecule’s structure with its odour and to live how close molecules are in terms of smell. 

The researchers only designed the model to link existing odours and molecules, but it seems ready to predict the odour of molecules that have not been smelled before, too. “The neural network seems to be learning some variety of representation of molecules that's more fundamental than what we expected,” says Mainland. 

The model was trained on two flavour and fragrance data sets for quite 5000 molecules. Mainland and his colleagues tested its abilities by getting it to explain how 320 different molecules would smell supported their structure and comparing this against smell-based descriptions from 15 people. The model performed additionally because the average person (bioRxiv, doi.org/jb72). 

Mainland’s colleagues also applied the AI to an information set of chemicals that repel mosquitoes to form a map relating molecular structure to how repellent the insects find the scents. 

This enabled the researchers to spot molecules that may be a minimum of as repellent as leading anti-mosquito products and which might be tested in trials (bioRxiv, doi.org/jb73). This method may be accustomed find molecules to discourage other disease-carrying organisms, as long as training data exists.

It's impressive work, says Barry Smith at the varsity of Advanced Study, University of London. “What the work does is skip neurobiology and check out to attach the structure of molecules to the perception of odours directly.” He says it helps us understand odours of single molecules, but that isn’t the entire story. “Nearly all of the smells we are attentive to – wine, coffee, soap, people, the ocean – are thanks to a combination of several hundred volatile molecules,” he says. 

It's going to even be a difficulty that the model can’t distinguish between enantiomers, mirror-image molecules with the identical structure, but different smells, says Smith. Mainland acknowledges this and says future work will specialise in models that may identify enantiomers and more complex mixes of molecules.