Google has used an artificial intelligence to produce a map that relates smells to the structures of molecules. It is as reliable as a human in describing the odour of a substance and the researchers behind the work say it is a crucial step towards digitising odours.
The hope is that the AI model could eventually be used to identify new scents for fragrances and flavours for food science, or come up with chemicals to repel organisms like mosquitoes that can carry diseases.
Mapping how our perceptions of smell relate to the physical source of an odour is difficult. Unlike the sensors in the human eye, which detect just three colours (red, green and blue), there are more than 300 scent receptors, so there is a far greater diversity of scents that we can potentially detect.
Advertisement
The differences in people鈥檚 capacity to smell, subjective opinions on what things smell like and a lack of any obvious 鈥減rimary鈥� scents, akin to what exists for colours, only complicate things further.
This makes digitising smells 鈥� encoding information so that any smell can be recreated from its digital signature, like the RGB digital scale for colours 鈥� an incredibly hard task.
Now, at the Monell Chemical Senses Center in Philadelphia, Pennsylvania, and his colleagues, including researchers at Google, have used a neural network to produce a map that links a molecule鈥檚 structure with the odour it will produce, as well as measure how close together molecules are in terms of their smell.
While the researchers only designed the model to link existing odours and molecules, it appears able to accurately predict the odour of molecules that have never been smelled before, too. 鈥淭he neural network seems to be learning some sort of representation of molecules that is more fundamental than what we expected,鈥� says Mainland.
The model was trained on a combination of two flavour and fragrance data sets for more than 5000 molecules. Mainland and his colleagues tested its abilities by getting it to describe how 320 different molecules would smell based on their structure and comparing this against smell-based descriptions from a group of 15 people. The model performed as well as the average person and the researchers suggest it could eventually be used to identify novel scents for use in fragrances and food science.
In a separate piece of work, Mainland鈥檚 colleagues applied the neural network to a data set of chemicals that repel mosquitoes to make a map relating their molecular structures to how repellent the insects find the scents.
This enabled the researchers to identify molecules that the model suggested would be at least as repellent as leading anti-mosquito products, and which could be tested in future trials. This method could, in theory, also be used to find molecules to deter other disease-carrying organisms, as long as training data exists.
鈥淚t is a very impressive piece of work and is an advance on what has already been done to try to connect the molecular structure of volatile compounds to the odour quality humans perceive them as having,鈥� says at the School of Advanced Study, University of London.
鈥淲hat the work does is skip neurobiology and try to connect the structure of molecules to the perception of odours directly,鈥� says Smith. 鈥淚mpressive if it can be done, but we will still have to fill in the biology eventually if we want to understand how humans perceive odours.鈥�
However, the model鈥檚 failure to distinguish between enantiomers, mirror-image molecules that have the same structure but cause vastly different smells, could be problem, he adds.
Mainland acknowledges this and says future work will focus on producing models that can identify enantiomers and more complex mixtures of molecules, rather than just the single molecules the current model works with.
Although the AI model could be used as part of a process to come up with new smells and flavours, says Smith, relying on it alone probably wouldn鈥檛 work.
鈥淭hese results could help us understand odours of single molecules, but most of the odours we smell are mixtures,鈥� says Smith. 鈥淣early all of the smells we are aware of 鈥� wine, coffee, soap, other people, the sea 鈥� are due to a mixture of several hundred volatile molecules.鈥�
There are other factors at play, too. 鈥淓ating food, there鈥檚 the saliva in our mouths, there鈥檚 the taste receptors contributing, the texture of the food,鈥� says Smith. 鈥淢any things are interacting to give you a multi-sensory experience. So I think we are still far away from simply predicting flavour from food molecules.鈥�
References: bioRxiv, DOI: ; DOI:
