A newly developed neural network links the chemical structure of odourants to their smell. There is no universally accepted way to quantify and categorise odours the way we express light in wavelengths and sound pitch in frequencies. A group of US researchers developed a special neural network to create an olfactory map that links a particular molecular structure to a particular olfactory experience. They have published their Principal Odour Map in Science. With this approach, they hope to eventually provide a way to store odours digitally, just like images and sounds.

The researchers started with a map in which a molecule is represented as a graph. They described each atom using criteria such as valence, hybridisation and atomic number; each bond with its number of neighbours, any aromaticity and whether the bond is in a ring structure. They fed the system 5000 odours and let it analyse what correlations there were between certain patterns in the chemical structure and a particular odour, which they plugged into the system as descriptors. Their programme could then associate one or more descriptive words, such as ‘garlic’ or ‘grass’, with an odour. The result was a Principal Odour Map (POM). Using this map, the system can predict how a new molecule will smell.

They compared the performance of their system with that of trained human noses. They had the POM predict how 400 different molecules would smell and then had a panel of trained human testers smell and rate them. The model proved to be as reliable as the average human panelist in describing odour quality.

The POM offers opportunities for research into new odours. The researchers made a list of more than 500,000 possible odours that have never been synthesised and plotted them in the POM to find out how they would smell. This would require 70 person-years of sniffing by trained noses. The next step is to analyse mixtures of odours. Ultimately, the researchers want to find a way to digitise smells, just as we do with images and sound.

Lee et al. (2023) Science https://doi.org/10.1126/science.ade4401