Credit: Rockefeller Univ.

Algorithms can predict a molecule's odour on the basis of its chemical structure.

Pablo Meyer at IBM's Computational Biology Center in Yorktown Heights, New York, and his colleagues, asked 49 people to smell hundreds of molecules (pictured) and rate them on intensity, pleasantness and 19 other descriptors, such as 'fruit', 'musky' and 'bakery'.

The researchers gave these ratings, along with information on the substances' chemical structures, to 22 teams of computational scientists, who competed to build the best predictive, machine-learning algorithms. After initially developing and training their algorithms on a partial data set, the teams tested their algorithms' abilities to predict people's perception of the remaining molecules.

Across all models, 'garlic' and 'fish' were the best-predicted attributes, at about 70% accuracy. Such tools could be used by the flavour and fragrance industry to formulate products, the authors say.

Science 355, 820–826 (2017)