With the 400 types of olfactory receptors present in olfactory nerve cells, humans can discriminate between as many as one trillion different odors. That keen sense of smell alerts us to dangers like gas leaks, infections, and spoiled food. It also helps to shape our experience of the world around us and adds to the enjoyment of good food. Given how important our sense of smell is to us, it is no wonder that the food, beauty, and wellness industries, for example, would be very interested in having the capability to tailor-make customized scents for use in their products.
Recent work by researchers at the Tokyo Institute of Technology has revealed a method that makes it possible to define a type of odor that one would like to create, and be given a molecular recipe to bring it to life. Using machine learning, they have developed a model that accepts a desired odor impression as an input, and predicts the mass spectrum signature of the molecules that can create it. This specifies all of the molecules that are needed, as well as their individual mixing ratios.
Before designing the machine learning algorithm, the team needed to track down the data that would be used to train it. They started with the mass spectra of 2,106 odorant molecules found in the NIST Chemistry WebBook database. They paired this data with the DREAM dataset, which has odor impression scores, assigned by humans, for various odors. Each odor was assigned one of 21 odor descriptors (e.g.: ‘sweet,' ‘fruit,' ‘fish,' ‘garlic,' ‘spices’). A total of 383 molecules overlapped between the two datasets — that is, there was matching molecular data and odor impression information — so the team chose this merged dataset to train the model.
The data was used to train a deep neural network that predicts the odor impression of a mixture of molecules, as defined by their mass spectrum. An iterative algorithm then seeks to find a mixture that best represents a target odor impression defined by the user. Once the error between the target and the predicted odor impression is sufficiently low, the mass spectrum data can be recovered, which serves as a recipe for creating the custom scent. It was observed that the prediction accuracy reached 0.90, with respect to the correlation coefficient.
To test out their methods, the researchers started with a known scent recipe for an apple odor. They then decided to try to increase the “sweet” and “fruit” components of the scent using their machine learning algorithm. The mass spectrum data produced with this algorithm showed novel peaks that indicate it should be perceived as smelling more fruity and sweet than a typical apple scent. The team indicated that it may have something like hints of fruit punch blended in. However, these scents have not actually been produced, let alone verified by a sensory test, so there is still additional work to be done to fully validate the results.
The team acknowledged that it is not always possible to create a desired target scent, for example, if the input target included both “warm” and “cold” odor impressions, the model would not be able to produce a result to accommodate the request. In any case, with some additional work and validation, this technique may prove invaluable to many industries in the future.