The sense of smell plays an important role in associating products such as fragrances and detergents with their brands for consumers. Consumers want new products with characteristic odors that are harmless to health and are environmentally sustainable. Therefore, it is crucial to identify substances that can be used for such products early in the development process to reduce costs. Machine learning (ML) can be utilized to leverage structure information of molecules and mixtures to predict their properties such as odors. This is a challenging problem that often requires specialized knowledge. Here, we present ML approaches we developed to classify and assist in odors assessment based on molecular structures that can be used in early discovery to identify molecule candidates. For this, we use the DREAMS Olfaction Challenge dataset and an in-house dataset consisting of molecule mixtures found in 16 whiskey samples. In previous work by Haug et al. 2023, the whiskey samples were analyzed by a semi-automated in-house software (Grasskamp et al. 2023) and their perceived aromas were acquired through a sensory panel. We created a data processing pipeline that uses a corpus of reference molecules to identify, extract and vectorize molecular features to train a convolutional neural network and predict the top 5 applicable whiskey aromas with an average accuracy of 85.3%. Moreover, building upon our previous work using 2D Olfactory Weighted classification algorithm – OWSum (Schicker et al. 2023), we also present a novel ML approach that uses 3D-electronic density distribution of molecules to learn their structures and infer properties such as their odor classes. We show that this approach can be further extended to other properties of interest for cosmetic industry such as classification of a molecule as a health hazard.
Funding: The research was funded by the Bavarian State Ministry for Economic Affairs, Regional Development and Energy via the project “Campus of the Senses”