Machine Learning Assisted Odor Assessment Based on Molecular Structures

The sense of smell is known to significantly influence human experiences, affecting food and product choices, memories and moods. This study explores the intricate relationship between molecular structures and the aromatic properties of molecules through the application of advanced machine learning (ML) techniques and emphasizes the necessity of accurately predicting the odors of both singular molecules and complex mixtures, particularly in industries such as cosmetics and food, where the development of new aromas is crucial. This research explores the inherent complexities in modeling olfactory properties, recognizing that molecular structures do not always correlate directly with the scent that the molecule elicits. For instance, closely related molecules, i.e., molecules with similar structures may evoke vastly different odors. This challenge underscores the necessity for sophisticated ML algorithms capable of discerning intricate relationships between molecular features and their corresponding olfactory properties, for both, mono-molecules and complex mixtures. This can also be extended to other molecular properties, for example, potential toxicity of molecules especially for food and cosmetic products. Thus, this dissertation specifically addresses the predictive modeling of aroma and potential toxicity of molecules, highlighting the importance of an integrated early safety and property assessment system in product development. A significant contribution of this work includes the development of three machine learning algorithms: the Olfactory Weighted Sum (OWSum), which classifies molecules based on their structural features; a 2D Convolutional Neural Network (CNN) for analyzing and predicting aroma of complex whisky mixtures; and a modified 3D UNet model (eDen) focused on predicting potential toxic nature of a molecule using 3D electron densities. These models leverage various molecular representations, including 3D electron densities, to enhance their predictive accuracies. The findings demonstrate that ML approaches can effectively streamline product development by allowing early detection of unsuitable substances and reducing trial-and-error during formulation. Furthermore, this research contributes to a deeper understanding of how molecular structures influence odor perception, potential toxicity and how these feature interactions can be captured using ML tools, ultimately aiming to facilitate the responsible and efficient creation of new aroma products while adhering to safety regulations. Overall, this dissertation positions machine learning as a transformative tool in the domain of olfactory science, paving the way for future innovations in aroma assessment and safety evaluation.