SAN2020

Unlike vision and hearing, the sense of smell remains poorly understood, difficult to model, and even harder to predict. Despite the vital evolutionary role of olfaction, crucial to identify safe food sources or detect nearby fires, we still lack a reliable and principled method to predict the smell of a molecule from its chemical structure: the only reliable way to determine the odour associated with a certain compound is to smell it. We propose a deep learning architecture to predict the odour of molecules from its representation as a chemical graph. The application of sequential convolutional operations is capable of identifying relevant patterns at multiple scales, allowing us to find important relationships that have eluded previous machine learning strategies based on more macroscopic approaches. We validated this approach using data from FlavorDB, a website with almost 26 thousand molecules associated with their respective tastes and smells. After representing the adjectives used to describe tastes and odors as semantic vectors, we were able to predict the projection values ​​in a reduced latent space by training our models using the chemical graphs of the respective molecules, outpeforming previous machine learning algorithms. The computational prediction of odours has potential to accelerate the design of new molecules to be used as fragrances, as well as to unveil the neurobiological principles associated with information processing in the olfactory system.