Machine learning-driven prediction of chemical compound sweetness based on molecular descriptors

Taste is a complex sensation related to the perception of food flavours, acting as a key control mechanism to defend the organism from poisons or noxious substances: the five basic taste sensations, namely bitter, sour, salty, sweet and umami, evoked either alone or in combination by food molecules, may stimulate the intake or rejection of the latter. Sugars (e.g., glucose, sucrose) constitute a prominent example of this mechanism, inducing a powerfully motivating sensory response toward food in light of their role as one of the primary energy sources for the body. On the other hand, the well-known correlation between sugar abuse and diseases development, e.g., diabetes, obesity or cardiovascular problems, is driving the tendency to substitute natural sugars with low-calorie sweeteners, to promote an overall healthier lifestyle. This consideration is spiking the interest in the development of artificial sweet molecules with a low calorie content but a strong sweetening ability, a process that can greatly benefit from the molecular-level identification of key chemical and physical features that ultimately result in a specific gustatory sensation. In this context, machine learning (ML) models can represent fast and readily deployable tools for the discovery of novel sweet compounds. With this goal in mind, in the present work, a novel set of both open-source and proprietary relevant molecular descriptors were extracted, starting from a database of 316 sweet molecules. After a thorough assessment of the state of the art, two ML models were constructed for the evaluation of the level of sweetness of a given query molecule. Given the importance of assuring the reliability of the predicted result, which can be jeopardized by the limited number of available molecules with known sweetness levels used to construct the models, an applicability domain of the sweetness predictor was also developed and is reported, taking the specific usage scenario into account. Overall, the present work provides a solid starting point for the future refinement of regressors as molecular predictors of the sweetness level of novel chemical compounds.