A Transfer Learning approach to develop natural molecules with specific flavor requirements

Abstract

In the past decades, the flavor industry's investment in research and development has increased to take innovative steps. Therefore, a new field to acknowledge the flavor industry challenges and concerns has arisen, developing innovative tools for the area of flavor engineering. Meanwhile, the lack of information and datasets regarding the flavored molecules and specific flavorings properties are obstacles to advances in this sector. In this context, this work presents the implementation of three Scientific Machine Learning techniques as an approach to specify flavoring characteristics in newly designed molecules. Therefore, this work brings an innovative methodology to design new natural flavor molecules with specific desired properties to product development. The Transfer Learning technique is presented, alongside a deep generative and a deep reinforcement learning models, to tackle the lack of data available when analyzing and studying flavor molecules and developing flavor-based products. This work brings as contributions the utilization of a web scrapper code to sample specific flavors’ databases, apply a generative model as well as a reinforcement learning one in a transfer learning context, integrates three Scientific Machine Learning techniques in a complex system as a framework, and approaches the transfer learning model training one-by-one keeping the parameters constant but training the neural networks specifically for each case. The deep transfer learning implementation in this purpose presented excellent results, regarding the generation of molecules based on specific flavor descriptors. Nine flavor descriptors were studied along this work and all of them presented more than 50% of new molecules generated within the outstanding results considered for the evaluation metric, Natural Product Likeness Score and Synthetic Accessibility Score. Finally, a discussion of the results is constructed based on the data availability, the presence in nature, and the multisensorial components of flavor impact for the specific flavors’ results.

Geom-SAC: Geometric Multi-Discrete Soft Actor Critic With Applications in De Novo Drug Design

“…This implies a notion of creativity in designing or optimizing molecules. In contrast to other approaches that use transfer learning ( [31], [32]) may suffer from limited creativity due to the pre-trained layers.…”

Scents of AI: Harnessing Graph Neural Networks to Craft Fragrances Based on Consumer Feedback

“…Research from other domains reveals that computer-aided molecular design can predict how a molecule relates to a target property (Queiroz et al, 2023c(Queiroz et al, , 2023b(Queiroz et al, , 2023a. This revelation unveils new avenues for the perfume sector.…”