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. Meanwhile, the lack of information regarding the flavored molecules and specific flavoring properties is an obstacle to advances in this sector. In this context, this work presents the implementation of three scientific machine learning techniques as an innovative methodology to design new natural flavor molecules with specific desired properties to product development. The transfer learning technique is presented to tackle the lack of data available when analyzing flavor molecules. Nine flavor descriptors were studied along this work, and all of them presented more than 50% of 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 flavor component impact for the specific flavors’ results.

Molecule Generation and Optimization for Efficient Fragrance Creation

“…A similar methodology was described in previous published works. 45,46 • Selection of Best-Fitting Molecules: Next, we focus on selecting molecules that align with the target fragrance profile. This step involves estimating the vapor pressure of the generated molecules and strategically choosing those with vapor pressures that closely match the original molecules.…”

FlavorMiner: A Machine Learning Platform for Extracting Molecular Flavor Profiles from Structural Data

“…Similar to several other knowledge domains, different Machine Learning (ML) models have been developed to perform in-silico flavor prediction from molecular structures with available flavor profiles 3 . This trend has been prompted by advancements in ML algorithms and the availability of large-scale molecular data [15][16][17] . This approach enables an efficient screening of potential flavors to prioritize compounds for validation using traditional experimental methods, saving time and resources 2,3,18 .…”

“…These flavor notes are of capital importance in fermented food processing (including coffee, beer, wine, chocolate, bread and others) 2 . Also, some models based on Generative Artificial Intelligence have been trained to generate new molecules with flavors potentially interesting for the food industry (including the abovementioned flavors), but they lack any classification capability 4,17 . A major challenge for predicting these flavor notes is the class imbalance, as the number of positive examples is significantly lower than the negatives 2,25 .…”