FlavorMiner: a machine learning platform for extracting molecular flavor profiles from structural data.

Area of Science:

• Food Science
• Computational Chemistry
• Bioinformatics

Background:

• Consumer acceptance of food products is primarily driven by flavor.
• Identifying flavor compounds in complex food matrices is challenging and costly.
• Machine learning (ML) offers a promising alternative for predicting flavor features.

Purpose of the Study:

• To develop and validate FlavorMiner, an ML-based multilabel predictor for molecular flavor features.
• To optimize ML approaches for flavor prediction by evaluating algorithm and representation combinations.
• To address class imbalance issues in flavor datasets.

Main Methods:

• FlavorMiner integrates various ML algorithms (Random Forest, K-Nearest Neighbors) with molecular descriptors (Extended Connectivity Fingerprint, RDKit).
• Class balance strategies, including resampling and weight balancing, were employed.
• The model's performance was evaluated using ROC AUC scores.

Main Results:

• FlavorMiner achieved an average ROC AUC score of 0.88, demonstrating high predictive accuracy.
• Combinations of Random Forest/K-Nearest Neighbors with Extended Connectivity Fingerprint/RDKit descriptors performed best.
• Resampling strategies were more effective than weight balancing for mitigating class imbalance.

Conclusions:

• FlavorMiner provides an accurate and efficient method for predicting molecular flavor features.
• The tool can analyze complex food metabolomics data, such as in cocoa.
• FlavorMiner offers a scalable solution for flavor mining across diverse food products, advancing food science and industry practices.