A survey of machine learning techniques in flavor prediction and analysis

Abstract

Background

Flavor prediction—the computational estimation of a compound's sensory profile from volatile and non-volatile compounds—is increasingly important for innovation in food, beverage, fragrance, and pharmaceutical industries. Traditional methods such as Gas Chromatography–Mass Spectrometry (GC–MS) and sensory evaluation remain valuable but face limitations in scalability, subjectivity, and cost. Taste perception, mediated by receptor–ligand interactions, fundamentally shapes the multisensory experience of flavor.

Scope and approach

This review examines diverse machine learning (ML) and deep learning (DL) approaches applied in flavor prediction, benchmarking their performance across sweet, bitter, umami/kokumi, pungency, and multi-class taste predictors. It spans traditional ML algorithms (SVM, RF, KNN) and advanced DL models (CNNs, RNNs, GANs, and LLMs), with additional focus on hybrid and ensemble strategies. A particular emphasis is placed on molecular representation (SMILES, SELFIES, fingerprints), the integration of multi-omics data (genomics, proteomics, and metabolomics) for a holistic understanding of flavor formation, genotype–phenotype–flavor relationships, and the crucial role of taste receptor information for enhancing model interpretability.

Key findings and conclusions

Model performance varies by taste modality: gradient boosting and consensus models excel for sweetness, SVMs and Graph Neural Networks (GNNs) for bitterness, and CNN + GNN hybrids for multi-class classification. Kokumi and umami benefit from GNN embeddings, iUmami-SCM, and transformer-based models with experimental validation, while pungency prediction—(though underexplored)-shows promise with ensembles and quantum chemistry-informed features. Challenges remain, including limited high-quality datasets, underfitting on novel molecules, and the interpretability gap in deep learning. Advances such as SELFIES and receptor–ligand–aware models (e.g., BitterX, BitterTranslate, KokumiPD) address these issues by ensuring structural validity and enhancing mechanistic insights.

Significance and novelty

The review identifies future directions in explainable AI, IoT-enabled data collection, multi-omics fusion, molecular representations, and hybrid modeling—offering a roadmap for scalable, interpretable flavor prediction in next-generation food, fragrance, and pharmaceutical innovation.