Machine Learning and Metabolomics for Predicting Nutrient and Phytochemical Bioavailability: A Systematic Review

David Anih and Kayode Arowora

This systematic review examined the integration of food composition data, processing conditions, in vitro digestion, untargeted metabolomics and artificial intelligence for predicting the bioaccessibility and bioavailability of nutrients and phytochemicals in food matrices. The literature search period spanned 2020 to 2025. Studies were retrieved from major biomedical, food science, chemical and computational databases and were eligible when they combined at least two of the three core elements of the review: in vitro digestion, untargeted metabolomics and machine learning or artificial intelligence. Following duplicate removal and two-stage screening, 40 studies met the inclusion criteria, with 21 forming the main evidence base for synthesis. The evidence showed that INFOGEST-based static digestion models remain the most widely used framework, although semi-dynamic and dynamic systems offer greater physiological realism. Untargeted liquid chromatography, high-resolution mass spectrometry and gas chromatography mass spectrometry were the principal analytical platforms, but variability in preprocessing, annotation confidence, quality control and reporting standards limited direct comparison across studies. Food processing methods such as heating, fermentation, milling and non-thermal technologies consistently influenced the bioaccessibility of carotenoids, polyphenols, glucosinolates and other bioactive compounds in a matrix-dependent manner. Machine learning approaches, including random forest, gradient boosting, neural networks, graph neural networks and physics-informed models, demonstrated promising capacity to predict bioaccessibility, metabolite transformation and semiquantitative concentration patterns. However, performance frequently declined during external validation, indicating overfitting, instrument effects and domain shift. Overall, the review indicates that artificial intelligence-supported prediction of nutrient and phytochemical bioavailability is scientifically feasible, but the field is not yet fully mature for broad translational use. Progress will depend on standardized benchmark datasets, harmonized digestion and metabolomics protocols, stronger quality control, transparent reporting, independent external validation and wider adoption of FAIR and reproducible workflows. These advances should support more reliable prediction systems for nutrition research, food processing optimization and industrial formulation.

View Fulltext Back