The review analysed 53 peer-reviewed studies published from 2015 to 2025 across AI, bioactive compounds, metabolomics, consumer modelling and functional-food optimisation.

Explainable AI adds a layer of interpretation: rather than returning only a prediction, it identifies variables that contributed to the result. Experts can then compare the output with biological plausibility, literature and formulation knowledge.

In an industrial workflow, XAI can help prioritise candidates, document why a combination was selected and detect signals of dependence on biased data or statistical shortcuts.

Interpretability is not the same as truth. Many models remain in silico, and an explanation may describe algorithm behaviour without proving causality. External data, experiments, controls and human review remain necessary.

Why it matters

For R&D, quality and regulatory teams, interpretability can turn an algorithmic score into a documented, challengeable hypothesis that is easier to test.

Evidence limit

The review notes that many applications remain in silico and lack experimental or clinical validation. A model explanation does not establish that a relationship is causal or biologically correct.

Editorial note. This content is intended for industry professionals. It does not constitute medical advice, therapeutic guidance or regulatory advice. Read our editorial method.

Read the original source · Food Chemistry: X ↗