Deep learning enable precision authentication of seasonal and processing signatures in tieguanyin tea.

Abstract

Authenticating specialty tea products remains a critical challenge in premium food markets, yet current analytical approaches are constrained by limited reproducibility and susceptibility to instrumental variation. Here, we present a deep learning framework that transforms liquid chromatography-mass spectrometry (LC-MS) metabolomic data into image representations, enabling robust authentication of tea products under real-world analytical conditions. Profiling 274 Tieguanyin tea samples across seasonal harvests (spring and autumn) and processing methods (light-scented and strong-scented), our approach achieved 90.9% (95% confidence interval [CI]: 80.4%-96.0%) classification accuracy-substantially outperforming conventional multivariate and machine learning methods (sPLS-DA: 85.5%; random forest: 87.3%). Critically, when subjected to chromatographic drift-a pervasive source of analytical irreproducibility-our model maintained 78.2% accuracy while traditional methods degraded to 69.1%. This framework addresses fundamental limitations in untargeted metabolomics, offering a generalizable solution for food authentication that extends beyond tea to broader applications in agricultural product verification and systems biology.