Multimodal Transformer Fusion of PTR-ToF-MS Volatiles and Targeted Non-Volatile Metabolites Enables Data-Driven Grading of Baimudan White Tea

Abstract

Rationale

Commercial grading of Baimudan white tea still relies largely on sensory evaluation and would benefit from a rapid, objective analytical approach. This study investigated whether PTR-ToF-MS volatile fingerprints combined with targeted non-volatile metabolite measurements could support data-driven grading of four official Baimudan grades.

Methods

Headspace volatiles were profiled by proton-transfer-reaction time-of-flight mass spectrometry (PTR-ToF-MS), and non-volatile metabolites were quantified by HPLC, amino acid analysis, and soluble sugar assay. A total of 120 samples (30 per grade) were divided into a training set (n = 96) and an independent prediction set (n = 24). A two-stream Transformer encoded the volatile and non-volatile modalities separately and fused them for four-class classification.

Results

The multimodal model achieved 95.8% accuracy (23/24) and a macro-F1 score of 0.958 on the independent prediction set, with a single misclassification between adjacent high grades. Repeated stratified five-fold cross-validation within the training set showed stable performance. SHAP and attention analyses indicated that high-grade teas were associated with floral/sweet volatile ions and higher amino acids and soluble sugars, whereas lower grades were associated with greener/woody volatile ions and kaempferol-related markers.

Conclusions

Within the current dataset, integrating PTR-ToF-MS volatile fingerprints with targeted non-volatile metabolites provides an objective and chemically interpretable approach for instrument-assisted grading of Baimudan white tea.