Reducing Structural Ambiguity in Food Compound Identification via Computational Refinement: A Case Study of Catechins and Their Oxidation Products in Chinese Black Tea

Technical Summary: Reducing Structural Ambiguity in Food Compound Identification via Computational Refinement

Problem Statement
The characterization of natural products and their metabolites in complex food matrices, specifically Chinese black tea, is hindered by three primary challenges: isomeric ambiguity, the limited availability of authentic chemical standards, and the presence of novel compounds absent from existing spectral libraries. Black tea contains a dynamic molecular landscape generated by the enzymatic and non-enzymatic oxidation, condensation, and polymerization of catechins. While well-characterized dimers like theaflavins exist, the majority of oxidation products, including polymeric thearubigins and novel intermediates, lack reference standards. Conventional identification workflows relying on spectral library matching (MSI Level 1) fail for these compounds, and standard computational tools often struggle to distinguish between chemically plausible regioisomers when training data is sparse, leading to high structural ambiguity at MSI Levels 3 and 4.

Methodology
The study proposes a confidence-stratified identification workflow aligned with the Metabolomics Standards Initiative (MSI) Levels 1–4, utilizing UHPLC-QTOF mass spectrometry. The approach integrates three distinct strategies based on the availability of structural evidence:

1. MSI Level 1 (Confirmed Identification): For compounds with authentic standards, identification relies on the concurrent matching of retention time (RT), accurate mass (<5 ppm), isotope patterns, and MS/MS fragmentation patterns against an in-house library. This strategy was used to resolve isomeric pairs (e.g., EGCG/GCG) where MS/MS spectra alone were insufficient, relying on chromatographic separation for discrimination.
2. MSI Level 2 (Putative Annotation): For compounds lacking standards but present in public databases, the workflow combines accurate mass querying with the SIRIUS software suite (v6.3.2). This integrates isotope pattern analysis (SIRIUS/ZODIAC) for molecular formula prediction, CSI:FingerID for structure ranking against public libraries (MassBank, COCONUT, PubChem), and CANOPUS for compound class annotation.
3. MSI Levels 3–4 (Tentative Structural Assignment): To address "known-unknowns" (compounds absent from all libraries), the authors developed a pathway-informed in silico strategy.
   • Small-scale refinement: A manually constructed candidate library was created based on known polyphenol oxidase (PPO) regioselectivity (e.g., oxygen insertion at benzotropolone vs. A-ring). Importing these specific structures into SIRIUS constrained the search space, allowing CSI:FingerID to rank chemically plausible regioisomers higher than unrelated candidates.
   • Large-scale generation: A computational pipeline was developed using Python and RDKit to generate a large-scale candidate library (3,272 predicted metabolites). This pipeline utilized two parallel branches: a biotransformation branch (applying oxidoreductase reaction rules to 41 parent polyphenols) and a chemical transformation branch (applying curated thermal processing rules). The resulting candidates were merged, deduplicated, and used to constrain the SIRIUS search space.

Key Results

• Confirmed Compounds: The workflow confirmed 27 compounds at MSI Level 1, including catechins, theaflavins, and related flavonoids. The study demonstrated that for isomeric pairs with identical MS/MS spectra (e.g., EGCG/GCG), chromatographic retention time is the decisive factor for Level 1 confirmation.
• Putative Annotations: 16 oxidation products were annotated at MSI Level 2. The integration of SIRIUS with public libraries successfully identified dimeric and galloylated derivatives (e.g., theasinensins, theacitrins). The study noted that while isomers like theasinensin B and epigallocatechin-(4β→8)-epigallocatechin-3-O-gallate share molecular formulas, differences in MS/MS fragmentation patterns allowed for tentative distinction.
• Reduction of Ambiguity: The pathway-informed candidate lists significantly improved annotation for compounds absent from public databases.
  • In the small-scale case study of theaflavin oxidation, restricting the search to chemically informed regioisomers promoted the correct benzotropolone-oxygenated isomer to Rank 1, whereas an unconstrained search yielded arbitrary rankings.
  • In the large-scale application, the custom pipeline enabled the upgrade of features from MSI Level 4 (molecular formula only) to Level 3 (tentative structure). For example, a feature at m/z 775.1140, which returned no structural candidates in a PubChem search, was assigned a tentative structure (Rank 1) when the custom candidate list was activated.
• Chemical Space Coverage: Analysis of the generated candidate list revealed that the biotransformation and chemical transformation pipelines are complementary. The biotransformation pipeline produced products with high scaffold similarity to parents (diversity ratio 0.199), while the chemical transformation pipeline generated highly diverse scaffolds (diversity ratio 0.744), covering a broader range of chemical space including larger condensation products and isomers.

Significance and Claims
The paper claims that this confidence-stratified workflow provides a systematic route to reduce structural ambiguity in the analysis of processed food matrices where authentic standards are unavailable. By integrating pathway knowledge into computational tools like SIRIUS, the study demonstrates that it is possible to move beyond simple molecular formula assignment (Level 4) to propose tentative structural candidates (Level 3) for processing-derived oxidation products.

The authors emphasize that while the manual construction of candidate lists is limited in throughput, the developed large-scale computational pipeline offers a scalable solution for characterizing "chemical dark matter" in complex food systems. The study positions this framework as a transferable template for the systematic characterization of oxidative polyphenol chemistry in other processed foods, such as oolong tea, coffee, and cocoa. The authors modestly acknowledge limitations, noting that Level 3–4 assignments remain tentative without authentic standard confirmation and that the annotation confidence is inherently constrained by the structural diversity of the training data used by CSI:FingerID.