From spectra to certainty: end-to-end workflows for confident metabolite identification

Significance of the topic

The identification of metabolites in untargeted LC–HRMS studies is a central challenge for metabolomics, exposomics, and small-molecule research. Accurate mass measurements generate large candidate lists, but orthogonal evidence (MS/MS fragmentation, adduct and isotope patterns, chromatographic behavior, and chemical-context relationships) is required to convert features into defensible annotations. Integrating acquisition strategies that preserve precursor integrity with advanced data-processing and AI/ML scoring improves decision boundaries between isomeric and near-isomeric candidates, enabling reproducible, reportable confidence levels for each annotation.

Objectives and overview of the study

The work demonstrates an end-to-end workflow that moves untargeted metabolomics from feature detection toward evidence-ranked metabolite annotation. The goals were to (1) exploit intelligent data acquisition to maximize useful MS/MS information, (2) combine multiple orthogonal lines of evidence in software, (3) apply AI/ML confidence scoring to separate hard cases such as positional isomers, and (4) produce automated, Schymanski-style confidence-level reporting for transparent results. The example case centers on distinguishing methylxanthine positional isomers and resolving caffeine from hundreds of exact-mass candidates in a human plasma reference material.

Methods and experimental design

Samples: NIST SRM 1950 human plasma extract prepared by methanol protein precipitation, centrifugation, evaporation, and reconstitution; injected plasma equivalent 4.583 µL per injection.

Chromatography and acquisition: Reversed-phase UHPLC using a C18 column (Hypersil GOLD VANQUISH, 150 × 2.1 mm, 1.9 µm) on a Vanquish Horizon system. Mobile phases were water and methanol, both with 0.1% formic acid; flow 0.30 mL/min; an 18 min gradient.

Mass spectrometry: Orbitrap Excedion operated in positive ion mode. Full-scan MS1: m/z 67–1000 at 120,000 resolving power with Enhanced Dynamic Range (eDR) using 16 auto-defined windows. MS/MS: AcquireX iterative Deep Scan workflow for targeted iterative fragmentation across replicate injections; stepped HCD collision energies 10%, 35%, 60%; MS2 resolution 30,000. Data processing employed Compound Discoverer 3.5, Thermo FreeStyle 1.8 SP2, and spectral libraries including mzCloud and mzVault; structure lookups via ChemSpider; molecular networking and an AI/ML confidence-scoring node were used to integrate evidence and assign confidence levels.

Used instrumentation

• Thermo Scientific Vanquish Horizon UHPLC System.
• Thermo Scientific Hypersil GOLD VANQUISH C18 UPLC Column (150 × 2.1 mm, 1.9 µm).
• Thermo Scientific Orbitrap Excedion Mass Spectrometer with OptaMax Plus ion source and features to reduce in-source fragmentation (labile-compound-optimized ion funnel, ion routing multipole, HCD fragmentation, enhanced dynamic range).
• Software: Thermo Scientific Compound Discoverer 3.5, FreeStyle 1.8 SP2; mzCloud and mzVault spectral libraries; ChemSpider for structure searches.

Main results and discussion

Accurate mass as a lone discriminator is insufficient: a monoisotopic mass search for caffeine [M+H]+ returned 920 candidate structures within ~5 ppm, illustrating high recall but low specificity of mass-only approaches. Incremental orthogonal evidence refined annotations: isotope pattern matching supported elemental composition, identification of sodiated adducts supported a single neutral entity, and high-quality MS/MS matched reference fragmentation for caffeine.

Positional-isomer resolution: For methylxanthine isomers (1-, 3-, and 7-methylxanthine) that share precursor m/z 167.05655 [M+H]+, traditional spectral-similarity metrics produced similar scores across candidates and sometimes ranked incorrect isomers highly. Visual inspection of mirror plots indicated 1-methylxanthine provided the best diagnostic fragment agreement. The AI/ML confidence score quantified this expert judgment: it assigned a high score (86.1) to 1-methylxanthine and low scores (~3.5–3.6) to 3- and 7-methylxanthine, thereby widening the decision boundary and enabling a defensible top annotation.

Molecular networking provided chemical-context evidence by grouping caffeine and related xanthine metabolites into a coherent subnetwork, reinforcing annotation plausibility via consistent transformation and fragmentation relationships. Automated conversion of integrated evidence into standardized confidence levels (following Schymanski-style reporting) allowed transparent, per-feature confidence statements.

Benefits and practical applications

• Improved confidence in untargeted annotation: Combining MS1, adduct/isotope patterns, MS/MS, molecular networking, and AI/ML scoring narrows candidate lists to single defensible annotations in many cases.
• Better handling of isomeric challenges: AI/ML scoring amplifies subtle but diagnostic MS/MS differences that conventional similarity scores may miss, reducing false positive isomer assignments.
• Traceable, reportable outputs: Automated confidence-level assignment and evidence ranking facilitate transparent reporting, reproducibility, and regulatory/QA workflows.
• Efficient acquisition: AcquireX iterative Deep Scan targets sample-specific precursors across injections, increasing MS/MS coverage without manual inclusion lists.

Future trends and potential applications

• Broader adoption of evidence-integration frameworks: Combining AI/ML scoring with molecular networking and curated libraries will become standard practice in untargeted metabolomics to improve annotation reliability.
• Model and library expansion: Larger, higher-quality reference spectral libraries and improved ML models trained on isomeric cases will increase discrimination power for positional isomers and stereoisomers.
• Automated decision support: Enhanced tools will translate multi-evidence scores into standardized reporting formats accepted across laboratories and studies, aiding meta-analysis and database curation.
• Acquisition innovations: Iterative and intelligent acquisition schemes that preserve precursor integrity and maximize informative fragmentation will further improve downstream annotation accuracy.

Conclusions

The presented Orbitrap Excedion–based workflow demonstrates that integrating intelligent acquisition, orthogonal MS1/MS2 evidence, molecular networking, and AI/ML confidence scoring markedly improves metabolite annotation confidence compared to mass-only approaches. In challenging cases such as methylxanthine positional isomers and caffeine candidate reduction, the combined approach reduces ambiguity, produces defensible single annotations, and enables standardized confidence-level reporting for untargeted metabolomics studies.

References

1. Schymanski EL, Jeon J, Gulde R, Fenner K, Ruff M, Singer HP, Hollender J. Identifying Small Molecules via High Resolution Mass Spectrometry: Communicating Confidence. Environ. Sci. Technol. 2014;48(4):2097–2098.
2. Krakko D, Tautenhahn R, Stutts WL. Implementing Annotation Confidence Scoring in Untargeted Mass Spectrometry Workflows for Small Molecule Analysis. Anal. Chem. 2026;98:10352–10359.