Tandem Ms Data Annotation And Curation

tandem-ms-data-annotation-and-curation

Summary

Assign chemical formulas and ionization adducts to tandem mass spectra (MS/MS) using energy-based neural network scoring without spectrum database reference. This skill ranks candidate formula–adduct pairs for unknown precursor masses by learning from fragmentation patterns in an end-to-end fashion.

When to use

You have an unknown tandem MS/MS spectrum (precursor m/z and fragment peaks) and need to infer the molecular formula and ionization mode (e.g., [M+H]+, [M+Na]+, [M+K]+, [M+NH4]+) in a de novo setting without access to spectral libraries. Apply this skill when you need to rank multiple plausible formula–adduct hypotheses and score their consistency with the observed fragment pattern.

When NOT to use

• Spectrum is in negative ionization mode (MIST-CF currently supports only positive mode; negative mode ionization is not yet supported)
• You have high-confidence spectral library matches already and only need confirmation rather than de novo ranking
• Input spectrum is a single MS1 (precursor mass only) without MS/MS fragment peaks; MIST-CF requires tandem mass spectrometry data

Inputs

• Tandem mass spectrum (precursor m/z and fragment peaks in MGF format)
• Precursor neutral mass (derived from m/z and charge state)
• Trained MIST-CF formula transformer model checkpoint
• Known adduct type(s) for hypothesis enumeration (positive mode: [M+H]+, [M+Na]+, [M+K]+, [M+NH4]+)

Outputs

• Ranked list of candidate formula–adduct pairs with scores
• Top-ranked chemical formula and adduct assignment
• Ranking metrics (rank of ground-truth adduct, recall@k)
• Energy-based model score reflecting spectrum–formula agreement

How to apply

Load a trained MIST-CF formula transformer model and prepare your spectrum in MGF format with precursor m/z and MS/MS fragment peaks. Use the internal chemical subformula assignment protocol to enumerate candidate formulas for the precursor mass (rather than SIRIUS fragmentation trees). For each candidate formula, consider multiple positive-mode adduct types ([M+H]+, [M+Na]+, [M+K]+, [M+NH4]+) to generate formula–adduct hypotheses. Score each hypothesis by passing the spectrum peaks and candidate formula through the transformer model to predict consistency. Rank hypotheses by score and report rank of ground-truth adduct, recall@k metrics, and the top-ranked formula–adduct assignment as the annotation. The model learns to score agreement between precursor formula candidate and spectrum in a data-dependent fashion without relying on fragmentation tree decomposition.

Related tools

• SIRIUS (Used to enumerate candidate chemical formulas via dynamic programming (SIRIUS decomp) for a given precursor m/z; MIST-CF uses an internal subformula assignment protocol instead of SIRIUS fragmentation trees for ranking) — https://bio.informatik.uni-jena.de/software/sirius/
• MIST-CF model (Trained formula transformer neural network that scores and ranks formula–adduct hypotheses against tandem MS/MS spectra) — https://github.com/samgoldman97/mist-cf

Examples

. quickstart/download_model.sh && . quickstart/run_model.sh

Evaluation signals

• Rank of ground-truth formula–adduct pair in the scored candidate list; lower rank indicates better calibration
• Recall@k metric: fraction of spectra for which correct formula–adduct pair appears in top-k predictions
• Comparison of ranking accuracy against baseline single-adduct mode [M+H]+ to quantify improvement from multi-adduct consideration
• Spectrum–formula consistency score magnitude and distribution across correct vs. incorrect candidates; non-overlapping distributions indicate discrimination power
• Schema validation: all output pairs have valid chemical formulas, plausible adduct types, and scores in expected numeric range

Limitations

• Currently supports only positive ionization mode; negative mode ionization is not yet supported
• Model performance may degrade on high-resolution instruments (e.g., Orbitrap) if trained only on public NPLIB1 dataset; NIST20-trained models are more performant but require license access
• Depends on SIRIUS for initial formula candidate enumeration; performance is upper-bounded by SIRIUS decomposition accuracy
• Assumes sufficient MS/MS peak information (experiments show few peaks are sufficient, but very low-abundance or sparse spectra may reduce ranking confidence)

Evidence

• [readme] ranks chemical formula and adduct assignments for an unknown mass spectrum using an end-to-end energy based modeling approach, without referencing any spectrum databases: "MIST-CF ranks chemical formula and adduct assignments for an unknown mass spectrum using an end-to-end energy based modeling approach, without referencing any spectrum databases."
• [readme] Considering multiple adduct types beyond [M+H]+ in positive mode to improve ranking: "Considering multiple adduct types beyond [M+H]+ (still only positive mode)"
• [readme] Uses an internal chemical subformula assignment protocol rather than SIRIUS fragmentation trees: "Utilizing an internal chemical subformula assignment protocol (rather than SIRIUS fragmentation trees)"
• [readme] Formula transformer neural network architecture that learns in a data-dependent fashion: "Instead of computing fragmentation trees, MIST-CF adopts a formula transformer neural network architecture and learns in a data dependent fashion."
• [other] Test dataset should contain tandem mass spectra annotated with known non-protonated adduct types: "prepare a test dataset of tandem mass spectra annotated with known non-protonated adduct types (e.g., [M+Na]+, [M+K]+, [M+NH4]+)"
• [other] Compute ranking metrics including rank of ground-truth adduct and recall@k: "Compute ranking metrics (e.g., rank of ground-truth adduct, recall@k) and collate results in a structured table."