De novo mass spectrum interpretation

de-novo-mass-spectrum-interpretation

Summary

Infer chemical formula and adduct type for an unknown metabolite directly from tandem mass spectra (MS/MS) without referencing spectral databases or computing fragmentation trees. This skill uses a neural network-based formula transformer to rank candidate formulas and assign subformulas to fragment peaks in a data-dependent fashion.

When to use

You have MS/MS spectra (centroided m/z and intensity pairs) and corresponding MS1 precursor masses but lack reference spectra or a priori formula information. Apply this skill when you need to annotate chemical formulas for unknowns in a de novo setting, particularly when rapid ranking of multiple formula–adduct candidates is required without external database lookups.

When NOT to use

• Input spectra are profile (uncentroided) or have very low mass resolution (<5 ppm); MIST-CF expects centroided data with accurate m/z assignments.
• Negative-mode MS/MS data; MIST-CF currently handles only positive-mode ionization.
• A reference database with high-confidence spectral annotations is available and can be queried; database matching may be faster and more reliable than de novo inference.
• Precursor mass is unknown or uncertain by more than ±10 ppm; formula enumeration depends on accurate precursor m/z.

Inputs

• MS/MS spectra (centroided m/z and intensity pairs in .mgf or similar format)
• MS1 precursor mass (or m/z value with charge state)
• Precursor formula candidate set (enumerated via SIRIUS decomp or equivalent)
• Candidate adduct types (e.g., [M+H]+, [M+Na]+, [M+K]+)

Outputs

• Ranked list of chemical formula–adduct assignments per spectrum
• Per-peak subformula predictions with confidence scores
• Energy-based model scores reflecting agreement between formula candidate and spectrum
• Exact-match accuracy and top-k ranking metrics on held-out test set

How to apply

Load MS/MS spectra and precursor masses into the MIST-CF pipeline. Use SIRIUS decomp to enumerate candidate formulas matching the precursor m/z within a specified mass tolerance (typical workflows use 5 ppm or better accuracy). Feed the precursor formula, fragment m/z values, and observed intensities into the formula transformer neural network, which ranks subformula assignments per peak conditioned on the precursor formula and m/z. The transformer learns peak-to-subformula mappings in a data-dependent manner without pre-computing fragmentation trees. Score and rank the candidate formulas using an energy-based model that aggregates peak-level predictions. Evaluate by comparing predicted formulas and adducts against ground truth using exact-match accuracy and top-k ranking metrics.

Related tools

• MIST-CF (Core neural network model for formula transformer inference; ranks formula–adduct candidates and assigns subformulas to peaks.) — https://github.com/samgoldman97/mist-cf
• SIRIUS (Generates candidate formulas via dynamic programming (SIRIUS decomp); enumerates all chemically valid formulas within mass tolerance of precursor.) — https://bio.informatik.uni-jena.de/software/sirius/
• SCARF (Provides sinusoidal formula embeddings used in the MIST-CF transformer architecture.) — https://arxiv.org/abs/2303.06470
• MIST (Original metabolite inference system with spectrum transformer; extended by MIST-CF for de novo formula assignment.) — https://www.nature.com/articles/s42256-023-00708-3

Examples

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

Evaluation signals

• Exact-match accuracy: fraction of spectra for which the top-ranked formula–adduct pair matches the ground truth label.
• Top-k ranking metric: fraction of correct formulas appearing in the model's top-k predictions (e.g., top-5 or top-10).
• Per-peak subformula accuracy: agreement between predicted subformula assignments and reference labels, aggregated across all peaks in the held-out test set.
• Comparison against SIRIUS baseline: MIST-CF should match or exceed SIRIUS ranking performance on the same spectrum set without database filtering.
• Energy-based model score distribution: ranked candidates should show clear separation between correct and incorrect formula–adduct assignments in the score distribution.

Limitations

• Currently supports positive-mode ionization only; negative-mode spectra require separate model training or adaptation.
• Performance may degrade on data from mass analyzers (e.g., Orbitrap, Q-TOF) not represented in the training set; models trained on NPLIB1 may be less performant on high-resolution Orbitrap data than models trained on NIST20.
• Depends on accurate precursor m/z and centroided peak detection; noisy or mis-calibrated spectra will reduce formula enumeration and ranking quality.
• Subformula assignment relies on learned neural network patterns; edge cases with unusual fragmentation modes or highly modified compounds may not generalize from training data.
• Computational cost scales with the number of candidate formulas; spectra with many near-isobaric candidates will require longer inference time.

Evidence

• [readme] 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: "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] Instead of computing fragmentation trees, MIST-CF adopts a formula transformer neural network architecture and 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"
• [readme] Utilizing an internal chemical subformula assignment protocol (rather than SIRIUS fragmentation trees): "Utilizing an internal chemical subformula assignment protocol (rather than SIRIUS fragmentation trees)"
• [readme] annotating MS1 precursor masses from MS/MS data in a de novo setting: "annotating MS1 precursor masses from MS/MS data in a de novo setting"
• [other] MIST-CF implements an internal chemical subformula assignment protocol as an alternative to SIRIUS fragmentation trees, enabling the formula transformer neural network to assign subformulas to peaks in a data-dependent fashion without external fragmentation tree computation.: "MIST-CF implements an internal chemical subformula assignment protocol as an alternative to SIRIUS fragmentation trees, enabling the formula transformer neural network to assign subformulas to peaks"
• [other] For each spectrum, generate per-peak subformula predictions without computing SIRIUS fragmentation trees or querying external spectrum databases.: "For each spectrum, generate per-peak subformula predictions without computing SIRIUS fragmentation trees or querying external spectrum databases"
• [readme] Model output will be saved in quickstart/mist_cf_out/. This model may be less performant than the model trained on the commercial NIST20 Library (particularly for Orbitrap or higher resolution data).: "This model may be less performant than the model trained on the commercial NIST20 Library (particularly for Orbitrap or higher resolution data)"
• [readme] Considering multiple adduct types beyond [M+H]+ (still only positive mode): "Considering multiple adduct types beyond [M+H]+ (still only positive mode)"