Energy based model inference

energy-based-model-inference

Summary

Use a learned energy-based neural network model to score and rank candidate chemical assignments (formula–adduct pairs) for an unknown tandem mass spectrum without relying on spectral databases. This approach learns data-dependent scoring directly from MS/MS fragmentation patterns rather than computing deterministic fragmentation trees.

When to use

When you have an unknown MS/MS spectrum (m/z and intensity pairs) and need to assign a chemical formula and ionization adduct to the precursor mass, particularly when spectrum database lookups are unavailable or when you want to exploit learned patterns in fragmentation rather than rule-based tree decomposition. Useful in de novo metabolomics workflows where the unknown compound is not in existing libraries.

When NOT to use

• Input is negative-mode MS/MS data; MIST-CF currently supports only positive-mode ionization.
• You have a high-confidence spectral library match already available; database lookup is faster and more reliable in that case.
• Input spectrum has < ~5 major fragment peaks; energy-based models may not learn robust patterns from very sparse data.

Inputs

• MS/MS spectrum as m/z–intensity pairs (MGF format or equivalent)
• Precursor m/z value
• Precursor charge state (typically +1 for positive mode)
• MIST-CF trained model weights (PyTorch or equivalent)
• SIRIUS binary or library for candidate formula enumeration

Outputs

• Ranked list of (chemical_formula, adduct_type, energy_score) tuples
• Top-1 or top-k predictions with formula and adduct assignment

How to apply

Load a pre-trained MIST-CF formula transformer model and prepare your MS/MS spectrum as m/z–intensity pairs. Use an internal chemical subformula assignment protocol (implemented via SIRIUS decomp) to enumerate all chemically feasible candidate formulas and adduct types (e.g., [M+H]+, [M+Na]+, [M+K]+ in positive mode) for the observed precursor m/z. Feed each formula–adduct candidate paired with the observed fragment peaks through the formula transformer neural network to compute an energy-based score reflecting agreement between the candidate and spectrum. Rank all candidates by descending score. The model learns to weight fragment patterns and neutral losses in a data-dependent fashion, making it more flexible than fixed fragmentation rules; higher scores indicate better candidate–spectrum fit.

Related tools

• MIST-CF (Core formula transformer model that scores formula–adduct candidate pairs against an observed MS/MS spectrum) — https://github.com/samgoldman97/mist-cf
• SIRIUS (Enumerates candidate chemical formulas via dynamic programming decomposition of the observed precursor m/z) — https://bio.informatik.uni-jena.de/software/sirius/

Examples

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

Evaluation signals

• Top-1 accuracy: Correct chemical formula appears as rank-1 prediction in ≥70–80% of test spectra (baseline dataset-dependent).
• Top-k recall: Correct formula ranked within top-5 or top-10 predictions; compare against SIRIUS or other tree-based baseline methods.
• Score distribution: Higher-ranked candidates should have significantly higher energy scores than lower-ranked ones; check for monotonicity or large score gaps at rank transitions.
• Adduct detection: [M+H]+ and other adduct assignments should match known ionization modes for the dataset (e.g., all [M+H]+ for a standard ESI+ run).
• Consistency across MS instruments: Apply trained model to data from different instrument types (Orbitrap, Q-TOF, etc.); expect stable ranking if the model generalizes well.

Limitations

• Model supports positive-mode ionization only; negative-mode spectra require retraining or a separate model.
• Performance may degrade on high-resolution data (e.g., Orbitrap) if trained on lower-resolution spectra; the paper notes that models trained on public NPLIB1 are 'less performant than the model trained on the commercial NIST20 Library (particularly for Orbitrap or higher resolution data)'.
• Relies on SIRIUS for candidate formula enumeration; if SIRIUS fails to enumerate the true formula, the model cannot rank it.
• Requires sufficient MS/MS peak intensity information; very low signal-to-noise or few peaks may limit model confidence.
• Energy-based scores are not directly calibrated to posterior probabilities; absolute score values are not interpretable without reference to the training distribution.

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] Considering multiple adduct types beyond [M+H]+ (still only positive mode): "Considering multiple adduct types beyond [M+H]+ (still only positive mode)"
• [readme] 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)."
• [other] Prepare input mass spectrum data in the required format (m/z and intensity pairs). Apply the internal chemical subformula assignment protocol to enumerate candidate formulas and adduct types (including [M+H]+ and other positive mode adducts). Feed candidate formula-adduct pairs through the formula transformer neural network to compute energy-based scores.: "Prepare input mass spectrum data in the required format (m/z and intensity pairs). Apply the internal chemical subformula assignment protocol to enumerate candidate formulas and adduct types"