# Tandem Mass Spectrum Prediction Fragment Level

> Use when you have a molecular structure (SMILES, InChI, or chemical formula) and need to predict its collision-induced dissociation (CID) tandem mass spectrum with fragment-level resolution. Use this when chemical-formula-level predictions (e.

- Skill: `holobiomicslab/tandem-mass-spectrum-prediction-fragment-level` (Agent Skill)
- Install (CLI): `npx skillmds@latest add holobiomicslab/tandem-mass-spectrum-prediction-fragment-level`
- Raw SKILL.md: https://api.skillmd.com/api/skills/holobiomicslab/tandem-mass-spectrum-prediction-fragment-level/raw
- Safety review: PASS (external: skill-scanner PASS, skillspector PASS)
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- License: CC-BY-4.0
- Author: HolobiomicsLab (https://skillmd.com/u/holobiomicslab)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/holobiomicslab/tandem-mass-spectrum-prediction-fragment-level

---


# tandem-mass-spectrum-prediction-fragment-level

## Summary

Use the ICEBERG model to predict tandem mass spectra at the level of molecular fragments, decomposing a query molecule into fragment structures and intensity estimates rather than chemical formulas alone. This enables more granular structural elucidation and retrieval-based compound identification.

## When to use

You have a molecular structure (SMILES, InChI, or chemical formula) and need to predict its collision-induced dissociation (CID) tandem mass spectrum with fragment-level resolution. Use this when chemical-formula-level predictions (e.g., SCARF) are insufficient and you need to rank candidate compounds from a library (e.g., PubChem) or interpret fragmentation patterns at substructural detail for structural elucidation campaigns.

## When NOT to use

- If you only need chemical-formula-level predictions (no substructural detail), use SCARF instead, as it is more lightweight and formula-agnostic.
- If your input is already an experimental tandem mass spectrum and you need to identify the compound directly without prediction, use retrieval matching or database search rather than prediction.
- If you lack pretrained model weights and cannot obtain them (requires NIST'20+ license or MassSpecGym public weights), local training requires two GPUs with ≥24 GB RAM each and significant computational time.

## Inputs

- molecular structure (SMILES string)
- molecular structure (InChI string)
- chemical formula
- configuration file (YAML) specifying model checkpoint paths, batch size, GPU devices

## Outputs

- predicted tandem mass spectrum (m/z peaks and intensities)
- fragment assignments (SMILES or InChI for each peak)
- retrieval ranking against reference library (e.g., PubChem candidates ranked by cosine similarity or spectral matching)

## How to apply

Input a molecular structure via the ICEBERG WebUI (http://iceberg-ms.mit.edu/) or via local Python API using the ms-pred repository. The ICEBERG pipeline internally trains a learned fragment generator to propose molecular substructures and their cleavage graph, then uses an intensity predictor (including optional contrastive finetuning on PubChem-SMILES mappings) to estimate peak intensities for each fragment. The model outputs predicted spectrum peaks annotated with fragment SMILES/InChI assignments and intensity values. For GPU-constrained environments, batch size and worker counts can be reduced (e.g., batch_size=8, num_workers=6 on 8GB RAM), or CPU-only inference is feasible by setting cuda_devices=None in the config file. Evaluate retrieval accuracy (top-1 hit rate) on held-out test sets or validate against experimental spectra.

## Related tools

- **ICEBERG WebUI** (user-friendly interface for fragment-level tandem mass spectrum prediction without local GPU or coding) — http://iceberg-ms.mit.edu/
- **ms-pred (coleygroup/ms-pred)** (core Python package implementing ICEBERG fragment generator, intensity predictor, and retrieval pipeline; supports local training and inference) — https://github.com/coleygroup/ms-pred
- **PubChem** (library of chemical structures and reference for retrieval-based structural elucidation; provides formula-to-(SMILES, InChIKey) mappings)
- **MAGMa algorithm** (annotates substructures and labels the molecular breakage process (directed acyclic graph) for ICEBERG training)
- **NIST'20 dataset** (commercial reference spectra with collision energy annotations; used for training and benchmarking ICEBERG (40% top-1 retrieval accuracy reported))

## Examples

```
python src/ms_pred/dag_pred/predict_smis.py --config configs/iceberg/iceberg_elucidation.yaml --smiles 'CC(=O)Oc1ccccc1C(=O)O' --output predictions.json
```

## Evaluation signals

- Top-1 retrieval accuracy on held-out test sets (ICEBERG achieves ~40% on NIST'20 [M+H]+ ions); higher accuracy indicates correct fragment and intensity predictions.
- Cosine similarity between predicted and experimental spectra (peaks and intensities); threshold of >0.7 indicates good match.
- Fragment assignments are chemically plausible: each predicted m/z corresponds to a valid substructure of the input molecule with no negative masses or impossible neutral losses.
- No GPU is required for WebUI inference; CPU-only predictions should complete in <2 minutes on a regular desktop.
- Predicted spectrum peaks fall within the mass-to-charge ratio range expected for the parent ion and plausible fragments (e.g., no peaks exceed parent m/z).

## Limitations

- Pretrained model weights require either a commercial NIST'20+ license or use of the public MassSpecGym weights (which have undergone less manual curation than NIST and will yield different predictions).
- Training ICEBERG from scratch requires two GPUs with ≥24 GB RAM each (e.g., NVIDIA A5000); smaller GPUs require batch size reduction and may skip contrastive finetuning, affecting performance.
- Fragment-level predictions depend on accurate MAGMa annotation of the molecular breakage process during training; errors in substructure labeling propagate to inference.
- Model predictions are collision-energy-dependent; prediction accuracy may vary across different ionization methods or collision energies not well represented in the training set.
- Retrieval performance degrades when query molecules are significantly different from training set chemistry or when the reference library (e.g., PubChem) does not contain true candidates.

## Evidence

- [intro] ICEBERG predicts spectra at the level of molecular fragments, whereas SCARF predicts spectra at the level of chemical formula.: "ICEBERG predicts spectra at the level of molecular fragments, whereas SCARF predicts spectra at the level of chemical formula."
- [readme] You can run ICEBERG structural elucidation easily at http://iceberg-ms.mit.edu/. By inputting the chemical formula and your experimental spectrum, the WebUI will rank it against all candidates from PubChem. No GPU is required.: "You can run ICEBERG structural elucidation easily at http://iceberg-ms.mit.edu/! By inputting the chemical formula and your experimental spectrum, the WebUI will rank it against all candidates from"
- [readme] ICEBERG is our recommended model with a 40% top-1 retrieval accuracy with [M+H]+, benchmarked on the NIST'20 dataset.: "ICEBERG is our recommended model with a 40% top-1 retrieval accuracy with [M+H]+, benchmarked on the NIST'20 dataset."
- [readme] You need two GPUs with at least 24GB RAM to train ICEBERG. If you are trying to train the model on a smaller GPU, try cutting down the batch size.: "You need two GPUs with at least 24GB RAM to train ICEBERG (we used NVIDIA A5000 for development). If you are trying to train the model on a smaller GPU, try cutting down the batch size"
- [readme] ICEBERG is trained in two parts: a learned fragment generator and an intensity predictor.: "ICEBERG is trained in two parts: a learned fragment generator and an intensity predictor. The pipeline for training and evaluating this model can be accessed in `run_scripts/iceberg/`."
- [readme] To train ICEBERG, we must annotate substructures and create a labeled dataset over the breakage process, which we do with the MAGMa algorithm.: "In addition to building processed subformulae, to train ICEBERG, we must annotate substructures and create a labeled dataset over the breakage process, which we do with the MAGMa algorithm."