- ICICLE
- Recommended checkpoint
- Repository layout
- Usage
- Model details
- Results
- Spectral similarity (full test set; Tables S9 and S11)
- Formula-match retrieval (50 same-formula PubChem isomers, entropy re-ranking; Tables S21 and S23)
- Full-PubChem retrieval (87,260,989 candidates, no formula, cosine ranking, seed 1 only; Tables S33, S34 and S37)
- RASSP
- Comparison with quantum chemistry (Table S13)
- ICICLE hyperparameters (Table S4)
- Limitations
- Citation
- Recommended checkpoint
ICICLE
Inferring Cleavage In silico via physics-Constrained Learning for EI-MS.
ICICLE is a fragment-grounded neural simulator for 70 eV electron ionization (EI) mass spectra. It works in two stages. First, MAGMa enumerates the molecule's full fragmentation graph (up to 3 bond cleavages, with hydrogen shifts of ±6). Then a Set Transformer assigns an intensity to each fragment, and a differentiable module expands every peak into its natural isotope envelope. Because every predicted peak comes from an explicit substructure, the predictions can be interpreted fragment by fragment.
This repo holds the trained ICICLE checkpoints and the retrained baselines (NEIMS, MassFormer, RASSP) used in the paper. All models were trained on the same NIST23 random and scaffold splits, with 3 seeds per split.
License notice: All checkpoints were trained on the NIST23 GC/EI-MS library. Please distribute these checkpoints only to people who hold a valid NIST23 license.
- Paper: Lederbauer, Wang, Coley. Fragment-Grounded Neural Simulation of Electron Ionization Mass Spectra at Library Scale. Analytical Chemistry (2026). doi:10.1021/acs.analchem.6c06136 (preprint: ChemRxiv)
- Code: https://github.com/coleygroup/icicle
- Predicted library for 87.3M PubChem compounds: Zenodo 10.5281/zenodo.22854374
Recommended checkpoint
icicle/random/s1/best.ckpt is the recommended default for general use, including new
chemistry. It was trained on the random split, which covers the broadest range of training
scaffolds, and it is the same checkpoint used to build the 87.3M-compound PubChem library on
Zenodo. The scaffold-split checkpoints are mainly for reproducing the out-of-distribution
benchmarks in the paper.
Repository layout
icicle/{random,scaffold}/s{1,2,3}/best.ckpt # ICICLE intensity predictor (PyTorch Lightning)
icicle/{random,scaffold}/s{1,2,3}/config.yaml # training config
neims/{random,scaffold}/s{1,2,3}/best_model.pt
massformer/{random,scaffold}/s{1,2,3}/chkpt.pkl
rassp/{random,scaffold}/s{1,2,3}/<run>.model + <run>.meta
s1–s3 are the three training seeds. All metrics in the paper are reported as the mean over these three seeds.
Usage
Loading the checkpoints requires the ICICLE source code:
git clone https://github.com/coleygroup/icicle && cd icicle
uv sync --extra cu124 # or: uv sync --extra cpu
uv pip install -e .
Command line
hf download CRG-MIT/ICICLE icicle/random/s1/best.ckpt --local-dir ckpts
uv run src/icicle/main.py \
--intensity-predictor ckpts/icicle/random/s1/best.ckpt \
--smiles "CCO" "BrCC(=O)NC1CCCCC1" \
--plot --output-format csv --output-folder predictions
Python
from huggingface_hub import hf_hub_download
from icicle.models.eims_predictor import EIMSPredictorFromFullEnumeration
ckpt = hf_hub_download(repo_id="CRG-MIT/ICICLE", filename="icicle/random/s1/best.ckpt")
model = EIMSPredictorFromFullEnumeration.load_from_checkpoint(intensity_predictor_checkpoint=ckpt)
spectrum = model.predict_from_smiles("BrCC(=O)NC1CCCCC1")
Large libraries (multi-GPU)
uv run src/icicle/batch_infer.py \
--intensity-predictor ckpts/icicle/random/s1/best.ckpt \
--input compounds.csv --smiles-col smiles \
--output predictions.hdf5 --num-gpus 8 --checkpoint-every 500
Prediction runs at about 100 spectra per second on a single RTX 2080 Ti.
Baselines
Each baseline runs in its own environment. Setup, inference and evaluation instructions
are in baselines/ of the code
repo, and every baseline writes predictions to the same HDF5 format, which
src/icicle/eval_from_predictions.py then scores.
Model details
- Input: a SMILES string (molecular mass < 750 Da; elements As, B, Br, C, Cl, Co, F, Fe, H, I, K, N, Na, O, P, S, Se, Si; stereochemistry is ignored).
- Output: a unit-resolution EI spectrum with 750 bins (m/z 1–750) that includes the isotope envelopes.
- Training data: NIST23 GC/EI-MS main library, 333,228 spectra after filtering, split 80/10/10 into train/validation/test as either a random split or a Bemis–Murcko scaffold split.
- Loss: spectral entropy distance. Optimizer: AdamW. Hyperparameters were chosen with Optuna TPE search.
Results
All results are on the NIST23 test set and are given as the mean ± 95% CI over 3 seeds, unless a table says otherwise. Table numbers refer to the paper's SI.
| Split | Train | Val | Test |
|---|---|---|---|
| Random | 266,361 | 33,331 | 33,536 |
| Scaffold | 266,169 | 33,354 | 33,705 |
Spectral similarity (full test set; Tables S9 and S11)
| Split | Model | Cosine | Entropy | Weighted cosine | Composite |
|---|---|---|---|---|---|
| Scaffold | ICICLE | 0.652 ± 0.002 | 0.655 ± 0.001 | 0.717 ± 0.001 | 0.635 ± 0.001 |
| NEIMS | 0.595 ± 0.013 | 0.610 ± 0.003 | 0.683 ± 0.001 | 0.637 ± 0.004 | |
| MassFormer | 0.536 ± 0.001 | 0.669 ± 0.001 | 0.675 ± 0.001 | 0.598 ± 0.001 | |
| Random | ICICLE | 0.805 ± 0.003 | 0.789 ± 0.002 | 0.838 ± 0.001 | 0.743 ± 0.001 |
| NEIMS | 0.813 ± 0.001 | 0.727 ± 0.004 | 0.755 ± 0.002 | 0.732 ± 0.001 | |
| MassFormer | 0.615 ± 0.001 | 0.736 ± 0.001 | 0.698 ± 0.001 | 0.621 ± 0.001 |
For reference, two experimental spectra of the same compound (replicates) have a mean cosine similarity of 0.948, and two constitutional isomers have 0.287 (Table S8).
Formula-match retrieval (50 same-formula PubChem isomers, entropy re-ranking; Tables S21 and S23)
| Split | Model | Top-1 | Top-5 | Top-10 | MRR |
|---|---|---|---|---|---|
| Scaffold | ICICLE | 43.0 ± 0.5% | 74.6 ± 0.2% | 84.7 ± 0.1% | 0.573 |
| NEIMS | 25.1 ± 2.9% | 53.2 ± 4.0% | 68.9 ± 3.4% | 0.389 | |
| MassFormer | 32.5 ± 0.3% | 66.2 ± 0.4% | 81.4 ± 0.2% | 0.478 | |
| Random | ICICLE | 56.2 ± 0.4% | 85.6 ± 0.2% | 92.6 ± 0.1% | 0.691 |
| NEIMS | 49.3 ± 1.6% | 79.8 ± 1.3% | 89.1 ± 0.5% | 0.628 | |
| MassFormer | 39.9 ± 0.2% | 74.2 ± 0.3% | 87.4 ± 0.0% | 0.550 |
Full-PubChem retrieval (87,260,989 candidates, no formula, cosine ranking, seed 1 only; Tables S33, S34 and S37)
| Split | Model | Top-1 | Top-10 | Top-50 | MRR |
|---|---|---|---|---|---|
| Scaffold | ICICLE | 3.3% | 8.7% | 14.2% | 0.052 |
| ICICLE + ±4 Da peak-mass window | 5.3% | 13.9% | – | 0.083 | |
| NEIMS | 1.1% | 4.0% | 8.0% | 0.021 | |
| MassFormer | 0.2% | 1.1% | 2.8% | 0.006 | |
| Random | ICICLE | 12.0% | 26.3% | 35.6% | 0.170 |
| NEIMS | 14.1% | 32.8% | 43.2% | 0.205 | |
| MassFormer | 1.0% | 4.4% | 9.1% | 0.022 |
On the scaffold split, the median rank of the true structure is 19,924 for ICICLE, 52,639 for NEIMS and 209,393 for MassFormer (Table S35). On the random split NEIMS leads on long aliphatic chains, esters and ethers, while ICICLE leads on halogenated and heteroatom-rich chemotypes such as alkyl/aryl halides, thioethers, nitriles and sulfonamides (Table S39).
RASSP
RASSP needs a formula-complexity-filtered subset of the data (197,840 entries). On that subset with the scaffold split, formula-match top-1 with entropy re-ranking is 38.1 ± 0.6% for ICICLE, 26.8 ± 0.1% for MassFormer, 19.9 ± 3.3% for NEIMS and 11.6 ± 0.3% for RASSP (Table S24). Entropy similarity on the same subset is 0.675 for ICICLE and 0.646 for RASSP (Table S12).
Comparison with quantum chemistry (Table S13)
On the 15 test molecules from Gorges and Grimme, with models trained on the random split with those molecules excluded, mean cosine similarity is 0.82 for ICICLE, 0.79 for NEIMS and 0.70 for wB97X-3c QCxMS2. ICICLE runs about seven orders of magnitude faster than QCxMS2.
ICICLE hyperparameters (Table S4)
Hidden size 256, 2 GNN layers, 1 intra-graph and 2 inter-fragment attention layers, 3 fragment-feature MLP layers, dropout 0.1, average pooling, batch size 256, learning rate 1e-3, weight decay 1e-5 and learning-rate decay 0.95.
Limitations
- Skeletal rearrangement products, which form new bonds, are not enumerated, so the peaks they produce can be missing from predictions.
- The models predict standard 70 eV spectra only and do not model how fragmentation changes with internal energy.
- All evaluations use one NIST reference spectrum per compound. Performance on spectra from other instruments was not tested.
Citation
@article{lederbauer2026icicle,
title = {Fragment-Grounded Neural Simulation of Electron Ionization Mass Spectra at Library Scale},
author = {Lederbauer, Magdalena and Wang, Runzhong and Coley, Connor W.},
journal = {Analytical Chemistry},
year = {2026},
doi = {10.1021/acs.analchem.6c06136}
}