dsadd4/UltraMS-Benchmark-Assets

UltraMS benchmark assets

Inputs for reproducing the UltraMS spectral-property and molecular-identification benchmarks. The GitHub benchmark guides provide the training and evaluation commands.

Path	Use
spectral_properties/masked_peak_reconstruction.pt	UltraMS checkpoint for masked peak reconstruction.
spectral_properties/massspecgym_labels.parquet	Frozen MassSpecGym property and neutral-loss training table used in Figure 2.
spectral_properties/heteroatom_count/	Selected UltraMS and DreaMS atom-query probe checkpoints for S, Cl, F and Br counts and element-bearing peak localization.
spectral_properties/negative_ion_structure/	Frozen validation and test spectra, molecular fingerprints, metadata, and split manifest for the negative-ion structure-family analysis.
spectral_properties/peak_neighbors/	Frozen MSnLib spectra, MAGMa annotations, and UltraMS/DreaMS peak embeddings for the peak-neighbor analysis.
molecular_identification/contrastive_candidate_ranking.pt	UltraMS contrastive checkpoint for candidate ranking.
molecular_identification/fragment_candidate_ranking.pt	UltraMS fragment-aware checkpoint for candidate ranking.
molecular_identification/single_spectrum_projection.pt	Selected UltraMS spectrum-to-molecule projection for the reported MSnLib single-spectrum identification result; load with the UltraMS Unsupervised encoder.
molecular_identification/single_spectrum_dreams_projection.pt	Selected DreaMS spectrum-to-molecule projection for the reported MSnLib comparison; load with the official DreaMS SSL encoder.
molecular_identification/single_spectrum_readouts/	Selected Linear, DeepSets, Fourier and Codebook readouts for the reported single-spectrum comparison, plus separate fresh-training weights.
molecular_identification/collision_energy_ultrams.pt	UltraMS fine-tuned checkpoint for collision-energy-aware library search.
molecular_identification/collision_energy_dreams.pt	DreaMS fine-tuned comparison checkpoint for the same task.
molecular_identification/nplib1/	Fixed NPLIB1 three-fold splits and candidate molecule table.
molecular_identification/single_spectrum_test_predictions/	Frozen Top-50 predictions for six methods on all 57,437 MSnLib test spectra; used for candidate voting and label-blind analyses.

The selected projections and readouts have public download and evaluation commands. The selected UltraMS projection was re-evaluated from public inputs on all 57,437 test spectra, recovering 15,597 Top-1 hits (27.15497%). The DreaMS selected projection and four readouts are released alongside the complete per-query results; independent full selected-weight evaluation is in progress.

The UltraMS benchmark checkpoints contain the exact model tensors used for the reported runs. The two collision-energy checkpoints also retain the exact fine-tuned encoder and projection-head tensors. The masked-peak checkpoint is separate from the three released UltraMS inference models.

Sources and reuse

The two heteroatom-count probe checkpoints contain the atom-query and regression-head tensors plus label normalization, with UltraMS and DreaMS backbones distributed separately. The selected evaluation snapshots are UltraMS epoch 13 and DreaMS epoch 1.

UltraMS checkpoint weights are released under Apache-2.0. The DreaMS comparison checkpoints and selected projection derive from the DreaMS pretrained model by Roman Bushuiev and colleagues, released under CC BY 4.0; cite and credit the DreaMS authors when using that checkpoint. The MassSpecGym table derives from the MassSpecGym public dataset. The negative-ion structure-family inputs derive from SpecBridge SpectraVerse by Yinkai Wang and colleagues (Zenodo record license: MIT), with the published exclusion protocol applied. The NPLIB1 splits and candidate molecules follow the public MIST NPLIB1 benchmark and its PubChem candidate data.

The spectral_properties/peak_neighbors/ inputs derive from MSnLib spectral libraries, released under CC BY 4.0, by selecting 10,000 spectra, adding MAGMa fragment annotations, and extracting UltraMS and DreaMS peak embeddings. Credit the original dataset authors when reusing derived data.

The six per-spectrum prediction tables derive from the public MSnLib spectra and candidate lists provided through SpecBridge by Yinkai Wang and colleagues (Zenodo record license: MIT). The underlying MSnLib spectral library is CC BY 4.0. Credit both data sources when using these tables. Each JSONL row contains one test spectrum’s identifier, candidate ranks, scores, and Top-50 candidate list for one method.