CSU-MS²

A spectral library can only identify a compound that somebody has already put through a mass spectrometer. Untargeted metabolomics acquires tens of thousands of MS/MS spectra per study and matches them against reference libraries covering a few hundred thousand compounds, while the structure databases those metabolites are actually drawn from hold tens of millions of entries. A molecule whose spectrum has never been recorded is invisible to spectral matching no matter how good the similarity score is.

CSU-MS² changes what a spectrum gets compared to. Rather than matching spectrum against spectrum, it trains two encoders — one reading MS/MS peak lists, one reading molecular graphs — into a single embedding space using a contrastive objective over spectrum–structure pairs. A query spectrum is then scored directly against candidate structures, so any molecule with a SMILES string is searchable whether or not a reference spectrum for it exists. The name is an acronym for the method: "contrastively spectral-structural Unification framework for MS/MS Spectra and Molecular Structures".

The model was built by Ting Xie and colleagues in the College of Chemistry and Chemical Engineering at Central South University and published in Analytical Chemistry in June 2025. It sits alongside a different bet on the same problem: where LSM-MS2 learns a representation of spectra on their own, CSU-MS² makes the spectrum-to-structure pairing itself the training signal.

#Key Features

• Cross-modal contrastive alignment: A symmetric NT-Xent objective over cosine similarities at temperature 0.1 pulls each spectrum toward its own structure and away from the rest of the batch, with the spectrum-to-structure direction weighted more heavily than the reverse.
• External Space Attention aggregation: Rather than mean-pooling atoms or peaks into a fixed vector, the ESA module learns a softmax weight per token, letting diagnostic fragment ions and the substructures they report dominate the embedding.
• In-silico pretraining, experimental fine-tuning: Pretraining uses large simulated MS/MS sets generated by CFM-ID and ICEBERG, which cover far more chemistry than measured libraries do; a fine-tuning pass on experimental spectra then corrects for the simulators' biases.
• Instrument- and energy-matched checkpoints: Six checkpoints ship in the repository — Orbitrap HCD and QTOF, each at low, medium and high collision energy — and their scores can be combined when the acquisition energy is uncertain.
• A searchable structure database: SSFDB, assembled from 23 structural databases and distributed as a single 14.9 GB SQLite file, backs retrieval so that a search is a nearest-neighbor lookup over precomputed structure embeddings.

#Technical Details

The structure tower is a five-layer graph isomorphism network with edge features (GINE) over 300-dimensional atom embeddings summed from atom type, chirality, hybridization, valence and degree, with bond type and direction carried on the edges. The spectrum tower embeds each peak's m/z with sinusoidal features spanning log-spaced frequencies, concatenates the intensity, and runs three transformer encoder layers with eight attention heads and a 1024-dimensional feed-forward block over the padded peak set. Both towers pass through their ESA head and a linear projection into a shared 256-dimensional space; the released checkpoints are about 21 MB each.

On the paper's benchmark — 1047 query spectra searched against a reference set of 1,001,047 compounds — CSU-MS² places the correct structure first for 75.45% of queries, against 68.38% for CFM-ID, 64.85% for SIRIUS, 48.59% for MetFrag and 30.47% for CMSSP. Three external evaluations, the MTBLS265 human metabolomics study, the PMhub plant metabolite resource and the CASMI 2022 challenge, test the same checkpoints without retraining, reaching Recall@10 of 91.67% on blood metabolites. These are retrieval numbers: they measure where the true structure lands in a ranked candidate list, not whether a structure was elucidated de novo.

#Applications

The intended use is annotation of unknown features in untargeted LC-MS/MS runs — clinical metabolite profiling, plant natural-product work, and any survey where most detected features never match a library entry. Because retrieval runs against a user-supplied structure file as readily as against SSFDB, a lab can narrow candidates to a biologically plausible set (the metabolites of one species, a drug-metabolite panel, a pesticide list) and get a ranked shortlist to take to authentic standards. The authors also released a Gradio web application for uploading spectra without a local install, though that Hugging Face Space has been paused since March 2026 and does not currently run; its source and the three Orbitrap HCD checkpoints it mirrors remain downloadable.

#Impact

CSU-MS² belongs to a shift in small-molecule mass spectrometry away from spectrum-to-spectrum matching and toward learned joint spaces where spectra and structures are directly comparable, a reframing that removes library coverage as the ceiling on what can be annotated. Later work on the same task now benchmarks against it rather than against library search alone. The practical limits are worth stating: the checkpoints are tied to particular instrument classes and collision-energy regimes, the paper is behind a subscription, and the code repository carries neither a license nor a model card, which constrains reuse more than the released weights suggest.