JESTR

A single chemical formula can match tens of thousands of real molecules — PubChem lists 44,374 structures for C12H18N2O2 alone. Untargeted metabolomics measures a tandem mass spectrum for each detected ion, and annotation is the problem of choosing which candidate produced it. Reference-library search answers that only when the spectrum has been recorded before: a survey of 15,327 MassIVE datasets against the 586,647 reference spectra in GNPS returned a positive identification rate of 2.3%.

The predictive alternatives all route through an explicit intermediate: mol-to-spec tools simulate a spectrum for every candidate and compare it against the query, while spec-to-fingerprint tools predict a molecular fingerprint and match it against candidate fingerprints. Either way a reconstruction step sits between the measurement and the answer, and its error propagates into the ranking. JESTR (Joint Embedding Space Technique for Ranking) removes the reconstruction entirely. It treats a molecule and its spectrum as two views of the same object — the framing contrastive language-image pretraining applies to images and captions — and trains a molecular encoder and a spectral encoder so that matching molecule–spectrum pairs land near each other in one shared space. Annotation becomes: embed the query spectrum, embed each candidate structure, sort by cosine similarity. There is no learned ranking head.

JESTR was developed by Apurva Kalia, Yan Zhou Chen and Soha Hassoun at Tufts University with Dilip Krishnan of Google DeepMind, posted as a preprint in November 2024 and published in Bioinformatics in June 2025. Where DreaMS represents spectra alone for library search and fingerprint prediction, JESTR's space is shared between both modalities — which is what lets an arbitrary PubChem structure be scored against a spectrum it was never measured with.

#Key Features

• Implicit annotation: Ranking is a cosine similarity in the joint space, with no predicted spectrum, fingerprint, or SMILES string in between.
• Regularization against lookalikes: Late in training, an auxiliary loss pushes away the PubChem candidates sharing a target's molecular formula — the exact molecules the model must later discriminate against — lifting rank@1 by 11.4%, 6.0% and 37.1% on NPLIB1, NIST2020 and MoNA.
• Asymmetric encoders: A graph convolutional network reads molecular structure while a small MLP reads a binned spectrum.
• Structure-disjoint evaluation: Splits are partitioned by InChIKey, so test structures never appear in training under any spectrum.
• Released checkpoints: The repository ships encoder weights for the NPLIB1 and MassSpecGym datasets alongside a ranking notebook that loads them directly.

#Technical Details

The molecular encoder is a three-layer graph convolutional network over atom-and-bond graphs (atom type, mass, valence, ring membership, formal charge, radical electrons, chirality, degree, hydrogen count, aromaticity; bond type, conjugation, stereo configuration), followed by max pooling and two fully connected layers. The spectral encoder bins peaks at 1 Da up to 1000 Da, sums intensities per bin, applies a log10/3 transform, and passes the 1000-dimensional vector through a three-layer MLP. Both produce 512-dimensional embeddings, trained with contrastive multiview coding's InfoNCE loss at temperature 0.05. Regularization runs over the last 3% of epochs at a loss weight of 0.1, drawing on roughly 7 million of the 15.86 million same-formula candidates retrieved from PubChem.

Weights are fit per spectral dataset rather than from one broad corpus, and the paper evaluates four. On NPLIB1 (8,030 spectra, 7,131 molecules), JESTR reaches 45.76% rank@1 and 95.77% rank@20, against 27.96% and 82.05% for MIST and 23.69% and 60.20% for ESP. On the commercial NIST2020 set (291,515 spectra, 22,001 molecules) it reaches 38.62% rank@1. On MassSpecGym, whose test fold is separated by maximum common edge subgraph distance and is correspondingly harder, it reaches 15.13% rank@1 with candidates drawn by mass and 11.85% by formula; restricted to spectra whose molecules are absent from SIRIUS's training data, it ranks 13.65% at position one against SIRIUS's 10.41%.

#Applications

JESTR targets the point in an untargeted LC-MS/MS workflow where a feature has a confident formula but no library match — the majority of features in most studies. Given candidates pulled from PubChem or a smaller biologically restricted database, it returns a ranked shortlist for manual inspection or pathway analysis, useful to natural-product chemists, exposomics groups, and clinical metabolomics labs working outside the well-covered regions of chemical space. The released checkpoint and ranking notebook let a user score their own spectra without retraining.

#Impact

JESTR is one of the clearer demonstrations that the contrastive joint-embedding recipe transfers from vision-language to spectrum-structure matching, and that a small architecture with a well-chosen similarity function can beat larger reconstruction-based pipelines: the paper shows that swapping cross-entropy over dot products for InfoNCE over cosine similarity recovers much of its margin over CMSSP, the closest prior joint-embedding annotator. The caveats are real: it ranks a supplied candidate list rather than proposing structures de novo, so it inherits whatever candidate retrieval misses; MIST still leads at rank@1 on MoNA; and of the evaluation data only NPLIB1 is redistributed with the code, the authors citing licensing agreements that prohibit releasing the rest — though the repository does ship weights for both NPLIB1 and MassSpecGym.