Metabolite annotation model aligning molecular graphs, fingerprints, MS/MS spectra and consensus spectra in one contrastive space to rank candidates.
No providers recorded yet. Browse all providers
A metabolite is rarely measured once. An LC-MS/MS run typically fragments the same ion at several collision energies, and sometimes as more than one adduct, so a single compound can arrive as three or four spectra, each showing a different slice of its fragmentation. Most annotation models take one spectrum at a time. A user holding several must either pick one or rank candidates separately for each and then reconcile the disagreeing lists.
MultiView Projection (MVP) is built to use those extra measurements directly. It extends the joint-embedding approach of its predecessor JESTR, in which a molecule and its spectrum are placed close together in one learned space and candidates are ranked by cosine similarity. MVP widens the set of views from two to four: a molecular graph and a Morgan fingerprint on the molecule side, and an individual spectrum and a consensus spectrum on the spectral side. The consensus spectrum merges every peak from all spectra sharing a molecular identity. Each view has its own encoder, and a contrastive loss is applied across all six pairs of views, so every representation of a compound is pulled toward every other. At inference the user chooses a ranking view: one spectrum against the graph, or a consensus spectrum against the graph. Merging first and ranking once ("aggregate-then-rank") replaces the reconciliation step.
MVP was developed by Yan Zhou Chen and Soha Hassoun at Tufts University and posted to bioRxiv in November 2025. It sits between JESTR and FLARE in the lab's line of spectrum-structure annotation models.
The molecular graph encoder is JESTR's: a three-layer graph convolutional network, pooling, and two fully connected layers. Fingerprints (1,024-bit Morgan, radius 5) pass through a three-layer MLP. For the spectral views, peaks are labelled with subformulae using MIST's annotation code, keeping the 60 most abundant peaks within a 20 ppm tolerance. Each peak becomes a 15-dimensional vector of counts over 14 elements plus normalized intensity. A three-layer MLP embeds each peak, and a two-head transformer encoder pools the peak set. Spectra and consensus spectra use encoders of the same architecture. Training uses Adam for 1,500 epochs at learning rate 7e-5, batch size 64, and contrastive temperature 0.05.
MVP was trained and tested on MassSpecGym, 194,119 training spectra of 25,046 molecules, with test molecules held at a maximum common edge subgraph distance of at least 10 from training. Querying with a single spectrum, it reaches 26.37% rank@1 with candidates retrieved by mass and 11.10% by formula. JESTR reaches 15.13% and 11.85%, and MIST 14.64% and 9.57%. Querying with a consensus spectrum raises this to 35.99% and 13.96%, with 91.92% rank@20 by mass. On formula-based candidates the binned-spectrum variant beats the formula variant at rank@1, which the authors attribute to formula features capturing composition more than structure.
MVP is suited to untargeted LC-MS/MS studies where a feature has a formula but no library match and several spectra of the same compound are available. The authors ran it on public Metabolomics Workbench data comparing cultured and conventional chicken meat, using MZmine3 for feature extraction and SIRIUS for formula and adduct assignment. MVP annotated 640 spectra covering 146 molecules, 73 of them not reported in the original study, and reproduced its pathway-level finding that amino acids and peptides are depleted in cultured meat. These annotations are putative and have not been confirmed experimentally.
MVP's main contribution is a demonstration that repeat measurements are better merged before annotation than after. It also shows that a contrastive framework can absorb added molecular and spectral views without architectural changes. Limitations are stated in the paper. MVP ranks a supplied candidate list and cannot propose structures outside it. Its ablations show no consistent benefit from extra views when trained on all of MassSpecGym, only on the [M+H]+ subset. The consensus advantage depends on having spectra at several collision energies. Formula-based peak encoding degrades when an element or count falls outside the training range. The work is a preprint and has not been peer reviewed.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.