Loading…
Metabolite-likeness scoring ranks any chemical structure by its distance from a learned hypersphere of known endogenous metabolites.
No providers recorded yet. Browse all providers
An untargeted LC–MS experiment measures an accurate mass, and searching that mass against a chemical database returns hundreds or thousands of structures sharing the formula — most of them industrial intermediates or drug-like scaffolds no cell has ever made. Narrowing the list needs a prior over what an endogenous metabolite looks like. The obvious way to learn one, a classifier trained on metabolites against non-metabolites, founders on the negative class: a compound absent from a metabolite database may simply be one nobody has found yet, so any labelled set of "non-metabolites" is guesswork.
DeepMet learns from the positive class alone. It adapts Deep SVDD, a one-class anomaly-detection objective, to chemical structures: a network maps known metabolites into a compact hypersphere in a 200-dimensional latent space, and an incoming molecule is scored by the squared distance of its embedding from that hypersphere's centre. Structures that land close resemble the training metabolites; structures that land far do not, and no counterexamples are needed at any stage.
The model was built by Jack Gisby and Ralf J. M. Weber at the University of Birmingham, with Gisby also at Imperial College London, and released in 2021 as the deepmet command-line tool. An unrelated Princeton chemical language model, DeepMet, carries the same name but generates candidate metabolite structures rather than scoring supplied ones.
The encoder is a set-transformer stack of three self-attention blocks (2,800 to 1,000 dimensions with 10 heads, 1,000 to 500 with 5 heads, 500 to 200 with 10 heads) followed by a pooling-by-multihead-attention head that collapses the sequence to the 200-dimensional representation — about 11.7 million parameters, distributed as a 46 MB PyTorch checkpoint inside the repository. Input width comes from feature selection: the concatenated fingerprint runs to roughly 13,000 bits, and columns that are constant or more than 90% a single value across the training set are dropped, leaving 2,800. That column list ships with the configuration and is replayed on every new molecule, so an incoming structure is projected onto exactly the features the checkpoint saw. Training used nu set to 0.1, twenty epochs of AMSGrad and a batch size of 2,000. It also ships the structure sets used to assess the model — endogenous metabolites from the HMDB as the normal class, ZINC12 structures as the non-normal class — so evaluation combines a held-out fold of metabolites with a disjoint database of non-metabolites. No manuscript accompanies the release, though the packaged notebook reports an AUC of 0.9886 against ZINC12 structures, alongside a ROC curve, a confusion matrix and isolation-forest and one-class-SVM baselines it beats. That figure belongs to a model the notebook retrains from the shipped compound lists at batch size 25, not to the distributed checkpoint, for which no metric is published anywhere.
The immediate use is triage inside untargeted metabolomics annotation. Given the candidates a formula or spectral search returns, the scoring command takes a CSV of identifiers and SMILES and returns one number per structure, letting an analyst discard implausible entries before committing instrument time and a purchased reference standard. Because the normal class is a free parameter, the same tool serves adjacent filtering problems, such as judging whether a screening hit resembles a lipid library. Installation carries one unusual requirement: the CDK fingerprints are computed in Java, so a runtime has to sit alongside the Python environment.
DeepMet argues, compactly, that one-class learning fits metabolite annotation better than binary classification does, and that a metabolite prior can ship as a fixed artifact instead of being refitted per study. It sits on the opposite side of the problem from generative approaches, which propose structures a measured mass might correspond to; a likeness score consumes the candidate lists those methods and conventional database searches produce. The release is modest in reach: no publication has appeared, the repository has seen little activity since 2024, and the PyPI, Bioconda and Read the Docs badges in its README point at targets that were never published, so the package is installed from the repository itself. What is there is complete, though — GPL-3.0 source, a trained checkpoint, the configuration that reproduces its feature space, and the evaluation sets, all in one tree.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.