The team behind MetaProViz, led by researchers at Heidelberg University and the University of Cologne, including Christina Schmidt, Christian Frezza and Julio Saez-Rodriguez, identified two stubborn obstacles that have long plagued metabolomics analysis. The first is the lack of standardised workflows. Although dozens of tools exist for individual steps, most are available only as web applications whose user-friendly interfaces hide critical parameter settings, undermining the reproducibility standards that modern science demands. The second obstacle is more subtle and arguably more damaging: metabolite annotations are frequently ambiguous. Mass spectrometry cannot reliably distinguish fine structural details such as stereoisomers or constitutional isomers, meaning a detected feature does not correspond to one fully defined chemical structure. Databases compound the problem by storing metabolites at different levels of specificity, so assigning a single unambiguous identifier to every measured peak is often impossible.

MetaProViz tackles these problems with a modular architecture consisting of five interlocking components: pre-processing, differential metabolite analysis, prior knowledge access and integration, functional analysis, and visualisation. The package operates on annotated intensity matrices, the format typically delivered by metabolomics facilities after peak assembly and annotation, and each module can be used interactively or as a building block within larger pipelines. Built-in vignettes with example datasets, automatic logging of parameters, warnings and errors, and publication-ready figures based on ggplot2 make the tool approachable for researchers with minimal coding experience while satisfying the audit trails demanded by experienced bioinformaticians. Automated checks and informative messages guide users through decisions that would otherwise require specialist knowledge, from normalisation strategy to statistical test selection, including Shapiro-Wilk tests that report whether the chosen statistical approach is appropriate for the data at hand.

At the heart of the package lies a curated knowledge resource called MetSigDB, the Metabolism Signature Database. This collection of annotated metabolite sets unifies pathway-metabolite sets from sources such as KEGG, Reactome and WikiPathways, chemical class-metabolite sets, cancer-metabolite sets from MACdb, and metabolite-protein interactions from MetalinksDB covering receptors, transporters and enzymes. MetSigDB also enables conversion of gene sets into metabolite sets using metabolic reactions drawn from a filtered version of the Recon-3D model. Crucially, MetaProViz systematically excludes non-specific metabolites such as water, carbon dioxide and xenobiotics, and its prior knowledge is accessed through the OmniPath ecosystem, giving the tool a sustainable backend that can evolve with the field.

The most technically innovative part of MetaProViz is its handling of metabolite identifier ambiguity, a problem the authors illustrate vividly with the amino acid alanine. When assigning identifiers from the Human Metabolome Database, both D-alanine and L-alanine must be attached because no unspecific HMDB entry exists, whereas ChEBI offers only a single generic identifier that fails to map to most standard pathway databases. Because enzyme stereoselectivity is homochiral, with living systems dominated by L-amino acids and D-sugars, these distinctions matter biologically. MetaProViz addresses the mismatch through an eight-step workflow: quality control of the feature identifier space, counting identifiers per feature, detecting identifier mismatches, translating identifiers from one database to another, traversing a metabolite identifier graph to harvest all connected identifiers, adding equivalent identifiers based on chirality, and then quantifying and flagging mapping ambiguities for manual review.

Applied to a published clear-cell renal cell carcinoma patient dataset, this workflow delivered striking results. Initially only 43 percent of features carried complete assignments across HMDB, KEGG and PubChem, and a quarter of features had nothing but a PubChem identifier. After MetaProViz cleaning, translation and graph traversal, the proportion of fully annotated features rose to 54 percent, dramatically expanding the overlap between measured data and prior knowledge. The team also catalogued the mapping scenarios that can silently distort enrichment analysis, from many-to-one mappings that deflate pathways to one-to-many mappings that inflate them, and flagged sixteen cases in the renal cancer data that required manual review. Without this systematic accounting, over-representation analysis could produce misleading or outright nonsensical conclusions, a known hazard when genomic-style pathway tools are naïvely applied to metabolites.

The package also breaks new ground in exometabolomics, the measurement of metabolites consumed from and released into the culture medium by cells, a capability the authors note is not offered by any other publicly available tool. MetaProViz normalises these data against media blanks and growth factors, yielding negative values for consumed metabolites and positive values for released ones, and replaces classical log2 fold changes with a log2 distance measure that captures the direction of exchange between conditions. Its biological regulatory clustering method, bioRCM, then integrates extracellular consumption-release patterns with intracellular metabolomics through logical regulatory rules, producing clusters of metabolites that carry clear biological meaning.

When the team applied this machinery to renal cancer cell lines, the results were biologically compelling. Primary and metastatic clear-cell renal cell carcinoma cells predominantly consumed amino acids from the medium while healthy renal epithelial cells released them. One cluster of particular interest contained metabolites that were consumed by the cancer cells and simultaneously depleted inside them. Querying this cluster against MetalinksDB revealed methionine as a standout: cancer cells consumed methionine from the media and showed reduced intracellular levels, and tumour tissue from patients similarly exhibited decreased methionine compared with adjacent normal tissue. The authors connect this increased methionine usage to the elevated DNA-hypermethylation landscape characteristic of clear-cell renal cell carcinoma, and the associated enzymes and transporters, including BHMT and SLC43A2, have previously been linked to patient survival. This is precisely the kind of mechanistic hypothesis, linking an extracellular metabolic exchange to an intracellular epigenetic mechanism, that the package was designed to surface.

Beyond cell lines, MetaProViz introduces a metadata analysis method for heterogeneous patient cohorts. By relating principal components of the metabolomics data to clinical variables through analysis of variance, the method identifies which patient characteristics drive metabolic variation. Applied to 138 matched tumour-normal pairs, it confirmed that tissue type and tumour stage explained the largest share of variance, but also revealed that a small percentage of variance in the fourth principal component separated patients by age. Subsequent multi-condition clustering comparing young and old patient subsets uncovered an age-specific signature of dipeptides, upregulated in older patients and downregulated in younger ones, accompanied by altered aminoacyl-tRNA biosynthesis and protein digestion and absorption pathways. The authors suggest such dipeptide patterns could hold promise as biomarkers, and demonstrate that metadata-driven analysis can extract metabolic patterns invisible to classical pathway enrichment alone.

The developers have positioned MetaProViz for the future as well as the present. The package is available through Bioconductor, ensuring reproducible and scalable installation, and the team anticipates integration into workflow management systems such as nf-core modules in Nextflow, which would embed it directly into the pipelines of metabolomics core facilities. As artificial intelligence accelerates metabolite identification through language-model-guided annotation and self-supervised learning of mass spectra, the number of confidently identified metabolites will grow sharply, making rigorous tools for functional interpretation ever more valuable. By combining curated prior knowledge, ambiguity-aware identifier handling, exometabolomics analysis and safeguarded statistics in one flexible framework, MetaProViz offers the metabolomics community what genomics has long enjoyed: a standardised path from raw data to biological understanding, and a systematic way to ask not just which metabolites change, but why it matters.

The statistical machinery underlying MetaProViz reflects a broader lesson about the distinct nature of metabolites as data objects. Unlike genes, whose counts lend themselves to established negative binomial models, metabolite intensities span wide dynamic ranges and their distributions can vary substantially across features, which is why the package evaluates normality per feature and warns users when a chosen test may be inappropriate. The background set problem in over-representation analysis is treated with similar care: because the measured metabolite list is shaped by the extraction method and instrumentation rather than by biology alone, the package allows users to define the statistical universe explicitly, a detail that genomic enrichment tools often take for granted but that can materially alter results in metabolomics.

The methionine findings in clear-cell renal cell carcinoma also sit within a well-documented biological context. Renal cancer cells are known to rely heavily on amino acid uptake, and methionine metabolism feeds directly into the one-carbon cycle that supplies methyl groups for DNA methylation. The observed combination of methionine consumption from the medium, intracellular depletion in tumours, and hypermethylation signatures is therefore internally coherent, and the survival associations reported for the linked enzymes and transporters suggest the pattern may have clinical relevance rather than being a mere artefact of cell culture. The authors are careful to frame these connections as hypotheses for experimental validation, which is precisely the intended output of the functional analysis modules.

The age-specific dipeptide signature uncovered in the patient cohort illustrates another strength of the metadata-driven approach. Dipeptides are products of protein turnover, and altered aminoacyl-tRNA biosynthesis and protein digestion pathways accompanying the age separation point toward shifts in protein metabolism with ageing. Because the analysis of variance framework relates principal components to clinical variables directly, such associations emerge without preselecting metabolites of interest, reducing the risk of confirmation bias. For laboratories adopting the package, the modular design means these analyses can be adopted incrementally, starting with quality control and differential analysis before progressing to prior knowledge integration, making the transition to reproducible metabolomics workflows manageable even for groups without dedicated bioinformatics support.

Subject of Research: An open-source metabolomics data analysis package, MetaProViz, that integrates prior knowledge to generate mechanistic hypotheses.

Article Title: Integrated metabolomics data analysis to generate mechanistic hypotheses with MetaProViz

Article References: Schmidt, C., Franken, J., Turei, D., Prymidis, D., Daley, M., Frezza, C., & Saez-Rodriguez, J. (2026). Integrated metabolomics data analysis to generate mechanistic hypotheses with MetaProViz. Molecular Systems Biology. https://doi.org/10.1038/s44320-026-00231-8

Image Credits: AI Generated

DOI: 10.1038/s44320-026-00231-8

Keywords: metabolomics, MetaProViz, bioinformatics, mass spectrometry, exometabolomics, pathway enrichment, MetSigDB, metabolite identifiers, clear-cell renal cell carcinoma, methionine metabolism, R package, prior knowledge integration

Metabolomics is designed to capture the small molecules that reflect what cells are doing at a particular moment. Unlike genomics, which describes biological potential, metabolomics records the chemical consequences of genetics, diet, exercise, disease, medication and environmental exposure. Liquid chromatography–tandem mass spectrometry, or LC–MS/MS, is one of the field’s most powerful tools. In this technique, liquid chromatography separates compounds according to their chemical interactions with a column, while mass spectrometry measures their mass-to-charge ratios and fragments selected molecules to generate structural clues. The resulting data contain thousands of peaks, each representing an ionized molecular feature. Some can be matched to reference standards or spectral libraries. Others have an accurate mass and perhaps a molecular formula, but no secure name, structure or biological interpretation.

The problem is more complicated than simply searching a larger database. Two compounds can share the same molecular formula while having different arrangements of atoms, known as structural or positional isomers. Stereoisomers can also possess the same connectivity but differ in three-dimensional orientation. These molecules may produce nearly identical masses and overlapping fragmentation patterns, while behaving differently in the chromatographic system and exerting different effects in the body. Unknown signals may also arise from chemical transformations that are poorly represented in databases, including compounds generated by gut microbes, noncanonical biochemical reactions or exposure to external chemicals. Instrumental artifacts, such as in-source fragmentation and ion rearrangement, can further create peaks that resemble genuine metabolites. As a result, a formula match alone is rarely enough to support a biologically meaningful annotation.

The researchers addressed this challenge by first defining the chemical landscape of the biological dataset itself. Their test data came from a randomized controlled trial investigating exercise during pregnancy in women with obesity. Plasma samples were collected from 119 sedentary pregnant participants, including women assigned to a combined aerobic and resistance-training program and others receiving standard care. In total, 235 samples were analyzed, including samples collected after submaximal exercise assessments. The study was not primarily intended to compare exercise and control groups in the present analysis. Instead, it provided a biologically relevant collection of plasma metabolites in which to develop and evaluate a strategy for interpreting unknown LC–MS features.

The samples were analyzed in both positive and negative electrospray ionization modes using an Orbitrap Exploris 480 mass spectrometer coupled to ultra-high-performance liquid chromatography. The instrument acquired high-resolution full-scan data across broad mass ranges and tandem spectra from pooled quality-control samples. After filtering for mass accuracy, signal quality, reproducibility and background contamination, the researchers retained 2,857 metabolite features. Of these, 1,021 had some level of annotation, ranging from relatively strong matches supported by an in-house standard and retention time to weaker assignments based only on accurate mass or molecular formula. The remaining 1,836 features lacked a chemical identity and were designated as metabolic dark matter. Rather than treating the partially annotated compounds as proven identifications, the researchers used them as landmarks defining regions of chemically plausible space.

To organize those landmarks, the team grouped the 1,021 known or partially characterized metabolites into ten structurally coherent clusters. The clustering process used molecular fingerprints, digital representations of the substructures present in each molecule. Specifically, the researchers generated Morgan circular fingerprints from SMILES chemical notation using the RDKit cheminformatics toolkit. These fingerprints encode local atomic environments as binary vectors. Structural similarity between two molecules was then measured with the Tanimoto coefficient, calculated from the number of shared fingerprint features relative to the total features present in either molecule. A score of 1 indicates identical fingerprint representations, while a score approaching 0 indicates little overlap. Importantly, the score reflects two-dimensional structural similarity and does not establish identical stereochemistry, tautomeric state or biological activity.

The next stage began with a deliberately broad search. For each of the 1,836 unannotated features, the researchers used its predicted molecular formula and molecular weight to retrieve possible structures from PubChem, allowing a relatively generous mass tolerance of plus or minus 0.5 daltons. This produced 569,115 candidate entries. After removing duplicates, malformed structures and records without usable chemical representations, 368,197 unique structures remained. The candidates were compared with the ten known-metabolite clusters, and those with Tanimoto similarity scores between 0.50 and 1.00 were retained. The lower threshold was intentionally permissive: a score of 0.5 was treated as evidence of a potentially useful chemical relationship, not as proof of identity. More stringent thresholds of 0.6 and 0.7 were also examined to show how the shortlist changed under increasingly conservative assumptions.

Similarity alone still left too many possibilities, so the researchers added a second, independent clue: retention time. In LC–MS, retention time is the point at which a compound emerges from the chromatography column. It depends on physicochemical properties such as hydrophobicity, polarity, hydrogen bonding and interactions with the stationary phase. Positional isomers with the same formula and mass may therefore separate at different times. The team used machine-learning models to predict retention times for candidate structures and compared those predictions with experimentally observed retention times for the unknown features. Candidates received higher priority when their predicted chromatographic behavior agreed with the measured signal. This step was especially valuable for distinguishing isomers that could not be separated confidently by formula matching or conventional spectral similarity.

Combining structural neighborhoods and retention-time agreement reduced the candidate space to 418 high-confidence candidate annotations. Among these, 83 were additionally supported by cross-referencing the Human Metabolome Database and the LIPID MAPS Structure Database. The final candidates were also placed in biological context through analyses involving absorption, distribution, metabolism and excretion properties, predicted protein targets, molecular docking and mapping to Kyoto Encyclopedia of Genes and Genomes pathways. Those analyses were applied to the known metabolites used to establish the reference landscape, helping the researchers interpret which chemical families and biological processes were represented in the dataset. The framework could thus prioritize not only molecules that resemble known compounds, but also candidates that fit the chemical and physiological environment of the samples.

The method is not a replacement for authentic standards, high-quality reference spectra or direct structural confirmation. A candidate structure that matches a formula, resembles a known metabolite and has a compatible retention time can still be wrong, particularly when stereoisomers or unusual ion forms are involved. The PubChem search tolerance was also broad enough to generate many chemically implausible possibilities before filtering, and the Tanimoto coefficient does not capture every aspect of molecular behavior. In addition, the test dataset came from a specific population and analytical platform, so retention-time models trained or calibrated in one laboratory may not transfer perfectly to another instrument, column or solvent gradient. Even so, the workflow provides a practical way to decide which unknown signals deserve scarce experimental resources first.

The broader significance is that metabolic dark matter is no longer being treated solely as an obstacle caused by incomplete databases. Unknown features can be interpreted as members of chemical neighborhoods, assessed according to how they travel through a chromatographic system and connected to biological pathways that may explain their origin or importance. By integrating formula-based retrieval, molecular fingerprints, retention-time prediction and biological context, the framework creates an auditable chain of reasoning between an unexplained mass-spectrometry peak and a testable molecular hypothesis. If validated with authentic standards and expanded across tissues, diseases, diets and environmental exposures, such approaches could accelerate the discovery of previously overlooked biomarkers and signaling molecules. For metabolomics, the viral idea is simple: thousands of mysterious peaks may not be noise waiting to be discarded, but clues waiting for the right combination of chemistry, computation and biology.