Chapter Four · failure evidence
What Metabolic Profiling & Metabolomics got wrong, from 41 dissertations
The records document methodological and practical challenges encountered across metabolic profiling and metabolomics studies. Researchers frequently faced predictive model overfitting, confounding from medications or background disease, analytical detection limits, and failures of in vitro signatures to generalize to clinical settings. These records come from PhD theses at 15 institutions, 2021 to 2026. Each links to its thesis. They were extracted by language models reading the full text, so treat each as a lead to read, not a verdict.
Supervised multivariate models fail validation due to severe overfitting and lack of predictive separation
Multivariate methods such as OPLS-DA and PCA repeatedly produced negative Q2 values or lacked predictive significance despite high apparent model fit. Across multiple biological matrices, metabolic profiles failed to separate treatment, disease, or control cohorts beyond chance.
Tried and failed
multivariate statistical classification using targeted metabolomics applied to dietary pattern differentiation in biofluids. Outcome: no signal. Reason: metabolite profiles exhibited insufficient discriminatory variance across dietary groups resulting in negative Q2 values
Investigating the role of different dietary patterns on gut metabolites · Imperial
Tried and failed
OPLS-DA on untargeted metabolomic profiles applied to distinguishing clinical disease subgroups. Outcome: no signal. Reason: metabolic profiles had insufficient separation between patient subgroups to build statistically significant discriminant models
Tried and failed
O-PLS-DA classification of metabolomic profiling data applied to differentiating disease state from controls. Outcome: no signal. Reason: metabolic features failed to show significant predictive discrimination between groups in cross-validation permutation testing
Metabolic profiling of traumatic brain injury, associated outcomes and treatment in rodents · Imperial
Tried and failed
1H NMR metabolomic profiling with multivariate analysis applied to longitudinal urinary biomarkers for physiological state. Outcome: no signal. Reason: Unsupervised and supervised multivariate models showed no predictive metabolic shift (CV-ANOVA p>0.05, Q2Y<0)
Tried and failed
unsupervised principal component analysis on metabolomic profiles applied to disease case-control stratification. Outcome: no signal. Reason: metabolic variance lacked distinct unsupervised clustering corresponding to clinical disease or outcome groups
Mapping the metabolome in the developing gut · Imperial
Tried and failed
supervised multivariate discriminant analysis of metabolite profiles applied to treatment group classification. Outcome: overfit. Reason: model showed high fit (R2Y) but lacked predictive significance (Q2 p>0.05) and group separation
Tried and failed
supervised multivariate discriminant analysis on untargeted metabolomics applied to longitudinal serum metabolomic profiles. Outcome: no signal. Reason: metabolic variance between surgery and control cohorts was indistinguishable across multiple gestational timepoints
Pregnancy following bariatric surgery: maternal considerations · Imperial
Tried and failed
supervised multivariate classification on metabolomic spectral profiles applied to intergenerational metabolic fermentation response profiles. Outcome: no signal. Reason: models failed to discriminate metabolic differences between mother and offspring cohorts
Tried and failed
untargeted NMR metabolomic profiling and classification applied to cord blood serum from postoperative cohorts. Outcome: no signal. Reason: models showed negative Q2 indicating severe overfitting and absence of discriminative metabolic differences between groups
Pregnancy following bariatric surgery: maternal considerations · Imperial
Candidate metabolite profiles fail to correlate with or differentiate disease progression and intervention outcomes
Metabolite levels showed no statistically significant differences across disease progression stages, clonal differences, or surgical and dietary intervention outcomes. Standalone metabolites and VIP-ranked pathway features also failed to correlate with disease severity scores or yield significant biological pathways.
Tried and failed
plasma metabolite biomarker correlation with disease phenotype applied to metabolic disorder epilepsy severity and drug-responsiveness. Outcome: no signal. Reason: circulating metabolite levels did not correlate with presence of seizures or treatment responsiveness
Tried and failed
untargeted metabolomic profiling and statistical testing applied to distinguishing genetic clones from non-clones. Outcome: no signal. Reason: Rank-sum test yielded no statistically differentiating metabolites between clonal and non-clonal samples.
Metabolomic characterization of coral species susceptible to stony coral tissue loss disease · Georgia Tech
Tried and failed
targeted liquid chromatography mass spectrometry metabolomics applied to stratifying disease severity stages. Outcome: no signal. Reason: metabolic profiles showed no statistically significant differences across disease progression stages
Upper gastrointestinal disease in familial adenomatous polyposis · Imperial
Tried and failed
untargeted metabolomic profiling via mass spectrometry applied to disease staging and demographic stratification. Outcome: no signal. Reason: no statistically significant metabolite differences detected across disease stages or patient demographics
Upper gastrointestinal disease in familial adenomatous polyposis · Imperial
Tried and failed
partial least squares regression applied to plasma metabolomics for clinical disease markers. Outcome: no signal. Reason: metabolite profiles showed minimal to no correlation with target disease burden and cognitive scores
Machine Learning Approaches for Characterizing ALS Disease Progression · MIT
Tried and failed
Baseline metabolomic profiling for outcome classification applied to surgical diabetes resolution prediction. Outcome: no signal. Reason: No statistically significant differences in baseline metabolite levels between resolution and non-resolution groups
Metabolic insights into bariatric surgical diabetes resolution · Imperial
Tried and failed
baseline biomarker profiling to predict intervention outcomes applied to phenotypic response to dietary intervention. Outcome: no signal. Reason: baseline clinical, metabolic, and endocrine characteristics showed no statistically significant correlation with phenotypic improvement
Tried and failed
metabolite set enrichment analysis on ranked metabolites applied to disease progression regression model features. Outcome: no signal. Reason: VIP-ranked metabolites yielded no statistically significant biological pathways
Machine Learning Approaches for Characterizing ALS Disease Progression · MIT
Tried and failed
univariate biomarker classification applied to necrotizing enterocolitis diagnosis. Outcome: no signal. Reason: individual pathway metabolites lacked sufficient discriminatory power as standalone diagnostic markers
Mapping the metabolome in the developing gut · Imperial
Computational and algorithmic assumptions fail during metabolite annotation, network representation, and feature matching
Assumed statistical independence in isotopic labeling, uniform matching constraints in optimal transport, and unpruned metabolite network topologies led to invalid representations and excessive connectivity. Furthermore, learned molecular fingerprints, simple mass spectrometry scoring heuristics, and in silico annotation tools suffered from poor discriminatory power or generated false structural identifications.
Tried and failed
binomial distribution modeling of metabolic isotopic labeling applied to partially labeled peptide mass spectra. Outcome: did not generalise. Reason: tracer incorporation is correlated across metabolic pathways rather than being statistically independent and uniformly scrambled
Tried and failed
deep learning based molecular fingerprints applied to metabolite structural similarity and phenotype prediction. Outcome: worse than baseline. Reason: Learned and hybrid representations underperformed simple predefined substructure fingerprints for structural grouping.
Computational prediction of health status from the human gut microbiome and metabolome · MIT
Considered and rejected
Considered and rejected: Rejected the Kolmogorov-Smirnov test MDS embedding procedure from MDITRE for metabolites because it failed to achieve substantial dimensionality reduction, replacing it with an explicit 90% variance threshold in PCoA.
Computational prediction of health status from the human gut microbiome and metabolome · MIT
Considered and rejected
Considered and rejected: Rejected standard balanced Gromov-Wasserstein for LC-MS feature matching because exact marginal constraints force all metabolites to match, failing when datasets have non-overlapping features.
Inference from Limited Observations in Statistical, Dynamical, and Functional Problems · MIT
Considered and rejected
Considered and rejected: Rejected using experimental MS Z-score, RT score, and PPM score to rank putative metabolites due to near-random discriminatory power (AUCs ~0.55-0.59).
Detection and analysis of binding sites and protein-ligand interactions · OpenBU
Considered and rejected
Considered and rejected: Rejected unpruned full genome-scale metabolite networks because non-specific ubiquitous metabolites caused excessive connectivity.
Exploring the network’s world: From omics-driven machine learning workflow for drug target identification to quantification of signaling model diversity. · IRIS - UNITN - prod
Considered and rejected
Considered and rejected: Rejected using incremental AUC (iAUC) because metabolite values dropped below baseline.
The effects of exogenous ketosis on human metabolism, physiology, and endurance exercise performance · Oxford
Tried and failed
in silico mass spectrometry metabolite annotation applied to bacterial natural product identification. Reason: generated a false positive structural identification disproven upon manual evaluation
Metabolomic characterization of coral species susceptible to stony coral tissue loss disease · Georgia Tech
Analytical limits and sample preparation complexities prevent reliable detection or quantification of low abundance metabolites
Target compounds such as specialized plant metabolites and circulating steroids went undetected or unquantified due to extraction complexity and low ionization intensities. High missing-value rates and instrument sensitivity thresholds in NMR and mass spectrometry resulted in the exclusion of substantial numbers of candidate metabolites.
Tried and failed
cross-platform metabolite quantification comparison applied to low-abundance urinary metabolites. Outcome: data insufficient. Reason: metabolites fell below NMR quantification limits, yielding too few valid measurements for cross-platform correlation
Lost to a baseline
Metabolomics analysis excluded 11 of 58 identified metabolites because they had 57-92% missing values, violating the 80% rule
Lost to a baseline
Single-quadrupole GCMS SIM methods failed to quantify low-abundance steroids and their metabolites in mouse serum due to extraction and derivatization complexity.
Gut bacterial metabolism of corticoids and other host-produced steroids · Harvard
Tried and failed
untargeted LC-MS metabolomics profiling applied to plant secondary metabolite detection. Outcome: no signal. Reason: failed to detect target specialized metabolites previously documented in literature for multiple taxa
Exploring quinolizidine alkaloids : built-in pesticide for the Genisteae tribe (Fabaceae) · UT Austin
Considered and rejected
Considered and rejected: SIRIUS in silico metabolite annotation was rejected for remaining nodes in the B. dendrobatidis network due to insignificant ion intensities in the crude extract.
Chemical Investigation of Chytrid Fungi and Amphibian Skin Microbiome Bacteria · Publikationssystem UB Tuebingen
Confounding clinical covariates and multiple testing corrections eliminate candidate metabolic biomarker signals
Metabolic variation was frequently dominated by confounding variables such as patient medications, non-specific end-stage tissue damage, or primary disease states rather than the target condition. Candidate associations identified during profiling also failed to survive adjustment for multiple hypothesis testing and metabolic covariates.
Tried and failed
mass-spectrometry metabolomic profiling with multiple testing correction applied to cancer risk biomarker discovery. Outcome: no signal. Reason: No individual metabolites remained statistically significant after adjusting for multiple hypothesis testing.
Tried and failed
multivariate statistical analysis of NMR metabolomic profiles applied to stratifying patients by body mass index. Outcome: no signal. Reason: metabolic variance was dominated by primary disease state rather than obesity
Metabonomic profiling in inflammatory bowel disease: application to real-life population of patients · Imperial
Tried and failed
multivariable biomarker association testing with multiple testing correction applied to circulating metabolite and inflammatory marker associations. Outcome: no signal. Reason: association lost statistical significance after adjusting for confounding metabolic covariates and multiple testing correction
Application of novel technologies to cardiovascular prevention · Harvard
Considered and rejected
Considered and rejected: Rejected sampling urine metabolites from diagnosed CKDnt cases because resulting metabolite profiles reflect non-specific structural end-stage damage rather than early causal etiology.
Role of endogenous and exogenous factors in chronic disease development and progression · OpenBU
Tried and failed
multivariate classification of NMR metabolic profiles applied to ulcerative colitis biomarker discovery. Outcome: no signal. Reason: separation was driven by medication metabolites rather than underlying disease pathophysiology
Metabonomic profiling in inflammatory bowel disease: application to real-life population of patients · Imperial
In vitro and controlled trial metabolic biomarkers fail to generalize to primary tissues and free living populations
Dietary biomarkers established in controlled intervention studies exhibited systematic bias and poor agreement when deployed in free-living individuals. Similarly, metabolic trends characterized in cell culture failed to replicate or showed opposite directions of effect in primary tumor tissues.
Tried and failed
Multivariate metabolic profiling for biomarker identification applied to personalized dietary intake assessment. Outcome: did not generalise. Reason: Identified biomarkers lacked specificity across diverse demographics, leading to contradictory and uninterpretable dietary feedback
Tried and failed
metabolomic profiling via GC-MS applied to primary tumour tissue validation. Outcome: did not generalise. Reason: metabolic trends observed in in vitro cell lines were opposite or absent in primary tumours
Metabolomic characterisation of malignant pleural mesothelioma cell lines and tumours · Imperial
Tried and failed
supervised multivariate regression on metabolic profiles applied to dietary assessment in free-living individuals. Outcome: did not generalise. Reason: controlled trial biomarkers had poor agreement and systematic bias when applied to free-living populations
Applying metabolic profiling as an objective dietary assessment method for personalised nutrition · Imperial
Left open by the authors
Problems the authors named and did not get to.
Left open
Validate associations between intra-individual metabolite variability and clinical risk profiles using longitudinal metabolomics data from cohorts with metabolic diseases. Blocker: Access to longitudinal, repeated-measurement clinical metabolomics datasets with disease outcome annotations.
Exploring intra-individual metabolite variation in metabolomic data · Research Repository UCD
Left open
Compute Spearman correlations between microbiome taxonomic abundances and urinary metabolite levels and generate heatmap correlograms. Blocker: Requires the private patient 16S sequencing and 1H-NMR metabolomics datasets generated in the thesis
Metabonomic profiling in inflammatory bowel disease: application to real-life population of patients · Imperial
Left open
Integrate microbial composition data with participant metabolite profiles to identify bacterial strains associated with specific metabolite surges. Blocker: Requires private paired microbial sequencing and participant-specific metabolomic data from the thesis study.
Investigating the role of the gut metabolome in appetite regulation and obesity · Imperial
Left open
Replicate and validate the 204-metabolite ultra-processed food signature in independent cohorts using alternative metabolomics platforms. Blocker: Requires access to controlled/restricted independent cohort metabolomics and dietary datasets.
Left open
Determine whether lysosome-related organelle-dependent modular metabolite biosynthesis is conserved in mammalian systems. Blocker: Requires wet-lab experimental biology and mammalian cell/tissue metabolomic profiling
COMBINATORIAL ASSEMBLY OF MODULAR METABOLITES VIA CARBOXYLESTERASES IN NEMATODES · Cornell
Left open
Identify and characterise unknown LC-MS metabolomic features (e.g., m/z 162.0533, 198.0386, 166.0131) elevated during Drosophila bacterial infection. Blocker: Requires original raw MS/MS spectra, physical reference standards, or wet-lab validation to confirm metabolite identities
A genetic and metabolomic analysis of the Drosophila melanogaster innate immune response · Imperial
Left open
Identify the specific metabolomic features driving Telmisartan's divergent metabolic separation over the 0-8 h time course dataset. Blocker: Requires private experimental mass spectrometry metabolomics time-course data from the thesis author's lab.
Development of Metabolomics Approaches to Decipher Chemical Interactions in Microbial Communities · Publikationssystem UB Tuebingen
Left open
Conduct metabolomic studies to evaluate whether taurine metabolites provide a better therapeutic effect than taurine itself in taiep rats. Blocker: Requires wet lab facilities, biological samples, and metabolomic profiling instruments (e.g., mass spectrometry).
Efecto de la administración de taurina sobre citoesqueleto, mielina, actividad motriz y neurogénesis en la rata taiep · Repositorio Institucional BUAP
Left open
Perform fecal metabolomic profiling to correlate excreted microbial metabolites with circulating and tissue metabolic outcomes in the mouse cohort. Blocker: Requires wet-lab metabolomics instrumentation and physical fecal samples from the animal cohort
Understanding the Interactions Between Maternal PBDE Exposure and the Functional Gut Microbiome in Developing Mouse Offspring · ResearchWorks
Left open
Perform metabolomic and metatranscriptomic profiling to identify and quantify novel secondary metabolite/PKS products linked to gene expression in Gambierdiscus microbiomes. Blocker: Requires wet-lab experiments for metabolomics and metatranscriptomic sequencing of Gambierdiscus microbiomes.
Novel microbes and viruses with roles in biogeochemical cycling and eukaryogenesis in marine systems · UT Austin
Checking a claim in this area?
We can run the same search on any method or claim. If nothing turns up, we will say so, and that proves nothing on its own.