Chapter Four · failure evidence
What Biomarker Discovery & Profiling got wrong, from 52 dissertations
Biomarker discovery and validation studies frequently fail to detect robust signals that differentiate disease states or correlate with clinical outcomes. Candidate signatures also repeatedly fail to generalize across independent cohorts, lose significance after multivariable adjustment, and provide no incremental value beyond existing clinical baselines. These records come from PhD theses at 17 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.
Candidate diagnostic biomarkers fail to differentiate disease cases from healthy controls
Evaluated metabolomic, proteomic, and microRNA candidates frequently showed no statistically significant differences between healthy individuals and disease groups. Individual markers and candidate multiplex panels exhibited weak predictive power and failed to provide meaningful diagnostic discrimination.
Tried and failed
statistical correlation analysis of circulating biomarkers applied to elevated disease plasma proteins. Outcome: no signal. Reason: biomarkers lacked statistically significant correlation despite all showing elevated expression levels
The impact of multiple myeloma and carfilzomib on the endothelium · 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
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
Considered and rejected
Considered and rejected: Rejected using CD25 as a standalone sepsis biomarker because it showed no significant difference between septic patients and healthy controls
Studies of Microfluidic Affinity Cell Separation for Bioanalysis and Clinical Application · Texas Tech
Tried and failed
predictive regression using single metabolomic biomarkers applied to perinatal depression prediction. Outcome: no signal. Reason: individual metabolite levels showed weak correlations and lacked statistical predictive power
Phenotyping perinatal depression : exploring interactions among biopsychosocial and behavioral determinants · UT Austin
Tried and failed
biomarker association regression analysis applied to metabolic and blood pressure outcomes. Outcome: no signal. Reason: circulating biomarker levels showed no significant correlation with glycemic control or blood pressure measures
VITAMIN B12, ONE-CARBON METABOLISM, AND METABOLIC HEALTH IN WOMEN OF REPRODUCTIVE AGE · Cornell
Tried and failed
spectrophotometric biomarker profiling of oxidative stress applied to distinguishing disease states in serum. Outcome: no signal. Reason: biomarker concentrations showed no statistically significant difference between healthy and diseased groups
Oxidative stress in critically ill neonatal foals · Iowa State
Tried and failed
multivariable linear regression of biomarker variance applied to predicting clinical health disorder prevalence. Outcome: no signal. Reason: biomarker mean and variability showed no statistically significant association with disease incidence
Tried and failed
candidate biomarker panel screening in biofluids applied to disease risk discrimination in saliva cohort. Outcome: no signal. Reason: candidate microRNA markers failed to show statistically significant differences or discriminative ability (AUC 0.5-0.6) in validation
A COMPARATIVE SPECIES APPROACH TO UNDERSTAND THE ROLE OF MIRNAS IN MAMMARY GLAND HEALTH AND DISEASE · Cornell
Tried and failed
nanopore barcoded probe multiplexed detection applied to circulating miRNA cancer biomarkers in serum. Outcome: no signal. Reason: multiplex panel showed no significant difference between healthy control and cancer serum samples
Candidate prognostic and predictive biomarkers fail to correlate with disease progression or treatment response
Baseline molecular and clinical candidate markers showed no statistically significant correlation with patient survival, tumor regression, or longitudinal motor and cognitive decline. These biomarkers also failed to predict therapeutic response or phenotypic improvement across evaluated clinical cohorts.
Tried and failed
apoptotic and cell cycle biomarker expression profiling applied to predicting radiotherapy tumour response. Outcome: no signal. Reason: biomarker expression levels failed to correlate significantly with tumour regression grade across evaluated datasets
A next-generation -omics analysis of the radiotherapy response in rectal cancer · Imperial
Tried and failed
immunofluorescence biomarker quantification on tissue microarrays applied to breast cancer clinical prognostic stratification. Outcome: no signal. Reason: No statistically significant correlation was found between histone mark intensity and subtype markers or patient survival
Tried and failed
evaluating metabolic gene expression as prognostic biomarker applied to cancer patient survival datasets. Outcome: no signal. Reason: gene expression levels did not statistically correlate with overall or disease-free survival
Delineating the role of malate dehydrogenase 2 in the hypoxia-driven metabolic reprogramming of metastatic cancers · Research Repository UCD
Tried and failed
statistical association of standard clinical prognostic biomarkers applied to disease progression and treatment response. Outcome: no signal. Reason: standard clinical variables lacked statistically significant correlation with minimal residual disease and post-induction clinical outcomes
High-efficiency multi-drug functional profiling in a microfluidic device towards personalized predictions in pediatric leukemia · Georgia Tech
Tried and failed
Correlation analysis for predictive biomarker discovery applied to synthetic lethality across cancer cell lines. Outcome: no signal. Reason: No robust lineage-agnostic or lineage-specific biomarkers or transcription factors correlated with dual dependency.
Systematic Interrogation of CBP/p300 Dependency in Cancer · Harvard
Considered and rejected
Considered and rejected: Rejected soluble full-length Ataxin-3 in PBMCs as a standalone prognostic disease biomarker due to weak, non-significant correlations with disease progression scores (SARA, INAS, CSDP) compared to mutant Ataxin-3 and NfL.
Parkin und Ataxin-3 als potenzielle Biomarker in der Spinozerebellären Ataxie Typ 3 (SCA3) · Publikationssystem UB Tuebingen
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
tertile-based biomarker stratification in regression models applied to longitudinal cognitive decline prediction. Outcome: no signal. Reason: Total baseline cerebrospinal biomarker levels did not differentiate categorical cognitive impairment outcomes over time
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
baseline biomarker quantification via sandwich ELISA applied to longitudinal neurodegenerative motor progression. Outcome: no signal. Reason: Total baseline biomarker levels showed no meaningful correlation with longitudinal disease score progression across 36 months.
Biomarker signatures fail to generalize across independent cohorts and diverse environments
Multi-gene signatures and metabolic biomarkers derived from controlled settings frequently showed directionally contradictory associations when applied to external populations or differing endpoints. In vitro functional screen candidates and imaging markers also failed to replicate in clinical signatures or independent validation cohorts.
Tried and failed
stratified biomarker survival analysis via step-function binarization applied to multi-omics cancer drug response data. Outcome: did not generalise. Reason: biomarker survival association directions contradicted established literature across different cohorts and treatments
Informatics Approaches for Identifying Drug-Specific Markers and Deciphering Genetic Regulation Mechanism in Cancer Treatment · Georgia Tech
Tried and failed
functional genetic screens for biomarker identification applied to cancer chemotherapy response prediction. Outcome: did not generalise. Reason: in vitro functional screen hits did not overlap with published clinical multi-gene prognostic signatures
CRISPRi screens to identify combination therapies for the improved treatment of ovarian cancer · MIT
Tried and failed
proteomic biomarker association analysis applied to cardiovascular disease progression. Outcome: did not generalise. Reason: associations were directionally inconsistent across different clinical endpoints and life stages
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
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
Tried and failed
automated vessel segmentation biomarker tracking across cohorts applied to longitudinal clinical magnetic resonance angiography. Outcome: did not generalise. Reason: vessel count reduction over time observed in one clinical trial cohort failed to reproduce in another
Automated precision computational image analysis and applied machine learning for experimental and clinical hematology applications · Georgia Tech
Tried and failed
evaluating published transcriptomic biomarker signatures directly applied to independent paediatric disease diagnosis. Outcome: did not generalise. Reason: original models failed to meet required clinical diagnostic performance thresholds on independent paediatric cohort
Tried and failed
differential expression analysis across independent cohorts applied to cross-dataset transcriptomic disease biomarker discovery. Outcome: did not generalise. Reason: yielded an overwhelming mass of significant genes with poor explainability and lack of generalizability
Improvements in the Modeling of High Dimension/Low Sample Size Imbalanced Clinical Datasets · Georgia Tech
Novel biomarker panels fail to provide incremental value over simple baselines or single markers
Combining multiple correlated biomarkers or supplementary assay platforms provided no predictive gain over single strongest markers or foundation model embeddings. Novel admission biomarkers and automated scoring systems were consistently matched or outperformed by standard clinical variables and manual expert assessments.
Tried and failed
adding clinical features to logistic regression applied to cancer risk prediction. Outcome: worse than baseline. Reason: additional biomarkers and demographic predictors provided no incremental discrimination over primary biomarker and age
The Utility of CA125 for the Detection of Ovarian Cancer in Primary Care · Cambridge
Tried and failed
combining multiple correlated biomarkers in survival forest applied to disease progression prediction. Outcome: no signal. Reason: multiplexing biomarkers provided no statistically significant predictive gain over the strongest standalone marker
Tried and failed
combining engineered spatial biomarkers with foundation model embeddings applied to survival prediction from histopathology images. Outcome: no signal. Reason: biomarker features provided no additional predictive signal beyond what foundation models already captured
Artificial Intelligence-Based Phenotyping of the Tumor Microenvironment in Hematoxylin and Eosin-Stained Images of Solid Tumors · Georgia Tech
Tried and failed
combining multi-platform assay features for classification applied to protein biomarker diagnostic signatures. Outcome: worse than baseline. Reason: supplementary platform measurements did not provide additional discriminative value beyond single-platform markers
Lost to a baseline
Automated CVS biomarkers (AUC 0.81-0.87) did not replicate the perfect discrimination (100% sensitivity/specificity) previously achieved by manual expert ratings of all lesions using 40% or 50% cutoffs.
Statistical Techniques For Addressing The Clinico-Radiological Paradox In Multiple Sclerosis · Penn
Lost to a baseline
For predicting adverse in-hospital events during AHF, all novel admission biomarkers had lower discrimination than baseline serum creatinine (AUC 0.65).
Validation of novel biomarkers of Acute Kidney Injury · Research Repository UCD
Low abundance, assay insensitivity, and biological variability hinder biofluid biomarker profiling
Candidate circulating microRNAs and proteins were frequently undetectable or lacked sufficient analytic sensitivity and proteomic coverage in biofluids. High intra-subject and inter-subject variability, along with discordance between tissue expression and systemic circulation, obscured meaningful biomarker signals.
Tried and failed
correlating circulating biomarker levels with tissue expression applied to colorectal cancer biomarker validation. Outcome: no signal. Reason: circulating plasma concentrations did not correlate with immunohistochemical tissue expression levels
Tried and failed
urinary multiplex biomarker profiling applied to disease severity stratification and monitoring. Outcome: unstable. Reason: High intra- and inter-subject biomarker variability obscured meaningful clinical signals
Considered and rejected
Considered and rejected: Rejected primary biomarker discovery in urine due to low proteomic coverage when total urine protein concentration is low
Maladaptive Repair in Acute Kidney Injury · JScholarship
Tried and failed
RT-qPCR validation of sequencing biomarker candidates applied to low-abundance biofluid microRNA biomarkers. Outcome: no signal. Reason: Target microRNAs were undetectable in more than half of the biofluid samples
MicroRNAs in extracellular vesicles as biomarkers for pancreaticobiliary cancers · Imperial
Tried and failed
small RNA sequencing for biomarker discovery applied to circulating cell-free RNA in plasma. Outcome: no signal. Reason: insufficient differentially expressed candidates detected in systemic circulation compared to local biofluids
MicroRNAs in extracellular vesicles as biomarkers for pancreaticobiliary cancers · Imperial
Considered and rejected
Considered and rejected: Rejected cardiac troponin from biomarker growth mixture modeling due to insufficient analytic sensitivity and zero-inflation (lack of variability) in the repository assay.
Apparent biomarker signals lose significance after adjusting for covariates and multiple testing
Univariate biomarker associations with clinical outcomes frequently lost significance once co-occurring clinical covariates were incorporated into multivariable models. High-dimensional metabolomic and proteomic features also failed to pass stability selection thresholds or survive multiple hypothesis testing adjustments.
Tried and failed
multivariable regression for prognostic biomarker identification applied to cancer genomic mutations. Outcome: no signal. Reason: Univariate biomarker associations lost statistical significance due to co-occurrence with confounding covariates.
Development of Methods for Cancer Genome Analysis and Clinical Applications · Harvard
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.
Considered and rejected
Considered and rejected: Rejected assessing high-order interactions among 15 biomarkers and baseline covariates due to multiple testing issues and model overspecification risk.
Measuring the impact of reduced antibiotic use in hospital settings using electronic health records · Oxford
Tried and failed
stability selection with LASSO logistic regression applied to high-dimensional proteomic biomarker discovery. Outcome: no signal. Reason: No individual proteomic features met stability selection thresholds after adjusting for clinical covariates.
Molecular phenotyping of severe asthma using statistical and machine learning models · Imperial
Tried and failed
regularised regression biomarker profiling applied to disease risk prediction from metabolomics. Outcome: no signal. Reason: derived dietary metabolomic signatures showed no statistically significant association with disease risk in multivariable models
Left open by the authors
Problems the authors named and did not get to.
Left open
Validate phase angle as a prognostic stratification biomarker in extensive longitudinal cohorts across different cancer types. Blocker: Requires collecting longitudinal clinical and bioimpedance data from large cancer patient cohorts
Left open
Identify clinical biomarkers predicting patient response to low-dose SFK inhibition combined with BRAF/MEK blockade. Blocker: Requires clinical patient cohorts, tissue samples, or wet-lab experimental validation.
Tumor cell-intrinsic signals promoting tolerance and adaptation to oncogenic kinase inhibition · MIT
Left open
Develop predictive biomarkers for patient tumor sensitivity to AHR or kynurenine pathway inhibition. Blocker: Requires clinical trial cohorts, patient tumor tissue samples, and wet lab experimental validation
Novel Functions of the Transcription Factor Aryl Hydrocarbon Receptor (AHR) and Its Tryptophan-Derived Ligands in Cancer Cells · DSpace at UTSWMED
Left open
Validate candidate molecular, genomic, and epigenetic biomarkers from mucosal tissue, serum, or stool for colorectal cancer risk stratification in commercial clinical practice. Blocker: Requires clinical patient biospecimens (tissue, serum, stool) and wet lab diagnostic infrastructure.
Optimising decision-making in the management of dysplasia in inflammatory bowel disease · Imperial
Left open
Validate lipoprotein-cholesterol biomarkers and test lipidomics/genomics profiles across larger, stratified cohorts of cancer patients. Blocker: Requires wet lab facilities, biological patient cohorts/samples, and clinical lipidomic/genomic profiling infrastructure
Protein Engineering for Personalized therapy and diagnosis · EPFL
Left open
Establish quantitative score thresholds in the Biomarker Toolkit to predict clinical translation success using historical biomarker evaluation datasets. Blocker: None
Left open
Validate the 34 lipid signatures associated with medulloblastoma metastasis across human patient cohorts. Blocker: Requires access to human patient cohorts and clinical wet-lab mass spectrometry facilities to validate lipid biomarkers.
Left open
Apply the Biomarker Toolkit scoring checklist to published non-cancer biomarker studies to validate rationale-related evaluation attributes across disease categories. Blocker: None
Left open
Investigate synaptic changes, inflammation, and mitochondrial dysfunction in prodromal Parkinson's disease cognitive deficits using biological or longitudinal assays. Blocker: Requires wet lab biological assays, biomarker data, or specialized longitudinal cohort samples measuring synaptic/mitochondrial markers
Prodromal Parkinson’s Disease and its Implications for Longitudinal Studies · Harvard
Left open
Analyze Parkinson's Disease Biomarkers Program (PDBP) proteomic data across cohorts to evaluate biological implications of molecular clusters. Blocker: PDBP multi-omics clinical and proteomic data requires controlled-access data use agreements and approval.
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.