The idea behind metabolic biomarkers is straightforward: measure molecules in a patient’s blood, urine, or breath, and use them to determine whether something is wrong. Metabolomics, the large-scale study of small molecules in biological systems, has been delivering on that promise in spectacular fashion in research settings for two decades. Yet the gap between discovering promising metabolic biomarkers and bringing them into clinical routine remains frustratingly wide. The reasons are more interesting than simply “science is slow.”
The Measurement Problem: Making Metabolic Biomarkers Reproducible
Before a molecule can be interpreted as a metabolic biomarker or disease signal, it has to be measured reliably.
That turns out to be much harder than it sounds. How a sample is collected, stored, and processed can change what gets detected. The solvent used to extract metabolites from a tissue pulls out some molecules and misses others. Freeze-thaw cycles degrade certain compounds. Even the time of day a blood draw happens shifts the metabolic landscape because of biological variabilityThe molecular universe inside a human, including circadian rhythms.”.
These are systematic, reproducible biases, but they are not biology. They are artifacts of the measurement pipeline, and they have sunk more than a few biomarker candidates when labs tried to reproduce each other’s results.
Breath analysis is where this problem is most visible, and most frustrating, because breath is also where the opportunity for metabolic biomarker discovery is so exciting. Breath is completely non-invasive. A single exhalation carries hundreds of volatile organic compounds reflecting processes occurring in the lungs, liver, and even gut.
For lung cancer, for example, early detection can dramatically change the clinical picture, so a reliable breath test could be transformative. Patients are not going to resist blowing into a tube. The technology exists. And yet, after decades of promising studies, breath-based lung cancer testing has not become part of routine clinical diagnosis.
The core issue is reproducibility: different collection devices, different ambient air conditions, different analytical platforms, and different statistical approaches produce results that look compelling in one lab and fall apart in the next. The biology may be real, but the measurement of it is not yet fully tamed.
The Dark Matter of Chemistry
Even if measurement were perfect, there is a more fundamental obstacle to developing metabolic biomarkers. The molecular universe inside a human body is almost incomprehensibly complex, and only a fraction of it has been confidently annotated.
When a biological sample is analyzed by liquid chromatography-mass spectrometry (LC-MS), the instrument detects signals for thousands of molecules. Roughly 93% of those signals cannot be matched to any known compound in any database. Many of those are real molecules (how many, exactly, is hotly debated), the instrument is capturing them, but their identity is unknown.
An unknown molecule cannot easily serve as a clinical metabolic biomarker.
This is not simply a matter of needing better databases, though better databases would help. It reflects genuine gaps in our knowledge of human biochemistry. Gas chromatography-mass spectrometry (GC-MS) has the opposite profile: most detected compounds are known, but the identification accuracy for individual molecules is around 25%. At the level of molecular families, the accuracy improves to roughly 75%, but 75% is still not good enough for a clinical test. When a physician orders a diagnostic, the test has to know exactly what it is measuring.
There are clear success stories among metabolic biomarkers, such as blood glucose for diabetes, nitric oxide in exhaled breath for asthma, or phenylalanine for phenylketonuria.
p-Cresyl sulfate, a gut microbiome-derived metabolite that accumulates when kidney function declines, is not yet a routine clinical marker, but it has a clear biological story connecting gut microbiome activity to uremic toxicity, and that story grows more compelling with each study.
These molecules work as biomarkers because their connection to disease is understood and the signal is strong enough to rise above the noise of individual variation.
Biology Is Not a Simple Circuit
The deeper problem in metabolic biomarker discovery is that most diseases do not arise from a single metabolic step going wrong in a predictable direction.
Cancer is the extreme case. Tumor cells have radically altered metabolism.
They burn through energy at high rates, generate enormous amounts of reactive oxygen species (ROS), and peroxidize lipids at a scale that floods the body with hydrocarbon breakdown products. In breath, this translates to elevated levels of alkanes, medium chain alcohols, and aldehydes.
The signal is detectable. But here is the catch: it is different for every tumor type, every stage of progression, and every patient.
Lung adenocarcinoma does not “smell” exactly like squamous cell carcinoma. Stage I does not look like Stage III. And every individual carries a personal baseline, a personal diet, and a personal stress response. No single molecule necessarily rises above all of that variation with enough consistency to become universally diagnostic.
Then add the microbiome.
Each person hosts several pounds of microorganisms, predominantly in the gut, producing a chemical output that dwarfs what human cells generate. A large fraction of the molecules detected in urine, blood, and breath are not human in origin. They come from or are transformed by bacteria, fungi, and archaea whose composition shifts with diet, antibiotics, stress, age, and other factors.
Two people with identical diseases can have completely different metabolic profiles because their microbial communities are doing completely different chemistry.
This is fascinating, the microbiome is essentially a personalized chemical factory embedded in every human, and it is also a nightmare for metabolic biomarker development. The signal worth detecting is embedded in noise that is itself biologically meaningful, just not necessarily relevant to the question being asked.
The Path Forward for Metabolic Biomarkers
Medicine has historically been designed to look for one thing. A single number outside a reference range. A pathogen present or absent.
That framework works beautifully when the disease has a clean molecular correlate. It fails when the disease state is better described as a pattern, a shift in the relationships among dozens or hundreds of molecules simultaneously.
The field of metabolic biomarker discovery is increasingly moving toward pattern-based approaches, including machine learning models that can integrate complex metabolic profiles without requiring that each component be individually interpretable. The results in research settings are often striking.
But there is a genuine tension here: a diagnostic test cannot be a black box. A clinician needs to understand why a result means what it means. Regulators need mechanistic justification, not just statistical performance. When a model flags a sample as cancerous based on a 500 dimensional metabolic pattern, the question “why?” cannot be answered with “because the training data said so.”
That understanding is slow to develop, and it is slowing deployment proportionally.
What is changing the picture is scale. Large, publicly accessible metabolomics repositories now hold data from tens of thousands of studies. Analyses that span entire repositories, rather than individual experiments, are beginning to reveal the kinds of consistent molecular patterns that no single study could detect. The picture of complex biological chemistry is getting clearer, and the tools to interpret it are getting sharper.
Still, as of now, routine clinical metabolomics is dominated by targeted approaches: a specific set of molecules, quantified precisely, with known reference ranges. Untargeted metabolomics, running everything and letting the pattern speak, is powerful in research but has not yet become routine in the clinic. Whether and when that changes depends on solving the annotation problem, the reproducibility problem, and the interpretability problem at the same time. Progress is real, but the timeline is unclear.
