Machine Learning for Biomarker Discovery: Finding Disease Signals in Biological Noise

Machine learning biomarker discovery is the work of teaching algorithms to find measurable biological signals, such as a protein level, a methylation pattern, or a metabolite ratio, that indicate disease, predict its course, or forecast who will respond to a drug. The data are enormous: one blood sample can yield thousands of protein measurements, and a genome adds millions of variants. The harder question is which signals the algorithms find are real, and which vanish in the next patient cohort.

Educational content, not medical advice. Talk to your doctor before changing any medication, treatment, or routine. Information here reflects current research but every person's situation differs.

Key takeaways

• Machine learning is well suited to omics data (genomics, proteomics, metabolomics) because these datasets measure far more variables than there are patients, a regime where classical statistics struggles.
• Liquid biopsy tests that read cell-free DNA methylation can detect many cancer types, but the 142,250-person NHS-Galleri trial did not meet its primary endpoint, which shows how far analytical accuracy sits from proven patient benefit.
• Proteomic "aging clocks" trained on tens of thousands of blood samples predict mortality and disease risk, though what to do with a biological-age number remains an open clinical question.
• Validation is the bottleneck. A model that looks excellent on its training data can perform at chance on an independent cohort.

What a biomarker is, and the three jobs it can do

The U.S. Food and Drug Administration (FDA) and the National Institutes of Health define a biomarker in a shared glossary called BEST as a measurable characteristic used as an indicator of normal biological processes, disease processes, or responses to an exposure or intervention. Blood pressure is a biomarker for cardiovascular risk. Hemoglobin A1c (HbA1c) is a biomarker for long-term blood sugar control. Crucially, a biomarker is not a measure of how a person feels or survives; it is a stand-in, and every stand-in has to earn its trust.

Biomarkers do different jobs, and confusing the jobs is a common source of misleading headlines. A diagnostic biomarker detects or confirms a disease, such as troponin in a suspected heart attack. A prognostic biomarker indicates how a disease is likely to progress regardless of treatment. A predictive biomarker identifies who is likely to respond to a particular therapy, as with HER2 status in breast cancer and the drug trastuzumab. A model trained to do one job often fails when asked to do another.

Why biomarkers are hard to find: too many variables, too few patients

Imagine searching a library of a million books for a sentence that only appears in books by one author, when you have read only 200 of the books. Almost any short phrase will, by pure chance, appear in a few of those 200 and not in the rest. That is the statistical trap of omics data. A plasma proteomics study can measure thousands of proteins in a few hundred people, so there are more candidate predictors than patients, and some combination will look like a perfect separator of sick from healthy by luck alone.

The old approach was hypothesis-driven. Researchers picked a protein already implicated in a disease pathway, measured it in patients and controls, and tested for association. That works when biology is well understood, but it can only find what someone already suspected. Machine learning enables agnostic discovery: measure everything, let an algorithm find the smallest set of variables that predicts the outcome, and then ask why.

To keep the search honest, practitioners lean on tools built for this regime. Regularization methods such as the lasso and elastic net add a penalty for every variable a model uses, so the model keeps only the predictors that earn their place. Random forests average many decision trees to smooth out noise; our explainer on how random forests reduce error shows the logic. Dimensionality reduction compresses thousands of correlated measurements into a few summary axes, as covered in our guide to PCA and dimensionality reduction.

None of these tools removes the underlying risk. They reduce the odds of fooling yourself; they do not eliminate them. That is why the rest of this article spends as much time on what was validated as on what was discovered.

The most common way to fool yourself: peeking

In 2002, Christophe Ambroise and Geoffrey McLachlan published a short paper in Proceedings of the National Academy of Sciences (PNAS) on selection bias in microarray studies. Several published studies had claimed that a prediction rule built from just a few genes could classify patients with almost no error. Ambroise and McLachlan showed that those error estimates ignored a subtle problem: the genes had been chosen using the same samples later used to test the rule, or the gene selection sat outside the cross-validation loop. When the selection step was properly included inside cross-validation, the apparent zero error disappeared.

Think of it as grading your own exam after seeing the answer key. If you use the test data, even indirectly, to decide which variables to keep, the test no longer measures how the model will do on new patients. The same failure appears under the name data leakage in other settings, and it is a reason our piece on why data leakage breaks AI evaluations matters for medicine as much as for software. The fix is procedural: lock the model, then evaluate it on data it has never touched.

Case study 1: metabolites that flagged diabetes risk

Metabolomics measures small molecules in blood or urine and gives a snapshot of current biochemistry. A frequently cited early example is a 2011 Nature Medicine paper by Thomas Wang and colleagues. Using liquid chromatography and tandem mass spectrometry, they profiled baseline blood from 2,422 people who had normal blood sugar and were followed for 12 years. Of these, 201 developed diabetes, and they were compared with matched controls.

Five amino acids stood out: isoleucine, leucine, and valine (the branched-chain amino acids) and tyrosine and phenylalanine (aromatic amino acids). A combination of three of them was associated with more than a fivefold higher risk of future diabetes for people in the top quartile, and the finding was replicated in an independent prospective cohort. That replication step is what made the result credible, and it remains the template for how a metabolic biomarker should be tested.

Two cautions apply. An association is not a mechanism: the amino acids might be signals of insulin resistance rather than causes of anything, and the paper does not claim otherwise. And a biomarker that predicts risk in a research cohort is not the same as a clinical test that changes what a doctor does. For readers curious about the biology underneath, our explainers on insulin and diabetes and the gut microbiome, which produces some circulating metabolites, supply the context.

Case study 2: finding cancer in blood, and the limits of accuracy

Tumors shed fragments of DNA into the bloodstream, and healthy tissues do too. Both carry chemical tags called methylation marks that differ by tissue type and by whether the cell is cancerous. A liquid biopsy sequences these cell-free DNA fragments and uses a machine learning classifier to decide whether a cancer signal is present and which organ it probably came from.

The leading example is the Galleri test from GRAIL, a company that Illumina acquired and then spun off as an independent public company on June 24, 2024. In a 2021 Annals of Oncology paper, Klein and colleagues reported results from an independent validation set of 4,077 participants, 2,823 with cancer and 1,254 without. The test had a specificity of 99.5 percent (meaning about 0.5 percent of people without cancer got a false alarm) and, in the 12 prespecified cancers, predicted the origin of a detected signal correctly 88.7 percent of the time. Overall sensitivity was 51.5 percent, and it climbed sharply with stage:

Cancer stage	Sensitivity in the 2021 validation study
Stage I	16.8%
Stage II	40.4%
Stage III	77.0%
Stage IV	90.1%

Source: Klein et al., Annals of Oncology, 2021, as summarized by The ASCO Post.

The pattern is the heart of the problem. The test is best at finding cancers that are already advanced, and a screening test is most valuable for cancers that are small. GRAIL has since sought FDA premarket approval for the test, which carries the agency's Breakthrough Device designation.

What the NHS-Galleri trial showed

The real-world test came in the NHS-Galleri trial, which enrolled 142,250 adults aged 50 to 77 in England and offered annual screening for three years. GRAIL presented the full results at the American Society of Clinical Oncology (ASCO) meeting on May 30, 2026; they had not yet been peer reviewed at the time of release. The trial's primary endpoint was a reduction in combined stage III and IV diagnoses across 12 prespecified cancers, and it was not met: the incidence rate ratio was 1.03 (95 percent confidence interval 0.92 to 1.14), a statistically insignificant 3 percent increase.

A secondary endpoint looked better. Stage IV diagnoses in those 12 cancers fell by 14 percent overall (rate ratio 0.86, confidence interval 0.744 to 0.998), described by GRAIL as nominally significant, with the reduction growing across screening rounds from 9 percent to 22 percent to 26 percent. Screen-detected cancers rose four-fold. Performance numbers were solid on paper: positive predictive value (the share of positive results that turn out to be cancer) of 52.0 percent, specificity of 99.55 percent, and cancer-origin accuracy of 92.5 percent. But episode sensitivity was 54.7 percent for the 12 prespecified cancers and 30.7 percent across all cancer types.

On September 23, 2026, an FDA advisory panel voted 10 to 0 that the test was safe, 6 to 4 that it was effective, and 7 to 2 with one abstention that its benefits outweighed its risks, for use as a prescription screening test in adults 50 and older alongside standard single-cancer screening. Concerns raised at the meeting included overall sensitivity of about 35 percent in the PATHFINDER 2 study and about 32 percent in NHS-Galleri, sensitivity of 13.6 percent for stage I disease, and the fact that, in the PATHFINDER 2 study, about 73 percent of participants with a cancer signal (159 of 218) underwent at least one invasive diagnostic procedure, including many who did not have cancer. As of the sources reviewed here, the FDA's final decision had not been announced, and the panel's votes are advisory.

Why a "99.5 percent specific" test can still raise many false alarms

A short worked example shows why positive predictive value is so sensitive to how common the disease is. The numbers below are illustrative arithmetic, not trial data. Suppose 100,000 people are screened, 1,000 of them (1 percent) have a detectable cancer, the test catches half of those, and the specificity is 99.5 percent.

• True positives: 50 percent of 1,000 = 500.
• False positives: 0.5 percent of 99,000 = about 495.
• Positive predictive value: 500 out of roughly 995 positives, or about 50 percent.

So even a very specific test produces about one false alarm for every true detection when the target is rare and the test misses half of cases. The real NHS-Galleri figure of 52 percent is in this ballpark. This is also why the false-positive burden, not just the headline specificity, belongs in any honest assessment. Our explainer on why 99 percent accuracy can be useless walks through the same logic for any imbalanced classifier.

Case study 3: aging clocks built from blood proteins

The blood proteome, meaning all the proteins in circulation, reflects the health of many organs at once. Modern affinity-based platforms can now measure thousands of proteins in a small plasma sample, which makes this a natural playground for machine learning.

In 2024, M. Austin Argentieri and colleagues published a proteomic aging clock in Nature Medicine. They used 45,441 UK Biobank participants aged 40 to 70 and a panel of 2,897 plasma proteins. A model of 204 proteins predicted chronological age with a Pearson correlation of 0.94, and a reduced set of 20 proteins retained 91 percent of the accuracy. The clock was then tested in the China Kadoorie Biobank (3,977 participants) and the Finnish FinnGen cohort (1,990 participants) with similar performance, and a person's proteomic age gap was associated with 18 major chronic diseases, multimorbidity, and all-cause mortality.

A 2023 Nature paper from Tony Wyss-Coray's group at Stanford took a different cut. Rather than one clock, Oh and colleagues estimated the age of 11 individual organs from organ-enriched plasma proteins across five independent cohorts totaling 5,676 adults. Nearly 20 percent of people showed strongly accelerated aging in one organ, 1.7 percent were "multi-organ agers," and accelerated organ aging was associated with 20 to 50 percent higher mortality risk. Accelerated heart aging was linked to a 250 percent increase in heart failure risk.

These are real results from large cohorts, and they are still research tools. A clock trained to predict calendar age too perfectly would have nothing left over to say about health, so the useful signal sits in the gap between biological and chronological age. Whether lowering that gap improves anyone's health is not something an observational clock can answer.

Other clocks: methylation and retinal photos

DNA methylation clocks came first. In 2013, Steve Horvath reported a predictor built from 353 methylation sites using about 8,000 samples from 82 datasets covering 51 healthy tissues and cell types. On test data it correlated 0.96 with chronological age with a median error of 3.6 years. A later generation, GrimAge, published in Aging in 2019 by Ake Lu, Horvath, and colleagues, used an elastic net Cox regression that combined methylation-based surrogates for smoking pack-years and 12 plasma proteins to predict time to death.

Images can carry biomarkers too. In 2018, Poplin and colleagues at Google and Verily trained deep-learning models on retinal photographs from 284,335 patients. Validated on two independent datasets of 12,026 and 999 patients, the models predicted age within a mean absolute error of 3.26 years, sex with an area under the curve (AUC, a score where 0.5 is chance and 1.0 is perfect) of 0.97, and major adverse cardiac events with an AUC of 0.70. That last figure is useful and modest, a reminder that "can be predicted" and "is well predicted" are different claims.

Case study 4: the Alzheimer's blood test, and a biomarker that did not come from agnostic ML

A good contrast with the cancer story is the first blood test cleared by the FDA to aid in diagnosing Alzheimer's disease. On May 16, 2025, the agency cleared the Lumipulse G pTau217/β-Amyloid 1-42 Plasma Ratio from Fujirebio through the 510(k) pathway, based on substantial equivalence to an earlier cerebrospinal fluid test.

The FDA reviewed a multicenter study of 499 plasma samples from cognitively impaired adults, compared against amyloid PET scans or spinal fluid results. Among people with positive results, 91.7 percent had confirmed amyloid plaques; among people with negative results, 97.3 percent were confirmed negative; and fewer than 20 percent of results were indeterminate. The test is for adults 55 and older with signs and symptoms of Alzheimer's in specialized care, and the agency says it is not intended as a screening or stand-alone diagnostic test.

Three lessons stand out. The biomarker came from decades of biology about amyloid and tau proteins, not from agnostic algorithmic mining, which is a reminder that hypothesis-driven work still delivers. The evidence was a head-to-head comparison against an established reference standard. And the regulatory claim is narrow, with a defined population and a defined use. Our overview of Alzheimer's disease in 2026 covers the clinical context.

Multi-omics integration: reading the whole orchestra

Each omics layer carries partial information. Genomics shows inherited predisposition, transcriptomics shows which genes are active, proteomics shows which proteins are present, and metabolomics shows current biochemical state. A gene variant might affect disease only through a specific metabolite, which no single-layer analysis would reveal.

Layer	What it measures	Time scale	Example in this article
Genomics	DNA sequence variants	Fixed from birth	Risk variants (not a focus here)
Epigenomics	Chemical tags such as methylation	Slow to moderate	Galleri cancer signal; Horvath clock
Proteomics	Proteins in tissue or blood	Days to weeks	Argentieri and Oh aging clocks; pTau217
Metabolomics	Small molecules	Minutes to hours	Amino acids and future diabetes

One widely used integration tool is Multi-Omics Factor Analysis (MOFA), described by Ricard Argelaguet and colleagues in Molecular Systems Biology in 2018, with the first author at the European Bioinformatics Institute (EMBL-EBI). It is an unsupervised method, meaning it is not told which patients are sick; it looks for the major sources of variation across several omics layers and reports them as latent factors, some shared across layers and some unique to one. The authors demonstrated it on a cohort of 200 chronic lymphocytic leukemia samples.

The analogy is a recording of an orchestra from a single microphone. You hear one tangled waveform, but the music is the sum of violins, cellos, and brass. A factor model tries to pull the sections apart, so a researcher can ask whether the "brass" section tracks with survival. Single-cell technologies add resolution: our piece on single-cell AI covers how cell-by-cell measurements are changing which biomarkers can be linked to which cell types.

Validation: the long road from signal to test

A biomarker has to clear a ladder of questions, and most candidates fall off at the first or second rung. Analytical validity asks whether the assay measures the thing accurately and repeatably. Clinical validity asks whether the measurement is associated with the disease in the intended population. Clinical utility asks whether using the test improves outcomes compared with not using it. The Galleri story is an example of strong analytical validity, plausible clinical validity, and clinical utility that is still being argued.

Rung	Question it answers	What can go wrong	Example in this article
Analytical validity	Does the assay measure accurately and repeatably?	Batch effects, instrument drift	Mass spectrometry and sequencing pipelines
Clinical validity	Does the signal track disease in the target population?	Overfitting, confounding by age or sex	Metabolite panel replicated in a second cohort
Clinical utility	Does acting on it improve outcomes?	Overdiagnosis, false-positive workups	NHS-Galleri primary endpoint

Two institutions try to keep this ladder honest. The FDA's Biomarker Qualification Program has three stages: a Letter of Intent, a Qualification Plan, and a Full Qualification Package. It qualifies a biomarker for a specific context of use, meaning a defined role in drug development, and once qualified it can be used for that purpose across drug programs. Meanwhile, the TRIPOD+AI reporting guideline, published in the BMJ on April 16, 2024, updates the earlier TRIPOD statement with a 27-item checklist for studies of prediction models built with regression or machine learning.

Good practice includes holding out data before any model selection, repeating selection steps inside each cross-validation fold (nested cross-validation), validating prospectively in a cohort collected after the model is locked, and checking whether performance holds across sexes, ancestries, and clinical sites. Our explainers on overfitting and underfitting and on why training and test data must be separated lay out the mechanics. A related caution on interpretation comes from how correlation differs from causation, because a biomarker that predicts disease may play no role in causing it.

Common misunderstandings

"The algorithm discovered the cause." Biomarkers are indicators, and most machine learning models detect correlations. A protein that predicts heart failure may be a symptom of early damage rather than a driver of it, and treating it as a target without further evidence can waste years.

"High accuracy means a useful test." Accuracy, sensitivity, and specificity each answer a different question, and the answer changes with disease prevalence. The worked example above shows a specific test generating about as many false alarms as true detections.

"A big sample guarantees a real result." Large cohorts such as the UK Biobank reduce random error, but they do not remove bias, and a clock that works in one population can degrade in another. The proteomic aging clock's authors tested theirs in Chinese and Finnish cohorts for this reason.

"A better score means a better outcome." The most important test of a screening tool is whether people who use it do better, which is exactly what the primary endpoint in NHS-Galleri measured. Detecting more cancers is not the same as lowering deaths or advanced-stage diagnoses.

"AI replaces the randomized trial." The question of whether acting on a biomarker helps patients is a causal one, and our explainer on why randomized trials are powerful shows why no amount of retrospective modeling substitutes for it.

The takeaway

Machine learning has changed what can be measured and searched, and it has produced working tools in cancer detection, Alzheimer's diagnosis, and biological aging research. But the field's best stories are also its best warnings. The Alzheimer's test succeeded because it was validated against a reference standard and approved for a narrow use. The multi-cancer test posts excellent specificity and still met only a secondary endpoint in a 142,250-person trial. The practical rule of thumb is to ask three questions of any biomarker headline: was it validated in people the model had never seen, what does a positive result actually mean at realistic disease rates, and did anyone show that acting on it improves outcomes?

Frequently asked questions

Q: What makes a biomarker clinically useful? A: It must be measurable reliably in an accessible sample, have a validated threshold that guides a decision, and improve outcomes when used to guide care. FDA and NIH's BEST glossary also notes that a biomarker is not a measure of how a person feels, functions, or survives, so usefulness has to be demonstrated, not assumed.

Q: What is a liquid biopsy and how does it detect cancer? A: A liquid biopsy analyzes tumor-derived material in blood, such as cell-free DNA fragments, and uses a classifier to detect a cancer signal and guess its tissue of origin. In the 2021 Klein study the Galleri test had 99.5 percent specificity but 51.5 percent overall sensitivity, with sensitivity rising from 16.8 percent at stage I to 90.1 percent at stage IV.

Q: How is multi-omics different from using one type of genomic data? A: Each layer captures different biology: DNA variants, gene activity, proteins, or small molecules. Integration tools such as MOFA look for patterns shared across layers, which can reveal relationships no single layer shows, though they add complexity and more ways to overfit.

Q: Why do so many biomarkers fail to replicate? A: Common reasons include small samples relative to the number of variables tested, selection bias when the test data influence which variables are chosen, and differences between cohorts. Ambroise and McLachlan showed in 2002 that apparent near-zero error disappeared once gene selection was handled properly inside cross-validation.

Q: What is an aging clock and how accurate is it? A: An aging clock is a model that predicts biological age from molecular data. Horvath's 2013 methylation clock correlated 0.96 with chronological age on test data, with a median error of 3.6 years, and a 2024 proteomic clock reached a correlation of 0.94. Accuracy for chronological age is not the same as accuracy for predicting health, which depends on the age gap.

Q: Has an FDA-cleared blood test for Alzheimer's disease used a biomarker like this? A: Yes, in May 2025 the FDA cleared the Lumipulse plasma pTau217 to amyloid-beta 1-42 ratio for adults 55 and older with symptoms. It matched amyloid PET or spinal fluid in 91.7 percent of positive results and 97.3 percent of negative results, but it is not a screening test.

Sources

• FDA. About Biomarkers and Qualification (BEST definition and the three-stage qualification process).
• Klein EA, et al. Clinical validation of a targeted methylation-based multi-cancer early detection test using an independent validation set. Annals of Oncology, 2021. Summary at The ASCO Post.
• GRAIL. Full results from the NHS-Galleri trial, May 30, 2026 (company release, not yet peer reviewed).
• Targeted Oncology. FDA panel on the Galleri multi-cancer early detection test, September 2026.
• Wang TJ, et al. Metabolite profiles and the risk of developing diabetes. Nature Medicine, 2011. Abstract summary at BIDMC.
• Argentieri MA, et al. Proteomic aging clock predicts mortality and risk of common age-related diseases in diverse populations. Nature Medicine, 2024. Summary at Harvard Medical School.
• Oh HS-H, et al. Organ aging signatures in the plasma proteome track health and disease. Nature, 2023. Abstract summary at Stanford.
• Horvath S. DNA methylation age of human tissues and cell types. Genome Biology, 2013. Full text on PMC.
• Poplin R, et al. Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning. Nature Biomedical Engineering, 2018. Google Research record.
• FDA. FDA clears first blood test used in diagnosing Alzheimer's disease, May 16, 2025.
• EMBL-EBI. Multi-Omics Factor Analysis (MOFA) training materials.
• Ambroise C, McLachlan GJ. Selection bias in gene extraction on the basis of microarray gene-expression data. PNAS, 2002. PubMed Central record.

---

Read next:

• AI Meets CRISPR: How Machine Learning Is Supercharging Genetic Engineering
• Machine Learning in Drug Discovery: From Molecules to Medicine Faster
• Machine Learning and Pandemic Prediction: Can AI Prevent the Next Outbreak?