Multi-cancer detection: a four-metabolite urine and plasma panel read by a label-free model reaches an AUC of 0.925 — but a 6.7% positive predictive value at 0.5% prevalence

A team at the US National Cancer Institute (NCI) measured four metabolites in the urine and plasma of 4,105 people, then applied a clustering algorithm that uses no "cancer / healthy" labels to separate cases of eight cancers from controls: AUC 0.925, versus 0.991 for a supervised random forest. The signal is real and the method is elegant. But this is a preprint and an internal case–control study, and at a realistic prevalence of 0.5% the positive predictive value falls to 6.7%: the authors themselves write that the panel is not suitable for population screening.

The context

"Multi-cancer" blood or urine tests are among the most debated topics in early detection: one sample, several cancers sought. The difficulty is well known. A screening test needs very high specificity (the ability not to alarm healthy people) because cancers are rare in the general population; and supervised models trained with case/control labels risk memorising the quirks of their cohorts rather than the biology.

This study, posted on medRxiv on 14 October 2026, tries to sidestep part of the problem. It starts from a narrow panel of four metabolites measurable in urine or plasma, already studied by the NCI's human carcinogenesis laboratory, and asks: if no label is used to build the model, how much of the signal remains?

The method

The data. 4,225 samples (3,875 urine, 350 plasma) from 4,105 individuals across six cohorts: 2,555 case samples (prostate 1,011, lung 661, liver 287, breast 233, colon 215, mesothelioma 93, glioblastoma 30, pancreas 25) and 1,670 controls. All samples were taken at or near diagnosis. The authors have no information on controls' benign conditions or comorbidities.

The measurements. Four metabolites are assayed by tandem mass spectrometry (UPLC-MS/MS, which separates then identifies molecules by mass): creatine riboside, N-acetylneuraminic acid (NANA), cortisol sulfate and a sterol glucuronide. Concentrations are log-transformed and standardised (Z-scores) within each sample type; a "combined score" is the mean of the available Z-scores, with no imputation of missing values (not every analyte was measured in every cohort).

The supervised model. Eight standard algorithms are compared (logistic regression, elastic net, random forest, gradient boosting, SVM, multilayer perceptron, naive Bayes, stacked ensemble), with five-fold cross-validation grouped by individual. The random forest (an ensemble of decision trees that vote) is the reference.

The "label-free" model. A Gaussian mixture model (GMM, which assumes the data come from several statistical "clouds" and recovers them unaided) is fitted to the combined score of all samples pooled. Components whose mean exceeds the median are declared "elevated"; the risk score is the probability of belonging to one of them, thresholded at 0.5. Labels are used only once, after the model is frozen, to compute performance. The authors add unsupervised comparators (k-means, hierarchical clustering, PAM), a partial-label model, and confounding, survival and conformal-prediction analyses.

The results

ML metrics. The random forest reaches an AUC of 0.991 (95% CI 0.989–0.993); the label-free model 0.925 (0.917–0.934), about 93% of supervised performance. (AUC is the probability that the model gives a higher score to a case than to a randomly drawn control; 0.5 is chance.) According to the tables, the label-free model has a sensitivity of 0.827 and a specificity of 0.941. Discrimination ranges from 0.996 (mesothelioma) to 0.707 (liver). Two analytes are enough on average to reach an AUC of 0.948; cortisol sulfate alone gives 0.963. Among cases, a higher score is associated with shorter survival (hazard ratio 1.33 per standard deviation, 95% CI 1.24–1.42).

Clinical translation. At the study's prevalence (about 60% cases, an artefact of the case–control design), per 1,000 samples the model would detect about 500 of 605 cases, miss 105, and wrongly flag 23 of 395 controls. At a screening prevalence of 0.5% (the authors' assumption) the picture is very different: per 1,000 people tested, about 4 of 5 cancers would be detected and 1 missed, but about 59 healthy people would receive a positive result (editorial calculations from sensitivity and specificity). The authors report a positive predictive value of 6.7%: about 67 cancers per 1,000 positive results, i.e. 14 to 15 work-ups per cancer found. The negative predictive value, 0.999, is excellent but mechanically driven by the rarity of the disease.

What is good

A construction without label leakage. Normalisation, fitting and thresholding of the mixture model use no labels; performance is computed afterwards. This is an honest answer to the overfitting critique and gives a more credible floor than the random forest's 0.991.

Checks against artefacts. The authors test whether the mere pattern of missing values (which reflects the source cohort) lets one guess the cancer type: a model using only that pattern reaches a balanced accuracy of 0.448 over eight types, above chance (0.125), showing the risk is real, and they report it. They also adjust for age, sex, ancestry and smoking (AUC 0.932) and compute an E-value (11.7).

Reasonable communication of uncertainty. Conformal prediction (a method that lets the model answer "uncertain" rather than commit) flags 7% of results as uncertain at a 10% error rate, concentrated in liver (45%) and breast (27%). The report follows TRIPOD+AI, funding is public (NIH intramural), and data are available on request.

What is less good

A case–control design with no external validation (spectrum bias, population bias). Cases are diagnosed, often advanced cancers, compared with controls whose benign diseases are unknown. This design almost always inflates performance, as the authors admit. There is no validation on an independent cohort or another measurement platform: generalisation is unknown.

Cohorts confounded with diagnosis (shortcut learning). Controls come from three sources, cases from six; some cancers (breast) are measured only in plasma and on two analytes, and pancreas comes from a commercial biobank. In such a set-up the model can learn "where the sample came from" instead of "who is ill". The missing-value test reduces this risk without removing it, and tissue-of-origin identification (balanced accuracy 0.56 for the label-free model) remains weak.

Figures that do not fully reconcile (misleading metric). The combined-score AUC is 0.922, 0.961 or 0.907 depending on the analysis; organ-identification accuracy is 0.847 in the text and 0.883 in the table; the per-100,000 projection implies a positive predictive value of about 7.5%, not 6.7%. These gaps probably reflect different subsets, but the paper does not always say so. Add small groups (glioblastoma 30, pancreas 25) and a declared conflict of interest: parts of the assay fall under an NCI technology-transfer arrangement.

What it changes

For the research community. The paper offers a useful protocol: establish first what an unsupervised model finds before training a heavy classifier. The logical next step is a prospective cohort with pre-diagnostic samples, controls with benign conditions, and a model frozen before testing.

For clinicians. Nothing changes today. This is a non-peer-reviewed preprint that declares itself unfit to guide clinical practice. The lesson lies elsewhere: a test that performs well in case–control data can remain unusable for screening, because prevalence decides predictive value.

For patients and the public. A multi-cancer urine test is not available and should not be sought on the basis of this work. Keep the reasoning: faced with any announced AUC of 0.9 or more, ask what share of the people testing positive are actually ill.

Going further

Preprint: Patel D.P. et al., "Label-free machine learning detects a multi-cancer signal in a noninvasive urine and plasma metabolite panel: a case–control study", medRxiv, 14 October 2026 (DOI 10.64898/2026.10.13.26365440). Authors: Daxesh P. Patel, Curtis C. Harris, Frank J. Gonzalez and colleagues, National Cancer Institute, NIH (Bethesda), with the University of Colorado Anschutz and Georgetown. Data on request (NCI data-use agreement); code announced on Zenodo upon acceptance. Funding: NIH/NCI intramural programme (ZIA BC 011492).

Editorial transparency: this article was produced with the help of artificial intelligence tools from the preprint's text; figures come from the paper, except the per-1,000 conversions, calculated by the editorial team. The English, Spanish and Chinese versions are machine translations.