Urine metabolites flag cancer without labels: a four-metabolite panel keeps most of supervised separation across eight cancers

A combination of four metabolites measured in urine and plasma separates cancer cases from cancer-free controls, and that separation appears in all eight cancer types studied. Most panels are developed and normalized with access to case-control labels, so it is rarely tested whether the separation is present in the metabolite data themselves. Across 4,225 samples (3,875 urine, 350 plasma) from 4,105 individuals, including 2,555 case samples from eight cancer types and 1,670 control samples, the combined score discriminated cases from controls with an AUC of 0.922, and case-control differences ranged from 2.56 SD in mesothelioma to 0.56 SD in liver cancer.

A Gaussian mixture model that uses no outcome labels recovers most of the case-control separation. A supervised random forest reached an AUC of 0.991 on the same features, and no metabolite study had compared a label-free and a supervised analysis on identical features. The label-free model reached an AUC of 0.925 (95% CI 0.917-0.934) with sensitivity 0.827 and specificity 0.941, retaining 93.4% of the supervised AUC; its normalization used the mean and SD of all samples of the same type, cases and controls combined, so no outcome labels entered any step.

Among cases sampled at diagnosis, a higher combined score was associated with shorter overall survival. The prognostic association had not been reported together across eight cancer types in a multi-cohort setting. Among 2,384 cases with follow-up, the hazard ratio for all-cause mortality was 1.33 per 1 SD increase in combined score (95% CI 1.24-1.42), with colon cancer at 2.85, mesothelioma 2.74, lung 2.02, prostate 1.73, breast 1.56, and liver 1.26, while glioblastoma and pancreatic cancer showed no association.

This work speaks to researchers in metabolite biomarkers and multi-cancer detection: it shows diagnostic-time case-control separation and that a label-free model recovers most of it. The authors propose next steps of prospective pre-diagnostic sampling, external validation in independent cohorts and on other analytical platforms, evaluation in defined high-risk populations with benign and comorbid controls, uniform measurement of all four analytes, and serial sampling within individuals.

Samples were collected at or near diagnosis, so the results describe diagnostic discrimination, and whether these metabolites rise before clinical diagnosis is untested. The case-control design likely inflates performance relative to a screening population, and information on benign disease and comorbidities was not available for controls. Validation was internal, so independent prospective cohorts and a different mass-spectrometry platform are still needed. Assay coverage differed by cohort and was linked to cancer type, the plasma set was small, and the breast cancer cohort was plasma-only with two analytes. Cause of death was not recorded, so the competing-risk analysis relied on a stage-based assumption and is supportive only.