Introduction Urine offers a noninvasive and low-cost source of disease biomarkers, yet most proteomic studies have targeted single conditions. Using deep proteomic profiling and machine learning, we evaluated whether urinary protein signatures distinguish early-stage cancer, neurologic, and metabolic diseases from healthy controls. Methods This case–control diagnostic accuracy study analyzed urine samples from the Ukraine Association of Biobank (UAB), collected during routine medical check-ups. The study included 22 patients each with kidney, bladder, melanoma, prostate, ovarian, endometrial, and cervical cancers; 22 with multiple sclerosis (MS); 22 with metabolic dysfunction–associated steatohepatitis (MASH); and 66 healthy controls, yielding 264 samples analyzed in September 2023. Proteomic assays were performed using the Olink Explore 3072 platform, with laboratory personnel blinded to disease status. Urine proteomes were profiled to identify disease-specific protein signatures. The primary outcome was diagnostic accuracy of multiprotein urine panels for each disease compared with healthy controls, expressed as the area under the receiver operating characteristic curve (AUC), along with sensitivity and specificity calculated at a prespecified threshold. Results Expected sex‑specific differences (KLK3, MSMB higher in males; KLK8, KLK13 higher in females) supported assay validity. Three expression patterns were observed: (1) few strong, symmetric signals in melanoma and endometrial cancer; (2) asymmetry with many up‑regulated proteins in cervical, ovarian, and prostate cancers and in MS; and (3) broad up‑regulation in kidney and bladder cancers and in MASH. Multiprotein models outperformed single proteins, plateauing at five to seven. Maximum AUCs ranged from 0.88 (MS) to 0.98 0.97 (ovarian cancer), with AUCs ≥ 0.95 for seven of nine diseases. Some proteins (e.g., C9orf40, PPY) showed cross‑disease importance. Conclusions Urine proteomics identified disease‑related signals across cancer and metabolic conditions and may enable accurate, noninvasive classification using multiprotein panels, but given the exploratory design and its limitations, the reported accuracies should be regarded as upper-bound estimates requiring prospective validation.
Budnik et al. (2026) studied this question.