2.1. Study design

This case‑control diagnostic accuracy study used urine samples from the Ukraine Association of Biobank (UAB). Clinical diagnoses were established before sample banking, and proteomic analyses were performed in September 2023 on stored specimens, blinded to disease status. Established in 2017, the UAB upholds high ethical standards and predominantly represents the Ukrainian population [21].

This analysis included 264 urine samples from the UAB, comprising 198 patients with early-stage, treatment-naïve disease and 66 healthy controls recruited during routine medical check-ups. Disease groups included kidney, bladder, prostate, ovarian, cervical, endometrial, and melanoma cancers (n = 22 each), multiple sclerosis (MS; n = 22), and metabolic dysfunction–associated steatohepatitis (MASH; n = 22). Cancer and MASH were biopsy confirmed, and MS was confirmed by a neurologist’s evaluation. The sample size was designed for signal discovery and to optimize use of the proteomic assay panel.

2.2. Samples

Urine samples, diluted fourfold to reduce salt content, were analyzed on the Olink Explore 3072 platform, which uses antibody-based detection with next-generation sequencing (NGS). The assay, organized into eight 384-plex panels, targeted 2,850 proteins and incorporated internal and external controls for variability reduction and normalization. Each sample was processed in a single run: overnight antibody–protein incubation, probe hybridization with extension and pre-amplification, and subsequent DNA indexing for library preparation. Protein abundances were reported as Normalized Protein eXpression (NPX) values on a log2 scale, with higher NPX values reflecting greater protein concentrations. Urine samples were not subjected to creatinine normalization prior to Olink assay; the Olink platform incorporates internal extension and amplification controls to reduce intra- and inter-run technical variability, but variation due to urine concentration, hydration status, or urinary tract inflammation cannot be excluded. Renal function measures and creatinine levels were not present in the accessible study data and could therefore not be incorporated into the analysis. All biospecimens were collected, processed, and stored by the UAB following standardized procedures in accordance with ESBB, ISBER, and NCI biobanking guidelines, with urine samples aliquoted and stored at −80 °C and detailed metadata—including storage temperature and freeze–thaw history—systematically recorded [21].

2.3. Protein Measurements

The protein levels in urine were measured using Olink Explore 3072 technology. A comprehensive list of the measured proteins, including their characteristics, is available in Table A in S1 File. All protein measurements for the samples were conducted in a single run. In summary, this technology relies on antibody-based detection to quantify the levels of target proteins. Antibodies, each conjugated with two complementary probes, were organized into eight distinct 384-plex panels. Within each panel, three control assays (interleukin-6 (IL6), interleukin-8 (CXCL8), and tumor necrosis factor (TNF)) were included for quality control purposes.

The assay process commenced with an overnight incubation to facilitate the binding of conjugated antibodies to the target proteins in the samples. This was followed by an extension and pre-amplification step, enabling the hybridization and extension of complementary probes. The extended DNA was subsequently amplified via PCR and indexed to prepare libraries, which were then subjected to sequencing on the Illumina NovaSeq platform. The resulting sequencing counts underwent a rigorous quality control and normalization procedure, incorporating internal controls to minimize intra-assay variability. These internal controls encompassed an incubation control with a non-human antigen, an extension control with a unique probe pair, and an amplification control with a double-stranded DNA sequence. Additionally, external controls, including a negative control (buffer sample) and plate controls (a plasma pool), were employed to establish the limit of detection and standardize measurements between plates. Finally, two known samples were used as controls to assess the precision of the measurements. Following quality control and normalization, the data was presented in NPX units, logarithmically transformed (log2 scale), with higher NPX values indicating higher protein levels. For quality purposes, proteins with more than 30% QC warnings, more than 30% assay warnings, or more than 50% missing frequency were excluded. Olink’s panel performance has been extensively validated for various analytical parameters, including sensitivity, dynamic range, specificity, precision, and scalability. The analytical measuring range was determined by establishing the lower limit of quantification (LLOQ) and upper limit of quantification (ULOQ) for each analyte, reported in pg/mL. Additionally, the potential occurrence of a high-dose hook effect, characterized by antigen excess in comparison to the reagent antibodies, leading to erroneously lower values, has been identified for all analytes (Table A in S1 File). All assays have undergone rigorous validation for precision (encompassing repeatability and reproducibility). Intra-assay variation (within a single run) was calculated by determining the mean coefficient of variation (CV) across 6 individual samples in each of 7 separate runs conducted during the validation studies. Likewise, inter-assay variation (between different runs) was computed as the mean CV across the same 6 individual samples, measured across 7 separate runs during the validation studies.

2.4. Statistical analysis

The analysis had two primary objectives: (1) to assess the bivariate relationship between disease and healthy samples using volcano plots of NPX averages and p-values, and (2) to identify a minimal protein set for disease classification. Given the limited sample size, a two -step approach was adopted to identify a minimal set of proteins capable of classifying samples within a stable model, ensuring general applicability.

In the first phase, random forest analysis was used for initial feature screening; in the second phase, LightGBM classifiers were trained on 100 bootstrap resamples to derive a stable minimal feature set, followed by additional LightGBM training to assess exploratory model stability using the selected biomarkers. The predicted probability for each sample was compared with a threshold to determine whether it was classified as healthy or diseased, with the threshold adjusted to achieve the desired level of specificity. Model performance was evaluated using the area under the receiver operating characteristic (ROC) curve (AUC) within an iterative leave-one-out validation process.

Feature importance scores derived from random forest analysis quantified each protein’s contribution to reducing classification impurity across disease–healthy comparisons. To evaluate cross-disease relevance, concatenated panels were reanalyzed across all disease groups, with normalized Gini-based importance scores reflecting each protein’s relative contribution to classification accuracy and the reduction of sample heterogeneity.

All preprocessing and modeling procedures were conducted in Python (version 3.9.13) using Scikit-learn (version 1.1.2), and supporting libraries including pandas (v1.4.4), numpy (v1.23.3), and statsmodels (v0.13.2) [22–25]. For data visualization, we used matplotlib (version 3.6.0), seaborn (version 0.12.0), and plotly (version 5.10.0) [26–28]. All protein-wise comparisons were two-sided. No correction for multiple comparisons was applied to the 2,850 protein-wise comparisons. The resulting p-values are used solely to rank and visualize proteins for downstream feature selection; they do not support inference regarding differential expression of any individual protein.

2.5. Pathway and functional enrichment analysis

Pathway and functional enrichment analyses were performed using the Reactome database (version 96) via over-representation analysis. For each disease-specific urinary protein panel, enrichment of Reactome pathways was assessed using a hypergeometric test comparing the observed number of proteins in each pathway to that expected by chance based on the human reference proteome. Resulting p-values were adjusted for multiple testing using the Benjamini–Hochberg false discovery rate (FDR) method, and pathways with FDR-adjusted p-values <0.05 were considered statistically significant. Analyses were restricted to Homo sapiens, and non-human identifiers were mapped to their human orthologs where applicable.

2.6. Patient and public involvement

Considering that this was an initial proof-of-concept study, patients and the public were not directly involved in its design. However, they will play a critical role in the dissemination of these preliminary results and in the subsequent validation phase of the research.

2.7. Ethics statement

This study analyzed archived, de-identified urine specimens obtained from the UAB. The UAB Institutional Review Board reviewed the study and determined that it was exempt from formal human-subjects review, as all biospecimens and associated data were fully anonymized prior to analysis.

The urine samples were originally collected by the UAB during routine medical check-ups, under their established ethical protocol that includes broad informed consent for future research use of de-identified materials. For this specific project, the requirement for additional individual informed consent was formally waived by the UAB IRB in accordance with institutional and national regulations.

No new participants were prospectively recruited for this research, and investigators had no access to personally identifiable information at any stage. The study adhered to the principles of the Declaration of Helsinki and all relevant ethical regulations for research involving human participants.

3.1. Urine proteome landscape in healthy individuals

We utilized the PEA technology to detect a total of 3,072 proteins in urine samples and successfully identified 2,850 of these proteins. This subset of 2,850 proteins formed the basis for the exploration of novel biomarkers in urine. Notably, all proteins were directly assessed from urine samples with a fourfold dilution, primarily aimed at minimizing the presence of salt in the samples. No further dilution was necessary for the PEA technology analysis, suggesting that the concentrations of the detected proteins generally remained at lower levels across the entire spectrum of proteins examined.

Our analysis indicated that the concentrations of certain proteins did not display significant differences in relation to the age of the individuals. As illustrated in Fig 1A, specific proteins such as VEGFB and IL11 exhibited relatively higher urinary concentrations in older participants, whereas proteins including BOLA1, ARHGEF5, and NFKB2 showed lower concentrations. Only a limited number of proteins demonstrated p < 0.01, and the lowest observed p-value was 0.002, indicating overall subtle age-related variation in urinary protein concentrations.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Variations in the urine proteome by age and sex.

Volcano plots depicting differential protein abundances in healthy adults aged above and below 35 years (a), and between adult males and females (b). Red points indicate proteins with the lowest p-values; p-values are not corrected for multiple testing and are presented for visualization rather than formal inference.

https://doi.org/10.1371/journal.pone.0354808.g001

Furthermore, urinary protein concentrations varied according to biological sex. Fig 1B depicts these differences in protein abundances between healthy males and females. Notably, Kallikreins were highly sensitive to sex differences: KLK3 and MSMB showed higher concentrations in male urine samples (both p < 1 × 10 ⁻ ⁷), while KLK8 and KLK13 were higher in female samples (both p < 1 × 10 ⁻⁵ ). In total, more than 20 proteins exhibited p < 0.001, suggesting a more pronounced proteomic contrast by sex than by age.

3.2. Urine proteome landscape across diseases

The volcano plots illustrated in Fig 2A provide an overview of differential protein expression across the diseases examined. Because these volcano plots are based on unadjusted p-values, they are intended solely as exploratory, descriptive, and non-inferential visualizations for hypothesis generation. Most conditions exhibited an asymmetric distribution, with an overrepresentation of proteins elevated in patients compared with healthy controls, except for melanoma. Three characteristic patterns emerged. First, in melanoma and, to a lesser extent, endometrial cancer, the volcano plots were largely symmetrical, showing few large-magnitude signals. Second, in cervical, ovarian, and prostate cancers as well as in MS, an asymmetric pattern was observed, although most protein changes were of small magnitude. Third, kidney, bladder, and MASH displayed pronounced asymmetry, with increases in most proteins. These findings underscore the influence of renal excretion on urinary proteomic profiles, which has important implications for disease biomarker discovery. The correlation heatmaps in Fig 2B further illustrate consistent clustering patterns within each disease group, supporting the observed disease-associated urinary proteomic patterns.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Disease-associated urinary protein signatures: volcano plots and correlation structure. 2A.

Volcano plots display differential protein abundances for 2,850 proteins per disease (x-axis, NPX difference between urine samples from patients and healthy controls, where NPX is a log₂-scaled relative abundance unit; y-axis, − log₁₀(p) of the difference between groups). They are exploratory, hypothesis-generating, and non-inferential. P values are unadjusted for multiple testing, and the plots are intended only for visualization and feature-selection purposes; they should not be interpreted as statistically confirmed evidence of differential protein expression. Red points represent the proteins with the lowest p-values. Conditions include bladder cancer (a), cervical cancer (b), endometrial cancer (c), kidney cancer (d), melanoma (e), multiple sclerosis (MS) (f), metabolic dysfunction–associated steatohepatitis (MASH) (g), ovarian cancer (h), and prostate cancer (i). 2B. Pairwise correlation heatmaps of protein abundances within each disease cohort, for the same conditions (a–i) as 2A.

https://doi.org/10.1371/journal.pone.0354808.g002

Pairwise correlation matrices among the top 100 proteins with the lowest p-values revealed distinct disease-specific correlation structures in both patient and healthy control groups (Fig A in S1 File). Venn diagrams of over- and under-represented proteins across conditions (Fig B in S1 File) further illustrated the overlap and specificity of urinary proteomic signatures across disease groups.

3.3. Optimal disease-specific protein signature

Across all disease groups, the classifier’s out-of-sample performance improved with the inclusion of additional proteins, reaching a plateau around five to seven proteins, suggesting a feasible basis for multiprotein urine tests. While this trend was generally consistent, performance saturation occurred more rapidly in certain conditions, potentially reflecting non-differential protein excretion patterns (Fig 3).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Relationship between the size of the protein panel and its diagnostic performance.

Boxplots display the area under the receiver operating characteristic curve (AUC) across 100 bootstrap iterations for varying panel sizes. The x-axis represents the number of proteins in the panel, and the y-axis shows model performance (AUC). Conditions include bladder (a), cervical (b), endometrial (c), kidney (d), melanoma (e), multiple sclerosis (f), metabolic dysfunction–associated steatohepatitis (MASH) (g), ovarian (h), and prostate (i). Y-axis scales are disease-specific and optimized to show the plateau behavior of each panel; they are not intended for direct cross-disease AUC comparison. For standardized AUC values and 95% confidence intervals across all disease groups, see Table B in S1 File.

https://doi.org/10.1371/journal.pone.0354808.g003

Optimal disease-specific performance, defined by AUC values exceeding 0.90, was typically achieved with panels containing at least seven proteins. Ovarian cancer represented an exception, requiring a 15-protein panel to reach maximal accuracy, whereas multiple sclerosis (MS) achieved a lower peak AUC of 0.88 with a five-protein panel, remaining stable up to 17 proteins.

The discriminative performance of each disease-specific multiprotein panel is illustrated in Fig 4, with ROC curves demonstrating strong classification accuracy across all nine disease groups. Across most diseases, the performance of the multiprotein panel exceeded that of any single protein, underscoring the additive contribution of each marker to overall model accuracy. The urinary panels achieved strong discrimination (AUC ≥ 0.95) for seven of nine diseases, with a measurable trade-off between sensitivity and specificity across conditions. The optimal panel sizes and corresponding AUC values for all disease groups are summarized in Table B in S1 File. For prostate cancer, MASH, and melanoma, relatively high sensitivity could be obtained at 99% specificity, whereas for cervical and endometrial cancers, high sensitivity required a substantial reduction in specificity. These findings confirm that disease-specific combinations of proteins provide robust yet variable discriminative performance across biological contexts.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. ROC curve performance of the selected multiprotein panels across disease groups.

Receiver operating characteristic (ROC) curves depict classification performance for each disease model: bladder (a), cervical (b), endometrial (c), kidney (d), melanoma (e), multiple sclerosis (f), metabolic dysfunction–associated steatohepatitis (MASH) (g), ovarian (h), and prostate (i). Blue lines represent individual protein classifiers, and orange lines represent the combined multiprotein panel. The dashed red line indicates random classification.

https://doi.org/10.1371/journal.pone.0354808.g004

Fig 5 presents a heatmap visualization of the normalized Gini impurity–derived feature importance values across different disease models. Each cell represents the relative contribution of an individual protein to disease classification accuracy, with color intensity reflecting normalized Gini importance scores on a relative scale from 0 (white) to the maximum observed value (dark red). This visualization highlights the unique significance of protein sets in distinguishing specific diseases.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Distinctive significance of protein sets for disease detection.

The heatmap displays normalized Gini impurity–derived feature importance values for each protein across disease-specific random forest classifiers. Each cell represents the relative contribution of a protein to model accuracy, with color intensity reflecting normalized Gini importance scores on a relative scale from 0 [white, no contribution] to the maximum observed value [dark red]. The figure highlights both disease-specific and cross-disease protein relevance among the optimized classification panels.

https://doi.org/10.1371/journal.pone.0354808.g005

Most disease detection sets typically comprised seven proteins, except for the ovarian cancer panel, which required fifteen proteins to achieve an acceptable level of specificity. Proteins exhibiting the highest predictive specificity for their respective diseases were the most dominant contributors. Notably, several proteins demonstrated cross-disease relevance. For example, VEGFD, included in the prostate cancer panel, also showed high importance in bladder cancer. Similarly, C9orf40 displayed substantial importance for bladder cancer, even though it was not retained in the final classifier, and PPY contributed equivalently to the detection of both melanoma and MS. Collectively, these findings underscore that the diagnostic performance of each panel arises not from single dominant proteins but from a coordinated combination of disease-specific and cross-disease informative markers.

3.4. Pathway and functional enrichment analysis

Pathway and functional enrichment analyses demonstrated significant enrichment of biologically coherent pathways for eight of the nine disease panels, with top-ranked pathways achieving FDR-adjusted p-values ranging from ~10 ⁻ ⁴ to <0.05. For the melanoma panel, no Reactome pathway reached the conventional FDR < 0.05 threshold; the top-ranked pathways (O₂/CO₂ exchange in erythrocytes, p38MAPK, RAS and NOTCH1 signaling, FDR ≈ 0.06–0.07) are reported as suggestive rather than confirmatory.

Cancer-associated panels were enriched for signal transduction and receptor tyrosine kinase pathways, including STAT3 nuclear events downstream of ALK signaling and broader receptor tyrosine kinase signaling, as well as hormone-related pathways such as estrogen-dependent signaling, glycoprotein hormone pathways, and peptide hormone biosynthesis in hormonally driven cancers. Additional significantly enriched pathways included cell adhesion and epithelial integrity processes (e.g., CDH1 regulation, Type I classical cadherin regulation, and cell–cell communication), transcriptional regulation programs (including NPAS4- and MITF-related pathways), apoptosis pathways (e.g., SMAC/XIAP-mediated caspase activation and apoptosome-related processes), and oxidative stress and detoxification pathways mediated by NFE2L2. Immune and cytokine signaling pathways (including interleukin-4 and interleukin-13 signaling) and cellular response-to-stimuli processes were also represented. In the endometrial panel, the majority of FDR < 0.05 hits derived from a single FLT3 identifier producing multiple near-identical TKI-resistance pathways, and this cluster should therefore be interpreted as a single mechanistic axis rather than independent enrichments.

In multiple sclerosis, the strongest enrichment was observed in NR1H2/NR1H3 (LXR)-mediated signaling and HDL/plasma lipoprotein remodeling (top-pathway FDR ≈ 10 ⁻ ⁴), consistent with the established role of nuclear-receptor–regulated cholesterol homeostasis in myelin biology; immune and antiviral pathways including NF-κB and interferon response signaling were also represented at lower ranks (FDR ≈ 0.035–0.04).

In metabolic dysfunction–associated steatohepatitis, the top enrichments included reversible carbon dioxide hydration (FDR ≈ 0.003) and FOXO-mediated transcription of oxidative-stress, metabolic, and neuronal genes (FDR ≈ 0.014), alongside interferon-α/β signaling (FDR ≈ 0.037).

Additional enriched pathways represented across panels included transport and small-molecule metabolism processes (e.g., fatty acid transport in cervical cancer), reflecting systemic physiological contributions. O₂/CO₂ exchange in erythrocytes was also represented in MASH and melanoma, although these pathways did not reach the conventional FDR < 0.05 threshold in either panel (FDR ≈ 0.05–0.07) and are therefore noted as trending rather than significantly enriched. While broader categories such as signal transduction were observed across panels, pathway-level enrichment patterns revealed distinct mechanistic profiles. Collectively, these findings support the biological relevance of the identified urinary protein panels and their alignment with established molecular pathways.

Limitations

Several limitations of this study should be considered when interpreting the findings. First, the case–control design using archived biobank specimens, while appropriate for discovery-phase diagnostic research, may overestimate diagnostic performance compared with prospective clinical settings. Case and control samples were drawn from distinct cohorts without matching or statistical adjustment for age and sex. Because cases and controls were recruited separately, cohort membership almost entirely coincides with disease status. The observed proteomic differences may therefore reflect cohort-related bias in pre-analytical handling, physiology, or demographic composition rather than disease biology. Age-related variation in urinary protein levels was minimal in this study, which limits age specifically as a source of such bias but does not address the other pathways through which cohort bias may act. Second, the sample size for each disease group was modest (n = 22 per group), which, although adequate for signal discovery and optimized for the proteomic assay panel, reduces statistical power and increases the risk of overfitting despite the leave-one-out validation strategy employed. Replication in larger, independently collected cohorts is necessary to confirm the robustness of the identified panels. The 100-iteration bootstrap resampling procedure was used to assess classifier stability, providing approximate rather than formal inferential confidence intervals. Mean AUC estimates were derived from leave-one-out cross-validation and are independent of the number of bootstrap iterations, although confidence interval precision may be limited by the relatively small number of resamples.

Third, urinary protein concentrations were not adjusted for hydration status, creatinine levels, osmolality, or specific gravity measurements — physiological variables known to influence protein excretion rates. This may introduce additional variability and potentially confound disease-specific signals, particularly for conditions not directly involving the kidney or urinary tract. Systematic measurement and correction of these variables — through creatinine normalization or osmolality adjustment — are identified as priorities for the planned prospective validation studies.

Fourth, all samples were derived from a single biobank representing predominantly the Ukrainian population, limiting ethnic and geographic diversity. The identified biomarker panels may not perform equivalently across populations with different genetic backgrounds, comorbidity profiles, or environmental exposures.

Finally, this study lacks external or prospective validation; all performance estimates are derived from the same biobank dataset, and independent validation in prospectively collected cohorts is required before any clinical utility can be established.