FFPE tissue specimens were procured from the Pathology Department of the Seoul National University Hospital. The diagnoses were reviewed by three board-certified pathologists (MJ, KM, and HR), using hematoxylin and eosin slides, according to the 2016 World Health Organization Classification (18). The IUPs included in this study showed inverted trabeculae, cords, or nests of thin urothelium with intact maturation pattern and no cytological atypia ( Supplementary Figure S1 ). Any patient with a previous history of bladder tumor and/or intravesical treatment was excluded. Clinical information was obtained from the medical records. The regional Institutional Review Board (IRB) approved the experimental protocols (IRB No. H-2009-163-1160).

For proteomic analysis, 31 tissue specimens, consisting of 9 IUP, 12 PUC, and 10 normal urothelium (NU), were included ( Supplementary Table S1 ). All IUP and PUC specimens were cystoscopically resected from the urinary bladder. All PUCs were non-invasive (stage Ta) and 83.3% (10/12) were high-grade. For validation, we performed immunohistochemical staining in an independent validation cohort composed of 25 IUP and 16 PUC with inverted growth ( Supplementary Table S1 ). The inverted growth pattern accounted for variable portions (mean ± S.D., 52 ± 32%) of PUCs with inverted growth. The overall demographics of the specimens for validation were similar to those of the proteomic cohort, except most (81.2%) PUCs with inverted growth in the validation cohort were low-grade. All patients with IUP, except for one, were followed up for 12–52 months (median, 31 months) by urine cytology, cystoscopy, or computed tomography, and no one showed recurrence.

Tandem mass spectrometry (LC–MS/MS) analysis was conducted following the methods used in our previous study (17). Briefly, target areas were macro-dissected from unstained FFPE slides. After filter-aided sample preparation and desalting procedures, a liquid chromatography-tandem mass spectrometry (LC–MS/MS)-based proteomic study was conducted, using a Q Exactive HF-X Hybrid Quadrupole-Orbitrap mass spectrometer (Thermo Fisher Scientific, Waltham, MA) and an Ultimate 3000 RSLC system (Dionex, Sunnyvale, CA), according to the instructions of the manufacturer. The MaxQuant.Live version 1.2 (Max Planck Institute of Biochemistry, Munich, Germany) was used to perform BoxCar acquisition (19). The MS1 resolution was set to 120,000 at m/z 200 for BoxCar, and the acquisition cycle comprised two BoxCar scans at 12 boxes (scaled width, 1 Th overlap) with a maximum ion injection time of 20.8 per box with the individual AGC target set to 250,000. The MaxQuant version 1.6.1.0 (RRID : SCR_014485) (20) was employed with the Andromeda engine (21) to process the MS raw files. In the global parameter, the BoxCar was set as the experimental type. All search parameters were set as the default parameter of the software. For label-free quantification, the iBAQ algorithm was used as part of the MaxQuant platform (22). Raw LC–MS/MS data were uploaded into the PRIDE database (RRID : SCR_003411; Accession ID: PXD027602).

Proteomic data were analyzed using the Perseus software (RRID : SCR_015753, Max Planck Institute of Biochemistry). For comparisons, we performed an analysis of variance (ANOVA) and a two-sided t-test with a permutation-based false discovery rate (FDR) at significance level <0.05. Gene Ontology-biologic process (GOBP) and Gene Ontology-molecular function (GOMF) annotations were explicated using the Toppgene Suite (RRID : SCR_005726) (23). Protein–protein interaction (PPI) network models were constructed from the String database (24) and were illustrated using Cytoscape (RRID : SCR_003032) (25). The canonical pathway data were analyzed through the use of Ingenuity Pathway Analysis (Qiagen, RRID : SCR_008653, Hilden, Germany) (26).

The TCGA bladder cancer (BLCA) dataset was used to determine the impact of the IUP-risk biomarkers on the prognosis of bladder cancer (27), under the R environment (R Foundation for Statistical Computing, RRID : SCR_001905, Vienna, Austria; packages “survival” and “survminer”). The RNA sequencing data were chosen for clinical validation because there is no publicly available cohort that contains both high-throughput proteomics and prognostic information in bladder neoplasms. For 405 patients, the gene expression and survival data were obtained from the cBioPortal for Cancer Genomics (https://docs.cbioportal.org/; RRID : SCR_014555) (28, 29). The prognostic effects of the log2-transformed gene expression levels were assessed by calculating a hazard ratio (HR) with 95% confidence interval (CI) using univariate Cox proportional hazard models. Kaplan–Meier analysis and log-rank test were used to compare overall survival outcomes, according to the low vs. high gene expression levels, using the median as a cutoff.

First, we used a feature selection method for supporting vector machines with radial basis function kernel to choose the proteome with discriminative power between IUP and PUC (30). Next, the high-ranked proteins, selected from the machine learning analysis, were screened using The Human Protein Atlas (https://www.proteinatlas.org/; RRID : SCR_006710) (31, 32). Briefly, the antibody staining intensities (high, score 4; medium, score 3; low, score 2; not detected, score 1) were multiplied by the positive samples proportion showing each staining and then the scores were summed. For multiple antibodies, these scores were averaged. The finalized immunoscores in bladder urothelial carcinoma were compared against the relative fold changes derived from the t-test between IUP and PUC; when a protein was relatively overexpressed in urothelial carcinoma according to the public database but downregulated in PUC compared to IUP in the proteomic analysis, the protein was excluded. The finally selected markers were validated using immunohistochemistry.

A validation cohort, consisting of IUP (n = 25) and PUC with inverted growth (n = 16), was utilized independently from those used for the proteomic analysis. Immunostaining assays for SERPINH1 (1:200, sc-5293, RRID : AB_627757, Santa Cruz Biotechnology, Dallas, TX) and PYGB (1:2,000, HPA031067, RRID : AB_2673722, Sigma-Aldrich, St. Louis, MO) were conducted in IUP and PUC with inverted growth, using an automated BenchMark ULTRA System (Roche Diagnostics, Rotkreuz, Switzerland). The immunostained glass slides were digitally scanned using an Aperio Digital Pathology Slide Scanner AT2 (Leica Biosystems, Buffalo Grove, IL). Expression of SERPINH1 and PYGB was quantified by “H-score” [1 ∗ (% cells 1+) + 2 ∗ (% cells 2+) + 3 ∗ (% cells 3+)], with an interpretation ranging from 0 to 300 (33), using the QuPath platform for bioimage analysis (RRID : SCR_018257) (34). The area under the receiver operating characteristic (AUROC) was calculated using MedCalc version 20.019 (MedCalc Software Ltd, RRID : SCR_015044, Ostend, Belgium) and the optimal level of H-score with corresponding sensitivity and specificity were estimated on the basis of the Youden index.

Overall, the LC–MS/MS proteomic assay identified 9,890 and quantified 5,057 proteins, which were present in ≥20% of all samples, from peptides with high confidence (FDR <0.01). The normalized protein abundance is provided as Supplementary Table S2 . Principal component analysis demonstrated that IUP was closer to PUC than NU ( Supplementary Figure S2 ). Using a one-way ANOVA test among the three groups ( Supplementary Figure S3A ), 698 differentially expressed proteins (DEPs) were identified, and these proteins were used to stratify the risks in IUP compared with PUC and NU ( Figure 1A , left). Specifically, ‘NU-like’ IUP signatures, namely, upregulation of SELENBP1, OGDH, and CKB (total, n = 66) and downregulation of TOP2B, NOC2L, and COA3 (total, n = 83), were similar between IUP and NU, as opposed to PUC ( Figure 1A , left). On the other hand, ‘PUC-like’ IUP signatures, namely, upregulation of TSTD1, EOGT, and CIT (total, n = 120) and downregulation of DPH6, VPS13D, and SHPRH (total, n = 429), were similar between IUP and PUC, as opposed to NU ( Figure 1A , left). We investigated the clinical significance of the 40 most significant proteins of the ‘NU-like’ and ‘PUC-like’ IUP signatures ( Figure 1A , right) through survival analysis using the TCGA database. Univariate Cox analysis identified the significance impact of OGDH (HR = 1.469, 95% CI = 1.105–1.954, p = 0.008), SPON1 (HR = 1.092, 95% CI = 1.019–1.170, p = 0.012), PYGB (HR = 1.277, 95% CI = 1.080–1.511, p = 0.004), EPHX1 (HR = 1.159, 95% CI = 1.026–1.309, p = 0.017), SRP68 (HR = 1.643, 95% CI = 1.118–2.415, p = 0.011), and SETD3 (HR = 1.647, 95% CI = 1.180–2.299, p = 0.003) expression on poor urothelial carcinoma outcomes ( Figure 1B ). Among these proteins, low expression of SRP68 and high expression of SETD3 were concordantly observed in the ‘NU-like’ and ‘PUC-like’ IUP signatures, respectively ( Figure 1A , right). These results are consistent with the ones derived from the BCLA survival analysis using the TCGA data; in the latter, low SRP68 expression was associated with favorable, whereas high SETD3 with poor prognosis ( Figure 1C ).

We investigated the molecular functions associated with upregulated and downregulated proteomes in each group. Overall, the GOBP analysis showed significant enrichment of metabolism in all signatures and especially in the upregulated ‘NU-like’ IUP signatures ( Figure 1D ). The downregulated proteins of the ‘NU-like’ IUP signatures were also enriched in immune response, the upregulated ‘PUC-like’ IUP signatures were biased towards transport/membrane, transcription, and translation, while the downregulated ‘PUC-like’ IUP signatures were enriched in cell–cell interaction, responses to stimuli, and transport/membrane functions ( Figure 1D ). Figure 1E summarizes the top 10 significant GOBP terms of each signature set that showed concordant membership. The PPI networks of the proteins included in these top 10 GOBPs consistently highlighted metabolism in the upregulated ‘NU-like’ IUP proteins, immune response and metabolism in the downregulated ‘NU-like’ IUP proteins, metabolism and biosynthesis in the upregulated ‘PUC-like’ IUP proteins, and cell–cell interaction in the downregulated ‘PUC-like’ IUP proteins ( Figure 1F ). Previous studies identified metabolism, cell proliferation, immune response, and intercellular communication as constitutively altered functions in urothelial carcinoma (35, 36). Enhanced metabolism/biosynthesis functions and decreased cell–cell interaction/adhesion, consistent with what was found in the ‘PUC-like’ IUP signatures, were previously shown to promote PUC by supporting cell proliferation and structural breakdown (35, 37–39). Therefore, the results suggested altered metabolism, biosynthesis, and cell–cell interaction functions were consistent with the ‘PUC-like’ (high-risk) IUP.

To further characterize the distinct pathobiology of IUP, we identified DEPs (permutation-based t-test FDR <0.05) between IUP and PUC, namely, PKP2, PYGB, SERPINH1, and TUBB, and those between IUP and NU, namely, ALDH1L1, JUP, COL14A1, and VIM ( Figure 2A and Supplementary Figures S3A, B ). GOBP-based 2D annotation enrichment analysis, as previously described (40), revealed that IUP was distinctly enhanced in the metabolism of amines and carboxylic acids yet repressed in cell response, transport/membrane, adhesion, interaction, and extracellular matrix (ECM) ( Figure 2B ). Similarly, IUP-common DEPs, or the intersecting proteins derived from the comparisons of IUP with PUC or NU, jointly coded for similar GOBP themes relevant to the oncologic significance of IUP as mentioned earlier, namely, metabolism, cell–cell interaction, cytoskeleton formation, and transport/membrane ( Figures 2C, D ). The representative IUP-common DEPs selected from the top 10 most significant GOBPs, namely, PKP2, ALDH1L1, CKB, SERPINH1, and TUBB, interacted towards upregulated processes related to desmosome formation or metabolism and downregulated processes related to cell–cell interaction, cell activation, and ECM ( Figure 2E ). In line with this, IPA for these IUP-common DEPs confirmed the activation of metabolism (z-score ≥2.0) and inhibition of cytoskeleton formation/cell–cell interaction (z-score ≤−2.0) as the constitutive pathways in IUP ( Figure 2F ). Figure 2G illustrates the activation of the TCA cycle and inhibition of the actin-cytoskeleton signaling in the IUP-common DEPs.

To translate the findings from proteomics to IUP diagnosis in practice, we selected proteome features discriminating IUP from PUC and NU, using support vector machine-based machine learning. The lowest error rates, along with keeping the protein lists short, were achieved at 0.4% between IUP and PUC, corresponding to 10 proteins ( Figure 3A ), and at 0.13% between IUP and NU, corresponding to 3 proteins ( Figure 3B ). We failed to find GOBPs implicated in these proteome sets due to the small numbers. However, GOMF analysis identified aldehyde, glycogen, and redox metabolism enriched for the proteomes of IUP compared to PUC ( Figure 3C ), whereas aldehyde metabolism, cytoskeleton, and cell–cell interaction overrepresented by the proteomes of IUP compared to NU ( Figure 3D ). The results corroborated the indispensable roles of metabolism and cytoskeleton/cell interaction-related functions in IUP compared to PUC and NU.

To prioritize protein biomarkers for IUP diagnosis, the top 10 candidates selected by machine learning analysis between IUP and PUC were further shortlisted, based on the t-test FDR and the machine learning rank order ( Figure 4A ), resulting in the five most robust proteins; SERPINH1, ALDH1L1, PKP2, OGDH, and PYGB ( Figure 4B ). These were additionally narrowed down, based on the similarity between the proteomic profiles and knowledge-based immuno-expression in bladder urothelial carcinoma ( Figure 4B , green heatmap); ALDH1L1, PKP2, and OGDH were excluded due to the discordancy. Finally, SERPINH1 and PYGB were selected as candidate biomarkers for the distinction between IUP and PUC ( Figure 4B ).

Differential diagnosis of IUP and PUC with inverted growth is one of the common pitfalls in pathological diagnosis of bladder neoplasms due to the morphologic similarity (3, 11, 12). We aimed to resolve this challenge by validating SERPINH1 and PYGB using immunohistochemistry in IUPs and PUCs with inverted growth. To achieve our goal, we additionally enrolled an independent cohort comprising IUP (n = 25) and PUC with inverted growth (n = 16) ( Figure 4C ). In a pilot test, discordantly to the proteomic analysis, SERPINH1 appeared to be diffusely expressed in IUPs compared to PUCs with inverted growth (Mann–Whitney p = 0.1333; Figure 4D ); SERPINH1 was precluded from further study. However, PYGB was significantly upregulated in IUP compared with PUC with inverted growth (Mann–Whitney p <0.0001; Figure 4E ), verifying the differential expression found in the proteomic data. The AUROC for the diagnosis of IUP vs. PUC with inverted growth using the PYGB H-score was 0.923 (p <0.0001) and the sensitivity and specificity were 72% (95% CI = 50.6–87.9%) and 100% (95% CI = 79.4–100%), respectively, when H-score 21.4 was set as the cutoff ( Supplementary Figure S4 ). The morphology and clinical follow-up of IUPs were similar regardless of whether the immunostaining to PYGB was positive or not. Therefore, we propose that PYGB might be a useful immunohistochemical biomarker for differentiation of IUP from low-grade PUC with inverted growth.