Dynamic urinary proteomics integrates single-cell and spatial transcriptomics to reveal tumour microenvironment and predict immunotherapy response in biliary tract cancer Authors: Shanshan Wang, Zhengguang Guo, Boyu Sun, Kai Liu, Jiashuo Chao, Ziyu Xun, Yunchao Wang, Zibo Xu, Ziyue Huang, Hao Wang, Yang Tan, Nan Zhang, Mingjian Piao, Longhao Zhang, Chengjie Li, Shuofeng Li, Jiongyuan Li, Haidan Sun, Feng Qi, Aiwei Wang, XiaoBo Yang, Chengpei Zhu, Hanping Wang, Wei Sun, Haitao Zhao Published: 2026 ABSTRACT AbstractBackgroundMost patients with biliary tract cancer (BTC) do not derive durable clinical benefit (DCB) from immune checkpoint inhibitors (ICIs), underscoring the urgent need for predictive biomarkers. While urinary proteomics represents a non-invasive approach for biomarker discovery and mechanism exploration, its utility in ICI-treated patients with cancer remains unexplored.ObjectiveWe aimed to establish urinary proteomics as a predictive tool for ICI responsiveness and to elucidate its relationship with tumour dynamics and tumour microenvironment (TME) remodelling in BTC.DesignWe performed a staged mass spectrometry (MS)-based discovery-validation proteomics workflow in 211 urine samples from 97 treatment-naïve patients with BTC undergoing ICI-based therapy. A machine learning model was developed based on baseline proteomic features for ICI response prediction. Single-cell transcriptomics of 11 pretreatment tumour biopsies and spatial transcriptomics were integrated to explore the link between urinary proteomics and TME.ResultsPatients achieving DCB exhibited enrichment of immune activation and systemic inflammatory pathways, whereas non-durable benefit was correlated with protumourigenic processes. Longitudinal urinary proteomic dynamics could mirror TME remodelling and tumour evolution. A machine learning-derived 4-urinary protein panel (protein tyrosine phosphatase non-receptor 13 (PTPN13), SUB1, MICAL-L1, VARS1) robustly predicted DCB and early responses. Subsequent external validation in an independent cohort (n=24) using parallel reaction monitoring-MS further confirms its generalisability. PTPN13+ malignant cells were identified as key regulators of proapoptotic TME states, contributing to sustained ICI responsiveness.ConclusionsThis study pioneers the application of urinary proteomics in immuno-oncology, providing a non-invasive approach to predict and monitor ICI responsiveness, while offering mechanistic insights into TME dynamics in BTC. WHAT IS ALREADY KNOWN ON THIS TOPIC Reliable biomarkers for predicting immune checkpoint inhibitor (ICI) responsiveness remain elusive. Reliable biomarkers for predicting immune checkpoint inhibitor (ICI) responsiveness remain elusive. Tissue-based biomarkers are limited by invasiveness, sampling bias and inability to capture the plasticity of antitumour immune responses during therapy. Tissue-based biomarkers are limited by invasiveness, sampling bias and inability to capture the plasticity of antitumour immune responses during therapy. Urinary proteomics offers a non-invasive window into real-time human physiological state but remains unexplored in tumour immunotherapy. Urinary proteomics offers a non-invasive window into real-time human physiological state but remains unexplored in tumour immunotherapy. WHAT THIS STUDY ADDS Our study pioneers the clinical investigation of urinary proteomics in tumour immunotherapy, offering a novel and non-invasive approach to predict ICI responsiveness and capture tumour dynamics and the plasticity of tumour microenvironment beyond tissue-based approaches. Our study pioneers the clinical investigation of urinary proteomics in tumour immunotherapy, offering a novel and non-invasive approach to predict ICI responsiveness and capture tumour dynamics and the plasticity of tumour microenvironment beyond tissue-based approaches. Protein tyrosine phosphatase non-receptor 13 emerges as a promising biomarker and potential therapeutic target for enhancing ICI outcomes. Protein tyrosine phosphatase non-receptor 13 emerges as a promising biomarker and potential therapeutic target for enhancing ICI outcomes. HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY Our study showed that proteomics-driven urine biomarker discovery provides a valuable reference for the liquid biopsy. Our study showed that proteomics-driven urine biomarker discovery provides a valuable reference for the liquid biopsy. These preliminary findings position urinary proteomics as a promising non-invasive tool for early prediction and longitudinal monitoring of ICI responsiveness in biliary tract cancer. This approach may enable timely therapeutic adjustments and facilitate broader application of urinary proteomic profiling in immuno-oncology research and clinical practice. These preliminary findings position urinary proteomics as a promising non-invasive tool for early prediction and longitudinal monitoring of ICI responsiveness in biliary tract cancer. This approach may enable timely therapeutic adjustments and facilitate broader application of urinary proteomic profiling in immuno-oncology research and clinical practice. Introduction Biliary tract cancers (BTCs) are rare but highly aggressive malignancies with an increasing global incidence.1 Approximately 70% of patients with BTC are diagnosed at advanced stages, where immune checkpoint inhibitors (ICIs) are the first-line therapy.1 2 However, most patients with BTC fail to derive durable benefit from ICIs.2 3 This clinical reality highlights the critical need for biomarkers to stratify responders and elucidate mechanisms of ICI resistance. Currently, most biomarkers are tissue-based and static, with limited predictive value.46 Their clinical utility is hindered by invasiveness, spatiotemporal heterogeneity and an inability to capture dynamic immune-tumour interactions.46 Furthermore, the efficacy of ICIs is intricately linked to the composition and dynamics of the tumour microenvironment (TME).7 However, the development of dynamic biomarkers to predict and monitor ICI responsiveness and TME plasticity is still in its infancy.8 9 Therefore, it is crucial to identify non-invasive biomarkers that can predict ICI responsiveness and capture the dynamic nature of tumour evolution and the TME under the selective pressure of ICIs. Urine offers a non-invasive and longitudinally accessible matrix for biomarker discovery.1012 Proteins released by malignant, apoptotic or necrotic cells, enter the bloodstream, are filtered by the kidneys and are ultimately excreted in urine.10 As the executors of human biological activities with good stability, proteins are particularly suitable for clinical biomarker development.13 14 Urinary proteomics provides valuable insights into the body’s immune response and TME, positioning urine as an excellent ‘window’ for monitoring pathophysiological changes.13 15 Urinary proteomics is advancing for biomarker discovery, treatment efficacy prediction and health monitoring, but its role in immuno-oncology remains underexplored.10 11 15 Mass spectrometry (MS)-based proteomics is a powerful tool for biomedical and clinical research.16 Data-independent acquisition (DIA)-MS enables unbiased, proteome-wide profiling, improving detection efficiency and sample reproducibility.17 Parallel reaction monitoring (PRM)-MS enables targeted validation with high accuracy.17 Together, the DIA+PRM workflow enables antibody-free, comprehensive validation of candidate biomarkers in a cost-effective manner, facilitating clinical translation and has become a widely adopted strategy for proteomic studies in diverse cohorts.14 16 18 Here, we employed our previously established staged DIA+PRM workflow in a prospective cohort of patients with BTC receiving first-line ICI treatment.11 We developed a urinary proteomic biomarker panel using machine learning to predict ICI responsiveness. Furthermore, by integrating single-cell and spatial transcriptomics, we investigated how urinary proteomic profiles reflect TME patterns, thereby providing molecular insights into the mechanisms underlying both response and resistance to immunotherapy. Materials and methods Detailed materials and methods are provided in the online supplemental materials and methods. Results Study design and characteristics of urine and tissue specimens The study design workflow is illustrated in figure 1. A total of 73 eligible individuals and 187 (64 pretreatment and 123 post-treatment) qualified urine samples were included for DIA-MS analyses (online supplemental figure S1 and table S5,S8, figure 2A). Detailed clinical information is presented in table 1 and figure 2B. Pretreatment tumour biopsy samples from 11 patients were analysed by single-cell RNA sequencing (scRNA-seq), with further clinical details provided in online supplemental table S4. Schematic overview of the study design. Overview of the study design. The illustration was created with BioRender.com. BTC, biliary tract cancer; CA19-9, carbohydrate antigen 19–9; DC, dendritic cell; DEPs, differentially expressed proteins; DIA, data-independent acquisition; ICI, immune checkpoint inhibitor; IS-PRM, internal standard parallel reaction monitoring; LC-MS, liquid chromatography-tandem mass spectrometry; mIF, multiplex immunofluorescence; N, number of patients; NK, natural killer; PD, progressive disease; PR, partial response; scRNA-seq, single-cell RNA sequencing; UMAP, Uniform Manifold Approximation and Projection; 4-UP, 4-urinary protein. Cohorts, samplings, survival stratified by clinical benefit and associations of clinical variables with clinical benefit. (A) Swimmer’s plot showing clinical course, sampling time points and outcome events of each patient in the study cohort. (B) Heatmap representing patient demographics and clinical characteristics. (C) Kaplan-Meier estimates of OS and PFS in the entire cohort stratified by clinical benefit (DCB and NDB). (D) Associations of clinical variables with DCB using simple logistic regression. The measure of centre is the parameter estimate and error bars represent 95% CIs. CA19-9, carbohydrate antigen 19-9; cTNM, clinical tumour, node, metastases; DCB, durable clinical benefit; ECC, extrahepatic cholangiocarcinoma; ECOG-PS, Eastern Cooperative Oncology Group Performance Status; GBC, gallbladder cancer; ICC, intrahepatic cholangiocarcinoma; ICI, immune checkpoint inhibitor; irAE, immune-related adverse event; MTT, molecular targeted therapy; NDB, non-durable clinical benefit; NR, not reached; OS, overall survival; PD, progressive disease; PD-L1, programmed cell death ligand 1; PFS, progression-free survival; PR, partial response; RECIST, Response Evaluation Criteria In Solid Tumors; SD, stable disease; TMB, tumour mutational burden. Table 1Baseline characteristics of the study participants in the DIA-MS cohortCharacteristicNo. (%)Total(N=73)DCB group(n=41)NDB group(n=32)P value*Age, years Median (IQR)63.0(58.0–68.5)60.0(57.5–68.0)64.5(58.5–69.8)0.14 ≥ 6045 (61.6)21 (51.2)24 (75.0)0.053 < 6028 (38.4)20 (48.8)8 (25.0)Female sex29 (39.7)14 (34.1)15 (46.9)0.34ECOG PS 065 (89.0)37 (90.2)28 (87.5)0.72 18 (11.0)4 (9.7)4 (12.5)Child-Pugh grade A71 (97.3)41 (100)30 (93.8)0.19 B2 (2.7)02 (6.2)Extent of disease Locally advanced†17 (23.3)8 (19.5)9 (28.1)0.42 Metastatic56 (76.7)33 (80.5)23 (71.9)Primary tumour site GBC21 (28.8)13 (31.7)8 (25.0)0.75 ICC42 (57.5)22 (53.4)20 (62.5) ECC10 (13.7)6 (14.6)4 (12.5)Hepatitis Yes10 (13.7)4 (9.8)6 (18.8)0.32 No63 (86.3)37 (90.2)26 (81.2)Baseline CA19-9 ≥37 U/mL47 (64.4)24 (58.5)23 (71.9)0.33 <37 U/mL26 (35.6)17 (41.5)9 (28.1)PD-L1‡0.68 Positive8 (11.0)6 (14.6)2 (6.3) Negative19 (26.0)12 (29.3)7 (21.9) Missing§46 (63.0)23 (56.1)23 (71.8)MSI0.67 High1 (1.4)1 (2.4)0 Stability45 (61.6)25 (61.0)20 (62.5) Missing§27 (37.0)15 (36.6)12 (37.5)TMB0.12 High5 (6.8)4 (9.8)1 (3.1) Middle218 (24.7)13 (31.7)5 (15.6) Low23 (31.5)9 (22.0)14 (43.8) Missing§27 (37.0)15 (36.5)12 (37.5)Regimen IM + Len35 (47.9)17 (41.5)18 (56.3)0.24 IM + Len + CHEMO38 (52.1)24 (58.5)14 (43.7)Response<0.001 Partial response25 (34.2)24 (58.5)1 (3.1) Stable disease31 (42.5)16 (39.0)15 (46.9) Progressive disease17 (23.3)1 (2.5)16 (50.0)PFS, months, median, (95% CI)6.1(4.6 to 7.6)11.5(5.3 to 17.7)3.0(2.0 to 4.0)<0.001OS, months, median, (95% CI)17.0(12.2 to 21.8)NR8.2(4.6 to 11.8)<0.001Percentages may not total 100 because of rounding.For non-normally distributed numeric variables (median (IQR)), two-sided Mann-Whitney U tests were applied. Χ² tests were used for categorical variables (n (%).*P values are obtained with non-norm, t, Wilcoxon or Fisher exact test.†Patient has only locally advanced sites of disease.‡PD-L1 positivity was defined as a tumour proportion score (TPS) ≥1%.§PD-L1/MSI/TMB status missing includes PD-L1/MSI/TMB-unknown and not tested.CA19-9, carbohydrate antigen 19–9; CHEMO, chemotherapy; DCB, durable clinical benefit; DIA-MS, data-independent acquisition-mass spectrometry; ECC, extrahepatic cholangiocarcinoma; GBC, gallbladder cancer; ICC, intrahepatic cholangiocarcinoma; IM, immunotherapy; Len, lenvatinib; MSI, microsatellite instability; NDB, non-durable clinical benefit; NR, not reach; OS, overall survival; PD-L1, programmed cell death ligand 1; PFS, progression-free survival; ECOG PS, Eastern Cooperative Oncology Group Performance Status; TMB, tumour mutational burden. Baseline characteristics of the study participants in the DIA-MS cohort Percentages may not total 100 because of rounding. For non-normally distributed numeric variables (median (IQR)), two-sided Mann-Whitney U tests were applied. Χ² tests were used for categorical variables (n (%). P values are obtained with non-norm, t, Wilcoxon or Fisher exact test. Patient has only locally advanced sites of disease. PD-L1 positivity was defined as a tumour proportion score (TPS) ≥1%. PD-L1/MSI/TMB status missing includes PD-L1/MSI/TMB-unknown and not tested. CA19-9, carbohydrate antigen 19–9; CHEMO, chemotherapy; DCB, durable clinical benefit; DIA-MS, data-independent acquisition-mass spectrometry; ECC, extrahepatic cholangiocarcinoma; GBC, gallbladder cancer; ICC, intrahepatic cholangiocarcinoma; IM, immunotherapy; Len, lenvatinib; MSI, microsatellite instability; NDB, non-durable clinical benefit; NR, not reach; OS, overall survival; PD-L1, programmed cell death ligand 1; PFS, progression-free survival; ECOG PS, Eastern Cooperative Oncology Group Performance Status; TMB, tumour mutational burden. As of the data cut-off of 23 June 2025, the median follow-up time was 20.3 (range 2.0–26.3) months. All patients received ICIs plus lenvatinib, and 38 received chemotherapy (table 1, online supplemental table S6). Durable clinical benefit (DCB) is defined as a complete response, partial response (PR) or stable disease lasting ≥6 months. Accordingly, 41 (56.2%) exhibited DCB, while 32 (43.8%) showed non-durable clinical benefit (NDB). The DCB group had significantly longer survival (figure 2C), with no significant differences in routine and metabolic-related indicators between the two groups (table 1, online supplemental table S7). Baseline clinical variables, except for age, were unrelated to immunotherapy response (figure 2D). Given the limited predictive power of existing clinical indicators, we performed a staged MS-based discovery-validation proteomics workflow to screen urinary proteomic biomarkers for predicting immunotherapy outcomes (figure 1). For the discovery cohort, 64 patients with pretreatment urine samples were randomly divided into the training and validation sets in a 2:1 ratio (online supplemental table S7). 24 individuals (test cohort) were included between 25 February 2024 and 25 September 2024 for external validation. Detailed clinical characteristics are provided in online supplemental table S1. Characteristics of urinary proteomics and single-cell transcriptomic data We performed proteome profiling of urine samples using a high-throughput DIA-MS-based proteomics strategy. In urine, 5965 proteins were identified using a 1% false discovery rate, with an average of 3385 proteins detected per sample (online supplemental figure S2A). After excluding proteins with over 50% missing values in all samples and quality control (QC)-MS samples, 3185 proteins were quantified for further analysis. The proteomic depth of the urine specimens exceeded that of most published studies (online supplemental table S9). The Pearson correlation coefficient among the QC-MS and QC-experiment samples was 0.90 and 0.88, respectively, indicating the MS platform’s stability and repeatability of sample preparation (online supplemental figure S2B,C). Nine samples were excluded after QC from further analysis (online supplemental figure S2D–G). Annotation using the Human Protein Atlas database revealed that most urinary proteins originated from intracellular sources (70.4%). These proteins were primarily specific to the brain and liver (online supplemental figure S2H,I), with liver-derived intracellular proteins enriched in the urine, consistent with previous reports.15 For the scRNA-seq data, 60 909 cells were retained after stringent filtering and assigned to 15 major clusters based on canonical marker gene expression (online supplemental figure S3A–F). The cell type-to-cell type correlation matrix illustrates the distinct characteristics of different cell subsets and their interconnections (online supplemental figure S3G). Immune cells, constituting a large proportion of the samples, were further clustered, with a focus on T cells (online supplemental figure S4A,C). CD4+ T cells were categorised into naive T (CD4+ Tn) and regulatory T (CD4+ Treg) cells. CD8+ T cells were reclustered based on the expression of naïve (CD8+ Tn), effector memory (CD8+ Tem), type 17 CD8+ T like (CD8+ Tc17), stress-related (CD8+ Tstr), exhausted (CD8+ Tex) and proliferative (CD8+ Tprolif) signatures (online supplemental figure S4B).19 Baseline urine proteomic patterns correlate with DCB from ICIs We initially analysed pretreatment urinary proteomics to evaluate whether specific proteomic features correlate with benefits from ICIs. Specifically, a comparative analysis between the DCB and NDB groups identified 133 differentially expressed proteins (DEPs) in the training cohort, with distinct inter-group expression patterns (figure 3A,B). These DEPs were predominantly intracellular, with a significant proportion annotated as liver-specific (figure 3C), highlighting unique potential as biomarkers for hepatobiliary tumours. Gene Ontology (GO) analysis highlighted immune activation processes in the DCB group (figure 3D,F). Conversely, the NDB group exhibited enrichment in protumourigenic pathways (figure 3D,F). Gene set enrichment analysis (GSEA) further corroborated these findings, showing upregulated lymphocyte-mediated immunity in the DCB group, whereas oncogenic pathways dominated in the NDB group (figure 3E). We also observed seven DEPs (SUB1, protein tyrosine phosphatase non-receptor 13 (PTPN13), IL13RA1, C5, TDRD15, ADGRG2 and YIPF3) significantly correlated with response states and baseline Carbohydrate antigen 19–9 (CA19-9) levels (figure 3G). Notably, CA19-9 is a validated diagnostic and prognostic biomarker for BTC, suggesting that these DEPs are associated with the biological characteristics of BTC.20 Additionally, 133 DEPs clustered into two groups corresponding to the DCB and NDB groups, revealing intrinsic correlations among the DEPs (online supplemental figure S5). Characteristics of differentially expressed urinary proteins between the DCB and NDB groups at baseline. (A) The heatmap displaying the 133 baseline DEPs between DCB and NDB groups in the training cohort. (B) Volcano plot of the DEPs when comparing pretreatment urine samples from DCB versus the NDB group. (C) Number of tissue-specific proteins identified in urine and percentages of identified urine proteins which are intracellular, secreted or membrane-bound. The protein annotations and tissue specificity information are retrieved from the Human Protein Atlas database (the Tissue Atlas). (D) GO analysis of upregulated baseline proteins in the DCB and NDB groups. (E) GSEA comparisons between the DCB and NDB groups. Blue bars represent pathways upregulated in the DCB group, while red bars represent pathways upregulated in the NDB group. (F) Schematic diagram of pathways related to immunotherapy response. Created with BioRender.com. (G) The heatmap showing the Spearman correlation between the 133 DEPs and clinical characteristics. DCB, durable clinical benefit; DEPs, differentially expressed proteins; ECM, extracellular matrix; FGFR, fibroblast growth factor receptor; GO, Gene Ontology; GSEA, gene set enrichment analysis; MAPK, mitogen-activated protein kinase; MHC, major histocompatibility complex; NDB, non-durable clinical benefit; NES, normalised enrichment score; PI3K, phosphatidylinositol 3-kinase; TNF, tumour necrosis factor. Collectively, these results demonstrate a strong association between baseline immune-related pathways and urinary proteomic features with treatment sensitivity. Urinary proteomics dynamics reflect antitumour immunity changes under ICI treatment Building on our prior finding that baseline urinary proteomics is linked to immunotherapy response and tumour microenvironment (TME),13 we hypothesised that longitudinal urine proteomic changes reflect systemic and TME adaptations during ICI treatment. We systematically investigated: (1) early proteomic shifts post-ICI initiation, (2) longitudinal dynamics during multicourse therapy and (3) resistance-associated signatures. Early ICI treatment induces response-specific proteomic remodelling To determine whether early ICI treatment induces characteristic proteomic changes, regardless of treatment response, we compared paired urinary proteomic profiles from pretreatment and post-first cycle across the cohort (n=38). Dramatic alterations in proteome composition were identified, with 181 DEPs detected (p<0.05), with liver-specific and intestine-specific proteins predominating (online supplemental figure S6A). GO analysis demonstrated enrichment of these DEPs in innate immune pathways (figure 4A), highlighting the impact of ICIs on the immune system. Urinary proteome alteration patterns under ICI therapeutic stress. (A) Radar plot showing differential pathway enrichment before and after treatment (T1) in the DCB and NDB groups, as well as in all patients (regardless of treatment response). (B) GO analysis of changed proteins after a dose of ICI in the DCB and NDB groups. (C) Downregulated proteins after treatment in the DCB group mapping to our scRNA-seq data. (D) Cluster analysis of proteins in the PR group at different timepoints (T0, T1, T2 and T3) using Mfuzz identified a continuous downregulation of proteins in Cluster 1 and a continuous upregulation of proteins in Cluster 5. (E) The comparison of KEGG pathway enrichment of Cluster 1 and Cluster 5 in the PR group. (F) Proteins in Cluster 1 and Cluster 5 in the PR group mapping to our scRNA-seq data. (G) Cluster analysis of proteins in the PD group at different time points (T0, T1, T2 and T3) using Mfuzz identified a continuous upregulation of proteins in Cluster 1 and a continuous downregulation of proteins in Cluster 4. (H) The comparison of KEGG pathway enrichment of Cluster 1 and Cluster 4 in the PD group. (I) Proteins in Cluster 1 and Cluster 4 in the PD group mapping to our scRNA-seq data. DC, dendritic cell; DCB, durable clinical benefit; FGFR, fibroblast growth factor receptor; GO, Gene Ontology; ICI, immune checkpoint inhibitor; IFN, interferon; KEGG, Kyoto Encyclopedia of Genes and Genomes; MAPK, mitogen-activated protein kinase; MHC, major histocompatibility complex; NDB, non-durable clinical benefit; NK, natural killer; PD, progressive disease; PR, partial response; scRNA-seq, single-cell RNA sequencing; TNF, tumour necrosis factor. Next, we compared proteomic changes between the DCB and NDB groups. GO enrichment analysis revealed divergent pathways: DCB was associated with NF-κB and TNF signalling, antigen presentation and T cell-related pathways, while NDB showed enrichment in angiogenesis and TGF-β/SMAD/BMP signalling (figure 4B). Notably, the hyperactive TGF-β signalling pathway, which promotes T cell exclusion, has been linked to ICI resistance.21 These results suggest that urinary proteomics reflects distinct changes in antitumour immunity following early ICI treatment, potentially serving as an early predictor of therapeutic response. The TME is a key determinant of antitumour immunity. While urinary proteomics has been linked to TME in renal cancer,13 evidence in BTC is lacking. To address this gap, we mapped ICI-induced proteomic changes to scRNA-seq of tumour tissues, assessing their association with antitumour immunity. In the DCB group, most downregulated proteins post-treatment were enriched in epithelial cells, likely reflecting reduced tumour burden and TME remodelling (figure 4C). These findings implicate immune regulatory signalling in ICI responsiveness, whereas TGF-β and angiogenesis pathways correlate with resistance—a pattern consistent with pan-cancer studies.21 22 Overall, these findings suggest that urinary proteomics can non-invasively capture dynamic changes in antitumour immunity and TME remodelling following early ICI treatment, positioning it as a biomarker for real-time treatment monitoring. Urinary proteomics dynamics reveal response-specific trajectories during multiple consecutive courses of ICIs Having characterised proteomic fluctuations after a single ICI infusion, we then explored whether urinary proteomics dynamics could track responses throughout continuous ICI treatment. Capturing extreme responses provides insights into biomarkers and mechanisms across a broader patient population.23 We focused on patients with extraordinary treatment responses, classified as PR or progressive disease (PD) at 3 months post-treatment, and examined urine proteome dynamics at baseline, before the second and third treatment cycles and 2–3 months post-treatment (online supplemental figure S6B). Correspondingly, we analysed 77 urine samples from 30 patients. In the PR group, clustering analysis of 244 altered proteins across the four time points using the Mfuzz method identified six distinct clusters (figure 4D, online supplemental figure S6C). Given the overall linear increase or decrease in Cluster 1 and Cluster 5, the two clusters were further analysed. Kyoto Encyclopedia of Genes and Genomes functional analysis indicated that Cluster 1 was primarily involved in metabolic processes, including one-carbon unit metabolism, vitamin B6 metabolism and mitogen-activated protein kinase (MAPK) cascade, all downregulated during the response (figure 4E). Proteins in Cluster 5 were strongly associated with B cell-related pathways and endocytosis, possibly reflecting the pathogenesis of tumour shrinkage (figure 4E). Moreover, scRNA-seq data linked Cluster 1 proteins to stromal cells and Cluster 5 proteins to immune cells (figure 4F), suggesting tumour burden reduction and immunity activation during treatment. Similarly, the 311 time-varying proteins in the PD group were classified into six clusters (figure 4G, online supplemental figure S6D). Interestingly, proteins in Cluster 1 of the PD group exhibited a gradual increase over time, with enriched pathways including complement, MAPK cascade and angiogenesis, resembling those in Cluster 1 of the PR group (figure 4H). Conversely, proteins in Cluster 4 of the PD group demonstrated a gradual decrease in expression, with pathways primarily associated with immune responses, similar to those in Cluster 5 of the PR group (figure 4H). In the PD group, proteins upregulated in Cluster 1 were primarily enriched in stromal cells, while those downregulated in Cluster 4 were mainly derived from immune cells, aligning with tumour progression (figure 4I). These opposing proteomic trajectories between PR and PD groups highlight conserved mechanisms underlying treatment efficacy, establishing an association between proteome patterns and therapy response. Resistance-associated proteomic signatures overlap with protumourigenic pathways Next, we aimed to determine whether the urinary proteomics reflects resistance-associated changes. We performed paired Mann-Whitney U tests to compare protein levels (pretreatment vs pre-PD; pre-PD vs on-PD) and 107 and 111 DEPs were identified, respectively. Enrichment analysis showed that DEPs between the on-PD and pre-PD stages were mainly enriched in pathways related to TGF-β, MAPK and ERBB2 signalling, complement activation and neutrophil extracellular traps (online supplemental figure S6E), consistent with prior studies.21 24 25 Similar changes were observed during progression compared with the baseline (online supplemental figure S6F). Collectively, these findings demonstrate that the urinary proteomics dynamically mirrors systemic and TME adaptations during ICI therapy, highlighting its utility as a non-invasive biomarker for predicting and monitoring immunotherapy response. Machine learning-based model for predicting durable response to immunotherapy before treatment Based on the aforementioned findings, we hypothesised that the urinary proteomics could effectively distinguish individuals likely to benefit from immunotherapy. To test this, we developed a machine learning model based on the 133 DEPs identified between DCB and NDB groups in the training set, as illustrated in figure 5A and online supplemental figure S7. We employed three feature selection methodologies (filtering, univariate logistic regression and AdaBoost algorithm) and identified 18 overlapping proteins. A logistic regression model was then trained to optimise biomarker combinations, ultimately yielding a 4-urinary protein (4-UP) prediction panel (PTPN13, SUB1, MICAL-L1, VARS1) for predicting DCB. This panel demonstrated superior accuracy compared with single-protein biomarkers (figure 5B, online supplemental table S10). Performance of the model was independently validated in the validation set, achieving areas under the curve (AUCs) of 0.950 (training) and 0.867 (validation) (figure 5C). The construction and validation of the 4-UP predictive model for immunotherapy response. (A) Diagram describing the construction and validation of the predictive model for DCB and NDB groups. (B) Heatmap showing the prediction of durable benefit from ICIs by each of the four proteins (P26640, P53999, Q8N3F8 and Q12923) and the 4-UP model in the training and validation sets with the criterion of the maximum of the Youden index J in the training set. Red represents the prediction of DCB, and white represents prediction of NDB. (C) ROC curves showing the predictive performance of this model in the training and validation sets, as well as in chemo-free and chemo-containing ICI-based combinations, ICC and GBC subtypes, and CA19-9-negative and positive patients across the discovery cohort based on DIA data. (D) ROC curves showing the predictive performance of this model in the training, validation sets and test cohort based on PRM data. (E) The accuracy of the 4-UP predictive model to predict early therapeutic response under different classification schemes based on DIA data. Classifications: (1) tumour enlargement (PD and SD+ (stable disease with diameters of target lesions increase <20%)) versus tumour shrinkage (PR and SD− (stable disease with diameters of target lesions decrease <30%)); (2) non-PD (PR and SD) versus PD; (3) PR versus PD. AUC, area under the curve; CA19-9, carbohydrate antigen 19–9; CFS, candidate feature selection; FMC, final model construction; DCB, durable clinical benefit; DEPs, differentially expressed proteins; DIA, data-independent acquisition; ECC, extrahepatic cholangiocarcinoma; GBC, gallbladder cancer; ICC, intrahepatic cholangiocarcinoma; ICI, immune checkpoint inhibitor; NDB, non-durable clinical benefit; PD, progressive disease; PR, partial response; PRM, parallel reaction monitoring; ROC, receiver operating characteristic; SD, stable disease; 4-UP, 4-urinary protein. To address potential overfitting, we performed internal validation using 1000 bootstrap resamples in the training set.26 The optimism-corrected AUC remained stable at 0.914, and the calibration curve showed strong agreement between predicted and observed outcomes with a Brier score of 0.086 (online supplemental figure S9A). The Hosmer-Lemeshow test yielded a non-significant p value of 0.79, supporting good calibration. Subgroup analyses further validated stable performance among clinically relevant strata (BTC subtypes and regimens) across the discovery cohort (both training and validation sets) (figure 5C). Notably, imaging-based assessment of BTC therapeutic efficacy is often challenging, and CA19-9 levels, though commonly used adjuncts, are unreliable in CA19-9-negative patients. To address this, we validated the panel in CA19-9-negative and CA19-9-positive cohorts. Strikingly, the panel consistently discriminated DCB from NDB patients across all subgroups, confirming its stability and generalisability under diverse clinical conditions. Next, to verify the authenticity of the proteins in the 4-UP panel, we employed a widely used targeted internal standard triggered-PRM14 16 18 to quantitatively verify the four proteins in both the discovery (n=53) and test (n=24) cohorts, which confirmed their differential expression between the DCB and NDB groups (online supplemental figure S8A–G). Validation in these independent PRM-MS discovery and test cohorts confirmed the generalisability of the model, yielding AUCs of 0.821 (training), 0.750 (validation) and 0.805 (test) with an optimism-corrected AUC of 0.756 and good calibration (Brier score=0.172, Hosmer-Lemeshow p=0.07) (figure 5D, online supplemental figure S9B). Subgroup analyses further validated stable performance among clinically relevant strata across both the PRM-MS discovery and test cohorts (online supplemental figure S8H). Collectively, this optimised urinary proteomic panel demonstrates robust performance across multiple validation frameworks and clinical subgroups, underlining its broad applicability potential in clinical practice. The 4-UP panel accurately predicts the early response to immunotherapy Over 20% of patients with BTC experience early disease progression,2 3 and early response evaluation could serve as an alternative approach for predicting the ultimate clinical benefit of ICIs.4 Thus, we assessed the 4-UP panel’s ability to predict responses 1.5–3 months post-treatment by stratifying patients into three clinically actionable groups: tumour enlargement versus tumour shrinkage, PD versus Non-PD and PR versus PD. The panel demonstrated acceptable predictive performance (AUC=0.697–0.878) in both DIA-MS and PRM-MS cohorts (figure 5E, online supplemental file 1). Taken together, these findings indicate that the 4-UP panel is a reliable tool for assessing early immunotherapy outcomes, enabling timely clinical interventions in diverse BTC populations. PTPN13+ epithelial cells are associated with responses to immunotherapy Building on our prior demonstration that baseline urinary proteomics predicts ICI responsiveness, we hypothesised that the proteins in the 4-UP model may provide deeper insights into the TME. To test this, we mapped the four proteins to scRNA-seq data of BTC TME, comparing their expression between DCB and NDB groups. Among the proteins analysed, PTPN13 stood out due to its significantly higher expression among epithelial cells of the DCB group (figure 6A,B). Quantification revealed a trend toward higher proportions of PTPN13+ epithelial cells in DCB patients (p=0.052; figure 6C), suggesting that PTPN13+ epithelial cells may confer sustained immunotherapy benefit. Integrative single-cell transcriptomic and functional validation of PTPN13 in BTC. (A) Heatmap displaying the expression levels of proteins in the 4-UP model across cell subsets in the DCB and NDB groups. (B) UMAP plot showing the expression of PTPN13 in the DCB and NDB groups. (C) Box plots comparing the proportions of PTPN13+ epithelial cells among all epithelial cells between the DCB and NDB groups (p=0.052, Wilcoxon rank-sum test). (D) Volcano plot showing differentially expressed genes between PTPN13+ and PTPN13− epithelial cells. (E) Two-sided bar plots exhibiting significantly enriched GO pathways between PTPN13+ and PTPN13− epithelial cells identified by GSEA. (F) Scatterplots showing the correlation between the proportion of PTPN13+ epithelial cells in the TME and serum CA19-9 levels, and between urine PTPN13 abundance and serum CA19-9 levels. Spearman’s correlation coefficient r and p value are displayed. (G) Kaplan-Meier estimates of OS stratified by the median PTPN13 mRNA expression, using data from GSE89747 (n=115). (H) Spatial feature plots showing the expression of ALB, KRT19 and PTPN13 in an ICC leading-edge sample. (I) Representative mIF images of tumour tissues immunostained with panCK (green) and PTPN13 (red) in NDB and DCB groups. (J) mRNA expression levels of PTPN13 in HUCCT1, QBC939 and RBE cell lines. (K) PTPN13 mRNA expression in RBE cells transfected with sh-NC (negative control) or sh-PTPN13 (sh1, sh2, sh3). (L) Representative subcutaneous tumours formed by stably transfected sh-NC or sh-PTPN13 RBE cell lines in nude mice at terminal endpoint, along with tumour growth curves and tumour volumes of each group. (M) The expression level of Ki67 in nude mice-derived xenograft tumour was determined by IHC, scale bar, 20 µm. (N) Representative images of TUNEL staining performed on paraffin-embedded sections of subcutaneous xenograft tumours harvested from nude mice. scale bar, 500 and 20 µm. Data are expressed as mean±SD, p>0.05, **p<0.05, ***p<0.001, ****p<0.0001 (unpaired t-test or ANOVA). ANOVA, analysis of variance; BTC, biliary tract cancer; CA19-9, carbohydrate antigen 19-9; DAPI, 4,6-diamidino-2-phenylindole; DC, dendritic cell; DCB, durable clinical benefit; GO, Gene Ontology; GSEA, gene set enrichment analysis; ICC, intrahepatic cholangiocarcinoma; IHC, immunohistochemistry; mIF, multiplex immunofluorescence; mRNA, messenger RNA; NDB, no durable benefit; NES, normalised enrichment score; NK, natural killer; ns, not significant; OS, overall survival; PTPN13, protein tyrosine phosphatase non-receptor 13; TUNEL, TdT-mediated dUTP Nick-End Labelling; UMAP, Uniform Manifold Approximation and Projection; 4-UP, 4-urinary protein. We then focused on the functional role of PTPN13+ epithelial cells and their relationship with tumour biology and immunotherapy response. Our BTC scRNA-seq data revealed that epithelial cells exhibited notable chromosomal amplifications and deletions, characteristic of malignant cells (online supplemental figure S10A). Specifically, PTPN13+ epithelial cells had slightly higher copy number variation (CNV) scores compared with PTPN13− cells, and both significantly exceeded hepatocyte CNV levels (online supplemental figure S10B). Moreover, PTPN13+ epithelial cells exhibited upregulated hepatocyte marker ALB levels (figure 6D), suggesting hepatocyte-like differentiation and reduced stemness.27 GSEA analysis highlighted enrichment of hepatocyte-specific functions, epithelial homeostasis and pro-apoptotic signatures in PTPN13+ cells, whereas PTPN13− cells exhibited antiapoptotic and interleukin (IL)-4/IL-10-mediated immune suppression (figure 6E). Additionally, hallmark pathway analysis revealed that PTPN13− epithelial cells were associated with the epithelial–mesenchymal transition (EMT) and hypoxia (online supplemental figure S10C), indicating a more aggressive phenotype. Single-cell regulatory network inference and clustering (SCENIC) analysis revealed differential transcription factor (TF) expression between PTPN13+ and PTPN13− epithelial cells. TFs enriched in PTPN13+ cells, such as HNF1B, ONECUT1, NFIC, ATF3 and KLF6, are known to support epithelial phenotype maintenance, prevent EMT and promote apoptosis (online supplemental figure S11A).2832 Conversely, TFs such as HCFC1, ETS1, BHLHE40, ZNF384, ETS2 and ELK3 were enriched in PTPN13− epithelial cells, which are associated with cell proliferation, tumour invasiveness and EMT promotion.3336 Notably, PTPN13 is a target gene of both ATF3 and ONECUT1, potentially contributing to the apoptosis-promoting effects seen in PTPN13+cells (online supplemental figure S11B). Given the importance of intercellular communication within the TME, we examined interactions between immune cells and PTPN13+/PTPN13− epithelial cells. We identified close interactions between immune cells and PTPN13+ epithelial cells via three classical proapoptotic pathways, including FAS (FASLG−FAS), TNF (TNF−TNFRSF1A and TNF−TNFRSF1B) and TRAIL (TNFSF10−TNFRSF10A) (online supplemental figure S11C,D). In contrast, PTPN13− cells interact with immune cells through ligand–receptor pairs like CD1D/CD8A−CEACAM5 and ADGRE5−CD55, which are linked to cancer cell migration and invasion.37 38 Additionally, PTPN13 expression inversely correlated with CA19-9 in tumours and urine (figure 6F). Analysis of the GSE89747 RNA-seq data set of cholangiocarcinoma from Jusakul et al39 revealed that high PTPN13 expression correlated with better prognosis (figure 6G). Spatial analysis of four intrahepatic cholangiocarcinoma specimens from Wu et al’s40 and Li et al’s41 studies showed that PTPN13 was predominantly expressed in tumour regions (KRT19) and minimally in normal regions (ALB) (figure 6H, online supplemental figure S12). Multiplex immunofluorescence confirmed its malignant cell specificity and enrichment in the DCB group (figure 6I). Based on the critical role of PTPN13 revealed in the aforementioned results, we next sought to investigate its function in BTC. Among three human BTC cell lines (RBE, HuCCT1, QBC939), we selected the RBE cell line for further study due to its highest PTPN13 expression (figure 6J). PTPN13 knockdown significantly promoted cell proliferation both in vitro and in vivo, along with reduced apoptosis of tumour cells (figure 6K–N, online supplemental figure S13). In summary, our analysis and in vitro/in vivo experiments demonstrate that PTPN13 in epithelial cells is associated with a proapoptotic phenotype, enhanced immunotherapy responsiveness and improved survival rates. These findings support its potential as both a predictive biomarker and a therapeutic target for future immunotherapy strategies. Discussion In this study, we systematically investigated the urinary proteomic features associated with ICI responsiveness in BTC and integrated these findings with single-cell and spatial transcriptomic data. We demonstrated that antitumour immunity, both systemically and within the TME, can be non-invasively monitored throughout ICI therapy. Importantly, we identified mechanisms underlying both response and resistance and developed a 4-UP panel through a multilayered validation strategy combining DIA-based discovery with PRM-based targeted proteomics. It robustly predicted DCB and early response. To our knowledge, this study represents the pioneering investigation of urinary proteomics in tumour immunotherapy, offering a novel non-invasive approach to dynamically monitor TME remodelling and predict therapeutic outcomes. Among the four proteins in the panel, PTPN13 was highlighted for its mechanistic significance. In this study, PTPN13+ malignant cells exhibited a proapoptotic phenotype, which was linked to improved ICI responsiveness and survival. PTPN13 acts as a regulator of key signalling pathways by dephosphorylating intracellular receptors and signalling molecules, thereby influencing proliferation, migration and apoptosis.42 In solid tumours, PTPN13 functions as a tumour suppressor with reduced expression correlating with enhanced invasiveness and poor prognosis.4347 The other three proteins in the 4-UP model also contribute importantly to BTC pathophysiology: MICAL-L1 is implicated in cytoskeleton remodelling and immune modulation,48 49 VARS1 affects immune-related pathways and may also serve as a predictor of response to anti-PD-1 immunotherapy,50 while SUB1 regulates transcription processes influencing tumour progression and therapeutic resistance.51 52 These proteins validate the biological relevance of the 4-UP panel, reflecting both tumour-related and immune-related processes, providing a mechanistic foundation for its predictive capacity in BTC. Further studies are necessary to fully elucidate the role of these proteins in the context of immunotherapy in BTC. Furthermore, we demonstrated other proteins such as C3, C5, Siglec-3, Dectin-1 and SULT2B1 may contribute to ICI resistance through immune suppression and metabolic reprogramming.255355 Notably, our analyses also highlight the role of metabolic pathways, particularly cholesterol and fatty acid metabolism, in shaping the TME and modulating treatment response. Pathway analyses showed enrichment of these pathways among proteins altered post-treatment in the DCB group. Cholesterol metabolism modulates programmed cell death ligand 1 (PD-L1) stability, T cell activation and macrophage polarisation, while fatty acid metabolism affects T cell function and immune suppression, emphasising the role of lipid metabolism in antitumour immunity.56 57 Together, these findings offer avenues for future functional studies and biomarker exploration. Pretreatment tissue-based predictive biomarkers, such as PD-L1, tumour mutational burden and microsatellite instability-high, may not capture the plasticity of systemic and local antitumour immunity during ICI therapy and therefore may not accurately reflect clinical outcomes.58 However, biofluids that reflect overall pathophysiological status changes may serve as valuable sources, and our study also suggested the urinary proteomic model outperformed these biomarkers in predicting ICI response (online supplemental figure S14 and table S11). Biofluid-based biomarkers such as circulating tumour DNA (ctDNA) show promise, but their clinical utility is constrained by low concentrations and technical challenges.8 13 Conversely, proteins in biofluids are abundant, stable and more amenable to development as clinically applicable biomarkers.13 14 Urinary proteomics reflects the genetic background, overcomes spatial heterogeneity and sensitively captures dynamic physiological changes, and its utility in mechanistic investigations, target screening and treatment prediction is well-established across various diseases.10 11 15 However, its application in tumour immunotherapy remains limited, with only a few ongoing trials (NCT04712305, NCT05928299, NCT04641936). Additionally, previous studies have shown that urinary proteomics can reflect the static TME, and here we demonstrate that it can also track dynamic TME changes during immunotherapy.10 11 13 We identified four clusters of urinary proteins whose levels changed linearly with tumour progression or regression. These proteins were associated with tumour burden and immune cell dynamics, revealing distinct proteomic trajectories: responders exhibited patterns consistent with immune activation and reduced tumour burden, while PD patients exhibited protumourigenic signalling. These findings support urine as a surrogate ‘window’ into systemic and local antitumour dynamics. Further studies with serial sampling are needed to provide deeper insights into urinary proteomic features correlated with specific biological processes during immunotherapy. This study has some limitations. (1) Despite employing a multilayered validation strategy, the single-centre design, limited sample size and heterogeneity across tumour subtypes and regimens limit the stability and generalisability of biomarker discovery. Multicentre studies comprising diverse populations (varying ethnic backgrounds, international standard regimens, etc) are required for further validation. (2) More extensive functional studies are ongoing to elucidate the mechanisms underlying these candidate molecules in the context of ICI therapy. (3) To optimise the panel for enhanced discriminatory power and reduced redundancy, several clinically significant proteins and pathways remain underexplored and will be explored in future research. In conclusion, we established urinary proteomics as a novel, non-invasive strategy for predicting and monitoring ICI responsiveness in BTC. By capturing both systemic and local TME dynamics, urinary proteomics offers unique insights into treatment response, resistance mechanisms, biomarker discovery and can serve as a potential tool for individual treatment decisions. Supplementary material