Serum proteomic profiling identifies pathways and protein signatures associated with early versus late recurrence in cholangiocarcinoma

Abstract

Cholangiocarcinoma (CCA) is characterized by frequent postoperative recurrence, yet reliable biomarkers for distinguishing early from late recurrence remain limited. This study aimed to identify serum protein biomarkers associated with early versus late recurrence and develop a proteomic recurrence score (PRS) to distinguish these recurrence groups after curative resection. Serum proteomic profiling was performed using LC-MS/MS in patients with resected CCA. Candidate proteins were identified through integrated univariate and multivariate analyses, followed by pathway enrichment and protein–protein interaction analyses. Proteins with ≥50% detection frequency were refined for model development, and elastic net – based feature selection within a repeated nested cross-validation framework was used to derive the proteins for construction of the PRS. A total of 232 candidate proteins associated with recurrence were identified, highlighting enrichment of immune/inflammatory signaling, metabolic reprogramming, and tumor progression pathways. Refinement yielded 33 consistently detectable proteins. During nested model-size optimization, the 20-protein candidate model achieved a patient-level OOF ROC-AUC of 0.937 (95% CI, 0.881–0.979). Based on feature-selection stability, the 20 most stable proteins were subsequently prioritized as a fixed exploratory signature for PRS construction, achieving a patient-level OOF ROC-AUC of 0.948 (95% CI, 0.895–0.988). A PRS derived from an exploratory 20-protein signature was significantly associated with clinicopathological variables and independently associated overall survival in multivariable Cox regression. Comprehensive serum proteomic profiling identified biological pathways associated with recurrence timing and enabled the prioritization of an exploratory 20-protein PRS for distinguishing early from late recurrence among patients with CCA. The PRS demonstrated promising internal discrimination and was independently associated with overall survival. External validation in independent cohorts is required before clinical application.

Cholangiocarcinoma (CCA) is characterized by frequent postoperative recurrence, yet reliable biomarkers for distinguishing early from late recurrence remain limited.

1. Introduction

Cholangiocarcinoma (CCA) is a highly aggressive malignancy of the biliary tract that is frequently diagnosed at advanced stages because of its asymptomatic early course.^1–3 Despite curative-intent surgery combined with adjuvant chemotherapy, long-term outcomes remain poor,^3,4 with postoperative recurrence occurring in approximately 60–70% of patients.^5,6 Recurrence is commonly classified as early or late according to the time of onset, with early recurrence (<12 months) associated with aggressive tumor biology and poor survival outcomes, whereas late recurrence is more closely linked to indolent tumor behavior and host-related factors, particularly immune responses.^6–9 Current clinical biomarkers, including carcinoembryonic antigen (CEA) and carbohydrate antigen 19-9 (CA19-9), are routinely used for diagnosis and disease monitoring; however, their clinical utility remains limited by low specificity, inter-individual variability, and insufficient sensitivity for early-stage disease. Moreover, CA19-9 can be elevated in benign hepatobiliary disorders and is undetectable in individuals with Lewis's antigen-negative status,^10,11 while CEA lacks cancer specificity.¹² Therefore, more reliable molecular biomarkers are urgently needed to improve recurrence risk stratification and guide postoperative management in patients with CCA.

Increasing evidence indicates that cancer recurrence is not solely determined by tumor-intrinsic properties but is also profoundly influenced by interactions within the tumor microenvironment (TME).^13–15 The TME contributes to recurrence through several interconnected mechanisms, including immune escape, persistence of residual or dormant tumor cells, inflammatory and chemokine-driven reactivation, and stromal or extracellular matrix remodeling, collectively creating a permissive niche that supports tumor survival, adaptation, and disease re-initiation.^14–16 Importantly, the balance between antitumor immune surveillance and immunosuppressive signaling critically determines whether residual tumor cells are eliminated or persist following resection.^16,17 In this context, molecular alterations at both the protein and metabolite levels may provide important insights into tumor–host interactions associated with recurrence. Proteomic profiling can identify dysregulated signaling pathways and microenvironmental components, whereas metabolomic analysis captures metabolic reprogramming that supports tumor adaptation and progression. Integrating these complementary approaches may therefore improve biological understanding and biomarker discovery in CCA.^18–22

Recent studies have highlighted the clinical potential of circulating molecular profiles in recurrent CCA. Thanasukarn et al. demonstrated using MALDI-TOF MS and LC-MS/MS serum peptidomics that peptide mass fingerprints could distinguish early from late recurrence, with early recurrence associated with pathways related to DNA damage response, proliferation, inflammation, and immunosuppressive signaling contributing to a tumor-promoting microenvironment.²³ In parallel, Prajumwongs et al. identified significant alterations in amino acid and lipid metabolism, including carnitines and glycerophospholipids, reflecting enhanced energy production, fatty acid β-oxidation, membrane remodeling, and immune-modulatory metabolic pathways linked to tumor progression and recurrence.²¹ Together, these findings support the potential of integrated peptidomic and metabolomic biomarkers for improving recurrence stratification and understanding recurrence biology in CCA. While previous peptidomic and metabolomic studies provided important insights into recurrence-associated molecular alterations in CCA, an important strength was that both analyses were conducted in the same patient cohort, enabling integrated interpretation across multiple molecular levels. However, these approaches primarily reflected downstream secreted peptides and metabolic products, whereas comprehensive proteomic profiling at the functional protein level, which may more directly capture signaling pathways and tumor–microenvironment interactions associated with recurrence, remains limited.

Therefore, using the same patient cohort from our previous peptidomic and metabolomic studies, this study aimed to identify serum protein biomarkers associated with early and late recurrence after curative resection and to characterize biological pathways related to recurrence timing, tumor progression, and tumor–microenvironment interactions. In addition, a proteomic recurrence score (PRS) to distinguish early from late recurrence and explored its associations with overall survival in patients with CCA who subsequently developed recurrence (Fig. 1).

Fig. 1. Overview of the study workflow for serum proteomic profiling, candidate protein discovery, and proteomic recurrence score construction and clinical integration in cholangiocarcinoma.

2. Materials and methods

2.1. Ethics approval and consent to participate

The study was conducted in accordance with the principles of Good Clinical Practice, the Declaration of Helsinki, and applicable national regulations. Ethical approval was obtained from the Khon Kaen University Ethics Committee for Human Research (approval number HE661318). Written informed consent was obtained from all participants prior to sample collection. All clinical data were de-identified prior to analysis, and no personally identifiable information was accessible to the researchers. Patient identifiers, including hospital numbers (HN), were removed or replaced with coded study identifiers to ensure that individual participants could not be identified.

2.2. Study population and data collection

Preoperative serum samples were obtained from patients with cholangiocarcinoma (CCA) who underwent curative-intent resection at Srinagarind Hospital, Khon Kaen University, Thailand. Samples were retrieved from the Cholangiocarcinoma Research Institute (CARI) biobank and derived from the same patient cohort previously used for peptidomic and metabolomic analyses, enabling integrative multi-omics interpretation. Clinical and pathological data from January 2017 to December 2021 were retrospectively collected from hospital records and the Isan Cohort database. Variables included demographic characteristics, tumor morphology and differentiation, surgical margin status, lymph node status, adjuvant chemotherapy, preoperative laboratory parameters, and tumor markers (CEA and CA19-9). Preoperative hematological parameters were obtained from routine complete blood count tests performed within one week before surgery. If more than one complete blood count result was available during this period, the result closest to the date of surgery was used. The neutrophil-to-lymphocyte ratio (NLR) was calculated by dividing the absolute neutrophil count by the absolute lymphocyte count. Patients were categorized into low- and high-NLR groups using the median NLR value of the study cohort as the cut-off. NLR values below the median were classified as low, whereas values equal to or above the median were classified as high. Histopathological evaluation was performed on H&E-stained sections according to the 2019 WHO classification, and tumor staging followed the 8th AJCC/UICC TNM system.²⁴ Stromal tumor-infiltrating lymphocyte (TIL) data were obtained from the previously published study by Naeklang et al.²⁵

2.3. Clinical outcome assessment

Patients who underwent surgical resection between January 1, 2017 and December 31, 2021 were followed from the date of surgery until death, the last documented clinical follow-up, or the administrative censoring date of December 31, 2023, whichever occurred first. Survival outcomes were assessed using hospital medical records and verified against the national database maintained by the Ministry of Interior. Routine follow-up visits were generally conducted every six months. The primary outcomes of interest were overall survival (OS) and disease-free survival (DFS). Recurrence was determined based on radiological and/or clinical evidence during follow-up and was classified as early or late according to the time interval from surgical resection to the first documented recurrence. Early recurrence was defined as the first documented recurrence occurring within 12 months after surgery (<12 months), whereas late recurrence was defined as recurrence occurring at or beyond 12 months (≥12 months). Patients without documented recurrence were censored at the date of the last clinical follow-up or the administrative censoring date, whichever occurred first.

2.4. Serum sample collection and storage

Peripheral blood samples were collected prior to surgical treatment via venipuncture into clot-activator tubes. After complete clot formation, samples were centrifuged at 3000–3500 rpm for 10 minutes at 4 °C to separate serum. The serum was carefully aliquoted into microcentrifuge tubes to avoid repeated freeze–thaw cycles and stored at −80 °C in the CARI biobank until analysis. Protein quantification was performed using the Lowry assay.

2.5. Serum proteomic profiling and proteomic recurrence score development

2.5.1. Serum proteomic analysis using LC-MS/MS

Serum proteomic profiling was performed using liquid chromatography–tandem mass spectrometry (LC-MS/MS). No selective depletion of highly abundant serum proteins was performed. Serum samples were first centrifuged to remove insoluble material and particulate debris. The resulting supernatants were then collected, diluted, and subjected to subsequent protein processing. Serum proteins were reduced with dithiothreitol (DTT), alkylated with iodoacetamide (IAA), and digested overnight with trypsin at 37 °C. The resulting peptides were desalted using C18 ZipTip (Millipore) before analysis. Peptide separation was carried out using an Ultimate 3000 Nano/Capillary LC system (Thermo Fisher Scientific) coupled to a Hybrid quadrupole Q-ToF Impact II™ mass spectrometer (Bruker Daltonics) equipped with a CaptiveSpray nano-electrospray ion source. For chromatographic separation, peptide digests were loaded onto a C18 µ-precolumn for enrichment and desalting, followed by separation on an Acclaim PepMap RSLC C18 analytical column (75 µm × 15 cm, 2 µm particle size, 100 Å pore size). Peptides were eluted using a binary solvent system consisting of 0.1% formic acid in water (solvent A) and 0.1% formic acid in 80% acetonitrile (solvent B) with a linear gradient of 5–55% solvent B over 30 min at a flow rate of 0.30 µL min⁻¹. MS and MS/MS spectra were acquired in positive-ion mode over an m/z range of 150–2200 using collision-induced dissociation (CID). Each biological sample was analyzed in triplicate to ensure analytical reproducibility. Raw LC-MS/MS data were acquired in technical triplicate for each sample, run in a single batch, and processed together in MaxQuant (version 2.5.0.0) using the integrated Andromeda search engine against the UniProt Homo sapiens reference proteome (downloaded 17 February 2026). Instrument performance was monitored throughout acquisition using a tryptic digest of bovine serum albumin as a quality-control standard. Technical replicates were combined at the MaxQuant analysis level by including each replicate as a separate raw file within a single search, with LFQ intensities from each replicate treated as independent values for downstream statistical analysis. Label-free quantification was performed using the MaxLFQ algorithm, with peptide and protein identification controlled at a 1% false discovery rate (FDR) at both the peptide-spectrum-match and protein levels. Search parameters were as follows: enzyme specificity, trypsin with up to two missed cleavages; mass tolerance, 20 ppm for the first search and 10 ppm for the main search; fixed modification, carbamidomethylation of cysteine; variable modifications, methionine oxidation and protein N-terminal acetylation; minimum peptide length, 7 amino acids; minimum unique peptides per protein, 1; “match between runs” enabled with a matching time window of 0.7 min; all other parameters were kept at MaxQuant default settings. Common contaminants were removed automatically using MaxQuant's built-in contaminants fasta database.

2.5.2. Candidate protein selection and functional enrichment analysis

Candidate proteins associated with recurrence were identified using the MetaboAnalyst platform (version 6.0)²⁶ through integrated univariate and multivariate analyses. Differential protein expression between groups was assessed using Welch's t-test with Benjamini–Hochberg false discovery rate (FDR) correction, and proteins with FDR-adjusted p < 0.05 were considered statistically significant. Fold change (FC) thresholds of ≥1.2 or ≤0.83 were applied to determine the magnitude of differential expression. Partial least squares–discriminant analysis (PLS-DA) was subsequently performed, and variable importance in projection (VIP) scores were used for feature ranking, with VIP ≥ 1.5 considered significant. Proteins meeting all combined criteria (FDR-adjusted p < 0.05, FC ≥ 1.2 or ≤0.83, and VIP ≥ 1.5) were retained for downstream analyses.

Functional enrichment analysis was performed using ShinyGO (version 0.77) to identify enriched Gene Ontology (GO) biological processes and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways. Enrichment significance was evaluated using the hypergeometric test with Benjamini–Hochberg FDR correction. Protein–protein interaction (PPI) network analysis was conducted using the STRING database (version 12.0).

2.5.3. Detection rate filtering and elastic net model development using repeated nested cross-validation

Candidate proteins identified from univariate analysis (adjusted p-value and fold change) and multivariate analysis (PLS-DA) were filtered based on detection frequency to prioritize consistently detectable candidates for biomarker development. Proteins detected in ≥50% of samples were retained to reduce the influence of missing values in label-free proteomic analysis.²⁷ Elastic net–regularized logistic regression was subsequently applied for feature set optimization. Candidate proteins identified from the preceding statistical analyses were further evaluated using an elastic net–regularized logistic regression model. To minimize overfitting and ensure unbiased estimation of model performance, model development was conducted using a repeated nested cross-validation framework. The outer loop consisted of repeated stratified five-fold cross-validation and was used exclusively for performance evaluation, whereas the inner loop employed stratified four-fold cross-validation for feature ranking, candidate protein selection, and hyperparameter optimization.

All preprocessing steps were performed independently within each training fold to prevent information leakage. Detailed preprocessing, scaling, model optimization, and evaluation procedures are summarized in SI Table S1. Proteins with no variability were excluded, and missing values were imputed using the median estimated from the corresponding training fold. Protein abundances were subsequently standardized using z-score normalization, with the mean and standard deviation calculated exclusively from the training data and then applied to the corresponding validation and outer-test folds without refitting. Candidate proteins were ranked according to the absolute coefficients of an L1-penalized logistic regression model, and multiple candidate panel sizes were evaluated within the inner cross-validation loop. The final classifier was an Elastic Net logistic regression model with optimized regularization strength and mixing parameter selected according to the highest mean receiver operating characteristic area under the curve (ROC–AUC) during inner cross-validation.

After model selection, the optimal classifier was refitted using the complete outer-training set and evaluated on the corresponding untouched outer-test fold. Because the outer cross-validation was repeated, each sample received multiple out-of-fold predicted probabilities, which were averaged to generate the final prediction. Model performance was assessed using ROC–AUC, area under the precision–recall curve (PR–AUC), Brier score, accuracy, sensitivity, specificity, precision, F1-score, and confusion-matrix statistics. Bootstrap resampling with 2000 iterations was used to estimate 95% confidence intervals for ROC–AUC and PR–AUC.

2.5.4. Proteomic recurrence score construction

A proteomic recurrence score (PRS) was constructed using a fixed candidate protein signature prioritized after feature-set size optimization and feature-selection stability analysis, together with elastic net–penalized logistic regression.^28,29 Protein expression values were preprocessed using median imputation and z-score standardization before model training. Hyperparameters were optimized using grid search with stratified 5-fold cross-validation based on ROC-AUC. The PRS for each patient was calculated as a weighted linear combination of standardized protein expression values and their corresponding regression coefficients derived from the elastic net model, according to the following equation:where x_j represents the standardized expression level of the j-the protein and β_j represents the corresponding regression coefficient.^28,30 Model performance was evaluated using out-of-fold predictions, and discriminative ability was assessed using ROC-AUC. The optimal classification cutoff was determined using the Youden index, and classification performance was evaluated using sensitivity, specificity, and accuracy.

2.6. Software and statistical analysis

All machine learning and bioinformatic analyses were performed using Python (version 3.13.5; Anaconda, Inc.) with pandas (version 2.2.3), NumPy (version 2.1.3), and scikit-learn (version 1.6.1). Clinical and survival analyses were performed using IBM SPSS Statistics version 27 (IBM Corp., Armonk, NY, USA). Continuous variables were compared using the Mann–Whitney U test, while categorical variables were analyzed using the chi-square or Fisher's exact test. Disease-free survival (DFS) and overall survival (OS) were estimated using the Kaplan–Meier method and compared using the log-rank test. Variables with p < 0.20 in univariate Cox regression were entered into multivariate Cox proportional hazards models. Hazard ratios (HRs) and 95% confidence intervals (CIs) were reported, and p < 0.05 was considered statistically significant.

3. Results

3.1. Proteomic profiling identifies candidate proteins and enables accurate classification of recurrence

This retrospective study included serum samples from 80 patients with cholangiocarcinoma (CCA) collected between January 1, 2017, and December 31, 2021. All patients underwent curative surgical resection and were followed up according to the protocol described in the Methods section. Recurrence status was defined based on the predefined clinical criteria. Patients who developed recurrence were further stratified into two groups: early recurrence, defined as disease-free survival (DFS) < 365 days (n = 34), and late recurrence, defined as DFS ≥365 days (n = 46). This cutoff was adopted based on previous publications.^4,5,21

To identify proteins associated with recurrence, both univariate and multivariate analyses were performed. LC-MS/MS initially identified 16 462 proteins across all samples. The PLS-DA score plot showed a clear separation between early and late recurrence groups, indicating distinct serum proteomic profiles (Fig. 2A). Consistently, VIP analysis identified several proteins with high discriminative importance (Fig. 2B), while differential expression analysis further revealed significantly altered proteins, as illustrated in the volcano plot (Fig. 2C). When the selection criteria were integrated, 232 proteins overlapped among the predefined thresholds of VIP ≥1.5, adjusted p < 0.05, and fold change (FC) ≥1.2 or ≤0.83, representing a robust set of recurrence-associated candidate proteins (Fig. 2D). Based on this refined feature set, hierarchical clustering demonstrated clear grouping of samples according to recurrence status (Fig. 2E). Similarly, PCA based on these candidate proteins showed improved separation between groups, confirming distinct global expression patterns (Fig. 2F).

Fig. 2. Proteomic profiling identifies candidate proteins and enables accurate classification of recurrence in CCA. (A) PLS-DA score plot showing separation between early and late recurrence groups (95% confidence ellipses). (B) VIP scores highlighting top discriminative proteins. (C) Volcano plot of differentially expressed proteins (FDR-adjusted p < 0.05; fold change ≥ 1.2 or ≤ 0.83). (D) Venn diagram showing overlap between VIP ≥ 1.5 and FDR < 0.05, identifying 232 candidate proteins. (E) Hierarchical clustering heatmap of candidate proteins showing distinct group-specific expression patterns. (F) PCA score plot based on candidate proteins. (G) PLS-DA score plot using the refined candidate protein set. (H) VIP scores from PLS-DA of candidate proteins. (I) Model performance (R² and Q²) and permutation test validating the PLS-DA model (p = 0.038). (J) Representative violin plots of selected proteins with corresponding FDR, fold change (FC), and VIP values.

To further evaluate the discriminatory ability of this candidate protein set, PLS-DA was re-performed using the 232 proteins. The resulting score plot showed even clearer separation between groups with reduced overlap (Fig. 2G), indicating improved discrimination after feature refinement. Correspondingly, VIP analysis highlighted the most influential proteins within the candidate proteins (Fig. 2H), suggesting their strong contribution to group separation. Model performance evaluation demonstrated high explanatory and predictive ability, with R² and Q² values indicating good model fit and predictive robustness (Fig. 2I). Permutation testing further confirmed that the model performance was statistically significant (p = 0.038), indicating that the observed separation was unlikely to occur by chance.

Among the 232 identified proteins, 134 were upregulated in the early recurrence group, whereas 98 were downregulated, or relatively upregulated in the late recurrence group. Representative candidate proteins exhibited consistent differential expression patterns between the two groups, with several showing markedly distinct abundance levels (Fig. 2J). The complete list of candidate proteins is provided in SI Data 1.

Beyond classification, these findings suggest that early and late recurrence in CCA may be associated with distinct serum proteomic patterns. The observed separation between the two groups raises the possibility that differences in recurrence timing may reflect underlying biological variation, potentially involving tumor-related processes, inflammatory signaling, and regulation of the tumor microenvironment. These potential biological associations were further explored through pathway and functional enrichment analyses.

3.2. Functional enrichment and protein interaction network reveal immune-associated and oncogenic mechanisms underlying recurrence

Functional enrichment analysis was performed to investigate the biological relevance of recurrence-associated candidate proteins. Gene Ontology (GO) biological process analysis revealed predominant enrichment of immune and inflammatory pathways, particularly cytokine-mediated signaling, indicating a central role of tumor–immune interactions in recurrence (Fig. 3A). Additional enriched processes included immune cell differentiation, T-cell and leukocyte regulation, apoptotic signaling, angiogenesis, vascular development, and cellular responses to DNA damage, suggesting involvement of immune modulation, tumor microenvironment remodeling, genomic instability, and survival adaptation in recurrence biology.

Fig. 3. Functional enrichment and protein–protein interaction network reveal recurrence-associated biological pathways in cholangiocarcinoma. (A) Gene Ontology (GO) biological process enrichment analysis of candidate proteins, ranked by fold enrichment; dot size indicates gene count and color represents statistical significance (−log10 FDR). (B) Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis showing key pathways associated with recurrence; dot size indicates gene count and color represents −log10(FDR). (C) Protein–protein interaction (PPI) network constructed using STRING with a high-confidence interaction score (≥0.7), highlighting functionally clustered modules, including inflammatory/immune signaling (e.g., IL6–JAK–STAT axis), complement/acute-phase response, metabolism/hypoxia, epithelial–mesenchymal transition, migration/invasion and extracellular matrix remodeling, RNA processing/miRNA regulation, and DNA damage and cellular stress response. Proteins circled in red indicate higher expression in early recurrence, whereas those circled in green indicate lower expression in early recurrence. Solid lines represent high-confidence interactions (≥0.7), and dashed lines indicate moderate–confidence interactions (≥0.4).

Consistent with these findings, Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analysis further highlighted enrichment of pathways broadly involved in tumor immunity and cancer progression (Fig. 3B). These pathways were predominantly associated with cytokine-mediated and immune regulatory signaling, with notable enrichment of JAK–STAT signaling and hypoxia-responsive pathways. The JAK–STAT pathway represents a major cytokine-driven signaling axis regulating immune responses, proliferation, and survival, whereas hypoxia-associated pathways mediated by HIF1A are linked to metabolic reprogramming, angiogenesis, and tumor progression. Collectively, these findings suggest that dysregulated signaling networks play important roles in recurrence-associated tumor behavior and immune evasion.

To further characterize molecular interactions among candidate proteins, a protein–protein interaction (PPI) network was constructed using STRING and visualized with functional clustering (Fig. 3C). In this network, proteins circled in red indicate higher expression in early recurrence, whereas proteins circled in green indicate lower expression in early recurrence. Edges represent functional associations, with solid lines indicating high-confidence interactions (≥0.7) and dashed lines indicating moderate–confidence interactions (≥0.4). The full interaction network at a moderate confidence threshold (≥0.4) is provided in SI Fig. 1.

The network was organized into several interconnected functional modules corresponding to established biological pathways.^31,32 At the center of the network, an inflammatory and immune signaling hub composed of IL6, JAK1, JAK3, STAT3, and STAT4 formed the major interaction cluster, with IL6 and STAT3 acting as central hub nodes. Several proteins within this module showed higher expression in early recurrence, supporting activation of the IL6–JAK–STAT signaling axis. Persistent IL-6/STAT3 signaling has been widely implicated in cancer progression through promotion of tumor cell survival, proliferation, angiogenesis, and immune evasion. STAT3-mediated signaling may also suppress effective antitumor immunity by impairing dendritic cell maturation and cytotoxic T-cell activity. Therefore, the central positioning of IL6, JAK1/JAK3, and STAT3/STAT4 in the present network suggests that inflammatory signaling may function as an important coordinating mechanism linking multiple recurrence-related processes rather than representing a secondary bystander phenomenon. These findings are consistent with previous reports demonstrating sustained IL-6/STAT3 activation in CCA and its association with aggressive tumor behavior and anti-apoptotic signaling.

Closely connected modules included extracellular matrix remodeling and invasion pathways involving MMP2, MMP9, CD44, and COL3A1, together with epithelial–mesenchymal transition (EMT)-related proteins such as ZEB1, SNAI1, SOX4, and SLC4A11. These findings support the biological relevance of invasive and migratory programs in recurrence. EMT-related plasticity facilitates dissemination and survival of residual tumor cells following surgical resection, whereas matrix remodeling promotes the establishment of a permissive microenvironment for tumor re-initiation.³¹ Collectively, these findings suggest that recurrence, particularly early recurrence, may reflect a biologically aggressive and highly invasive phenotype rather than simply earlier detection of recurrent disease.

Another interconnected cluster represented metabolism- and hypoxia-related pathways centered on HIF1A and PKM, together with metabolic regulators including FASN and HK2. HIF-1 signaling is a key regulator of hypoxic adaptation and metabolic reprogramming, whereas PKM contributes to glycolysis and anabolic metabolism required for tumor progression.^18,33 These observations suggest that hypoxia-driven metabolic rewiring may represent an important component of recurrence biology in CCA. Importantly, metabolic reprogramming may also interact with inflammatory and immune-related pathways, thereby linking metabolic adaptation with immune suppression within the tumor microenvironment.

A complement and acute-phase response module containing C3, CFH, HPX, and ADAMTS13 was also identified, indicating dysregulation of innate immune responses. Complement signaling is increasingly recognized as an important component of the tumor microenvironment that can promote inflammation, tumor progression, metastasis, and immune modulation.^34–36 In particular, CFH may contribute to immune evasion by limiting complement-mediated cytotoxicity and protecting tumor cells from innate immune clearance. These findings suggest that complement-associated innate immune remodeling may represent an additional layer of recurrence biology in CCA, although further functional validation is required. Additional peripheral subnetworks included DNA damage and cellular stress response proteins, such as TOP2A, FANCD2, FANCB, TIMELESS, and PLK2, and RNA processing or miRNA-related proteins such as RSL24D1, KRR1, and SNRNP200. Collectively, these analyses demonstrate coordinated activation of inflammatory, invasive, metabolic, complement-related, and immune-associated pathways in recurrent CCA.

The current findings are also concordant with our previous serum peptidomic and metabolomic studies conducted in the same patient cohort.^21,23 Previous peptidomic analysis demonstrated that early recurrence was associated with DNA damage response, proliferative stress, and inflammatory signaling, whereas late recurrence was more closely linked to immune regulation and complement-related pathways.²³ The present proteomic analysis extends these observations at the intact protein level by identifying inflammatory signaling, complement regulation, EMT, and metabolic adaptation as interconnected recurrence-associated processes. Similarly, our prior metabolomic study identified alterations in amino acid metabolism, lipid-derived carnitines, glycerophospholipids, and kynurenine-related metabolites, suggesting enhanced metabolic flexibility in recurrent tumors.²¹ Integrating these multi-omics findings, the present study supports an immunometabolic model of recurrence in which inflammatory signaling, hypoxia-associated metabolic reprogramming, and immune modulation converge to promote tumor persistence and recurrence after surgery.

3.3. Detection frequency–guided feature refinement and nested cross-validated elastic net modeling identify an exploratory protein signature

Although the initial set of 232 candidate proteins demonstrated discriminatory ability for recurrence classification, not all proteins were equally suitable for clinical application because reliable biomarkers should be consistently detectable across samples. Therefore, a detection frequency–guided feature refinement strategy was incorporated into the analytical pipeline to prioritize robust and clinically applicable proteins while maintaining predictive performance (Fig. 4A). This strategy was intended to reduce the impact of missingness and technical variability, as low-detection features may be less suitable for future clinical translation.^27,37

Fig. 4. Detection frequency–guided feature refinement and elastic net modeling for clinically applicable protein set selection. (A) Schematic overview of the analytical pipeline, including detection filtering, feature selection, proteomic recurrence score construction, nested cross-validation, and survival analysis. (B) Protein detection rate rank plot with a 50% cutoff for feature filtering. (C) Hierarchical clustering heatmap of filtered proteins showing separation between early and late recurrence groups. (D) Elastic net performance across protein sets (top 5, 10, 15, 20, and full) during nested feature-set size optimization, evaluated by precision, recall, F1-score, and accuracy. (E) ROC curves across feature sets. (F) Feature stability based on selection frequency across repeated nested cross-validation. (G) Elastic net coefficients indicating direction and magnitude of protein associations. (H) SHAP-like plot showing feature contributions to model predictions.

Proteins were ranked according to detection frequency across samples, revealing a gradual decline in detectability (Fig. 4B). Applying a ≥50% detection rate threshold reduced the candidate set to 33 consistently detectable proteins (SI Data 2). To investigate whether the excluded low-detection proteins nevertheless showed potentially informative group-associated detection patterns, the 199 proteins detected in fewer than 50% of the cohort were examined separately. Detection frequencies were compared between the early- and late-recurrence groups using Fisher's exact test, followed by Benjamini–Hochberg correction across the 199 comparisons. Although 114 proteins showed nominal unadjusted p < 0.05, none met the FDR-adjusted significance threshold of 0.05. Forty-two proteins showed an absolute between-group difference in detection frequency of at least 20 percentage points and were reported as exploratory low-detection candidates in SI Data 3. Hierarchical clustering based on the 33 retained proteins showed partial grouping of early- and late-recurrence cases (Fig. 4C), indicating that group-associated abundance patterns remained detectable after detection-frequency filtering.

To prioritize a practical proteins for subsequent targeted verification, elastic net–regularized logistic regression was evaluated within a nested cross-validation framework using progressively sized feature sets comprising the top 5, 10, 15, and 20 ranked proteins, together with the full 33-protein set (SI Data 4). Within each outer training fold, model preprocessing, hyperparameter tuning, protein ranking, and feature selection were performed without access to the corresponding outer test fold. Thus, the composition of each reduced protein panel, including the 20-protein set, was determined independently within each training fold and could vary across folds. Patient-level out-of-fold predictions were subsequently aggregated to estimate the internal discriminatory performance of each protein set. During this nested panel-size optimization, the model performance improved progressively as additional proteins were incorporated into the classifier (Fig. 4D). The 5-protein feature set achieved a patient-level out-of-fold ROC-AUC of 0.836 (95% CI, 0.743–0.917) and a PR-AUC of 0.798 (95% CI, 0.656–0.912), with an accuracy of 78.8%. Expanding the feature set to 10 proteins improved the ROC-AUC to 0.887 (95% CI, 0.807–0.955) and the accuracy to 81.3%, while the 15-protein feature set further increased the ROC-AUC to 0.919 (95% CI, 0.850–0.970) and achieved an accuracy of 82.5%. The model based on 20 candidate proteins showed additional improvement, yielding a ROC-AUC of 0.937 (95% CI, 0.881–0.979) (Fig. 4E) and a PR-AUC of 0.918 (95% CI, 0.828–0.976), together with an accuracy of 83.8%, sensitivity of 85.3%, specificity of 82.6%, and an F1-score of 0.817. The corresponding confusion matrix comprised 29 true positives, 38 true negatives, 8 false positives, and 5 false negatives. The model incorporating all 33 retained proteins achieved the highest internal performance, with a patient-level out-of-fold ROC-AUC of 0.949 (95% CI, 0.895–0.987) and a PR-AUC of 0.929 (95% CI, 0.846–0.984). This model achieved an accuracy of 87.5%, sensitivity of 91.2%, specificity of 84.8%, and an F1-score of 0.861, corresponding to 31 true positives, 39 true negatives, 7 false positives, and 3 false negatives (SI Data 4).

Although the model incorporating all 33 retained proteins achieved the highest internal discriminatory performance, the gain over the model based on 20 candidate proteins was modest, with absolute increases of only 0.012 in ROC-AUC and 0.010 in PR-AUC. In contrast, limiting the model to 20 proteins reduced the number of candidate analytes by approximately 39% while retaining comparable internal discrimination. Because inclusion of the additional 13 proteins would increase the complexity of subsequent targeted MRM development, including peptide selection, transition optimization, and assessment of analytical interference, the 20-protein feature set was selected as a more parsimonious approach that balanced predictive performance with the practical feasibility of future targeted verification. These performance estimates were derived from nested cross-validation and therefore represent internal validation within the present cohort.

Feature-selection stability analysis across the nested cross-validation procedure showed that the prioritized candidate proteins were repeatedly selected during model development (Fig. 4F). CARD6, MYH10, ZEB1, ARHGF, PKM, HHLA3, IL6, C3, CD44, and VEGF-D were selected in all outer-fold models, whereas MMP9, MMP2, ADAM8, DYH10, HIF1A, SNAI1, and DNHD1 were selected in 96% of the models. STAT4 and STAT3 were selected in 92% of the models, while CFH was selected in 88%. Overall, the 20 prioritized candidate proteins demonstrated high internal selection stability across the nested cross-validation iterations (SI Data 5), supporting their prioritization for subsequent targeted verification.

To further characterize their multivariable contributions, an elastic net model was descriptively refitted using the 20 most stable candidate proteins identified through the nested cross-validation analysis. Positive standardized coefficients were observed for MYH10, ARHGF, CARD6, PKM, ZEB1, VEGF-D, STAT3, ADAM8, HHLA3, MMP9, C3, IL6, MMP2, DYH10, HIF1A, SNAI1, and CD44, whereas CFH, STAT4, and DNHD1 had negative coefficients (Fig. 4G). The corresponding patient-level contribution plot illustrated the direction and relative magnitude of each protein's contribution to the fitted model predictions (Fig. 4H). Because this descriptive model was refitted using the complete cohort, the standardized coefficients should be interpreted as conditional multivariable contributions within the fitted model rather than as independently validated effect estimates or direct between-group differences in protein abundance. Collectively, these analyses supported prioritization of the 20 most stable proteins for subsequent proteomic recurrence score construction and targeted MRM verification. Independent analytical and clinical validation will nevertheless be required before potential clinical application.³⁸

3.4. Proteomic recurrence score stratifies early from late recurrence and was associated with adverse clinicopathological and survival outcomes

To evaluate the discriminatory ability of the proteins prioritized by the preceding stability analysis, the 20 most stable proteins were defined as a fixed exploratory signature and modeled using elastic net–regularized logistic regression to derive a proteomic recurrence score (PRS). Patient-level out-of-fold probabilities were generated using repeated nested five-fold cross-validation, with each patient receiving five outer-fold test predictions that were averaged to obtain the final OOF probability. The PRS was defined as the logit transformation of this averaged probability, with higher values indicating a greater predicted probability of early recurrence. All preprocessing steps, including median imputation, standardization, and hyperparameter tuning, were performed within the corresponding outer training folds (SI Data 6).

The fixed 20 candidate proteins achieved a patient-level OOF ROC-AUC of 0.948 for distinguishing early from late recurrence. For distinguishing early from late recurrence. An exploratory decision threshold was derived from the pooled out-of-fold predictions using the Youden index. The resulting predicted-probability threshold was 0.565, corresponding to a logit-transformed PRS cutoff of 0.261. With early recurrence treated as the positive class, this threshold yielded an accuracy of 92.5%, sensitivity of 91.2%, specificity of 93.5%, precision of 91.2%, and an F1-score of 0.912. Overall, 74 of the 80 patients were correctly classified, including 31 of 34 patients with early recurrence and 43 of 46 patients with late recurrence; three patients in each group were misclassified. Because the threshold was derived and evaluated using the pooled out-of-fold predictions from the same cohort, these cutoff-based performance estimates should be considered exploratory and require confirmation in an independent cohort (SI Data 7).

Correlation analyses further demonstrated associations between PRS and immune-related variables (Fig. 5B and C). PRS showed a positive correlation with neutrophil-to-lymphocyte ratio (NLR) (Spearman's r = 0.48, p = 3.17 × 10⁻⁵), indicating that higher proteomic risk was associated with increased systemic inflammation. Conversely, PRS was negatively correlated with stromal tumor-infiltrating lymphocytes (%TILs) (r = −0.48, p = 2.79 × 10⁻⁵), suggesting reduced local immune infiltration in high-risk tumors. NLR was also negatively correlated with %TILs (r = −0.57, p = 3.73 × 10⁻⁷), supporting an inverse relationship between systemic inflammation and antitumor immune response. Visual clustering patterns further suggested that patients with high PRS tended to exhibit lower %TILs and higher NLR values. Because stromal %TIL data were available for 69 of the 80 patients, baseline clinicopathological characteristics were compared between patients with evaluable and missing %TIL data. No statistically significant differences were observed across the variables examined (SI Table S2), providing some reassurance that the missing %TIL data did not introduce substantial systematic differences between the two groups.

Fig. 5. Proteomic recurrence score (PRS) associates with immune context and clinical outcomes in cholangiocarcinoma. (A) Out-of-fold (OOF) PRS distribution with Youden cutoff for risk stratification. (B) Spearman correlation heatmap between PRS and clinical variables; sample sizes vary due to missing values across variables (n = 69 for %TILs, n = 74 for CEA, and n = 72 for CA19-9). (C) Relationship among PRS, neutrophil-to-lymphocyte ratio (NLR), and tumor-infiltrating lymphocytes (%TILs); bubble size/color indicate %TILs. (D) Kaplan–Meier analysis of overall survival by PRS groups. (E) Multivariate Cox analysis for overall survival (hazard ratios and 95% CI).

These molecular and clinicopathological findings are biologically consistent with the functional pathways identified in the present study. High PRS appears to reflect a tumor microenvironment characterized by systemic inflammation and reduced local immune surveillance. Elevated NLR has consistently been associated with poor prognosis across multiple solid tumors, including CCA, whereas higher TIL density generally reflects a more active antitumor immune response.^39,40 The observed positive association between PRS and NLR, together with the negative association between PRS and stromal TILs, supports the interpretation that the PRS captures an immune-suppressive and inflammation-dominant recurrence phenotype. These observations are also consistent with the enrichment of IL6–JAK–STAT signaling and hypoxia-associated metabolic reprogramming, both of which may contribute to immune evasion and tumor progression.^{18,19,32,33,41}

Because PRS showed a significant positive correlation with NLR, we next investigated whether the proteomic signature provided discriminatory information beyond this routinely available inflammatory marker. Using the same repeated nested cross-validation framework, three prespecified models were evaluated: NLR alone, the fixed 20 candidate proteins, and the combined NLR plus 20 candidate proteins. The NLR-only model demonstrated moderate discrimination between early and late recurrence, with a ROC-AUC of 0.836 (95% CI, 0.740–0.917) and a PR-AUC of 0.818 (95% CI, 0.693–0.916). The 20 candidate proteins achieved substantially higher discrimination, with a ROC-AUC of 0.948 (95% CI, 0.895–0.988) and a PR-AUC of 0.928 (95% CI, 0.847–0.986). Paired bootstrap analysis confirmed that the 20 candidate proteins significantly outperformed the NLR-only model in both ROC-AUC (ΔROC-AUC = 0.112, 95% CI, 0.015–0.217; p = 0.021), whereas the improvement in PR-AUC did not reach statistical significance (ΔPR-AUC = 0.110, 95% CI, −0.007 to 0.229; p = 0.062). Adding NLR to the 20 candidate proteins did not provide additional discriminatory benefit. The combined model achieved a ROC-AUC of 0.965 (95% CI, 0.922–0.995) and a PR-AUC of 0.944 (95% CI, 0.867–0.992) and significantly outperformed NLR alone in both ROC-AUC (ΔROC-AUC = 0.129, 95% CI, 0.047–0.217; p < 0.001) and PR-AUC (ΔPR-AUC = 0.126, 95% CI, 0.030–0.232; p = 0.009). However, adding NLR to the 20-protein signature did not significantly improve discrimination over the proteomic signature alone, either for ROC-AUC (ΔROC-AUC = 0.017, 95% CI, −0.012 to 0.049; p = 0.239) or PR-AUC (ΔPR-AUC = 0.016, 95% CI, −0.026 to 0.064; p = 0.451) (SI Fig. 2 and Data 8).

Although PRS, derived from the 20 candidate proteins developed to distinguish early from late recurrence, was moderately correlated with NLR, NLR alone also demonstrated meaningful discrimination between the two groups. This finding supports the relevance of systemic inflammation to recurrence timing and suggests that part of the information captured by PRS may reflect inflammatory variation represented by NLR. However, the Top 20 proteomic model significantly outperformed the NLR-only model in both ROC-AUC and PR-AUC, indicating that the proteomic signature captured broader discriminatory information beyond NLR alone. By integrating signals across 20 candidate proteins, the model may reflect multiple biological processes associated with recurrence, including inflammation, immune regulation, cellular stress, and metabolic alterations.

The lack of improvement after adding NLR to the Top20 model may reflect overlap in the predictive information captured by the two measures. The moderate correlation between PRS and NLR suggests that systemic inflammatory variation represented by NLR may already be partly incorporated within the proteomic signature. In addition, the high discrimination achieved by the Top20 model left limited scope for further improvement. Thus, although NLR showed useful standalone discrimination, it provided little additional information after accounting for the 20-protein feature set. Adding NLR slightly increased specificity while reducing sensitivity, but did not improve overall accuracy, ROC-AUC, or PR-AUC. The negligible change in Brier score further suggests that any improvement in probability estimation was minimal.

These findings should not be interpreted as indicating that NLR lacks clinical or prognostic relevance. Rather, its discriminatory information appeared to be largely encompassed by the proteomic model for distinguishing early from late recurrence in this cohort. Both NLR and circulating protein concentrations may also be influenced by biliary obstruction, cholangitis, biliary drainage, infection, and impaired liver function.^42,43 Because these factors were not consistently documented, residual confounding cannot be excluded. The relatively small sample size also limits the ability to rule out a modest incremental benefit. Future prospective studies should adjust for these clinical factors and validate the incremental value of the proteomic signature over routine inflammatory markers in larger, independent cohorts.

We next examined whether the PRS was associated with clinicopathological features and patient survival. Patients were stratified into low- and high-PRS groups using a Youden index-derived logit cutoff of 0.26, resulting in 45 patients in the low-PRS group (PRS < 0.26) and 35 (PRS ≥ 0.26) in the high-PRS group. NLR was dichotomized using the cohort median of 3.12, with values < 3.12 classified as low NLR and values ≥ 3.12 as high NLR. The PRS was significantly associated with several adverse clinicopathological characteristics (Table 1).

Table 1. Association of the proteomic recurrence score and clinicopathological characteristics with overall survival: univariate Cox proportional hazards analysis^a.

Variable	n = 80	PRS (Youden logit = 0.26)			OS (month)	Univariate
		Low (n = 45)	High (n = 35)	p-Value		HR (95%CI)	p-Value
Age (year)
<62	39	22	17	0.978	22.9	1	0.062
≥62	41	23	18		18.2	1.63 (0.97–2.72)
Gender
Male	34	21	13	0.393	22.9	1	0.087
Female	46	24	22		18	1.57 (0.93–2.65)
Tumor location
iCCA	50	26	24	0.225$	19.7	1	0.454
pCCA	30	19	11		18.0	1.22 (0.72–2.08)
Tumor morphology
With ID	34	20	14	0.690	26.8	1	0.103
Without ID	46	25	21		18	1.54 (0.91–2.60)
R status
R0	45	29	16	0.074$	22.9	1	0.042
R1	35	16	19		15.9	1.69 (1.01–2.82)
Tumor differentiation
Well-	64	37	27	0.573	22.0	1	0.033
Moderate-/poorly-	16	8	8		16.6	1.88 (1.04–3.39)
N status
N0	44	30	14	0.015$	22.9	1	0.050
N1	36	15	21		15.9	1.66 (0.99–2.76)
TNM stage
0-II	20	16	4	0.013	38.5	1	0.009
III-IV	60	29	31		16.6	2.30 (1.21–4.35)
Adjuvant CMT
No	30	15	15	0.261$	10.7	1	0.008
Yes	50	30	20		21.1	0.51 (0.31–0.85)
NLR (median = 3.12)
Low NLR: <3.12	40	37	3	<0.001	41.5	1	<0.001
High NLR: ≥ 3.12	40	8	32		10.7	3.23 (1.91–5.46)
PRS (Youden logit = 0.26)
Low PRS: < 0.26	45	—	—	—	32.8	1	<0.001
High PRS: ≥ 0.26	35	—	—		9	5.86 (3.86–10.26)

^aCategorical variables were analyzed using the chi-square test or Fisher's exact test (expected counts < 5, indicated by $). OS were assessed using the Kaplan–Meier method with log-rank tests. Univariate analysis was performed using Cox regression, and variables with p < 0.2 were included in multivariate models. Abbreviations: PRS = proteomic recurrence score; OS = overall survival; iCCA = intrahepatic cholangiocarcinoma; pCCA = perihilar cholangiocarcinoma; R status = resection margin status; R0 = microscopically margin-negative resection; R1 = microscopically margin-positive resection; N status = lymph nodes status; N0 = no lymph node metastasis; N1 = lymph node metastasis; CMT = chemotherapy; NLR = neutrophil-to-lymphocyte ratio.

Patients in the high PRS group had a higher prevalence of lymph node metastasis than those in the low PRS group (N1: 60.0% vs. 33.3%; p = 0.015), advanced TNM stage (stage III–IV: 88.6% vs. 64.4%; p = 0.013), and high NLR (91.4% vs. 17.8%; p < 0.001). R1 resection was also more frequent in the high-PRS group than in the low-PRS group, although the association was of borderline statistical significance (54.3% vs. 35.6%; p = 0.074). No significant associations were observed between PRS group and age, sex, tumor location, tumor morphology, tumor differentiation, or receipt of adjuvant chemotherapy (all p > 0.05).

Survival analyses showed that the PRS was associated with overall survival. Kaplan–Meier analysis showed that patients in the high PRS group had significantly shorter OS than those in the low PRS group (median OS: 9.0 vs. 32.8 months; log-rank p < 0.001; Fig. 5D).

In univariate Cox proportional hazards analysis, R1 resection (HR = 1.69, 95% CI: 1.01–2.82; p = 0.042), moderate/poor tumor differentiation (HR = 1.88, 95% CI: 1.04–3.39; p = 0.033), advanced TNM stage (HR = 2.30, 95% CI: 1.21–4.35; p = 0.009), receipt of adjuvant chemotherapy (HR = 0.51, 95% CI: 0.31–0.85; p = 0.008), elevated NLR (HR = 3.23, 95% CI: 1.91–5.46; p < 0.001), and a high PRS (HR = 5.86, 95% CI: 3.86–10.26; p < 0.001) were significantly associated with overall survival, whereas age (p = 0.062), sex (p = 0.087), and N status (p = 0.050) showed borderline significance (Table 1). Variables with a univariate p < 0.20 were subsequently included in the multivariable Cox proportional hazards model.

In multivariable Cox proportional hazards analysis, a high PRS remained independently associated with poorer overall survival after adjustment for clinicopathological variables (HR = 4.32, 95% CI: 1.62–11.50, p = 0.003). Older age (HR = 2.37, 95% CI: 1.31–4.28, p = 0.004), R1 (HR = 2.14, 95% CI: 1.22–3.78, p = 0.008), elevated NLR (HR = 2.71, 95% CI: 1.07–6.85, p = 0.036), and receipt of adjuvant chemotherapy (HR = 0.24, 95% CI: 0.13–0.47, p < 0.001) were also independently associated with overall survival (Fig. 5E). In contrast, sex, tumor morphology, tumor differentiation, lymph node status, and TNM stage were not associated with survival after multivariable adjustment. These findings indicate that the PRS provides prognostic information independent of established clinicopathological factors in this cohort. From a translational perspective, the PRS may support individualized postoperative management by identifying patients at high risk of early recurrence and poor survival. Such patients may benefit from closer surveillance, risk-adapted follow-up strategies, and consideration of more intensive adjuvant or investigational therapeutic approaches. However, further validation is required before clinical implementation.

Although the present study demonstrates the potential of serum proteomic profiling for distinguishing early and late recurrence after curative resection, several limitations should be acknowledged. First, the 232 candidate proteins were identified before nested cross-validation. Thus, although downstream feature selection and model tuning were performed within the nested framework, the reported performance remains conditional on this upstream preselection and may be optimistic. A fully nested discovery-to-model workflow and independent external validation will be required for a more stringent assessment of model performance. Second, this was a retrospective single-center study with a relatively small sample size, which may limit the generalizability of the findings despite the use of repeated nested cross-validation to minimize overfitting. Third, although the proteomic signature showed strong discriminatory performance, external validation in independent multicenter cohorts is required before clinical implementation. Fourth, detailed information regarding inflammatory conditions that may influence circulating protein levels, including cholangitis, biliary drainage procedures, active infection, and the severity of biliary obstruction, was not consistently available, and therefore residual confounding cannot be completely excluded. Finally, the present study was designed as a discovery-phase investigation using untargeted serum proteomics. Future studies employing targeted quantitative assays, such as multiple reaction monitoring, together with prospective validation, will be necessary to confirm the robustness and clinical applicability of the identified protein signature.

4. Conclusion

Comprehensive serum proteomic profiling identified proteins and biological pathways associated with early versus late recurrence in CCA, involving immune regulation, inflammatory signaling, metabolic reprogramming, and tumor progression. Through detection frequency–guided refinement and nested cross-validated elastic net modeling, we prioritized a parsimonious 20 protein signature for subsequent targeted verification. The resulting exploratory PRS showed promising internal discrimination between early and late recurrence and was moderately correlated with NLR and inversely correlated with stromal TIL levels. The PRS was also associated with adverse clinicopathological features and shorter overall survival, and a high PRS remained independently associated with poorer overall survival after adjustment for clinicopathological factors. These findings support the potential of serum proteomic signatures for postoperative risk stratification and prognostic assessment in CCA. Targeted analytical verification and external validation in independent prospective cohorts are required before clinical application.

Abbreviations

CCA

Cholangiocarcinoma

DFS

Disease-free survival

FC

Fold change

FDR

False discovery rate

GO

Gene ontology

KEGG

Kyoto encyclopedia of genes and genomes

LC-MS/MS

Liquid chromatography–tandem mass spectrometry

LFQ

Label-free quantification

NLR

Neutrophil-to-lymphocyte ratio

OS

Overall survival

PLS-DA

Partial least squares–discriminant analysis

PPI

Protein–protein interaction

PR-AUC

Area under the precision–recall curve

PRS

Proteomic recurrence score

ROC-AUC

Area under the receiver operating characteristic curve

TILs

Tumor-infiltrating lymphocytes

TME

Tumor microenvironment

VIP

Variable importance in projection

Author contributions

Conceptualization, methodology, formal analysis, investigation, data curation, visualization, writing – original draft, writing – review and editing, P. P.; methodology, formal analysis, investigation, resources, data curation, supervision, writing – review and editing, A. T.; methodology, formal analysis, investigation, data curation, visualization, writing – review and editing, V. T.; methodology, writing – review and editing, A. J.; methodology, writing – review and editing, N. K.; methodology, writing – review and editing, N. N.; methodology, writing – review and editing, P. K.; methodology, writing – review and editing, A. W.; methodology, writing – review and editing, J. C.; methodology, writing – review and editing, P. S.; methodology, visualization, writing – review and editing, N.M.; methodology, writing – review and editing, S. C.; resources, supervision, writing – review and editing, S. R.; conceptualization, methodology, formal analysis, investigation, resources, supervision, project administration, funding acquisition, writing – review and editing, W. L. All the authors have read and agreed to the published version of the manuscript.

Conflicts of interest

There are no conflicts to declare.

Data availability

The raw MS/MS spectral data have been deposited in the ProteomeXchange Consortium repository under accession numbers JPST004606 (https://repository.jpostdb.org/preview/35886702569fd0cea8bcba; access key: 4085) and PXD078141. Clinical data supporting the findings of this study are available from the corresponding author upon reasonable request.

Supplementary information (SI) is available. See DOI: https://doi.org/10.1039/d6ra05199e.