Penile cancer (PeCa), predominantly squamous cell carcinoma (SCC), is a rare but aggressive malignancy. The global incidence is approximately 1 per 100.000 men in developed regions, with higher rates reported in parts of South America, Africa, and Asia (1). In Sweden, incidence has increased over the past two decades, now exceeding 2 per 100.000 men (2, 3). The median age at diagnosis is 68 years and established risk factors include human papillomavirus (HPV) infection (4, 5).

PeCa metastasizes primarily to inguinal and subsequently pelvic lymph nodes, and nodal involvement remains the most important prognostic factor (4). A recent Swedish study demonstrated a three-fold increase in disease-specific mortality among patients with lymph node metastases (LNM) (6), while distant metastases (M1) are uniformly fatal (7).

Current imaging lacks sensitivity for detecting occult LNM (8), and no validated predictive biomarkers are available (9–11), hence, clinically node-negative (cN0) patients are stratified into risk-groups based on histopathological features of the primary tumor (12, 13). High-risk patients undergo lymph node staging using dynamic sentinel node biopsy (DSNB), while therapeutic lymph node dissection, due to considerable morbidity, is reserved for those with confirmed metastases (pN+) (14). Hence, an unmet need for sensitive and accessible biomarkers enabling early identification of LNM remains.

Stimulatory and inhibitory immune checkpoint proteins (ICs) are essential regulators of immune homeostasis and play a fundamental role in tumor immune surveillance (15, 16). While benign inflammatory conditions, such as benign prostatic hyperplasia (BPH), may increase circulating ICs, cancer-related immune dysregulation is likely to produce distinct expression patterns or combinations of checkpoints, which may still carry diagnostic or prognostic value. Elevated expression of inhibitory ICs is commonly associated with tumor development, contributing to immune evasion. In contrast, increased expression of stimulatory ICs is more frequently linked to benign inflammatory conditions, reflecting immune activation rather than tumor immune escape. Traditionally, ICs such as PD-1, PD-L1, and CTLA-4, have been studied in their membrane-bound form, mediating cell-to-cell signaling within the tumor microenvironment. Increasing attention has been directed towards their soluble forms (sICs), detectable in plasma and other body fluids (17). These soluble forms retain functional capabilities and can interact with corresponding receptors or ligands at distant sites, potentially modulating systemic anti-tumor immunity.

Several studies have identified sICs as promising biomarkers for both prognosis and prediction of lymph node metastases in various malignancies, including osteosarcoma, pancreatic, kidney, and lung cancers (18–21). However, their role in PeCa remains unexplored, despite documented IC expression in PeCa tissue and growing interest in immunotherapy for advanced disease (11, 15, 22–24). Notably, recent studies have reported high PD-L1 expression, particularly among HPV-positive tumors, supporting further evaluation of immune-related biomarkers in PeCa (22–26).

To date, research on ICs in PeCa has been limited by small cohort sizes and a narrow range of ICs assessed. Larger studies are needed to clarify the biomarker potential of these proteins in PeCa. Given the non-invasive nature and clinical utility of liquid biopsies, evaluating sICs could enhance current diagnostic and prognostic strategies. Ultimately, sICs may emerge as valuable prognostic tools, facilitating earlier detection of LNM, thereby enabling less mutilating treatments and optimally improving survival.

To our knowledge, this is the first study to systematically investigate a broader panel of sICs in PeCa. The primary aim was to evaluate their predictive potential as biomarkers for identifying LNM. Additionally, plasma sIC levels were compared between men with and without PeCa. This study aims to establish a foundation for non-invasive, immune-based risk stratification in PeCa, with potential implications for clinical decision-making in patients with clinically localized disease.

From the initial cohort, 284 PeCa patients and 45 cancer-free controls met the inclusion criteria. Cases were divided into a training set (n=202) and a test set (n=82) (Figure 1). Clinicopathological characteristics are presented in Table 1. Pathologically confirmed LNM (pN+) were present in 24.1% of the training set and 35.9% of the test set. HPV positivity was observed in 41.6% and 45.6% of cases in the training and test sets, respectively. No statistically significant clinicopathological differences were observed between the two sets.

Plasma levels of 14 sICs were initially measured in all cases and controls. PD-L1 was excluded due to detectability below the predefined detection threshold (>50% of samples) since 96% of samples did not have detectable PD-L1. Thirteen sICs were thus included for further analyses.

A logistic regression model was developed to predict LNM based on the 13 included sICs. In the training set, the model achieved an accuracy of 77.5% (95% CI: 70.9 – 83.3) and a balanced accuracy of 54.1%. Specificity was high (99.3%), but sensitivity was low (8.9%), with a PPV of 80.0% and a NPV of 77.5% (Table 2). When pT-stages and tumor grades were added to the model, accuracy and specificity decreased slightly to 69.6% (95% CI: 60.2 – 78.0) and 85.5% (95% CI: 75.0 – 92.8), respectively, while sensitivity increased to 44.2% (95% CI: 29.1 – 60.1).

When validated in the test set, model performance declined, yielding an accuracy of 62.2% (95% CI: 50.8 – 72.7), sensitivity of 17.9%, and specificity of 85.2% (Table 2). The AUC was 67.1% (95% CI: 58.0 – 76.3) in the training set and 54.2% (95% CI: 40.3 – 68.0) in the test set (Figure 2), indicating limited generalizability. Applying the model incorporating pT-stage and tumor grade to the test data resulted in a lower accuracy of 46.9% (95% CI: 32.5 – 61.7) and specificity of 60.7% (95% CI: 40.6 – 78.5), while sensitivity increased to 28.6% (95% CI: 11.3 – 52.2).

To further explore the model´s limited predictive performance, plasma levels of the 13 included sICs were compared between patients with and without LNM within the training set. No statistically significant differences were observed for any sIC (Table 3a). Moreover, sIC levels did not vary significantly when the PeCa training set was stratified by pT-stage, tumor grade, or HPV-status (p > 0.05, Supplementary Table 1 and 2).

Penile cancer is a rare but aggressive disease, with prognosis strongly influenced by the presence of LNM. Current clinical tools often fail to predict nodal involvement accurately, highlighting the need for reliable, preferably non-invasive biomarkers. In this study, the predictive potential of 14 sICs for identifying LNM in patients with PeCa was evaluated. A logistic regression model incorporating 13 sICs achieved high specificity in the training set but showed low sensitivity and moderate balanced accuracy. Performance declined further in the independent internal test set, with low AUC values in both cohorts. No significant associations were observed between sIC levels and tumor characteristics, including pT-stage, tumor grade, HPV-status, or LNM. Although plasma sIC profiling remains a non-invasive biomarker approach, its predictive utility for LNM in PeCa appears limited, suggesting that systemic levels of sICs are not primarily driven by tumor aggressiveness.

The predictive model based on 13 detectable sICs demonstrated limited performance already in the training set, with low sensitivity and moderate accuracy despite high specificity. The AUC values (0.67 in the training set and 0.54 in the test set) further indicate poor discriminative capacity, suggesting that systemic levels of sICs alone provide insufficient predictive information for identifying LNM in PeCa. These results imply that sIC profiles may not adequately reflect the local immune dynamics driving metastatic progression.

When pT-stage and tumor grade were added to the model sensitivity increased, whereas accuracy and specificity decreased. This indicates that incorporating established clinicopathological variables slightly improved the model´s ability to identify LNM-positive cases but at the cost of reduced overall accuracy. Thus, even with additional clinical parameters, the model´s predictive capacity remained insufficient for clinical application. The small effect size between LNM-positive and LNM-negative patients likely limited discriminative power, while the pronounced class imbalance (substantially more LMN-negative than LNM-positive cases) may have biased the model toward the majority class, reducing sensitivity for LNM. Although strategies to mitigate class imbalance were explored, these did not substantially improve model performance or sensitivity. Moreover, recent methodological studies indicate that correcting for class imbalance may introduce miscalibration through risk overestimation that is not always correctable, suggesting that such corrections are not necessarily required and may even be detrimental for clinical prediction models (29).

In addition, the training and test sets were analyzed using different multiplex platforms, which may introduce inter-platform variability and influence model validation. However, as no major discrepancy in performance was observed between the training and test sets, this suggests that platform-related differences did not substantially affect overall classification performance, while also providing an initial indication of model robustness across platforms.

Even though categorization of sICs is commonly seen in the literature, and many regression models for predicting outcome are based on categorized sICs, we analyzed them as continuous variables. This decision was based on the absence of biologically or clinically validated cut-offs for the investigated sICs. Arbitrary categorization lacks biological plausibility and leads to loss of information and reduced statistical power (30). Maintaining continuous variables therefore improved the reliability of statistical inference.

Our results differ in some respects from previous studies that have evaluated the potential of using sICs as biomarkers in other solid cancers. In renal cell carcinoma, Wang et al. found elevated sTIM-3 and sLAG-3 levels correlated with advanced disease, sPD-L2 with tumor recurrence, and sTIM-3 and sBTLA with decreased survival (20). Furthermore, Bian et al. found that increased levels of sBTLA, sPD-1, and sPD-L1 were significantly associated with poor prognosis in patients with pancreatic adenocarcinoma and Li et al. demonstrated a correlation between elevated levels of sBTLA and sTIM-3 and disease progression in osteosarcoma patients (18, 19). The observed discrepancies could be explained by methodological variations, such as differences in the assays or detection techniques used to measure sICs. It may also suggest that sICs may reflect tumor aggressiveness in a cancer-specific manner.

In PeCa, immune checkpoint expression has mainly been studied in tumor tissue. PD-L1 expression ranges from 32% to 75% and is linked to adverse features and poor survival (23–26). Davidsson et al. and Ottenhof et al. identified PD-L1 positivity as a predictor of poor survival, with Ottenhof also linking its pattern to LNM (23, 24). These findings were confirmed, showing high PD-L1, TIGIT, and CD155 expression in LNM-positive PeCa, as well as better survival in PD-L1-negative patients (25, 26).

The discrepancy between blood- and tissue-based IC assessments may reflect both biological and methodological differences. Soluble proteins can be affected by systemic inflammation and individual immune variability, and their circulating levels may not reflect the immune interactions within the tumor microenvironment. In addition, the lack of detectable sPD-L1 in this study is most likely due to methodological constraints rather than true biological scarcity. Previous studies using the ProcartaPlex panel have similarly reported low or undetectable sPD-L1 levels (31, 32). In our prior work, sPD-L1 was readily detected in urinary bladder cancer and renal cell carcinoma samples using ELISA, whereas detection was markedly reduced when using the ProcartaPlex panel on the same material (unpublished data). This suggests that differences in assay sensitivity and limit of detection (lower limit of quantification 5.62 pg/mL), rather than limited biological availability, likely explain the low detectability observed here.

Tissue-based assessments, by contrast, allow direct quantification of IC expression in tumor cells and tumor-infiltrating lymphocytes, providing localized insight into immune evasion mechanisms relevant to metastatic progression and greater analytical specificity and sensitivity.

When comparing PeCa patients and cancer-free controls, we found that four inhibitory sICs (IDO, TIM-3, CD80, and CTLA-4) were significantly elevated in the PeCa group. This supports the concept that tumor development is associated with a shift towards an immunosuppressive environment, as inhibitory ICs (such as IDO, TIM-3, CD80, and CTLA-4) are key regulators that limit immune activation. The increased levels observed in patients with PeCa indicate that penile tumors may exploit these immune regulatory pathways to evade immune surveillance. Furthermore, the concurrent elevation of multiple ICs suggests that penile tumors may use multiple pathways to suppress antitumor immune responses and that systemic immunomodulation may occur early in PeCa development as it appeared independent of tumor burden. Although based on a limited sample size, and thus requiring cautious interpretation, our findings are consistent with previous reports across various malignancies, where elevated inhibitory sIC levels are thought to reflect chronic tumor-induced immunosuppression (17–20).

Strengths of this study include its relatively large cohort and inclusion of a cancer-free control group, enhancing comprehensive analyses. Furthermore, we used a standardized multiplex assay for multianalyte profiling, and the study design incorporated both a training set and an independent test set of PeCa patients, which strengthens the reliability of our findings. Limitations include the small number of LNM-positive cases, reducing statistical power.

In conclusion, the evaluated panel of 13 sICs measured in plasma demonstrated limited utility for predicting LNM in PeCa, due to low sensitivity, modest accuracy, and weak discriminative power between LNM-positive and LNM-negative patients. This study nevertheless represents the first systemic evaluation of a broad sIC panel in PeCa. Four inhibitory sICs were significantly elevated in PeCa patients compared to cancer-free controls, suggesting systemic immunosuppression associated with tumor presence, consistent with findings in other malignancies. Future studies assessing larger cohorts and integrating both tissue-based and soluble ICs are warranted to clarify their prognostic significance in PeCa.