Prognostic Implications of Nitrogen Metabolism-Related Gene SMS in Hepatocellular Carcinoma: An Integrated Bioinformatics and Experimental Research

Abstract

Objective

To identify nitrogen metabolism-related genes associated with hepatocellular carcinoma (HCC) prognosis, focusing on spermine synthase (SMS), and to evaluate its prognostic and risk-stratification value.

Methods

Transcriptomic and clinical data from TCGA-LIHC and GSE14520 and single-cell RNA-sequencing data from GSE242889 were analyzed. Prognostic genes were screened using univariate Cox, LASSO-Cox, and random survival forest analyses. The prioritized candidate, SMS, was subsequently evaluated by immunohistochemistry in an independent cohort of 144 patients with HCC after curative resection. Associations with clinicopathological features, recurrence-free survival (RFS), and overall survival (OS) were assessed. Prognostic models were evaluated using time-dependent ROC analysis, and risk stratification was examined by combining SMS expression with microvascular invasion (MVI) status and Barcelona Clinic Liver Cancer (BCLC) stage.

Results

Integrated bioinformatics and machine-learning analyses prioritized SMS, a key enzyme in polyamine biosynthesis, as the top-ranked nitrogen metabolism-related prognostic candidate, and its prognostic association was validated in GSE14520. In the clinical cohort, high SMS expression was significantly associated with multiple tumors, MVI positivity, and BCLC stage B/C. High SMS expression was also associated with shorter RFS and OS and remained independently associated with unfavorable RFS and OS in multivariable Cox analyses. The combined clinical+SMS model showed better prognostic discrimination than the SMS-only model, whereas its improvement over the clinical model was not statistically significant. Combined stratification showed particularly unfavorable outcomes in patients with high SMS expression and MVI positivity, while SMS further differentiated prognosis within the BCLC A and BCLC B/C subgroups.

Conclusion

This study provides independent clinical validation of the prognostic relevance of SMS in HCC. High SMS expression was associated with aggressive clinicopathological features and unfavorable postoperative outcomes. Integrating SMS expression with MVI status and BCLC stage may provide complementary information for postoperative risk stratification.

Graphical Abstract

Introduction

Liver cancer ranks among the most prevalent malignancies worldwide; the annual number of incident cases is projected to nearly double by 2050, increasing from 870,000 in 2022 to 1.52 million. Hepatocellular carcinoma (HCC) represents the predominant histological subtype, accounting for approximately 80% of all primary liver cancers.¹ Surgical resection remains the most effective treatment for liver cancer; however, because early-stage disease typically lacks characteristic clinical manifestations, eligibility for surgery has often been lost by the time diagnosis is established. Although chemotherapy, radiotherapy, and liver transplantation may be used to prolong survival, outcomes remain unfavorable because of high recurrence rates, intrahepatic dissemination, and other risk factors.² Accurate assessment of postoperative recurrence and mortality risk is therefore of considerable clinical relevance for developing individualized follow-up strategies, identifying high-risk patients, and guiding adjuvant therapy. Owing to the marked heterogeneity of liver cancer, conventional prognostic indicators, including TNM stage, age, and sex, have substantial limitations. Tumor markers such as AFP are often dependent on tumor burden and have limited utility for early prediction.³ A biomarker capable of accurately predicting prognosis in HCC patients is urgently required.

Recent studies have further highlighted the molecular and therapeutic heterogeneity of HCC, including pathway-based prognostic signatures and immune-microenvironment characterization, as well as evolving immunotherapeutic strategies.^4,5 Other studies have identified specific molecular regulatory axes, such as miR-3130-5p/FDX1, and applied integrative computational and experimental approaches to identify candidate biomarkers and therapeutic targets in HCC.^6,7 Integrative molecular studies in other solid tumors have likewise demonstrated that non-coding RNA regulatory networks may influence malignant cell behavior and epithelial–immune interactions.⁸

Nitrogen metabolism is integral to the maintenance of nucleotide synthesis, amino acid metabolism, polyamine synthesis, redox homeostasis, and related biological processes. Marked dysregulation of nitrogen metabolism has been reported across multiple tumor cell types, accompanied by increased consumption of nitrogen sources by tumor cells. Accordingly, investigation of nitrogen metabolism-related genes (NMRGs) is of considerable value for accurate prognostic prediction in HCC. Systematic screening of nitrogen metabolism-related prognostic genes in HCC may identify candidate biomarkers and provide additional molecular information for prognostic assessment.

Spermine synthase is a key enzyme in polyamine biosynthesis that catalyzes the conversion of spermidine to spermine. Previous studies have demonstrated that SMS is upregulated in HCC and that elevated SMS expression is associated with unfavorable prognosis and an immunosuppressive tumor microenvironment. In particular, Xiang et al reported associations between high SMS expression, poor survival, altered immune-cell infiltration, and immune-checkpoint expression in HCC.⁹ More recently, Sun et al provided mechanistic evidence that SMS-derived spermine promotes M2 polarization of tumor-associated macrophages through the PI3K–Akt–mTOR–S6K pathway and consequently impairs CD8⁺ T-cell antitumor activity.¹⁰ These studies have established important prognostic and immunomodulatory roles of SMS in HCC. However, the clinical value of SMS protein expression in an independent postoperative cohort, particularly its potential to provide complementary prognostic information when integrated with established clinicopathological indicators such as MVI status and BCLC stage, remains insufficiently characterized.

Microvascular invasion (MVI) is an important pathological feature that reflects the aggressive biological behavior of HCC and is closely associated with early postoperative recurrence, tumor metastasis, and unfavorable survival outcomes. The Barcelona Clinic Liver Cancer (BCLC) staging system integrates tumor burden, liver function, and patient performance status and is widely used for clinical treatment decision-making and prognostic assessment in HCC. However, both MVI and the BCLC system have inherent limitations as prognostic indicators for HCC, highlighting the potential value of additional molecular markers that may complement established clinicopathological indicators in prognostic risk stratification.

In this study, bioinformatics analysis and machine learning were integrated to prioritize SMS as a nitrogen metabolism-related prognostic candidate in HCC, and its protein expression was further validated by IHC in an independent clinical cohort. Furthermore, prognostic models and risk-stratification analyses were performed by integrating SMS expression with clinicopathological characteristics, including MVI status and BCLC stage, to evaluate its complementary value for postoperative prognostic assessment. This integrated strategy combines cellular-level characterization, large-scale transcriptomic screening, machine-learning-based prioritization, external validation, and clinical tissue verification. The complementary use of these approaches was intended to improve the robustness and clinical relevance of candidate biomarker evaluation.

Methods

Data Collection

The overall study flowchart is presented in Graphical Abstract. The TCGA-LIHC dataset, which included 374 HCC samples, 50 normal samples, and corresponding clinical information, was downloaded from the Xena database (xena.ucsc.edu/). HCC microarray data (GSE14520), comprising 250 tumor samples and associated clinical information, and a single-cell sequencing dataset (GSE242889), containing 5 tumor tissue samples and 5 normal tissue samples, were obtained from the GEO database. After exclusion of cases with incomplete survival information, 370 TCGA-LIHC cases were available for survival analyses. Multivariable Cox regression was performed using 342 cases with complete required clinicopathological data; missing values were not imputed. For GSE14520, the primary external validation analysis was restricted to the GPL3921 subset to maintain a platform-homogeneous validation cohort. After exclusion of cases with incomplete survival or required clinical information, 221 patients were included in the final analysis. In addition, 70 NMRGs were obtained from the nitrogen metabolism-related gene set reported by Zhang, Z et al.¹¹

Single-Cell Data Processing

The single-cell sequencing dataset was processed using the R package Seurat (Version 5.3.0). Quality control criteria were defined as follows: each feature was required to be detected in at least three cells, and at least 200 features were required for each cell. In addition, the number of genes detected per cell was restricted to 300–6000, and outlier cells were excluded according to the proportions of mitochondrial and ribosomal genes (percent.mt ≤ 20, percent.rb ≤ 20).

Batch effects were corrected using the Harmony algorithm with group.by.vars = “Type”; parameters not explicitly specified were left at their default settings. The top 2000 highly variable features were identified using the FindVariableFeatures function. PCA was performed using the RunPCA function, followed by Uniform Manifold Approximation and Projection for dimensionality reduction and clustering. Each cell cluster was manually annotated. Cluster-specific cell markers were as follows: endothelial cells (BMPER, GPR182, CLEC14A), fibroblasts (MYOCD, SMOC2, CNN1), B cells (IGHGP, IGHG3, IGLC2), NK/T cells (CXCR6, IFNG, XCL1, CCNA2, DLGAP5, MKI67), epithelial cells (GSTA2, GSTA1, HP), myeloid cells (CD300E, BST1, SLC11A1), and other cells. Five methods, including AddModuleScore, ssGSEA, UCell, AUCell, and singscore, were independently applied to score the NMRG set in each cell. For AUCell, aucMaxRank was set to 10% of the ranked genes. UCell and singscore were implemented using irGSEA.score based on the RNA data slot, with a fixed seed of 123. ssGSEA was implemented using the GSVA package with a Gaussian kernel and abs.ranking = TRUE, whereas AddModuleScore was applied using its default settings. The score generated by each method was first standardized using Z-score transformation and subsequently rescaled to 0–1 using min–max normalization. The five normalized scores were then summed with equal weight to generate the final integrated NM score. In addition, CopyKAT was applied to epithelial cells using its default settings to distinguish malignant from non-malignant epithelial cell populations based on inferred copy-number alterations.

Identification of Nitrogen Metabolism-Related Prognostic Genes

Relative expression levels of the 70 NMRGs were extracted from the TCGA-LIHC dataset. Univariate Cox regression analysis was first performed to identify prognosis-associated NMRGs. The significant genes were subsequently subjected to LASSO-Cox regression. Ten-fold cross-validation was performed using cv.glmnet, and the value of λ corresponding to the minimum cross-validation error (lambda.min) was selected. Genes with non-zero coefficients at lambda.min were retained for subsequent analysis. The LASSO-selected genes were then evaluated using random survival forest analysis with the randomForestSRC package (version 3.4.3). The main model parameters were set as ntree = 1500, mtry = 2, nodesize = 10, and splitrule = “logrank”. Variable importance was evaluated using the built-in anti-split VIMP procedure, and genes were ranked according to their VIMP values. After SMS was prioritized for further analysis, patients were stratified into high- and low-SMS-expression groups according to the median SMS expression level. Kaplan–Meier analyses were subsequently performed in the TCGA-LIHC and GSE14520 cohorts.

Construction of the Clinical Prediction Model

Expression levels of key prognostic genes, along with sex, age, histological grade, and pathological stage for each sample, were extracted from the TCGA-LIHC dataset. Categorical variables were processed as follows: sex was incorporated as a categorical variable, with the female group used as the reference. Histological grade was reclassified as follows: Grade 1 was defined as well differentiated, Grade 2 as moderately differentiated, and Grades 3 and 4 were combined into a poorly differentiated/undifferentiated category; the well-differentiated group was used as the reference. Pathological stage was also reclassified: Stages I and II were combined into an early/intermediate-stage group, whereas Stages III and IV were combined into an advanced-stage group; the Stage I/II group was used as the reference. These categorical variables were incorporated into the Cox regression model, with their respective reference groups used as baselines.

Immune Microenvironment

Immune infiltration analysis of the TCGA-LIHC dataset was performed using the TIMER algorithm through the “Immunedeconv” R package. Gene sets associated with immune functions were scored using the “GSVA” R package. Comparative analysis identified significant differences (P < 0.05) in immune cell infiltration and immune functions between the high- and low-SMS-expression groups, and the results were visualized. In addition, immune checkpoint gene expression levels were compared between the high- and low-SMS-expression groups.

Clinical Sample Collection

Tissue sections from 144 HCC patients treated at Baoding No. 1 Central Hospital were collected for this study. Inclusion criteria: 1) curative liver resection performed as the initial antitumor treatment; 2) radical resection performed with postoperative pathological confirmation of HCC; and 3) availability of complete clinical and follow-up data. Exclusion criteria: 1) prior anti-tumor treatment, including surgical resection, TACE, local ablation, targeted immunotherapy, or liver transplantation; 2) incomplete clinical or follow-up data; and 3) concurrent malignancies. This study was approved by the hospital Ethics Committee (Ethics Approval No.[2025] 215). The requirement for informed consent was waived, and all procedures followed the ethical principles of the Declaration of Helsinki. All eligible patients who met the predefined inclusion and exclusion criteria and had available tumor tissue, clinical information, and follow-up data during the study period were included. Because this was a retrospective exploratory cohort based on all eligible available cases, no formal a priori sample-size or power calculation was performed.

Immunohistochemistry

IHC staining was performed on tissue sections from 144 patients with HCC. The primary antibody against SMS (E-AB-18693) was purchased from Elabscience Biotechnology Co., Ltd. (Wuhan, China). Paraffin-embedded tissue samples were sectioned at 4 μm and baked at 68°C for 1 hour. The sections were deparaffinized and rehydrated through graded ethanol. Antigen retrieval was conducted using citrate buffer, endogenous peroxidase activity was blocked with 3% hydrogen peroxide, and nonspecific binding was inhibited with 5% BSA. The sections were incubated with the primary antibody overnight at 4°C. Primary antibody detection was performed using an HRP-labeled secondary antibody, with DAB used as the chromogenic substrate. The results were independently evaluated by two experienced pathologists; discrepant assessments were adjudicated by a third pathologist. The positive-cell area was scored as follows: ≤5% stained cells = 0; 6–25% stained cells = 1; 26–50% stained cells = 2; 51–75% stained cells = 3; and 76–100% stained cells = 4. Staining intensity was scored as follows: - = 0; + = 1; ++ = 2; and +++ = 3. The IHC score was calculated by multiplying the staining intensity score (0–3) by the proportion score of stained cells (0–4). Cumulative scores of 0–4 were defined as low expression, whereas scores of 6–12 were defined as high expression.

Collection of Clinicopathological Data and Follow-Up

The collected clinicopathological data included age, sex, serum AFP level, HBV infection status, cirrhosis, tumor size, tumor number, pathological differentiation grade, MVI status, and BCLC stage. MVI status was determined by two independent pathologists based on postoperative pathological sections. BCLC stage was assessed according to international guidelines. Postoperative follow-up was conducted through outpatient visits, telephone interviews, and the hospital electronic medical record system. The follow-up cutoff date was May 2018. Follow-up data included survival status, recurrence, and cause of death. Recurrence-free survival (RFS) was defined as the interval from the date of surgery to tumor recurrence, death from any cause, or the last follow-up. Overall survival (OS) was defined as the interval from the date of surgery to death from any cause or the last follow-up.

Prognostic Analysis

Survival curves were generated using the Kaplan-Meier method to compare RFS and OS between the high- and low-SMS-expression groups. In addition, multiple clinical variables were incorporated to determine whether SMS expression functioned as an independent prognostic factor. Univariate Cox regression analyses were first performed to evaluate the associations of SMS expression and clinicopathological variables with RFS and OS. Variables with P < 0.05 in univariate Cox analysis were subsequently entered into the multivariable Cox regression models, P-value < 0.05 was considered statistically significant, and hazard ratios (HRs) with 95% confidence intervals (CIs) were reported. To improve numerical stability, the multivariable Cox regression models were fitted with an L2 penalization term (penalizer = 0.1). The proportional hazards assumption for the multivariable Cox regression models was assessed using Schoenfeld residual-based tests with rank-transformed time for each covariate included in the final models. A P value > 0.05 was considered to indicate no statistically significant evidence of violation of the proportional hazards assumption. To assess prognostic discrimination in the clinical cohort, three models were constructed: an SMS-only model based on SMS expression, a clinical model incorporating clinicopathological variables that were statistically significant in univariate Cox analysis, and a combined model integrating SMS expression with the variables included in the clinical model. Time-dependent ROC curves were generated to evaluate 1-, 3-, and 5-year RFS and OS discrimination, and AUCs were compared between models using DeLong tests.

Risk Stratification

To further assess the prognostic stratification value of SMS expression combined with MVI status and BCLC stage in HCC patients, BCLC stage was dichotomized as BCLC A vs BCLC B/C for the present exploratory analysis. This grouping was selected with reference to previous surgical studies evaluating postoperative outcomes across BCLC stages and to the survival pattern observed in the current cohort. It was intended to contrast patients with BCLC A disease with those in BCLC B/C, who generally have greater tumor burden and less favorable postoperative prognosis. Because no BCLC stage 0 cases were included in the 144-patient cohort, BCLC A constituted the lower-stage subgroup in this analysis. This analysis-specific dichotomization was used solely for risk-stratification purposes and was not intended to redefine the conventional BCLC classification, in which stage B is considered intermediate and stage C advanced.^12–14 The following risk stratification analyses were performed: 1) patients were divided into four groups according to SMS expression level and MVI status (positive vs negative); and 2) patients were divided into four groups according to SMS expression level and BCLC stage (BCLC A vs BCLC B/C). Kaplan–Meier curves were generated for RFS and OS across the stratified groups, and survival differences between groups were compared using the Log-rank test. Pairwise comparisons were performed using the Log-rank test with Bonferroni correction.

Results

Aberrant Nitrogen Metabolism in HCC Tissues

After batch-effect correction (Figure 1A), dim = 18 and resolution = 0.3 were selected for subsequent analyses. Seven cell types were identified in Figure 1B: B cells, epithelial cells, macrophages, NK/T cells, endothelial cells, fibroblasts, and other cells. As shown in Figure 1C and D, nitrogen metabolism scores were significantly increased in epithelial cells from tumor samples, indicating NMRG dysregulation in HCC. The CopyKat algorithm was applied to identify epithelial cell subpopulations. As shown in Figure 1E and F, epithelial cells in the tumor group were classified into malignant and normal subpopulations, and nitrogen metabolism scores were significantly higher in malignant cells than in normal cells.

Figure 1.

Single-cell sequencing analysis: (A) Batch-effect correction; (B) Cell subpopulation annotation; (C) NM scores across cell subpopulations; (D) Comparison of NM scores between normal and HCC tissues; (E) Identification of malignant epithelial cells using CopyKAT; (F) Comparison of NM scores between normal and malignant epithelial cells.

Nitrogen Metabolism-Related Prognostic Genes

Univariate Cox analysis was performed for 70 NMRGs (Figure 2A), identifying 32 prognosis-related genes, including 17 risk factors and 15 protective factors. Seven key genes were identified using LASSO-Cox regression (Figure 2B and C). Random survival forest analysis further ranked these genes according to variable importance (VIMP): SMS (0.181), GCLM (0.096), HSP90AA1 (0.066), CPS1 (0.041), GOT2 (0.030), ALDH5A1 (0.020), and ASNSL1 (0.012). SMS showed the highest VIMP value and was therefore prioritized for further investigation (Figure 2D). As shown in Figure 3A and B, patients were stratified into high- and low-SMS-expression groups according to the median SMS expression level; Kaplan-Meier curves demonstrated significantly shorter survival in the high-SMS-expression group. As shown in Figure 3C and D, the AUC values for the 1-, 3-, and 5-year ROC curves in the TCGA-LIHC cohort were 0.76, 0.69, and 0.70, respectively, indicating that SMS exhibited prognostic predictive capacity, with moderate discriminatory performance for 1-year survival. In the GSE14520 cohort, the AUC values for the 1-, 3-, and 5-year ROC curves were 0.62, 0.60, and 0.55, respectively. As shown in Figure 3E and F, independent prognostic analysis indicated that SMS expression and tumor stage (I/II vs III/IV) remained significant in both univariate and multivariate Cox analyses, suggesting that SMS expression and tumor stage remained significantly associated with prognosis in both univariate and multivariable Cox analyses, supporting the independent prognostic relevance of SMS in the TCGA-LIHC cohort. Although SMS expression may be associated with HCC prognosis, its stability and generalizability as a standalone prognostic marker remain limited; therefore, integration with clinicopathological features or other molecular indicators is warranted to construct a comprehensive predictive model with improved accuracy.

Figure 2.

Screening of key prognostic genes: (A) Univariate Cox regression analysis; (B–C) LASSO-Cox regression analysis; (D) Variable importance ranking based on random survival forest.

Figure 3.

Prognostic value of SMS expression in HCC patients. (A and B) Kaplan-Meier survival curves for high- and low-SMS expression groups in the TCGA-LIHC and GSE14520 datasets. (C and D) Time-dependent ROC curves for predicting patient prognosis based on SMS expression in the TCGA-LIHC and GSE14520 datasets. (E) Univariate Cox regression analysis of SMS expression and clinical characteristics in TCGA-LIHC. (F) Multivariate Cox regression analysis of SMS expression and clinical characteristics in TCGA-LIHC.

Analysis of Immune Infiltration and Immune Functions

As shown in Figure 4A, the infiltration proportions of regulatory T cells, T follicular helper cells, M0 macrophages, and resting dendritic cells were significantly increased in the SMS high-expression group. As shown in Figure 4B, significantly higher enrichment scores for immune functions, including APC co-inhibition, HLA, MHC class I, and Type II IFN response, were observed in the SMS high-expression group, whereas the Type I IFN response score was significantly decreased. SMS expression levels appeared to be closely associated with immune-related functions in liver cancer, including antigen presentation, interferon responses, and APC-mediated immune regulation.

Figure 4.

Immune landscape analysis. (A) Comparison of immune cell infiltration proportions between SMS high- and low-expression groups. (B) Comparison of immune-related functional scores between SMS high- and low-expression groups. Asterisks indicate the statistical significance of comparisons between the high- and low-SMS expression groups: ns, not significant; *P < 0.05; **P < 0.01; ***P < 0.001.

Baseline Characteristics of the Clinical Cohort

A total of 144 HCC patients who underwent curative liver resection were included, comprising 114 males (79.2%) and 30 females (20.8%). Patient age ranged from 29 to 83 years (median: 57 years). The median follow-up duration was 43.50 months (range: 1–114 months). IHC staining results for SMS in HCC tissues are shown in Figure 5A and B. According to SMS protein expression levels, patients were classified into an SMS low-expression group (n = 54, 37.5%) and an SMS high-expression group (n = 90, 62.5%). As shown in Table 1, the high-SMS-expression group had a significantly larger tumor size and higher proportions of multiple tumors (≥2), MVI positivity, and BCLC stage B/C disease than the low-SMS-expression group. These results indicate that high SMS expression is associated with more aggressive clinicopathological features.

Figure 5.

Representative results of SMS gene immunohistochemical staining: (A and B) High SMS gene expression; (C and D) Low SMS gene expression.

Table 1.

Baseline Characteristics of the SMS High- and Low-Expression Groups

Variable	Low (N=54)	High (N=90)	Statistic	P
BCLC, n (%)			χ2=12.701	P<0.05
A	32 (59.3%)	25 (27.8%)
B/C	22 (40.7%)	65 (72.2%)
Gender, n (%)			χ2=0.281	P>0.05
Male	41 (75.9%)	73 (81.1%)
Female	13 (24.1%)	17 (18.9%)
Age M[Q1, Q3] (years)	57.00 [48.00, 60.75]	57.50 [51.25, 65.00]	Z=−1.048	P>0.05
Tumor number M [Q1, Q3]	1.00 [1.00, 1.00]	1.00[1.00, 2.00]	Z=−3.355	P<0.05
Tumor size M [Q1, Q3] (cm)	3.75 [3.00, 4.95]	5.50 [3.50, 9.00]	Z=−3.728	P<0.05
Differentiation, n (%)			χ2=1.716	P>0.05
Well-differentiated	14 (25.9%)	24 (26.7%)
Moderately-differentiated	35 (64.8%)	51 (56.7%)
Poorly-differentiated	5 (9.3%)	15 (16.7%)
MVI, n (%)			χ2=5.438	P<0.05
NO	32 (59.3%)	34 (37.8%)
YES	22 (40.7%)	56 (62.2%)
Vascular tumor thrombus, n (%)			χ2=0.031	P>0.05
NO	22 (40.7%)	34 (37.8%)
YES	32 (59.3%)	56 (62.2%)
Satellite lesion, n (%)			χ2=0.151	P>0.05
NO	28 (51.9%)	51 (56.7%)
YES	26 (48.1%)	39 (43.3%)
Cirrhosis, n (%)			χ2=0.314	P>0.05
NO	23 (42.6%)	44 (48.9%)
YES	31 (57.4%)	46 (51.1%)
PT M [Q1, Q3] (s)	11.20 [10.72, 12.00]	11.10 [10.60, 11.67]	Z=0.978	P>0.05
AFP M [Q1, Q3] (ng/mL)	13.93 [4.46, 327.32]	30.36 [5.26, 595.95]	Z=−0.856	P>0.05
CEA M [Q1, Q3] (ng/mL)	2.50 [1.52, 3.71]	2.67 [1.77, 3.53]	Z=−0.675	P>0.05
CA199 M [Q1, Q3] (U/mL)	14.90 [7.95, 30.48]	21.39 [9.67, 32.92]	Z=−1.267	P>0.05
ALT M [Q1, Q3] (U/L)	32.00 [19.25, 49.00]	30.00 [20.00, 45.75]	Z=0.312	P>0.05
AST M [Q1, Q3] (U/L)	30.00 [24.00, 42.75]	34.00 [24.25, 46.75]	Z=−0.910	P>0.05
ALB M [Q1, Q3] (g/L)	43.55 [40.67, 46.55]	42.15 [39.02, 45.10]	Z=1.188	P>0.05
TB M [Q1, Q3] (μmol/L)	16.40 [12.72, 23.38]	16.60 [13.60, 21.38]	Z=−0.192	P>0.05

SMS Expression and Prognosis in HCC Patients

Kaplan-Meier survival curves (Figure 6A and B) demonstrated significantly lower RFS and OS in patients with high SMS expression than in those with low SMS expression. High SMS expression was therefore closely associated with an increased risk of postoperative recurrence and reduced survival in HCC patients, suggesting its potential utility as a prognostic biomarker. As shown in Figure 6C and E, univariate and multivariate Cox analyses identified SMS expression and AFP remained significantly associated with RFS. Similarly, Figure 6D and F showed that SMS expression, AFP, and CA19-9 remained significantly associated with OS. The proportional hazards assumption was assessed using Schoenfeld residual-based tests. No statistically significant violations were detected for any covariate included in either the RFS or OS multivariable Cox model (all P > 0.05), supporting the proportional hazards assumption for both models. Collectively, high SMS expression remained significantly associated with poor postoperative RFS and OS after adjustment for the clinicopathological variables included in the multivariable Cox models. To assess the prognostic value of integrating SMS expression with clinicopathological features, three models were constructed: an SMS-only model based solely on SMS expression, a clinical model incorporating clinicopathological variables that reached statistical significance in univariate Cox analysis, and a combined model integrating SMS expression with the aforementioned clinicopathological variables. Time-dependent ROC curve analysis (Figure 7A and B) showed that, for RFS prediction, the combined model yielded AUC values of 0.801, 0.734, and 0.789 at 1, 3, and 5 years, respectively, which were numerically higher than those of the SMS-only model (0.628, 0.670, and 0.674) and the clinical model (0.757, 0.692, and 0.710). DeLong tests identified statistically significant differences between the combined and SMS-only models at all time points (1-year: P < 0.05; 3-year: P < 0.05; 5-year: P < 0.05), whereas no statistically significant differences were observed between the combined and clinical models (1-year: P = 0.318; 3-year: P = 0.190; 5-year: P = 0.533).For OS prediction, the combined model achieved AUC values of 0.854, 0.732, and 0.821 at 1, 3, and 5 years, respectively, which were numerically higher than those of the SMS-only model (0.629, 0.658, and 0.666) and the clinical model (0.848, 0.686, and 0.746). DeLong tests demonstrated statistically significant differences between the combined and SMS-only models at all three time points (1-year: P < 0.05; 3-year: P < 0.05; 5-year: P < 0.05), whereas the differences between the combined and clinical models were not statistically significant (1-year: P = 0.740; 3-year: P = 0.157; 5-year: P = 0.250).Overall, the combined model showed significantly better discrimination than the SMS-only model, whereas its time-dependent AUCs were not significantly different from those of the clinical model. Therefore, although SMS remained independently associated with RFS and OS in multivariable Cox analyses, the present ROC analyses did not demonstrate statistically significant incremental discriminatory value of SMS beyond the established clinicopathological variables included in the clinical model.

Figure 6.

Prediction of RFS and OS in HCC patients based on SMS expression in the local cohort: (A and B) Kaplan-Meier survival curves for RFS and OS in patients with high and low SMS expression; (C) Univariate Cox regression analyses for RFS; (D) Univariate Cox regression analysis for OS; (E) Multivariate Cox regression analysis for RFS; (F) Multivariate Cox regression analysis for OS. Asterisks indicate the statistical significance of comparisons between the high- and low-SMS expression groups: ns, not significant; *P < 0.05; **P < 0.01; ***P < 0.001.

Figure 7.

Predictive performance of the model combining the SMS and clinicopathological features for the prognosis of HCC patients: (A) Time-dependent ROC curves of the three models for predicting 1-, 3-, and 5-year recurrence-free survival (RFS) in HCC patients; (B) Time-dependent ROC curves of the three models for predicting 1-, 3-, and 5-year overall survival (OS) in HCC patients.

Risk Stratification Analysis Combining SMS Expression with MVI and BCLC Staging

In the local cohort, patient stratification revealed significant differences in RFS among the four groups (Figure 8A and B). The low SMS expression/MVI-negative group had the most favorable RFS, whereas the high SMS expression/MVI-positive group had the poorest RFS. Similarly, significant differences in OS were observed among the four groups; the low SMS expression/MVI-negative group had the most favorable OS, whereas the high SMS expression/MVI-positive group had the poorest OS. To assess the prognostic stratification value of combining SMS expression with BCLC stage, patients were classified into four groups according to these parameters (Figure 9A and B). Significant differences in RFS were observed among the groups: the low SMS expression/BCLC A group had the most favorable RFS, whereas the high SMS expression/BCLC B/C group had the poorest RFS. Likewise, significant differences in OS were identified, with the low SMS expression/BCLC A group showing the most favorable OS and the high SMS expression/BCLC B/C group showing the poorest OS.

Figure 8.

Prognostic risk stratification value of combined SMS expression and MVI status in HCC patients: (A) Comparison of recurrence-free survival (RFS) and overall survival (OS) among patient subgroups stratified by combined SMS expression levels and MVI status; (B) Heatmap of log-rank tests for RFS and OS across different SMS expression and MVI status subgroups. Asterisks indicate the statistical significance of comparisons between the high- and low-SMS expression groups: ns, not significant; *P < 0.05; **P < 0.01; ***P < 0.001.

Figure 9.

Prognostic risk stratification based on combined SMS expression and BCLC stage in HCC patients: (A) Comparison of RFS and OS across patient subgroups stratified by combined SMS expression level and BCLC stage; (B) Differences in RFS and OS between the SMS high- and low-expression groups within the BCLC stage a and BCLC stage B/C cohorts. Asterisks indicate the statistical significance of comparisons between the high- and low-SMS expression groups: ns, not significant; *P < 0.05; **P < 0.01; ***P < 0.001.

Discussion

Metabolic reprogramming is a defining biological feature of malignant tumors. Compared with normal cells, tumor cells exhibit markedly increased uptake and utilization of nitrogen sources to satisfy the demands of nucleotide synthesis, amino acid metabolism, and macromolecular biosynthesis required for rapid proliferation.^15,16 Analysis of single-cell sequencing data in this study showed that NMRG expression scores were significantly higher in malignant HCC epithelial cells than in normal epithelial cells; moreover, NMRG levels were significantly elevated in tumor samples compared with controls. These results provide cellular-level evidence of nitrogen metabolism dysregulation in HCC, consistent with the concept of nitrogen metabolism reprogramming in tumor cells.

A key downstream pathway of nitrogen metabolism is the polyamine biosynthetic pathway. Polyamines, including putrescine, spermidine, and spermine, are nitrogen-containing organic cations that have central roles in cellular growth. Dysregulated polyamine metabolism and altered polyamine demand are common in many cancers and are primarily driven by upregulated polyamine synthases, resulting in increased polyamine levels required for malignant transformation and tumor progression. In HCC, several enzymes involved in polyamine metabolism have been associated with tumor progression and prognosis. ODC1, which catalyzes the conversion of ornithine to putrescine, has been reported to be upregulated in HCC and to promote tumor-cell proliferation, migration, and invasion through mechanisms involving the AKT/GSK3β/β-catenin pathway and the acidotic tumor microenvironment.¹⁷ Elevated expression of another key polyamine metabolic enzyme, AMD1 (S-adenosylmethionine decarboxylase), is also associated with poor prognosis in HCC.¹⁸ SRM, which catalyzes spermidine synthesis upstream of SMS, has likewise been linked to HCC progression and unfavorable prognosis; recent experimental evidence further suggests that SRM can participate in tumor-macrophage metabolic crosstalk and promote a tumor-supportive macrophage phenotype.¹⁹ Compared with these upstream enzymes, SMS catalyzes the conversion of spermidine to spermine and therefore occupies a distinct downstream position in the polyamine biosynthetic pathway. Taken together, these findings indicate that the prognostic and functional relevance of polyamine metabolism in HCC is distributed across multiple enzymatic components rather than being unique to SMS. This broader pathway context also highlights a limitation of using a single metabolic enzyme as a prognostic marker, because one marker may not fully capture the biological heterogeneity of polyamine metabolism. Accordingly, SMS may be more appropriately considered as a complementary molecular marker within a broader metabolic network rather than as a standalone determinant of HCC prognosis. Previous research by Guo, Ye, and colleagues demonstrated that SMS was overexpressed in colorectal cancer (CRC) cells and promoted their growth and proliferation; targeted disruption of SMS resulted in spermidine accumulation and induced apoptosis in CRC cells.²⁰ These findings support the broader oncogenic relevance of dysregulated SMS-mediated polyamine metabolism.

Previous studies have directly linked SMS to unfavorable prognosis and immune dysregulation in HCC, while mechanistic evidence has further implicated SMS-derived spermine in macrophage polarization and tumor immune escape. Against this background, the present study extends previous findings through protein-level evaluation of SMS in an independent postoperative HCC cohort.

Consistent with these previous observations, high SMS expression in the present study was associated with unfavorable RFS and OS and remained significantly associated with postoperative outcomes in multivariable Cox analyses. The present study provides additional clinical evidence through protein-level evaluation of SMS by IHC in an independent cohort of 144 patients with HCC who underwent curative resection. High SMS expression was also significantly associated with aggressive clinicopathological features, including larger tumor size, multiple tumors, MVI positivity, and BCLC stage B/C. These findings provide clinical validation of the prognostic association of SMS and extend previous transcriptomic and mechanistic studies by evaluating its relationship with postoperative outcomes and established clinicopathological indicators.

Although SMS remained significantly associated with RFS and OS in multivariable Cox analyses, statistical independence does not necessarily imply incremental predictive value beyond established clinicopathological factors. In the present cohort, high SMS expression was also associated with several aggressive clinicopathological features, suggesting that SMS and conventional clinical variables may capture partially overlapping prognostic information. Consistent with this possibility, the combined clinical+SMS model showed numerically higher time-dependent AUC values than the clinical model; however, these differences did not reach statistical significance in DeLong tests. Therefore, our findings support an independent prognostic association of SMS, whereas its additional discriminatory value beyond the clinicopathological variables included in the clinical model was not statistically established. The limited incremental improvement may, at least in part, reflect overlap between SMS expression and aggressive tumor characteristics, although this interpretation remains exploratory and cannot be established from the present analysis. Further validation in larger independent cohorts is required to determine whether SMS meaningfully improves prognostic prediction when added to established clinical models. In addition, because the prognostic models, including the combined clinical+SMS model, were developed and evaluated within the same retrospective cohort without independent external validation, their observed performance may be optimistic. Moreover, although time-dependent ROC analysis was used to assess discrimination, the present study did not evaluate clinical net benefit across decision thresholds. Therefore, the current findings should not be interpreted as establishing the clinical utility of the combined model. Prospective validation in larger independent cohorts, together with decision-curve analysis where clinically relevant risk thresholds can be defined, will be necessary before the model can be considered for clinical application.

Previous mechanistic findings also suggest that SMS-related polyamine metabolism may have therapeutic relevance in HCC. However, no treatment-response or intervention data were available in our study. Therefore, the therapeutic implications of SMS remain hypothesis-generating and require validation in mechanistic studies and prospective clinical cohorts. Future studies should determine whether SMS expression can identify patient subgroups that may benefit from therapeutic strategies targeting polyamine metabolism or related metabolic pathways.

MVI and BCLC stage are currently among the most widely used pathological and clinical indicators for HCC prognostic assessment. However, because they are based on clinicopathological characteristics, they do not capture aberrant molecular and metabolic alterations within tumor cells. With the advent of precision medicine, incorporation of molecular markers into traditional staging systems to construct multidimensional risk assessment frameworks has become an important approach to improving prognostic accuracy and advancing individualized therapy. For example, integration of AFP levels with BCLC stage improves prognostic stratification in patients with BCLC stage B disease.²¹ However, previous studies have largely focused on single classes of molecular markers, such as serological indicators, genetic mutations, or immunological markers, whereas relatively few have investigated prognostic markers from the perspective of metabolic reprogramming.

In this study, SMS, a molecular marker associated with nitrogen metabolism, was integrated with conventional clinicopathological indicators in risk-stratification analyses based on SMS+MVI and SMS+BCLC. Compared with previous SMS-HCC studies, which primarily characterized the prognostic, immune, or mechanistic roles of SMS, the present analysis extends these findings by evaluating whether SMS expression provides complementary prognostic information when interpreted together with established clinicopathological indicators. Combined stratification by SMS expression and MVI status identified subgroups with distinct postoperative outcomes, with particularly unfavorable outcomes observed in patients with high SMS expression and MVI positivity; SMS expression also further differentiated prognosis within the BCLC A and BCLC B/C subgroups. These findings suggest that SMS may provide additional prognostic information when interpreted together with established clinicopathological indicators, although the stratification results require validation in larger independent cohorts. MVI reflects the capacity of tumor cells to breach the basement membrane and invade the microvascular lumen, an essential step in hematogenous metastasis.^22,23 Existing evidence indicates that hypoxic tumor cells exhibit increased polyamine synthesis; specifically, spermine promotes VEGF expression by stabilizing HIF-1α, thereby enhancing tumor angiogenesis.²⁴ These previous mechanistic observations provide a possible biological context for the association observed in our cohort but do not establish a causal relationship between SMS expression and MVI. The present study demonstrates an association between high SMS expression and MVI positivity as well as other aggressive clinicopathological features, rather than evidence that SMS directly promotes these phenotypes. Whether SMS-driven spermine synthesis contributes directly to MVI, angiogenesis, or tumor invasion requires further functional validation. The potential basis for the additional prognostic information provided by SMS within the BCLC A and BCLC B/C subgroups may relate to the fact that BCLC staging is primarily based on anatomical tumor burden, hepatic functional reserve, and patient performance status and does not directly incorporate molecular or metabolic characteristics of the tumor. Previous studies have shown that metabolically active tumor cells often exhibit greater proliferative capacity, invasiveness, and therapeutic resistance.²⁵ Therefore, SMS expression may provide information on molecular or metabolic heterogeneity that is not directly captured by BCLC stage, although the biological basis underlying this association requires further investigation.

Polyamines have important roles in immunomodulation by regulating immune cell behavior, antigen presentation, and cytokine signaling. Aberrant polyamine metabolism may not only enhance tumor-cell adaptability but also reshape immune interactions, thereby contributing to immune evasion and tumor progression. Immune microenvironment analysis in the present study revealed significantly increased Treg infiltration in the high-SMS group, which is consistent with an immunosuppressive immune context. Xiang et al previously reported that high SMS expression was associated with an immunosuppressive HCC microenvironment characterized by altered immune-cell infiltration and increased expression of immune checkpoint-related molecules, which is broadly consistent with our observation of increased Treg infiltration and altered immune-related functions in the high-SMS group.⁹ Tregs can suppress the antitumor activity of effector T cells through inhibitory cytokines, including IL-10 and TGF-β, as well as immune checkpoint-related mechanisms involving molecules such as CTLA-4 and PD-1. Previous studies have also shown that polyamines can promote Treg differentiation and immunosuppressive activity.^26,27 Notably, our computational immune-deconvolution analysis showed increased estimated M0 macrophage infiltration but decreased estimated M2 macrophage infiltration in the high-SMS group. This observation differs from the mechanistic findings of Sun et al, who experimentally demonstrated that SMS-derived spermine promotes M2 polarization of tumor-associated macrophages.¹⁰ This apparent discrepancy may reflect differences in study populations, analytical platforms, cell-state definitions, and the inherent limitations of computational immune-cell deconvolution. Consequently, the immune-infiltration findings of the present study should be interpreted as associative rather than mechanistic evidence and should not be taken to indicate that SMS directly inhibits M2 macrophage polarization. The biological significance of the macrophage infiltration pattern observed in our computational analysis therefore remains uncertain and requires further experimental investigation. The higher APC co-inhibition enrichment score observed in the high-SMS group was consistent with greater enrichment of co-inhibitory immune-related signaling and an immunosuppressive immune context. The differential pattern of increased Type II IFN response and decreased Type I IFN response is also notable: sustained IFN-γ signaling can promote immune checkpoint molecule expression and immune exhaustion, whereas reduced Type I IFN response may weaken innate immune responses and promote tumor immune escape.²⁸

Limitations

This study has several limitations: 1) As a single-center retrospective study, potential selection bias related to sample sources and patient populations may have been introduced, and the generalizability of the results requires further validation in multicenter prospective cohorts. 2) The sample size was relatively limited (N=144), and small subgroup sizes in certain analyses may have reduced statistical power. Accordingly, the subgroup survival analyses should be considered exploratory and interpreted cautiously, particularly for groups with limited sample sizes. In addition, the prognostic models were developed and evaluated in the same retrospective cohort without independent external validation, which may have resulted in optimistic estimates of model performance. Clinical net benefit was not assessed using decision-curve analysis; therefore, the clinical utility of the combined model remains to be established. 3) This study primarily focused on the prognostic value of SMS. Although SMS protein expression was evaluated by IHC, additional validation using qPCR and Western blotting was not performed. Moreover, the specific functional mechanisms and signaling pathways involved in HCC were not extensively validated experimentally. Because the clinical analyses were observational, the associations between SMS expression and MVI or other aggressive clinicopathological features should not be interpreted as evidence of causality. Further molecular, cellular, and animal studies are needed to clarify the biological mechanisms linking SMS-related polyamine metabolism to HCC aggressiveness. 4) The follow-up duration was relatively short for some patients, which may have affected the accuracy of long-term survival estimates. 5) Data on postoperative adjuvant therapy were not included; therefore, the influence of SMS expression on responses to different treatment regimens could not be evaluated. In summary, large multicenter cohorts should be used in future studies to further validate the prognostic value of SMS, and functional experiments should be conducted to clarify the molecular mechanisms underlying its role in HCC progression.

Conclusions

Aberrant expression of NMRGs in malignant epithelial cells of HCC was identified through single-cell sequencing data analysis, after which machine learning was applied to prioritize SMS as a candidate gene closely associated with HCC prognosis. By analyzing clinical data from HCC patients who underwent curative liver resection, SMS protein expression in tumor tissues was validated by IHC, and high SMS expression remained independently associated with unfavorable RFS and OS in multivariable analyses. Notably, combining SMS expression with MVI status and BCLC stage identified subgroups with distinct postoperative outcomes. These findings provide independent clinical validation of the prognostic relevance of SMS and support its potential value as a complementary molecular marker for postoperative risk stratification in HCC.

Funding Statement

This work was supported by the Baoding Key Laboratory of Gastrointestinal Tumor Diagnosis and Treatment, Baoding, Hebei, China.

Abbreviations

HCC, Hepatocellular carcinoma; SMS, Spermine synthase; IHC, Immunohistochemistry; TCGA, The Cancer Genome Atlas; LIHC, Liver Hepatocellular Carcinoma; GEO, Gene Expression Omnibus; RNA, Ribonucleic acid; NMRGs, Nitrogen metabolism-related genes; NM, Nitrogen metabolism; PCA, Principal component analysis; UMAP, Uniform Manifold Approximation and Projection; ssGSEA, Single-sample Gene Set Enrichment Analysis; AUCell, Area Under the Curve Cell scoring; Cox, Cox proportional hazards regression; LASSO, Least Absolute Shrinkage and Selection Operator; Kaplan-Meier/KM, Kaplan-Meier survival analysis; TNM, Tumor-Node-Metastasis; T, Tumor; N, Node; M, Metastasis; TIMER, Tumor Immune Estimation Resource; GSVA, Gene Set Variation Analysis; MVI, Microvascular invasion; BCLC, Barcelona Clinic Liver Cancer; RFS, Recurrence-free survival; OS, Overall survival.

Data Sharing Statement

The datasets analysed during the current study are available in the GEO repository (http://www.ncbi.nlm.nih.gov/geo), and the Xena repository (https://xena.ucsc.edu/).

Ethics Approval and Informed Consent

This study was approved by the Ethics Committee of Baoding No.1 Central Hospital (Ethics Approval No.[2025] 215). This was a retrospective study using archived formalin-fixed paraffin-embedded (FFPE) tissue samples from hepatocellular carcinoma patients who underwent surgery between 2010 and 2013. Given the retrospective nature of the study and the fact that most patients were lost to follow-up or deceased, it was not feasible to obtain informed consent from all participants. The Ethics Committee of Baoding No.1 Central Hospital waived the requirement for informed consent for this study. All patient data were anonymized prior to analysis to protect patient privacy.

Author Contributions

TY.G, C.G, and H.W contributed equally to this work and share first authorship. All authors made a significant contribution to the conception, study design, execution, acquisition of data, analysis, and interpretation of the work. Specifically, HQ.Z and XL.L designed the study; TY.G, C.G, and H.W collaborated on the work and drafted the manuscript; X.L, HN.W, and BJ.X contributed to interpreting the data and reviewing the literature. All authors made a significant contribution to the work reported, whether that is in the conception, study design, execution, acquisition of data, analysis and interpretation, or in all these areas; took part in drafting, revising or critically reviewing the article; gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work.

Correspondence and requests for materials should be addressed to Huiqing Zhang and Xiaolong Li.

Disclosure

The authors declare no competing interests.