Construction of a Glycosylation-Related Prognostic Signature for Predicting Prognosis, Tumor Microenvironment, and Immune Response in Soft Tissue Sarcoma

Frontiers in Oncology 2025 AI 8 Explanations View Original
Original Paper (PDF)

Unable to display PDF. Download it here or view on PMC.

Plain-English Explanations
Pages 1-2
Glycosylation as an Untapped Prognostic Signal in Soft Tissue Sarcoma

Soft tissue sarcoma (STS) is a rare mesenchymal malignancy accounting for roughly 1% of all adult cancers. Its clinical management is complicated by extreme histological heterogeneity, with dozens of recognized subtypes that differ in cellular origin, molecular drivers, and treatment responsiveness. Despite standardized anthracycline-based chemotherapy and advances in surgical technique, up to 40% of patients with localized disease develop metastatic progression, which is frequently fatal. Five-year survival rates for metastatic STS remain below 20%, and the pace of therapeutic improvement has been slow relative to other solid tumors.

The glycosylation gap: Glycosylation, the enzymatic attachment of sugar moieties to proteins and lipids, is the most abundant post-translational modification in eukaryotic cells. In cancer, dysregulated glycosylation touches virtually every hallmark of malignancy: altered glycan patterns promote tumor cell proliferation, invasion, metastasis, angiogenesis, immune evasion, and resistance to chemotherapy-induced apoptosis. Common cancer-associated glycosylation changes include O-glycan truncation, aberrant sialylation, fucosylation, and N-glycan branching. Despite this centrality to tumor biology, glycosylation-related genes had received comparatively little systematic attention in STS research, which has focused more heavily on recurrent mutations in RAS/MAPK and PI3K/AKT/mTOR pathways and on immune checkpoint expression.

Study rationale: This 2025 study published in Frontiers in Oncology addresses that gap by conducting the first comprehensive machine learning analysis of glycosylation-related gene (GRG) expression in STS using large-scale transcriptomic datasets. The authors aimed to build a glycosylation-related prognostic signature (GRPS) capable of stratifying patients by survival risk, characterizing the immune microenvironment associated with each risk group, predicting therapeutic sensitivity, and identifying specific glycosyltransferases that could serve as therapeutic targets.

The study combines bioinformatics analysis of public datasets with in vitro functional experiments, bridging computational prediction and bench-level validation. The primary training cohort was TCGA-SARC, comprising 265 STS cases with comprehensive clinical and transcriptomic data. An independent validation cohort from the Gene Expression Omnibus (GSE17674, 32 cases) was used to test model generalizability.

TL;DR: STS affects 1% of adults, with 40% developing metastatic disease and poor survival. Glycosylation dysregulation drives proliferation, invasion, and immune evasion in cancer but had not been systematically studied in STS. This paper builds a machine learning prognostic model from glycosylation-related genes using 265 TCGA-SARC training cases and 32 GEO validation cases.
Pages 2-4
A 101-Algorithm Machine Learning Framework for Feature Selection

The methodological core of this study is a comprehensive machine learning framework that evaluates 101 distinct algorithm combinations to identify the optimal prognostic model from a starting set of glycosylation-related genes. Differential expression analysis between STS tissue samples and normal tissue in the GSE21122 dataset identified 2,092 differentially expressed genes (1,100 upregulated, 992 downregulated). The intersection of these DEGs with a curated set of 240 glycosylation-related genes yielded 65 glycosylation-related differentially expressed genes. Univariate Cox regression with P less than 0.05 further filtered this set to 21 genes significantly associated with STS prognosis.

The 101-algorithm approach: To avoid the bias inherent in selecting a single modeling strategy, the authors systematically tested 101 combinations of 10 base algorithms: Lasso (least absolute shrinkage and selection operator), StepCox (stepwise Cox regression), Elastic Net (Enet), Ridge regression, Random Survival Forest (RSF), Partial Least Squares Cox Regression (plsRcox), CoxBoost, Gradient Boosting Machine (GBM), Supervised Principal Component Analysis (SuperPC), and Survival Support Vector Machine (Survival-SVM). Each combination was evaluated using 10-fold cross-validation in the TCGA training cohort, with the concordance index (C-index) serving as the primary performance metric. The combination that achieved the highest C-index across both the training and independent validation cohort was selected as the optimal model.

NMF molecular subtyping: Prior to prognostic model construction, non-negative matrix factorization (NMF) clustering was applied to the TCGA cohort using the differentially expressed GRGs to identify distinct STS molecular subtypes. The optimal cluster number k=2 was selected based on cophenetic, dispersion, and silhouette metrics. This unsupervised approach revealed two biologically distinct STS subgroups with significantly different survival outcomes and immune profiles, providing biological context for the downstream prognostic model.

Immune and functional analyses: Immune cell infiltration was quantified using single-sample gene set enrichment analysis (ssGSEA) and the ESTIMATE algorithm, which computes stromal score, immune score, and tumor purity from transcriptomic data. Drug sensitivity was predicted using the OncoPredict R package trained on the GDSC2 dataset, with half-maximal inhibitory concentration (IC50) values estimated for each sample. Immunotherapy response was externally validated in the IMvigor210 cohort, comprising patients with advanced urothelial carcinoma treated with atezolizumab (anti-PD-L1). Tumor mutational burden (TMB) was analyzed from TCGA somatic mutation data using the maftools R package.

TL;DR: From 240 glycosylation-related genes, 65 were differentially expressed in STS, narrowed to 21 with prognostic significance by univariate Cox regression. A 101-algorithm framework (Lasso, RSF, GBM, Survival-SVM, and 6 others in 101 combinations) with 10-fold cross-validation selected the optimal model by C-index. NMF clustering identified 2 molecular STS subtypes. Drug sensitivity, immune infiltration, and immunotherapy response were assessed using OncoPredict, ssGSEA, ESTIMATE, and the IMvigor210 cohort.
Pages 4-6
NMF Clustering Reveals Two Biologically Distinct STS Subtypes

NMF clustering of the 265 TCGA-SARC patients based on differentially expressed glycosylation-related genes produced a stable two-cluster solution (k=2), confirmed by PCA scatter plots showing clear subgroup separation. Kaplan-Meier survival analysis demonstrated that Cluster 1 patients had significantly worse overall survival than Cluster 2 patients, establishing that glycosylation gene expression patterns are not merely epiphenomenal but directly track clinically meaningful differences in patient outcomes.

Pathway enrichment between clusters: Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis revealed distinct biological profiles between the two clusters. Cluster 1, associated with poorer prognosis, showed enrichment in pathways linked to cytokine-cytokine receptor interaction and immune suppression. Gene Set Variation Analysis (GSVA) using HALLMARK gene sets from the GSEA database further distinguished the immune activity landscape of each cluster, with Cluster 2 showing stronger immune activation signatures.

Copy number variations: Analysis of copy number variations (CNVs) among the glycosylation-related genes revealed interesting patterns. Genes including PIGC, RPN2, ALG6, DDOST, EIF2B3, and B4GALT2 showed increased CNV frequency (gains), while ST3GAL4, STT3A, CHPF, B3GAT3, B3GALNT1, and GLT8D1 showed decreased CNV (losses). These CNV patterns suggest that genomic alterations in glycosylation regulators actively contribute to STS tumor biology and are not simply byproducts of chromosomal instability. The chromosomal locations of the 21 differentially expressed GRGs were annotated across 23 chromosomes, with a network plot demonstrating extensive molecular interactions among glycosylation-related gene products.

The NMF-derived molecular subtypes contextualize the downstream risk model: the prognostic risk score is capturing a biologically real axis of variation in glycosylation gene expression that divides patients into groups with meaningfully different survival trajectories, immune microenvironments, and genomic characteristics.

TL;DR: NMF clustering of 265 TCGA STS cases by glycosylation gene expression produced 2 stable subtypes with significantly different overall survival. Cluster 1 (poor prognosis) showed immune suppression and cytokine receptor pathway enrichment. CNV analysis revealed amplification in PIGC, RPN2, ALG6, DDOST, EIF2B3, and B4GALT2, with deletion in STT3A, B3GAT3, and GLT8D1.
Pages 6-8
The 12-Gene GRPS: Construction, Validation, and Clinical Nomogram

Among all 101 algorithm combinations tested, the Lasso + plsRcox combination achieved the highest C-index in both the TCGA training cohort and the independent GEO validation cohort (GSE17674, n=32). This hybrid approach combines Lasso regularization, which performs variable selection by shrinking coefficients toward zero and eliminating weaker predictors, with partial least squares Cox regression (plsRcox), which handles multicollinearity among predictors by constructing latent components that maximize covariance with survival outcomes. The resulting model retained 12 glycosylation-related genes: RPN2, ALG6, XYLT2, DDOST, MFNG, PIGC, EIF2B3, B3GAT3, B3GNT4, B4GALT2, GLT8D1, and STT3A.

Risk stratification performance: Multivariate Cox regression on the 12 selected genes generated regression coefficients used to calculate individual risk scores. Patients were stratified into high-risk and low-risk groups based on the median risk score cutoff in each cohort. Risk curves and survival status plots confirmed a clear separation between groups in both cohorts. ROC curve analysis demonstrated AUC values exceeding 0.8 for 1-year, 3-year, and 5-year overall survival prediction in both the TCGA training set and the GEO validation set, indicating strong and generalizable prognostic performance.

Independent prognostic value: Univariate and multivariate Cox regression analyses incorporating clinical covariates (gender, main tumor site, metastatic status, specific tumor location, and risk score) confirmed that metastasis status and the GRPS risk score were the only variables with independent prognostic significance (P less than 0.05). Age, sex, and tumor location did not reach significance in multivariate analysis, underscoring the added value of the molecular risk score beyond standard clinical variables.

Clinical nomogram: To translate the prognostic model into a clinician-friendly tool, a nomogram was constructed incorporating the risk score alongside significant clinical variables for individualized 1-, 3-, and 5-year OS probability estimation. The nomogram's C-index was higher than that of the risk score alone or any individual clinical variable, confirming the value of integration. Calibration plots showed close alignment between predicted and actual survival rates at all three time points. Decision curve analysis (DCA) demonstrated that the nomogram provided net clinical benefit across a range of decision thresholds, supporting its potential utility in clinical risk communication.

TL;DR: Lasso + plsRcox outperformed all other 101 algorithm combinations by C-index. The resulting 12-gene GRPS (RPN2, ALG6, XYLT2, DDOST, MFNG, PIGC, EIF2B3, B3GAT3, B3GNT4, B4GALT2, GLT8D1, STT3A) achieved AUC above 0.8 for 1-, 3-, and 5-year OS in both TCGA and GEO cohorts. GRPS risk score and metastasis status were the only independent prognostic variables by multivariate Cox regression. A clinical nomogram combining risk score and clinical variables outperformed either alone.
Pages 8-10
Immune Landscape: High-Risk Patients Carry an Immunosuppressive Microenvironment

The tumor microenvironment (TME) is a critical determinant of STS behavior and responsiveness to immunotherapy. ssGSEA analysis quantified the relative abundance of 28 immune cell types across the high-risk and low-risk GRPS groups, revealing clear and statistically significant differences. The low-risk group showed significantly greater enrichment of activated B cells, activated CD8+ T cells, and immature dendritic cells, all of which are associated with productive antitumor immune responses. The high-risk group, by contrast, showed patterns consistent with immune exclusion or exhaustion, with enrichment of immunosuppressive cell types including regulatory T cells (Tregs) and M2-polarized macrophages.

Immune function differences: Immune function analysis identified significant between-group differences in type II interferon (IFN-gamma) response, antigen-presenting cell (APC) co-stimulation, cytokine-cytokine receptor (CCR) signaling, parainflammation, and immune checkpoint activity. Collectively, these data indicate that low-risk patients mount stronger and more coordinated antitumor immune responses, while high-risk patients operate in a suppressed immune context. Importantly, this divergence in immune activity is linked to the glycosylation risk score, suggesting that glycosylation dysregulation actively shapes the immune landscape rather than being a secondary correlate.

ESTIMATE scores and tumor purity: ESTIMATE analysis computed stromal scores, immune scores, and total ESTIMATE scores for each patient. Violin plot comparisons between risk groups showed that the high-risk group had significantly different stromal and immune scores, with correlation analysis demonstrating that the GRPS risk score was significantly associated with reduced ImmuneScore and increased tumor purity, consistent with a "cold" immune phenotype. Correlation heatmaps linking each of the 12 GRPS genes individually to ESTIMATE scores highlighted that STT3A showed particularly strong correlations with immune suppression, motivating its selection for downstream functional validation.

Immune checkpoint gene expression: Differential analysis of immune checkpoint-related genes between risk groups revealed that PD-L1, PD-1, and BTLA expression levels were significantly elevated in the low-risk group compared to the high-risk group. This counterintuitive finding, where checkpoint expression is higher in low-risk patients, reflects the fact that high checkpoint expression in a context of active immune infiltration (the low-risk group) signals an ongoing but partially suppressed immune response rather than immunosuppression itself. It also suggests that low-risk patients may be more likely to respond to immune checkpoint inhibitor (ICI) therapy.

TL;DR: Low-risk GRPS patients showed enrichment of activated CD8+ T cells, B cells, and immature dendritic cells by ssGSEA, along with stronger IFN-gamma response and APC co-stimulation. High-risk patients had immunosuppressive infiltrates (Tregs, M2 macrophages) and lower ESTIMATE immune scores. PD-L1, PD-1, and BTLA were higher in the low-risk group, suggesting better ICI candidacy for low-risk patients.
Pages 10-12
Drug Sensitivity Predictions and Real-World Immunotherapy Validation

Drug sensitivity analysis using the OncoPredict R package, trained on the GDSC2 pharmacogenomics dataset, predicted IC50 values for hundreds of compounds across TCGA-SARC samples. A total of 39 compounds showed statistically significant IC50 differences between high-risk and low-risk groups (P less than 0.001). The high-risk group demonstrated increased predicted sensitivity (lower IC50) to Epirubicin, Cyclophosphamide, Docetaxel, Temozolomide, Dactinomycin, Bortezomib, AZD6738 (an ATR inhibitor), and Wee1 Inhibitor. These findings are clinically actionable: Epirubicin and Docetaxel are already used in STS management, and the predicted differential sensitivity between risk groups could inform treatment selection decisions.

ATR and Wee1 inhibitor sensitivity: The predicted sensitivity of the high-risk group to AZD6738 (ATR inhibitor) and Wee1 inhibitor is particularly noteworthy. Both agents target DNA damage response kinases and are being evaluated in clinical trials for STS and other solid tumors. Higher predicted sensitivity in the high-risk group could reflect greater replication stress in glycosylation-dysregulated tumors, a biologically plausible mechanism given the role of N-linked glycosylation in protein quality control within the endoplasmic reticulum. Disruption of N-glycosylation by aberrant glycosyltransferase expression could increase ER stress and proteotoxic burden, making cells more reliant on DNA damage checkpoint pathways.

IMvigor210 immunotherapy validation: The predictive relevance of the GRPS risk score for immunotherapy response was tested in the IMvigor210 cohort, comprising patients with advanced urothelial carcinoma treated with atezolizumab (anti-PD-L1). While this is a different cancer type, it provides a large externally characterized immunotherapy dataset not available for STS specifically. In this cohort, low-risk patients demonstrated significantly extended overall survival compared to high-risk patients (P = 0.00065). Risk score distributions showed that patients achieving complete or partial response (CR/PR) had significantly lower risk scores than those with stable or progressive disease (SD/PD). Moreover, the high-risk group contained a higher proportion of SD/PD patients, providing direct evidence that the GRPS risk score tracks immunotherapy responsiveness independent of tumor type.

The drug sensitivity and immunotherapy analyses together reinforce the clinical utility of the GRPS: high-risk patients are less likely to respond to immune checkpoint blockade but may be more sensitive to certain cytotoxic and targeted agents, providing a framework for differential treatment selection based on GRPS classification.

TL;DR: OncoPredict identified 39 drugs with significantly different IC50 values by risk group. High-risk patients showed greater predicted sensitivity to Epirubicin, Cyclophosphamide, Docetaxel, Dactinomycin, Bortezomib, AZD6738, and Wee1 Inhibitor. In the IMvigor210 atezolizumab cohort (n not stated in paper), low-risk patients had significantly longer survival (P = 0.00065) and higher rates of CR/PR, validating the risk score as an immunotherapy response predictor.
Pages 12-14
STT3A Silencing Suppresses Sarcoma Cell Proliferation and Migration

Among the 12 GRPS genes, STT3A was selected for in vitro functional validation based on its consistent upregulation in STS cell lines by RT-qPCR and its strong bioinformatic associations with poor prognosis and immune suppression. STT3A encodes the catalytic subunit of the oligosaccharyltransferase (OST) complex, which is located in the endoplasmic reticulum and mediates co-translational N-linked glycosylation, transferring oligosaccharide chains from dolichol-linked donors to asparagine residues on newly synthesized polypeptides. This core glycosylation function makes STT3A a central node in the protein quality control and tumor immunology networks.

Cell lines and knockdown: Functional experiments were conducted in two STS cell lines: A673 (human rhabdomyosarcoma, Ewing sarcoma subtype) and SW872 (human liposarcoma). Human bone marrow-derived mesenchymal stem cells (MSCs) served as non-tumor controls. RT-qPCR in these three cell types confirmed STT3A overexpression in both STS lines relative to MSCs, validating the transcriptomic finding. siRNA-mediated knockdown of STT3A was confirmed at both the mRNA level (qPCR) and protein level (Western blotting using anti-STT3A and anti-beta-actin antibodies) in both cell lines.

Proliferation and viability assays: CCK-8 assays assessing cell viability over multiple time points demonstrated that STT3A knockdown significantly suppressed proliferation in both A673 and SW872 cells compared to siNC controls. Colony formation assays confirmed this finding, with STT3A-silenced cells forming significantly fewer colonies after one week of culture. These results demonstrate that STT3A is not merely a passive correlate of sarcoma malignancy but an active contributor to tumor cell growth.

Migration and invasion assays: Wound healing (scratch) assays showed that STT3A knockdown significantly reduced the migration rate of both A673 and SW872 cells over 24 hours. Transwell assays without Matrigel confirmed impaired migration, while Matrigel-coated Transwell assays demonstrated significantly reduced invasion capacity following STT3A silencing. Statistical analyses were performed using one-way ANOVA with a significance threshold of P less than 0.05, and each experiment was conducted in triplicate. All cell functional endpoints (proliferation, colony formation, migration, and invasion) showed consistent, statistically significant suppression upon STT3A knockdown, making a strong case for STT3A as a functionally relevant oncogene in STS.

TL;DR: STT3A was overexpressed in A673 (rhabdomyosarcoma) and SW872 (liposarcoma) cell lines versus MSC controls by RT-qPCR. siRNA knockdown confirmed by qPCR and Western blotting significantly reduced proliferation (CCK-8, colony formation), migration (wound healing, Transwell), and invasion (Matrigel Transwell) in both lines, establishing STT3A as a functional oncogene in STS.
Pages 14-16
Model Boundaries and the Path Toward Clinical Translation

Retrospective and database-only design: The GRPS was constructed and validated exclusively using retrospective public transcriptomic datasets (TCGA-SARC and GSE17674) without access to clinically sourced fresh tissue samples from independent institutional cohorts. Public datasets, while comprehensive, may not capture the full heterogeneity of real-world STS populations treated across diverse clinical settings. Batch effects and differences in RNA sequencing platforms, library preparation protocols, and clinical data completeness between TCGA and GEO datasets can introduce noise that is difficult to fully correct, potentially inflating apparent generalizability.

Histologic subtype heterogeneity: Soft tissue sarcoma encompasses dozens of distinct histologic subtypes with different molecular drivers, clinical behaviors, and treatment sensitivities. The current analysis treats STS as a single entity for risk stratification without stratifying by subtype (e.g., liposarcoma, leiomyosarcoma, synovial sarcoma, undifferentiated pleomorphic sarcoma). It is plausible that the GRPS has different prognostic accuracy across subtypes, or that subtype-specific glycosylation signatures would outperform the pan-STS model. Future analyses should explicitly assess subtype-stratified performance.

Limited functional validation: Among the 12 GRPS genes, in vitro functional validation was conducted only for STT3A in two cell lines. The mechanistic contributions of the other 11 genes (RPN2, ALG6, XYLT2, DDOST, MFNG, PIGC, EIF2B3, B3GAT3, B3GNT4, B4GALT2, and GLT8D1) to STS biology have not been experimentally characterized in this study. While each gene has prior literature supporting its role in glycosylation pathways and cancer biology, the specific mechanisms through which they contribute to STS prognosis remain to be elucidated. Additionally, in vivo tumor models are needed to confirm the therapeutic relevance of STT3A targeting in STS before clinical investigation is warranted.

Path forward: Prospective multicenter validation in clinically annotated STS cohorts, ideally stratified by histologic subtype and treatment regimen, is the critical next step. Integration of the GRPS with protein-level glycosylation assays (glycoproteomics, lectin arrays) could provide orthogonal validation of the transcriptomic risk score. The drug sensitivity predictions for AZD6738 and Wee1 inhibitor in high-risk patients provide testable hypotheses for combination therapy trials. Developing a clinically deployable risk scoring tool (such as a web-based nomogram calculator) would facilitate adoption in prospective studies and ultimately in personalized STS treatment planning.

TL;DR: Key limitations include retrospective public-dataset design (TCGA + GEO only), lack of histologic subtype stratification in a heterogeneous cancer, and functional validation restricted to STT3A in 2 cell lines without in vivo confirmation. Prospective multicenter validation, subtype-stratified analysis, and clinical-grade glycoproteomics assays are the necessary next steps. Drug sensitivity predictions for AZD6738 and Wee1 inhibitor in high-risk patients provide targets for translational studies.