Ewing sarcoma (ES) is one of the most aggressive bone and soft-tissue tumors in children, adolescents, and young adults. Advances in multimodal treatment, combining surgery, radiotherapy, and intensive chemotherapy, have improved outcomes considerably over the past two decades. The 5-year survival rate for patients with localized ES now exceeds 70%, but this figure drops sharply to only 20-30% for patients who present with metastatic disease or experience relapse. Despite this wide prognostic gap, no molecular tool in routine clinical use reliably separates patients with localized disease who are at high risk of recurrence from those who are not, and treatment intensification decisions remain largely guided by anatomic staging alone.
The role of the immune microenvironment: Accumulating evidence shows that ES prognosis is tightly linked to the immune landscape of the tumor. CD8+ cytotoxic T cells recognize ES-specific neoantigens such as enhancer of zeste 2 (EZH2) and chondromodulin 3 and can directly kill tumor cells. Natural killer (NK) cells exert cytotoxicity against ES cells through NKG2D and DNAM-1 receptor pathways, and allogeneic NK cell transfer produces stronger anti-tumor responses than autologous NK cells. Macrophages, mast cells, dendritic cells, and regulatory T cells have all been implicated in ES biology, though their precise roles are incompletely defined. Inflammatory cytokines including interleukin-6 (IL-6), IL-10, and killer cell lectin-like receptor K1 have been shown to modulate the ES tumor microenvironment and correlate with clinical outcomes.
The lncRNA opportunity: Long non-coding RNAs (lncRNAs) are transcripts longer than 200 nucleotides that do not encode proteins but regulate gene expression through signaling, decoy, guide, and scaffold mechanisms. A growing body of research demonstrates that lncRNAs are not only involved in tumor initiation and progression, but also play regulatory roles in immune cell biology. Immune-associated lncRNA signatures have been validated as prognostic biomarkers in glioblastoma, breast cancer, and bladder cancer, but no equivalent tool had been developed for ES prior to this study.
This 2021 paper, published in Frontiers in Cell and Developmental Biology, represents the first systematic construction of an immune-related lncRNA prognostic signature for Ewing sarcoma. The authors combine machine learning iterative lasso regression with transcriptome data from publicly available databases to build and validate an 11-lncRNA model that stratifies ES patients into distinct risk groups with significantly different survival outcomes.
The study draws from two public databases. The primary training set is the Gene Expression Omnibus dataset GSE17679, which contains transcriptome and matched clinical data from 88 Ewing sarcoma samples and 18 normal skeletal muscle samples. The external validation set consists of 58 ES cases downloaded from the International Cancer Genome Consortium (ICGC) database. Using two independent cohorts from different sources is an important methodological strength, since it tests generalizability beyond the training institution.
ESTIMATE algorithm for immune scoring: To stratify the 88 ES samples by immune infiltration level, the authors applied the ESTIMATE algorithm, which uses gene expression signatures to infer immune score, stromal score, and tumor purity from bulk RNA sequencing data. The 88 samples were divided into high and low immune infiltration groups (n = 44 each). The high-infiltration group showed significantly elevated ESTIMATE scores (p less than 0.001) and stromal scores (p less than 0.001), and significantly lower tumor purity (p less than 0.001), confirming that the immune scoring effectively captured real biological differences between groups. Principal component analysis further verified clear separation between the high and low immune infiltration groups.
Filtering for immune-related differentially expressed lncRNAs: The authors used the limma R package to identify lncRNAs that were differentially expressed both between the high and low immune infiltration groups (yielding 262 immune-related lncRNAs) and between ES samples and healthy skeletal muscle (yielding 884 differentially expressed lncRNAs). The intersection of these two lists produced 171 lncRNAs that were both immune-related and differentially expressed in ES. This two-step filtering strategy ensures that candidate lncRNAs are genuinely relevant to the immune microenvironment of the tumor rather than merely correlated with broad transcriptional differences between cancer and normal tissue.
Conditional probability survival analysis: Before building the prognostic model, the authors computed conditional survival curves to understand how prognosis evolves dynamically with time since resection. Starting from a 49% probability of surviving from the point of resection, the 5-year conditional survival probability rose progressively: to 58% at 1 year post-resection, 72% at 2 years, 84% at 3 years, and 100% at 4 years. This finding is clinically useful because it demonstrates that patients who survive longer periods after surgery face progressively lower cumulative risk, which has implications for surveillance intensity and treatment decision-making in long-term survivors.
The machine learning approach at the core of this study is an iterative lasso (least absolute shrinkage and selection operator) regression applied 500 times to the pool of candidate lncRNAs. Lasso regression is well suited to high-dimensional genomic data because it applies L1 regularization, shrinking less informative coefficients toward zero and thereby performing automatic feature selection. In each of the 500 iterations, the algorithm selects a different subset of lncRNAs from the 35 that passed univariate Cox regression filtering (p less than 0.05), producing a different candidate signature for each run. The combination with the largest AUC of the receiver operating characteristic (ROC) curve for overall survival prediction is retained as the optimal signature.
Advantage over stepwise regression: Standard stepwise regression, which is commonly used for constructing prognostic gene signatures, adds or removes variables one at a time based on statistical criteria and can be highly sensitive to the initial variable set and to overfitting. The iterative lasso approach avoids these pitfalls by systematically evaluating all possible combinations over hundreds of runs, incorporating both the individual prognostic value of each lncRNA and the redundancy between lncRNAs. This yields a more parsimonious and generalizable signature.
The 11 lncRNAs in the final signature: The optimal model retains 11 immune-related lncRNAs. Four of these, ARHGAP26 antisense RNA 1 (ARHGAP26-AS1), FUT8 antisense RNA 1 (FUT8-AS1), FOXC1 upstream transcript (FOXCUT), and chromosome 5 putative open reading frame 64 (C5orf64), are highly expressed in the high-risk group. The remaining seven, NAV2 antisense RNA 2 (NAV2-AS2), LINC00408, SEC24B antisense RNA 1 (SEC24B-AS1), LINC01343, LINC01398, LINC01197, and DPP10 antisense RNA 3 (DPP10-AS3), are lowly expressed in the high-risk group. Several of these lncRNAs have established roles in other cancers: FOXCUT promotes colorectal cancer metastasis via the FOXC1/PI3K/Akt pathway, and serves as a prognostic marker in esophageal squamous cell carcinoma, breast cancer, nasopharyngeal carcinoma, and gastric adenocarcinoma. NAV2-AS2 and SEC24B-AS1 are prognostic biomarkers for lung adenocarcinoma and non-small cell lung cancer, respectively.
Each patient receives a composite risk score derived from the weighted expression values of all 11 lncRNAs. Patients above the median risk score are designated high-risk; those below are low-risk. This continuous risk scoring approach allows for more nuanced stratification than binary classification.
In the training set (GSE17679, n = 88), ROC analysis of the 11-lncRNA signature yielded an AUC of 0.819 for overall survival prediction, indicating strong discriminative ability. Kaplan-Meier survival analysis confirmed that the high-risk group had significantly worse prognosis compared to the low-risk group (log-rank p less than 0.001), with clearly diverging survival curves. Both the overall risk scores and the raw death counts were substantially higher in the high-risk group, visually validating the model's ability to separate patients into clinically meaningful strata.
External validation on the ICGC cohort: The 11-lncRNA signature was applied without modification to 58 ES cases from the ICGC database, which was entirely independent from the training cohort. Time-dependent ROC analysis assessed predictive accuracy at three clinically relevant timepoints: 3-year AUC = 0.71, 5-year AUC = 0.68, and 8-year AUC = 0.75. The slightly lower performance relative to the training set is expected and characteristic of external validation; the fact that AUC remains above 0.65 at all three timepoints confirms that the model generalizes meaningfully beyond the training data.
Comparison with established ES prognostic biomarkers: The authors benchmarked the 11-lncRNA signature against four published ES prognostic markers: BCL2 interacting killer (BIK), epidermal growth factor receptor (EGFR), CD44 molecule (Indian blood group), and leucine-rich repeat containing G protein-coupled receptor 5 (LGR5). The lncRNA signature outperformed all four individual biomarkers in the external validation set, demonstrating that the composite machine learning model captures prognostic information that single-gene markers miss. This direct comparison is methodologically important because it grounds the new signature's performance in a clinical reference frame rather than presenting it in isolation.
Independence from clinical variables: A critical test of any molecular prognostic signature is whether it adds value beyond basic clinical characteristics. Kaplan-Meier analysis within subgroups defined by age (above and below median), sex (male and female), and metastatic status showed that the high-risk group retained significantly worse prognosis in all subgroups (p less than 0.05 in each). Time-dependent concordance index (C-index) analysis confirmed that the lncRNA signature alone outperformed models based on age alone, sex alone, and age plus sex, and that adding clinical variables to the lncRNA signature did not substantially improve predictive accuracy.
To explore the biological mechanisms underlying the risk stratification, the authors used CIBERSORT, a linear support vector regression deconvolution algorithm, to estimate the relative abundance of 22 immune cell types in each tumor sample from bulk transcriptome data. CIBERSORT is widely used in tumor immunology because it can resolve the cellular composition of complex mixtures without requiring physical separation of cell populations. Principal component analysis of the CIBERSORT outputs clearly separated the high- and low-risk groups, confirming that the two groups have distinct immune microenvironments.
Immune cell correlations within the tumor microenvironment: Correlation analysis among 22 immune cell types revealed several notable relationships. Plasma cells were positively correlated with M1 macrophages and resting mast cells, but negatively correlated with M2 macrophages, suggesting that higher plasma cell infiltration tracks with a more pro-inflammatory (M1-polarized) macrophage environment. Activated CD4 memory T cells were positively correlated with gamma-delta (gd) T cells and negatively correlated with M0 macrophages. The cell types with the broadest interaction networks were M2 macrophages, resting NK cells, and activated NK cells, while monocytes, naive B cells, and resting dendritic cells had the fewest significant connections to other immune populations.
Prognostic immune cells, favorable and unfavorable: Kaplan-Meier analysis of individual immune cell infiltration levels identified prognostically relevant populations. Higher infiltration by regulatory T cells (Tregs, p = 0.001) and activated CD4 memory T cells (p = 0.001) was associated with favorable prognosis. In contrast, higher infiltration by activated dendritic cells (p = 0.009), M2 macrophages (p = 0.001), monocytes (p less than 0.001), resting mast cells (p less than 0.001), and gd T cells (p less than 0.001) was associated with poor prognosis. Memory B cells and activated NK cells were more abundant in the high-risk group compared to the low-risk group.
Biological context for macrophage and mast cell findings: The finding that M2 macrophages predict poor prognosis is consistent with established tumor biology. Tumor-associated macrophages in the M2 phenotype secrete immunosuppressive cytokines, promote angiogenesis, and facilitate invasion and metastasis. Mast cell infiltration of primary malignant bone tumors, including osteosarcoma, has been associated with poor outcomes in prior studies. The current results suggest that mast cells play a similar role in ES, though the precise mechanisms remain to be characterized through biological experiments.
Beyond identifying which immune cells are prognostically relevant in ES, the study mapped the specific correlations between the 11 lncRNAs in the signature and the prognosis-related immune cell populations using Pearson correlation analysis. This approach illuminates potential mechanisms by which the lncRNAs may exert their prognostic influence, pointing toward biological hypotheses that can guide future wet-lab experiments.
DPP10-AS3 and its immune cell correlates: DPP10 antisense RNA 3 (DPP10-AS3) was positively correlated with resting dendritic cell infiltration, neutrophil infiltration, and gd T cell infiltration. Since all three of these cell types are associated with poor prognosis in the CIBERSORT analysis, the positive correlation between DPP10-AS3 expression and these unfavorable immune populations suggests that higher DPP10-AS3 levels may promote an immunosuppressive or tumor-permissive microenvironment. Interestingly, DPP10-AS3 is among the lncRNAs that are lowly expressed in the high-risk group, meaning its absence may paradoxically be associated with specific immune cell compositions that promote worse outcomes.
LINC01398 and macrophage polarization: LINC01398 was negatively correlated with resting dendritic cell infiltration and M2 macrophage infiltration. Because M2 macrophages are associated with poor prognosis in ES, the negative correlation between LINC01398 and M2 macrophage abundance implies that higher LINC01398 expression may suppress pro-tumorigenic macrophage polarization or recruit resting dendritic cells that maintain immune surveillance. LINC01398 is also among the lncRNAs lowly expressed in the high-risk group, suggesting that its downregulation in high-risk patients may partly explain the shift toward an M2-polarized immunosuppressive microenvironment in these cases.
These correlations are observational and computed from bulk transcriptome data, which limits causal interpretation. CIBERSORT deconvolution estimates cell fractions from aggregate gene expression and may not perfectly reflect single-cell resolution immune compositions. However, the patterns are internally consistent with the known biology of both the immune cell types involved and the general principles of tumor immunoediting, making them plausible starting points for mechanistic validation in ES cell lines and mouse models.
To identify which biological pathways are most enriched in high-risk ES patients (and therefore likely regulated by the 11 immune-related lncRNAs), the authors performed gene set enrichment analysis (GSEA) using GSEA 4.0.3 software. Two reference gene set collections were used: h.all.v7.1.symbols.gmt (Hallmark gene sets covering curated biological processes) and c7.all.v7.1.symbols.gmt (immunologic gene sets derived from immunology experiments). Results were considered significant at nominal p less than 0.05 and false discovery rate (FDR) less than 0.05.
Pathways enriched in the high-risk group: GSEA identified several immune-related gene sets as significantly enriched in high-risk ES patients. These included gene sets related to IL-4 activation of CD4 T cells at 48 hours, LAIV influenza vaccine responses in peripheral blood mononuclear cells, and bone marrow-derived dendritic cell responses to PAM3CSK4 (a toll-like receptor 1/2 agonist). Gene set variation analysis (GSVA), using the Hallmark gene set collection, confirmed activation of IL-2/STAT5 signaling, protein secretion, complement signaling, and PI3K/Akt/mTOR signaling pathways in the high-risk group. The complement system is directly relevant: complement component C5 is activated in ES and high C5 expression is associated with better prognosis, suggesting that complement dysregulation in the high-risk group may reflect a switch away from this protective complement activity.
Immune checkpoint marker expression: The authors examined expression of 17 common immune checkpoint genes in the high- and low-risk groups. Three checkpoints showed statistically significant differential expression. CD40 (p = 0.01) and CD70 (p = 0.019) were higher in the high-risk group, while CD276 (p = 0.019) was higher in the low-risk group. CD40 has previously been shown to be highly expressed in osteosarcoma and ES cell lines, and its expression is closely associated with ES prognosis. CD70, a ligand for the co-stimulatory receptor CD27, is a known therapeutic target in osteosarcoma, though its role in ES had not been reported before this study. CD276 (also known as B7-H3) is an established immunotherapy target in peritoneal cancer, glioma, and central nervous system tumors.
These findings suggest that the 11-lncRNA signature may influence ES prognosis in part by modulating immune checkpoint expression patterns, with potential implications for predicting responsiveness to checkpoint inhibitor immunotherapy in high-risk versus low-risk ES patients.
Computational versus experimental validation: The study is entirely bioinformatic and observational in nature. All findings, including the correlations between specific lncRNAs and immune cell populations, the pathway enrichments, and the checkpoint associations, are derived from publicly available transcriptome datasets and require biological validation in cell lines and animal models. The specific molecular mechanisms by which the 11 lncRNAs influence ES prognosis remain entirely unexplored. This is a fundamental limitation because gene expression correlations can reflect co-regulation without implying functional causality, and lncRNAs can have complex, context-specific functions that correlative analyses cannot capture.
Dataset size constraints: The training set consists of only 88 ES samples, and the external validation set contains 58 cases. While both cohorts are drawn from established public databases, Ewing sarcoma is a rare malignancy and these sample sizes are modest by the standards of modern machine learning validation. Small datasets increase the risk of overfitting even with regularization methods like lasso, and may not adequately represent the full biological heterogeneity of ES across different patient populations, treatment protocols, and institutions. The performance drop observed between training (AUC = 0.819) and external validation (AUC = 0.68-0.75) is partly attributable to this small-sample problem.
Biological unknowns for individual lncRNAs: For the majority of the 11 lncRNAs in the signature, their functional roles specifically in Ewing sarcoma are entirely uncharacterized. Several have established roles in other cancer types (FOXCUT in colorectal and esophageal cancer, NAV2-AS2 and SEC24B-AS1 in lung cancer), but direct evidence for functional activity in ES cells is absent. The specific regulatory relationships between the 11-lncRNA signature and the immune checkpoint targets CD40, CD70, and CD276 also require experimental elucidation. Large-scale wet-lab studies will be needed before these lncRNAs can be considered actionable targets.
Future directions: Prospective multicenter validation using newly collected ES cohorts with standardized treatment protocols and longer follow-up would substantially strengthen confidence in the signature's clinical utility. Single-cell RNA sequencing could resolve the immune cell composition of ES tumors at higher resolution than CIBERSORT deconvolution allows, potentially revealing cell-type-specific lncRNA expression patterns. Functional experiments knocking down or overexpressing the 11 lncRNAs in ES cell lines and xenograft models would determine whether they causally regulate the immune pathways identified here. Integration of the lncRNA signature with existing clinical staging systems and molecular markers could improve prognostic precision further and help select patients most likely to benefit from immune checkpoint inhibitor therapy in future trials.