Soft tissue sarcomas (STS) represent a clinically demanding group of malignancies derived from embryonic mesoderm, encompassing more than 70 recognized histological subtypes. They account for approximately 1% to 2% of all new adult cancer cases annually, yet their rarity is matched by their aggressive biology. Even with standard-of-care approaches combining complete surgical resection, adjuvant or neoadjuvant radiotherapy, and systemic chemotherapy, a substantial proportion of patients experience local recurrence or distant metastasis. The resulting prognosis is poor for advanced and metastatic cases, and tumor heterogeneity both across and within individual tumors continues to limit the predictive power of current risk stratification tools.
The case for molecular biomarkers: Because STS is so heterogeneous, a single clinical variable such as tumor size, grade, or histological subtype captures only a fraction of prognostic information. Polygenic signatures, panels of genes whose combined expression pattern predicts outcome, have emerged as a more powerful approach. However, existing polygenic signatures for STS have faced repeated criticism for underutilizing available data, relying on inappropriate machine learning methods, failing to validate across independent cohorts, and lacking clinical-grade experimental testing. These limitations have prevented widespread adoption.
Two biological targets: NETs and lncRNAs: This 2024 study published in Frontiers in Immunology focuses on two intersecting biology layers that have not previously been integrated for STS prognosis. Neutrophil Extracellular Traps (NETs) are lattice-like structures secreted by activated neutrophils, composed of DNA fibers, histones, and antibacterial proteins. Initially described as an antimicrobial defense, NETs have since been shown to promote tumor growth, progression, angiogenesis, metastasis, and cancer-associated thrombosis in multiple solid tumor types. Long non-coding RNAs (lncRNAs), RNA transcripts longer than 200 nucleotides that do not encode proteins, regulate cancer-relevant processes including proliferation, migration, invasion, and chemoresistance. The authors hypothesize that identifying lncRNAs functionally linked to NETs activity (called NETsLnc) would yield a biologically grounded, more stable prognostic signature for STS.
The study also tests whether a machine learning consensus strategy, combining 96 algorithmic frameworks derived from 10 different algorithms, can identify the most robust version of this signature across multiple independent patient cohorts. Prior work by the same group had demonstrated that this "AI consensus" approach outperformed any single algorithm for similar signatures in melanoma and colorectal cancer.
The study assembled a multi-cohort dataset of 969 total STS samples drawn from four publicly available sources: 259 samples from The Cancer Genome Atlas (TCGA), 88 from the Therapeutically Applicable Research to Generate Effective Treatments (TARGET) database, 310 from the GSE21050 GEO dataset, and 312 from the GSE71118 GEO dataset. For each sample, transcriptome expression data, somatic mutation data, copy number variation (CNV) data, and clinical information (including overall survival and disease-free survival) were collected. Samples lacking survival time or survival status data were excluded. RNA-seq expression values were log2-transformed for normalization, while microarray expression data were normalized using the Robust Multiarray Average (RMA) method. Two additional immunotherapy datasets, IMvigor210 (bladder cancer patients treated with anti-PD-L1 atezolizumab) and the Liu David melanoma immunotherapy cohort, were incorporated specifically for immunotherapy response validation.
Identifying NETsLnc via WGCNA: The first analytical step was identifying which lncRNAs co-express with NETs-related gene programs. The authors used Weighted Gene Co-expression Network Analysis (WGCNA), an R package that constructs a scale-free gene co-expression network. After removing outlier samples through hierarchical clustering and setting the soft-power threshold at beta=9 to achieve scale-free topology, the WGCNA algorithm identified 8 co-expression modules in the combined transcriptome data. Module-trait correlation analysis was then performed to determine which modules were most strongly correlated with NETs scores (calculated using single-sample Gene Set Enrichment Analysis, ssGSEA), immune scores, and stromal scores. Two modules, designated "yellow" and "black" based on color coding in the WGCNA output, showed the strongest correlation with NETs activity and were selected. The genes within these modules constituted the candidate NETsLnc pool for further analysis.
The 96-algorithm consensus pipeline: From the WGCNA-identified NETsLnc pool, univariate Cox regression was used to filter to 87 lncRNAs significantly associated with overall survival. These 87 candidates were then entered into an integrated machine learning pipeline combining 96 algorithm configurations derived from 10 base algorithms, including LASSO, Ridge regression, ElasticNet, Random Survival Forest (RSF), stepwise Cox regression, gradient boosting machines, support vector machines, and several others. For each of the 96 configurations, tenfold cross-validation was applied to the TCGA training cohort (the largest available dataset), and the mean concordance index (c-index) was computed across all four cohorts. The algorithm combination with the highest average c-index was selected as the final model. All statistical analyses used R version 4.0.1; ROC curves were generated using the "timeROC" package; nomograms used the "regplot" and "rms" packages.
Before constructing the NETsLnc signature, the authors first characterized the relationship between NETs activity and clinical outcomes in the STS patient cohorts. Using ssGSEA to calculate per-patient NETs scores based on the expression of established NETs-related genes, they found that STS patients with higher NETs scores had significantly improved overall survival on Kaplan-Meier analysis. This counterintuitive finding, in which higher NETs activity correlates with better prognosis, runs counter to some findings in other cancer types where NETs promote metastasis, and suggests that the relationship between NETs and STS biology is nuanced and likely context-dependent.
Immune cell correlations: Higher NETs scores were significantly associated with increased infiltration of several immune cell populations within the tumor microenvironment, including CD8+ T cells, M2 macrophages, and neutrophils themselves, as assessed by Spearman correlation analysis. Notably, CD8+ T cell infiltration is generally associated with improved immunotherapy response and prognosis across solid tumors. The co-enrichment of multiple immune cell types in the high-NETs-score group suggests that NETs activity in STS may reflect a more immunologically active tumor microenvironment rather than a purely pro-tumorigenic state.
Immune checkpoint correlations: A positive correlation was identified between NETs scores and expression of several immune checkpoint molecules, including TIM3, PD-1, PD-L2, and CTLA4. Higher expression of these checkpoints in the high-NETs-score group implies that while the tumor microenvironment is more immunologically active, it is also subject to immunosuppressive braking mechanisms that could be pharmacologically targeted. Additionally, NETs scores correlated strongly with both stromal scores and immune scores computed by the ESTIMATE algorithm, further supporting the idea that NETs activity is intertwined with the composition and state of the TME in STS.
These TME characterization results motivate the core hypothesis: that lncRNAs co-expressed with NETs-related genes will capture biologically meaningful variation in the STS immune landscape, and that this variation will be prognostically informative.
After filtering 87 survival-associated NETsLnc candidates through univariate Cox regression and testing all 96 machine learning algorithm combinations, the combination of Random Survival Forest (RSF) and LASSO regression achieved the highest mean c-index across all four patient cohorts and was selected as the final model. The resulting NETsLnc signature contained 23 lncRNA features, with their coefficients derived from the LASSO regularization step to minimize overfitting. Each patient was assigned a continuous NETsLnc risk score based on the weighted expression of these 23 features. Patients were then dichotomized into high and low NETsLnc score groups using the median risk score as the cutoff.
Prognostic performance: In the TCGA training cohort, Kaplan-Meier analysis showed significantly inferior overall survival in the high NETsLnc score group compared to the low NETsLnc score group. Time-dependent ROC analysis in the TCGA cohort yielded AUC values of 0.943 for 1-year OS prediction, 0.991 for 3-year OS prediction, and 0.988 for 5-year OS prediction, reflecting exceptional discrimination in the training cohort. In the TARGET validation cohort, AUC values were 0.563 (1-year), 0.641 (3-year), and 0.746 (5-year), showing meaningful but substantially lower performance. In the two GEO cohorts (GSE21050 and GSE71118), AUC values were close to 0.5 (near random chance), ranging from 0.471 to 0.522, which the authors attribute to the fact that these datasets report disease-free survival (DFS) rather than overall survival (OS), creating an outcome mismatch with the OS-trained model.
Independent prognostic value: The NETsLnc signature outperformed conventional clinical prognostic variables including age, sex, and metastatic status in both the TCGA and GEO cohorts. Multivariate Cox regression analysis confirmed that the NETsLnc score was an independent predictor of overall survival, maintaining statistical significance after adjustment for these clinical covariates. A nomogram integrating the NETsLnc signature score with clinical characteristics was developed using the "regplot" and "rms" R packages. Calibration curves and AUC analysis confirmed good concordance between nomogram-predicted survival probabilities and actual observed outcomes.
To understand why high NETsLnc scores associate with worse prognosis, the authors examined the biological programs enriched in each risk group. Analysis of the cancer-immunity cycle, a 7-step framework describing the sequential steps by which immune cells recognize, infiltrate, and kill cancer cells, revealed that the high NETsLnc group exhibited increased activity across multiple steps of this cycle. This is consistent with the TME characterization showing greater immune cell infiltration, though the presence of elevated checkpoint expression suggests that this immune activity is not translating into effective anti-tumor killing.
Hallmark and KEGG pathway enrichment: NETsLnc scores showed positive correlation with several classical cancer-driving pathways: MYC targets, Wnt/beta-catenin signaling, and TGF-beta signaling. The association with Wnt/beta-catenin is particularly notable for STS because this pathway is constitutively activated in multiple STS subtypes including liposarcoma, leiomyosarcoma, synovial sarcoma, and fibrosarcoma, where CDC25A has been identified as a key downstream target. TGF-beta signaling is a well-established driver of both tumor progression and immunosuppression in the tumor microenvironment.
Distinct enrichment patterns between groups: Gene Set Enrichment Analysis (GSEA) using KEGG gene sets revealed that the low NETsLnc group was enriched in metabolic and translational pathways, including aminoacyl tRNA biosynthesis, cysteine and methionine metabolism, ribosome biogenesis, RNA polymerase activity, and spliceosome function. By contrast, the high NETsLnc group showed enrichment in calcium signaling, complement and coagulation cascades, hematopoietic cell lineage pathways, and metabolic pathways involving nicotinate, nicotinamide, and phenylalanine. The immune atlas radar map analysis demonstrated significant upregulation of cytolytic activity, inflammation promotion, and APC co-inhibition markers in the low NETsLnc group, patterns associated with a more active anti-tumor immune response and potentially better immunotherapy candidacy.
The authors examined whether the NETsLnc risk groups differed in genomic instability and immune cell composition, since both factors are clinically relevant for treatment selection. Waterfall plots of the top 20 most frequently mutated genes (FMGs) in the STS cohort were generated using the "maftools" R package. Among these common FMGs, ATRX and DNAH14 showed significantly elevated mutation frequencies in the high NETsLnc group compared to the low NETsLnc group. ATRX mutations are particularly relevant in STS: they are found in a substantial fraction of leiomyosarcomas and are associated with the Alternative Lengthening of Telomeres (ALT) mechanism, a telomerase-independent pathway for telomere maintenance that correlates with high genomic instability.
Copy number variation burden: Analysis of copy number variation (CNV) data derived from GISTIC2.0 demonstrated that patients in the high NETsLnc group had more pronounced amplification and deletion changes across chromosomal arms. NETsLnc scores showed positive correlation with both deletion burden and amplification burden. Although tumor mutational burden (TMB) did not differ significantly between the high and low NETsLnc groups, patients with high NETsLnc scores exhibited elevated homologous recombination deficiency (HRD) burden and diminished microsatellite instability (MSI) scores. Together, these findings suggest that high NETsLnc scores characterize a genomically unstable subtype with copy number complexity and HRD, while low NETsLnc scores reflect a more stable genomic background.
ESTIMATE and CIBERSORT immune profiling: Using the ESTIMATE algorithm on expression data, the high NETsLnc group showed higher tumor purity scores and lower stromal, immune, and ESTIMATE composite scores, indicating that tumors in this group have fewer infiltrating stromal and immune cells per unit of tumor mass. CIBERSORT deconvolution of the 22-immune-cell panel showed that high NETsLnc scores correlated with greater M0 macrophage infiltration, while M1 macrophage infiltration, resting mast cells, and various immunomodulatory markers were more abundant in the low NETsLnc group. The high NETsLnc group also showed greater intratumor heterogeneity, a recognized contributor to treatment resistance. The low NETsLnc group had lower leukocyte fraction and lymphocyte infiltration signature scores by immune cell scoring algorithms.
Given that the low NETsLnc group showed enrichment of immune-active markers and higher immune cell infiltration, the authors tested whether NETsLnc risk scores could predict response to actual immunotherapy treatment. Two independent immunotherapy datasets were used for validation: the IMvigor210 cohort (bladder cancer patients receiving atezolizumab anti-PD-L1 therapy) and the Liu David dataset (metastatic melanoma patients receiving immune checkpoint blockade). NETsLnc scores were computed for each patient in these datasets, and survival and response outcomes were compared between high and low NETsLnc score groups.
Immunotherapy response validation: In both the IMvigor210 and Liu David cohorts, patients with low NETsLnc scores had significantly improved overall survival compared to those with high NETsLnc scores. Critically, NETsLnc scores were significantly higher in patients with stable disease (SD) and progressive disease (PD) compared to those achieving complete response (CR) or partial response (PR). The Tumor Immune Dysfunction and Exclusion (TIDE) algorithm and the Submap pathway mapping approach both independently supported these findings: patients with low NETsLnc scores were predicted to have higher response rates to anti-PD-1 therapy in the TCGA and TARGET cohorts. The presence of elevated PD-1, PD-L2, LAG-3, and TIM3 immune checkpoint expression in the low NETsLnc group explains why these patients may particularly benefit from checkpoint blockade, as the immune cells are present but suppressed by these braking molecules.
Drug sensitivity screening: The Connectivity Map (CMap) database was queried for compounds whose gene expression signatures oppose the NETsLnc high-risk expression profile, identifying 10 candidate therapeutic agents. The top three ranked drugs were VEGF-receptor-2-kinase-inhibitor-IV, teniposide, and SIB-1893. CTRP dataset analysis of hundreds of cancer cell lines identified BI2536, GSK461364, KX2-391, paclitaxel, SB-74392, and vincristine as showing significantly lower AUC values (indicating greater drug sensitivity) in the high NETsLnc group. PRISM dataset analysis identified echinomycin, LY2606368, vincristine, and YM-155 as similarly more effective against the high-NETsLnc phenotype. Vincristine appeared in both CTRP and PRISM results, supporting its potential as a particularly relevant agent for high-risk STS patients, consistent with its existing role as a first-line antineoplastic drug for sarcoma.
The authors acknowledge several meaningful limitations. First and most importantly, all patient cohorts used in this study are retrospective, drawn from existing databases rather than prospectively enrolled. This introduces selection biases inherent to database-derived samples, including non-uniform treatment histories, variable data completeness, and the absence of standardized specimen collection protocols. Prospective cohort validation, ideally including patients treated under a defined protocol, would be necessary before the NETsLnc signature could be considered for clinical use.
Outcome measure mismatch: A specific technical limitation is that the GEO validation cohorts GSE21050 and GSE71118 report disease-free survival rather than overall survival. Since the model was trained on OS in TCGA, this outcome mismatch explains the near-random-chance AUC values (0.47-0.52) observed in these cohorts, and cannot be interpreted as evidence that the model fails to generalize. Curating STS datasets that uniformly report OS and have sufficient follow-up time would strengthen future validation.
Limited in vitro and in vivo mechanistic validation: The study provides RT-qPCR confirmation that 8 of the top signature NETsLnc (including LINC00491, LINC00703, ARRG1, JARID2-AS1, DDC-AS1, MCHR2-AS1, LINC00330, and TTTY13) show significant differential expression between STS cell lines (hSS-005R and SYO-1) and normal human skin fibroblasts (HSF). This validates that the signature genes are biologically active in STS cellular models. However, functional mechanistic studies, such as knockdown or overexpression experiments to test whether individual NETsLnc actually drive the observed phenotypes, have not been performed. The specific roles of these lncRNAs in STS biology remain largely unexplored.
Drug validation gap: The drug sensitivity predictions from CMap, CTRP, and PRISM are computational and based on cell line data, which is known to imperfectly recapitulate in vivo tumor biology. The identified drug candidates, particularly vincristine, have biological plausibility but require validation in STS-specific drug treatment cohorts or dedicated preclinical models. Translation of the signature into a clinically deployable assay would also require prospective assessment of which sequencing platform and normalization approach is optimal for routine clinical use.