Sarcomas are malignant tumors of mesenchymal origin, collectively representing about 1% of adult solid tumors across more than 70 histological subtypes. Within this already rare category, angiosarcomas (ASs) stand out as among the most aggressive, carrying notoriously poor progression-free survival (PFS) and overall survival (OS). ASs account for 1 to 2% of all soft tissue sarcomas and are most common in older adults (60 to 80 years old). The skin and subcutaneous tissue are the predominant site (50 to 60% of cases), with more than half of those occurring on the scalp and face. Mammary angiosarcomas are less frequent (5 to 10%), typically arising as a complication of prior radiotherapy for breast cancer. Visceral forms (hepatic, splenic) are rare, representing under 5% of cases.
Primary vs. secondary forms: ASs are classified as primary (arising de novo) or secondary. Secondary ASs often develop in the context of prior radiotherapy or chronic lymphedema, and this secondary group is characterized by distinct molecular alterations, most notably MYC amplification. Advanced or metastatic ASs respond poorly to conventional cytotoxic therapies, creating an urgent need for alternative treatment strategies.
The immunotherapy opportunity: Immune checkpoint inhibitors (ICIs) have transformed treatment for multiple solid tumors. Recent studies suggest that a subset of ASs, particularly those located in the head and neck region or lacking MYC amplification, may harbor a more immunogenic tumor microenvironment (TME), potentially making them candidates for ICI-based therapies. However, no established biomarker or scoring system has existed to prospectively identify which angiosarcoma or sarcoma patients are most likely to respond to immunotherapy.
This 2025 study, published in Cancers, addresses this gap by developing and validating two complementary immune-based tools: the AS score (a machine learning-derived prognostic score built from immune gene expression) and the Immunoscore (a measure of immune cell infiltration estimated from transcriptomic data). The combined use of these two scores is proposed as an ancillary framework for identifying sarcoma patients who may benefit from immunotherapy.
The study was structured as a two-phase validation analysis. The training cohort consisted of 25 formalin-fixed, paraffin-embedded (FFPE) angiosarcoma tissue samples collected from patients diagnosed between 2000 and 2015, retrieved from pathology archives across multiple regional hospitals. Both cutaneous and non-cutaneous forms (including soft tissue and visceral ASs) were represented. Patient-level clinical data captured included age, sex, tumor size, anatomical location, prior medical history, treatment modalities (surgery, chemotherapy, radiotherapy), surgical margin status, and clinical outcomes (alive or deceased from disease). IRB approval was obtained in accordance with the Declaration of Helsinki.
Validation cohort: For external validation, the authors applied the AS score to the soft tissue sarcoma cohort from The Cancer Genome Atlas (TCGA), comprising 253 adult cases. The TCGA sarcoma cohort spans six histological subtypes: dedifferentiated liposarcoma, leiomyosarcoma, undifferentiated pleomorphic sarcoma, myxofibrosarcoma, malignant peripheral nerve sheath tumor (MPNST), and synovial sarcoma. Because TCGA does not contain a dedicated angiosarcoma series, this broader sarcoma cohort served as a surrogate validation set, deliberately testing whether the AS score has prognostic generalizability beyond vascular sarcomas. Transcriptomic and clinical data were accessed via the publicly available Firebrowse platform.
Gene expression platform: For the training cohort, gene expression was quantified using the HTG EdgeSeq Precision Immuno-Oncology Assay (HTG Molecular Diagnostics, Tucson, AZ), a targeted panel that measures 1,392 immune-related mRNAs spanning the immune contexture of the TME. Expression values were normalized within and between samples using median-of-ratios normalization via DESeq2 (v1.42.1), reducing technical variability while preserving biologically meaningful differential expression. For the TCGA validation, gene expression data were retrieved as raw read counts and median-normalized using the same DESeq2 framework to enable cross-cohort comparability.
The AS score construction followed a multi-step analytical approach designed to reduce dimensionality from 1,392 candidate immune genes to a parsimonious, clinically applicable model. In the first step, the Maxstat algorithm (R package Maxstat, v0.7-25) was applied to each gene individually. Maxstat determines the optimal expression cut-off for each gene by maximizing the log-rank test statistic for overall survival, identifying the threshold that best separates patients into distinct survival groups. This approach avoids arbitrary median-based dichotomization and directly prioritizes genes with the strongest univariate associations with patient outcomes.
Gene selection pipeline: A stringent p-value threshold of 0.002 (from the log-rank test) was applied to select the ten most statistically significant genes from the initial 1,392. These ten top-ranked candidates were then entered into a multivariate Cox proportional hazards regression to assess their independent prognostic contributions and derive regression coefficients (weights). Four genes survived this multivariate selection: IGF1R (insulin-like growth factor 1 receptor), MAP2K1 (MEK1, a MAPK signaling kinase), SERPINE1 (plasminogen activator inhibitor-1, PAI-1), and TCF12 (transcription factor 12, a helix-loop-helix protein involved in immune cell differentiation). Each of these genes has documented roles in tumor immunology and progression.
Score formula and cut-off: The AS score was defined as the weighted sum of the normalized expression values of these four genes, with each weight corresponding to the gene's Cox regression coefficient. A cut-off value of -1.9525 was determined as the value that best stratified the training cohort by OS (again using the Maxstat approach). Patients above this threshold were classified as high-risk; those below as low-risk. For the TCGA validation cohort, the cut-off was re-established at 0.85 to account for differences in normalization and cohort composition, preserving the model's interpretive framework while adapting to the independent dataset.
Immunoscore derivation: The Immunoscore was computed using the ESTIMATE R package (v1.0.13), which infers the level of immune cell infiltration within the tumor from gene expression signatures. ESTIMATE generates an immune infiltration score reflecting the relative abundance of immune cells in the TME. Immunoscore values were normalized and dichotomized into high and low groups using the Maxstat algorithm relative to OS. All analyses used RStudio v4.3.3, with nonparametric tests (Wilcoxon rank-sum, Kruskal-Wallis) for continuous variables, Kaplan-Meier curves with log-rank tests for survival, and multivariate Cox models for independent prognostic assessment.
Applied to the 25-case angiosarcoma training cohort, the AS score separated patients into two prognostically distinct groups with strong statistical significance (p = 0.00012). Patients with high AS scores, which included both cutaneous and non-cutaneous localizations, exhibited significantly poorer overall survival. In contrast, low-score patients were predominantly cutaneous angiosarcomas, a subtype generally regarded as somewhat more favorable than deep soft tissue or visceral forms. This initial stratification demonstrated that an immune gene expression signature can capture clinically meaningful prognostic heterogeneity even within a small, histologically defined cohort.
Univariate analysis: In univariate Cox regression, three clinical variables reached statistical significance for OS: primary tumor site (cutaneous vs. non-cutaneous; p = 0.0012), surgical margin status (p = 0.0049), and receipt of chemotherapy (p = 0.0003). These findings are consistent with the established clinical literature on angiosarcoma, where anatomical location and the ability to achieve clear surgical margins are key determinants of survival.
Multivariate independence: Critically, when the AS score and the three significant clinical variables were entered simultaneously into a multivariate Cox model, the AS score emerged as the only independent predictor of poor prognosis (HR = 7.0; 95% CI: 2.4 to 21; p = 0.000444). Tumor site, surgical margins, and chemotherapy all lost statistical significance in this multivariate context. A hazard ratio of 7.0 indicates that patients in the high-AS-score group faced approximately seven-fold higher risk of death compared to low-score patients after accounting for conventional clinicopathological features. This finding substantially elevates the clinical relevance of the AS score, demonstrating that it captures prognostic information orthogonal to what standard clinical variables provide.
The AS score was applied to the 253-case TCGA sarcoma cohort, spanning six histological subtypes well beyond the angiosarcoma subtype used for model development. Using the recalibrated cut-off of 0.85, the AS score significantly distinguished two prognostic groups (p = 0.0006), with high-risk patients consistently showing shorter survival. The replication of prognostic stratification in an independent cohort of histologically diverse sarcomas is a meaningful validation milestone, demonstrating that the four-gene immune signature captures biology that is relevant across mesenchymal malignancies rather than being specific to vascular tumors alone.
Immunoscore performance: Independently of the AS score, the Immunoscore (derived from the ESTIMATE algorithm) also demonstrated significant prognostic value in TCGA sarcoma (p = 0.0029). Patients with a high Immunoscore, reflecting greater immune cell infiltration in the TME, showed significantly improved overall survival compared to those with a low Immunoscore. This aligns with the broader oncology literature: tumors with substantial immune infiltration (often described as "immune-hot") generally respond better to immune checkpoint inhibition and tend to have more favorable natural histories in some histological contexts.
Correlation between scores: A notable finding was a significant positive correlation between the AS score and the Immunoscore when treated as continuous variables (p = 2.9 x 10^-8). Counterintuitively, this means that tumors with greater immune infiltration (high Immunoscore) also tended to carry higher AS scores, which predict poor outcomes. This apparent paradox reflects the existence of a biologically distinct tumor subgroup: "immune-hot, high-risk" sarcomas that are simultaneously immunologically active and clinically aggressive. The authors hypothesize that these tumors, despite poor prognosis under standard therapy, may harbor actionable immunogenicity making them prime candidates for immune checkpoint inhibition.
The most clinically actionable result of this study is the combined stratification analysis, in which patients were divided into four groups based on their categorical AS score and Immunoscore status: Group 1 (high AS score / high Immunoscore), Group 2 (high AS score / low Immunoscore), Group 3 (low AS score / high Immunoscore), and Group 4 (low AS score / low Immunoscore). The four-group combined model demonstrated robust prognostic stratification across the TCGA cohort (p = 0.00021), substantially refining survival separation compared to either score alone.
Context-dependent prognostic impact of the AS score: A key finding was that the prognostic power of the AS score was not uniform across all patients but rather was conditional on immune context. Among patients with a low Immunoscore (immune-cold tumors), the AS score did not significantly stratify outcomes. In contrast, among tumors with a high Immunoscore (immune-hot tumors), the AS score showed very strong prognostic discrimination (p < 0.0001). This interaction indicates that the four-gene AS score functions as a prognostic and potentially predictive biomarker specifically within the immunologically active tumor compartment.
Group 1 as an immunotherapy candidate pool: Patients in Group 1 (high AS score / high Immunoscore) carry both the worst prognosis and the most immunologically active TME. The authors argue that this combination represents a window of therapeutic opportunity: these tumors already exhibit immune cell infiltration that checkpoint inhibitors might be able to enhance, yet conventional treatment is failing them. Identifying this subgroup prospectively using the combined scoring framework could enrich clinical trial populations for immunotherapy studies in sarcoma, increasing the statistical power to detect meaningful responses in a disease where immunotherapy data remain limited.
Multivariate independence of the combined model: Multivariate Cox regression including both the AS score and Immunoscore as covariates confirmed that each score provides independent prognostic information. In an extended multivariate model adjusting for age, sex, tumor size, mitotic rate, radiotherapy, and surgical margin status in TCGA, both the AS score (HR = 2.47; 95% CI: 1.42 to 4.30; p = 0.0014) and the Immunoscore (HR = 0.42; 95% CI: 0.25 to 0.70; p = 0.0009) retained statistical significance, while none of the clinical variables remained significant in the model.
Small, retrospective training cohort: The AS score was developed on only 25 angiosarcoma cases, a sample size constrained by the rarity of the disease. Retrospective designs introduce well-known biases, including selection effects, incomplete follow-up, and variability in treatment practices over the 15-year accrual window (2000 to 2015). The small training cohort limits the power of the multivariate Cox regression and increases the risk that the selected gene panel, while biologically plausible, reflects a degree of overfitting to a specific institutional patient population.
No dedicated AS validation cohort in TCGA: The external validation relied on the TCGA sarcoma cohort, which does not include angiosarcoma as a represented subtype. Using a six-subtype soft tissue sarcoma cohort as a surrogate for angiosarcoma-specific validation is a meaningful limitation because the immune biology of leiomyosarcoma, liposarcoma, or synovial sarcoma may differ substantially from that of vascular tumors. The authors acknowledge this, noting that the AS score was applied without histology-specific recalibration of cut-off thresholds, which may affect performance across biologically heterogeneous sarcoma subtypes.
No immunotherapy outcome data: While the study proposes that Group 1 (high AS score / high Immunoscore) patients are candidates for immunotherapy, no patients in either cohort received immune checkpoint inhibitors in a controlled manner. The designation of this subgroup as "immunotherapy-eligible" is thus a hypothesis grounded in biological reasoning rather than empirical outcome data from treated patients. This is a critical gap because the correlation between immune-hot TME and ICI response is well-established in some tumor types but remains poorly characterized in sarcoma.
Conventional statistical methods vs. deep learning: The authors explicitly note that while their framework uses rigorous conventional statistical tools (Maxstat, Cox regression, ESTIMATE), more advanced deep learning approaches applied to the same transcriptomic data could potentially uncover richer representations of immune microenvironment biology. The study authors identify this as a direction for future work, acknowledging that neural networks and attention-based architectures may extract non-linear patterns that Cox regression cannot capture from high-dimensional immune gene expression data.
Prospective validation in dedicated angiosarcoma cohorts: The authors identify prospective validation in independent, histology-specific angiosarcoma series as the highest-priority next step. Given the rarity of ASs, this will require multi-institutional collaboration and data sharing. Prospective application of the AS score and Immunoscore at diagnosis, with subsequent tracking of treatment outcomes including ICI response rates, would directly test whether the Group 1 designation carries predictive value for immunotherapy beyond its demonstrated prognostic value.
Subtype-specific recalibration: The AS score was validated in a mixed sarcoma cohort without histology-specific threshold adjustment. Future work should recalibrate the model separately within leiomyosarcoma, undifferentiated pleomorphic sarcoma, and other common subtypes to assess whether the four-gene immune signature retains prognostic value uniformly or requires subtype-specific tuning. This is particularly relevant for subtypes such as MPNST and synovial sarcoma, which have distinct immune contextures compared to angiosarcoma.
Deep learning and advanced machine learning: The authors explicitly call for future studies to apply deep learning to the same transcriptomic data. Architectures such as attention-based neural networks, graph neural networks modeling gene co-expression relationships, or transformer-based models pre-trained on large multi-cancer transcriptomic datasets could capture immune biology features beyond the linear Cox framework. Applying such approaches to the 1,392-gene HTG EdgeSeq panel might reveal more nuanced immune gene co-expression modules or interaction terms that further refine patient stratification.
Clinical trial enrichment: The combined AS score and Immunoscore framework is positioned as a patient selection tool for sarcoma immunotherapy trials. Given the historically poor response rates to ICI monotherapy in unselected sarcoma populations (response rates around 10 to 18% in the SARC028 trial with pembrolizumab), enriching trial populations using the Group 1 classification could meaningfully increase the probability of detecting clinical benefit. Integration of the combined scoring system into future Phase 2 or basket sarcoma immunotherapy trials represents a concrete translational pathway from this exploratory study to clinical impact.