Transcriptomic-Based Classification Identifies Prognostic Subtypes and Therapeutic Strategies in Soft Tissue Sarcomas

Cancers 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
Why STS Classification Needs a Molecular Overhaul

Soft tissue sarcomas (STSs) are a heterogeneous group of rare mesenchymal malignancies with a classification system that has been fragmenting the disease into 50 to 150 histological subtypes, approximately 20% of which qualify as "ultra-rare" with incidences below 1 in 1,000,000. This extreme fragmentation is not simply a taxonomic inconvenience. It directly undermines prognosis estimation, treatment planning, drug discovery, and clinical trial design, since no single institution or trial can assemble adequately powered cohorts for so many individual subtypes.

The classification problem: The current histopathological classification is fundamentally morphological, relying on the resemblance of neoplastic tissue to a presumed normal tissue counterpart. This approach is indirect and non-specific, and it carries real-world consequences: overall diagnostic discrepancy rates of 28.2 to 56% and major diagnostic discrepancy rates of 16.4 to 37% have been documented between referring and tertiary centers. In some series, histopathological reclassification altered treatment strategies in up to 15% of cases, a significant patient safety concern.

The molecular opportunity: Comprehensive genomic and transcriptomic profiling studies have already demonstrated that molecular approaches can reclassify 7% of STS histopathological diagnoses, identify treatment-relevant variants in 15% of cases, and uncover actionable molecular alterations in up to 37.2% of patients with 31.2% receiving personalized treatment based on those alterations. Prognostic molecular signatures, including CINSARC (a 67-gene expression signature related to mitosis and chromosomal integrity) and clinical nomograms (such as the SARCULATOR application), exist but have significant performance limitations.

This study, published in Cancers (2025) from a collaboration across Portuguese oncology centers, IST, F. Hoffmann-LaRoche, and Foundation Medicine, presents a novel multi-omics analysis of 102 high-grade STS samples to identify molecular subtypes with superior prognostic and predictive utility compared to existing tools.

TL;DR: Current STS classification is morphological, with diagnostic discrepancy rates of 28-56% between institutions and major reclassification in up to 15% of cases affecting treatment. Molecular approaches identify actionable alterations in 31-37% of patients. This study uses unsupervised machine learning on RNA-seq and DNA-seq data from 102 high-grade STS samples to derive a superior classification system.
Pages 3-5
Study Design: Multi-Omics Profiling with Clinically Validated Sequencing Assays

The study enrolled 101 patients diagnosed and treated at Instituto Portugues de Oncologia de Lisboa Francisco Gentil (IPOLFG), a tertiary oncological center and European sarcoma reference center, between April 2013 and September 2022. The sample pool comprised 26 dedifferentiated liposarcoma (DDLPS) samples, 25 high-grade leiomyosarcoma (LMS) samples, and 51 undifferentiated pleomorphic sarcoma (UPS) samples, all formalin-fixed paraffin-embedded (FFPE). These three histotypes represent the most common non-translocation-associated STS subtypes and account for a substantial proportion of all STS cases.

Sequencing platform: All 102 FFPE samples were characterized using FoundationOne CDx (F1CDx) for DNA sequencing (DNA-seq) and FoundationOne RNA (F1RNA) for RNA sequencing (RNA-seq), performed at a CLIA-certified, CAP-accredited, New York State-approved laboratory at Foundation Medicine (Cambridge, MA). F1CDx is a next-generation sequencing (NGS) assay detecting short variants, copy number alterations, large genomic rearrangements, and complex biomarkers including microsatellite instability (MSI) and tumor mutational burden (TMB) across 324 cancer-associated genes. F1RNA is a hybrid-capture-based targeted RNA-seq assay covering fusion and rearrangement detection for 318 genes and gene expression profiling (GEP) for 1,517 genes.

QC passage rates and downstream cohort: Of 102 samples submitted, 79 passed F1CDx quality control for DNA analysis and 75 passed F1RNA QC for RNA-seq expression analysis. One additional RNA-seq sample was excluded as an outlier on principal component analysis (PCA), yielding a final RNA-seq expression analysis cohort of 74 samples (16 DDLPS, 15 LMS, 43 UPS). For fusion/rearrangement detection specifically, only 22 of 75 RNA-seq samples passed the more stringent post-sequencing QC metrics, reflecting the challenges of archival FFPE tissue with variable degradation.

Bioinformatics pipeline: RNA-seq data from 74 samples was processed using the edgeR filterByExpression method (v4.2.1) for lowly expressed gene removal, followed by Voom normalization. Genes were then filtered based on Mean Absolute Deviation, retaining the top 55% most variable genes. Consensus clustering from the ConsensusClusterPlus package (v1.68.0) was applied, with the optimal number of clusters determined using the Elbow method. Differential gene expression analysis was conducted pairwise between clusters using the Limma package (v3.60.4), with p-values adjusted using the Benjamini-Hochberg False Discovery Rate (BH-FDR). Genomic alteration actionability was evaluated using the Molecular Tumor Board Portal (MTBP), which applies the ESMO Scale of Clinical Actionability for Molecular Targets (ESCAT) tiering framework.

TL;DR: 102 FFPE samples (26 DDLPS, 25 LMS, 51 UPS) from a European sarcoma reference center were profiled with FoundationOne CDx (324 genes, DNA-seq) and FoundationOne RNA (1,517 genes, RNA-seq). Final RNA analysis cohort: 74 samples. Consensus clustering (ConsensusClusterPlus, Elbow method) on the top 55% most variable genes identified four clusters; Limma with BH-FDR correction confirmed differential expression. Actionability was classified using ESCAT tiers via the MTBP portal.
Pages 5-8
Clinical Profile: High-Grade, Mostly Localized STS with Aggressive Outcomes

The 101 patients displayed a median age of 67 years (IQR 19.7) with a balanced gender distribution (50.5% male). Tumors were predominantly located in the lower limb (n=49, 48.5%) and retroperitoneum (n=31, 30.5%), with a median primary tumor size of 13 cm (IQR 10.0). All samples were high-grade (Grade 3). The great majority (n=96, 95%) presented with localized disease, with only 5% metastatic at diagnosis, primarily with pulmonary metastases (80% of the metastatic cases).

Treatment and surgical outcomes: Surgery was the primary treatment modality in 98% of patients (n=99). Among those operated with curative intent at IPOLFG (n=94), resection margin status was R0/R1 in 96.7% of cases. Only 3.2% received neoadjuvant treatment (two with doxorubicin-ifosfamide chemotherapy, one with external radiotherapy at 50 Gy/25 fractions), while 63% received adjuvant treatment, primarily external radiotherapy (94.8% of adjuvant cases). This treatment profile reflects current evidence-based standard of care for localized high-grade STS.

Oncological outcomes: Among the 89 patients with R0/R1 resection followed over a median of 27 months (IQR 51.3), outcomes were challenging across all dimensions. The local recurrence rate was 46.1% (n=41), with a median time to local recurrence of 14 months (IQR 29.0). Distant metastasis occurred in 41.7% (n=40) of the 96 patients without initial metastases, predominantly to the lungs (85%), with a median time to distant metastasis of 13 months (IQR 17.2). The 5-year metastasis-free survival (MFS) rate was 37% and the 5-year overall survival (OS) rate was 46%, underscoring the poor prognosis of high-grade STS and the urgent need for better prognostic and predictive tools.

Twenty-nine percent of patients (n=29) had prior local treatment at another institution before IPOLFG, reflecting the referral patterns typical of a tertiary sarcoma center. These real-world outcomes contextualize why the investigators sought a molecular classification system that could better stratify individual patient risk and guide treatment intensity decisions.

TL;DR: 101 patients, median age 67, 95% localized disease, median tumor size 13 cm, all Grade 3. Five-year OS 46%, 5-year MFS 37%. Local recurrence rate 46.1% and distant metastasis rate 41.7% over median 27-month follow-up. R0/R1 resection in 96.7%, adjuvant radiotherapy in 63%. These aggressive outcomes reinforce the clinical need for better prognostic stratification tools.
Pages 8-11
Four Transcriptomic Subtypes with Distinct Molecular Identities

Unsupervised consensus clustering of RNA-seq data from 74 samples identified four transcriptomic clusters (TCs), each with a discrete molecular identity defined by the differential expression of specific gene sets and associated biological pathways. The Elbow method confirmed four as the statistically optimal cluster number. Each cluster incorporated all three STS histotypes (DDLPS, LMS, UPS), though with varying proportional representation, establishing that the subtypes are not simply a reorganization along histological lines but represent genuinely cross-histotype molecular patterns.

Cluster 1 (C1): DNA Repair-Deficient (HRD-like/Hypermutant). C1 is primarily defined by the under expression of an extensive array of homologous recombination repair (HRR) genes, including BRCA1, BRCA2, FANCD2, PALB2, RAD51, CHEK1, and BRIP1. The over expression of CDK4 and CCND2 also characterizes this cluster. The HRR gene under expression pattern suggests a homologous recombination deficiency (HRD)-like phenotype, potentially generating chromosomal instability and hypermutability. This cluster is predominantly composed of DDLPS samples (52.4%), with UPS (28.6%) and LMS (19.0%) also represented.

Cluster 2 (C2): Cancer Testis Antigens-Enriched (Immunogenic). C2 is defined by strong over expression of multiple cancer testis antigen (CTA) genes, particularly MAGE family members (MAGEA2B, MAGEA3, MAGEA12, MAGEB1, MAGEB2, MAGEC2) and SSX family genes (SSX1, SSX2, SSX2B, SSX3). The over expression of CTNNB1 and transcriptional regulation pathways also characterizes this cluster. The MAGE and SSX gene co-over expression pattern is notable because the study cohort did not include synovial sarcoma or myxoid/round cell liposarcoma samples, subtypes classically associated with SSX expression, indicating this is an independent biological phenomenon in DDLPS, LMS, and UPS. C2 is predominantly UPS (58.3%) with LMS (29.2%) and DDLPS (12.5%).

Cluster 3 (C3): HLA-High (Immune Activated). C3 is specifically characterized by the over expression of Major Histocompatibility Complex (MHC) class II/HLA class II genes (HLA-DMA, HLA-DMB, HLA-DOA, HLA-DQA1, HLA-DRA, HLA-DRB1) alongside TGFbeta1, ETV5, BTK, and BATF. This immune activation signature is paired with under expression of CDKN1C, CDKN2A, FGFR2, and FGFR3, as well as under expression of the beta-catenin pathway. Samples are predominantly UPS (85.0%), with LMS (10.0%) and DDLPS (5.0%). Cluster 4 (C4): Claudin-High (Structural). C4 is defined by over expression of structural/adhesion proteins including claudin 4 (CLDN4), CLCA2, GAS7, SMAD3, and PDGFD, with under expression of ACTN1. The preponderance of UPS (66.7%) with LMS (22.2%) and DDLPS (11.1%) makes this cluster also cross-histotypic.

TL;DR: Four transcriptomic clusters identified by consensus clustering: C1 (HRD-like, under-expressed BRCA1/2, FANCD2, RAD51, PALB2; predominantly DDLPS), C2 (Immunogenic, over-expressed MAGE and SSX families; predominantly UPS), C3 (Immune Activated, over-expressed HLA class II genes including HLA-DRA and HLA-DRB1; predominantly UPS), and C4 (Structural, over-expressed CLDN4 and adhesion proteins; predominantly UPS). All four clusters contain representatives of all three histotypes.
Pages 11-14
Transcriptomic Subtypes as the Strongest Predictor of OS and DFS

The four transcriptomic subtypes were tested as prognostic variables within a Cox Proportional Hazards Model that also incorporated demographic, clinical, and histopathological variables (including the histopathological classification, FNCLCC grade, gender, age, distant metastasis, and treatment modality). This analysis of the study cohort identified C2, C3, and C4 as independent negative prognostic factors for overall survival, with hazard ratios of C2 (HR 5.10; 95% CI 1.81-14.34; p=0.002), C3 (HR 4.47; 95% CI 1.39-14.45; p=0.01), and C4 (HR 7.66; 95% CI 2.06-28.53; p=0.002). An Analysis of Variance (ANOVA) test applied to this Cox model confirmed that transcriptomic cluster/subtype was the variable with the most significant correlation with OS (p less than 0.01), surpassing all other variables tested.

External validation in TCGA-SARC: Independent validation was conducted using the TCGA-SARC dataset, restricted to the same three histotypes (DDLPS, LMS, UPS, n=127). Normalized gene expression data was used to assign each TCGA-SARC patient to a transcriptomic subtype via single-sample Gene Set Enrichment Analysis (ssGSEA) using the C1_under and C3_over gene signatures as reference sets. The analysis confirmed that C3-enriched TCGA-SARC patients had a significantly worse prognosis (HR 2.08; 95% CI 1.11-3.9; p=0.022), with ANOVA confirming the transcriptomic cluster classification as the most significant OS predictor (p=0.0165) in the external cohort. Median follow-up for censored patients was 37.8 months in the study cohort and 37.2 months in TCGA-SARC, supporting comparability.

DFS analysis: For disease-free survival in the study cohort, C2 (HR 3.69; 95% CI 1.33-10.20; p=0.012) and C3 (HR 3.68; 95% CI 1.15-11.77; p=0.028) were confirmed as negative prognostic factors, while neoadjuvant/adjuvant treatment was a positive prognostic factor (HR 0.32; 95% CI 0.14-0.74; p less than 0.01). The ANOVA test confirmed transcriptomic clusters as a significant DFS predictor (p=0.042) alongside adjuvant treatment (p=0.012). External validation in TCGA-SARC showed a trend toward worse DFS for C3-enriched patients (HR 1.43; 95% CI 0.96-2.1; p=0.078) with Kaplan-Meier log-rank testing confirming significantly worse DFS for C3 vs. C1 (log rank p=0.043).

Independence from UPS prevalence: A key validation step addressed the concern that the transcriptomic subtypes might merely reflect the over-representation of UPS samples (43 of 74) in the analysis cohort. When UPS patients were removed from the TCGA-SARC validation dataset, the molecular enrichment of the remaining population in C1 and C3 and the statistically significant correlation between the TC-based classification and OS were preserved, ruling out UPS "contamination" as an explanation for the findings.

TL;DR: C2 (HR 5.10, p=0.002), C3 (HR 4.47, p=0.01), and C4 (HR 7.66, p=0.002) are independent negative prognostic factors for OS in the study cohort. ANOVA confirms TC classification as the strongest OS predictor (p less than 0.01). External TCGA-SARC validation confirms C3 as a negative OS predictor (HR 2.08, p=0.022). C2 and C3 are also negative DFS predictors (HR 3.69 and 3.68, both p less than 0.03). Findings hold after removing UPS samples.
Pages 14-17
Outperforming SARCULATOR and CINSARC in Prognostic Accuracy

A central claim of this paper is that the new TC-based classification provides superior prognostic accuracy compared to the two most widely used STS prognostication tools: SARCULATOR (a clinical nomogram incorporating histological subtype, grade, size, depth, age, and anatomical site to estimate 5-year OS probability) and CINSARC (a 67-gene expression signature capturing mitotic and chromosomal instability, validated for predicting metastasis-free survival). Concordance indices (C-indexes) from Cox Proportional Hazards Models for OS were systematically compared across multiple model combinations.

vs. SARCULATOR in the study cohort: Among the 67 patients for whom SARCULATOR nomograms could be applied (median predicted 5-year OS 57%, IQR 26.5%), the C-indexes were: SARCULATOR alone (C-index 0.62), TC alone (C-index 0.63), TC combined with SARCULATOR (C-index 0.65), and TC combined with age (C-index 0.70). The TC-based model outperformed SARCULATOR despite being neither specifically designed nor trained to predict OS, and without including age as an input variable (unlike SARCULATOR, which requires age). The TC + AGE model showed the strongest prognostic accuracy overall.

vs. SARCULATOR and CINSARC in TCGA-SARC: External validation using TCGA-SARC confirmed that TC (C-index 0.61) marginally outperformed SARCULATOR (C-index 0.60) and clearly outperformed CINSARC (C-index 0.49) and CINSARC + AGE (C-index 0.53) for OS prediction. The best-performing model was TC combined with SARCULATOR and CINSARC (C-index 0.67). Critically, while CINSARC showed its known strength for MFS prediction (Kaplan-Meier log rank p=0.018 in TCGA-SARC), it failed entirely to distinguish OS profiles (log rank p=0.930), whereas the TC-based classification showed a significant OS survival difference between C1 and C3 (log rank p=0.017).

Sub-stratification within SARCULATOR-defined groups: A practically important finding was that the TC classification could sub-stratify patients within the SARCULATOR-defined unfavorable prognostic group (predicted 5-year OS below or equal to 60%), identifying C1 patients with significantly better relative outcomes than non-C1 patients (Kaplan-Meier log rank p=0.018). This capacity to refine risk within an already high-risk group has direct clinical implications for decisions about adjuvant therapy intensity and surveillance intervals.

TL;DR: TC alone beats SARCULATOR in the study cohort (C-index 0.63 vs. 0.62) and TCGA-SARC (0.61 vs. 0.60). TC clearly outperforms CINSARC (0.61 vs. 0.49 for OS; CINSARC log rank p=0.930, TC log rank p=0.017 for OS in TCGA-SARC). Best model is TC + SARCULATOR + CINSARC (C-index 0.67). TC also sub-stratifies patients within SARCULATOR's unfavorable risk group (p=0.018), a capability neither SARCULATOR nor CINSARC provides.
Pages 17-20
DNA-seq Reveals 151 Actionable Variants with Subtype-Specific Therapeutic Implications

DNA-seq data from patients in each of the four transcriptomic clusters was analyzed using FoundationOne CDx, and all identified genomic alterations were systematically tiered for actionability using the Molecular Tumor Board Portal (MTBP) following the ESCAT framework. Tier 2 designates investigational evidence (alteration-drug match with known antitumor activity), Tier 3 designates hypothetical targets with associated antitumor activity but unknown benefit magnitude, and Tier 4 designates targets with preclinical evidence only. In total, 151 gene variants classified with ESCAT tiers 2 to 4 were identified: 29 in C1, 51 in C2, 56 in C3, and 15 in C4.

Cross-cluster targets: MDM2 amplifications (conferring sensitivity to Brigimadlin and Milademetan) and TP53 alterations (mainly missense mutations conferring sensitivity to Pazopanib and Vorinostat) were found across all four clusters as Tier 2 variants. MTAP deletions (conferring sensitivity to MRTX1719 and AMG193) were present in three of four clusters. TSC2 mutations (conferring sensitivity to ABI-009) appeared in two clusters.

C1-specific targets (HRD-like): Consistent with the HRR gene under expression that defines C1, actionable alterations in HRR pathway genes were enriched: RAD51B frameshift mutations (Tier 3), ATM missense mutations (Tier 3), and BRIP1 missense mutations (Tier 3), all conferring sensitivity to the PARP inhibitor Olaparib. NF1 mutations (Tier 3 and Tier 4, conferring sensitivity to Selumetinib and Trametinib/Cobimetinib) and CDK4 amplifications (Tier 4) also characterize this cluster. C2-specific targets (Immunogenic): ERBB2 amplifications (Tier 2, conferring sensitivity to Trastuzumab Deruxtecan) represent a never-before-documented alteration in DDLPS, LMS, and UPS. RET missense mutations (Tier 3, conferring sensitivity to Selpercatinib and Pralsetinib) as tumor-agnostic targets, POLE missense mutations (Tier 3, conferring sensitivity to Pembrolizumab via hypermutation), and FGFR1 mutations (Tier 3/4) also appear uniquely in C2.

C3-specific targets (Immune Activated): C3 displays the richest actionable landscape with 56 variants. Beyond the cross-cluster targets, C3 harbors PIK3CA missense mutations (Tier 2, conferring sensitivity to Capivasertib and Copanlisib), KRAS and NRAS missense mutations (Tier 3 and 4), MET amplifications (Tier 3, conferring sensitivity to Capmatinib, Tepotinib, Telisotuzumab Vedotin, and Crizotinib), PTEN frameshift mutations (Tier 3, conferring sensitivity to Capivasertib), and VHL missense mutations (Tier 3, conferring sensitivity to Everolimus). RNA-seq also detected two fusions not identified by DNA-seq: an HMGA2::TPH2 fusion in a DDLPS case and a NOTCH3::BRD4 fusion in a UPS case, both missed because breakpoints occurred in intronic regions not covered by F1CDx.

TL;DR: 151 ESCAT-tiered actionable variants across 4 clusters: 29 (C1), 51 (C2), 56 (C3), 15 (C4). MDM2 amplification and TP53 mutations are universal. C1 harbors HRR gene mutations (RAD51B, ATM, BRIP1) predicting PARP inhibitor sensitivity. C2 uniquely harbors ERBB2 amplification (Trastuzumab Deruxtecan) and RET mutations (Selpercatinib). C3 carries PIK3CA, MET amplification, KRAS/NRAS, and PTEN mutations. RNA-seq detected two additional fusions missed by DNA-seq.
Pages 20-23
Study Limitations, Therapeutic Hypotheses, and the Road to Prospective Validation

The authors candidly enumerate several limitations. The study is retrospective and single-centered, even though IPOLFG is a European sarcoma reference center with patients of diverse ethnic backgrounds. The sample pool is restricted to three STS histotypes (DDLPS, LMS, UPS), excluding ultra-rare subtypes for which molecularly guided prognostication may be even more critical due to the impossibility of assembling large cohorts. The cohort is also predominantly (95%) composed of early-stage/localized disease, limiting the representation of metastatic or advanced STS, and none of the advanced-stage patients were included in the RNA-seq subtyping analysis.

Technical limitations: The FFPE samples were collected across a wide timespan (2013-2022) and shipped in three batches, introducing heterogeneous sample chronological ages and differential degradation risk. This contributed to the high rate of RNA-seq QC failure for clinical-grade fusion detection (53 of 75 RNA samples failed post-sequencing QC for rearrangement detection), though 73.5% passed QC for the less stringent research-grade GEP assay. F1CDx and F1RNA are targeted sequencing assays, not whole-genome or whole-exome approaches, limiting coverage. Specifically, F1RNA does not cover 32 (48%) of the 67 genes in the CINSARC signature, which is why direct CINSARC comparison in the study cohort was not feasible and had to rely entirely on the TCGA-SARC external cohort.

Therapeutic hypotheses by subtype: The molecular identities of each cluster suggest specific treatment sensitivity hypotheses. C1 (HRD-like) is predicted to show sensitivity to PARP inhibitors (Olaparib), CDK 4/6 inhibitors, and MDM2 antagonists. C2 (Immunogenic) is hypothesized to be amenable to cancer testis antigen-directed T cell receptor therapies (analogous to afamitresgene autoleucel for MAGE-A4 in synovial sarcoma) and cancer testis antigen-directed vaccines targeting the MAGE and SSX antigens uniquely over-expressed in this cluster. C3 (Immune Activated) is predicted to show increased sensitivity to immune checkpoint inhibitors and immunomodulatory agents, given its HLA class II overexpression and immune pathway enrichment. C4 (Claudin-High) may be amenable to claudin-directed agents currently in development for other cancer types.

The investigators note that a prospective multicenter clinical trial is required to validate both the prognostic utility (guiding adjuvant treatment decisions) and predictive utility (guiding treatment selection for advanced disease) of this classification. The group is actively designing such a trial. A particularly notable pre-operative applicability advantage is that TC classification requires only biopsy specimens, unlike SARCULATOR nomograms, which depend on surgical variables (tumor size measured at operation, resection margins) that are not available preoperatively, meaning TC-based risk stratification could influence decisions before surgery.

TL;DR: Limitations: retrospective, single-center, three histotypes, 95% localized disease, FFPE degradation, and targeted (not whole-genome) sequencing. Therapeutic hypotheses: C1 responds to PARP inhibitors and CDK 4/6 inhibitors; C2 to CTA-directed T cell therapies and vaccines; C3 to immune checkpoint inhibitors; C4 to claudin-directed agents. Prospective multicenter trial is the required next step. TC classification is preoperatively applicable, unlike SARCULATOR.