Myxofibrosarcoma (MFS) is a soft tissue sarcoma defined by WHO classification as the myxoid variant of malignant fibrous histiocytoma (MFH) in which the myxoid component exceeds 50% of the tumor mass. The classification was first established in 2002 and revised in 2013, and MFS is now recognized as one of the most common adult soft tissue sarcomas. Histologically, it is characterized by curvilinear vessels within a variable myxoid stroma and a multinodular growth pattern of spindle to polygonal cells. Among all subtypes of soft tissue sarcoma, MFS accounts for approximately 5% of cases, making it a relatively rare but clinically significant entity.
Clinical behavior: MFS carries a notably poor prognosis, with an overall local recurrence rate of 50-60%. Tumors harbor highly complex karyotypes that often overlap with genomic aberrations seen in leiomyosarcoma (LMS) and undifferentiated pleomorphic sarcoma (UPS). The most common copy number gain or amplification in MFS occurs on chromosome 5p, present in approximately 60% of cases. Key oncogenes in this region include SKP2 and AMACR, which have been described as having cumulative effects on MFS tumorigenesis. Additional proteins including MET, CD109, and Ezrin (EZR) have been reported as biomarkers for aggressive behavior, but formal risk stratification remains poorly developed compared to other sarcoma subtypes.
The biomarker gap: Despite documented molecular heterogeneity and clinical urgency, prognostic biomarkers for MFS remain far less developed than for more common cancers. Public omics platforms such as OncoLnc, UALCAN, and Kaplan-Meier Plotter have enabled gene-survival correlation analyses across TCGA and GEO datasets for many cancer types, but prior to this publication no dedicated tool existed for MFS specifically. This study addresses that gap by constructing OSmfs, the first web-based interactive survival analysis tool built exclusively for Myxofibrosarcoma.
The authors frame their tool within a broader trend in computational oncology toward leveraging publicly available gene expression and clinical outcome databases to discover and validate prognostic biomarkers without the logistical burden of collecting new patient cohorts. Their previous tools for leiomyosarcoma (OSlms) and bladder cancer (OSblca) established the methodological template that OSmfs follows.
Building a prognostic tool for a rare disease requires aggregating data across multiple sources, since no single institution or database holds a sufficient number of cases. The authors searched TCGA (The Cancer Genome Atlas) and GEO (Gene Expression Omnibus) using the keywords "Myxofibrosarcoma" combined with "survival" or "prognosis," identifying three independent datasets with matched gene expression and clinical follow-up information. Together these datasets contribute 128 unique MFS cases, which represents a meaningful but still limited sample size given the disease's rarity.
TCGA cohort: The TCGA dataset contributes 25 MFS samples profiled by RNA sequencing, with a gender distribution of 11 male and 14 female patients. The median age is 60.0 years (SD 14.78). This cohort records five clinical endpoints: overall survival (OS), disease-free interval (DFI), progression-free interval (PFI), disease-specific survival (DSS), and progression-free survival (PFS). Seven death events are recorded in this cohort, reflecting the relatively short follow-up or small cohort size.
GSE71118 cohort: This GEO dataset (GPL570-55999 microarray platform) contributes 39 MFS samples with 10 death events. Notably, it is the only dataset in OSmfs that includes metastasis information, with 10 samples classified as metastatic MFS, allowing metastasis-free survival (MFS endpoint, not to be confused with the disease abbreviation) analysis. Gender and age data are not available for this cohort.
GSE72545 cohort: The largest contributing dataset at 64 MFS samples, profiled on GPL96-57554 microarray platforms. Gender distribution is 23 male and 41 female, with median age 63.5 years (SD 16.67). This cohort records 21 death events and supports overall survival analysis. Importantly, this dataset does not include metastasis data, but records tumor depth (52 deep, 12 superficial), a known prognostic variable in soft tissue sarcomas.
OSmfs is designed as a freely accessible web application that allows users, whether clinicians or researchers, to query the prognostic significance of any gene in the MFS datasets without requiring bioinformatics expertise. The architecture follows the same pattern established by the authors' prior tools (OSlms, OSblca) and is built on a standard three-tier web application stack.
Technical stack: The front-end interfaces are built in HTML 5.0 and hosted via Apache Tomcat on a Windows server. Server-side logic is written in Java, which controls the flow of analysis and rendering of results. Statistical analyses are handled through R, with the "Rserve" package functioning as middleware to bridge R and the Java application layer. Gene expression profiles and clinical data are stored and retrieved from a SQL Server database backend. This architecture allows real-time, on-demand survival analysis triggered by a simple gene symbol input from the user.
User interface and inputs: To run an analysis, users enter an official HGNC gene symbol (or multiple genes, one per line for gene signature analysis). They then select from four data source options: TCGA, GSE71118, GSE72545, or Combined (which pools all three cohorts with each dataset analyzed separately before pooling). Survival endpoint choices include OS, DFI, PFI, PFS, DSS, and metastasis-free survival. Users can also specify the expression cutoff method from eleven options, including upper quartile (25%), upper tertile (30%), median split (50%), or comparison of upper vs. lower quantile subgroups.
Clinical covariates and output: For TCGA and GSE72545, analyses can be further stratified by gender (all, male, female), tumor depth (all, deep, superficial), and age range, enabling multivariate analysis. For GSE71118, metastasis status can be used as a stratification variable. The output consists of Kaplan-Meier survival curves annotated with hazard ratio (HR), 95% confidence interval, and log-rank p-value, directly displayed in the browser without requiring any local software installation.
Before using OSmfs to discover new biomarkers, the authors first assessed its reliability by testing whether the tool could reproduce known associations. They searched PubMed for genes previously reported as poor prognostic markers in MFS at either the mRNA or protein level, identifying 12 such genes from prior literature. Seven of these 12 genes were successfully validated as significant prognosticators within OSmfs (p < 0.05), while the remaining five were not statistically significant in the OSmfs datasets.
Successfully validated genes (7/12): ITGA10 (Integrin Subunit alpha 10) showed OS hazard ratio of 6.66 (95% CI: 2.76-16.08, p < 0.0001) in GSE72545. CD109 showed OS HR of 5.03 (95% CI: 1.11-22.92, p = 0.0366) and PFS HR of 4.26 (95% CI: 1.40-12.96, p = 0.0105) in TCGA. CDK6 (Cyclin Dependent Kinase 6) was validated in GSE71118 for metastasis-free survival with HR of 4.14-5.04. CDKN2A, MET, CCND1 (Cyclin D1), and EZR (Ezrin) all showed significant adverse survival associations consistent with previously published reports.
Genes not confirmed (5/12): AMACR, SKP2, KRAS, EGFR, and ASS1 did not reach statistical significance in OSmfs. The authors attribute this to several methodological factors: differences in dataset size or clinical composition between prior studies and the OSmfs cohorts; the fact that OSmfs uses mRNA-level expression data whereas some prior reports were based on protein-level immunohistochemistry; and the 2013 WHO reclassification of MFS, which changed the diagnostic criteria and therefore the case composition of historical cohorts. Notably, three of the five unconfirmed biomarkers were originally reported before the 2013 WHO reclassification, making diagnostic comparability uncertain.
This validation exercise demonstrates both the utility and the known limitations of OSmfs: it reliably reproduces associations supported by transcriptomic data in its three cohorts, but may miss associations that depend on protein-level biology or that were established under a different diagnostic framework.
The core scientific contribution of OSmfs goes beyond tool development: the authors used Cox proportional hazard regression applied to all three datasets independently to identify genes with prognostic significance (p < 0.05) in each cohort, then used Venn diagram analysis (Venny 2.0.2) to find genes that were prognostically significant across all three datasets simultaneously. This three-way intersection identified seven genes as consistently and independently associated with poor prognosis in MFS, none of which had previously been reported as prognostic markers in this disease.
LYPLA1 (Lysophospholipase 1): High expression predicted worse overall survival in TCGA (p = 0.0223, HR = 5.83, 95% CI: 1.29-26.43), adverse metastasis-free survival in GSE71118 (p = 0.0108, HR = 5.17, 95% CI: 1.46-18.30), and worse OS in GSE72545 (p = 0.0067, HR = 3.34, 95% CI: 1.40-7.97). Combined analysis across all three cohorts produced p = 0.0004 with HR = 3.81, representing the most statistically robust signal. LYPLA1 was previously known to promote tumor progression in non-small cell lung cancer through depalmitoylating activity.
MMP13 (Matrix Metallopeptidase 13): MMP13 showed the most dramatic metastasis-free survival association in GSE71118 (p = 0.0003, HR = 13.13, 95% CI: 3.31-52.13), alongside significant OS associations in both TCGA (p = 0.0264, HR = 5.62) and GSE72545 (p = 0.0216, HR = 2.76). MMP13 plays a well-established role in extracellular matrix breakdown and has been associated with poor prognosis in gastric cancer, colorectal cancer, and oral squamous cell carcinoma, but had not been examined in MFS.
PLK1 and TMEM158: PLK1 (Polo Like Kinase 1) showed a remarkable HR of 19.61 (95% CI: 2.12-181.55, p = 0.0088) in TCGA, with additional significance in both GEO cohorts. PLK1 is a serine-threonine kinase critical for mitotic initiation and completion, and its overexpression has been associated with poor prognosis across multiple cancer types. TMEM158 (Transmembrane Protein 158) demonstrated an HR of 18.57 in TCGA (95% CI: 1.79-192.47) and an especially strong GSE72545 association (p < 0.0001, HR = 5.98, 95% CI: 2.50-14.30), and was previously implicated in pancreatic cancer progression via TGF-beta1 and PI3K/AKT pathways.
DBF4B, WNT5B, and RUNX2: DBF4B (DBF4 Zinc Finger B), a serine-threonine kinase linking cell cycle regulation to genome duplication, showed TCGA HR of 7.42 (p = 0.0099). WNT5B (Wnt Family Member 5B), previously reported as a prognostic biomarker in hepatocellular carcinoma, produced TCGA HR of 7.42 and GSE72545 OS HR of 5.14 (p = 0.0002). RUNX2, a transcription factor essential for osteoblast differentiation and implicated in breast cancer bone metastasis, showed TCGA HR of 8.17 (95% CI: 1.32-50.71, p = 0.0242).
To complement the Cox regression results, the authors performed receiver operating characteristic (ROC) analysis on the seven candidate biomarker genes using OS event state (alive vs. dead) as the binary outcome in both the TCGA and GSE72545 cohorts. ROC analysis provides an additional measure of how well each gene's expression level can discriminate between patients who died and those who remained alive over the follow-up period, quantified by the area under the ROC curve (AUC). An AUC of 0.5 indicates no discriminative ability, while 1.0 represents perfect discrimination.
TCGA ROC results: Among the seven genes tested in TCGA, MMP13 demonstrated the highest AUC at 0.8413, indicating strong discriminative capacity for OS outcome. WNT5B followed closely with AUC of 0.8016. PLK1 reached AUC 0.7460, DBF4B reached 0.7619, and TMEM158 reached 0.6746. LYPLA1 yielded AUC of 0.6270 and RUNX2 the lowest at 0.5952, though all seven exceeded the 0.5 threshold, confirming that each gene carries at least some predictive signal beyond chance.
GSE72545 ROC results: In the larger GSE72545 microarray cohort, the performance hierarchy shifted somewhat, reflecting dataset-specific biology and platform differences. WNT5B showed AUC of 0.7486 for the 221029_s_at probe. DBF4B reached AUC 0.6955, and LYPLA1 probes ranged from 0.6600 to 0.6667. PLK1 reached 0.6069 in this cohort. RUNX2, probed at three positions (221282_x_at, 221283_at, 216994_s_at), showed AUC values of 0.6567, 0.6401, and 0.6190, confirming modest but consistent discriminative capacity across probe sets.
The authors emphasize that ROC analysis provides complementary information to Cox regression: while Cox regression quantifies the effect size of high vs. low expression on survival time (HR), ROC analysis quantifies binary outcome discrimination regardless of time. Together, these two analytical frameworks present a more complete characterization of each biomarker's prognostic utility than either approach alone. The consistent finding that all seven genes exceed AUC 0.5 across both cohorts, despite the dataset heterogeneity, reinforces their candidacy as genuine prognostic signals rather than dataset-specific artifacts.
Each of the seven genes identified by OSmfs has established biological roles in cancer or developmental biology, lending mechanistic plausibility to their association with MFS prognosis. The authors discuss the known functions of each gene in the context of how they might contribute to aggressive sarcoma behavior.
Cell cycle and genomic stability regulators: PLK1 (Polo Like Kinase 1) is a serine-threonine kinase with established roles in initiating, maintaining, and completing mitosis. Dysfunction of PLK1 has been shown to promote cancerous transformation and drive tumor progression across multiple cancer types, and PLK1 overexpression is consistently linked to poor prognosis. DBF4B, a related serine-threonine kinase, links cell cycle regulation to genome duplication and is required for colon cancer cell proliferation and genomic stability maintenance. The strong HR values for PLK1 and DBF4B in TCGA (19.61 and 7.42, respectively) suggest that genomic instability pathways may be particularly consequential in MFS prognosis.
Extracellular matrix remodeling: MMP13 (Matrix Metallopeptidase 13) is involved in the breakdown of extracellular matrix components in both normal physiological processes and disease. Its prognostic relevance in gastric cancer, colorectal cancer, and oral squamous cell carcinoma is well documented, and its extremely high HR in GSE71118 (13.13 for metastasis-free survival) suggests it may play a particularly aggressive role in MFS metastatic progression, consistent with its known matrix-remodeling function enabling tumor invasion.
Signaling pathways: TMEM158 (Transmembrane Protein 158) promotes aggressiveness in pancreatic cancer via activation of TGF-beta1 and PI3K/AKT signaling pathways, pathways broadly relevant to sarcoma biology. WNT5B (Wnt Family Member 5B) has been implicated in oncogenesis and developmental patterning; it was previously identified as a prognostic biomarker in hepatocellular carcinoma. RUNX2, a transcription factor essential for osteoblast differentiation, also promotes breast cancer bone metastasis through integrin-alpha5 mediated colonization, raising the question of whether its expression in MFS reflects aberrant mesenchymal differentiation programs common to soft tissue tumors. LYPLA1, exhibiting both depalmitoylating and lysophospholipase activity, plays a tumor-promotor role in non-small cell lung cancer by modulating lipid metabolism.
The authors explicitly acknowledge several limitations of OSmfs that affect the strength of conclusions that can be drawn from its analyses. These limitations are inherent to the rare disease context and the reliance on publicly available data rather than purpose-built clinical trial cohorts.
Small and heterogeneous datasets: The three cohorts together contribute only 128 MFS samples, a small number by oncology standards. Sarcomas in general have low incidence relative to carcinomas, and MFS represents only about 5% of all soft tissue sarcomas, making large prospective cohorts exceptionally difficult to assemble. The TCGA cohort, with only 25 cases and 7 deaths, is particularly underpowered for robust multivariate analysis, which is reflected in the very wide confidence intervals seen for some biomarkers (e.g., PLK1 HR 95% CI: 2.12-181.55; TMEM158 HR 95% CI: 1.79-192.47). Wide CIs are not inherently invalidating but indicate the estimates carry high uncertainty and require confirmation in larger cohorts.
Platform heterogeneity: The datasets come from different technology platforms: RNA sequencing (TCGA) versus two distinct microarray platforms (GSE71118 and GSE72545). Cross-platform batch effects are a recognized concern in transcriptomic meta-analyses. The authors address this by analyzing each cohort separately rather than directly merging raw expression values, with "Combined" analysis pooling subgroups (high vs. low expression) defined within each cohort independently. While this approach mitigates the most severe batch effect problems, it cannot fully eliminate platform-specific technical variation.
Diagnostic uncertainty and classification evolution: The 2013 WHO reclassification of MFS changed the diagnostic criteria, making the case composition of datasets assembled under the older classification potentially discordant with current diagnostic standards. The authors note this as a likely explanation for why some historical biomarkers cannot be reproduced in OSmfs. As the WHO classification continues to evolve with molecular subtyping, future versions of OSmfs will need to incorporate cases diagnosed under current criteria.
mRNA vs. protein discordance: OSmfs operates entirely on transcriptomic data, but prognostic relevance at the mRNA level does not always translate to protein level, where post-translational modifications, protein stability, and subcellular localization govern actual biological function. Prospective validation of the seven novel candidate biomarkers at the protein level using immunohistochemistry on independent tissue microarrays, ideally from multi-institutional sarcoma referral centers, represents the necessary next step before any of these markers can enter clinical consideration.