Sarcomas are malignant tumours of soft tissue and bone that collectively affect adults, adolescents, and children, but each individual subtype is rare. There are roughly 100 recognized histological subtypes under the WHO classification, and many of them look similar under the microscope while responding to very different treatments. This morphological overlap is the central clinical problem: a pathologist in a regional hospital who sees one synovial sarcoma per year cannot accumulate the same pattern-recognition fluency as a sarcoma specialist at a referral center. Population-level studies have documented discrepancy rates of 14% or more when sarcoma diagnoses made at primary institutions are reviewed by specialist pathologists, meaning a meaningful proportion of patients receive the wrong diagnosis from the start.
The molecular testing gap: For roughly half of sarcoma entities, pathologists can fall back on characteristic molecular alterations, most commonly chromosomal translocations that generate diagnostic gene fusions. Techniques like FISH (fluorescence in situ hybridization) or RNA-based fusion detection can confirm entities such as synovial sarcoma (SS18-SSX fusions) or Ewing sarcoma (EWSR1 fusions). However, approximately half of sarcoma entities lack such definitive molecular hallmarks. Even in fusion-defined tumours, technical failures, poor RNA quality in formalin-fixed paraffin-embedded (FFPE) tissue, or unusual fusion partners can prevent a clean molecular diagnosis. This diagnostic vacuum affects treatment selection, trial eligibility, and prognosis.
Inspiration from brain tumour pathology: The authors drew directly on a successful precedent: DNA methylation-based classification of brain tumours. The Heidelberg brain tumour classifier, applied across roughly 3,000 cases in the initial study, demonstrated that individual tumour types have stable, reproducible epigenetic signatures that can be read from array data and fed into a machine learning algorithm. Importantly, it identified entirely new tumour entities that had been misclassified by conventional methods. The current paper attempts to replicate this strategy across the heterogeneous landscape of soft tissue and bone sarcomas.
The classifier's foundation is a curated reference cohort of 1,077 tumour samples drawn from cases that had received rigorous expert pathological diagnosis combined with entity-specific molecular testing whenever feasible. The goal was to anchor each methylation profile to a diagnostically confirmed ground truth. Specimens were predominantly formalin-fixed paraffin-embedded (FFPE) tissue blocks, which is the standard archival format in clinical pathology and is notoriously challenging for most molecular assays, but is compatible with Illumina methylation arrays. DNA methylation was measured using either the Infinium HumanMethylation450K BeadChip or the newer EPIC array, which covers roughly 850,000 CpG sites across the genome.
Quality control and probe filtering: Raw IDAT files were processed with the minfi Bioconductor package. Each sample underwent background correction (shifting the 5th percentile of negative control probe intensities to zero) and dye-bias correction. A batch correction step using the limma package's removeBatchEffect function addressed systematic differences introduced by tissue type (FFPE versus frozen) and array platform (450K versus EPIC). Probes targeting sex chromosomes (n = 11,551), probes with common SNPs within 5 base pairs of the CpG site (n = 7,998), and probes mapping non-uniquely to the hg19 reference genome were removed, leaving 428,230 probes for downstream analysis. Samples in the lowest 10th percentile for bisulfite conversion controls or with noise levels above 3 were excluded.
From profiles to classes: Unsupervised analysis used two complementary methods. Hierarchical clustering was computed from the 20,000 most variably methylated CpG sites (by median absolute deviation), using Euclidean distance and Ward's linkage. t-SNE (t-Distributed Stochastic Neighbour Embedding) was computed from the 10,000 most variable probes using 3,000 iterations and a perplexity of 30. Tumour entities that clustered together consistently were designated methylation classes. A minimum of seven cases per class was required to define a stable group, which proved empirically sufficient to train the downstream classifier. This analysis produced 62 tumour methylation classes plus three non-neoplastic control classes (skeletal muscle, reactive soft tissue, and leukocytes).
WHO relationship mapping: The 62 tumour classes were assigned to one of four relationship categories relative to WHO classification. Category 1 (equivalent to a WHO entity): 48 classes, covering 45 mesenchymal entities plus cutaneous melanoma, cutaneous squamous cell carcinoma, and Langerhans cell histiocytosis. Category 2 (subgroup within a WHO entity): 9 classes, including four chondrosarcoma subgroups and rhabdomyosarcoma with MYOD1 mutation. Category 3 (combining WHO entities): 3 classes, such as atypical fibroxanthoma combined with pleomorphic dermal sarcoma. Category 4 (novel entities not yet WHO-defined): 2 classes, a CNS sarcoma with rhabdomyoblast-like morphology and DICER1 mutations, and an MPNST-like entity retaining H3K27 trimethylation.
The classification algorithm is a Random Forest, a machine learning method that generates an ensemble of binary decision trees. In this implementation, 10,000 decision trees were trained, each incorporating genome-wide methylation data from all 1,077 reference samples across the 65 classes (62 tumour classes plus 3 controls). The input features were the 10,000 CpG probes with the highest variable importance as determined by the Random Forest itself during training. To address the inherently unequal class sizes (some sarcoma subtypes are far more common than others), downsampling was applied so that larger classes could not dominate the tree-building process.
Raw scores to calibrated scores: Each tree assigns an incoming sample to one of the 65 classes. Aggregating over 10,000 trees produces raw probability scores per class. Raw scores are internally consistent within a single run but are not directly comparable across classes because of differences in class size and separability. The authors applied L2-penalized multinomial logistic regression (using the glmnet R package) to recalibrate these raw scores into calibrated scores that represent more accurate class probabilities and are inter-class comparable. The calibration model was trained on cross-validation-generated scores, not re-using training data, to avoid systematic bias. A nested threefold cross-validation loop was embedded within the outer threefold cross-validation to generate the independent scores needed for calibration model training.
Methylation class families: The main sources of misclassification in the cross-validation occurred within seven groups of biologically related tumours. Conventional chondrosarcoma split into four methylation subclasses that could be confused with one another, and three BCOR-altered sarcoma classes had overlapping profiles. To handle these edge cases, the authors defined methylation class families (MCFs) by summing calibrated scores within a group. This allowed the algorithm to declare a confident family-level call even when subtype resolution was ambiguous.
Cross-validation performance: Threefold cross-validation on the 1,077-sample training set produced an estimated error rate of 1.95% on raw scores, reducing to 0.65% when using calibrated scores with the MCF correction. The multi-class area under the ROC curve was 0.99 and the Brier score was 0.05. A cut-off of calibrated score 0.9 was established as the threshold for confident class assignment, optimized by tuning the L2 penalty parameter to maximize the Youden index at this threshold across 500 resampling iterations. Any sample scoring below 0.9 is flagged as non-classifiable rather than forced into a category.
To test how the classifier performs outside its training environment, the authors assembled an independent validation cohort of 428 cases. These cases came predominantly from four precision oncology trials: MNP2.0, PTT2.0, INFORM, and NCT MASTER, all focused on relapsed or refractory tumours with extensive molecular profiling. Because these trials enroll patients based on disease severity and molecular testing rationale, the validation cohort skewed toward high-grade and diagnostically challenging cases rather than representing the typical sarcoma population. This selection bias is important context when interpreting the results.
Classification rate: Of the 428 validation cases, 322 (75%) received a calibrated classifier score of 0.9 or higher, meaning the algorithm made a confident call. The remaining 106 cases (25%) fell below the 0.9 threshold and were considered non-classifiable. t-SNE visualization showed that most non-classifiable cases projected into the periphery or outside the reference methylation class clusters, consistent with either unusual sarcoma biology not covered by the training set or the presence of subtypes that had not been included.
Concordance with institutional diagnosis: Among the 322 confidently classified cases, the predicted methylation class matched the institutional diagnosis in 263 cases (82% of classified cases; 61% of the full cohort of 428). In 59 cases (18% of classified), the classifier disagreed with the institutional pathologist. These 59 discordant cases were then subjected to additional molecular validation using RNA sequencing and whole-exome sequencing or gene panel NGS, depending on the trial protocol.
Breakdown of discordant cases: Of the 59 discrepant predictions, 29 cases were reclassified in favour of the methylation class prediction after molecular validation confirmed entity-specific alterations matching the classifier's call. In 26 cases, the discrepancy could not be resolved because no entity-specific molecular alteration was identifiable to serve as an arbiter. Four cases were labelled misleading: the institutional diagnosis was retained despite a confident classifier prediction, meaning the classifier was wrong. All four misleading cases passed quality control, and the mechanism of the error remains unclear. This gives an effective misleading rate of roughly 1% of the full validation cohort.
One underappreciated feature of high-density DNA methylation arrays is that they simultaneously generate copy number information at genome-wide resolution. The intensity values across the 428,000+ probes reflect both methylation state and probe-level coverage, allowing inference of large-scale chromosomal gains and deletions from the same experiment that produces the methylation profile. The authors used the conumee R package with additional baseline correction to generate copy number variation (CNV) plots for all reference and validation cases, producing a second diagnostic layer derived from no additional sequencing cost or tissue requirements.
Entity-specific CNV signatures: Specific alterations recurrently associated with sarcoma entities were detectable in the CNV plots. MDM2 amplification on chromosome 12q, a hallmark of well-differentiated and dedifferentiated liposarcoma, appeared in the expected class. MYC amplification on chromosome 8q was present in radiation-induced angiosarcoma. Segmental deletions on chromosome 22q encompassing SMARCB1, characteristic of malignant rhabdoid tumours, were clearly visible. These patterns align with the entities identified by the methylation classifier, reinforcing confidence in borderline cases where the calibrated score approaches but does not clearly exceed the 0.9 threshold.
Limitations of CNV interpretation: While these alterations are characteristic of specific entities, they are not pathognomonic in isolation. MDM2 amplification, for example, occurs predominantly in liposarcoma but can appear in other settings. The clinical value of CNV data therefore lies in the combination with methylation class assignment: a classifier call for well-differentiated liposarcoma accompanied by MDM2 amplification on the CNV plot is far more compelling than either data point alone. Summary CNV profiles for each of the 62 methylation classes were compiled and made available through the online platform.
Clinical report format: The complete output of the analysis, including calibrated class scores, methylation class family scores, and the CNV plot, is packaged as a molecular classifier report. This report is designed to be interpretable by clinicians and pathologists without bioinformatics training, placing the quantitative confidence metrics alongside a visual representation of the copy number landscape. The classifier platform accepts uploads of locally generated IDAT files and returns this report automatically.
DNA methylation classifiers read the average methylation pattern across all cells in a sample. When a biopsy contains a mixture of tumour cells and non-neoplastic inflammatory cells, stromal fibroblasts, or normal tissue, the measured profile represents a composite of all those cell populations, not just the tumour. Sarcomas frequently have high inflammatory infiltrates, making this contamination problem particularly relevant. The authors used the RF_Purify R package to estimate tumour cell purity from the methylation array data itself, leveraging the fact that normal and tumour cells have systematically different methylation landscapes.
Purity comparison between classifiable and non-classifiable cases: The mean estimated tumour cell purity in the classifiable cases was 51.3%. For the 106 non-classifiable cases (calibrated score below 0.9), the mean purity was 47.4%, a slightly lower figure that does not fully explain the classification failure on its own. However, the effect of purity is not uniform across entities: for high-grade osteosarcoma, for instance, plotting calibrated score against tumour purity showed a clear positive correlation, with samples above roughly 50% purity reliably achieving classification and those below frequently failing. Other sarcoma subtypes may have different purity thresholds.
Alternative explanations for non-classifiable cases: The authors are careful not to attribute all failures to purity. Some non-classifiable cases carried unique molecular alterations, such as ONECUT1-NUTM1 or EWSR1-TFCP2 gene fusions, that define entities not yet represented in the training set. These cases may be biologically distinct entities rather than technical failures. The practical recommendation from the training data is to aim for 70% or more tumour cell content in reference-quality samples, achievable by selecting representative regions on H&E-stained sections and taking targeted punches from corresponding FFPE blocks.
Potential mitigation strategies: One approach discussed is computational subtraction of methylation patterns characteristic of lymphocytes and other common non-tumour cell types, effectively deconvoluting the mixed profile to amplify the tumour-specific signal. This is technically feasible and has been applied in other epigenomic contexts but requires careful validation before deployment in the sarcoma classifier. Larger datasets enabling subtype-specific purity thresholds represent another avenue for improving classification rates in samples with moderate tumour content.
The classifier does not cover the full spectrum of sarcoma subtypes. The 62 tumour methylation classes represent a broad but incomplete range. Entities with very few documented cases cannot generate the minimum seven examples required to define a stable methylation class, and very rare subtypes may have been absent from the training data entirely. This is a structural limitation of any supervised or semi-supervised classifier: its coverage is bounded by what it has been trained on. Cases outside the coverage of the classifier will fall below the 0.9 threshold and receive no prediction, which is the correct behaviour, but it means the tool remains an adjunct rather than a comprehensive replacement for expert pathological assessment.
Validation cohort selection bias: The 428-case validation cohort was drawn from precision oncology trials that specifically enrolled relapsed, refractory, and molecularly unusual cases. These are systematically harder to classify than newly diagnosed sarcomas in a general population. The 75% classification rate and the 82% concordance among classified cases should be interpreted in that context. In a prospective study enrolling consecutive newly diagnosed sarcomas with adequate tumour content, performance would likely be higher on both metrics.
No centralized pathological review: The validation cohort did not undergo a central expert pathological review before comparison with the classifier output. The authors note that a French Sarcoma Group study documented a 14% reclassification rate after central expert review. If the institutional diagnoses used as the gold standard in this study contain that level of error themselves, the true concordance of the classifier with ground-truth pathology may be higher than the reported 82%, and some of the 26 unresolved discordant cases may actually represent classifier-correct, institution-incorrect calls.
Platform and regulatory considerations: The classifier requires Illumina 450K or EPIC array data. This places a cost and infrastructure requirement on institutions that may not routinely perform methylation arrays. RNA expression profiling, a competing technology, requires fresh or high-quality frozen tissue and would not work on most archival FFPE samples. The methylation platform's FFPE compatibility is a major practical advantage, but acquisition of array-quality DNA from small biopsies with high necrosis or decalcification (common for bone sarcomas) can be challenging. The online platform at www.molecularsarcomapathology.org accepts IDAT file uploads but operates as a research tool, not a CE-IVD or FDA-cleared diagnostic device.
The authors released the classifier as an open-access web platform at www.molecularsarcomapathology.org. Pathologists and researchers can upload locally generated IDAT files and receive back a molecular classifier report containing calibrated prediction scores for each methylation class, the methylation class family score if applicable, and a copy number variation plot. The report is formatted for clinical communication, displaying prediction confidence numerically. This design mirrors the architecture of the Heidelberg brain tumour classifier platform, which has been used in thousands of cases worldwide and has materially influenced clinical management decisions.
Iterative training improvement: The platform is designed to be a living classifier. Cases uploaded by contributing institutions, subject to permission and data sharing agreements, will be incorporated into future training iterations. This creates a feedback loop: the more cases are submitted from diverse institutions, the broader the coverage becomes, and rare entities that currently lack sufficient representation will eventually accumulate enough profiles to define new methylation classes. The precedent from brain tumour classification, where entities like pilocytic astrocytoma and medulloblastoma have progressively subdivided into more refined methylation subgroups as data accumulates, illustrates how this iterative approach can deepen biological understanding alongside classification accuracy.
Expanding entity coverage and subtype resolution: The chondrosarcoma example is instructive: conventional chondrosarcoma already splits into four methylation subclasses in the current classifier. Larger datasets may reveal additional substructure within entities currently treated as homogeneous, potentially identifying clinically meaningful subsets with different prognoses or treatment sensitivities. Similarly, the two novel Category 4 methylation classes identified in this study point toward entities that lack WHO recognition but may have distinct biology. Future characterization of these classes through molecular profiling and clinical outcome data could lead to revised WHO classification criteria.
Integration with other molecular data: The classifier is built on methylation alone but outputs a report alongside CNV data from the same array. Future versions could integrate RNA sequencing or targeted gene panel NGS results directly into the prediction framework, building a multi-omic model that assigns higher confidence to cases where methylation, CNV, and transcriptomic evidence converge. The trial infrastructure used for the validation cohort, particularly NCT MASTER and INFORM, already routinely generates all three data types, making this integration technically achievable within existing research settings. The ultimate goal is a fully automated, prospectively validated, and regulatory-approved diagnostic tool that functions as a second opinion engine for sarcoma pathology at institutions with limited specialist access.