Machine Learning–Supported Diagnosis of Small Blue Round Cell Sarcomas Using Targeted RNA Sequencing

Lea D Schlieben; Maria Giulia Carta; Evgeny A Moskalev; Robert Stöhr; Markus Metzler; Manuel Besendörfer; Norbert Meidenbauer; Sabine Semrau; Rolf Janka; Robert Grützmann; Stefan Wiemann; Arndt Hartmann; Abbas Agaimy; Florian Haller; Fulvia Ferrazzi

doi:10.1016/j.jmoldx.2024.02.002

. 2024 Feb 21;26(5):387–398. doi: 10.1016/j.jmoldx.2024.02.002

Machine Learning–Supported Diagnosis of Small Blue Round Cell Sarcomas Using Targeted RNA Sequencing

Lea D Schlieben ^∗,^†,^‡, Maria Giulia Carta ^∗,^†,^‡, Evgeny A Moskalev ^∗,^†,^‡, Robert Stöhr ^∗,^†,^‡, Markus Metzler ^†,^‡,^§, Manuel Besendörfer ^¶, Norbert Meidenbauer ^†,^‡,^‖, Sabine Semrau ^†,^‡,^∗∗, Rolf Janka ^††, Robert Grützmann ^†,^‡,^¶,^‡‡, Stefan Wiemann ^§§, Arndt Hartmann ^∗,^†,^‡, Abbas Agaimy ^∗,^†,^‡, Florian Haller ^∗,^†,^‡, Fulvia Ferrazzi ^∗,^†,^‡,^¶¶,^∗

PMCID: PMC12178384 PMID: 38395409

Abstract

Small blue round cell sarcomas (SBRCSs) are a heterogeneous group of tumors with overlapping morphologic features but markedly varying prognosis. They are characterized by distinct chromosomal alterations, particularly rearrangements leading to gene fusions, whose detection currently represents the most reliable diagnostic marker. Ewing sarcomas are the most common SBRCSs, defined by gene fusions involving EWSR1 and transcription factors of the ETS family, and the most frequent non–EWSR1-rearranged SBRCSs harbor a CIC rearrangement. Unfortunately, currently the identification of CIC::DUX4 translocation events, the most common CIC rearrangement, is challenging. Here, we present a machine-learning approach to support SBRCS diagnosis that relies on gene expression profiles measured via targeted sequencing. The analyses on a curated cohort of 69 soft-tissue tumors showed markedly distinct expression patterns for SBRCS subgroups. A random forest classifier trained on Ewing sarcoma and CIC-rearranged cases predicted probabilities of being CIC-rearranged >0.9 for CIC-rearranged–like sarcomas and <0.6 for other SBRCSs. Testing on a retrospective cohort of 1335 routine diagnostic cases identified 15 candidate CIC-rearranged tumors with a probability >0.75, all of which were supported by expert histopathologic reassessment. Furthermore, the multigene random forest classifier appeared advantageous over using high ETV4 expression alone, previously proposed as a surrogate to identify CIC rearrangement. Taken together, the expression-based classifier can offer valuable support for SBRCS pathologic diagnosis.

Key Points

•
Small blue round cell sarcomas (SBRCSs) are a heterogeneous group of tumors with overlapping morphologic features but divergent genotypes and markedly varying prognosis.
•
The detection of recurrent gene fusions represents the most reliable diagnostic marker for SBRCS. However, it can be challenging with currently available methods, as in the case of the CIC::DUX4 translocation events, the most common CIC rearrangement in sarcomas.
•
Bioinformatics analyses of the gene expression profiles measured via targeted sequencing show that SBRCS subgroups are characterized by markedly distinct expression patterns.
•
The machine-learning classifier based on gene expression profiles that is presented here is a valuable tool to support the pathologic diagnosis of SBRCSs.

Key Points.

•
Small blue round cell sarcomas (SBRCSs) are a heterogeneous group of tumors with overlapping morphologic features but divergent genotypes and markedly varying prognosis.
•
The detection of recurrent gene fusions represents the most reliable diagnostic marker for SBRCSs. However, it can be challenging with currently available methods, as in the case of the CIC::DUX4 translocation events, the most common CIC rearrangement in sarcomas.
•
Bioinformatics analyses of the gene expression profiles measured via targeted sequencing show that SBRCS subgroups are characterized by markedly distinct expression patterns.
•
The machine-learning classifier based on gene expression profiles that is presented here is a valuable tool to support the pathologic diagnosis of SBRCSs.

Small blue round cell sarcomas (SBRCSs) are a heterogeneous group of mesenchymal neoplasms predominantly affecting children and young adults. According to the current World Health Organization classification of soft-tissue tumors (STTs), SBRCSs are composed of relatively undifferentiated cells with small blue nuclei and a round to ovoid cytomorphology in hematoxylin and eosin staining.¹ Overlapping histologic and morphologic characteristics make the routine morphology-based classification of SBRCSs into specific subgroups a diagnostic challenge. These subgroups are characterized by different progression and sometimes treatment strategies, necessitating tumor stratification. Advances in both cytogenetics and molecular genetics have led to the detection of tumor-specific alterations in subsets of SBRCSs, which are characterized by particular recurrent chromosomal rearrangements and associated specific gene fusions.² The tumorigenic activity of gene fusions is exerted either by the formation of encoded chimeric fusion proteins or by the aberrant expression of a normal gene product through hijacking of regulatory elements of the translocation partner.³^,⁴ Molecular pathologic identification of characteristic gene fusion events has gained increasing importance in sarcoma diagnostics,⁵ and RNA sequencing (RNA-seq)–based methods now are used extensively in the molecular diagnostic routine workup of STTs.

Ewing sarcomas (ESs) are the most common type of SBRCSs and harbor gene fusions involving EWSR1 and transcription factors of the ETS family, predominantly FLI1 (85%) and ERG (10%). In contrast, atypical ES or Ewing-like sarcomas are a group of undifferentiated round cell sarcomas that resemble classic ES morphologically to a variable extent, but lack the molecular hallmark for ES, the EWSR1::ETS fusion. Thus far, four main types of Ewing-like sarcomas have been described: CIC-rearranged sarcomas,⁶ BCOR-rearranged sarcomas,⁷ sarcomas with a fusion between EWSR1 and a gene not belonging to the ETS family members, and unclassified round cell sarcomas.⁸ SBRCSs with CIC rearrangement represent the most prevalent group within non–EWSR1-rearranged SBRCSs. The frequency of CIC-rearranged sarcomas has been estimated as 3% to 5% of SBRCSs in both adult and pediatric populations.⁹^,¹⁰ From a clinical perspective, these tumors present most commonly as deep STTs in the extremities.⁹ Patients generally are older than those with classic ES, with a peak incidence in the fourth decade of life, in contrast to ES, in which an onset in the second to third decade of life is common.⁹^,¹¹^,¹² CIC-rearranged sarcomas have a more aggressive behavior when compared with classic ES, although the available data are limited to retrospective case reports and case series.⁹^,¹¹^,¹³ The most common CIC rearrangement events are CIC::DUX4 gene fusions. However, given the complex structure of the DUX gene locus, which harbors multiple gene copies with highly similar sequences, both fluorescence in situ hybridization (FISH)- and next-generation sequencing–based methods have difficulties in the correct identification of CIC::DUX4 translocation events.¹⁴^,¹⁵ In particular, an inability of different RNA-seq approaches to identify the CIC::DUX4 gene fusion has been observed.15, 16, 17 Yet, gene fusions involving transcription factors are expected to have a significant effect on gene expression patterns. Indeed, the CIC::DUX4 fusion protein has been shown to up-regulate the expression of the polyomavirus enhancer activator 3 (PEA3) subfamily of ETS transcription factors (eg, ETV1, ETV4, and ETV5).⁶ Thus, the up-regulated expression of ETV1, ETV4, and WT1 have been introduced as diagnostic markers for CIC-rearranged sarcomas by immunohistochemistry,¹⁰^,¹⁸^,¹⁹ RNA expression analysis,¹⁰^,¹⁵ and RNA in situ hybridization.²⁰ In particular, previous studies have shown that immunohistochemical stainings for ETV4 and WT1 are helpful diagnostic markers in the workup of SBRCSs, yet not sufficient to identify all CIC-rearranged cases in a specific way.¹⁸^,¹⁹ On the other hand, the molecular pathologic detection of CIC rearrangements, which still represents the only reliable diagnostic marker, remains challenging.

In this study it was hypothesized that the use of gene expression profiles extracted from a widely used routine diagnostic RNA hybrid capture–based next-generation sequencing panel might offer effective support in the identification of CIC-rearranged sarcomas and be advantageous with respect to relying only on the expression of the PEA3 subfamily of ETS transcription factors. Machine learning methods applied to gene expression profiles have allowed the successful development of diagnostic and prognostic tools for different cancer entities. For example, different breast cancer expression-based classifiers are used in clinical practice,²¹^,²² and several expression-based signatures have been proposed for lymphoid neoplasms, such as for peripheral T-cell lymphoma²³ or B-cell lymphomas.²⁴^,²⁵ Here, an expression-based classifier to predict the likelihood of any given SBRCS of harboring a CIC rearrangement was developed and tested on a large retrospective cohort of routine molecular pathology diagnostic cases.

Materials and Methods

Patient Samples

A total number of 1404 tumors analyzed at the Institute of Pathology (Erlangen, Germany) were used for this study, including mainly STTs, salivary gland carcinomas, and kidney tumors. A cohort of 69 STTs with a confirmed diagnosis by an expert STT pathologist (A.A.) was used for the development of the machine learning classifier. These 69 STT samples hereafter are referred to as the curated cohort, whereas the remaining 1335 cases were used as the test cohort. The curated cohort comprised 31 tumors within the morphologic spectrum of SBRCSs: 13 classic ESs with EWSR1 gene fusions, 3 BCOR-rearranged sarcoma samples with BCOR gene fusions confirmed by immunostaining, and 15 SBRCSs with morphologic and immunophenotypic features of CIC-rearranged sarcomas that were negative for EWSR1 gene rearrangements by FISH. Two of the BCOR-rearranged samples were two distinct sarcoma samples from the same patient, 2 years apart. The additional 38 STTs included 12 synovial sarcomas, 10 solitary fibrous tumors, 6 myxoid liposarcomas, and 10 cases of dermatofibrosarcoma protuberans. All samples were used in accordance with the ethical guidelines for the use of retrospective tissue samples provided by the local ethics committee of the Friedrich-Alexander-Universität Erlangen-Nuremberg (Ethics Committee statements 24.01.2005 and 18.01.2012).

Targeted RNA Sequencing

Tumor RNA was isolated from microdissected formalin-fixed, paraffin-embedded tissues (approximately five sections, 6- to 8-μm thick) using the RNeasy formalin-fixed, paraffin-embedded RNA kit (Qiagen, Hilden, Germany). The quality of RNA samples was assessed using RNA ScreenTape on a TapeStation 4200 (Agilent, Santa Clara, CA). Only samples with >30% of RNA fragments exceeding 200 nucleotides were included for library preparation. Libraries for sequencing were prepared using the Illumina TruSight RNA Fusion Panel (Illumina, Inc., San Diego, CA), which enriches for 507 genes frequently involved in gene fusions²⁶ and covers an approximately 1.44-megabase region. A total of 500 ng RNA was used as input for library preparation, prepared according to the manufacturer's instructions. Amplified libraries were sequenced either on a MiSeq instrument or on a NextSeq 550 instrument (Illumina) with 151-bp paired-end reads with generally >3 million reads per sample. Illumina's sequenced libraries were preprocessed via the Illumina RNA-seq Alignment BaseSpace App. Within the Illumina App, sequencing reads were aligned to the human reference genome (University of California Santa Cruz, hg19) relying on the STAR aligner,²⁷ and the Manta gene fusion caller²⁸ was used for gene fusion detection.

Additional libraries for a subset of samples were prepared with the Archer FusionPlex Sarcoma Kit (Archer, Boulder, CO) according to the manufacturer's instructions. Library was prepared using 250 ng total RNA. Libraries were sequenced with 151-bp paired-end reads on a MiSeq instrument, leading to a minimum of 1.5 million reads per sample. Sequencing data were analyzed using Archer Analysis Suite v6.2.7.

FISH Analysis

Freshly cut tissue sections were used for the detection of CIC gene locus rearrangements using a commercial probe according to the manufacturer's instructions (Zytolight SPEC CIC Dual Color Break Apart Probe, Z-2285-50; Zytovision GmbH, Bremerhaven, Germany). A CIC-rearranged sarcoma (CIC_R_11) from the curated cohort was used as a positive control, whereas a BCOR-rearranged sarcoma (BCOR_02) from the curated cohort was used as a negative control. Fifty nuclei were counted per case. A signal was considered to be rearranged if the distance between the green and orange signals was at least the size of a single signal. A case was considered rearranged if >20% of the signals were rearranged. Image acquisition and high-resolution microscopy were performed using a LSM800 confocal laser scanning microscope equipped with an Airyscan detector (Zeiss, Oberkochen, Germany) and the ZEN 2.3 (blue edition) software with Airyscan image processing.

Expression Data Analysis

Illumina targeted RNA-seq data were analyzed within the R version 4.1.2 (R Foundation for Statistical Computing, Vienna, Austria, https://www.R-project.org)/Bioconductor version 3.13 environment,²⁹ starting from the raw counts obtained via the Illumina RNA-seq Alignment BaseSpace App. First, only the curated cohort samples were considered. Relying on the DESeq2 package version 1.34.0,³⁰ size factor normalization was performed for the raw counts and variance-stabilized transformed (VST) counts were obtained. A variance-stabilizing transformation is a data transformation such that the variance of transformed values is approximately independent of the mean.³¹ The use of VST counts facilitates visualization and analyses of gene expression such as clustering or classification, which generally work better with variables with similar distributions. In this study, VST counts were used as the expression measure. Principal component analysis was performed on the VST counts of the 100 genes with the highest variance across all samples and the plot of the first two principal components was generated with ggplot2 v.3.3.6.³² Expression heatmaps were generated with the pheatmap package 1.0.12 using mean-centered VST counts (ie, VST counts from which, for each gene, the mean VST count across samples was subtracted). Dot plots of the VST counts of single genes were generated using ggplot2. Differential expression analysis of SBRCS samples was performed, relying on the DESeq2 package, starting from the raw count data of only the SBRCS samples. Genes with a Benjamini-Hochberg–adjusted P value smaller than 0.01 were considered differentially expressed.

Random Forest Classifier

A random forest (RF) classifier³³ was trained using the randomForest version 4.7-1.1 package.³⁴ To build the training data set, only the CIC-rearranged and ES samples contained in the curated cohort were considered. The associated raw gene counts were transformed into VST counts and the top 100 variable genes (ie, the 100 genes with the highest variance based on their VST counts) were identified. The VST counts of these 100 genes were used as training data for the RF classifier, with the number of trees set to 100 and the number of genes randomly sampled at each tree split set to 10. Gene importance was assessed by the mean decrease in the Gini index³⁵ and used to select a subset of top important genes to be explored further. The Gini index is a measure of the impurity of a node in a tree, which equals 0 when all samples associated to a node belong to one class, and reaches the maximum (corresponding to 0.5 in a binary setting) when the classes are represented equally. During training of a tree, the ability of the expression value of a gene to separate samples belonging to the two classes is assessed via the decrease in the Gini index. Thus, in a RF, the overall mean decrease in the Gini index of a gene offers a measure of the gene's discriminative power. Given the small size of the training set and the presence of multiple genes likely to be able to discriminate well between the two classes, to assess the stability of the top important genes, another nine RFs were trained, each time using a different random seed to ensure different random initializations as well as reproducibility. Then, the median, as well as the minimum and maximum, variable importance (mean decrease in the Gini index) of the top identified genes was calculated across the 10 trained RFs. In addition, for each of the top genes, the number of RFs (minimum number of RFs = 1, maximum number of RFs = 10) in which the gene was found within the top important genes corresponding to each RF was assessed. Internal assessment of the RF classifier performance occurred via out-of-bag evaluation.³³ Indeed, because each decision tree in a RF typically is trained on a bootstrap sample of the training set, out-of-bag evaluation utilizes, for each sample, the predicted classes by the decision trees that did not use the sample during training. For each sample, a classification (CIC-rearranged or Ewing) is provided by each tree in the RF. The predicted probability of being CIC-rearranged [hereafter indicated as prob(CIC-rearranged)] is taken equal to the proportion of trees in the RF classifying the case as CIC rearranged.

Afterward, the RF model was used to calculate the predicted prob(CIC-rearranged) for all samples of the curated cohort not included in the training. Finally, the RF model was used to obtain predictions on the 1335 test cases. In both cases, the transformed counts to be used for the RF prediction were obtained for each sample as follows: first, raw counts were normalized using the size factor estimated using the gene-wise geometric means calculated on the training data as pseudo-reference sample; then, VST counts were obtained using the dispersion function pre-estimated on the training data. For each case of the test cohort, prob(CIC-rearranged) was assessed, and a probability cut-off value chosen on the basis of the distribution of all probability values was used to select candidate CIC-rearranged cases. The stability of the predicted probabilities for the candidate cases was explored by assessing the range of their predicted probabilities across the 10 trained RFs, as well as by assessing whether the median probability was higher than the selected cut-off value.

Results

Detection of Gene Fusions via Targeted RNA Sequencing

Gene fusions were analyzed in a curated cohort of 69 STT samples, including 31 SBRCSs and 38 other types of STTs (Supplemental Table S1), via targeted RNA-seq (Illumina). In all 38 STTs, RNA-seq revealed the expected gene fusion given their histopathologic diagnosis. Specifically, 8 and 4 of the 12 synovial sarcomas, respectively, were identified to harbor a SS18::SSX1 and SS18::SSX2 gene fusion, all 10 solitary fibrous tumors had a NAB2::STAT6 gene fusion, all 6 myxoid liposarcomas had a FUS::DDIT3 gene fusion, and all 10 dermatofibrosarcoma protuberans showed a COL1A1::PDGFB gene fusion.

The 31 SBRCS samples included 13 classic ESs, 3 BCOR-rearranged sarcoma samples with positive BCOR immunostaining, and 15 SBRCSs with morphologic features of CIC-rearranged sarcomas. Targeted RNA-seq allowed the identification of gene fusions in all 13 ESs, with 10 samples harboring a EWSR1::FLI1 gene fusion, and the other 3 samples harboring a EWSR1::ERG, EWSR1::FEV, and FUS::ERG gene fusion, respectively. All three BCOR-rearranged sarcoma samples harbored a BCOR::CCNB3 gene fusion. However, no fusion was detected for the 15 SBRCSs with morphologic features of CIC-rearranged sarcomas. These SBRCSs additionally were analyzed by a second RNA-seq panel (Archer). The use of this panel allowed identification of a CIC::DUX4 gene fusion in 10 samples, which thus were classified as CIC-rearranged sarcomas. In addition, a CIC rearrangement was confirmed by FISH in one case (CIC_R_11) (Supplemental Table S1) with available paraffin material. No fusion could be identified for the five remaining SBRCSs with the second RNA-seq panel. Because these samples had morphologic features of CIC rearrangement, they are hereafter referred to as CIC-rearranged-like sarcomas. Furthermore, a CIC rearrangement was found by FISH in two of these CIC-rearranged cases (CIC_R_like_01 and CIC_R_like_05) (Supplemental Table S1) with available paraffin material, while there was no sufficient paraffin material for additional FISH analysis in the three other cases.

STT Subtypes Show Distinct Gene Expression Profiles

To explore whether the expression profiles measured with the Illumina TruSight targeted RNA-seq panel could be used to support the identification of CIC-rearranged sarcomas, unsupervised analyses of the SBRCS expression profiles were performed. A plot of the first two components obtained by principal component analysis of the expression of the top 100 variable genes revealed that tumor entities clustered according to diagnosis (Figure 1A). Notably, the five SBRCSs with morphologic features of CIC rearrangement clustered together with the 10 CIC-rearranged tumor samples. In addition, hierarchical clustering analysis of the expression profiles of the same 100 genes confirmed that tumor samples with the same diagnosis clustered together (Figure 1B). Furthermore, CIC-rearranged sarcomas and CIC-rearranged-like sarcomas displayed remarkably similar expression profiles characterized by up-regulation of the expression of PEA3 transcription factor genes (ETV1, ETV4, ETV5) and WT1. At the single-gene level, several STT entities were characterized by up-regulated genes involved in the respective gene fusion (eg, there was a high expression of CCNB3 in BCOR-rearranged sarcomas, of PDGFB in dermatofibrosarcoma protuberans, and of SSX1 in synovial sarcomas) (Figure 2). Similarly, CIC-rearranged sarcomas and CIC-rearranged-like sarcomas showed a distinct and similar up-regulated expression of the known CIC::DUX4 target genes ETV1, ETV4, and WT1 (Figure 2). Yet, up-regulation was not always seen in all cases of one STT entity or was not entirely specific to one entity. Thus, the expression pattern of a larger number of genes might be more predictive.

Unsupervised expression data analysis. A: Principal component (PC) analysis plot based on the expression [variance-stabilized transformed (VST) counts] of the 100 genes with the highest variability across the 69 soft-tissue tumors of the curated cohort. B: Expression heatmap of the same top 100 variable genes. Gene expression values in the heatmap are mean-centered VST counts and are color-coded so that red represents up-regulation and blue represents down-regulation. DFSP, dermatofibrosarcoma protuberans; MLS, myxoid liposarcoma; SFT, solitary fibrous tumor.

Expression profiles of selected genes. Expression (variance-stabilized transformed counts) dot plots of selected genes showing up-regulation in distinct soft-tissue tumor entities: *CCNB3*, *PDGFB*, *SSX1*, *ETV1*, *ETV4*, and *WT1*. *BCOR*-r., *BCOR*-rearranged sarcoma; *CIC*-r., *CIC*-rearranged sarcoma; *CIC*-r.-like, *CIC*-rearranged-like sarcoma; DFSP, dermatofibrosarcoma protuberans; MLS, myxoid liposarcoma; SFT, solitary fibrous tumor.

Differential gene expression analysis between the different SBRCS groups confirmed that CIC-rearranged and CIC-rearranged-like sarcomas showed similar expression profiles that were markedly different from the other SBRCSs. Indeed, 68 genes were differentially expressed between CIC-rearranged and BCOR-rearranged sarcomas (adjusted P < 0.01) (Supplemental Table S2), and 137 between CIC-rearranged and ESs (adjusted P < 0.01) (Supplemental Table S3). In both comparisons, ETV1, ETV4, and WT1 were among the differentially expressed genes, showing up-regulation in CIC-rearranged sarcomas. Instead, only one gene (LINC00982) was differentially expressed (adjusted P < 0.01) (Supplemental Table S4) when comparing the CIC-rearranged sarcomas with the CIC-rearranged-like sarcoma samples.

Machine Learning–Based Prediction of CIC-Rearranged Cases

To explore the possibility of using gene expression profiles to support the diagnosis of SBRCSs, a gene expression-based classifier was trained. Because in the pathologic diagnostics the most critical differential diagnosis is between CIC-rearranged sarcoma (Figure 3A) and ES (Figure 3B), and these are the two most frequent SBRCSs, the classifier focused on these two SBRCSs. An RF classifier was trained on the expression profiles of the top 100 variable genes identified considering the CIC-rearranged and ES samples of the curated cohort. The predicted out-of-bag probabilities of being CIC-rearranged were greater than 0.85 for all CIC-rearranged samples and less than 0.3 for all Ewing samples (Figure 3C). Thus, the classifier could effectively support the differential diagnosis between CIC-rearranged and ESs. Results showed that only 39 genes had an importance (assessed via the mean decrease in the Gini index) greater than zero. Of these 39 genes used by the RF classifier, the top 20 important ones (ie, those with the highest mean decrease in the Gini index) included several of the genes identified as differentially expressed when comparing CIC-rearranged versus ESs, including ETV4, ETV1, and WT1 (Figure 3E). The stability of these 20 top predictors was assessed by considering their importance measure obtained in nine additional RFs. Despite fluctuations in the variable importance measure, 15 of 20 genes were among the top 20 predictors in at least eight RFs, 4 of 20 genes were among the top 20 predictors in five to seven RFs, and only 1 gene was not found in the top 20 predictors in other RFs. To explore the possibility of using the classifier to identify candidate CIC-rearranged cases in a more varied SBRCS cohort, the RF was used to predict the probability of being CIC-rearranged for all other samples in the curated cohort. The predicted probabilities were greater than 0.9 for all five CIC-rearranged-like sarcoma samples and less than 0.6 for all other samples (Figure 3D). This validation suggested that the RF classifier effectively supports the identification of candidate CIC-rearranged sarcomas.

Expression-based classifier to predict the probability of being *CIC*-rearranged. Histopathologic hematoxylin and eosin stain of an exemplary *CIC*-rearranged (A) and Ewing sarcoma (B) sample. Dot plots of out-of-bag probabilities of being *CIC*-rearranged in the training set (C) and predicted probabilities of being *CIC*-rearranged for curated cohort samples not included in the training (D). E: Variable importance plot of the top 20 genes, sorted according to their median importance (assessed by mean decrease in the Gini index) across the 10 random forests (RFs). For each gene, the **dot** shows the median variable importance across the 10 RFs, and the **squares** show the minimum and maximum variable importance. Dot color is associated with the number of RFs in which the gene was included in the top 20 predictors. Scale bars = 100 μm. DFSP, dermatofibrosarcoma protuberans; MLS, myxoid liposarcoma; SFT, solitary fibrous tumor.

To test the classifier, a large collection of 1335 tumor samples was used, mainly including STTs, salivary gland carcinomas, and kidney tumors. These samples were analyzed during routine diagnostics at the Institute of Pathology (Erlangen, Germany) using the same Illumina TruSight Panel. On the basis of visual inspection of the distribution of predicted probabilities of being CIC-rearranged, a probability threshold of 0.75 was adopted to identify 15 samples as candidate CIC-rearranged cases (Figure 4A and Table 1). The predicted probabilities appeared stable because the 15 samples had a median predicted probability of being CIC-rearranged higher than 0.75 over the 10 trained RFs (Supplemental Figure S1A), and a predicted probability higher than 0.75 in most of the RFs (Supplemental Figure S1B). For 14 of these cases no candidate gene fusion was identified during routine diagnostics, whereas 1 case harbored a CIC::NUTM1 gene fusion. Careful review of these cases by STT expert pathologists (A.A. and F.H.) showed that all cases were compatible with a diagnosis of CIC-rearranged sarcoma. In addition, in eight of the cases a diagnosis of most likely CIC-rearranged sarcoma already had been made based on histomorphology and immunohistochemistry in the first diagnostic setting without RNA fusion analysis (Table 1). Furthermore, 1 of these 15 cases (C4) (Table 1) was available for FISH analysis and showed a CIC rearrangement (Supplemental Figure S2).

Test of the expression-based classifier on a cohort of 1335 routine diagnostics cases. A: Histogram of the predicted probabilities of being *CIC*-rearranged in the test cohort samples (N = 1335). B: Expression values [variance-stabilized transformed (VST) counts] of *ETV4* and *WT1* in the test cohort samples. The **red dotted line** indicates the threshold used to identify high-*ETV4-*expression samples (*ETV4* > 11 VST counts, corresponding to the 97th percentile). A and B: Purple color indicates the candidate *CIC*-rearranged samples predicted by the random forest classifier, that is, the test samples with probability of being *CIC*-rearranged [prob(*CIC*-rearranged)] > 0.75; dark yellow indicates test samples with high-*ETV4*-expression and prob(*CIC*-rearranged) ≤ 0.75. C: Expression heatmap based on the top 20 predictors identified by the random forest. Green indicates the *CIC*-rearranged samples of the training set; pink indicates the Ewing sarcoma samples of the training set; purple indicates the candidate *CIC*-rearranged samples; and dark yellow indicates the high-*ETV4* expression samples with prob(*CIC*-rearranged) ≤0.75. The **red asterisk** indicates the false-negative *CIC*-rearranged case [ie, the case (S_09) compatible with *CIC*-rearranged sarcoma diagnosis that was predicted with a probability of 0.72]. Gene expression values in the heatmap are mean-centered VST counts and are color-coded so that red represents up-regulation and blue represents down-regulation.

Table 1.

Candidate CIC-Rearranged Cases Identified by the Classifier

Sample identifier	Predicted probability of being CIC-rearranged	Identified candidate fusion	Age, years/sex/localization	Initial diagnosis in pathology report
C1	0.99	None	64/F/lower jaw	Unclassified highly malignant epithelioid neoplasm
C2	0.98	None	38/M/thoracic wall	Ewing-like sarcoma, most likely CIC-rearranged
C3	0.97	None	54/M/tongue	Undifferentiated round cell sarcoma, NOS
C4	0.94	None	19/F/upper arm	Unclassified highly malignant round and spindled cell sarcoma
C5	0.94	None	32/F/neck	Ewing-like sarcoma, most likely CIC-rearranged
C6	0.93	None	25/F/thoracic wall	Ewing-like sarcoma, most likely CIC-rearranged
C7	0.92	None	48/F/lower arm	Ewing-like sarcoma, most likely CIC-rearranged
C8	0.89	None	21/F/foot	Ewing-like sarcoma, most likely CIC-rearranged
C9	0.87	None	42/M/arm	Unclassified epithelioid soft tissue neoplasm of unknown histogenesis
C10	0.86	CIC::NUTM1	44/F/neck	CIC::NUTM1 rearranged neoplasm
C11	0.85	None	34/F/knee	Ewing-like sarcoma, most likely CIC-rearranged
C12	0.84	None	17/F/retroperitoneum	Unclassified small round cell sarcoma
C13	0.82	None	25/F/leg	Ewing-like sarcoma, most likely CIC-rearranged
C14	0.82	None	65/F/arm	Ewing-like sarcoma, most likely CIC-rearranged
C15	0.8	None	14/F/axilla	Unclassified highly malignant epithelioid neoplasm

Open in a new tab

The table columns show sample identifier, predicted probability of being CIC-rearranged, identified gene fusion during primary routine pathology workup, age, sex, anatomic localization, and initial diagnosis in the pathology report.

F, female; M, male.

Unsupervised analyses based on the expression of the 20 top predictors identified via the RF classifier showed that the 15 candidate cases clustered together with the CIC-rearranged samples of the training set (Figure 4C). ETV4 expression was very high in all 15 candidate cases, validating the approach. However, these cases were not those with the highest WT1 expression (Figure 4B). This suggested that ETV4 expression was a key element in the classifier predictions, whereas the use of WT1 expression alone might have missed these cases. To explore further whether the multigene RF classifier is superior to using ETV4 expression alone, all 37 test samples with high-ETV4-expression (ie, ETV4 expression higher than the 97th percentile of ETV4 gene expression across all test samples) were investigated. This ETV4 expression threshold was chosen on the basis of the curated cohort as one that could well separate CIC-rearranged and CIC-rearranged-like cases from the others (Figure 2). These 37 high-ETV4-expression cases included the 15 candidate cases identified by the RF and 22 additional cases. Of these 22 cases, 1 case clustered with the CIC-rearranged samples of the training set and 21 cases had discordant expression profiles, which differed from the CIC-rearranged cases (Figure 4C). The case clustering with the CIC-rearranged samples (S_09) (Table 2) was an additional candidate CIC-rearranged case on the basis of its expression profile. This high-ETV4-expression case had a predicted probability of 0.72 and indeed was compatible with a CIC-rearranged sarcoma on the basis of the STT expert pathologists' review. Because this case was an external case that had been sent for consultation, no paraffin material was available for FISH analysis. However, all of the other 21 tumors with a probability less than 0.75 but high ETV4 expression were not compatible with CIC-rearranged sarcomas, and included carcinomas of the salivary glands, sinonasal undifferentiated carcinoma, and epithelioid sarcoma of proximal type. A reassessment of all 1335 cases on the basis of their clinicopathologic data and morphology suggested that the 15 candidate CIC-rearranged samples identified by the classifier with prob(CIC-rearranged) > 0.75 and the additional sample S_09 were the only 16 cases with a diagnosis compatible with CIC-rearranged sarcoma. Taken together, these results show that the RF classifier performs better than ETV4 expression alone.

Table 2.

Additional Candidate CIC-Rearranged Case Identified on the Basis of Its Expression Profile

Sample identifier	Predicted probability of being CIC-rearranged	Identified candidate fusion	Age, years/sex/localization	Initial diagnosis in pathology report
S_09	0.72	None	47/F/hand	Unclassified small round cell sarcoma

Open in a new tab

F, female.

Discussion

Here, the usefulness of gene expression profiles measured with the Illumina TruSight RNA Fusion panel to support the diagnosis of STTs was evaluated. In the curated cohort of 69 STTs, unsupervised analysis of expression data revealed that tumor entities characterized by distinct gene fusions clustered together according to their gene expression profiles. Notably, the 10 CIC-rearranged sarcomas and the five SBRCSs with a CIC-rearranged-like histomorphology showed very similar expression patterns. In addition, all 15 samples showed an up-regulation of PEA3 transcription factor genes (ETV1, ETV4, ETV5) and WT1, in agreement with previous findings that PEA3 genes are transcriptional targets of the chimeric CIC::DUX4 fusion protein.⁶^,¹⁰^,¹⁵^,18, 19, 20 Given the observed subtype-specific transcriptional profiles revealed by the targeted RNA-seq panel, the possibility of building an expression-based RF classifier focusing on CIC-rearranged and ES cases, the two most common SBRCS subgroups and those with the most critical differential diagnosis, was explored further. For the other samples in the curated cohort, the RF predicted probabilities greater than 0.9 for the CIC-rearranged-like cases and less than 0.6 for the other SBRCSs. When tested on a large test cohort of 1335 routine diagnostic cases, the classifier identified 15 cases with high probability of being CIC-rearranged, including 1 case with a CIC::NUTM1 gene fusion.³⁶ The compatibility with a CIC-rearranged sarcoma diagnosis was confirmed for all of them by their re-evaluation by expert soft-tissue pathologists. In addition, one case with available paraffin material was confirmed to harbor CIC rearrangement by FISH.

Results confirmed the importance of ETV4 as a marker for CIC-rearranged sarcomas, however, they also highlighted advantages of using a multigene assessment. Indeed, all 15 candidate cases had very high ETV4 expression value. However, when considering the additional 22 samples with ETV4 expression higher than approximately the 97th percentile, the predicted probabilities by the RF in most cases were markedly lower, correctly indicating that these samples were not CIC-rearranged. The expression profiles of 21 of these 22 cases differed from those of the training CIC-rearranged cases and re-evaluation by the expert STT pathologists indeed ruled out a CIC-rearranged sarcoma diagnosis for all 21. On the other hand, only one case of potential CIC-rearranged sarcoma with a probability of 0.72 would have been interpreted as a false negative. A reassessment of this case revealed that this tumor had been classified as likely CIC-rearranged sarcoma in the initial pathology report based on morphology and immunohistochemistry. However, the sample harbored very poor RNA quality, which might have affected the measured expression profiles and in turn the predicted probability by the RF. In summary, a prediction based on ETV4 expression alone, with an expression threshold of approximately the 97th ETV4 expression percentile, would have identified one more CIC-rearranged case correctly, but with a precision (ie, percentage of correctly identified CIC-rearranged cases out of all cases predicted as CIC-rearranged) of 43% (16 of 37). Instead, the RF classifier achieved a precision of 100% (15 of 15). This suggests that more specific diagnostic support might be achievable by using the RF multigene classifier, while high ETV4 expression is a sensitive but less-specific screening approach. Furthermore, the RF provides a quantitative and observer-independent prediction score. To assess the sensitivity and specificity of the RF properly, it would be necessary to have an alternative, ground-truth criterion able to identify all positive cases (ie, all CIC-rearranged sarcomas), in the test set of 1335 cases. The reassessment of all 1335 cases on the basis of their clinicopathologic data and morphology suggested that these 16 samples were the only cases compatible with CIC-rearranged sarcoma. Thus, a threshold of prob(CIC-rearranged) > 0.75 identified 15 of 16 cases with no false-positive results, whereas a threshold ≥0.72 would have identified all 16 cases with one false-positive case. Accordingly, the identification of CIC-rearranged cases with the RF classifier appears highly specific. Yet, even a case with high prob(CIC-rearranged) by the RF classifier needs critical interpretation on the basis of morphology, immunoprofiling, and clinicopathologic data by the pathologist.

Taken together, this study shows that the bioinformatical analysis of targeted RNA-seq expression patterns can effectively support a SBRCS diagnosis. The failure to detect CIC::DUX4 rearrangements by next-generation sequencing–based RNA fusion analysis emphasizes the need for careful evaluation of novel diagnostic assays using known rearranged cases and expertise of a laboratory specialized in STTs. In the future, the availability of a larger and well-curated training cohort of STTs could mitigate the well-known curse of dimensionality issue, associated with having a number of features (genes) significantly higher than the number of samples. A larger training set also could allow extension of the approach to a multiclass prediction tool, able to predict probabilities associated with different STT subtypes. In addition, this approach may be extended to separate STT entities without specific gene fusion events, such as dedifferentiated liposarcoma.

Disclosure Statement

None declared.

Acknowledgments

We thank Dr. Silvia Vergarajauregui and Prof. Felix Engel for support with confocal imaging, and Simone Hebele and Nadine Pohl for excellent technical assistance.

Author Contributions

L.D.S. acquired and analyzed the data; M.G.C. performed bioinformatical analyses; L.D.S., M.G.C., F.H., and F.F. interpreted the data and wrote the manuscript; E.A.M. acquired and analyzed the data; R.S., M.M., M.B., N.M., S.S., R.J., R.G., S.W., and A.H. participated in patient collection and/or provided resources; A.A. participated in patient collection, analyzed the data, and wrote the manuscript; F.H. and F.F. conceived and designed the study; F.H. supervised data acquisition and analysis; and F.F. supervised bioinformatical analyses. All authors read and approved the final version of the manuscript.

Footnotes

Supported by Deutsche Forschungsgemeinschaft (German Research Foundation) grants TRR 305 [projects Z01 (F.F.) and Z02 (A.H.)] and TRR 374 (project INF) (F.F.); and confocal imaging was funded by Deutsche Forschungsgemeinschaft grant INST 410/91-1 FUGG.

L.D.S. and M.G.C. contributed equally to this work.

Current address of L.D.S., School of Medicine, Institute of Human Genetics, Klinikum Rechts der Isar, Technical University of Munich, Munich; and Institute of Neurogenomics, Computational Health Center, Helmholtz Zentrum München, Neuherberg, Germany.

Supplemental material for this article can be found at http://doi.org/10.1016/j.jmoldx.2024.02.002.

Supplemental Data

Supplemental Figure S1

Stability of the predictions on the test cohort. A: Histogram of the median predicted probabilities of being CIC-rearranged in the test cohort samples (N = 1335), calculated across the 10 repeated random forests (RFs). B: For each of the 15 candidate CIC-rearranged cases, the median probability of being CIC-rearranged (black dot), as well as the minimum and maximum probabilities (white squares) in the 10 RFs, are shown.

mmc1.pdf^{(153.8KB, pdf)}

Supplemental Figure S2

Confirmation of CIC-rearrangement by fluorescence in situ hybridization. High-resolution confocal image of CIC break apart FISH in one selected CIC-rearranged-like case (CIC_R_like_05) (Supplemental Table S1). The three channels represent the following: orange, break apart probe proximal/centromeric to the CIC breakpoint region; green, break apart probe distal/telomeric to the CIC breakpoint region; and blue, DAPI. A: Overlaid image with all three channels. B: Image containing only orange/green channels. Scale bars = 10 μm.

mmc2.pdf^{(329KB, pdf)}

Supplemental Table S1

mmc3.xlsx^{(11.4KB, xlsx)}

Supplemental Table S2

mmc4.xlsx^{(60.3KB, xlsx)}

Supplemental Table S3

mmc5.xlsx^{(61.1KB, xlsx)}

Supplemental Table S4

mmc6.xlsx^{(58.6KB, xlsx)}

References

1.WHO Classification of Tumours Editorial Board . ed 5. Vol 3. IARC Press; Lyon, France: 2020. (Soft Tissue and Bone Tumours). [Google Scholar]
2.Antonescu C. Round cell sarcomas beyond Ewing: emerging entities. Histopathology. 2014;64:26–37. doi: 10.1111/his.12281. [DOI] [PubMed] [Google Scholar]
3.Antonescu C.R. The role of genetic testing in soft tissue sarcoma. Histopathology. 2006;48:13–21. doi: 10.1111/j.1365-2559.2005.02285.x. [DOI] [PubMed] [Google Scholar]
4.Mitelman F., Johansson B., Mertens F. The impact of translocations and gene fusions on cancer causation. Nat Rev Cancer. 2007;7:233–245. doi: 10.1038/nrc2091. [DOI] [PubMed] [Google Scholar]
5.Machado I., Yoshida A., Morales M.G.N., Abrahao-Machado L.F., Navarro S., Cruz J., Lavernia J., Parafioriti A., Picci P., Llombart-Bosch A. Review with novel markers facilitates precise categorization of 41 cases of diagnostically challenging, "undifferentiated small round cell tumors". A clinicopathologic, immunophenotypic and molecular analysis. Ann Diagn Pathol. 2018;34:1–12. doi: 10.1016/j.anndiagpath.2017.11.011. [DOI] [PubMed] [Google Scholar]
6.Kawamura-Saito M., Yamazaki Y., Kaneko K., Kawaguchi N., Kanda H., Mukai H., Gotoh T., Motoi T., Fukayama M., Aburatani H., Takizawa T., Nakamura T. Fusion between CIC and DUX4 up-regulates PEA3 family genes in Ewing-like sarcomas with t(4;19)(q35;q13) translocation. Hum Mol Genet. 2006;15:2125–2137. doi: 10.1093/hmg/ddl136. [DOI] [PubMed] [Google Scholar]
7.Pierron G., Tirode F., Lucchesi C., Reynaud S., Ballet S., Cohen-Gogo S., Perrin V., Coindre J.M., Delattre O. A new subtype of bone sarcoma defined by BCOR-CCNB3 gene fusion. Nat Genet. 2012;44:461–466. doi: 10.1038/ng.1107. [DOI] [PubMed] [Google Scholar]
8.Sbaraglia M., Righi A., Gambarotti M., Dei Tos A.P. Ewing sarcoma and Ewing-like tumors. Virchows Arch. 2020;476:109–119. doi: 10.1007/s00428-019-02720-8. [DOI] [PubMed] [Google Scholar]
9.Antonescu C.R., Owosho A.A., Zhang L., Chen S., Deniz K., Huryn J.M., Kao Y.C., Huang S.C., Singer S., Tap W., Schaefer I.M., Fletcher C.D. Sarcomas with CIC-rearrangements are a distinct pathologic entity with aggressive outcome: a clinicopathologic and molecular study of 115 cases. Am J Surg Pathol. 2017;41:941–949. doi: 10.1097/PAS.0000000000000846. [DOI] [PMC free article] [PubMed] [Google Scholar]
10.Specht K., Sung Y.S., Zhang L., Richter G.H., Fletcher C.D., Antonescu C.R. Distinct transcriptional signature and immunoprofile of CIC-DUX4 fusion-positive round cell tumors compared to EWSR1-rearranged Ewing sarcomas: further evidence toward distinct pathologic entities. Genes Chromosomes Cancer. 2014;53:622–633. doi: 10.1002/gcc.22172. [DOI] [PMC free article] [PubMed] [Google Scholar]
11.Italiano A., Sung Y.S., Zhang L., Singer S., Maki R.G., Coindre J.M., Antonescu C.R. High prevalence of CIC fusion with double-homeobox (DUX4) transcription factors in EWSR1-negative undifferentiated small blue round cell sarcomas. Genes Chromosomes Cancer. 2012;51:207–218. doi: 10.1002/gcc.20945. [DOI] [PMC free article] [PubMed] [Google Scholar]
12.Gaspar N., Hawkins D.S., Dirksen U., Lewis I.J., Ferrari S., Le Deley M.C., Kovar H., Grimer R., Whelan J., Claude L., Delattre O., Paulussen M., Picci P., Sundby Hall K., van den Berg H., Ladenstein R., Michon J., Hjorth L., Judson I., Luksch R., Bernstein M.L., Marec-Berard P., Brennan B., Craft A.W., Womer R.B., Juergens H., Oberlin O. Ewing sarcoma: current management and future approaches through collaboration. J Clin Oncol. 2015;33:3036–3046. doi: 10.1200/JCO.2014.59.5256. [DOI] [PubMed] [Google Scholar]
13.Yoshimoto T., Tanaka M., Homme M., Yamazaki Y., Takazawa Y., Antonescu C.R., Nakamura T. CIC-DUX4 induces small round cell sarcomas distinct from Ewing sarcoma. Cancer Res. 2017;77:2927–2937. doi: 10.1158/0008-5472.CAN-16-3351. [DOI] [PMC free article] [PubMed] [Google Scholar]
14.Yoshida A., Arai Y., Kobayashi E., Yonemori K., Ogura K., Hama N., Mukai W., Motoi T., Kawai A., Shibata T., Hiraoka N. CIC break-apart fluorescence in-situ hybridization misses a subset of CIC-DUX4 sarcomas: a clinicopathological and molecular study. Histopathology. 2017;71:461–469. doi: 10.1111/his.13252. [DOI] [PubMed] [Google Scholar]
15.Kao Y.C., Sung Y.S., Chen C.L., Zhang L., Dickson B.C., Swanson D., Vaiyapuri S., Latif F., Alholle A., Huang S.C., Hornick J.L., Antonescu C.R. ETV transcriptional upregulation is more reliable than RNA sequencing algorithms and FISH in diagnosing round cell sarcomas with CIC gene rearrangements. Genes Chromosomes Cancer. 2017;56:501–510. doi: 10.1002/gcc.22454. [DOI] [PMC free article] [PubMed] [Google Scholar]
16.Machado I., Navarro L., Pellin A., Navarro S., Agaimy A., Tardio J.C., Karseladze A., Petrov S., Scotlandi K., Picci P., Llombart-Bosch A. Defining Ewing and Ewing-like small round cell tumors (SRCT): the need for molecular techniques in their categorization and differential diagnosis. A study of 200 cases. Ann Diagn Pathol. 2016;22:25–32. doi: 10.1016/j.anndiagpath.2016.03.002. [DOI] [PubMed] [Google Scholar]
17.Panagopoulos I., Bjerkehagen B., Gorunova L., Berner J.M., Boye K., Heim S. Several fusion genes identified by whole transcriptome sequencing in a spindle cell sarcoma with rearrangements of chromosome arm 12q and MDM2 amplification. Int J Oncol. 2014;45:1829–1836. doi: 10.3892/ijo.2014.2605. [DOI] [PMC free article] [PubMed] [Google Scholar]
18.Hung Y.P., Fletcher C.D., Hornick J.L. Evaluation of ETV4 and WT1 expression in CIC-rearranged sarcomas and histologic mimics. Mod Pathol. 2016;29:1324–1334. doi: 10.1038/modpathol.2016.140. [DOI] [PubMed] [Google Scholar]
19.Le Guellec S., Velasco V., Perot G., Watson S., Tirode F., Coindre J.M. ETV4 is a useful marker for the diagnosis of CIC-rearranged undifferentiated round-cell sarcomas: a study of 127 cases including mimicking lesions. Mod Pathol. 2016;29:1523–1531. doi: 10.1038/modpathol.2016.155. [DOI] [PubMed] [Google Scholar]
20.Smith S.C., Palanisamy N., Martin E., Almenara J., McHugh J.B., Choi E.K., Lucas D.R., Betz B.L., Thomas D., Patel R.M. The utility of ETV1, ETV4 and ETV5 RNA in-situ hybridization in the diagnosis of CIC-DUX sarcomas. Histopathology. 2017;70:657–663. doi: 10.1111/his.13112. [DOI] [PubMed] [Google Scholar]
21.Prat A., Ellis M.J., Perou C.M. Practical implications of gene-expression-based assays for breast oncologists. Nat Rev Clin Oncol. 2011;9:48–57. doi: 10.1038/nrclinonc.2011.178. [DOI] [PMC free article] [PubMed] [Google Scholar]
22.Taherian-Fard A., Srihari S., Ragan M.A. Breast cancer classification: linking molecular mechanisms to disease prognosis. Brief Bioinform. 2015;16:461–474. doi: 10.1093/bib/bbu020. [DOI] [PubMed] [Google Scholar]
23.Iqbal J., Wright G., Wang C., Rosenwald A., Gascoyne R.D., Weisenburger D.D., et al. Gene expression signatures delineate biological and prognostic subgroups in peripheral T-cell lymphoma. Blood. 2014;123:2915–2923. doi: 10.1182/blood-2013-11-536359. [DOI] [PMC free article] [PubMed] [Google Scholar]
24.Bobee V., Drieux F., Marchand V., Sater V., Veresezan L., Picquenot J.M., Viailly P.J., Lanic M.D., Viennot M., Bohers E., Oberic L., Copie-Bergman C., Molina T.J., Gaulard P., Haioun C., Salles G., Tilly H., Jardin F., Ruminy P. Combining gene expression profiling and machine learning to diagnose B-cell non-Hodgkin lymphoma. Blood Cancer J. 2020;10:59. doi: 10.1038/s41408-020-0322-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
25.Zhang W., Ao Q., Guan Y., Zhu Z., Kuang D., Li M.M.Q., Shen K., Zhang M., Wang J., Yang L., Cai H., Wang Y., Young K.H., Zhou J., Xiao M. A novel diagnostic approach for the classification of small B-cell lymphoid neoplasms based on the NanoString platform. Mod Pathol. 2022;35:632–639. doi: 10.1038/s41379-021-00954-z. [DOI] [PMC free article] [PubMed] [Google Scholar]
26.Sun L., McNulty S.N., Evenson M.J., Zhu X., Robinson J.A., Mann P.R., Duncavage E.J., Pfeifer J.D. Clinical implications of a targeted RNA-sequencing panel in the detection of gene fusions in solid tumors. J Mol Diagn. 2021;23:1749–1760. doi: 10.1016/j.jmoldx.2021.08.009. [DOI] [PubMed] [Google Scholar]
27.Dobin A., Davis C.A., Schlesinger F., Drenkow J., Zaleski C., Jha S., Batut P., Chaisson M., Gingeras T.R. STAR: ultrafast universal RNA-seq aligner. Bioinformatics. 2013;29:15–21. doi: 10.1093/bioinformatics/bts635. [DOI] [PMC free article] [PubMed] [Google Scholar]
28.Chen X., Schulz-Trieglaff O., Shaw R., Barnes B., Schlesinger F., Kallberg M., Cox A.J., Kruglyak S., Saunders C.T. Manta: rapid detection of structural variants and indels for germline and cancer sequencing applications. Bioinformatics. 2016;32:1220–1222. doi: 10.1093/bioinformatics/btv710. [DOI] [PubMed] [Google Scholar]
29.Gentleman R.C., Carey V.J., Bates D.M., Bolstad B., Dettling M., Dudoit S., Ellis B., Gautier L., Ge Y., Gentry J., Hornik K., Hothorn T., Huber W., Iacus S., Irizarry R., Leisch F., Li C., Maechler M., Rossini A.J., Sawitzki G., Smith C., Smyth G., Tierney L., Yang J.Y.H., Zhang J. Bioconductor: open software development for computational biology and bioinformatics. Genome Biol. 2004;5:R80. doi: 10.1186/gb-2004-5-10-r80. [DOI] [PMC free article] [PubMed] [Google Scholar]
30.Love M.I., Huber W., Anders S. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol. 2014;15:550. doi: 10.1186/s13059-014-0550-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
31.Anders S., Huber W. Differential expression analysis for sequence count data. Genome Biol. 2010;11:R106. doi: 10.1186/gb-2010-11-10-r106. [DOI] [PMC free article] [PubMed] [Google Scholar]
32.Wickham H. Springer-Verlag; New York: 2016. ggplot2: Elegant Graphics for Data Analysis. [Google Scholar]
33.Breiman L. Random forests. Mach Learn. 2001;45:5–32. [Google Scholar]
34.Liaw A., Wiener M. Classification and regression by randomForest. R News. 2002;2:18–22. [Google Scholar]
35.Krzywinski M., Altman N. Classification and regression trees. Nat Methods. 2017;14:755–756. [Google Scholar]
36.Schaefer I.M., Dal Cin P., Landry L.M., Fletcher C.D.M., Hanna G.J., French C.A. CIC-NUTM1 fusion: a case which expands the spectrum of NUT-rearranged epithelioid malignancies. Genes Chromosomes Cancer. 2018;57:446–451. doi: 10.1002/gcc.3. [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supplemental Figure S1

mmc1.pdf^{(153.8KB, pdf)}

Supplemental Figure S2

mmc2.pdf^{(329KB, pdf)}

Supplemental Table S1

mmc3.xlsx^{(11.4KB, xlsx)}

Supplemental Table S2

mmc4.xlsx^{(60.3KB, xlsx)}

Supplemental Table S3

mmc5.xlsx^{(61.1KB, xlsx)}

Supplemental Table S4

mmc6.xlsx^{(58.6KB, xlsx)}

[bib1] 1.WHO Classification of Tumours Editorial Board . ed 5. Vol 3. IARC Press; Lyon, France: 2020. (Soft Tissue and Bone Tumours). [Google Scholar]

[bib2] 2.Antonescu C. Round cell sarcomas beyond Ewing: emerging entities. Histopathology. 2014;64:26–37. doi: 10.1111/his.12281. [DOI] [PubMed] [Google Scholar]

[bib3] 3.Antonescu C.R. The role of genetic testing in soft tissue sarcoma. Histopathology. 2006;48:13–21. doi: 10.1111/j.1365-2559.2005.02285.x. [DOI] [PubMed] [Google Scholar]

[bib4] 4.Mitelman F., Johansson B., Mertens F. The impact of translocations and gene fusions on cancer causation. Nat Rev Cancer. 2007;7:233–245. doi: 10.1038/nrc2091. [DOI] [PubMed] [Google Scholar]

[bib5] 5.Machado I., Yoshida A., Morales M.G.N., Abrahao-Machado L.F., Navarro S., Cruz J., Lavernia J., Parafioriti A., Picci P., Llombart-Bosch A. Review with novel markers facilitates precise categorization of 41 cases of diagnostically challenging, "undifferentiated small round cell tumors". A clinicopathologic, immunophenotypic and molecular analysis. Ann Diagn Pathol. 2018;34:1–12. doi: 10.1016/j.anndiagpath.2017.11.011. [DOI] [PubMed] [Google Scholar]

[bib6] 6.Kawamura-Saito M., Yamazaki Y., Kaneko K., Kawaguchi N., Kanda H., Mukai H., Gotoh T., Motoi T., Fukayama M., Aburatani H., Takizawa T., Nakamura T. Fusion between CIC and DUX4 up-regulates PEA3 family genes in Ewing-like sarcomas with t(4;19)(q35;q13) translocation. Hum Mol Genet. 2006;15:2125–2137. doi: 10.1093/hmg/ddl136. [DOI] [PubMed] [Google Scholar]

[bib7] 7.Pierron G., Tirode F., Lucchesi C., Reynaud S., Ballet S., Cohen-Gogo S., Perrin V., Coindre J.M., Delattre O. A new subtype of bone sarcoma defined by BCOR-CCNB3 gene fusion. Nat Genet. 2012;44:461–466. doi: 10.1038/ng.1107. [DOI] [PubMed] [Google Scholar]

[bib8] 8.Sbaraglia M., Righi A., Gambarotti M., Dei Tos A.P. Ewing sarcoma and Ewing-like tumors. Virchows Arch. 2020;476:109–119. doi: 10.1007/s00428-019-02720-8. [DOI] [PubMed] [Google Scholar]

[bib9] 9.Antonescu C.R., Owosho A.A., Zhang L., Chen S., Deniz K., Huryn J.M., Kao Y.C., Huang S.C., Singer S., Tap W., Schaefer I.M., Fletcher C.D. Sarcomas with CIC-rearrangements are a distinct pathologic entity with aggressive outcome: a clinicopathologic and molecular study of 115 cases. Am J Surg Pathol. 2017;41:941–949. doi: 10.1097/PAS.0000000000000846. [DOI] [PMC free article] [PubMed] [Google Scholar]

[bib10] 10.Specht K., Sung Y.S., Zhang L., Richter G.H., Fletcher C.D., Antonescu C.R. Distinct transcriptional signature and immunoprofile of CIC-DUX4 fusion-positive round cell tumors compared to EWSR1-rearranged Ewing sarcomas: further evidence toward distinct pathologic entities. Genes Chromosomes Cancer. 2014;53:622–633. doi: 10.1002/gcc.22172. [DOI] [PMC free article] [PubMed] [Google Scholar]

[bib11] 11.Italiano A., Sung Y.S., Zhang L., Singer S., Maki R.G., Coindre J.M., Antonescu C.R. High prevalence of CIC fusion with double-homeobox (DUX4) transcription factors in EWSR1-negative undifferentiated small blue round cell sarcomas. Genes Chromosomes Cancer. 2012;51:207–218. doi: 10.1002/gcc.20945. [DOI] [PMC free article] [PubMed] [Google Scholar]

[bib12] 12.Gaspar N., Hawkins D.S., Dirksen U., Lewis I.J., Ferrari S., Le Deley M.C., Kovar H., Grimer R., Whelan J., Claude L., Delattre O., Paulussen M., Picci P., Sundby Hall K., van den Berg H., Ladenstein R., Michon J., Hjorth L., Judson I., Luksch R., Bernstein M.L., Marec-Berard P., Brennan B., Craft A.W., Womer R.B., Juergens H., Oberlin O. Ewing sarcoma: current management and future approaches through collaboration. J Clin Oncol. 2015;33:3036–3046. doi: 10.1200/JCO.2014.59.5256. [DOI] [PubMed] [Google Scholar]

[bib13] 13.Yoshimoto T., Tanaka M., Homme M., Yamazaki Y., Takazawa Y., Antonescu C.R., Nakamura T. CIC-DUX4 induces small round cell sarcomas distinct from Ewing sarcoma. Cancer Res. 2017;77:2927–2937. doi: 10.1158/0008-5472.CAN-16-3351. [DOI] [PMC free article] [PubMed] [Google Scholar]

[bib14] 14.Yoshida A., Arai Y., Kobayashi E., Yonemori K., Ogura K., Hama N., Mukai W., Motoi T., Kawai A., Shibata T., Hiraoka N. CIC break-apart fluorescence in-situ hybridization misses a subset of CIC-DUX4 sarcomas: a clinicopathological and molecular study. Histopathology. 2017;71:461–469. doi: 10.1111/his.13252. [DOI] [PubMed] [Google Scholar]

[bib15] 15.Kao Y.C., Sung Y.S., Chen C.L., Zhang L., Dickson B.C., Swanson D., Vaiyapuri S., Latif F., Alholle A., Huang S.C., Hornick J.L., Antonescu C.R. ETV transcriptional upregulation is more reliable than RNA sequencing algorithms and FISH in diagnosing round cell sarcomas with CIC gene rearrangements. Genes Chromosomes Cancer. 2017;56:501–510. doi: 10.1002/gcc.22454. [DOI] [PMC free article] [PubMed] [Google Scholar]

[bib16] 16.Machado I., Navarro L., Pellin A., Navarro S., Agaimy A., Tardio J.C., Karseladze A., Petrov S., Scotlandi K., Picci P., Llombart-Bosch A. Defining Ewing and Ewing-like small round cell tumors (SRCT): the need for molecular techniques in their categorization and differential diagnosis. A study of 200 cases. Ann Diagn Pathol. 2016;22:25–32. doi: 10.1016/j.anndiagpath.2016.03.002. [DOI] [PubMed] [Google Scholar]

[bib17] 17.Panagopoulos I., Bjerkehagen B., Gorunova L., Berner J.M., Boye K., Heim S. Several fusion genes identified by whole transcriptome sequencing in a spindle cell sarcoma with rearrangements of chromosome arm 12q and MDM2 amplification. Int J Oncol. 2014;45:1829–1836. doi: 10.3892/ijo.2014.2605. [DOI] [PMC free article] [PubMed] [Google Scholar]

[bib18] 18.Hung Y.P., Fletcher C.D., Hornick J.L. Evaluation of ETV4 and WT1 expression in CIC-rearranged sarcomas and histologic mimics. Mod Pathol. 2016;29:1324–1334. doi: 10.1038/modpathol.2016.140. [DOI] [PubMed] [Google Scholar]

[bib19] 19.Le Guellec S., Velasco V., Perot G., Watson S., Tirode F., Coindre J.M. ETV4 is a useful marker for the diagnosis of CIC-rearranged undifferentiated round-cell sarcomas: a study of 127 cases including mimicking lesions. Mod Pathol. 2016;29:1523–1531. doi: 10.1038/modpathol.2016.155. [DOI] [PubMed] [Google Scholar]

[bib20] 20.Smith S.C., Palanisamy N., Martin E., Almenara J., McHugh J.B., Choi E.K., Lucas D.R., Betz B.L., Thomas D., Patel R.M. The utility of ETV1, ETV4 and ETV5 RNA in-situ hybridization in the diagnosis of CIC-DUX sarcomas. Histopathology. 2017;70:657–663. doi: 10.1111/his.13112. [DOI] [PubMed] [Google Scholar]

[bib21] 21.Prat A., Ellis M.J., Perou C.M. Practical implications of gene-expression-based assays for breast oncologists. Nat Rev Clin Oncol. 2011;9:48–57. doi: 10.1038/nrclinonc.2011.178. [DOI] [PMC free article] [PubMed] [Google Scholar]

[bib22] 22.Taherian-Fard A., Srihari S., Ragan M.A. Breast cancer classification: linking molecular mechanisms to disease prognosis. Brief Bioinform. 2015;16:461–474. doi: 10.1093/bib/bbu020. [DOI] [PubMed] [Google Scholar]

[bib23] 23.Iqbal J., Wright G., Wang C., Rosenwald A., Gascoyne R.D., Weisenburger D.D., et al. Gene expression signatures delineate biological and prognostic subgroups in peripheral T-cell lymphoma. Blood. 2014;123:2915–2923. doi: 10.1182/blood-2013-11-536359. [DOI] [PMC free article] [PubMed] [Google Scholar]

[bib24] 24.Bobee V., Drieux F., Marchand V., Sater V., Veresezan L., Picquenot J.M., Viailly P.J., Lanic M.D., Viennot M., Bohers E., Oberic L., Copie-Bergman C., Molina T.J., Gaulard P., Haioun C., Salles G., Tilly H., Jardin F., Ruminy P. Combining gene expression profiling and machine learning to diagnose B-cell non-Hodgkin lymphoma. Blood Cancer J. 2020;10:59. doi: 10.1038/s41408-020-0322-5. [DOI] [PMC free article] [PubMed] [Google Scholar]

[bib25] 25.Zhang W., Ao Q., Guan Y., Zhu Z., Kuang D., Li M.M.Q., Shen K., Zhang M., Wang J., Yang L., Cai H., Wang Y., Young K.H., Zhou J., Xiao M. A novel diagnostic approach for the classification of small B-cell lymphoid neoplasms based on the NanoString platform. Mod Pathol. 2022;35:632–639. doi: 10.1038/s41379-021-00954-z. [DOI] [PMC free article] [PubMed] [Google Scholar]

[bib26] 26.Sun L., McNulty S.N., Evenson M.J., Zhu X., Robinson J.A., Mann P.R., Duncavage E.J., Pfeifer J.D. Clinical implications of a targeted RNA-sequencing panel in the detection of gene fusions in solid tumors. J Mol Diagn. 2021;23:1749–1760. doi: 10.1016/j.jmoldx.2021.08.009. [DOI] [PubMed] [Google Scholar]

[bib27] 27.Dobin A., Davis C.A., Schlesinger F., Drenkow J., Zaleski C., Jha S., Batut P., Chaisson M., Gingeras T.R. STAR: ultrafast universal RNA-seq aligner. Bioinformatics. 2013;29:15–21. doi: 10.1093/bioinformatics/bts635. [DOI] [PMC free article] [PubMed] [Google Scholar]

[bib28] 28.Chen X., Schulz-Trieglaff O., Shaw R., Barnes B., Schlesinger F., Kallberg M., Cox A.J., Kruglyak S., Saunders C.T. Manta: rapid detection of structural variants and indels for germline and cancer sequencing applications. Bioinformatics. 2016;32:1220–1222. doi: 10.1093/bioinformatics/btv710. [DOI] [PubMed] [Google Scholar]

[bib30] 29.Gentleman R.C., Carey V.J., Bates D.M., Bolstad B., Dettling M., Dudoit S., Ellis B., Gautier L., Ge Y., Gentry J., Hornik K., Hothorn T., Huber W., Iacus S., Irizarry R., Leisch F., Li C., Maechler M., Rossini A.J., Sawitzki G., Smith C., Smyth G., Tierney L., Yang J.Y.H., Zhang J. Bioconductor: open software development for computational biology and bioinformatics. Genome Biol. 2004;5:R80. doi: 10.1186/gb-2004-5-10-r80. [DOI] [PMC free article] [PubMed] [Google Scholar]

[bib31] 30.Love M.I., Huber W., Anders S. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol. 2014;15:550. doi: 10.1186/s13059-014-0550-8. [DOI] [PMC free article] [PubMed] [Google Scholar]

[bib32] 31.Anders S., Huber W. Differential expression analysis for sequence count data. Genome Biol. 2010;11:R106. doi: 10.1186/gb-2010-11-10-r106. [DOI] [PMC free article] [PubMed] [Google Scholar]

[bib33] 32.Wickham H. Springer-Verlag; New York: 2016. ggplot2: Elegant Graphics for Data Analysis. [Google Scholar]

[bib34] 33.Breiman L. Random forests. Mach Learn. 2001;45:5–32. [Google Scholar]

[bib35] 34.Liaw A., Wiener M. Classification and regression by randomForest. R News. 2002;2:18–22. [Google Scholar]

[bib36] 35.Krzywinski M., Altman N. Classification and regression trees. Nat Methods. 2017;14:755–756. [Google Scholar]

[bib37] 36.Schaefer I.M., Dal Cin P., Landry L.M., Fletcher C.D.M., Hanna G.J., French C.A. CIC-NUTM1 fusion: a case which expands the spectrum of NUT-rearranged epithelioid malignancies. Genes Chromosomes Cancer. 2018;57:446–451. doi: 10.1002/gcc.3. [DOI] [PMC free article] [PubMed] [Google Scholar]

PERMALINK

Machine Learning–Supported Diagnosis of Small Blue Round Cell Sarcomas Using Targeted RNA Sequencing

Lea D Schlieben

Maria Giulia Carta

Evgeny A Moskalev

Robert Stöhr

Markus Metzler

Manuel Besendörfer

Norbert Meidenbauer

Sabine Semrau

Rolf Janka

Robert Grützmann

Stefan Wiemann

Arndt Hartmann

Abbas Agaimy

Florian Haller

Fulvia Ferrazzi

Abstract

Key Points

Key Points.

Materials and Methods

Patient Samples

Targeted RNA Sequencing

FISH Analysis

Expression Data Analysis

Random Forest Classifier

Results

Detection of Gene Fusions via Targeted RNA Sequencing

STT Subtypes Show Distinct Gene Expression Profiles

Figure 1.

Figure 2.

Machine Learning–Based Prediction of CIC-Rearranged Cases

Figure 3.

Figure 4.

Table 1.

Table 2.

Discussion

Disclosure Statement

Acknowledgments

Author Contributions

Footnotes

Supplemental Data

References

Associated Data

Supplementary Materials

ACTIONS

PERMALINK

RESOURCES

Similar articles

Cited by other articles

Links to NCBI Databases