Accurate 3-gene-signature for early diagnosis of liposarcoma progression

Total RNA was isolated using Allprep DNA/RNA/miRNA Kit (Qiagen, cat No. 80224). In four of the 20 DDLPS samples, total RNA was extracted from the WD component. The cDNA synthesis and the RT-qPCR were performed using TaqMan gene expression Assay (Applied Biosystems). For the normalization, two house-keeping genes were initially tested, B2M and TUBA1A. Both genes gave similar relative expression level of target genes. Subsequently, for the space saving, internal control genes were reduced to only B2M, which had lower Ct values. All primers were purchased from Applied Biosystems.

Receiver operator characteristics (ROC) analyses were employed to evaluate the ability of univariate and multivariate models to correctly classify samples in the discovery cohort as either WD- or DDLPS, as defined by the pathologist. For each gene, a cut-off value was set as the mean of the two least different ΔCt values between the WD- and DDLPS groups in the discovery cohort.

Multivariate models were created using binary logistic regression with differential diagnosis as binary outcome variable (WDLPS vs. DDLPS), and the generated predictive probabilities were evaluated by ROC analysis in R (v3.4.3). Hierarchical clustering heatmaps were generated in R, using the “heatmap.2” package with Euclidean measure for distance matrix and complete agglomeration method for clustering. Gene ΔCt values were scale normalized (mean = 0 and standard deviation = 1) yielding Z-scores in R, and hierarchical clustering was applied patient-wise. Wilcoxon rank sum tests were applied to evaluate the statistical significance comparing ΔCt values from samples with differential histology. χ² and Student’s t-tests were applied evaluate differences in distribution of clinicopathological data. A two-sided P value < 0.05 was considered statistically significant.

High quality DNA was isolated using the Promega Wizard Genomic DNA Purification Kit (Promega, Wisconsin, United States) and the QIAamp DNA FFPE Tissue kit (Qiagen, Venlo, Netherlands) as previously described. One microgram of genomic DNA was used to produce exome-captured sequencing libraries using the Agilent SureSelect Human All Exon v5 kit (Agilent Technologies, California, United States). Paired-end 100-bp sequencing of each exome capture library was done using an Illumina HiSeq 2500 instrument and Illumina’s TruSeq SBS v3 chemistry (Illumina, California, United States).

Reads from tumor and matched normal blood sample were aligned separately to the human NCBI Build GRCh37 reference genome using Novoalign (Novocraft Technologies, Selangor, Malaysia) with default parameters. PCR duplicates, improper pairs and ambiguously mapped reads were removed using in-house scripts. SNVs were called using MuTect [13, 14]. Variants annotation was done using Oncotator.

Scale-normalized PNPLA2 expression levels, copy number variation (CNV) data and clinical parameters from DDLPS cases within the Adult Soft Tissues Sarcoma were downloaded from cBioPortal (refs 23550210 and DOI: https://​doi.​org/​10.​1158/​2159-8290.​cd-12-0095) and visualized using GraphPad Prism 5. Statistical analysis (student’s t-test) on medians and variances was performed in GraphPad Prism 5 [15].

Tissue was collected from 37 patients diagnosed with WD/ALT or DDLPS between 2014 and 2017 at Oslo University Hospital and Leiden University Medical Centre. 17 patients were diagnosed with WDLPS/ALT, and 20 with DDLPS, according to the criteria defined by WHO. Eight lipoma specimens from different patients, a healthy adipose tissue sample (commercial pooled total RNA from 18 individuals), and adipose tissue from one random non-sarcoma patients were also included in the study as controls (Fig. 1). The median age at presentation was 63 years (range 39–80), with similar distribution in each group (P = 0.51). Genders were equally represented in the groups. All eight ALT were in non-abdominal locations such as extremities or thoracic wall, the nine WDLPS were located in scrotum (n = 4) or retroperitoneum (Table 1).

Microscopic characterization of WDLPS/ALT and DDLPS was performed by a sarcoma reference pathologist (BB) on tumor sections after haematoxylin and eosin staining. Patients were diagnosed with ALT, WDLPS or DDLPS according to the current World Health Organization classification [16]. The WHO definition of ALT is a mesenchymal neoplasm consisting of entirely or partly of mature adipocytes with variation in size and at least focal nuclear atypia. The finding of multivacuolated lipoblasts and hyperchromatic stromal cells contributes to the diagnosis. WDLPS/ALT tumors were composed of scattered lipoblasts and sheets of adipocytes of variable sizes. Cytoplasmic multivacuolization, atypical nuclei and deposition of collagen was observed. The diagnosis ALT versus WDLS is clinically based on location and the possibility to resect the tumor. WDLPS is therefore used for tumors in the retroperitoneum and other locations where it is difficult to remove the tumor completely, but the two entities share the same morphology and genetics.

Classical DDLPS have a biphasic appearance, where one component is WD and another is non-lipogenic area appearing either in the primary or in a recurrent tumor. While the WD components were highly lipogenic, with densely packed adipocytes, DD components contained spindle-shaped, partly pleomorphic tumor cells, with only scarce adipose-like differentiation. In selected cases, to differentiate from lipomas and other sarcomas, analyses of MDM2 amplification were performed. Details on each sample and the disease stage are given (Additional file 1: Table S1).

The malignancy grading of the tumor was performed using the Tumor differentiation score of sarcomas in the French Federation of Cancer Centres Sarcoma Group System (FNCLCC, Federation Nationale des Centres de Lutte Contre le Cancer) [17, 18]. The three-tiered grading scheme is based on evaluation of tumor differentiation, mitotic count and the amount of tumor necrosis. The total score of these parameters are given as the grade. Grade two and grade three are considered as high grade sarcomas.

The cases were reviewed for diagnosis, grade, size, location and MDM2 status if analyzed.

One of the hallmarks of cancer is altered cell metabolism. Adipose tissue is a dynamic metabolic tissue characterized by the expression of very specific metabolic genes. In WDLPS/ALT there are alterations of transcription factors and cell-cycle related genes, but tissue identity is preserved. Instead, in DDLPS dramatic histologic changes occur, indicating the deregulation of tissue-specific metabolic genes. Thus, we decided to focus specifically on genes involved in fat metabolism, hypothesizing that the changes affecting the expression of these genes may precede major histological changes, allowing the detection of the transitional state from WD to higher grade malignancy. In addition to the tissue-specific metabolic genes, we included genes involved in antioxidant defense as they are strictly connected to cellular metabolic processes.

Gene expression levels of 11 genes coding for proteins involved in adipocyte-specific, lipid and redox metabolism were evaluated for their ability to distinguish WDLPS/ALT- from DDLPS in a discovery cohort (8 WDLPS/ALT and 3 DDLPS samples). The choice to include only 3 DDLPS samples in the discovery cohort was due to the limited number of samples we had, and with the purpose to leave more samples for the validation cohort. Gene expression levels of thioredoxin-interacting protein (TXIP), superoxide dismutase 2 (SOD2) and glutaminase (GLS) showed fair to poor accuracies towards distinguishing WD- from DDLPS (ROC AUCs of 0.813, 0.734 and 0.5, respectively), and were thus disregarded from further validation. Although heme oxygenase 1 (HMOX1) and glucose 6-phosphate dehydrogenase (G6PD) were highly accurate (ROC AUCs of 0.937 and 0.937 respectively), the intra-group variances in ΔCt values were relatively large (Additional file 2: Figure S1), and these genes were therefore also disregarded.

We next proceeded with the evaluation of genes involved in fatty acid metabolism, namely stearoyl-CoA desaturase 1 (SCD1), fatty acid synthase (FASN), glucose transporter 4 (SLC2A4), hormone-sensitive lipase (LIPE), adipose triglyceride lipase (PNPLA2), as well as the lipid droplet structural protein perilipin 1 (PLIN1). Except for FASN, all showed significant differences in the expression level between WD- and DDLPS (Wilcoxon rank sum tests P < 0.05; Additional file 2: Figure S1), and as each gave sensitivities and specificities of 1 in the discovery cohort (Table 2a) they were all chosen for further validation.

Validation was done in an independent cohort consisting of 9 WDLPS/ALT and 17 DDLPS samples, and the cutoff values determined in the discovery cohort were used to classify the samples based on each gene separately. The pre-determined cutoff values generally yielded correct classifications, with ROC AUC values ranging from 0.824 to 1.0 (Table 2b). All the genes showed statistically significantly different ΔCt values between WDLPS and DDLPS in the validation cohort (Additional file 3: Figure S2). A binary logistic regression model built from the three genes with the highest AUC values (LIPE, PNPLA2 and PLIN1, hereby termed 3M) in the discovery cohort could correctly classify all the patients in the validation cohort (AUC = 1.0).

Due to the limited sample size, we chose to reanalyze the data after pooling the discovery and validation cohorts, including in the analysis two healthy controls (commercial pooled total RNA from healthy adipose tissue of 18 individuals and chest adipose tissue from one random non-sarcoma patient) and eight lipoma samples. This strategy allowed us to eventually observe any transitional change of the gene expression from healthy adipose tissue towards DDLPS. The gene expression profiles from healthy adipose tissue samples and lipomas were indistinguishable. The expression of both LIPE and PNPLA2 showed a gradual decrease from lipomas towards DDLPS, with significantly lower expression already in WDLPS/ALT compared to lipomas (Fig. 2a). All five genes had significantly lower expression in DDLPS compared to WDLPS (Fig. 2a). Interestingly, the expression level of fat-specific gene PLIN1 was not changed between lipomas and WDLPS, consistently with the presence of comparable amount of fat in these two types of specimen. This indicates that the downregulation of PNPLA2 and LIPE in WDLPS cannot be attributed to a loss of lipid content. Due to the high accuracies of LIPE, PNPLA2 and PLIN1 (the 3M signature) in distinguishing WDLPS from DDLPS, we evaluated the expression patterns of only these three genes by hierarchical clustering heatmap analysis (Fig. 2b). The normalized expression values of these genes showed distinct clustering patterns, which distinguished WD from DD LPS samples. Interestingly, in 3 cases where RNA was available only from WD component of DDLPS tumors, gene expression profile of the investigated genes clustered together with that obtained from DD components of other specimens. The only exception was PLIN1, which had higher expression in WD component, consistently with the presence of lipid droplets (Fig. 2a, red dots). In one case, where RNA was available from both WD and DD components of the same tumor, the expression level of LIPE and PNPLA2 again was similar in both components [Fig. 2a, blue dots (WD) and green dots (DD)].

Because several adipose tissue-specific metabolic genes are directly controlled by the PPARγ master regulator of adipocyte differentiation, we measured its expression in all the patient specimens to assess whether the observed decrease in the expression of metabolic genes could be a consequence of reduced PPARG expression. Although the majority of DDLPS cases displayed strongly down-regulated PPARG expression, some cases had PPARG expression within the normal range of the WDLPS cases (Fig. 3). Among these cases with “WD-like” PPARG expression, one specimen belonged to the patient with multiple relapses and metastatic disease (referred to as M18).

Since WD/DDLPS carry many copy number variants (CNVs), we investigated if any of the 3M genes could be mutated in the case M18, for which exome data were available. This specimen derived from a patient with metastatic disease. The region carrying PNPLA2 was deleted in this case (Fig. 4, circles), but not in five randomly chosen WDLPS samples (data not shown).

We also performed in silico reanalysis of publicly available datasets containing sequencing data and clinical parameters of DDLPS cases (TCGA PanCanAtlas, https://​www.​cbioportal.​org/​study/​summary?​id=​sarc_​tcga_​pub). First, we checked the copy number variation (CNVs) and found that PNPLA2 was deleted in 22 cases (44%). The deletions were hemizygous and were associated with significantly lower PNPLA2 expression level (P < 0.01) (Fig. 5a). In three cases there was either CN gain or gene amplification, which however were not associated with gene expression increase.

To assess a clinical relevance of PNPLA2, LIPE and PLIN1 expression levels in DDLPS, we investigated the clinical parameters in relation to the gene expression level. Strikingly, we found that samples with R0 surgical margins had significantly (for both means and variances, P < 0.05) higher expression level of PNPLA2, while R1 margins were associated with lower PNPLA2 expression (Fig. 5b). None of the other two genes, LIPE and PLIN1, displayed similar associations. Notably, based on TCGA data, the oncogenes with the highest CNV in DDLPS like MDM2, CDK4 or FRS2 did not associate with surgical margins, disease progression or overall survival (data not shown).

Collectively, these data highlight PNPLA2 as a promising diagnostic and prognostic biomarker that is likely to play a role in liposarcoma pathogenesis.

Next, we investigated whether there was a difference in the expression levels across non-retroperitoneal and retroperitoneal location of ALT/WDLPS tumors in the pooled cohort. We found no significant intra-group differences between the expression levels of the five genes in the different anatomical sites (Additional file 4: Figure S3). We then performed binary logistic regression to evaluate the independence of the five metabolic genes of anatomical site. Here, SLC2A4, PNPLA2 and SCD1 were independently associated with differential diagnosis (OR = 1.09, P = 0.014; OR = 1.52, P = 0.016; and OR = 1.45, P = 0.020, respectively). No apparent differences in gene expression levels were observed in samples from primary and relapse surgeries or across gender (data not shown). Anatomical site (abdominal vs. non-abdominal) yielded an AUC of 0.75, age (continuous) 0.59, and the two-combined reached an AUC of 0.81 towards distinguishing WDLPS from DDLPS (Additional file 5: Figure S4).