Experimental sets used to define the gene expression signature associated to NF-to-MPNST transition

To define a unique MPNST vs. NF gene signature that integrates data from diverse expression studies, we sought cited accessions in GEO and ArrayExpress databases that include MPNST and NF samples, and found five microarray studies (four human and one murine). These studies are referenced in Fig 1 and the heterogeneity of these data sets is detailed in Table 1 (bold). Despite the different number of genes represented by each microarray platform, all of them satisfied the selection criteria, segregated properly NF and MPNST samples, and supported relevant works that identified potential biomarkers and therapeutic targets [3,10,16,17] or discriminated neurofibromas and MPNST from other mesenchymal tumors [15]. Based on expression profile similarity [3], we grouped dNF and pNF samples from studies E-TABM-69, GSE41747 (human) and GSE66743, respectively. We also grouped segregated sporadic and NF1-associated MPNST samples from the GSE66743 study, as there were no differentially expressed genes between samples [10]. The final number of MPNST and NF samples compared is shown in Table 1; see Materials and methods for details of sample selection and preprocessing, and probe mapping to ENSEMBL genes as common identifiers.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Prisma flow diagram.

https://doi.org/10.1371/journal.pone.0178316.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Microarray studies selected from public databases and included in the MPNST vs. NF meta-analysis.

https://doi.org/10.1371/journal.pone.0178316.t001

MPNST vs. NF gene expression signature

To integrate the five data sets, we calculated a score for each gene in each experiment (Fig 2, Microarray processing, and S1 Table), based on a formula used to combine data from heterogeneous sources [23]. To obtain a single score (s_i) for each gene, we added individual scores (s_ij) across human and mouse data sets (Fig 2, Meta-analysis). The final MPNST vs. NF gene signature contained 579 unique ENSEMBL human genes with non-null score and absolute median logFC value >0.99. The 336 up- and 243 downregulated genes included in this signature are highlighted in Table A in S2 Table (bold), embedded in the larger list from which the gene signature was filtered. The list includes the starting 7064 unique ENSEMBL human genes (4059 up- and 3005 downregulated) with non-null score in at least one independent study for which s_ij was computed. The comparison of non-null score profiles of this unfiltered integrative gene list and the lists derived from each individual study revealed comparable and overlapping patterns of gene scores (S1 Fig). Genes with the highest absolute scores in each study were at both ends of the unfiltered list, and were thus included in the final gene signature. However, some mice genes showed opposite behavior to that set in human, which corroborated differences in transcriptional responses between human and mouse models, particularly in neurodegenerative diseases [27]. Given the difficulty of interspecies comparison of results, we restricted the number of murine genes used to compute final scores (for score computation details, see Materials and methods).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. General schema of the integration of studies to generate the MPNST vs. NF gene signature and its processing by NFFinder.

The upper panel shows the five microarray studies used to compute the MPNST vs. NF gene signature by meta-analysis (central panel). Score values calculated for each gene in each study (s_ij) and for each signature gene (s_i) are shown next to rectangles that indicate the respective gene size effect or fold change (FC_ij and FC_i); red for upregulated genes and green for downregulated ones. The lower panel describes the results obtained from NFFinder when GEO and CMap-DrugMatrix databases are interrogated for direct or inverse matching of gene expression patterns, respectively, by using the signature gene names as input. Experimental validation should verify the hypotheses generated by this in-silico predictive workflow.

https://doi.org/10.1371/journal.pone.0178316.g002

The most promising biomarkers, 20 up- and 20 downregulated genes with the highest and the lowest score values from the MPNST vs. NF signature, were extracted from Table A in S2 Table (Table 2). In addition to gene score values, we determined attributes for estimating the relevance of each gene in the list, such as the median value of effect size, computed regarding the mean (logFC) or the median (logFC_m) across studies, as well as to assess the bias across the studies, like the number of studies in which each gene is represented, the inclusion or exclusion of mouse data, and the Bhattacharyya distance (BD) ratio, which replaced the entropy used by Rasche et al. [23] as an estimator of the homogenous contribution of each study to the final score (see Materials and methods) (Table A in S2 Table). The relationships between gene scores and additional attributes (Results A in S1 Appendix) and the final contribution of each individual experiment to the gene signature (Results B in S1 Appendix) are also evaluated.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. List of 20 genes with the highest and the lowest scores of the MPNST vs. NF signature¹.

https://doi.org/10.1371/journal.pone.0178316.t002

In silico search for therapeutic drugs to reverse malignant phenotype

Analysis of similar or opposite gene signatures is one of the most important applications of gene signatures for generating hypotheses for the study of NF1 and other rare diseases. We used NFFinder [20] to explore repurposing of drugs that might reverse the NF1-associated MPNST malignant phenotype by inspecting CMap and DrugMatrix databases for gene expression patterns opposite to the MPNST vs. NF gene signature (Fig 2, NFFinder processing). The complete score-ranked list of drugs with pval <0.005 can be seen at goo.gl/IdyV1N; Table A in S3 Table shows the first 50 drug entries retrieved. In the top two positions was Entinostat MS-275, a histone deacetylase (HDAC) inhibitor selective for class I HDAC. The non-specific HDAC inhibitor Trichostatin A (TSA) appeared in 39 of the top 50 entries retrieved. Two other HDAC inhibitors, the class I-selective HC-toxin and the non-specific HDAC inhibitor Scriptaid, were also on the shortlist. Other anti-cancer compounds were rifabutin, an antibiotic effective against lung cancer cells, PNU-0251126, which correlated positively with drugs for leukemia treatment, the protein phosphatase 2A inhibitor cantharidin, which induces cell death, the anti-inflammatory steroid medrysone, the topoisomerase II inhibitor ellipticine, a potent antineoplastic agent, and the non-steroidal selective estrogen receptor (ER) modulator tamoxifen, used to treat ER-positive breast cancer. To identify conditions similar to the NF-to-MPNST transformation that might share therapeutic treatments, we used NFFinder to search for disease gene signatures from GEO experiments with expression patterns resembling the MPNST vs. NF gene signature; results are shown in goo.gl/36xlB1 and the first 50 GEO studies are summarized in Table B in S3 Table. The most similar diseases were other types of cancer (58%), of which 73% were solid tumors, especially in prostate and breast, 20% were leukemias and lymphomas, and 7%, tumor cell lines. We also found premalignant neoplasias of epithelial tissue in endometrium and kidney (12%), lipid metabolism conditions (6%), pulmonary diseases (4%), muscular dystrophy (4%) and neuronal conditions (2%).

The experimental validation of the therapeutic effectiveness of drugs retrieved by NFFinder to treat NF1-associated MPNST was reported for HDAC inhibitors [28], cantharidin [29] and tamoxifen [30], and clinical trials have been suggested for all of them. The conclusion of our in-silico workflow (Fig 2) with these experimental data confirms the predictive value of the MPNST vs. NF gene signature and its usefulness to identify robust biomarkers and therapeutic agents in neurofibromatosis disease.

Functional characterization of the MPNST vs. NF gene signature

We provide detailed lists of GO term enrichment of MPNST vs. NF up- and downregulated genes (Tables A to F in S4 Table), as well as over-represented pathways derived by analysis of KEGG (Tables G and H), Wiki (Tables I and J), and Reactome (Tables K and L) databases. Results C in S1 Appendix details genes in the pathways identified.

For upregulated genes, DNA replication and cell cycle pathways were shared in the three pathway databases; 70% of GO terms associated to these genes were thus involved mainly in processes that underlie positive regulation of cell proliferation, mitosis and meiosis. We also found terms related to morphogenesis and development of skeletal, nervous, cardiovascular and digestive systems. GO terms directly related to malignancy of proliferating cells included collagen catabolic process, which can affect extracellular matrix (ECM) organization, a process involved in epithelial to mesenchymal transition (EMT) and cell migration. The KEGG ECM-receptor interaction pathway was also over-represented, which supports this observation. In accordance with GO terms and pathways, components from nucleus and cytoplasm were associated predominantly with upregulated genes.

For the downregulated genes, all three pathway databases identified the immunity pathway linked to complement activation, including regulatory elements and complement components. Downregulated genes also appeared to control peripheral nervous system development, particularly via myelination and axonogenesis. Other downregulated genes participated in the response to steroid hormones and in cell migration, chemotaxis, and cell adhesion to another cell or to a substrate such as the ECM. Unlike upregulated genes, downregulated genes accumulated GO terms related to plasma membrane and its intrinsic components, including elements from proteinaceous ECM involved in cell junction formation. When we analyzed the unfiltered list of 3005 downregulated genes, from which we derived the MPNST vs. NF downregulated gene signature, we found that specific KEGG pathways in cancer were over-represented (Table P in S4 Table).

In accordance with this functional characterization, most upregulated genes from Table 2 are involved in cell cycle progression, have been linked with carcinogenic processes in which they are upregulated, and many have also been described as diagnostic biomarkers (S5 and S6 Tables). HMGA2 is the only gene product that is an architectural transcription factor, although some other genes act as regulatory elements in mitosis, especially cyclins CCNB2, CCNB1, and kinase CDK1. As anticipated, and in contrast to the mitosis-related products of upregulated genes, which were found mainly in nucleus and cytoskeleton, the 20 most-downregulated gene products (Table 2) act mainly in the extracellular space. Whereas upregulated genes are overexpressed in cancer, some downregulated genes can be silenced or overexpressed, and thus have dual roles in cancer, as tumor suppressors or cell proliferation promoters. Some of these genes appear to be involved in signal transduction and development in the nervous system (PLEKHB1, SOX10, CHL1, PTGDS), in maintenance of structural integrity of cells and tissues, e.g., ECM formation or myelin synthesis (CRYAB, CBX7, GSN, MAL, TNXB, CDH19, PMP2, ANGPTL7, ADIRF), or to have various roles in the immune response (ENDOD1, P2RY14, FGL2, S100B). SOX10 and S100B coregulate Schwann cell proliferation and myelination [31]. CBX7 is involved in epigenetic transcriptional repression. The metabolic proteins ASPA and ADH1B are also in the most-downregulated group. ASPA participates in increasing the pool of acetate, an essential precursor for histone acetylation reactions. ADH1B oxidizes alcohol, thus also helping to generate acetate precursors, as well as retinol, an early step in synthesis of retinoic acid, a basic molecule in epithelial tissue growth and differentiation.

Chromosome distribution of genes in the MPNST vs. NF signature

To establish whether there are significant distribution differences in human chromosomes for genes in the MNPST vs. NF signature (336 up- and 243 downregulated genes), we computed frequencies of the number of genes per chromosome arm from the gene signature and the whole human genome (Fig 3). Upregulated genes accumulated significantly in chromosome arms 4q, 7p, 8q and 17q, whereas downregulated genes were over-represented mainly in 1q, 3p, 5q and 11q. Another binomial distribution identified chromosome bands in which both up- and downregulated genes were over-represented. The highest concentration of upregulated genes was found in bands 3q25.33 and 15q15.1, with others concentrated in bands 7p15.3 and Xq22.1; the largest number of bands with over-representation of upregulated genes was seen in arm 15q and the distal part of 17q. The highest significant concentration of downregulated genes was found in bands 1q24.3 and 19q13.12; arm 1q showed the largest number of over-represented bands. The function of genes in over-represented chromosome regions is shown in S7 Table. Enrichment in GO biological process terms coincided with the functions of up- and downregulated genes in the MPNST vs. NF signature, as described above.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Chromosome distribution of the MPNST vs. NF gene signature.

The gene signature distribution was calculated from the 336 genes with positive score (a), and from the 243 genes with negative score (b). Bar diagrams compare the observed distribution of MPNST vs. NF gene percentage in the human chromosome arms (dark bars) with the expected distribution according to the human ENSEMBL database (light bars). Statistical significance of the gene signature over-represented chromosome arms is above the bars. Over-represented human chromosome bands in the MPNST vs. NF gene signature are shown below each chart. Their statistical significance is shown at the top right side of band names. (****) P(X≥x) < 0.0001, (***) 0.0001< P(X≥x) < 0.001, (**) 0.001< P(X≥x) < 0.01, (*) 0.01< P(X≥x) < 0.05.

https://doi.org/10.1371/journal.pone.0178316.g003

Comparison of chromosome region enrichment from the MPNST vs. NF gene signature with that in S2 Fig for the unfiltered gene list indicated some shared over-represented regions; chromosome arms 7p, 8q and 17q accumulated upregulated genes on the unfiltered list, whereas arms 3p and 11q concentrated downregulated genes. The distal part of chromosome arm 17q showed the largest number of over-represented bands. In addition, the unfiltered list showed over-represented upregulated genes in arms 1q, 2p, 2q, 6p, 7q, 19p and Xq, and downregulated genes in 1p, 9q, 10q, 14q, 17p and 20p.

Panels of biomarkers with similar expression profiles in the context of NF1 disease

Given the greater potential prognostic capacity of biomarker panels compared to individual biomarkers, we built robust gene profiles to group functionally related genes with similar expression patterns. To obtain gene profiles, we extended the gene score methodology to identify other average gene signatures in the NF1 context. We performed five additional differential expression analyses; Table 1 groups the accession numbers of the studies used to determine gene signatures for the five new comparisons. Two of these new comparisons involved nerve tumors, whereas the other three comprised cell cultures. For tumor tissue, we carried out two meta-analyses that integrated human and mouse nerve tumors by comparing NF vs. control (nerve tissue) and MPNST vs. control. For cell cultures, we made three comparisons. Only the comparison of MPNST cell lines vs. control (NHSC; normal human Schwann cells) integrated data from two studies; for the other two comparisons, primary NFSC (neurofibroma Schwann cells) vs. control (NHSC) and MPNST cell lines vs. primary NFSC, we used standard differential expression analyses, as only one study was available. These additional gene signatures were added as Tables B to J in S2 Table. Their composition, and a comparison with the MPNST vs. NF gene signature are detailed in Results D in S1 Appendix.

To search for function-related genes in the context of neurofibromatosis, we sought genes with similar expression patterns in the six comparisons of NF, MPNST and control phenotypes in cell cultures and nerve tumors; S8 Table shows the list of 2209 genes obtained by gathering signatures of these comparisons. Comparisons (clustered by tissue type) and genes grouped independently, and dendrograms were determined of the comparisons and of the genes, grouped in 46 clusters represented by their self-organizing tree algorithm (SOTA) centroid vectors (Fig 4). Divergence between sample tissue types coincided with dissimilarities in the functional characterization of genes in the cell culture and nerve tumor signatures (Results D in S1 Appendix). As predicted, the second line of sample divergence separated MPNST samples (vs. NF or vs. control) from NF vs. control comparisons. Results E in S1 Appendix evaluates gene pattern homogeneity in each cluster.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Clustering of phenotype comparisons and of NF1-related genes.

Comparisons among MPNST, NF and control phenotypes were grouped through hierarchical clustering. Cell culture and nerve tumor comparisons are shown on the top of the upper dendrogram, and the node height scale is detailed on the left of this tree. The hierarchical relationship among gene clusters obtained by grouping the logFC_m values of 2209 NF1-related genes by the Self-Organizing Tree Algorithm (SOTA) is represented by the dendrogram on the left. Clusters are described by their SOTA centroid vectors. Color scale of logFC_m values is shown below. The right side of the Fig details the number of genes in each cluster, the gene number of the MPNST vs. NF signature in each cluster (*, > 15 genes), the percentage of genes of this signature in each cluster (grey scale shown below), the number of biological process GO terms over-represented in each cluster, and the summary of the GO term enrichment as functional characterization of clusters. A complete list of terms is shown in S9 Table.

https://doi.org/10.1371/journal.pone.0178316.g004

Although almost every cluster contained genes from the MPNST vs. NF tumor tissue signature (highlighted in S8 Table), the most populated clusters (>15 genes) appeared in four cluster areas, from which were derived four biomarker panels that included the majority of the 20 most up- and downregulated genes (Fig 4). Cluster 3, which includes genes involved in development of the peripheral nervous system, contained six genes of the 20 most-downregulated (Table 2), CRYAB, SOX10, FGL2, CDH19, PMP2 and S100B; SOX10 and S100B are markers used to diagnose neural crest-derived tumors [32]. Cluster 17 incorporated three of the most-downregulated genes, PLEKHB1, GSN and CHL1. Cluster 25 included 153 genes and 181 GO terms, and thus concentrated the highest functional enrichment. This cluster, with adjacent clusters 23 and 24, accumulated the majority of genes involved in cell proliferation. Cluster 25 included 17 of the 20 genes with the highest scores from the MPNST vs. NF comparison (Tables 2 and S8), particularly the genes that encode the putative NF1 prognostic markers TOP2A and BIRC5 [10]. Finally, cluster 34 contained five of the 20 most-downregulated genes in the MPNST vs. NF signature, CBX7, P2RY14, TNXB, ADIRF and PTGDS. Details for gene accumulation in clusters from the MPNST vs. NF signature, and for the five additional signatures, are shown in S3 Fig.

Panel of hypermethylated biomarkers in the context of NF1 disease

Aberrant epigenetic changes are associated to most cancers, particularly hypomethylation of oncogenes and hypermethylation of tumor suppressors. To establish relationships between gene expression and aberrant promoter methylation, we associated our six gene expression signatures with the methylation status of gene promoters. In GEO and ArrayExpress databases, we sought studies that provide raw methylation data relative to NF1 evolution. GEO accession GSE21714 genome-wide methylome analysis compared immunoprecipitated DNA from 10 pooled MPNST samples, 10 pooled NF samples, and 6 pooled control Schwann cell samples [13]; for the process used to identify methylated promoters, see Materials and methods. The edgeR.logFC and edgeR.adj.pval (<0.1) values obtained for MPNST vs. NF (Table A in S2 Table), MPNST vs. control (Table B), and NF vs. control (Table C) methylome comparisons showed 428, 594 and 9 differentially methylated gene promoters, respectively (right columns); most were hypermethylated. MPNST vs. NF and MPNST vs. control comparisons showed 181 promoters in common, and MPNST vs. control and NF vs. control Schwann cells indicated 4 shared promoters.

Hypermethylation of the RASSF1A promoter identified a NF1-associated MPNST subgroup with poor prognosis [14]; we therefore tested RASSF1 promoter methylation status as a control of our analysis. The RASSF1 promoter was differentially hypermethylated in MPNST vs. NF and in MPNST vs. control Schwann cells, but not in NF vs. control Schwann cells. To find other potential regulatory elements with aberrant promoter methylation, we searched for other downregulated genes with hypermethylated promoters in MPNST vs. NF and MPNST vs. control Schwann cell comparisons. Using the unfiltered gene lists for all comparisons (Table J in S2 Table), we selected genes whose expression pattern in tumor tissue included a negative MPNST vs. NF score, as well as logFC values < -0.5 for MPNST vs. NF and MPNST vs. control comparisons. Genes selected also showed edgeR.logFC and edgeR.adj.pval promoter methylation values >1.5 and <0.1, respectively. We found 56 gene promoters that fit these selection conditions. The RASSF1 gene and 9 of the selected genes found in the MPNST vs. NF gene signature are listed in Table 3; the full list is given in S10 Table.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. MPNST vs. NF signature genes potentially down-regulated by promoter hypermethylation¹.

https://doi.org/10.1371/journal.pone.0178316.t003

MPNST vs. NF gene expression signature

Here we report the integrative MPNST vs. NF gene signature associated to the NF-to-MPNST transition in the context of neurofibromatosis type 1 disease. For this signature, filtered from a larger list of all differentially expressed genes from five studies comparing MPNST to NF, we included genes with the highest scores, size effects and consensus among individual studies, and used it as a core to study biomarkers and drugs that might control evolution to malignity in NF1 patients.

The 20 highest and 20 lowest score-ranked genes suggested as promising biomarkers (Table 2) are all implicated in carcinogenic processes and many are suggested biomarkers. Among the top 20 genes in the MPNST vs. NF signature are 4 genes reported as induced in the transition from benign NF to MPNST (S11 Table); their products are the topoisomerase TOP2A, needed for correct chromosome segregation in mitosis, the apoptosis inhibitor BIRC5, a member of the mitotic chromosome passenger complex, the architectural transcription regulator HMGA2, and TPX2, which is essential for correct mitotic spindle assembly and activates the AURKA kinase to control cell cycle progression. Among the 20 most-downregulated genes are 4 NF- and/or MPNST-associated genes (S12 Table), the transcriptional regulator SOX10, needed for neural crest multipotent cell and peripheral nervous system development, the extracellular matrix glycoprotein TNXB, the cell-cell adhesion Ca²⁺-dependent cadherin CDH19, specific to myelin-forming cells, and the Ca²⁺-binding S100B, whose inhibition is associated to MPNST transformation [33].

The NF1 gene, the basic reference of reliability for the gene signature generated, had a null score and was thus absent from the MPNST vs. NF signature. This suggested that most samples of sporadic MPNST had somatic deactivating NF1 mutations, which corroborates previous studies [34]. Our results coincide with gene expression data that compare MPNST with benign NF, although given the strict score threshold, some known MPNST-related genes on the unfiltered gene list were lacking in the gene signature, such as SOX9, which encodes a developmental transcription factor [3], and TNC, which is involved in axon guidance during development [16]. Other cancer-related genes were absent from the signature for similar reasons, such as CDH1, which codes for the tumor suppressor cadherin 1. CDH1 silencing is a marker of the EMT associated to cell proliferation, invasion and metastasis [35].

We found that despite some exceptions, our results for over-represented chromosome regions in the MPNST vs. NF gene signature corroborated previous data on chromosome region enrichment. In addition, we identified other regions such as chromosome arm 3p, in which many downregulated genes are concentrated. A detailed comparison of these regions and reported aberrant chromosome modifications is shown in Discussion A in S1 Appendix.

Biomarker panels that share NF1-related gene expression profiles

In addition to the study of individual genes as potential biomarkers, tests of groups of genes in the MPNST vs. NF signature can improve prospects for diagnosis or prognosis. Kolberg et al. [10] defined a prognostic test for post-tumor resection MPNST patients, based on expression of three proteins encoded by genes located in the distal part of chromosome 17q (BIRC5, TK1, TOP2A). These genes were included in a 31-gene cell cycle progression (CCP) signature that is a robust predictor of clinical outcome for prostate cancer patients [36].

Given the limitation of the MPNST vs. NF score-ranked gene signature for grouping genes with similar function, we built profiles that integrated other NF1-related signatures. These signatures, obtained from tumor tissue and cultured cells, were based on one or two studies each and were thus less robust than our original MPNST vs. NF, built from five studies. Genes were grouped by their profiles using SOTA, which allows hierarchical clustering by a neural network. We obtained a large number of SOTA clusters containing homogeneous groups of functionally related genes that shared similar expression profiles.

Of the upregulated genes in the MPNST vs. NF signature, most grouped to Cluster 25 and are linked to cell cycle progression. Mechanisms that regulate the complex process of cell cycling are precisely regulated, and deregulation drives to aberrant cell proliferation and cancer development [37]. Not surprisingly, many genes from the CCP signature [38] were included in the cluster 25. Our group of the 20 most-upregulated genes contained 9 of the 31 CCP signature genes (RRM2, TOP2A, KIAA0101, BIRC5, NUSAP1, CENPF, PRC1, ASPM, DTL), and 14 more CCP genes were upregulated in the MPNST vs. NF gene signature (FOXM1, TK1, CDC20, BUB1B, PBK, CDKN3, ASF1B, CEP55, DLGAP5, RAD51, KIF11, KIF20A, PTTG1, CDCA8). Due to its potential as MPNST biomarker [39], we propose the inclusion of HMGA2 in any similar list built to interrogate the expression status of genes involved in cell cycle progression in MPNST prognosis. A panel from cluster 3 should contain at least CRYAB, SOX10, FGL2, CDH19, PMP2, and S100B. Cluster 17 genes to be evaluated would be PLEKHB1, GSN and CHL1. A cluster 34 panel should include at least CBX7, P2RY14, TNXB, ADIRF and PTGDS genes.

The divergence between expression values in comparisons of cell cultures and nerve tumors indicated clear differences due to the nature of the samples. Many essential characteristics of cancers in which the ECM plays a fundamental role, such as those involving EMT, invasion and metastasis, cannot be appropriately mimicked by cultured cells. In contrast to MPNST, which contain a mixture of different cell types, cultured cells derive from a single cell type. Differences between cultured cells and nerve tumors, exemplified by SOX9, SUZ12, EGFR, SPP1 and BMP2 genes, are detailed in Discussion B in S1 Appendix.

Genes potentially silenced by hypermethylation of their CpG-island promoter region

To identify potential biomarkers whose expression is inhibited by hypermethylation, we sought genes that were downregulated and showed promoter hypermethylation in tumor tissue for MPNST vs. NF and MPNST vs. control comparisons; 42% of the differentially methylated promoters were shared, most of which were hypermethylated. This result correlated with gene expression data, as more than 58% of genes from the MPNST vs. NF showed similar behavior in the MPNST vs. control comparison. In addition to the tumor suppressor RASSF1, whose differential promoter hypermethylation status and expression inhibition are associated in MPNST and other cancers [14], we found additional regulatory genes. The genes identified by common expression and methylation patterns were implicated mainly in cancer, which corroborates the importance of epigenetic changes in the control and evolution of cancerous processes; they were also linked to immune response, nervous system development, lipid metabolism, metabolic energy balance and detoxification (S10 Table). Many act as mediators in signal transduction pathways, have a structural role, or interact with other proteins to control various biological processes related to cell proliferation and apoptosis. Most genes in the MPNST vs. NF signature that are potentially silenced by promoter hypermethylation are associated with structural functions in the peripheral nervous system (PRIMA1, ST6GALNAC2, ITGB4, MBP, MPZ) or with signal transduction pathways (RABGAP1L, ACKR1, S100B) (S10 Table, bold). Differentially methylated genes are discussed in detail in Discussion C in S1 Appendix.

From MPNST vs. NF gene signature to working hypotheses: Therapeutic applications of HDAC inhibitors, cantharidin and tamoxifen

To identify diseases or biological problems similar to malignant NF evolution, or to define potential therapeutic drugs that could reverse malignant phenotype, we used NFFinder to compare the MPNST vs. NF signature with similar or contrasting signatures. Gene signatures in GEO studies most similar to that of MPNST vs. NF were associated with cancer, mainly solid tumors, followed by other diseases that share several phenotypic alterations with neurofibromatosis. Most signatures in CMap and DrugMatrix databases that contrasted from the MPNST vs. NF signature indicated that the reported as effective compounds to treat MPNST, cantharidin and tamoxifen, and especially HDAC inhibitors, could potentially reverse the malignant phenotype. Our in-silico prediction corroborates these data, which indicates the reliability of our gene signature as representative of NF evolution to malignancy.

HDAC inhibitors control gene expression by blocking deacetylation of histone and non-histone proteins. These inhibitors modify the chromatin condensation status [40] and also control several chromatin structure-independent processes that alter gene expression, such as transcription factor activity, miRNA expression, and signal transduction [41]. Some HDAC inhibitors act as epigenetic regulators by modifying the DNA methylation status, which reveals crosstalk between acetylation and methylation [42]. This ability of HDAC inhibitors to reverse epigenetic aberrations make them effective therapeutic agents in cancer, as well as in neurological and immune disorders [43]. In cancer, HDAC inhibitors impair cell proliferation, neoangiogenesis and metastasis, and increase differentiation and apoptosis. Blockade of tumor angiogenesis nonetheless hinders drug delivery and limits the use of these inhibitors for solid tumors [44]. Combination therapies are thus necessary.

Based on the sensitivity of Ras signaling tumors to HDAC inhibitors, López et al. assessed these compounds in MPNST, both in vitro and in tumor xenografts [28]. The strong reaction of NF1-associated MPNST cell lines to HDAC inhibitors led the authors to suggest their therapeutic value for inclusion in clinical trials. NFFinder also found that the pan-HDAC inhibitor TSA can replace the effective combination of synergy-acting compounds PD-901 and JQ1 to kill MPNST cells and shrink tumors [12,20]. Robust biomarkers are nonetheless needed to predict the effectiveness of TSA and other HDAC inhibitors, alone or in combination, in clinical trials for treatment of NF1-associated MPNST.

Among the 20 most up- and downregulated, several genes alter their expression in response to HDAC inhibitors such as the upregulated genes TOP2A, BIRC5, HMGA2, CCNB1, CCNB2, TPX5 and CDK1 [45–49], and the downregulated genes SOX10 [50] and GSN [51]. As components of the epigenetic regulation system, HMGA2 and GSN are main targets of HDAC inhibitors, and suggest their utility as predictive treatment markers.

HMGA2 controls gene transcription directly via chromatin remodeling, or indirectly, by altering the binding affinity of regulators and nuclear proteins to DNA through protein-protein interactions. Involved in the control of fetal development, HMGA2 also has a central role in tumor growth and metastasis. HMGA2 protein levels rise acutely in malignancies [52], and its overexpression correlates with poor prognosis in colon, lung, pancreas, ovary and gastric cancers [53,54]. Besides use of HMGA2 as a diagnostic/prognostic biomarker in NF progression to malignity, the reported silencing of HMGA2 with HDAC inhibitors justifies its use as a biomarker of treatment effectiveness of these drugs [55].

In addition to its structural role as one of the most abundant actin-binding proteins, the multifunctional regulator gelsolin GSN is involved in apoptosis and regulates processes related to pathological states such as amyloidosis, inflammation, Alzheimer's disease, cardiovascular diseases, cancer and aging [56]. In cancer, GSN has a dual effect as a promoter of cell growth and invasion [57] and as a tumor suppressor that inhibits metastasis. Its tumor suppressor effect is reported for most cancers, and it is downregulated in all of them [58]. GSN transcriptional repression is associated with epigenetic control through DNA methylation and histone deacetylation, and addition of HDAC inhibitors increases GSN expression [51], which supports its use as a biomarker for the effectiveness of MPNST treatment with HDAC inhibitors.

HMGA2 and GSN, as well as EZH2 and CBX7, respective members of PRC2 and PRC1 epigenetic repressor complexes, probably impaired during NF malignant transformation, are involved in the control of cell proliferation and metastatic phenotype by regulation of the tumor suppressor protein CDH1 and other markers of EMT, whose expression is reversed after treatment with HDAC inhibitors, particularly TSA [59]. HDAC inhibitor activity counteracts CBX7 silencing and EZH2 protein overexpression over the CDH1 promoter, as discussed in Discussion D in S1 Appendix.

In addition to genes involved in epigenetic regulation as possible HDAC inhibitor targets, we found two other genes related to acetate metabolism and availability (ASPA, ADH1B) among the 20 most-underexpressed genes of the MPNST vs. NF signature. Supplementation with acetate precursors as coadjuvant chemotherapy to complement two metabolic pathways that involve ASPA and ADH1B in acetate synthesis is considered in Discussion E in S1 Appendix.

Cantharidin and tamoxifen were also retrieved by NFFinder and confirmed experimentally as effective drugs to inhibit MPNST cultured cell proliferation and survival. The protein phosphatase 2A inhibitor cantharidin was found by screening from a library of 472 small bioactive compound library [29] and shown to avoid growth of NF1-associated MPNST cultured cells, though additional studies should clarify the relevance of cantharidin in vivo.

Due to MPNST are not sex steroid hormone sensitive, the ER modulator tamoxifen inhibits MPNST cultured cell growth independently of ER [30,60]. Combination therapies of tamoxifen and trifluoperazine have shown to be effective on treatment of sporadic and NF1-associated MPNST, suggesting the quick repurposing of these drugs for clinical and prophylactic uses. Tamoxifen, which induces NF1-associated MPNST cell death mediated by autophagy in a K-Ras degradation dependent-process [61], might replace cloroquine and combine with HDAC inhibitors to induce apoptosis-oriented autophagy in “resistant” sporadic MPNST cells [28]. Tamoxifen and HDAC inhibitors might thus boost productive autophagy in a similar way as the combination of tamoxifen with panobinostat has been proposed for treatment of other solid tumors like hepatocellular carcinoma [62].

Meta-analysis method to define the MPNST vs. NF gene signature

Various meta-analysis approaches have been developed to integrate data from independent studies. Due to their inherent difficulties, most meta-analyses restrict the inclusion of studies to a small number of different platforms or to a single platform. Integration of genomics data for rare diseases presents even greater challenges. Scanty heterogeneous data, common for rare diseases, might hide consistent information and patterns potentially present across studies. The meta-analysis method used here, based on previously described gene scores [23,26], identifies differentially expressed genes between two conditions, by integrating data from platforms of distinct size. Although we used a similar formula, our gene score differed slightly from that published in five aspects; a) scores were computed for genes as up- and downregulated, b) the logFC computed, based on the mean of expression ratios, was replaced by the logFC_m based on the median value of expression ratios, which increased score robustness, c) logFC_m values were normalized to the interval [0,1] for upregulated and [-1,0] for downregulated genes to avoid expression ratio bias among studies, d) the use of MAD (median absolute deviation) rather than standard error values to compute the penalty term due to expression deviation among replicates, which resulted in lower values for the penalty term and thus in more stringent scores, and e) stringency was reinforced by a third condition imposed for calculation of scores (B factor >0). Computation of the gene final score by the addition of each gene study score might favor genes represented in a higher number of platforms. The last filter step selected 10% of the genes on the complete list, which yielded the MPNST vs. NF signature; this is nonetheless a more balanced representation of the gene list for each of the starting studies, as it contains a larger proportion of genes common to various studies.

In addition to problems frequently encountered when comparing transcriptomes, integration of human and mouse data has drawbacks inherent to species-specific differences. As mouse models do not appear to mimic human neurodegenerative diseases at molecular level in every respect [27], we prioritized the human over the mouse transcriptome. Although doubts remained regarding the incorporation of mouse data in the meta-analysis, we included those data since robust genes expressed similarly in humans and mice could be candidates for preclinical studies. As the gene signature includes the behavior of each gene regarding the species and tissue type, it provides comprehensive information for the design of experiments.

The integrative gene signature associated to neurofibroma malignant evolution is the main contribution of this work. This signature constitutes the first step to generate working hypotheses concerning biomarkers of disease evolution and treatment effectiveness, as well as therapeutic drugs. Our in silico methodology used to define the signature and derive hypotheses for experimental purposes could be applied in the study of orphan diseases other than NF1.

Study selection

Microarray and DNA methylation studies were selected from GEO and ArrayExpress public databases using the key words NF1, MPNST, neurofibromatosis, and neurofibroma; we included high-throughput sequencing, microarray and methylation data. Additional selection criteria were 1) for microarray studies, only data from Affymetrix, Agilent, ABI and Illumina platforms were allowed, and 2) we only considered studies accepted for publication, with raw data from samples from cell cultures or nerve tumors.

Microarray data preprocessing

Each microarray was preprocessed and evaluated independently using the R/BioConductor software environment [63]. R packages used in pre-processing steps are detailed in Materials and methods A in S1 Appendix. After grouping dNF and pNF samples as NF, and sporadic and NF1-associated as MPNST, the resulting expression set was filtered to discard features with FDR values >0.05 obtained by ANOVA. The expression set derived was used to assess appropriate grouping of samples from the independent studies by principal component analysis (PCA), as shown in S5 and S6 Figs. The final derived expression set was used to analyze gene differential expression.

Probe name translation to human ENSEMBL and HUGO gene IDs and mapping in human chromosome arms

As integration of scores from different platforms required common identifiers, we translated probe names from each study to ENSEMBL human gene IDs according to Ensembl Archive release 82 (September 2015). To obtain a unique score for each gene, we selected the score of the probe with the greatest variation when the identifier was present more than once. This variation was calculated as the variance of normalized expression values of all samples from the two phenotypes compared. Mouse gene IDs were also translated to human homologous gene IDs. Human ENSEMBL gene IDs were associated with HUGO gene nomenclature IDs and mapped to human chromosomes and bands. Details of ENSEMBL gene ID association to human chromosome arms are shown in Materials and methods B in S1 Appendix.

Gene scores across studies and final individual gene score in a two-phenotype comparison

Gene scores were also computed in the R/Bioconductor environment [63]; Materials and methods C in S1 Appendix details the formula and computation method used. In an example illustrating the factors involved in gene score computation, S13 Table shows values for the differentially expressed genes obtained by comparing MPNST vs. NF from the GSE66743 study. The final score across studies of a single species was calculated as the sum of standardized individual s_ij scores. Final human and mouse scores were weighted for human results. Genes exclusive to mouse were ignored and mouse data that differed from human data in the expression ratio sign were discarded. The human score was calculated by adding scores from human studies. Inclusion of murine data was limited to genes for which the sign for the final human and mouse scores were identical, and final human scores were not null.

logFC and logFC_m of genes across studies

Gene logFC were calculated across the studies as the median value of logFC_ij; logFC_ij values for each gene (i) in each individual study (j) were obtained in the phenotype comparison. Gene logFC_m were computed as the median value of logFC_m_ij; we have defined logFC_m_ij as the log₂ of the median value of all expression ratios between phenotypes. To avoid inconsistency, the same restriction imposed on the final score was used for logFC and logFC_m values. The signs for human median values of computed logFC/logFC_m were compared with those of mouse. If human and mouse signs were equal, the mouse logFC/ logFC_m was used to compute the final logFC/logFC_m. If the signs were different, mouse values were ignored and human values were used as the final logFC/logFC_m.

Final filters to obtain the gene signature

To avoid inconsistency, the resulting list of genes with non-null score was screened to remove genes showing score signs opposite to those of logFC values (S4 Fig, third filter). To select genes with the highest score and median size effect values, a final filter was applied to the previous gene list (S4 Fig, fourth filter). This filter screened 10% of genes with the highest scores (genes with positive score) and 10% of genes with the lowest score (genes with negative score). In both cases, the absolute logFC value should be >0.99.

Computation of bias in score values among studies: Bhattacharyya distance (BD) ratio

To determine the relative influence of individual experiments on final score values, we computed for each gene the BD between the observed distribution of score frequency and the expected discrete distribution if each study contributed equally to the final score. The lowest BD values thus correlated with similar contributions of individual studies to final score values, while the highest values indicated disparity in the contribution of individual studies. To compare BD without considering the number of studies available for each gene, we computed the BD ratio by dividing the BD by the maximum possible BD value considering the number of studies for each gene. The lowest and highest BD ratio values indicated the similar or dissimilar contribution of each study to each gene score, respectively. A BD ratio equal to zero was applied exceptionally to those genes for which only one study was available, to distinguish these genes from those for which there was more than one study, although only one showed a non-null score. Computation details are given in Materials and methods D in S1 Appendix.

Statistical analyses

To determine the statistical significance of the different gene distribution frequencies in human chromosome regions, we calculated the accumulated probability of a binomial distribution P(X≥x) using the pbinom function from the R stats package; **** P ≤0.0001, *** 0.0001<P≤0.001, ** 0.001 <P≤0.01, * 0.01<P≤0.01. The expected frequency of gene distribution in human chromosome arms/bands was calculated considering the association of ENSEMBL gene IDs to human chromosome arms/bands. To assess correlation between values of two vectors, Pearson's product-moment correlation was calculated using the cor.test function from the R stats package.

Gene and pathway functional enrichment

The over-representation test for GO terms from GO Ontology database release 2016_09_24 (The Gene Ontology Consortium, 2015) was performed through PANTHER v.10 (release 20160715; [64]) using ENSEMBL IDs as input, Homo sapiens (all genes in the database) as reference list, GO biological process, GO molecular function and GO cellular component complete as Annotation Data Set, and the Bonferroni correction for multiple testing (P <0.05). The EnrichNet tool [65] was used to inspect KEGG, BioCarta, Wiki Pathways and Reactome databases to search for pathways in which a differentially expressed group of genes is enriched compared to the whole human genome; ENSEMBL IDs were used as input. Only pathways with a significant XD-score or Fischer q-value <0.1 were considered. The search for opposite and similar gene signatures in CMap-DrugMatrix and GEO databases in NFFinder was performed as described [20].

Gene clustering and visualization

Genes were grouped using SOTA [66] implemented in the analysis and visualization tool MultiexperimentViewer MeV v. 4.9 [67]. Parameters were fit by default, including distance (Pearson correlation), and Max. Cycles (100). Sample tree was carried out using hierarchical clustering to optimize sample leaf order. Pearson correlation was selected as the distance metric, and average linkage clustering as the linkage method.

DNA methylation analysis

Methylation data were downloaded from GEO, and inspected and analyzed using the MEDIPS R package [68]. Linux command lines, R code based on Lienhard et al. [68], and details of the analysis are shown in Materials and methods E in S1 Appendix.