An imprinted non-coding genomic cluster at 14q32 defines clinically relevant molecular subtypes in osteosarcoma across multiple independent datasets

While miRNA profiles were prognostic of outcome in recent studies, the precise clinical utility of miRNAs located in the 14q32 region for individualized patient outcome prediction has not yet been determined. We used two previously published genomic datasets with outcome annotation (called “Boston” and “Utah” datasets) and studied the clinical prognostic utility of the 14q32 miRNAs via the time-dependent receiver operator characteristic (tdROC) curve method. A summary table for the two datasets, the details of which have been previously reported [ 6, 9], is provided in Additional file 1. In order to minimize overfitting that is frequently associated with the construction of complex multivariate models, we used the simple, yet robust “signed average” approach. First, we generated a tdROC curve based on the signed average of the top 5 prognostic miRNAs residing at the 14q32 locus, which were previously identified in the Boston dataset [ 6] (and were able to be mapped on the Agilent platform used in the Utah dataset). These were miR-495, miR-329, miR-487b, miR-410, and miR-656, and the resulting tdROC curve was highly discriminatory for recurrence at 120 months (area under the curve (AUC) = 0.743; Fig.  1a). Then, we mapped this five-miRNA profile on the Agilent array and assessed its performance in the Utah cohort again using the signed average method. In this analysis, the prognostic model was fully frozen (selected miRNA features and signs were predefined in the Boston cohort) and applied to the Utah cohort, and we found that the 5-miRNA profile maintained a very strong discriminatory power for overall survival at 60 months (AUC = 0.723, permutation p = 0.03; Fig.  2a). The 60-month endpoint was chosen due to the much shorter follow-up in the Utah cohort.

Given the rich miRNA content of the 14q32 chromosomal region, we were interested to test if the power for prognostic discrimination extended beyond the top 5 miRNA prognostic markers. Thus, we extended this analysis to a group of 18 miRNAs on the same locus that were part of a larger prognostic profile (that included the top 5) in the previous study and found that that they still accurately predicted an individual patient’s risk to recur in both the Boston and Utah cohorts (Figs.  1b and 2b). Then, we tested all 62 miRNAs that are located on 14q32, and we found that the collection of these miRNAs offered good discriminatory capacity as well in the Boston cohort (Fig.  1c). We could not perform this analysis in the Utah cohort due to significant differences in the probe content between miRNA DASL (cDNA-mediated annealing, selection, extension, and ligation) and Agilent miRNA arrays, such that a significant subset of the total group of 14q32 miRNAs could not be mapped on the arrays from the Utah cohort. These results demonstrate that significant prognostic power resides on the entire 14q32 non-coding region, and miRNA subsets from this locus can be used for very accurate prognostic discrimination, using a simple method that minimizes model overfitting.

For the patients who experience suboptimal response to standard preoperative methotrexate, doxorubicin, and cisplatin (MAP) chemotherapy (defined as <90% necrosis in the operative specimen), there remains uncertainly as to whether adding alternate chemotherapy regimens such as ifosfamide/etoposide (IE) offer any benefit, with studies to date, including a recent large randomized trial, failing to show survival benefit. We were interested to assess if the miRNA profiles may have prognostic interaction with chemotherapy choice. In the Boston dataset (the only one for which details of postoperative alternative chemotherapy regimens were available), we constructed multivariate models using the signed averaged expression values for the 5-miRNA and 18-miRNA profiles, together with two clinicopathologic covariates, namely chemotherapy-induced necrosis and use of postoperative alternate chemotherapy regimen in addition to conventional MAP chemotherapy. tdROC analysis showed improved prognostic power with the combined models (AUC = 0.852, permutation p < 0.01, and AUC = 0.854, permutation p = 0.07, respectively) at the 60-month follow-up time (Fig.  4) while use of chemotherapy as the sole prognostic variable was (as expected) not prognostic, and analysis combining the profiles with use or not of alternate chemotherapy did not show any improvement in prognostic power. (Analysis at a follow-up time of 120 months produced similar, only slightly lower, AUC values compared to the 60-month time point analysis (Additional file 2). This observation raises the possibility that these miRNA profiles may affect outcomes in the context of pathologic necrosis and use of alternate adjuvant chemotherapy, although this would need to be proven in a larger prospectively designed study, where predictive interaction with chemotherapy can be statistically assessed.

We then considered the hypothesis that 14q32 miRNAs are not simply markers of prognosis but signify previously unrecognized distinct molecular subtypes of osteosarcoma. To address this, we first ascertained that sample assignments as “high” or “low” risk were not sensitive to the precise classification algorithm that was used. For example, class assignments were highly similar whether the 5-miRNA profile was used in the signed averaged expression model (described in the previous paragraph) or in a clustering-based grouping (Fisher’s exact p < 0.001) or in the supervised multivariate 5-miRNA model previously published (Fisher’s p < 0.001). This observation also held true in the Utah dataset (Fisher’s p < 0.001, for the signed average-based versus clustering-based grouping). Further, the classification was stable when we used a larger set of 18 miRNAs from this locus to classify samples using any of these methods in both datasets (Fig.  5a, b). Finally, similar concordance was observed in the Texas data, where the 3-miRNA model classification was highly associated with the groups generated by hierarchical clustering using 3, 5, and 23 of the prognostic miRNAs mapped on the Taqman platform (Fisher’s p = 0.04, 0.07, and 0.03, respectively). Confirmation of strong similarity in terms of risk group assignments for the samples independent of the specific number of miRNAs and grouping method suggests that the miRNA markers track potential underlying molecular phenotype.

If the two prognostic risk groups represent molecular subtypes, one might expect that they display large-scale molecular differences in addition to the marker 14q32 miRNAs. Thus, we performed global miRNA differential expression analysis between the high- and low-risk groups in the Boston dataset (which were defined in our previous report) and found that 492 miRNA probes (64%) were differentially expressed across the risk subtypes ( t test p < 0.05; Benjamini-Hochberg false discovery rate (FDR) <0.07; Fig.  6a, Additional file 3). Since the proposed subtypes are not yet established, and in order to ascertain that these subtype-related miRNA expression differences are not due to a statistical artifact, we generated 100 random splits of the samples and tested them for differential miRNA expression. None of these randomly generated groups yielded a similar level of global miRNA differential expression. Multidimensional scaling (principal component analysis) was then performed with the global miRNA content (all miRNAs only filtered by low variance), and the three-dimensional sample groups were highly associated with the 5-miRNA model risk predictions (Fisher’s p < 0.001; Fig.  7a). Of interest, hierarchical clustering the samples using the global miRNA content provided a trend for prognostic discrimination (median RFS 126 versus 151 months; log-rank p = 0.092; Fig.  7b), though it did not reach the level of statistical significance achieved when using the 14q32 miRNA markers. We then assessed the correlation of the miRNAs with DICER, a key gene involved in the miRNA biogenesis and processing machinery. We found that only a small fraction of the miRNAs that were differentially expressed between the subtypes (5%) were moderately or strongly correlated with DICER1 (Spearman rank correlation coefficient = 0.44–0.51; p < 0.01; Additional file 4), suggesting that DICER may account for (only) a small part of the miRNA deregulation in the subtypes.

These findings were reproduced in the Utah dataset, where 546 miRNA probes (36%) were differentially expressed across survival risk groups ( p < 0.05; FDR <0.139; Fig.  6a, Additional file 3), but with no similar differences observed in any of the 100 random sample splits. Multidimensional scaling/principal component analysis using the global miRNA content also demonstrated that the risk groups were significantly associated with the 5-miRNA signed averaged expression-based predictions (Fisher’s p = 0.05; Fig.  7c) and classification was highly concordant regardless of the number of miRNAs used in clustering. Finally, in the Texas dataset, hierarchical clustering using different numbers of miRNAs (3, 5, 23) produced highly similar groups, which were also similar to clustering using the 14q32 miRNAs or even the global miRNA content (Fisher’s p < 0.05 for all comparisons). These results from the three datasets support the notion that these risk groups represent true molecular osteosarcoma subtypes with substantial underlying molecular differences beyond the set of 14q32 miRNAs.

We then explored this hypothesis in a publicly available miRNA dataset from 19 osteosarcoma cell lines [ 11]. In this dataset, experimental data were available, grouping the cell lines according to their proliferative capacity. We compared highly aggressive and less aggressive cell lines (by virtue of high proliferation versus low proliferation, as previously published [ 12]) and again found a substantial amount of differential miRNA expression. Importantly, we also found a moderate to strong amount of overlap in subtype-specific differential expression among the two clinical and the cell line datasets. Figure  6 shows the scale of differential miRNA expression between high- and low-risk subtypes and the degree of overlap between the two clinical datasets, which was highly significant using a hypergeometric distribution test. Generally, there was a much stronger concordance in miRNAs upregulated in the aggressive phenotype than in miRNAs downregulated in the aggressive phenotype among the datasets. Details of the cell line miRNA differential expression analysis are provided in Additional file 3. The observation of reproducible large-scale miRNA overexpression in aggressive samples in two clinical and one in vitro datasets further supports the hypothesis of distinct molecular osteosarcoma subtypes with different biologic and clinical behavior.

We also studied global mRNA transcript changes (other than microRNA), which were available for a subset of 37 samples from the Boston dataset, with respect to the risk subtypes, and found that 1362 mRNA probes (13%) were differentially expressed across risk subtypes ( p < 0.05; FDR <0.275) in the Boston dataset. Although these changes were statistically more modest compared to the respective miRNA changes in supervised univariate analysis (partly due to the smaller sample size of the mRNA dataset), they still are consistent with the notion of largely different transcriptional programs between the two subtypes. Also, unsupervised genome-wide gene expression (mRNA) clustering generated subgroups (Additional file 5) that showed a strong trend for association with both the 5- and 18-miRNA previously defined model risk predictions (Fisher’s exact p = 0.05–0.08 for different association tests). These observations further speak to the possible existence of distinct molecular phenotypes. Interestingly, the list of differentially expressed genes was enriched for all the transcripts on the 14q32 genomic locus, taken collectively as a single gene set or potential functional unit (KS/LS enrichment p = 0.06). This collection includes 124 coding genes and 95 non-coding RNAs (long non-coding RNAs and small nucleolar RNAs (snoRNAs), which were included in the DASL whole genome array), raising the possibility of a regional mechanism of coordinated regulation affecting both coding and non-coding RNA elements on the 14q32 locus.

In a hypothesis-based approach, we tested the 14q32 miRNAs and found that several of them were also highly correlated with other non-coding genes located on 14q32. For example, MEG3, a long non-coding RNA, was also highly correlated with several of the miRNAs, as were numerous snoRNAs located near the 14q32 miRNA cluster ( p < 0.05, uncorrected because of the small number of variables). These correlations (Table  1) suggest a potentially highly coordinated mechanism of expression regulation including coding as well as non-coding elements within the larger imprinted 14q32 chromosomal region. A list of all significant correlations of the 14q32 genes with prognostic miRNAs is shown in Additional file 6.

We used the 19-osteosarcoma cell line dataset (introduced above) to test the hypothesis that 14q32 miRNAs define distinct osteosarcoma subtypes in vitro, correlated with tumor aggressiveness as predicted by our genomic analysis of the clinical osteosarcoma cohorts. In this study, five variables were used as metrics of cancer cells’ aggressiveness: tumorigenicity, colony-forming ability, invasion, migration, and proliferation [ 13]. We selected all miRNAs on the 14q32 locus, which were associated with recurrence in univariate analysis, in our previous study of the clinical Boston cohort (Cox regression p < 0.05), resulting in a matrix of 27 miRNAs (a subset of which is the 5-miRNA profile shown above).

We studied the association between the 14q32 prognostic miRNAs with cell line “aggressiveness” metrics, as reported in the public dataset. We found that the median expression levels of the 27 prognostic miRNAs (in aggregate) were higher in the aggressive cell lines (defined by proliferative capacity, Mann-Whitney p = 0.02). In addition, unsupervised hierarchical clustering of the cell lines using the expression patterns of these miRNAs distinguished a cluster of mainly aggressive cell lines based on a composite metric of migration/invasion/colony-forming capacity, which can be viewed as an in vitro surrogate for metastatic potential (Fig.  8, Fisher’s p = 0.05 for association between the composite metric and miRNA-based cluster groups).

Further, we found a number of significant or strongly trending ( p < 0.1) associations testing each individual miRNA in relation to the different aggressiveness metrics of the cell lines (Additional files 7 and 8). In a two-group differential expression analysis (high versus low proliferative cell lines), 13 miRNAs were upregulated in the more proliferative cell lines. Eight miRNAs were upregulated in highly invasive cell lines, while one was downregulated. Three miRNAs were upregulated in cell lines with increased migratory ability, while two were downregulated. Two miRNAs were upregulated in cell lines with higher colony-forming ability, while two miRNAs were upregulated in cell lines with higher tumorigenicity. Generally, the large majority of the miRNAs appear upregulated in the more aggressive phenotype, with the exception of hsa-miR-493 and hsa-miR-411*, which were consistently downregulated in the aggressive cell lines.

To circumvent possible pitfalls of binarization in cell line assays, we also analyzed proliferation as a continuous variable. In this analysis more (17) of the 27 prognostic miRNAs were highly or moderately correlated either positively (Spearman coefficient = 0.394 to 0.646) or negatively (Spearman coefficient = −0.403 to −0.582) with greater levels of proliferation at 72 h, which was the end time point of the proliferation experiments in the public data (all at p < 0.1). Two miRNAs were highly or moderately correlated with colony-forming ability (Spearman = 0.618, 0.653; p < 0.1) while three miRNAs were highly or moderately correlated with invasiveness (Spearman = 0.405, −0.418, −0.54; p < 0.1). Again, the majority of these correlations were positive, while hsa-miR-493 expression was negatively correlated with migration capacity at a significant or trending level (Spearman = −0.453; p < 0.1). This miRNA was the only one consistently negatively associated with the other more aggressive phenotypes in all types of statistical analysis. Proliferation appears to be the single individual attribute better capturing the extent of association between 14q32 miRNAs and cell line aggressiveness. However, the composite invasion/migration/colony-forming capacity metric also showed a clear association with aggressiveness, as a possible surrogate for clinical metastatic potential.

The hallmark of 14q32 is allele-specific methylation (imprinting). In order to explore the genomic context as it relates to imprinting control on miRNAs, we first considered a map of the locus containing the genes, non-coding RNAs (miRNAs or others), and CpG densities localized on 14q32 (Fig.  9a). Most of the prognostic miRNAs (previously defined in the clinical cohorts) are generally clustered in a 350-kb region of 14q32, which also contains the majority of all other non-coding RNAs on this locus, including 41 out of 47 snoRNAs. The non-coding RNA cluster also includes the DLK1-DIO3 differentially methylated region (DMR), which includes both the intergenic DMR (IG-DMR) and the MEG3 DMR and controls imprinting of this locus (Fig.  9b). Upon further exploration, we noted that there are several CpG islands (CGIs) within the 14q32 miRNA/non-coding cluster. CGIs are unmethylated in “normal tissue”; however, variable degrees of CGI methylation have been associated with various disease states including cancer [ 14].

Based on these observations, we were interested to study the association between miRNA expression and DNA methylation. The non-coding RNA cluster lies in the DLK1-DIO3 imprinted region, which also includes the long non-coding RNAs, MEG3, MEG8, and MEG9 (Fig.  8), and is regulated by at least two DMRs, the intergenic DMR and the MEG3 DMR [ 15– 17], with involvement of CCCTC zinc finger-binding factor (CTCF). Recent data suggest a process of loss of imprinting (LOI) in this region, which is involved in cancer development [ 18], and a correlation between methylation patterns in the MEG3 promoter and expression of several miRNAs in the 14q32 ncRNA cluster has been suggested. The direction of the correlation may be different for methylation of CTCF binding sites (positive) versus conventional promoter methylation (negative) [ 16]. This is consistent with the known role of CTCF as a transcription enhancer blocker. We hypothesized that this regulatory pattern may hold true in osteosarcoma, and we tested this hypothesis using publicly available miRNA expression (Agilent array) and methylation data (Illumina 27K array) from the previously mentioned 19-osteosarcoma cell line dataset.

We focused on seven CpG sites interrogated by the 27K array, which are located upstream of the miRNAs and proximal to the MEG3 promoter (no intergenic probes are interrogated by this array). One of these CpG sites, corresponding to the probe cg09280976, is located within a known CTCF binding site in the MEG3 promoter, and another one, corresponding to the probe cg04291079, is located in the MEG3 gene body, in a region where there is no known CTCF binding site. The other five probes are located close to but not entirely within CTCF binding sites. In our analysis, we found that 7 of the 27 14q32 prognostic miRNAs showed moderately or highly positive correlation with the methylation probe cg09280976 at a significant or strongly trending level (Spearman = 0.596–0.404; p < 0.1; Table  2). Eight of these 27 miRNAs were moderately or highly negatively correlated with the methylation probe cg04291079 at a significant or strongly trending level (Spearman = −0.393 to −0.552; p < 0.1; Table  2), while two were moderately positively correlated (Spearman = 0.554–0.400; p < 0.1; Table  2). Methylation probes located near, but not entirely within, CTCF binding sites showed variable, positive and negative, correlations with miRNAs (Table  2).

We also found that expression of 23 of the 27 prognostic miRNAs was also positively correlated with MEG3 expression (Spearman = 0.808–0.409; p < 0.1) while only one miRNA was negatively correlated (Spearman = −0.582; p < 0.1). Finally, four of the methylation probes in the non-coding RNA cluster were positively correlated with MEG3 expression (Spearman = 0.528–0.41; p < 0.1). These results, taken together, suggested a possible methylation-based regulation of both short and long non-coding RNAs in this imprinted genomic region.

We then sought to validate these observations using a genomic resource that recently became available by the NIH. The NCI Therapeutically Applicable Research to Generate Effective Treatments (TARGET) osteosarcoma project generated several genomic profiles for a large number of patient samples and includes DNA methylation data (on the more advanced Illumina 450K methylation array) as well as miRNA TaqMan qRTPCR expression data. For most of the MEG3 CpG sites targeted by methylation probes, we found remarkable similarity between the moderate or moderately strong methylation/miRNA correlations observed in the cell lines (Table  2) and the correlations seen in the TARGET data (Fig.  10). This was particularly striking, for the two methylation sites were previously identified as clearly within (targeted by probe cg09280976) or clearly not within a CTCF binding site (targeted by probe cg04291079), both showing the same positive or negative association with miRNA expression seen in the cell line analysis. The rest of the MEG3 methylation sites showed generally similar variable associations seen in the cell line data with the exception of two probes within the gene body. These differences could be possibly related to the small sample size and inherently heterogeneous nature of the cell lines, as well as the fact that the effect of gene body methylation is much less clear than that of promoter methylation. Also, one promoter site (probe cg1656), which did not show significant associations in the cell lines, did show moderately strong associations in the TARGET data, which probably further supports the original hypothesis as it is located very near a CTCF binding site and it should be expected to show a clear positive correlation with expression. Due to publication restrictions currently in place by the NCI on the osteosarcoma TARGET data (see “ Methods”), we are only allowed to provide numerical details of the correlation coefficients for a small subset of the miRNAs. Therefore, as an example, we show the average correlation coefficients for the following four miRNAs: miR-495, miR-329, miR-656, and miR-411*, which were some our top miRNA prognostic markers in the three previously analyzed clinical datasets (Boston, Utah, Texas). For these four miRNAs, the average correlation coefficients were as follows: cg09971646, 0.2105; cg16567044, 0.408; cg09280976, 0.323; cg25836301, 0.219; cg05711886, 0.185; and cg15101633, 0.199. These correlation coefficients were largely unchanged when the larger group of prognostic 14q32 miRNAs was considered.

Of interest, additional three CpG sites targeted by probes in the arrays and located in the DLK1 gene locus showed moderate, variable (positive/negative), or no association with the miRNAs, potentially to be expected given their localization ~100 kb upstream of the MEG3 transcription start site (TSS), further underscoring the potential specificity of the MEG3 methylation effect on the non-coding RNA cluster.

Next, we used the same cell line data to determine if methylation patterns in the non-coding RNA cluster correlated with aggressiveness phenotypes. First, we performed hierarchical clustering using the methylation patterns of the seven MEG3 methylation probes analyzed in the previous section and examined the resulting clusters for the 19 cell lines. We observed that the cell line clusters were significantly associated with aggressiveness (measured by proliferation, Fisher’s p < 0.05 when using either two or three proliferation groups, high/low or high/intermediate/low, as previously published) (Fig.  11a). Similar results were obtained when we repeated the clustering including additional four probes from the 27K array located in the DLK1 promoter and gene body (recognizing that methylation effects often represent an aggregate of changes in a broader genomic region, especially when CTCF binding is involved).

In a two-group differential methylation analysis between two proliferation phenotypes (high/low), 7 of 11 CpG sites had higher methylation in cell lines with increased colony-forming ability ( p < 0.1) (high and low categories as previously published). Four CpG sites had higher methylation in more invasive cell lines, cell lines with greater migratory capability ( p < 0.1) as well as cell lines with higher tumorigenicity ( p < 0.1), compared to their respective less aggressive counterparts. In a continuous variable correlation analysis, methylation at five CpG sites was positively correlated with colony-forming ability (Spearman = 0.634–0.392; p < 0.1). Methylation intensity at one CpG site near DLK1 was also positively correlated with invasive ability (Spearman = 0.599); trending p < 0.1) while another CpG site was positively correlated with migration (Spearman = 0.5; p < 0.05). Detailed results of the differential methylation analyses and the continuous variable correlation analyses are included in Additional file 9.

We then hypothesized that there may be a three-way association between prognostic miRNA expression, methylation, and cell line aggressiveness. As an example, we focused on two CpG sites represented by Illumina 27K array probes cg09280976 and cg04291079, because, as explained above, their corresponding locations are known to be clearly within or clearly outside known CTCF binding sites and it was therefore conceptually easier to predict biologic associations. For each probe, we plotted CpG methylation versus average expression of miRNAs that were associated with that probe and with proliferative capacity at a Spearman p < 0.1 versus proliferative capacity as a continuous variable. The resulting 3-D plots (Fig.  11b, c) recapitulate possible biologic interactions in this imprinted region. For the probe that is located within a CTCF binding site, we observe two cell line clusters in the “corners” of the plot displaying the patterns “high methylation/high expression/high aggressiveness” and “low methylation/low expression/low aggressiveness,” respectively, consistent with the known transcription enhancer-blocking role of CTCF in the imprinted locus (Fig.  11b). For the probe that is located outside CTCF binding sites, we also observe two cell line clusters in diagonally opposed corners as compared to panel b. These clusters display patterns of “high methylation/low expression/low aggressiveness” and “low methylation/high expression/high aggressiveness,” respectively, consistent with a conventional negative regulatory effect of promoter methylation on expression. Associations between individual probe methylation and aggressiveness showed a strong trend for significance in panel b, while in panel c they did not, possibly related to the possible heterogeneity in the cell line group and the presence of some cell lines that did not necessarily follow these clear biologic patterns. Also, it is unlikely that one methylation probe, in isolation, can fully account for overall tumor behavior especially when its regulatory effect would be contrary to the broader methylation effect in this same region.

Expression levels of miRNAs do not fully reflect their biologic activity. We previously reported on miRNA gene targets with gene set enrichment and found modest evidence that some miRNAs appear to exert regulatory influence [ 6]. However, this approach assumes that the regulators affect each gene equally and does not take into account multiple sources of information on gene regulation. Inference on biologic activity can be drawn by analyzing miRNA regulation of target gene expression in the context of a gene regulatory network. Our group recently described a network inference method for network that models the regulatory effects of transcription factors. This method (called “PANDA”) integrates different “omics” data types to infer network edges that accurately estimate gene expression regulation [ 19, 20]. Here, we implemented a modification of PANDA (called “PUMA,” PANDA Using MicroRNA Associations), which estimates regulatory effects of miRNAs on gene expression. We implemented PUMA to reconstruct networks for the high- and low-risk osteosarcoma subtypes using mRNA expression data from the Boston clinical cohort and “prior” data from the STRING database, JASPAR, and TargetScan. The gene regulatory networks showed substantial differences in gene targeting between the two risk groups. This is interesting in light of our earlier finding of a large number of differentially expressed miRNAs between the high- and low-risk patient groups, especially because our network construction model did not incorporate miRNA expression levels. In order to place this finding in a wider context, we analyzed chemotherapy response phenotypes in a similar manner and found much less differential miRNA targeting between groups of patients with optimal versus suboptimal chemotherapy-induced necrosis (Fig.  12a). We then focused on a network module derived by the 5-miRNA prognostic profile to determine what differences in gene regulation may be driven by these miRNAs (Fig.  12b, c). We identified significant differences in multiple edges of these networks using a permutation test on sample labels (Additional file 10). Patients with poor prognosis showed a particularly active module targeted by miR-495. The top differential edges of this network included at least two tumor suppressor genes ( GAS1 and CD9). Although miR-495 showed the strongest overall statistical signal for perturbed regulation, the other four miRNAs also contributed a number of statistically significant differential edges. These included multiple genes of known or potential significance including, for example, another known tumor suppressor gene in osteosarcoma ( RASF5), which was differentially regulated by two of five miRNAs in this network module. Thus, increased targeting of tumor suppressor genes may be a mechanism by which these prognostic miRNAs individually or synergistically could affect tumor behavior and prognosis. Furthermore, Gene Ontology analysis of target genes in the network modules suggested that these miRNAs may differentially regulate pathways of cellular senescence, insulin signaling, osteoblast differentiation, and processes related to cellular proliferation and cell cycle and growth signaling (Additional file 11).

We then curated the list of all significant edges from the 5-miRNA network (in all 652 edges/637 genes) and screened it against the Drug Gene Interaction Database, a previously published interactive database of gene drug interaction supported by clinical or preclinical evidence. Figure  12d shows a selected subset of these interactions, suggesting possible hypotheses for therapeutic development including histone deacetylase (HDAC) inhibitors, proteasome inhibitors, or mTOR inhibitors, some of which are already in clinical or preclinical development in sarcoma or osteosarcoma. A detailed list of possible gene-drug interactions from the networks is provided in Additional file 12.

We used three previously published clinically annotated human osteosarcoma datasets. One consisted of DASL miRNA expression data from 65 and mRNA data from 37 diagnostic biopsy specimens from Beth Israel Deaconess Medical Center and Boston Children’s Hospital (called “Boston dataset,” GEO accession GSE39040), and another consisted of Agilent miRNA expression data from 27 frozen tissue specimens from the University of Utah (called “Utah dataset,” Array Express accession E-MTAB-1136). The third dataset consisted of ABI TaqMan human microRNA qRTPCR data from 25 frozen diagnostic biopsy samples from the University of Texas Health Science Center (called “Texas dataset,” GEO accession GSE79181, details of which were published before [ 10]). Cell line miRNA data (Agilent arrays) and methylation data (Illumina 27K array) were derived from 19 osteosarcoma cell lines (published by the Institute of Cancer Research, Oslo University, GEO accession GSE28425, GSE36004). Transcription and methylation array details as well as clinical cohort and cell line annotations have been previously reported [ 6, 9, 11, 47].

We also used new data provided by the NCI, which launched the TARGET initiative, producing a repository of large-scale genomic data from a number of rare pediatric cancers. Within the osteosarcoma TARGET project, methylation profiles for 86 osteosarcoma patients (Illumina 450K array) and miRNA expression profiles for 89 osteosarcoma patients (MegaPlex TaqMan) became recently publicly available ( http://​target.​nci.​nih.​gov/​dataMatrix/​TARGET_​DataMatrix.​html, retrieved November 10, 2016). The NCI has currently placed a limitation on publishing findings from analyzing osteosarcoma TARGET data. ( https://​ocg.​cancer.​gov/​programs/​target/​target-publication-guidelines). This limitation allows investigators to only publish data from a focused analysis of a handful of genes, until the primary osteosarcoma project TARGET investigator team publishes their first “global” analysis of the genomic data. In complying with this limitation, any TARGET-derived data we present here are only related to methylation and expression analysis of a very small number of genes and miRNAs, all focused on the 14q32 locus. Therefore, our analysis is in no way similar in scope to the (currently unpublished) global genomic investigation of osteosarcoma undertaken by the TARGET initiative.

For any global expression analysis in the Boston, Utah, and Texas datasets, variance filtering was performed excluding 33% of probes with the lowest variance for the miRNA arrays and 66% of probes with the lowest variance for mRNA arrays, before performing any genome-wide (global) analyses. MiRNA expression values were quantile normalized before being subjected to further statistical analysis. For analyses involving methylation intensity, we utilized the M value (a transformation of the conventional beta value) as it has been shown to possess better statistical properties for differential analysis [ 48]. Further methodological details related to processing and analyzing these data are provided in Additional file 13.

In order to avoid overfitting, we used the signed average method and leave-one-out cross-validation in all survival analyses. In this approach, we averaged the expression levels of individual miRNA features in each candidate profile, weighted only by the sign of their hazard ratio (positive or negative) in univariate Cox regression analysis, and the resulting signed averaged metric was used as the prognostic index [ 49– 51]. Kaplan-Meier analysis with log-rank test and Cox regression were used to analyze or model recurrence and survival as necessary. In the Utah dataset, time-censored recurrence data were not available, so we used overall survival as the time-censored endpoint. Time-dependent receiver operating characteristic (tdROC) and area under the curve (AUC) analyses were performed as previously described. For the Boston dataset, the tdROC endpoint was 120 months, while for the Utah dataset, it was set to 60 months due to the shorter follow-up available in that study. Penalized multivariate Cox regression analysis was performed as previously described [ 52].

Two-class comparison of continuous variables was done using the t test, and p values were corrected with the Benjamini-Hochberg FDR test while the Mann-Whitney test was used for non-parametric two-class continuous variable comparisons. In the case of testing a small number of variables based on prespecified hypotheses, we present unadjusted univariate p values. Univariate two-sided p values <0.05 were considered significant, except in some exploratory analyses in cell lines involving a limited number of preselected methylation and expression probes, where we also report associations at a p < 0.1 as suggestive of biologic inference. When examining the proposed molecular tumor subtypes, not previously established in the literature, we generated 100 random splits of the samples and considered that the expression differences between the proposed subtypes are significant if less than 5% of the random splits showed the same degree of differential expression as the proposed subtypes. We performed gene set analysis using the KS/LS statistic to determine if the expression profiles of the high- and low-risk groups were enriched for gene sets or miRNA targets of interest, according to the functional class scoring method [ 53]. Unsupervised hierarchical clustering was performed with the average linkage method as previously described [ 54]. We also used multidimensional scaling to construct 3-D representation of sample locations in the multivariate expression space.

Data plots showing gene coordinates and CpG island frequencies were plotted using Circos [ 55]. Chr14q32 was defined as the region between the co-ordinates 89,800,000 and 109,000,000 on chromosome 14. The location of genes, non-coding RNA, and miRNAs within these co-ordinates were derived from UCSC hg19 gene annotation tables. UCSC bed file annotations of CpG island locations in hg19 were used to determine CpG island location. CPG frequency was determined by counting the number of CpG islands within 100,000 bp upstream a given location. CTCF binding sites in MEG3 were determined in previous studies [ 16, 17]. These sites were cross-referenced with the University of Tennessee CTCF binding site database, which includes data from the UCSC genome browser and the Broad Institute [ 56]. Search for enhancer sites was performed using the VISTA Enhancer Browser (Lawrence Berkeley National Laboratory) as previously described [ 33].

We used a method called PUMA (PANDA using MicroRNA Associations, Kuijjer et al., in preparation). This is a network reconstruction algorithm that models gene regulation by miRNAs and transcription factors and is an extension of our previously published network reconstruction method PANDA [ 19], which uses message passing between regulatory, protein-protein interaction, and gene expression data to model information flow between regulators and their target genes. PUMA extension works similar to PANDA, but in the message-passing steps does not allow microRNAs to form edges in the protein-protein interaction network, thereby indirectly influencing microRNA-target gene edges. We reconstructed networks for both subtypes and generated a set of “background” permuted networks to estimate edge significance (defined as activity score outside the four standard deviation range of the background distribution. Details are provided in Additional file 14, and implementation of the PUMA algorithm is available at https://​github.​com/​mararie/​PUMA. For drug target identification, we utilized the Drug Gene Interaction Database (DGIdb) as previously described [ 57].

We used the NCI BRB-ArrayTools software (developed by Dr Richard Simon and the BRB-ArrayTools Development Team) and the SPSS software, version 18 (IBM Corporation, NY).