We, and others, previously reported that a collection of miRNAs on the 14q32 locus is associated with prognosis in osteosarcoma [
]. Prognostic association via a Kaplan-Meier analysis does not guarantee the value of a test for predicting individual patient course. Thus, our tdROC analysis complements prior findings and demonstrates an excellent discriminatory power for miRNAs in this locus, which was replicated in three separate miRNA datasets. This replication is quite significant when one considers the very different characteristics of the three datasets (different array platforms, frozen versus paraffin tissue material, somewhat different clinical characteristics with respect to age distribution and the mix of metastatic/nonmetastatic cases, and a relatively small number of recurrence/death events in two of the datasets), which may also explain any minor differences in the findings among the three datasets. New studies will be required, in larger multi-institutional cohorts, in order to select the optimal subset of miRNAs among the 50–60 present at 14q32, for a clinically useful test to be developed. While potentially different subsets of these miRNAs could also be appropriate markers, we believe that the 5-miRNA or 3-miRNA profiles described in this study would be excellent candidates.
Evidence to date has been disappointing with respect to the use of alternate chemotherapy regimens, such as ifosfamide/etoposide, for patients who do not achieve an optimal pathologic response to standard preoperative chemotherapy with the MAP regimen. This was further underscored by the recent publication of the results from the large international randomized EURAMOS-1 study. This may signify true lack of clinical benefit from these regimens, but it could also suggest lack of an optimal stratification approach for selection of the right patient subset. Our findings suggest that these miRNAs may also be candidates for further study as possible markers for selecting patients for such alternate regimens, in stratification schemes that may include miRNA expression and conventional pathologic necrosis in the operative specimen. Proof of this concept will require analysis in a large cohort with prospectively randomized chemotherapy regimen allocation, such that a formal test of interaction can be performed. It was inherently not possible to perform such analysis in our data, but our findings would perhaps justify such a study in the future.
We suggest that, in addition to being prognostic markers, these miRNAs may track previously unidentified osteosarcoma molecular subtypes, potentially related to imprinting defects at 14q32. We found that the subgroups of patients defined by the miRNA risk profile harbor substantially different global (genome-wide) miRNA expression patterns, as well as different mRNA expression patterns. The miRNA patterns were more strongly associated with patients’ outcome compared to mRNA patterns. It is quite possible that miRNAs are better surrogates for tumor behavior, given their capacity to regulate large numbers of coding genes. In addition, miRNA detection may be more degradation proof than mRNA detection in banked tissue, though our results are possibly confounded by the smaller sample size of the mRNA dataset (compared to the miRNA dataset). That having been said, the characterization of patient samples as “high risk” versus “low risk” was consistent regardless of which particular method (supervised or unsupervised) or subset of miRNAs was used for sample risk prediction, and clustering-based groups were associated whether using either the global miRNA or the global mRNA data. Furthermore, the large-scale miRNA differences between high- and low-risk samples were also reproducible in two additional clinical and one in vitro datasets. Taken together, these observations support the notion of robust molecular subtypes in osteosarcoma with very different transcriptional programs, coding and non-coding.
Using the largest (published to date) osteosarcoma cell line dataset with genome-wide molecular information, we found that the expression patterns of the 14q32 miRNAs as a function of cell line aggressiveness were largely concordant with the findings in the clinical datasets. Cell line proliferation seemed to be the single best correlate of individual miRNA expression, although a composite metric of invasion/migration/colony formation that was used as a surrogate for metastatic potential also showed a clear association with collective 14q32 miRNA patterns. These findings notwithstanding, it should be acknowledged that attributes such as invasion and migration and overall clinical metastatic potential may also be more heavily dependent on tumor-stroma interactions and other elements, which are generally lost in cell line systems and which our current bioinformatics analysis in the cell lines may not adequately capture. Some of the individual miRNA associations presented are only strongly trending as opposed to nominally significant in the cell lines; however, the cell line datasets are often imperfect correlates of clinical observations, and the cell line dataset was much smaller compared to the clinical datasets, where the associations have been shown to be more robust. Also, 14q32 miRNAs can probably be considered functionally interrelated, and when we analyzed them in aggregate via clustering, their association with cell line aggressiveness was robust. Previous reports have demonstrated both tumor-promoting and tumor-suppressing effects of 14q32 miRNAs in different settings, but the outcome-related findings in our study are consistent with previously reported effects of these miRNAs in other malignancies such as leukemia, lung cancer, and liver cancer as well as their growth-promoting effect in mouse pluripotent stem cells [
Network analysis using inferred miRNA-mRNA regulatory events by integrating target prediction with mRNA data provided insights into potential 14q32 miRNA-driven mechanisms in tumors with different prognosis and suggested that these may include perturbation of tsumor suppressor genes [
]. Furthermore, this analysis recapitulated the strong correlation between the miRNAs and tumor aggressiveness showing that their target networks include a large number of proliferation and cell cycle-related pathways and further supporting the notion of distinct regulatory programs affecting cellular behavior in the proposed osteosarcoma subtypes. To place these results in context, we observed a much wider network perturbation related to tumor aggressiveness and recurrence, compared to chemotherapy response, an endpoint that is more proximal in the natural history of the tumor. This leads us to speculate that overcoming short-term chemotherapy resistance may prove easier than achieving long-term remission or cure in osteosarcoma. This network analysis is limited by the fact that it provides in silico evidence, requiring future functional elucidation. Due to the multidimensional and multi-interactive nature of the networks, complex experimental designs will be needed in order to provide optimal in vitro systems for functional exploration of these networks. However, this is one of the first large-scale attempts to generate such networks in human specimen cohorts in this rare tumor.
The 14q32 chromosomal band uniquely contains a very large cluster (>100) of non-coding RNAs, including snoRNAs, microRNAs, and long non-coding RNAs, which is the largest known miRNA cluster in the genome (54 miRNAs). Genetic defects at 14q32 have been associated with severe developmental abnormalities, suggesting a very tight regulatory role in early tissue growth and differentiation. In our study, we noted a high degree of coordinated expression between miRNAs, snoRNAs, and long non-coding RNAs, suggesting perhaps an integrated mechanism of expression regulation in this region. Furthermore, 14q32 is an imprinted genomic region [
]. Imprinting is defined as allele-specific expression and is found in genomic regions critical for tissue growth and embryonic development. Typically, non-coding RNAs are expressed on the maternal allele while coding RNAs are expressed on the paternal allele, controlled by allele-specific methylation in genomic areas called imprinting control regions, or differentially methylated regions (DMRs). Disruption of this mechanism, called “loss of imprinting” (LOI), has been described not only in developmental abnormalities but also in cancer [
]. An enhancer-blocking factor, CCTC zinc finger-binding factor (CTCF), is also involved in gene and miRNA expression control in imprinted regions such as the
and 14q32 loci, in a methylation-sensitive manner. Specifically, it binds to unmethylated insulator sequences on DNA, preventing active enhancer-promoter interactions, thereby reducing transcription [
Our data support a three-way interaction between methylation, miRNA expression, and the phenotype (tumor aggressiveness), possibly contingent upon CTCF binding activity. Specifically, expression patterns of a large subset of the 14q32 miRNAs were associated with methylation patterns in the
G3 DMR region, and both were associated with tumor aggressiveness. The mechanism behind this interaction is likely complex. We found that hypermethylation of
DMR sites within a CTCF binding domain is associated with increased miRNA expression (potentially by inhibiting CTCF binding and its enhancer-blocking effect as described above) and higher tumor aggressiveness. In contrast, hypermethylation of
DMR sites outside a CTCF binding domain is associated with decreased miRNA expression and decreased tumor aggressiveness. A very similar mechanism of methylation-sensitive CTCF binding and miRNA expression control at 14q32 was recently described in acute promyelocytic leukemia, and other reports have shown regulation of this imprinted domain by allele-specific enhancer activity in human embryonic stem cells [
]. This genomic region has been reported to contain a large number of enhancer elements [
], and further studies will be required in order to identify which of them may be involved in the regulation of non-coding 14q32 genes in osteosarcoma.
The basic methylation/expression associations discovered in the cell lines were also reproducible in the large clinical osteosarcoma TARGET dataset. Perhaps more importantly, findings related to methylation sites within or outside known CTCF binding domains were highly similar between the in vitro and the clinical TARGET data. Future functional characterization and validation of the CTCF-related methylation/expression loop will require elaborate targeted designs including perhaps CRISPR approaches, but our initial observations herein, both in an in vitro and in a large clinical dataset, provide first evidence in support of this hypothesis.
Both coding and non-coding 14q32 genes, taken collectively as a single gene set, were enriched in the overall transcription program differences characterizing the subtypes. We also found evidence of coordinated expression regulation of coding and non-coding genes on 14q32, most strikingly a strong positive correlation between snoRNAs and miRNAs. Whether all these changes or changes in genes like
(a known miRNA processing gene) are functional elements of a wider pathogenetic mechanism controlled by imprinting and related to the biology of the osteosarcoma subtypes remains to be clarified. While very little is known on the function of snoRNAs in general, it is interesting to note that overexpression of 14q32 snoRNAs has been reported to promote tumor growth in acute leukemia [
The methylation platform employed in the cell line dataset was not comprehensive enough to allow assessment of all relevant
DMR CpG sites as well as the intergenic DMR sites. In addition, recent reports have suggested that the effect of methylation on gene expression in osteosarcoma may be different for different genomic “compartments” such as promoter CGIs, CGI shores, enhancers, or intergenic regions. Pilot analysis in a small clinical cohort showed either hypermethylation or hypomethylation in osteosarcoma tumors with high recurrence potential, and one other recent study also provided initial evidence of and insight into the effect of 14q32 methylation patterns in osteosarcoma [
]. Our findings do not contradict but rather complement these reports and indicate that both IG-DMR and
DMR methylation in conjunction with loss of imprinting, possible gene enhancer function, and expression of the entire 14q32 non-coding RNA cluster should be thoroughly studied in relation to osteosarcoma biology and outcome.
It is unclear what fraction of osteosarcoma tumors harbors the methylator phenotype proposed here, and studies in small clinical cohorts may underestimate its incidence. Though methylation data on the clinical cohorts were not available, the miRNA-defined high-risk patient group appears enriched for this phenotype, and the cell line data suggest that it could affect about 20–25% of the cases. This would be consistent with data in other tumor settings where the CpG island methylator phenotype (CIMP) is a relatively rare epigenetic phenomenon [
]. In addition, methylation changes in different imprinting control regions or genomic compartments may lead to different phenotypes. Integrated studies utilizing methyl sequencing and global non-coding and coding gene profiling in large national cohorts, such as the NCI TARGET osteosarcoma initiative, will hopefully provide answers to these questions. In addition, these comprehensive studies can also fully address the possible influence of chromosomal amplification or copy number variation on miRNA expression. This was not examined in our study, though prior literature suggests that genomic copy number variation may not be a big factor in the regulation of gene expression at 14q32 [
Ultimately, this line of research will allow the development of a set of mechanistically relevant clinical biomarkers based on loss of imprinting and/or non-coding RNA expression, with important therapeutic implications. Randomized trials have tested the addition of interferon or ifosfamide and etoposide to the standard neoadjuvant chemotherapy consisting of cisplatin/doxorubicin/methotrexate in localized osteosarcoma patients, using chemotherapy-induced pathologic necrosis as a risk stratification marker [
]. While results to date point toward a lack of additional, or uncertain, benefit with either intervention, it is conceivable that rather than absolute lack of antitumor activity, these results reflect an imperfect marker for treatment stratification. Our prior work has shown that 14q32 miRNAs confer prognostic value independent of chemotherapy response, with the added benefit that they can be obtained early, at the time of diagnosis, as opposed to after 10 weeks of preoperative treatment [
]. As suggested above, a strategy that combines pathologic necrosis with miRNA biomarkers may allow for better treatment stratification in the future. Furthermore, development of methylation 14q32 biomarkers may enhance the discriminatory power of the miRNA assays, or even perhaps surpass them, as previous reports have suggested that methylation markers may be more stable and less susceptible to random variations over time in human cancer specimens [
]. Testing all these hypotheses in archived clinical trial material, such as that of the NCI TARGET initiative, could revolutionize an approach to adjuvant treatment as well as treatment with demethylating agents, such as decitabine, currently in clinical trials in metastatic osteosarcoma. In addition, new studies employing RNA and DNA methylation sequencing approaches will provide further depth in our understanding of the clinical and biologic effects addressed in this study. Finally, our drug-gene interaction screen, though computational in nature, was based on multiple sources of public experimental and clinical data, providing additional testable hypotheses for therapeutic development. Some of the drugs from our screen are already in clinical development for osteosarcoma (such as HDAC inhibitors), and an altered mTOR/PI3K/PTEN pathway was identified as a therapeutic target in 25% of tumors in a recently reported osteosarcoma genomic study [
]. One might envision a possible combinatorial and stratified application of these drugs, based on the molecular subtypes presented here.
Results published here are in part based upon data generated by the Therapeutically Applicable Research to Generate Effective Treatments (TARGET) initiative managed by the NCI. Information about TARGET can be found at
. Specifically, we wish to acknowledge the Children’s Oncology Group (COG) for collecting the tissues for TARGET as part of COG clinical and biological protocols (Peter C. Adamson, MD, Children’s Hospital of Philadelphia, Philadelphia, PA, USA) and the principal investigators for the Osteosarcoma project (Ching Lau, MD, PhD, Texas Children’s Hospital, Houston, TX, USA; Paul Meltzer, MD, PhD, National Cancer Institute, National Institutes of Health, Bethesda, MD, USA).
We also wish to thank Dr Yidong Chen (University of Texas Health Science Center at San Antonio) for assistance in the use of the osteosarcoma data generated by her group (GEO accession CSE79181).
This study was supported by National Cancer Institute grant 1R01CA178908 to DS and grant 1R35CA197449 to JQ. BHK was supported by the Gattuso-Slaight Personalized Cancer Medicine Fund at Princess Margaret Cancer Centre, the Cancer Research Society (Canada), Ministry of Economic Development, Employment and Infrastructure and the Ministry of Innovation of the Government of Ontario, and the Canadian Institutes of Health Research. DS and EC wish to acknowledge the generous support of Wendell Colson and Joanne Casper to the MGH Sarcoma Research Program.
The funders had no role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Availability of data and materials
KEH, ADK, and DS conceived the study. KEH, ADK, MLK, WB, AR, CG, HK, JG, APA, and DS performed the studies and analyzed the data. KEH, ADK, MLK, WB, AR, CG, HK, KJ, MCG, MSA, GC, FH, EC, ZD, JQ, BHK, and DS wrote the manuscript. All authors read and approved the final manuscript.
The authors declare that they have no competing interests.
Consent for publication
Ethics approval and consent to participate
All data were obtained from publicly available datasets. Ethics approval prior to performance of the previously published studies had been previously obtained at all participating institutions.
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.