Combined modality treatment in osteosarcoma has led to survival gains and fewer amputations, but outcomes have remained unchanged for over 20 years [2, 40]. Adoption of novel therapies is complicated by the lack of reliable means to prognostically stratify patients. Even though it is a useful surrogate, pathologically assessed chemotherapy-induced tumor necrosis assessed at the time of definitive resection, the only accepted prognostic variable, is imperfectly correlated with distant outcome especially for sub-optimally responsive patients [2, 41]. Genome-wide studies have provided valuable data on chemoresponse [5–7, 42, 43] but there have been no studies examining miRNA and mRNA expression profiles using continuous time-censored recurrence and survival as endpoints. Given the rarity of well-annotated frozen tissue repositories, we sought to develop clinical outcome predictors using FFPE tissue. Our successful attempt implies clinical applicability, and establishes FFPE tissue as an appropriate substrate for such studies in osteosarcoma, particularly for miRNA profiling.
We found a strong relationship between miRNA expression profiles and RFS, the first such observation in this disease. Using established methods , we developed several models predictive of recurrence independent of chemotherapy response, although future development would naturally focus on the smaller, simpler models (for example, the five miRNA profile; Figure 2H). We also demonstrated that miRNA models are prognostic independent of potential confounding by known prognostic factors, such as chemoresponse, tumor location, presence of metastasis at presentation or chemotherapeutic regimen variation (although the latter has not been definitively shown to be prognostic in osteosarcoma to this date). Interestingly, risk prediction improved when chemotherapy response and miRNA risk profiles were combined, suggesting that chemosensitivity and miRNA profiles capture non-redundant prognostic information. Indeed, many patients in our cohort who did not respond optimally to chemotherapy either had no recurrence, or had recurrent disease after a long remission. These findings are highly relevant to clinicians wishing to provide robust prognostic information and prioritize patients for different or novel treatment approaches. For example, the largest ongoing randomized study in osteosarcoma (AOST0331, targeting 1,400 patients worldwide) is studying the modification of chemotherapy postoperatively, in the case of suboptimal response to preoperative chemotherapy. Thus the impetus to incorporate powerful marker profiles, non-overlapping with chemotherapy response, for improved patient risk stratification is increasing. The prognostic association between miRNA profiles and OS was weaker, however the power of this secondary analysis was limited by a smaller number of events (deaths) and we still detected potential prognostic synergy between miRNAs and chemoresponse with respect to OS. A larger future study will allow a more definitive prognostic analysis with respect to OS to be performed.
In addition to a stringent internal cross-validation, we achieved external validation using the only other public miRNA dataset that recently reported outcome data . Although widespread differences between the two studies were challenging (FFPE versus frozen tissue, DASL versus Agilent array platform, continuous time-censored versus binary recurrence outcome, and few events), we were nonetheless able to independently validate the prognostic value of a large subset of miRNAs from our predictive models and were able to develop signatures, using these miRNAs (preselected from our discovery set) with impressive survival distinction, using death as an endpoint. Future studies with larger sample sizes and standardized platforms will take the next step to assess the wider reproducibility and performance of a fully specified model. However, these studies will require a long time to assemble multiple adequately sized tumor cohorts in such a rare disease. Thus, our data provide strong and necessary initial evidence supporting the wider reproducibility of prognostic miRNA profiles in osteosarcoma.
Some of our profile miRNAs have been previously reported [10, 15] but not in association with osteosarcoma clinical outcome. These were predominantly located at 14q32 in the genome - a locus associated with osteosarcoma and Paget's disease (a known risk factor for osteosarcoma) - strengthening the biological plausibility of our results. It has been suggested that rearrangements of chromosome 14 may play a role in altered miRNA expression in osteosarcoma . Because miRNA target-genesets of some of 14q32 prognostic miRNAs were also associated with outcome, potentially indicating miRNA activation, our data suggest that this genomic locus may have a significant role in osteosarcoma progression and outcome. Although reports in other tumor systems have suggested a proliferative and pro-invasive role , further studies are needed to characterize the precise mechanism by which some of these miRNAs may modulate outcome.
We also investigated the relevance of miRNAs to chemotherapy response and identified novel miRNA signatures predictive of chemosensitivity. These signatures are largely non-overlapping with overall recurrence and survival profiles, supporting the notion that chemoresponse and tumor progression and metastasis may be regulated by non-overlapping molecular networks. We also performed a 'dynamic' analysis, revealing miRNA expression changes following chemotherapy in 'resistant' samples (those with viable tumor at the time of resection). Although this analysis is exploratory and requires further validation in a larger study with additional controls, it is interesting to note that some of the miRNAs identified in this dynamic analysis - for instance miR-15b, and miR-132 - have also previously been reported in relation to chemoresponse [7, 11], which is consistent with our results.
mRNA profiles were weaker and did not show additive value. This was possibly due to a smaller sample size (fewer samples run on whole genome expression assays than miRNA assays) and/or unique susceptibility to mRNA damage in FFPE osteosarcoma tissues (compared to miRNA stability) related to the fixation or decalcification process, which might have blunted the biologic signal. Nonetheless, miRNA and mRNA prognostic synergy should be explored more efficiently in future larger studies. Another possible limitation of our study is the inclusion of adult together with pediatric cases. However, only two patients were older than 35 years, making it very unlikely that their potential biologically unique nature affected our results.