Improved risk stratification in myeloma using a microRNA-based classifier

Authors


Correspondence: Gareth J. Morgan, Section of Haemato-Oncology, Institute of Cancer Research, 15 Cotswold Road, Sutton, Surrey SM2 5NG, UK.

E-mail: gareth.morgan@icr.ac.uk

and

Antonino Neri, Department of Clinical Sciences and Community Health, University of Milan and Hematology 1, Ca' Granda Foundation, Policlinico Hospital IRCCS, Via F. Sforza 35, 20122 Milan, Italy.

E-mail: antonino.neri@unimi.it

Summary

Multiple myeloma (MM) is a heterogeneous disease. International Staging System/fluorescence hybridization (ISS/FISH)-based model and gene expression profiles (GEP) are effective approaches to define clinical outcome, although yet to be improved. The discovery of a class of small non-coding RNAs (micro RNAs, miRNAs) has revealed a new level of biological complexity underlying the regulation of gene expression. In this work, 163 presenting samples from MM patients were analysed by global miRNA profiling, and distinct miRNA expression characteristics in molecular subgroups with prognostic relevance (4p16, MAF and 11q13 translocations) were identified. Furthermore we developed an “outcome classifier”, based on the expression of two miRNAs (MIR17 and MIR886-5p), which is able to stratify patients into three risk groups (median OS 19·4, 40·6 and 65·3 months, = 0·001). The miRNA-based classifier significantly improved the predictive power of the ISS/FISH approach (= 0·0004), and was independent of GEP-derived prognostic signatures (< 0·002). Through integrative genomics analysis, we outlined the potential biological relevance of the miRNAs included in the classifier and their putative roles in regulating a large number of genes involved in MM biology. This is the first report showing that miRNAs can be built into molecular diagnostic strategies for risk stratification in MM.

Multiple myeloma (MM) is a B-cell malignancy characterized by the clonal expansion of plasma cells within the bone marrow. At a molecular level it is highly heterogeneous, and many efforts are still on-going to derive reliable prognostic stratification based on these features. The Translocation/Cyclin D (TC) classification (Bergsagel et al, 2005) has been used to define biologically and prognostically important subgroups in MM including the unfavourable 4p16 and MAF (v-maf musculoaponeurotic fibrosarcoma oncogene homolog (avian)) groups as well as the favourable 11q13 group. Prognostic subgroups can be further defined using signatures derived from gene expression profiling (GEP) (Shaughnessy et al, 2007; Decaux et al, 2008; Dickens et al, 2010). However, these approaches do not capture all of the variability defining clinical behaviour; therefore there is potential for their utility to be improved. In this context the discovery of microRNAs (miRNAs) has revealed a new level of biological complexity that has the potential to better define prognostically important subgroups. miRNAs belong to a class of non-coding small RNAs (~ 22 nucleotides) that can bind to the 3′-untranslated region of mRNAs and negatively regulate the post-transcriptional expression of their targets, and more than 60% of human protein-coding genes are subject to miRNA regulation (Friedman et al, 2009). Many of the known miRNAs are clustered in the genome, suggesting that they might work in combination to achieve their biological function. miRNAs are involved in the regulation of normal haematopoiesis including B-cell differentiation (Zhang et al, 2009), and are also linked to the initiation and progression of lymphoproliferative diseases and myeloma (Calin & Croce, 2006a; Pichiorri et al, 2008; Zhou et al, 2010; Chi et al, 2011; Lionetti et al, 2012). Various mechanisms underlying the pathological deregulation of miRNAs have been reported including copy number changes, mutations, transcriptional activation, defective processing and promoter hypermethylation (Calin & Croce, 2006b). It is not surprising, therefore, that miRNA expression profiles can also be used to classify cancer, often with a greater degree of accuracy than traditional GEP (Lu et al, 2005; Calin & Croce, 2006a). In this work, we test the hypothesis that the patterns of miRNA expression in myeloma could correlate with survival providing important new molecular biomarkers of outcome.

Materials and methods

Patient samples

Bone marrow aspirates were obtained from newly diagnosed myeloma patients in the UK Medical Research Council (MRC) Myeloma IX study during standard diagnostic procedures following informed consent. The study was approved by the National Research Ethics Service (MREC 02/8/95) and registered under ISRCTN68454111. The design, patient evaluation and end points of this trial have been reported previously (Morgan et al, 2010). In summary the trial recruited 1960 patients, who were allocated to two main treatment pathways (intensive or non-intensive) at the discretion of the treating physician taking account of the age and performance status. The current median follow-up of this trial is 5·9 years. 153 patients with GEP have been stratified according to the TC classification described previously (Bergsagel et al, 2005).

miRNA profiling

Plasma cells were selected to a purity of >90% as determined by microscopy from bone marrow aspirate samples using CD138 magnetic bead sorting; subsequently small RNA was extracted and enriched using a modified protocol for Qiagen (Germantown, MD, USA) Allprep kit or Invitrogen (Paisley, UK) Trizol kit. miRNA expression profiling was then carried out in 185 cases according to the Affymetrix (High Wycombe, UK) recommended protocol. Briefly, the enriched small RNA was processed using the FlashTag labelling kit, which uses a tailing reaction followed by ligation of the biotinylated signal molecule to the target RNA sample. The labelled RNA was then hybridized to Affymetrix GeneChip® microRNA arrays v1.0 and scanned using a GeneChip® scanner 3000 7G. Expression values for 847 human miRNAs were extracted from CEL files using Affymetrix miRNA QC tool software (RMA normalized and log2-transformed). After quality control using R package affyPLM, 163 samples were included for analysis with NUSE (Normalized Unscaled Standard Error) values <1·05. The microarray data have been deposited in the Gene Expression Omnibus (GEO) under accession number GSE41276. The patients included in the miRNA expression analysis were representative of patients entered into the overall trial (Table SI).

Gene expression and copy number analysis

GEP of 261 samples was generated on Affymetrix HG-U133 Plus 2.0 arrays as previously described (Dickens et al, 2010; Wu et al, 2011), and the expression values were RMA normalized and log2-transformed. Affymetrix GeneChip Mapping 500K Array sets were performed as previously described (Dickens et al, 2010; Walker et al, 2010), and the copy number values were obtained using GTYPE and dChip and were inferred against the normal germ-line counterpart for each sample. A total of 153 of GEP samples and 72 genotyping samples had miRNA profiling data available for integrative analysis. The associated microarray datasets have been deposited into GEO under accession number GSE15695.

Quantitative RT-PCR

The expression of MIR886-5p, MIR17 and MIR18A was analysed in purified CD138+ cells by means of real-time quantitative polymerase chain reaction (Q-RT-PCR) using the TaqMan® microRNA assays (Applied Biosystems, Foster City, CA, USA) in accordance with the manufacturer's protocol. The measurement of transcript expression was performed using the Applied Biosystems StepONE Real-Time PCR System. All of the RNA samples were run in duplicate and normalized on the basis of the expression of MIR103 (Peltier & Latham, 2008). The threshold cycle (Ct) was defined as the fractional cycle number at which the fluorescence passed the fixed threshold. All signals with Ct ≥ 35 were manually set to undetermined. Data were expressed as 2−ΔCt (Applied User Bulletin No. 2) as previously described (Lionetti et al, 2009).

Statistical analysis

All analyses were performed in R 2.12.2 and Bioconductor 2·7. Multivariate Cox regression analyses were conducted to investigate the association of miRNA expression with progression-free (PFS) and overall survival (OS), where the expression level of each miRNA was used as continuous variable and treatment pathways (intensive or non-intensive) as covariate. miRNAs with a P-value < 0·05 were selected as being associated with survival irrespective of treatment pathway. For robustness only the miRNAs that remained significant after being corrected for multiple testing (Benjamin and Hochberg's method, < 0·05) were used to construct an outcome classifier to divide patients into different risk groups. The distribution of OS between risk groups of patients was estimated using the Kaplan-Meier method (log-rank test). Internal cross-validation was performed via bootstrapping on the final prognostic model (1000 replications). The independence of the risk groups defined by miRNA expression from other important risk predicting factors was tested using multivariate Cox regression. Performance of predictive models was compared by likelihood-ratio test (R package anova.coxph). The association of trend between the ISS/FISH risk groups and the miRNA-derived risk groups was investigated using linear by linear association test (R package coin).

Differentially expressed miRNAs between a particular TC subgroup of interest and the other subgroups were selected using significance analysis of microarray (SAM) (Bioconductor package samr), with a 1000-permutation adjustment and 5% false discovery rate (FDR).

Among the 163 samples there were also 72 cases with single nucelotide polymorphism-based mapping array data and 152 cases with fluorescence in situ hybridization (FISH) results. Integrative analyses were carried out to explore the mechanisms of miRNA deregulation. Either Wilcoxon or Kruska-Wallis test was used to look at the associations between miRNA expression levels and corresponding copy number values, as well as FISH lesions.

miRNA target prediction and correlation with gene expression

In order to be identified as putative targets of a particular miRNA, genes have to fulfil the following criteria: (i) The targets are predicted by at least 3 of the 11 programs in miRecords (Xiao et al, 2009), a resource for animal miRNA-target interactions which integrates the following target prediction tools: DIANA-microT, MicroInspector, miRanda, MirTarget2, miTarget, NBmiRTar, PicTar, PITA, RNA22, RNAhybrid and TargetScan, (ii) The targets are statistically associated with OS (< 0·05). The association between GEP and OS was tested following the same work flow as described above to produce two lists of genes associated with shorter OS (= 1569) and longer OS (= 1311) respectively, (iii) A significant inverse Pearson correlation needs to be identified between the expression of an miRNA and its targets (< 0·05). Correlation analyses between gene expression and miRNA expression were carried out among the 153 patients where GEP was available, only those interactions with negative correlation coefficients (r) were selected.

Results

miRNA expression profiles associated with the prognostic groups based on TC classification

In an initial analysis of the dataset, we assessed the miRNA expression patterns in TC classification groups with prognostic relevance (4p16, MAF and 11q13 groups). In the 153 cases with both GEP and miRNA profiling data, there were 26 cases of 4p16p, seven cases of MAF, 26 cases of 11q13, 42 cases of D1, 31 cases of D2, 12 cases of D1+D2, one case of D3 and eight cases of unknown classification; subsequently, the single D3 case and the unclassified cases were excluded from further analyses. We performed a one-to-one comparison between the test groups with the other major subtypes using SAM in 144 cases. The resulting lists were examined for intersections to find the miRNAs consistently being upregulated (or downregulated) in the subgroup of interest (Table SII). The 144 samples were grouped into 4p16, 11q13, MAF and others (comprising D1, D2 and D1+D2) based on TC classification, and the expression characteristics of the 4 subgroups were visualized using a heatmap (Fig 1). A distinct upregulation of the miRNA cluster MIR99b/MIRLET7E/MIR125A on 19q was identified in TC 4p16 cases, as well as MIR150/MIR155/MIR34a upregulation in MAF subgroup, largely confirming what has been seen previously by us in a smaller series (Lionetti et al, 2009). In addition, upregulation of MIR1275 and downregulation of MIR138 were observed in 11q13 cases.

Figure 1.

MicroRNA signatures for TC classification. Heatmap showing the distinct expression characteristics of the favourable 11q13 group, unfavourable 4p16 and MAF groups in contrast to the rest of cases (D1, D2, D1+D2) according to the 8 differentially expressed miRNAs. The TC subgroups and FISH abnormalities (Green, no chromosomal abnormality; red, chromosomal abnormality; grey, not known) are shown in colours above the heatmap. The colour scale bar in the heatmap represents the relative miRNA expression with red representing upregulation and blue representing downregulation.

miRNA expression associated with OS

After removal of those miRNAs with a percentage detection call (defined by Affymetrix QC Tool) of <2% across the samples, the expression values of 38 miRNAs were identified as being associated with OS as continuous variables according to Cox regression analyses (< 0·05, Table 1). Three clusters of miRNAs located at cytobands 13q31.1 (MIR17HG: MIR17, MIR18A, MIR20A, MIR19B1, MIR92A-1), Xq26.2 (MIR106A˜363: MIR106A, MIR18B, MIR20B, MIR19B2, MIR92A-2) and Xq26.3 (MIR503˜424: MIR503, MIR542-5p, MIR424-star) were identified as being associated with survival, comprising 13 of the 38 potentially deregulated miRNAs. Clusters MIR17HG and MIR106A˜363 are of particular interest, as the association between members of these clusters and OS remained significant or borderline significant (< 0·06) after being corrected by multiple testing (Benjamin and Hochberg's method). Some miRNAs located within these two clusters are homologous in sequence and, therefore, are classified as members of the same family (Fig 2A). The miRNAs within these clusters were co-expressed in our data, suggesting that they are subject to common regulatory mechanisms (Figure S1). An exception is MIR363 located within the MIR106A˜363 cluster, which might be due to the presence of an alternative/additional regulatory mechanism. The expression levels of MIR17 and MIR106A, despite belonging to separate clusters located on different chromosomes, are highly correlated (r = 0·94). The higher expression of MIR886-5p and MIR886-3p, originating from opposite arms of the same pre-miRNA, are both associated with shorter OS.

Table 1. miRNAs associated with OS (< 0·05) (= 38). 24 of them were upregulated in shorter survivors (HR > 1·0) and 14 were downregulated (HR < 1·0). Those labelled * remained significant after correction for multiple testing
miRNAsP valueP value after multiple testing correctionHazard Ratio (HR)Cytoband
* MIR886-5p0·00020·03851·745q31.1
*MIR18A 0·00020·03851·3913q31.1
* MIR17 0·00030·03851·5413q31.1
MIR501-3p0·00050·05062·21Xp11.23
MIR1260 0·00060·05060·3114q24.3
MIR18B 0·00070·05062·24Xq26.2
MIR106A 0·00080·05271·52Xq26.2
MIR17-star0·00310·17621·4513q31.1
MIR339-3p0·00480·20900·757p22.3
MIR503 0·00500·20901·57Xq26.3
MIR92A 0·00560·20901·2813q31.1/Xq26.2
MIR20A 0·00560·20901·3513q31.1
MIR20B 0·00610·21151·38Xq26.2
MIR129-3p0·01110·34772·497q32.1/11p11.2
MIR19B 0·01160·34771·3013q31.1/Xq26.2
MIR494 0·01240·34821·6414q32.31
MIR575 0·01370·36171·434q21.22
MIR615-3p0·01530·38330·4012q13.13
MIR31 0·01790·41561·789p21.3
MIR1308 0·01880·41561·20p22.11
MIR629 0·02010·41561·6215q23
MIR542-5p0·02030·41562·32Xq26.3
MIR424-star 0·02170·42391·60Xq26.3
MIR891B 0·02470·46270·43Xq27.3
MIR152 0·02580·46510·8517q21.32
MIR155 0·03000·51080·9121q21.3
MIR890 0·03060·51081·65Xq27.3
MIR1271 0·03530·53872·075q35
MIR650 0·03610·53871·2122q11.22
MIR886-3p0·03740·53871·285q31.1
MIR16-1-star0·03760·53870·4013q14.2
MIR541 0·03830·53870·4914q32.31
MIR1324 0·04180·56980·493p12.3
MIRLET7G-star0·04490·57570·543p21.1
MIR491-5p0·04550·57570·639p21.3
MIR216B 0·04680·57570·542p16.1
MIR92A-2-star 0·04810·57570·56Xq26.2
MIR570 0·04860·57570·673q29
Figure 2.

MIR17HG and MIR106˜363 clusters and their roles in myeloma pathogenesis. (A) Schematic of the genomic structure of the miRNA clusters located on Chr 13 and X. The colours indicate sequence homology between the individual miRNAs. miRNAs labelled with * are upregulated in short survivors (< 0·05). (B) A proposed model of miRNA/MYC/E2F interaction and downstream targets in MM (O'Donnell et al, 2005; Dews et al, 2006; Novotny et al, 2007; Pichiorri et al, 2008; Xiao et al, 2008; Zhou et al, 2010).

miRNA expression associated with PFS

Using the same work-flow as for OS, 35 miRNA were identified as being associated with PFS (Table SIII). Although after correction for multiple testing none of these miRNAs remained significant, it is worth mentioning that the members of the cluster at 19q13 (MIRLET7E, MIR125A-5p and MIR99B) being identified strongly associated with TC 4p16, were among the top six most differentially expressed miRNAs associated with PFS, adding further evidence to the global prognostic importance of the t(4;14) in myeloma.

Construction of a miRNA-based classifier for OS

Three upregulated miRNAs (MIR886-5p, MIR17 and MIR18A) were significantly associated with OS after correction for multiple testing (Benjamin and Hochberg's method, < 0·05). The expression of these miRNAs was validated by means of Q-RT-PCR in a fraction of samples (58 cases). A very good concordance with microarray data was found for all transcripts (Pearson correlation coefficients of the expression of each miRNA as determined by microarray or Q-RT-PCR 0·86, 0·72 and 0·75 respectively; data not shown). Unsupervised K-means clustering was applied to each miRNA across 163 samples to define a threshold splitting samples with higher expression from those with lower expression (Figure S2). After stepwise selection in a multivariate Cox regression model, using treatment pathway as a covariate, MIR886-5p and MIR17 were shown to have the strongest discriminative power for OS (Figure S3); consequently these two miRNAs were used to construct an outcome classifier. The proportion of samples defined as having higher expression of MIR886-5p and MIR17 were 24·5% and 56·4% respectively. Based on the expression levels of these two miRNAs, 163 patients were divided into three groups: a high risk group (both expression levels high) comprising 13·5% of the patients, a median risk group (either high) comprising 54% cases and a low risk group (both low) comprising 32·5% cases (Fig 3A). These three groups had significantly differential OS (log-rank test = 0·001, median OS 19·4 months vs. 40·6 months vs. 65·3 months). The median and high risk groups hada hazard ratio (HR) of 1·79 (95% Confidence Intervals (CI):1·15–2·78) and 2·89 (95% CI 1·60–5·20), respectively, relative to the low risk group. The stability of the miRNA-based classifier was assessed using bootstrap resampling. Based on 1000 replicates the mean significance was 0·004 with a standard error of 0·02, and the majority (98·4%) of the P values were <0·05. It is not surprising that the OS classifier based on the expression of two miRNAs was not associated with PFS in this dataset, because neither MIR17 nor MIR886-5p expression was associated with PFS, although the pathogenic role of MIR17HG in myeloma has been well demonstrated (Pichiorri et al, 2008; Zhou et al, 2010). To exclude the possibility that the association of miRNA classifier with OS is due to non-myeloma-related mortalities, 22 cases that died from reasons other than progressive myeloma (mostly other cancers, heart disease, stroke and infection) were censored at the time of death. The results showed that the three risk groups still had significantly differential effect on myeloma-specific survival (log-rank test = 0·002, median survival 28·2 months vs. 51·5 months vs. not reached, Figure S4). Further analysis on post-relapse survival for 141 relapsed cases showed a remarkable differential effect among the three risk groups (log-rank test = 2·4 × 10−7, median survival 6·1 months vs. 18·1months vs. 35·1 months, Fig 3B), which largely accounted for its impact on OS while lacking the significance on PFS.

Figure 3.

Survival of patients was stratified according to the miRNA-based classifier. (A) Patients (= 163) were divided into three groups: high risk (both miRNA expression high), median risk (either high) and low risk (both low) based on expression levels of MIR17 and MIR886-5p. (B) Further analysis on post-relapse survival for 141 relapsed cases showed a remarkable differential effect among the three risk groups. (C) The miRNA-based classifier is also able to identify subgroups within 22 t(4;14) cases, which had differential OS (median 13·8, 25·3 and 71·0 months respectively, = 0·005). (D) In 45 patients classified as being at low risk by ISS plus FISH abnormalities those with high expression of at least one of the two miRNAs (MIR17 and MIR886-5p) have shorter OS compared to rest of the cases (median 47·6 months vs not reached, = 0·01).

The miRNA-based classifier improves the ISS/FISH based risk stratification and is independent of GEP signatures

International Staging System (ISS) and FISH abnormalities including adverse IGH translocations [t(4;14), t(14;16) or t(14;20)], gain(1q) and del(17p) have been previously identified as independent prognostic factors (Boyd et al, 2012), as was the treatment pathway (log-rank test < 0·001) in the MRC Myeloma IX dataset examined here. Therefore, multivariate Cox regression analysis was carried out to test the independence of the miRNA-based risk groups from these important predictive factors. The results of this analysis confirmed the independent prognostic value of the miRNA-based classifier (= 0·0004, Table 2). This model showed a significant improvement of predictive capability compared to that without the miRNA-based classifier (likelihood-ratio test = 0·0004). Furthermore the miRNA-based classifier was able to identify subgroups within t(4;14) cases with different outcome (median OS 13·8, 25·3 and 71·0 months respectively, = 0·005) (Fig 3C), while no other FISH lesions could identify prognostically significant subgroups within these t(4;14) cases (> 0·05, data not shown).

Table 2. Multivariate Cox regression analysis showing the independence of the miRNA-defined risk groups from other important predictive factors in 97 patients with all variables available
 HRHR-95%CIP-value 
  1. ***< 0·001; *< 0·05.

MIR-groups2·111·40–3·200·0004 ***
ISS1·370·97–1·950·08
Adverse_translocations2·251·18–4·280·01 *
Gain(1q)0·980·57–1·680·94 
Del(17p)3·141·16–8·520·02 *
Path (nonintensive)2·671·53–4·640·0005 ***

Recently, we have developed a prognostic model based on the co-segregation of adverse prognostic FISH lesions and the ISS (Boyd et al, 2012). We investigated whether the miRNA classifier could be usefully incorporated into this type of risk stratification approach. We found a positive association between the ISS/FISH risk groups and the risk groups defined by the miRNA-based classifier (linear by linear association test = 0·005) (Table 3). Importantly, despite this association, the ISS/FISH low-risk patients who showed high expression levels of at least one of these two miRNAs, had shorter OS (median 47·6 months) compared to the remaining patients, 70% of whom remained alive after 7 years (= 0·01) (Fig 3D).

Table 3. Matrix depicting a positive correlation between the risk groups based on FISH abnormalities & ISS and the risk groups defined by miRNA-based classifier (linear by linear association test = 0·005)
Risk groups based on FISH abnormalities and ISSRisk groups defined by miRNA-based classifier
Low (%)Median (%)High (%)Grand total (%)
Low20 (44·4)19 (42·2)6 (13·3)45 (100)
Intermediate10 (27·8)22 (61·1)4 (11·1)36 (100)
High012 (75)4 (25)16 (100)
Grand total30531497

The expression levels of some individual miRNAs belonging to the MIR17HG and MIR106A˜363 clusters have previously been reported as being associated with the University of Arkansas for Medical Sciences (UAMS) GEP-defined risk score (Zhou et al, 2010). By using the published method (Shaughnessy et al, 2007) we applied this 70-gene signature to our series of 153 samples with matching GEP data to stratify them into high- and low-risk groups based on the gene-risk model. Then, we constructed a multivariate model including both 70-gene-defined risk groups and miRNA-defined risk groups, showing that miRNA-based classifier retained independent prognostic significance from UAMS' gene-risk model (= 0·002). In a similar fashion, the prognostic value of the 2-miRNA model was also confirmed to be independent from both the Intergroupe Francophone du Myelome (IFM)-signature (Decaux et al, 2008) and the Myeloma IX 6-gene signature (Dickens et al, 2010) with even more significant effect (< 0·001).

Putative targets of OS-associated miRNAs

Given that miRNAs have been shown to exert the functional effects via cleavage of the mRNAs of their target genes (Bagga et al, 2005; Jing et al, 2005), we looked at the putative targets of the OS-associated miRNAs to gain insights into potential mechanistic associations. In this context it is known that the members of MIR17HG and MIR106A˜363 clusters share sequence homology and therefore could potentially target the same genes. Validated targets of these miRNAs include CDKN1A, SOCS1 and BCL2L11 in myeloma cell lines (Pichiorri et al, 2008; Zhou et al, 2010), together with the pro-apoptotic genes PTEN, E2F1 and the anti-angiogenic genes CTGF, THBS1 in other cell types (O'Donnell et al, 2005; Dews et al, 2006; Novotny et al, 2007; Xiao et al, 2008). We assessed the correlation of expression of these miRNAs with their potential target genes (PTEN, E2F1, CTGF and THBS1) in 153 patient samples for whom both miRNA and gene expression profiling data were available. The results showed that there were significant inverse correlations between E2F1 expression and at least one of the cluster members (< 0·05). Trends have also been observed for the expression levels of CTGF, THBS1 and PTEN being inversely correlated with members of these two clusters. This observation suggests that the expression levels of these genes may be pathologically relevant to the adverse prognosis associated with the expression of these miRNAs.

Using the selection criteria described in the methods section, the putative targets of MIR886-5p, the other miRNA forming the classifier, were identified (Table SIV). Figure S5 is an example showing that lower expression of the top candidate target NR3C1 is associated with shorter OS.

The correlation of OS-associated miRNAs with cytogenetic abnormalities, copy numbers and transcriptional regulation

We examined the association of expression level of MIR17 and MIR886-5p with FISH abnormalities including del(13q), del(17p), t(4;14), del(1p) and gain(1q). The results of this analysis showed that MIR17 expression was significantly associated with del(1p) and gain(1q) (< 0·05, Figure S6). However, although MIR17 is located on 13q, which is frequently deleted in myeloma patients, we did not find any correlation between 13q deletion and the expression level of this miRNA. The expression of MIR886-5p was associated with t(4;14) and del(13q) (< 0·05, Figure S6). This is not surprising as these abnormalities are very tightly linked. In order to explore further how these two miRNAs are deregulated in myeloma, we investigated the correlation between their expression and the tumour acquired DNA-based copy number at their chromosomal locations. Our data indicated that neither MIR17 nor MIR886-5p expression levels were copy number sensitive (Figure S7), suggesting that other mechanisms could be responsible. As the MIR17HG and MIR106A˜363 clusters have previously been shown to be activated by MYC and E2F3 (O'Donnell et al, 2005; Woods et al, 2007). We evaluated the correlations between the gene expression and the expression of the miRNAs within these two clusters, and significant positive correlations were identified for both genes (< 0·05).

Discussion

In the present study, we have comprehensively analysed the miRNA expression profiling in a prospective cohort including 163 cases from the MRC Myeloma IX l trial, and correlated the miRNAs expression pattern with outcome in order to outline their possible role in MM prognostication. In myeloma, the TC classification has been used to define subgroups with distinct prognoses with the 11q13 translocated group being linked to favourable outcome and 4p16 and MAF translocated groups being linked to unfavourable outcome. In our study 8 miRNAs were identified as being deregulated distinctly in these three subgroups, suggesting that these miRNAs could play an important role in the pathogenesis of these distinct molecular subgroups.

Among the differentially expressed miRNAs, MIR125A, MIRLET7E, MIR150, MIR34A (positively associated with either TC 4p16 or MAF) and MIR138 (negatively associated with TC 11q13) have also been shown to be upregulated in myeloma cells in comparison to their normal counterparts (Pichiorri et al, 2008; Lionetti et al, 2009; Zhou et al, 2010; Chi et al, 2011). The cluster of miRNAs that was strongly correlated with TC 4p16, including MIR125A, MIRLET7E and MIR99B, was shown to be associated with shorter PFS in our dataset. Interestingly MIR125A has been demonstrated to have a role in hematopoietic stem cells, increasing their number both in vivo and in vitro (Guo et al, 2010; Gerrits et al, 2012), suggesting possible relevance to myeloma stem cell biology. In addition, transgenic mice ectopically expressing the MAF-associated miRNAs (MIR150 or MIR155) either show dramatic impairment of B cell differentiation (Xiao et al, 2007) or develop high-grade B cell lymphoma (Costinean et al, 2006). CCND1 has been previously identified as direct target of MIR138 (Liu et al, 2012), which was significantly downregulated in TC 11q13.

We found three clusters of miRNAs associated with adverse OS outcome in myeloma patients: MIR503˜424 (on Xq26.3), MIR17HG (on 13q31.1) and MIR106A˜363 (on Xq26.2). The expression level of MIR503˜424 has previously been found to be upregulated in malignant tissues and has been associated with impaired survival in a number of cancers (Zhao et al, 2009; Corbetta et al, 2010; Ozata et al, 2011), supporting its potential relevance. Previous studies have shown that members of the clusters MIR17HG and MIR106A˜363 are downregulated during the normal germinal centre B-cell to plasma cell transition (Jima et al, 2010), suggesting that upregulation of these variants in plasma cells may adversely affect their biological behaviour. Indeed, previous work on miRNA expression in myeloma has shown that the upregulation of these miRNAs is associated with either the transformation from monoclonal gammopathy of undetermined significance to myeloma or with an mRNA-based risk score (Pichiorri et al, 2008; Zhou et al, 2010). In this study, for the first time, we have shown that a high level of expression of these two clusters is associated with an adverse clinical outcome in a series of well-characterized clinical samples.

Currently, the ISS plus the FISH–based abnormalities are used to define prognosis; however, this approach does not capture all the clinical variability and there is potential for it to be improved. By combining the expression level of MIR17 and MIR886-5p, we classified patients into three subgroups associated with significant differences in OS, which were retained in multivariate analyses when taking into account the ISS and FISH–based model. Notably, when the miRNA-based classifier was removed from this model, the predictive power was significantly reduced (= 0·0004). The robustness of the miRNA-based classifier has been validated using 1000 bootstrap replications with an estimated error rate of 1·6%. Importantly the definition of the aggressiveness of the clinical behaviour of newly presenting cases can be improved by the incorporation of this miRNA-based classifier into currently used strategies. In this context, we have shown that, within the group classified as being at low risk using ISS/FISH approach, the expression of these miRNAs can define a further group (comprising half of the cases) with a significantly worse clinical outcome, which really should belong to intermediate risk group. The miRNA-based risk classifier is also able to identify prognostically important subgroups within t(4;14) cases; notably the t(4;14) cases with low expression level of both these two miRNAs show prolonged median OS of 71 months. These findings are supported by a large body of literature showing that miRNA-based classifiers predict survival in various types of cancers, which are independent from currently known clinicopathological features (Yu et al, 2008; Caramuta et al, 2010; Montes-Moreno et al, 2011; Srinivasan et al, 2011; Jamieson et al, 2012; Lee et al, 2012).

Although there is no independent validation dataset available, one of the two miRNAs comprising the classifier, MIR17, was associated with GEP risk score (Zhou et al, 2010); the classifier, therefore, could be considered as being partially validated. The other miRNA, MIR886-5p, is not present on the array reported by Zhou et al (2010). The classifier identified in this study is not significantly associated with PFS, and we have confirmed that its differential effect on OS is largely due to the impact on post-relapse survival. OS and PFS are known to be different endpoints; therefore, the strongest predictors for each of them are not necessarily the same. Furthermore OS remains the gold standard for demonstrating clinical benefit in myeloma patients. Notably the oncogenic role of the clusters MIR17HG and MIR106A˜363 in myeloma has been well demonstrated (Pichiorri et al, 2008; Zhou et al, 2010), although none of the members is associated with PFS. To this regard, it could be conceivable that the deregulation of specific genes/miRNAs might have long-term effects on myeloma cells and/or their interactions with other environmental components, which might not be reflected by PFS. Indeed, miRNAs have recently been recognized as key regulators in the neoplastic microenvironment (Wentz-Hunter & Potashkin, 2011).

An important question is whether these miRNA clusters are biologically relevant and actually mediate the biological changes associated with the poor prognosis in MM. Recently, two studies provided functional evidence that miRNAs within MIR17HG and MIR106A˜363 clusters target critical genes including BCL2L11, SOCS1 and CDKN1A, which are known to be involved in both myeloma cell proliferation and apoptosis (Pichiorri et al, 2008; Zhou et al, 2010). In one of these studies the oncogenic role of MIR19A/B was also confirmed in a nude mice model in which regression of transplanted tumours after treatment with an antagonist was achieved (Pichiorri et al, 2008). BCL2L11 and CDKN1A are also two main downstream effectors of transforming growth factor-β (TGFβ, TGFB) signalling, the inactivation of which is a major step in the development of a variety of human tumours (Bierie & Moses, 2006). The proapoptotic genes PTEN, E2F1, and anti-angiogenic genes CTGF, THBS1 have also been previously shown to be targets of these two clusters of miRNAs (O'Donnell et al, 2005; Dews et al, 2006; Novotny et al, 2007; Xiao et al, 2008). In our study we observed the inverse correlations of the expression level between these genes and their regulating miRNAs, suggesting that an important interaction could also exist in myeloma.

The association of MIR17HG expression and copy number of chromosome 13q in MM is currently a controversial issue. In keeping with another report (Chen et al, 2011), our analysis did not identify an association between expression of MIR17HG and del(13q); however, a few other studies observed an at least partial correlation (Lionetti et al, 2009; Gutierrez et al, 2010; Chi et al, 2011). This discordance may be due to the variability of the clinical samples. MYC deregulation is important in myeloma (Selvanayagam et al, 1988; Chng et al, 2011), and recent evidence suggested that MYC not only regulates expression of protein-coding genes directly, but also controls the expression of a large set of miRNAs (O'Donnell et al, 2005; Chang et al, 2008). In particular, MYC upregulation has been shown to directly activate the MIR17HG and MIR106A˜363 clusters (O'Donnell et al, 2005), suggesting that it may play an important role in miRNA deregulation in MM. Indeed in this study we identified a significant positive correlation between MYC expression and the expression levels of individual miRNAs within these two clusters. Despite being on different chromosomes, the high correlation between MIR17 and MIR106A may support them being co-regulated. However, as the correlations between the expression of MYC and members of the two clusters are modest (R value up to 0·3), other genes may also be important in their deregulation. MIR17HG cluster has previously been shown to be activated by the E2F family member E2F3 (Woods et al, 2007), and E2F and MYC are known to transactivate each other, suggesting a possible complex regulatory signal for MIR17HG expression (Fig 2B). These observations also highlight the potential importance of MYC/E2F/MIR17HG negative feedback loop in cancer.

The deregulated genes downstream of the deregulated miRNAs could mediate the prognostic effect of these miRNAs; therefore, we developed an approach to identify the putative targets of MIR886-5p, the other miRNA comprising the classifier. The pathogenic role of MIR886-5p has not been previously reported in myeloma; however, it was shown to be upregulated in Burkitt Lymphoma compared to some other lymphoma types (Zhang et al, 2009; Iqbal et al, 2012). MIR886-5p inhibits apoptosis of cervical cancer cells by down-regulating the production of Bax (Li et al, 2011) It also has been shown to confer chemo-resistance in colorectal cancer cell lines (Jensen et al, 2012). Increased MIR886-3p, which originates from the same pre-miRNA, has recently been associated with chemo-resistance in bladder cancer, which was translated to impaired overall survival (Nordentoft et al, 2012). As the top target for MIR886-5p, NR3C1 is the glucocorticoid receptor gene and its downregulation has been associated with glucocorticoid resistance and inferior prognosis in MM (Sanchez-Vega & Gandhi, 2009; van Rhee et al, 2010); the association of the expression of MIR886-5p with prognosis was confirmed in our dataset. One of the other potential targets ICOSLG (inducible co-stimulator ligand) is expressed on tumour cells and has been reported to have an important role in tumour immunity; it also induces B-cell differentiation into plasma cells. It has been demonstrated that cytotoxic T cells play a critical role in myeloma cell elimination (Deng et al, 2004; Michalek et al, 2010); therefore, it is not surprising that ICOSLG expression has an effect on OS of myeloma patients. The precursor of MIR886-5p and MIR886-3p, previously proposed to be a vault RNA, a component of the vault complex implicated in cancer drug resistance, was recently shown neither to be a genuine pre-miRNA nor a vault RNA (Lee et al, 2011). MIR886 binds directly to PKR (Protein Kinase RNA-activated) and silencing of miR-886 activates PKR and its downstream pathways, eIF2α phosphorylation and the nuclear factor-κB pathway, leading to impaired cell proliferation (Lee et al, 2011). The association of MIR886-5p expression level with OS warrants validation and additional studies to investigate its potential roles in MM pathogenesis.

Multiple myeloma is a genetically complex disease with a well-described heterogeneity in clinical outcome. Recent research highlights the contribution of a new class of non-coding genes, miRNA, in myeloma pathogenesis. In this work, we have developed a 2-miRNA-based classifier able to stratify MM patients into three risk groups. The classifier significantly improves the predictive power of an outcome predictor comprising ISS and FISH-based abnormalities; therefore, it may represent a complementary prognostic tool in clinical practice after being validated using independent dataset. The miRNAs related to the classifier are biologically relevant, and integrative analyses indicate that they are putative candidates regulating a large number of genes involved in MM biology such as proliferation, apoptosis, angiogenesis and drug resistance. In this context, miRNAs can be built into molecular diagnostic strategies for risk stratification as well as being used as treatment targets in MM.

Acknowledgements

Myeloma UK; Cancer Research UK; the Bud Flanagan Leukaemia Fund; The Biological Research Centre of the National Institute for Health Research at the Royal Marsden Hospital; Associazione Italiana Ricerca sul Cancro (AIRC) grants IG10136; AIRC “Special Program Molecular Clinical Oncology- 5 per mille” n. 9980, 2010/15; Ministero Italiano dell' Istruzione, Università̀ e Ricerca (MIUR) grant 2009PKMYA2. L M is supported by a fellowship of the Fondazione Italiana Ricerca sul Cancro (FIRC).

Author contributions

P.W. designed and performed research, analysed data and wrote the paper. L.A. performed research, analysed data and wrote the paper. B.A.W, K.T. and M.L. performed research and analysed data. D.C.J., M.K., F.M. and C.W. performed research. W.M.G. and D.B. analysed data. F.E.D. designed research. A.N. and G.J.M designed research and wrote the paper.

Conflict-of-interest disclosure

The authors declare no competing financial interests.

Ancillary