Skip to main content

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

Myeloma

A gene expression signature for high-risk multiple myeloma

A Corrigendum to this article was published on 07 May 2014

Abstract

There is a strong need to better predict the survival of patients with newly diagnosed multiple myeloma (MM). As gene expression profiles (GEPs) reflect the biology of MM in individual patients, we built a prognostic signature based on GEPs. GEPs obtained from newly diagnosed MM patients included in the HOVON65/GMMG-HD4 trial (n=290) were used as training data. Using this set, a prognostic signature of 92 genes (EMC-92-gene signature) was generated by supervised principal component analysis combined with simulated annealing. Performance of the EMC-92-gene signature was confirmed in independent validation sets of newly diagnosed (total therapy (TT)2, n=351; TT3, n=142; MRC-IX, n=247) and relapsed patients (APEX, n=264). In all the sets, patients defined as high-risk by the EMC-92-gene signature show a clearly reduced overall survival (OS) with a hazard ratio (HR) of 3.40 (95% confidence interval (CI): 2.19–5.29) for the TT2 study, 5.23 (95% CI: 2.46–11.13) for the TT3 study, 2.38 (95% CI: 1.65–3.43) for the MRC-IX study and 3.01 (95% CI: 2.06–4.39) for the APEX study (P<0.0001 in all studies). In multivariate analyses this signature was proven to be independent of the currently used prognostic factors. The EMC-92-gene signature is better or comparable to previously published signatures. This signature contributes to risk assessment in clinical trials and could provide a tool for treatment choices in high-risk MM patients.

Introduction

Multiple myeloma (MM) is characterized by accumulation of malignant monoclonal plasma cells in the bone marrow. The median overall survival (OS) for newly diagnosed patients treated with high-dose therapy varies from 4 to 10 years.1, 2

The International Staging System (ISS), based on serum β2-microglobulin and albumin, is widely used as a prognostic system for patients with newly diagnosed MM. ISS has been confirmed as a solid prognostic factor in clinical trials.1 Additional clinical factors to define high-risk disease have not been consistently reproduced, with the exception of the extensive disease represented by renal failure and plasma cell leukemia.2, 3 In addition to ISS, cytogenetic aberrations such as deletion of 17p (del(17p)), and translocations t(4;14) and t(14;16) were shown to be associated with an adverse prognosis. The combination of prognostic markers t(4;14), del(17p) and ISS enabled further delineation of patients into prognostic subgroups.4

A strategy to include the genetic characteristics of MM is the translocation and cyclin D classification, which distinguishes eight subgroups based on genes that are deregulated by primary immunoglobulin H translocations and transcriptional activation of cyclin D genes.5 Subsequently, the University of Arkansas for Medical Sciences (UAMS) generated a molecular classification of myeloma based on GEPs of patients included in their local trials. The UAMS molecular classification of myeloma identifies seven distinct gene expression clusters, including the translocation clusters MS, MF and CD-1/2, a hyperdiploid cluster, a cluster with proliferation-associated genes (PR) and a cluster characterized by a low percentage of bone disease (LB).6 More recently, we extended this classification based on the HOVON-65/GMMG-HD4 prospective clinical trial and identified additional molecular clusters, that is, NFκB, CTA and PRL3.7 Because these clusters were discriminated based on disease-specific GEPs, we and others hypothesized that they may be relevant for therapy outcome. Indeed, the UAMS-defined clusters MF, MS and PR were found to identify high-risk disease in the total therapy (TT)2 trial.6

Several survival signatures were developed based on samples from clinical trials, such as the UAMS-70, the related UAMS-17 and the recently published UAMS-80 signature, which have value in prognostication of MM.8, 9, 10 Other signatures include the Medical Research Council (MRC) gene signature based on the MRC-IX trial, the French Intergroupe Francophone du Myélome (IFM) signature and the Millennium signature based on relapse patients.11, 12, 13 Recently, a GEP-based proliferation index was reported.14 So far, none of these signatures have been introduced in general clinical practice.

The additional and independent prognostic significance of a prognosticator based on gene expression has been acknowledged in mSMART (Mayo Stratification for Myeloma And Risk-adapted Therapy). Hereby, a high-risk MM population can be defined, for which alternative treatment is proposed, although this has not been validated in prospective clinical trials.15

The aim of the present study was to develop a prognostic signature for OS in MM patients. This investigation was prospectively included as a secondary analysis of a randomized clinical trial for newly diagnosed, transplant-eligible patients with MM (HOVON-65/GMMG-HD4).

Patients and methods

Patients

As a training set the HOVON-65/GMMG-HD4 study (ISRCTN64455289) was used. Details of the training set are given in the Supplemental Document A.16 Informed consent to treatment protocols and sample procurement was obtained for all cases included in this study, in accordance with the Declaration of Helsinki. Use of diagnostic tumor material was approved by the institutional review board of the Erasmus Medical Center. Arrays that were used for analysis passed extensive quality controls, as described previously.7 Of the 328 gene arrays deposited at the NCBI-GEO repository, clinical outcome data were available for 290 patients (accession number: GSE19784).

Four independent data sets were used as validation, of which both survival data were available as well as GEPs of purified plasma cells obtained from bone marrow aspirates of myeloma patients. The data sets total therapy 2 (UAMS-TT2; n=351; GSE2658; NCT00573391), total therapy 3 (UAMS-TT3; n=142; E-TABM-1138; NCT00081939) and MRC-IX (n=247; GSE15695; ISRCTN68454111) were obtained from newly diagnosed patients. The APEX data set (n=264; GSE9782; registered under M34100-024, M34100-025 and NCT00049478/NCT00048230) consisted of relapsed myeloma cases (Supplemental Document A).11, 17, 18, 19, 20, 21, 22, 23

Gene expression pre-processing

To allow gene expression analysis in the HOVON-65/GMMG-HD4, plasma cells were purified from bone marrow aspirates obtained at diagnosis, using immunomagnetic beads. Only samples with a plasma cell purity of 80% were used. Gene expression was determined on an Affymetrix GeneChip Human Genome U133 Plus 2.0 Array (Affymetrix, Santa Clara, CA, USA).

To allow for validation across different studies, only probe sets present on both the U133 Plus 2.0 and U133 A/B platforms were included (n=44 754). Probe sets having an expression value below the lowest 1% bioB hybridization control in more than 95% of the samples are excluded. This resulted in 27 680 probe sets being analyzed. All data were MAS5 normalized, log2 transformed and mean-variance scaled, using default settings in the Affy package in Bioconductor.24

The normalized validation gene expression data sets were downloaded from the repositories NCBI-GEO (APEX, MRC-IX, UAMS-TT2) and ArrayExpress (UAMS-TT3). Data sets UAMS-TT2, UAMS-TT3 and MRC-IX were generated using the U133 Plus 2.0 platform (Affymetrix), whereas the Affymetrix HG U133 A/B platform was used in the APEX study. The IFM data set was not included in our analysis owing to an incompatible, custom platform.

The strong batch effect that exists between these GEP studies was successfully removed by ComBat using the non-parametric correction option.25 APEX was run on a different array platform with an incomplete overlap in probe sets with the other data sets, and as a result ComBat correction was applied in two separate runs with one run, for all analyses involving the APEX data set and an additional run for all other analyses.

Survival signature

The HOVON-65/GMMG-HD4 data were used as a training set. GEP and progression-free survival (PFS) data were combined for building a GEP-based survival classifier. PFS was used for generating a classifier for OS as PFS was the primary endpoint of the HOVON-65/GMMG-HD4 study and PFS demonstrated a higher number of events compared with OS (179 PFS vs 99 OS events in total in the HOVON-65/GMMG-HD4 data). All evaluations of the signature are based on OS data in training and validation sets. Analyses were performed using R with the survival package for survival analyses.26 Out of 27 680 probe sets tested, 1093 probe sets were associated to PFS in univariate Cox regression analyses (FDR <10%; for probe sets and survival data, see Supplemental Document B). Subsequently, this set was used as input into a supervised principal component analysis framework in combination with simulated annealing (Supplemental Documents A and B).27 This analysis yielded a model of 92 probe sets, termed the EMC-92 signature. The survival signature is a continuous score, that is, the sum of standardized expression values multiplied by the probe set-specific weighting coefficient (Supplemental Table S1 and R-script, Supplemental Document C). High-risk disease was defined as the proportion of patients with an OS of less than 2 years in the training set.

Validation of the EMC-92 signature

A multivariate Cox regression analysis was performed for patients with available covariates. Covariates with <10% of the data missing were used as input in a backward stepwise selection procedure (P<0.05).

The EMC-92 signature together with seven previously described, external signatures for OS in MM has been analyzed in a pair-wise comparison using a multivariate Cox regression analysis. This analysis was performed for all pair-wise comparisons on the pooled datasets excluding the training sets for the signatures being tested. The models were stratified for study.

Pathway analysis

Pathway analysis was performed using the 92 genes corresponding to the EMC-92 signature, as well as the 1093 genes generated by univariate-PFS analysis (FDR <10%), with the probe sets used as input for the analysis as a reference set (n=27 680; Ingenuity Systems, www.Ingenuity.com). P-values were derived from right-tailed Fisher exact tests and corrected for multiple testing by a Benjamini−Hochberg correction.28

Results

The EMC-92 signature

GEPs obtained from the newly diagnosed MM patients were analyzed in relation to survival data, in order to generate a classifier to distinguish high-risk from standard-risk disease. We used the HOVON-65/GMMG-HD4 data as a training set.7 After filtering for probe set intensity, using internal Affymetrix control probe sets, 27 680 probe sets were analyzed in a univariate Cox regression analysis with PFS as survival end point. This resulted in 1093 probe sets being associated with PFS with a false discovery rate of <10% (Supplemental Document B). Based on these 1093 probe sets, a supervised principal-components-analysis-based model was built, in which simulated annealing was applied to generate the optimal model settings in a 20-fold cross-validation. The final predictive model consisted of 92 probe sets with specific weighting coefficients. The sum of normalized intensity values multiplied by this weighting is the output of the signature. This model was termed the EMC-92 signature. A positive weighting coefficient indicates that increased expression contributes to a higher value for the EMC-92 signature value and thus a higher risk for poor survival. The majority of the probe sets are annotated genes (n=85, with one of the genes represented by two probe sets). The remaining probe sets are open reading frames (n=3), expressed sequence tags (n=2) and one additional probe set without annotation. Several known cancer genes are among these genes, of which FGFR3 (weighting coefficient: 0.06), STAT1 (weighting coefficient: 0.05) and BIRC5 (weighting coefficient: 0.02) were described in detail in relation to myeloma (Supplemental Table S1; all Supplemental Tables are given in the Supplemental Document A).29, 30, 31

To define a high-risk population, the cut-off threshold for the continuous signature score was set to a value of 0.827 based on the proportion of patients in the training set who had an OS of less than 2 years (63 out of 290 patients (21.7%); Supplemental Figure S2).

Four independent validation data sets were available: UAMS-TT2, UAMS-TT3, MRC-IX and APEX. Gene expression data sets UAMS-TT2 and UAMS-TT3 consisted of 351 and 142 transplant-eligible patients, whereas the MRC-IX data set contained both transplant-eligible and non-transplant-eligible MM patients (n=247). In the APEX data set, GEPs of 264 relapse patients were collected. The results of the EMC-92 signature in the validation sets are shown in Figure 1 and Supplemental Table S2. In the UAMS-TT2 data set, the EMC-92 signature identified a high-risk population of 19.4% with a hazard ratio (HR) of 3.40, 95% confidence interval (CI)=2.19–5.29 (P=5.7 × 10−8). In the UAMS-TT3 data set, 16.2% of patients were identified as high-risk with a HR of 5.23, 95% CI=2.46–11.13 (P=1.8 × 10−5). In the MRC-IX data set, 20.2% of patients were identified as high-risk with a HR of 2.38 and 95% CI=1.65–3.43 (P=3.6 × 10−6). The high-risk signature was able to identify patients with significantly shorter survival in both the transplant-eligible and non-transplant-eligible patients included in the MRC-IX study. In non-transplant-eligible patients, 23.9% high-risk patients were identified with a HR of 2.38, 95% CI=1.47–3.86, (P=4.3 × 10−4), whereas 16.8% of transplant-eligible patients were high-risk with a HR of 2.54, 95% CI=1.43–4.52 (P=1.5 × 10−3; Figures 1d and e). The signature was not restricted to newly diagnosed patients, as 16.3% of patients included in the APEX relapse dataset were designated high-risk with a HR of 3.01, 95% CI=2.06–4.39 (P=1.26 × 10−8; Figures 1f and 2e).

Figure 1
figure1

Kaplan-Meier OS curves for EMC-92 signature defined high-risk patients versus standard-risk patients in five validation sets. The cut-off value is fixed at 0.827 based on the proportion of patients with OS <2 years in the HOVON-65/GMMG-HD4 set. In the MRC-IX one patient had an unknown treatment status and was disregarded in (d and e). (a) UAMS TT2. (b) UAMS TT3. (c) MRC-IX. (d) MRC-IX transplant-eligible patients. (e) MRC-IX non-transplant-eligible. (f) APEX. Events, number of events; HR, hazard ratio; median, median survival time; N, number of patients; Wald P, P value for equality to standard-risk group.

Figure 2
figure2

Performance per signature in available data sets. For every signature the hazard ratio (high-risk versus standard-risk) is shown with 95% confidence interval. Gray lines indicate results on the training set. (a) HOVON-65/GMMG-HD4. (b) UAMS-TT2. (c) UAMS-TT3. (d) MRC-IX. (e) APEX. P, P value for equal survival in high and standard-risk groups; proportion, proportion of high-risk-defined patients.

To assess the relation between EMC-92 signature outcome and treatment, we evaluated whether there is evidence for differences in survival between treatment arms in the high-risk group and the standard-risk group. Within the high-risk patients of the HOVON-65/GMMG-HD4 trial, the survival of bortezomib treated patients was longer than patients treated with conventional chemotherapy (VAD) (30 months compared with 19 months), albeit not significant (P=0.06; number of bortezomib-treated patients : 26 vs 37 in the VAD arm). Within the high-risk patients of MRC-IX, no difference was observed between the treatment arms (P=0.5: MRC-IX non-transplant eligible: CTDA n=14 vs MP n=12) and P=1.0 (MRC-IX transplant eligible; CTD n=16 vs CVAD n=7). For the standard-risk patients no differences in survival between treatment arms were found in either trial.

Multivariate analysis was performed in the training set and in the APEX and MRC-IX validation sets, for which information on a large number of variables was available. This showed that in addition to the EMC-92 signature, del(17p) was an independent predictor in HOVON-65/GMMG-HD4. Furthermore, in both HOVON-65/GMMG-HD4 and the APEX multivariate analysis, a component of the ISS was an additional independent prognostic predictor (beta-2-microglobulin for the HOVON65/GMMG-HD4 data set and serum albumin for the APEX data set). Trial-specific covariates were seen in each multivariate analysis, such as a substudy in the APEX data set and the MP treatment arm in the MRC-IX set. In conclusion, in all the three data sets of newly diagnosed and relapse MM patients, the EMC-92 signature proved to be the strongest predictor for survival after inclusion of available covariates (Table 1). For univariate associations to survival, see Supplemental Table S3.1Supplemental Table S3.3.

Table 1 Multivariate analysis for the EMC-92-gene in the HOVON-65/GMMG-HD4 (1.1), APEX (1.2) and MRC-IX (1.3)

Using the nearest neighbor classification method, all patients in the validation sets were classified into molecular clusters based on the HOVON-65/GMMG-HD4 classification.7 A clear enrichment of the MF, MS, PR clusters and decreased proportion of the hyperdiploid cluster was found in the pooled high-risk populations of all validation sets (Supplemental Table S4).

To define the biological relevance of the EMC-92 signature and 1093 probe sets found by initial univariate ranking, pathway analysis of the EMC-92 and the 1093 probe sets was performed. Significant functions for the EMC-92 signature included multiple ‘cell cycle’ pathways (P=1.82 × 10−3–4.93 × 10−2; Supplemental Table S5), including genes such as BIRC5, TOP2A and CENPE. The 1093 probe sets indicated functions such as ‘protein synthesis’ (P=9.45 × 10−31–1.54 × 10−12), ‘cancer’ (P=4.82 × 10−12–4.91 × 10−2) and ‘cell cycle’ (P=3.74 × 10−9–4.91 × 10−2; Supplemental Table S6). Next, we compared the chromosomal locations of the probe sets within the EMC-92 signature to the expected proportion represented on the Affymetrix chip (Supplemental Table S7). None of the chromosomes demonstrated a significant enrichment in the EMC-92 signature, while all somatic chromosomes are represented. Within the set of 1093 probe sets, which formed the basis of the EMC-92 signature and were identified by univariate survival analyses, chromosomes 1 and 4 were found to be significantly overrepresented. Further analysis of chromosome 1 demonstrated a clear enrichment of the long arm of chromosome 1 in this set of genes (Supplemental Table S8).

Comparison with published gene signatures

We set out to evaluate the performance of the EMC-92 signature in relation to available GEP-based prognostic signatures for OS in MM. To this end, the following signatures were evaluated: UAMS-70, UAMS-17, UAMS-80, IFM-15, gene proliferation index (GPI-50), MRC-IX-6 and MILLENNIUM-100.9, 10, 11, 12, 13, 14

These signatures were evaluated as continuous variables as well as using the cut-off values as published (Figures 2a–e, Supplemental Figure S2 and Supplemental Documents A and B). Overall, the performance of the EMC-92 signature is robust, consistent and compares favorably with previously published signatures. Specifically, the EMC-92, UAMS, MRC-IX and GPI-50 signatures demonstrated significance in all validation sets tested both for the dichotomized and for the continuous values of the signatures. Significance was reached in three out of five studies for the IFM-15 signature using a dichotomized model, whereas the MILLENNIUM-100 signature had significant performance in the dichotomized model in one out of four independent studies. Thus, performance was less robust for the IFM-15 and MILLENNIUM-100 signatures. Although the proliferation index GPI-50 was found to be significant in all validation sets tested, the proportion of high-risk patients was much lower compared with the proportion found using either the EMC-92 or the UAMS-80 signatures. Ranked, weighted high-risk proportions are GPI: 10.0%, UAMS-17: 12.4%, UAMS-70: 13.0%, MRC-IX-6: 13.3%, EMC-92: 19.1% and UAMS-80: 23.4%. To determine which signature best explained the observed survival, pair-wise comparisons were performed. For every comparison the EMC-92 was the strongest predictor for OS tested in an independent environment (Figure 3 and Supplemental Table S9).

Figure 3
figure3

Pair-wise comparison for all signatures. To find the signature best fitting the underlying data sets, Cox regression models (high-risk versus standard-risk) were made for all pair-wise signatures. These models are based on pooled independent datasets (i.e., excluding training sets) and stratified for study. The two paired hazard ratios associated with the signatures derived per model are shown in the two cells within the square panels. Only hazard ratios within one panel can be compared because these are based on the same dataset. Blue cells indicate signifcant hazard ratios (Bonferroni-Holm corrected P-value); red cells denote non-significant findings. For the bottom right panel (i.e., UAMS-70 vs EMC-92 signatures) the underlying model is given. All other models can be found in Supplemental Table S9.

There is a varying degree of overlapping probe sets between all signatures. Overlapping genes are shown in Supplemental Figure S3. Seven out of fifty probe sets present in the GPI-50 overlap with the EMC-92 signature (BIRC5, FANCI, ESPL1, MCM6, NCAPG, SPAG5 and ZWINT). One of the six MRC-IX genes (ITM2B) is also seen in the EMC-92. Overlap between EMC-92 and the remaining signatures is limited (EMC92 vs UAMS17/70: BIRC5 and LTBP1; EMC-92 vs MILLENNIUM-100: MAGEA6 and TMEM97; and EMC-92 vs IFM-15: FAM49A).

Combined risk classifiers

The performance of the EMC-92 signature was in line with the UAMS signatures, although they were derived from quite different patient populations. The intersection of high-risk patients between the EMC-92 and UAMS-70 signatures was 8% of the total population on the pooled data sets that were independent of both our training set and the UAMS-70 training set (that is, MRC-IX, TT3 and APEX; Supplemental Table S11). Approximately 13% of patients were classified as high-risk by either one of these signatures. The intersecting high-risk group had the highest HR as compared with the intersecting standard-risk group (HR=3.87, 95% CI=2.76–5.42, P=3.6 × 10−15). Patients classified as high-risk by either signature showed an intermediate risk, that is, with a HR of 2.42, 95% CI=1.76–3.32, for the EMC-92 signature (P=5.1 × 10−8) and a HR of 2.22, 95% CI=1.20–4.11, for the UAMS-70 signature (P=1.1 × 10−2; Supplemental Table S12). To test whether there is evidence for better performance if outcomes of two dichotomous predictors are merged, we took the models made in the pair-wise comparison (Supplemental Table S9) and tested these in a likelihood-ratio test against a single signature outcome model. Merging the EMC-92 with UAMS-80 (P=2.19 × 10−3), UAMS-17 (P=9.36 × 10−3), GPI-50 (P=2.95 × 10−2), MRC-IX-6 (P=1.58 × 10−2) and UAMS-70 (P=3.96 × 10−2) demonstrated a better fit to the data than any of the single models (Supplemental Table S10).

EMC-92 signature and FISH

To compare the high-risk populations composition as defined by the EMC-92 and the UAMS-70 signatures, cytogenetic aberration frequencies in both populations were determined using an independent set for which cytogenetic variables were known, that is, MRC-IX (Figure 4 and Supplemental Table S13). As expected, poor prognostic cytogenetic aberrations 1q gain, del(17p), t(4;14), t(14;16), t(14;20) and del(13q) were enriched in the high-risk populations (Figure 5), whereas the standard-risk cytogenetic aberrations such as t(11;14) were diminished in the high-risk populations. In contrast, only 15% (6 out of 39) of MRC-IX cases with high-risk status as determined by the EMC-92 signature showed absence of any poor prognostic cytogenetic aberrations, as opposed to 44% (74 out of 168) in standard-risk cases (P=1.8 × 10−3). Similarly, of the UAMS-70-defined high-risk patients 4% (1 out of 23) did not have any poor prognostic cytogenetics, whereas of the UAMS-70 defined standard risk patients this proportion was 43% (79 out of 183) (P=5.3 × 10−3).

Figure 4
figure4

Distributions of high-risk and standard-risk patients per FISH marker in the MRC-IX data set. Distribution of FISH markers within the high-risk (top panels) and standard-risk (bottom panels) groups for the EMC-92 and UAMS-70 signatures. The EMC-92 and UAMS-70 identified 50 and 42 patients out of 247 as high-risk, respectively. Blue, absence of an aberration; OR, odds-ratio; P, Fisher exact P-value; red, presence of an aberration; white, missing data. Details are given in Supplemental Table S13.

Figure 5
figure5

Poor prognostic cytogenetic aberrations in comparison with the EMC-92 signature in MRC-IX patients. Each horizontal line represents one patient. The first column denotes the distinction between high-risk (in red, n=50) and standard risk (in blue, n=197). Columns 2−7 represent cytogenetic aberrations as shown. Red, presence of an aberration; blue, absence; and white, missing data. More than half of the EMC-92 standard risk patients are affected by one or more poor FISH markers.

Discussion

Here we report on the generation and validation of the EMC-92 signature, which was based on the HOVON65/GMMG-HD4 clinical trial. Conventional prognostic markers such as ISS stage and adverse cytogenetics have been augmented by signatures based on gene expression in order to increase accuracy in outcome prediction in MM. More accurate prognosis may lead to the development of treatment schedules that are specifically aimed at improving survival of high-risk MM patients. Prognostic signatures for MM include the UAMS-70, the UAMS-17, the UAMS-80, the IFM-15, the gene proliferation index (GPI-50), the MRC-IX-6 and the MILLENNIUM-100 signatures.

For clinical relevance, a signature must have both the ability to separate risk groups as clearly as possible and to predict stable groups of relevant size. The EMC-92 signature meets both criteria. In all validation sets a high-risk group of patients can be significantly determined and the proportion of high-risk patients is stable across the validation sets. The validation sets represent different drug regimens, including thalidomide (MRC-IX, TT2) and bortezomib (APEX, TT3). Also, the signature is relevant to both transplant-eligible (for example, TT3) and non-transplant-eligible patients (subset of MRC-IX), as well as newly diagnosed (for example, TT2) and relapsed patients (APEX). In contrast, the predictions of the IFM-15 and MILLENNIUM-100 signatures in the validation sets fail to reach significance in independent data sets such as MRC-IX and TT3. The differences in gene expression platform may have contributed to this. Indeed, the IFM signature is based on a custom cDNA-based gene expression platform, rather than the Affymetrix GeneChips, which have become common for MM GEP studies.32 The cDNA platforms have been reported to be difficult to compare with the Affymetrix oligonucleotide platform.12 Although the MILLENNIUM signature was generated using Affymetrix GeneChips, the use of an earlier version of this platform may have contributed to the limited performance of this signature.11 The performance of the EMC-92 signature is comparable to the UAMS-derived signatures, MRC-IX-6 and GPI-50, as measured by the significance of prediction in validation sets. For UAMS-70 and GPI-50, the proportion of high-risk patients appears more variable, which may hinder clinical interpretation, especially when the high-risk proportion is <10%. Importantly, pair-wise comparisons of all the signatures evaluated in this paper demonstrated that the EMC-92 has the best fit to the observed survival times in independent sets.

Strikingly, we found that performance can be improved by simply combining signatures (for example, EMC-92 with UAMS-80). However, this analysis is only an indication of the possibilities of combining signatures, and future work involving more complex combined signatures is in progress.

It is important to note that the genes within the signature reflect optimal performance of the signature rather than a biological definition of survival in MM. The initially selected 1093 probe sets, which were found to be associated with PFS in univariate testing, are more likely to give a good representation of myeloma biology, as indicated for instance by the protein synthesis-related pathways. Although an extended biological discussion is outside the scope of this paper, a number of interesting genes are included in the signature. BIRC5 was found in four signatures evaluated in this paper: EMC-92, UAMS-17, UAMS-70 and GPI-50. This gene is a member of the inhibitor of apoptosis gene family, which encodes negative regulatory proteins that prevent apoptotic cell death, and upregulation has been described to be associated with lower EFS and OS in newly diagnosed MM patients.11, 12, 31 Other important myeloma genes include FGFR3 and STAT1. FGFR3 is deregulated as a result of translocation t(4;14), which is an adverse prognostic cytogenetic event.30 FGFR3, a transmembrane receptor tyrosine kinase, is involved in the regulation of cell growth and proliferation.33 STAT1, an important component of the JAK/STAT signaling, is involved in multiple pathways, including apoptosis induced by interferon signaling.29

A clear enrichment of the long arm of chromosome 1 was observed in the 1093 probe sets in this study. Previously the importance of chromosome 1 was reported for the UAMS-70 signature. Genes on 1q in the UAMS-70 signature include CKS1B and PSMD4, both of which were not in the EMC- 92 signature, although CKS1B was found to be associated with PFS in our set and thus in the 1093 set.9, 10 The EMC-92 signature did contain nine genes on 1q, of which S100A6 has been described in relation to 1q21 amplification in MM and other cancer types.34 This may also be part of the explanation why, despite the use of the same GEP platform, the overlap between different signatures is limited. Indeed, multiple genes are found within the 1q21 amplicon with downstream factors possibly overexpressed as a result of this. Which gene will be linked most significantly to survival in a specific set is most likely determined by factors such as variability in data sets, to which population differences and differences in used techniques may contribute.

Other reasons may be found in the difference in treatment strategies used, in which other genes could be responsible for adverse prognosis.

To characterize the high-risk group in depth, we have demonstrated that in the MRC-IX study high-risk patients are enriched for poor cytogenetic aberrations such as 1q gain, del(17p), t(4;14), t(14;16), t(14;20) and del(13q). Still more than half of the patients in the standard-risk group showed one or more poor prognostic cytogenetic markers, indicating that the occurrence of a single poor-risk marker does not have very strong prognostic value.

Clinical use of a gene signature (UAMS-70) has recently been incorporated in the mSMART risk stratification, which additionally includes FISH, metaphase cytogenetics and plasma cell labeling index. The mSMART risk stratification is the first risk stratification system adjusting treatment regimens according to risk status, although this has not been validated in prospective clinical trials.15, 35 Ultimately, clinical use of any signature has been proven to be of use in prospective clinical trials, which will allow treatment choice based on risk assessment. This will yield clinical guidelines to improve treatment of patients with a poor PFS and OS on novel therapies. For practical application of the EMC-92 signature it is essential to stress that this signature has not been designed for classification of a single patient. However, collection of a set of more than 25 patients will result in reliable prediction, and each additional patient can be predicted as soon as it is tested.

In conclusion, we have developed a risk signature that is highly discriminative for patients with high-risk vs standard-risk MM, irrespective of treatment regime, age and relapse setting. Use of this signature in the clinical setting may lead to a more informed treatment choice and potentially better outcome for the patient.

Accession codes

Accessions

Gene Expression Omnibus

References

  1. 1

    Avet-Loiseau H . Ultra high-risk myeloma. Am Soc Hematol Educ Program 2010, 489–493.

    Article  Google Scholar 

  2. 2

    Palumbo A, Anderson K . Multiple myeloma. N Engl J Med 2011; 17: 1046–1060.

    Article  Google Scholar 

  3. 3

    Munshi NC, Anderson KC, Bergsagel PL, Shaughnessy J, Palumbo A, Durie B et al. Consensus recommendations for risk stratification in multiple myeloma: report of the International Myeloma Workshop Consensus Panel 2. Blood 2011; 117: 4696–4700.

    CAS  Article  Google Scholar 

  4. 4

    Neben K, Jauch A, Bertsch U, Heiss C, Hielscher T, Seckinger A et al. Combining information regarding chromosomal aberrations t(4;14) and del(17p13) with the International Staging System classification allows stratification of myeloma patients undergoing autologous stem cell transplantation. Haematologica 2010; 95: 1150–1157.

    CAS  Article  Google Scholar 

  5. 5

    Bergsagel PL, Kuehl WM . Molecular pathogenesis and a consequent classification of multiple myeloma. J Clin Oncol 2005; 23: 6333–6338.

    CAS  Article  Google Scholar 

  6. 6

    Zhan F, Huang Y, Colla S, Stewart JP, Hanamura I, Gupta S et al. The molecular classification of multiple myeloma. Blood 2006; 108: 2020–2028.

    CAS  Article  Google Scholar 

  7. 7

    Broyl A, Hose D, Lokhorst H, de Knegt Y, Peeters J, Jauch A et al. Gene expression profiling for molecular classification of multiple myeloma in newly diagnosed patients. Blood 2010; 116: 2543–2553.

    CAS  Article  Google Scholar 

  8. 8

    Chng WJ, Kuehl WM, Bergsagel PL, Fonseca R . Translocation t(4;14) retains prognostic significance even in the setting of high-risk molecular signature. Leukemia 2008; 22: 459–461.

    CAS  Article  Google Scholar 

  9. 9

    Shaughnessy JD, Zhan F, Burington BE, Huang Y, Colla S, Hanamura I et al. A validated gene expression model of high-risk multiple myeloma is defined by deregulated expression of genes mapping to chromosome 1. Blood 2007; 109: 2276–2284.

    CAS  Article  Google Scholar 

  10. 10

    Shaughnessy JD, Qu P, Usmani S, Heuck CJ, Zhang Q, Zhou Y et al. Pharmacogenomics of bortezomib test-dosing identifies hyperexpression of proteasome genes, especially PSMD4, as novel high-risk feature in myeloma treated with total therapy 3. Blood 2011; 118: 3512–3524.

    CAS  Article  Google Scholar 

  11. 11

    Mulligan G, Mitsiades C, Bryant B, Zhan F, Chng WJ, Roels S et al. Gene expression profiling and correlation with outcome in clinical trials of the proteasome inhibitor bortezomib. Blood 2007; 109: 3177–3188.

    CAS  Article  Google Scholar 

  12. 12

    Decaux O, Lode L, Magrangeas F, Charbonnel C, Gouraud W, Jezequel P et al. Prediction of survival in multiple myeloma based on gene expression profiles reveals cell cycle and chromosomal instability signatures in high-risk patients and hyperdiploid signatures in low-risk patients: a study of the Intergroupe Francophone du Myelome. J Clin Oncol 2008; 26: 4798–4805.

    CAS  Article  Google Scholar 

  13. 13

    Dickens NJ, Walker BA, Leone PE, Johnson DC, Brito JL, Zeisig A et al. Homozygous deletion mapping in myeloma samples identifies genes and an expression signature relevant to pathogenesis and outcome. Clin Cancer Res 2010; 16: 1856–1864.

    CAS  Article  Google Scholar 

  14. 14

    Hose D, Reme T, Hielscher T, Moreaux J, Messner T, Seckinger A et al. Proliferation is a central independent prognostic factor and target for personalized and risk-adapted treatment in multiple myeloma. Haematologica 2011; 96: 87–95.

    Article  Google Scholar 

  15. 15

    Dispenzieri A, Rajkumar SV, Gertz MA, Fonseca R, Lacy MQ, Bergsagel PL et al. Treatment of newly diagnosed multiple myeloma based on Mayo Stratification of Myeloma and Risk-adapted Therapy (mSMART): consensus statement. Mayo Clin Proc 2007; 82: 323–341.

    CAS  Article  Google Scholar 

  16. 16

    Sonneveld P, Schmidt-Wolf I, van der Holt B, Jarari Le, Bertsch U, Salwender H et al. HOVON-65/GMMG-HD4 randomized phase III trial comparing bortezomib, doxorubicin, dexamethasone (PAD) vs VAD followed by high-dose melphalan (HDM) and maintenance with bortezomib or thalidomide in patients with newly diagnosed multiple myeloma (MM). Blood 2010; 116: 40.

    Google Scholar 

  17. 17

    Barlogie B, Pineda-Roman M, van Rhee F, Haessler J, Anaissie E, Hollmig K et al. Thalidomide arm of Total Therapy 2 improves complete remission duration and survival in myeloma patients with metaphase cytogenetic abnormalities. Blood 2008; 112: 3115–3121.

    CAS  Article  Google Scholar 

  18. 18

    Pineda-Roman M, Zangari M, Haessler J, Anaissie E, Tricot G, van Rhee F et al. Sustained complete remissions in multiple myeloma linked to bortezomib in total therapy 3: comparison with total therapy 2. Br J Haematol 2008; 140: 625–634.

    CAS  Article  Google Scholar 

  19. 19

    Morgan GJ, Davies FE, Gregory WM, Bell SE, Szubert AJ, Navarro-Coy N et al. Thalidomide maintenance significantly improves progression-free survival (PFS) and overall survival (OS) of myeloma patients when effective relapse treatments are used: MRC myeloma IX results. Blood 2010; 116: 623–623.

    Google Scholar 

  20. 20

    Morgan GJ, Davies FE, Owen RG, Rawstron AC, Bell S, Cocks K et al. Thalidomide Combinations improve response rates; results from the MRC IX study. Blood 2007; 110: 3593–3593.

    Google Scholar 

  21. 21

    Jagannath S, Barlogie B, Berenson J, Siegel D, Irwin D, Richardson PG et al. A phase 2 study of two doses of bortezomib in relapsed or refractory myeloma. Br J Haematol 2004; 127: 165–172.

    CAS  Article  Google Scholar 

  22. 22

    Richardson PG, Sonneveld P, Schuster MW, Irwin D, Stadtmauer EA, Facon T et al. Bortezomib or high-dose dexamethasone for relapsed multiple myeloma. N Engl J Med 2005; 352: 2487–2498.

    CAS  Article  Google Scholar 

  23. 23

    Richardson PG, Barlogie B, Berenson J, Singhal S, Jagannath S, Irwin D et al. A phase 2 study of bortezomib in relapsed, refractory myeloma. N Engl J Med 2003; 348: 2609–2617.

    CAS  Article  Google Scholar 

  24. 24

    Gentleman R, Carey V, Bates D . Bioconductor: Open software development for computational biology and bioinformatics. Genome Biol 2004; 5: R80.

    Article  Google Scholar 

  25. 25

    Johnson WE, Li C, Rabinovic A . Adjusting batch effects in microarray expression data using empirical Bayes methods. Biostatistics 2007; 8: 118–127.

    Article  Google Scholar 

  26. 26

    Therneau T, Lumley T . Survival: survival analysis, including penalised likelihood. R package version 236-2. 2010.

  27. 27

    Bair E, Hastie T, Paul D, Tibshirani R . Prediction by supervised principal components. J Amer Statistical Assoc 2006; 101: 119–137.

    CAS  Article  Google Scholar 

  28. 28

    Benjamini Y, Hochberg Y . Controlling the false discovery rate: a practical and powerful approach to multiple testing. J Roy Stat Soc B Met 1995; 57: 289–300.

    Google Scholar 

  29. 29

    Arulampalam V, Kolosenko I, Hjortsberg L, Bjorklund AC, Grander D, Tamm KP . Activation of STAT1 is required for interferon-alpha-mediated cell death. Exp Cell Res 2011; 317: 9–19.

    CAS  Article  Google Scholar 

  30. 30

    Chesi M, Nardini E, Brents LA, Schrock E, Ried T, Kuehl WM et al. Frequent translocation t(4;14)(p16.3;q32.3) in multiple myeloma is associated with increased expression and activating mutations of fibroblast growth factor receptor 3. Nat Genet 1997; 16: 260–264.

    CAS  Article  Google Scholar 

  31. 31

    Hideshima T, Catley L, Raje N, Chauhan D, Podar K, Mitsiades C et al. Inhibition of Akt induces significant downregulation of survivin and cytotoxicity in human multiple myeloma cells. Br J Haematol 2007; 138: 783–791.

    CAS  Article  Google Scholar 

  32. 32

    Mah N, Thelin A, Lu T, Nikolaus S, Kuhbacher T, Gurbuz Y et al. A comparison of oligonucleotide and cDNA-based microarray systems. Physiol Genomics 2004; 16: 361–370.

    CAS  Article  Google Scholar 

  33. 33

    Trudel S, Ely S, Farooqi Y, Affer M, Robbiani DF, Chesi M et al. Inhibition of fibroblast growth factor receptor 3 induces differentiation and apoptosis in t(4;14) myeloma. Blood 2004; 103: 3521–3528.

    CAS  Article  Google Scholar 

  34. 34

    Inoue J, Otsuki T, Hirasawa A, Imoto I, Matsuo Y, Shimizu S et al. Overexpression of PDZK1 within the 1q12-q22 amplicon is likely to be associated with drug-resistance phenotype in multiple myeloma. The Am J Pathol 2004; 165: 71–81.

    CAS  Article  Google Scholar 

  35. 35

    Kumar SK, Mikhael JR, Buadi FK, Dingli D, Dispenzieri A, Fonseca R et al. Management of newly diagnosed symptomatic multiple myeloma: updated Mayo Stratification of Myeloma and Risk-Adapted Therapy (mSMART) consensus guidelines. Mayo Clin Proc 2009; 84: 1095–1110.

    CAS  Article  Google Scholar 

Download references

Acknowledgements

We would like to thank John Shaughnessy for kindly providing the data. Peter Valk and Mathijs Sanders are acknowledged for insightful discussions. Financial support: This work was supported by the Center for Translational Molecular Medicine (BioCHIP), a clinical research grant from the European Hematology Association and an EMCR Translational Research Grant from Skyline Diagnostics, Janssen and MSCNET.

Author contributions

RK, AB, MVD and PS designed the research. AB, YDK performed the research and collected the data. RK developed the classifier, AB, RK, MVD analyzed and interpreted the data and wrote the manuscript. MVV, EVB analyzed and interpreted the data and critically reviewed the paper. BVDH and LEJ performed data management and statistical analyses. GM provided the CEL files from the APEX data set and critically reviewed the paper. GM and WG provided CEL files and clinical data from the MRC-IX data set and critically reviewed the paper. HG is principal investigator of the performed research in the German part of the H65/GMMG-HD4 and critically reviewed the paper. HL organized the trial and critically reviewed the paper. PS organized the trial, is principal investigator of the performed research in the Dutch part of the H65/GMMG-HD4 and critically reviewed the paper.

Author information

Affiliations

Authors

Corresponding author

Correspondence to P Sonneveld.

Ethics declarations

Competing interests

GM has declared a financial interest in Millennium Pharmaceuticals Inc., whose product was an object of study in the HOVON 65/ GMMG-HD4. GM is currently employed by Millennium Pharmaceuticals Inc. HG has served on the advisory boards of Johnson & Johnson. HL is on the advisory Boards of Celgene and Genmab. EVB and MVV are employees of Skyline Diagnostics. PS is on the advisory boards of Skyline Diagnostics, Janssen and Celgene. The other authors declare no conflict of interest.

Additional information

Supplementary Information accompanies the paper on the Leukemia website

Supplementary information

Rights and permissions

Reprints and Permissions

About this article

Cite this article

Kuiper, R., Broyl, A., de Knegt, Y. et al. A gene expression signature for high-risk multiple myeloma. Leukemia 26, 2406–2413 (2012). https://doi.org/10.1038/leu.2012.127

Download citation

Keywords

  • multiple myeloma
  • gene expression
  • signature
  • prognosis
  • survival
  • comparison

Further reading

Search

Quick links