Dr Christoph Thorns, Institute for Pathology, University Clinic Schleswig-Holstein, Campus Luebeck, Ratzeburger Allee 160, 23538 Luebeck, Germany. E-mail: firstname.lastname@example.org
MicroRNAs (miRNA, miR) are negative regulators of gene expression that play an important role in diverse biological processes such as development, cell growth, apoptosis and haematopoiesis, suggesting their association with cancer. Here we analysed the expression signatures of 157 miRNAs in 58 diffuse large B-cell lymphoma (DLBCL), 46 follicular lymphoma (FL) and seven non-neoplastic lymph nodes (LN). Comparison of the possible combinations of DLBCL-, FL- and LN resulted in specific DLBCL- and FL-signatures, which include miRNAs with previously published function in haematopoiesis (MIRN150 and MIRN155) or tumour development (MIRN210, MIRN10A, MIRN17-5P and MIRN145). As compared to LN, some miRNAs are differentially regulated in both lymphoma types (MIRN155, MIRN210, MIRN106A, MIRN149 and MIRN139). Conversely, some miRNAs show lymphoma-specific aberrant expression, such as MIRN9/9*, MIRN301, MIRN338 and MIRN213 in FL and MIRN150, MIRN17-5P, MIRN145, MIRN328 and others in DLBCL. A classification tree was computed using four miRNAs (MIRN330, MIRN17-5P, MIRN106a and MIRN210) to correctly identify 98% of all 111 cases that were analysed in this study. Finally, eight miRNAs were found to correlate with event-free and overall survival in DLBCL including known tumour suppressors (MIRN21, MIRN127 and MIRN34a) and oncogenes (MIRN195 and MIRNLET7G).
In this report, follicular lymphoma (FL), as a tumour of germinal centre origin, was compared with GCL-DLBCL and non-GCB-DLBCL. We present specific DLBCL and FL miRNA signatures. Similarities between lymphoma profiles were identified and entity-specific miRNA expression was singled out. An outstanding role for MIRN155 in lymphoma development could be confirmed. A classification tree was developed that correctly identified 98% of all DLBCL-, FL- and LN cases. Finally, a correlation between miRNA expression levels and outcome (event-free and overall survival) was found in a multivariate analysis.
The study was conducted in accordance with the Helsinki declaration. Patients were eligible if they had previously untreated, biopsy-confirmed aggressive non-Hodgkin’s lymphoma according to the World Health Organization criteria. Tissue samples of 58 DLBCL were included in this study. These patients were treated within the NHL-B1 and NHL-B2 trials of the German High-Grade Non-Hodgkin’s Lymphoma Study Group (Pfreundschuh et al, 2004a,b). Seven lymph nodes (LN) were included as non-neoplastic lymphatic tissue. These included LN with only minor reactive changes. 46 samples of low grade FL were retrieved from the files of the pathology departments (UKS-H, Campus Luebeck and Semmelweis University, Budapest). All FL samples were grade 1 and 2 tumours and no transformation in DLBCL was noted.
Total RNA was extracted from four 20 μm sections of formalin-fixed paraffin-embedded tissues using the RecoverAll kit (Ambion, Austin, Texas, USA) according to the manufacturer’s protocol. RNA concentrations were subsequently quantified using a NanoDrop Spectrophotometer (NanoDrop Technologies, Wilmington, Delaware, USA). MiRNA quantification took place in two steps: (i) miRNA-specific reverse transcription (RT); (ii) quantitative PCR using diluted miRNA-specific cDNA as a template.
cDNA was synthesized from total RNA using gene-specific stem-loop primers according to the TaqMan MicroRNA Assay protocol (PE Applied Biosystems, Foster City, California, USA). Each 15 μl RT-reaction contained 10 ng of total RNA, 3 μl stem-loop RT primer, 1·5 μl 10 × RT buffer, 0·15 μl of dNTPs (100 mM), 1 μl MultiScribe reverse transcriptase (50 U/μl) and 0·188 μl RNase Inhibitor (20 U/μl) (TaqMan MicroRNA Reverse Transcription Kit, PE Applied Biosystems) and 4·162 μl nuclease-free water (Ambion, Austin, Texas, USA). They were incubated in a T1 Thermocylcer (Biometra, Goettingen, Germany) using a 96-well format and the following settings: 30 min at 16°C, 30 min at 42°C, 5 min at 85°C and held at 4°C. The resulting cDNA was then diluted 1:15 using nuclease-free water (Ambion). Real-time PCR was performed using an Applied Biosystems 7900HT Fast Real-Time PCR System. The 20 μl PCR included 1·33 μl diluted RT product, 1× TaqMan Universal PCR master mix and 2 μl of primers and probe mix of the TaqMan MicroRNA Assay protocol (PE Applied Biosystems). The reactions were incubated in a 96-well optical plate (PE Applied Biosystems) at 95°C for 10 min, followed by 40 cycles of 95°C for 15 s and 60°C for 1 min. The Ct was determined using default threshold settings.
Follicular lymphoma miRNA-profiles were measured on the LightCycler® 480 Instrument (Roche, Basel, Switzerland). The following program was used: (i) Enzyme Activation: 95°C 10 min; (ii) Amplification (45 cycles): 95°C 15 s (ramp: 4·4°C/s, analysis mode: quantification), 60°C 1 min (ramp: 2·2°C/s); (iii) Cooling: 40°C 30 s (ramp: 2°C/s). The detection format was set to ‘Mono Colour Hydrolysis Probe’ and the second derivative maximum method was used for absolute quantification. In order to demonstrate that miRNA profiles obtained from the LightCycler® 480 and the 7900HT instrument are highly similar, all seven LN profiles were measured and analysed on both real-time machines. Confidence intervals (95%) for a paired t-test comparing both PCR techniques include 134 miRNA species in the null hypothesis value of zero difference and no significant differential expression after correction for multiplicity.
Immunohistochemical (IHC) stainings were performed according to a standard three-step immunoperoxidase technique with diaminobenzidine as chromogen.
Fifteen different IHC stainings were performed on every DLBCL case previously analysed (BCL6, CD10, MUM1, IgM, BCL2, CD138, Ki-B5, Ki-B3, CD23, OPD4, HLA-DR, IgD, CD38, Ig light chain κ and light chain λ). The following antibodies were used: mouse monoclonal anti-BCL6 (1:10, Dako, Glostrup, Denmark), rabbit polyclonal anti-CD10 (1:10, Dako), mouse monoclonal anti-MUM1 (1:100, Dako), rabbit polyclonal anti-IgM (1:500, Dako), mouse monoclonal anti-BCL2 (1:50, Biocarta, Hamburg, Germany), mouse monoclonal anti-OPD4 (1:50, Dako), mouse monoclonal anti-HLA-DR (1:200, Dako), rabbit polyclonal anti-IgD (1:30, Dako), rabbit polyclonal anti-κ (1:1000, Dako) and rabbit polyclonal anti-λ (1:1000, Dako). All secondary antibodies and all reagents were purchased from Dako.
Except outcome analysis, all analyses are based on standard statistical functions available from the R-language (http://www.R-project.org).
Quantile normalisation was used to normalise the miRNA profiles (modified from Bolstad et al, 2003; as implemented in the R-package limma). Values of Ct = 40 (saturation) were deleted before this step and added again afterwards.
Differential miRNA expression
Differential expression was tested both on a global profile level (Goeman et al, 2004) and for all single miRNA species using Welsh two-sided t-tests. Multiplicity was corrected taking a False Discovery Rate (FDR) approach (Benjamini et al, 2001). miRNAs were called differentially expressed for FDR < 0·05 and ΔCt ≥ 1·5.
Unsupervised data mining
MiRNA profiles correlation structures were analysed using principal component analyses. Scores plots for profile data were colour-coded using the sample type information [LN, FL, DLBCL (GCB and non-GCB)].
Partial Least Squares regression discriminance analysis (PLS-DA) was performed as a supervised learning method to find combinations of miRNAs differentiating the different tumour types as opposed to the LN samples. The work flow was as follows: First, in a cross-validation approach we sought to determine if PLS regression using different numbers of latent variables (i.e. linear combinations of miRNA profiles) arrived at acceptable classification results. As a quality measure, we computed the error of prediction normalized with the variance of the variable in question (Q2) (Eriksson et al, 1999). From these results, the optimal number of latent PLS components was assessed. Second, via permutation analysis, we assessed the significance of the PLS-DA based on the distribution of Q2-values for 100 permutated response vectors. Third, candidate miRNAs with exceptional high contributions to the combination optimized for covariance with the response were selected using the variable importance plot (VIP) score (Chong & Jun, 2005). Candidates for combinatorial differentiation of the tumour types under investigation could be filtered out by comparing lists of differential expression against VIP-scores.
A classification tree was trained on a prefiltered set of miRNA profiles. MiRNA profiles used for this analysis were MIRN9, MIRN301, MIRN320, MIRN149, MIRN150, MIRN155, MIRN145, MIRN330, MIRN92, MIRN338– as taken from top of the lists of differential expression between the three sample groups (LN, FL, DLBCL) in this study (Fig 1). The classification tree was calculated with cross-entropy as measure for node impurity. A tree is grown by binary recursive partitioning using Ct values and choosing splits from the set of miRNAs described above. The split which maximizes the reduction in impurity is chosen, the data set split and the process repeated. Splitting continues until the terminal nodes are too small or too few to be split.
The data of 53 patients were available for outcome analysis. From 153 miRNAs 51 with a range ≤ 3 Ct were excluded. The median for each of the miRNAs (group 0: <median, group 1: ≥median; except for MIRN21: group 0: ≤median, group 1: >median) was applied as the cut-off points. Event-free survival (EFS) was defined as the time from the beginning of therapy to either disease progression, initiation of additional (off-protocol) or salvage therapy, relapse or death. Overall survival (OS) was defined as time from the beginning of therapy to death for any cause. EFS and OS were estimated according to Kaplan and Meier (Kaplan & Meier, 1958). In a univariate analysis logrank tests were performed and miRNAs with P < 0·1 were included in the multivariate analysis. Proportional hazard models for each of the selected miRNAs separately were fitted adjusted for the prognostic factors of the International Prognostic Index (age > 60 years, lactate dehydrogenase [LDH] > normal, Eastern Cooperative Oncology Group [ECOG] performance score > 1, stage III/IV and extranodal involvement > 1). For miRNAs with P-values <0·05 for EFS or OS within the Coxmodels, the relative risk with 95% confidence interval (CI) is shown in a forest plot. Because of the explorative character of the analysis and the small number of patients the P-values were not adjusted for multiple comparisons. Outcome analyses were performed with the Statistical Package for the Social Sciences (SPSS) v. 15.0.
MiRNAs differentially expressed between DLBCL and FL
By comparing miRNA profiles of seven LN and 58 DLBCL, we identified 15 differentially expressed miRNAs (Fig 1A). Four miRNAs (MIRN210, MIRN155, MIRN106A, MIRN17-5P) were expressed at a significantly higher level in DLBCL than in normal tissue. In contrast, the remaining 11 miRNAs (MIRN150, MIRN145, MIRN328, MIRN139, MIRN99a, MIRN10a, MIRN95, MIRN149, MIRN320, MIRN151, MIRNLET7E) were expressed at a significantly lower level in DLBCL. Overexpression of MIRN210 was highest (7-fold) and expression of MIRN150 was most strongly reduced (7-fold).
Similarly, comparison of the miRNA profiles of seven LN and 46 cases of FL identified 12 differentially expressed miRNAs (Fig 1B). Nine miRNAs (MIRN9, MIRN301, MIRN213, MIRN9*, MIRN330, MIRN106A, MIRN338, MIRN155, MIRN210) were significantly up- and three miRNAs (MIRN320, MIRN149, MIRN139) down-regulated in the lymphoma samples. Expression of MIRN9 was most strongly increased (12-fold) and of MIRN320 most strongly decreased (3-fold). Interindividual differences of MIRN9/MIRN9* expression was noticed to be particularly high in DLBCL, FL and LN (up to 32-fold).
Comparing DLBCL and FL with another (Fig 1C), two miRNAs (MIRN17-5P, MIRN92) were overexpressed and eight miRNAs (MIRN330, MIRN338, MIRN135A, MIRN150, MIRN125B, MIRN301, MIRN126, MIRN213) were down-regulated in DLBCL.
Identifying miRNAs of general importance in lymphoma genesis
To identify miRNAs with a general role in lymphoma genesis, 20 miRNAs that showed the strongest differential expression between DLBCL and FL were removed and obtained a new list of miRNAs with highly similar expression of the remaining 137 miRNA species in both lymphoma entities. This new list was then used for a ‘lymphoma versus reactive lymphatic tissue’ comparison (Fig 1D) and taking a PLS-DA approach. A VIP was computed in order to identify miRNAs that might play a general role in lymphoma development (Fig 2A). Amongst others, MIRN155, MIRN210, MIRN106A, MIRN149 and MIRN139 most strongly contribute to differences in miRNA-profiles of lymphomas and reactive lymphatic tissue (Fig 1D). ‘Leave one out’ (LOO) cross validation supported the validity of the identified miRNAs. The hit rate was determined to be 96% (Fig 2B).
Using miRNAs for DLBCL subgroup identification
In order to assign the 58 cases of DLBCL the GCB and non-GCB subgroups, we analysed IHC stains of CD10, BCL6 and MUM1 were analysed (Hans et al, 2004). 25 cases were assigned to GCB-DLBCL and non-GCB-DLBCL, respectively. Eight cases with unclear GCB-status, due to non-evaluable IHC stains, were removed from further analysis.
It was now possible to address the four entities, a) reactive LN, b) FL, c) GCB-DLBCL and d) non-GCB-DLBCL. Principal components analysis indicated that miRNA profiles created separate groups for LN, FL and DLBCL (Fig 3A, green, red and blue dots, respectively). However, GCB- and non-GCB-DLBCL cases (Fig 3A, light blue dots and dark blue dots, respectively) appeared to be hard to separate. MiRNAs important for discriminating GCB versus non-GCB cases were filtered using a t-test approach. Nine miRNAs were identified as candidates for differential expression taking into account an FDR cut-off of 20% (Fig 3B). Interestingly, MIRN155 was identified as one of the discriminating markers, as previously reported (Lawrie et al, 2007).
Four miRNAs accurately diagnose DLBCL and FL
The minimal number of miRNAs that can be used to accurately diagnose and discriminate DLBCL and FL cases needed to be determined. MIRN330, MIRN17-5P, MIRN106A and MIRN210 were the most discriminatory miRNAs. The classification tree (Fig 4A) showed that expression of MIRN330 alone (Ct ≤ 38·6) could separate FL from DLBCL and non-neoplastic LN with high confidence (40 of 46 cases), while erroneously assigning one additional case of DLBCL to this group. Similarly, using expression results of three additional miRNAs [MIRN17-5P (Ct < 32·17), MIRN106A (Ct < 33·45) and MIRN210 (Ct < 33·66)] correctly grouped 57 of 58 cases of DLBCL, six of seven LN and all (46 of 46) FL. In summary, the classification tree, shown in Fig 4(A), utilized miRNA expression results of four miRNAs to discriminate DLBCL, FL and LN with an overall accuracy of 98% (109 of 111 cases).
MiRNAs with prognostic capability in DLBCL
The Characteristics of the patients with diagnosis DLBCL are shown in Table I. The median age was 63 years, with 3- and 5-year OS rates of 71·2% and 66·9%, respectively. The median follow-up time was 72 months.
Table I. Patient characteristics at initial diagnosis (n = 53).
The relationship between expression of single miRNAs and survival prognosis was examined using a multivariate Cox regression analysis including factors of the International Prognostic Index (IPI) (Table I). MiRNAs with significance level P < 0·1 for EFS or OS in a univariate analysis (EFS: MIRN19A, MIRN29A, MIRN92, MIRN127, MIRN195, MIRN222, MIRNLET7G; OS: MIRN21, MIRN23A, MIRN27A, MIRN142-5P, MIRN182*, MIRN296; both: MIRN34A, MIRN125A, MIRN186) were used for multivariate analysis. Based on this, eight miRNAs (MIRN19A, MIRN21, MIRN23A, MIRN27A, MIRN34A, MIRN127, MIRN195 and MIRLET7G) with P-values <0·05 for EFS or OS were identified (Fig 5A). Only one miRNA, MIRN127, was found to significantly influence EFS and OS (Fig 5B). In a multivariate analysis (Fig 5A) patients with low expression of MIRN127 (Ct ≥ 37·73) had a poor prognosis for OS (RR = 4·3, 95% CI, 1·2–15·3, P = 0·023) and EFS (RR = 4·9, 95% CI, 1·7–14·1, P = 0·003). The remaining seven miRNAs were significantly correlated with only one of them. Patients with down-regulated MIRN21, MIRN23A, MIRN27A and MIRN34A expression levels had inferior OS. In contrast, EFS was influenced by low expression levels of MIRN19A (shorter EFS), MIRN195 and MIRNLET7G (longer EFS, respectively). A poor OS was most strongly correlated with decreased expression of MIRN21 (RR = 4·5, 95% CI, 1·4–14·0, P = 0·010) and MIRN27A (RR = 4·6, 95% CI, 1·5–13·6, P = 0·007). In addition to MIRN127, EFS was most strongly influenced by MIRNLET7G and MIRN19A. A reduced expression level of MIRNLET7G was contributed to significantly longer EFS (RR = 0·2, 95%, 0·1–0·6, P = 0·002) whereas a reduced expression level of MIRN19A correlated with significantly shorter EFS (RR = 4·2, 95%, 1·5–11·8, P = 0·005).
Prior studies have shown that a small subset of miRNAs may define tumour entities better than microarray expression data from thousands of mRNAs (Lu et al, 2005). The present study used a qRT-PCR based method to characterize signatures for 157 miRNAs in DLBCL and FL. We found that miRNA expression differentiated both lymphoma entities from normal tissue and DLBCL from FL. In addition, a small number of miRNAs were differentially expressed in DLBCL as well as in FL compared to normal tissue. Furthermore, a number of individual miRNAs were associated with clinicopathological factors. This study formed a basis for developing miRNA expression signatures as diagnostic and prognostic tools for DLBCL and FL and also adds to our understanding of the role of miRNAs in cancerogenesis.
Comparison of miRNA profiles from DLBCL and FL to reactive LN revealed 15 differentially expressed miRNAs in DLBCL and 12 differentially expressed miRNAs in FL. Interestingly, most differentially expressed miRNAs were down-regulated in DLBCL (11 of 15; Fig 1A) whereas the opposite was true for FL, which showed mostly up-regulated miRNAs (nine of 12; Fig 1B). Despite this apparent difference there was a strong overlap of miRNA expression. MIRN210, MIRN155 and MIRN106A were up-regulated in both DLBCL and FL as compared to reactive LN. Conversely, expression of MIRN149 and MIRN139 was reduced in both entities (Fig 1A and B): It is tempting to speculate that these five miRNAs might be of general importance in the development of lymphatic malignancies. In accordance with our findings, MIRN155 was previously reported to play a critical role in B-cell maturation and lymphocyte activation (Rodriquez et al, 2007; Thai et al, 2007). Transgenic mouse studies demonstrated that B-cell targeted expression of MIRN155 leads to the development of B-cell malignancies (Costinean et al, 2006). Further, a number of miRNA profiling studies have shown elevation of MIRN155 in a wide array of cancers including lymphomas (Van den Berg et al, 2003; Eis et al, 2005; Kluiver et al, 2005; Costinean et al, 2006; Volinia et al, 2006; Yanaihara et al, 2006; Lawrie et al, 2007). Recently, a correlation between MIRN155 and NF-κB expression was found in DLBCL cell lines and patients (Rai et al, 2008). Another dysregulated miRNA, MIRN210, has been suggested to play an important role in tumour onset as a key regulator of the cell cycle (Giannakakis et al, 2007). Even less is known about expression of the other listed miRNAs in cancer, making speculation about their role in lymphoma difficult at this time.
Next, we focussed on miRNAs that showed specific differential regulation in the two lymphoma entities analysed here. In comparison of DLBCL- to FL-profiles, ten miRNAs were differentially expressed. Again, DLBCL was characterized mostly via down-regulation of miRNAs (eight of 10; Fig 1C). Resuming these data, and for the comparison performed in this article (Fig 1A–D), we conclude that expression levels of MIRN9/9*, MIRN301, MIRN213, MIRN330 and MIRN338 are characteristic for FL, while MIRN150, MIRN17-5P, MIRN145, MIRN328, MIRN99A, MIRN10A, MIRN95, MIRN151 and MIRNLET7E add up to specific DLBCL signature. Some of these miRNAs have already been found to play an important role in tumour development. For instance, MIRN150– most strongly downregulated in DLBCL – has been described to control B-cell development and is significantly up-regulated in patients with chronic lymphocytic leukaemia (CLL) (Fulci et al, 2007; Xiao et al, 2007; Zhou et al, 2007). This miRNA might have a highly specific role in the development of different lymphatic neoplasias. A significant deregulation of MIRN10A in AML and of MIRN17-5P and as MIRN145 is known in various other cancer types (Bandrés et al, 2006; Volinia et al, 2006; Debernardi et al, 2007; Gramantieri et al, 2007; Iorio et al, 2007; Sempere et al, 2007; Slaby et al, 2007). Regulation of transcription factor E2F1 – a target of proto-oncogene MYC– by MIRN17-5P has already been shown (O’Donnell et al, 2005). Therefore, it seems to be possible that this miRNA has a general impact on tumour development.
It has been described that cases of the ABC subtype of DLBCL show an inferior prognosis as compared to GCB subtype. Lawrie et al (2007) were able to discriminate the GCB- and the non-GCB subgroups of DLBCL using the classifiers MIRN155, MIRN21 and MIRN221. In accordance with these results, MIRN155 was previously reported to be overexpressed in ABC-DLBCL (Eis et al, 2005). Although we used a similar approach to that of Lawrie et al (2007), we could not confirm these results. Given the number of cases of GCB and non-GCB (25 in each group), it seems remarkable that we were not able to select a combination of miRNA profiles that reliably discriminated between these two subtypes (Fig 3A). This, however, may be caused by the variable reproducibility of IHC stains and interpretations. Our list of candidates for differential expression contained MIRN155; however the listing of the other potentially differentially expressed miRNAs comes with a FDR of 20%.
Additional studies are required to improve our knowledge regarding the role of miRNA expression and cancer development to determine the potential of these small RNAs as either biomarkers or therapeutic targets.
This work was supported by the German High-Grade Non-Hodgkin Lymphoma Study Group. We thank Biggi Branke for the skilled and dedicated technical assistance.