Explorative data analysis of MCL reveals gene expression networks implicated in survival and prognosis supported by explorative CGH analysis

MCL gene expression data (n = 71) were obtained from cDNA arrays containing genes preferentially expressed in lymphoid cells or genes known or presumed to be part of cancer development or immune function ("Lymphochip" microarrays [7]; data have been deposited at NCBI's Gene Expression Omnibus data repository under GEO series accession number GSE10793. We give also the resulting gene expression ratios [see Additional file 5] and the prognosis assigned to patients [see Additional file 6]. The dataset is completed by comparative genomic hybridization (CGH) data for each patient (n = 71). The samples were collected from cyclin D1-positive patients of several hospitals in the "Lymphoma and Leukemia Molecular Profiling Project" (LLMPP) [6].

Most of the statistical analyses were performed using the "Genome Expression Pathway Analysis Tool" (GEPAT). This is a web-based platform for annotation (allowing also extensive re-annotation of the data), analysis and visualization of microarray gene expression data [8] including genomic, proteomic and metabolic features.

The database performs the analyses applying Bioconductor [9], an open source software for the analysis and comprehension of genomic data, based on the R programming language [10].

For identification of differentially expressed genes, GEPAT uses the "limma" package which offers moderate t-statistics [11, 12]. It fits linear models on the gene expression values of each gene with respect to the groups which are compared. After that empirical Bayes shrinkage of the standard errors is performed. Due to its robustness the method can be applied to experiments with a small number of samples. To correct for multiple testing it offers three options, we chose the method by Benjamini and Hochberg [13].

For identifying all protein-protein network interactions GEPAT uses the "Search Tool for the Retrieval of Interacting Genes/Proteins" (STRING) [14]. The STRING database comprises known and predicted protein-protein interactions. The interaction information arises from genomic context, experiments, other databases, coexpression and textmining.

For explorative correspondence analysis and principal component analysis, functions from the R package "Modern Applied Statistics with S" (MASS) was applied [15]. A constrained or canonical correspondence analysis (CCA) [16] was performed using the vegan package [17].

The Wilcoxon rank sum test [18], a non-parametric statistical test, was applied to the CGH data. It tests here each of the chosen bands against the null hypothesis that there is no statistically significant difference between our proposed two MCL patients "b" and "s". The R package "survival" is used to calculate all Cox regression hazard models [19, 20]. It examines the correlation between the given measurements and the survival data. For the exploratory analysis of the CGH data as well as for the new predictor of MCL overall survival, we used the Wald test to determine the significance of the association between the model and the outcome.

The survival time itself is the most obvious and biological meaningful parameter in which subgroups should show a big difference for realizing individual clinical treatment. We selected 3.000 genes with the highest variance and applied correspondence analysis (Figure 1). We found (71 MCL patients) that already the second axis separated almost perfectly the longer and the shorter living patients above and below the median of survival. Furthermore, this coincides well with the median of the proliferation signature [6] values in a multidimensional data space (see Methods). This finding was re-examined by exploratory data analysis of the genes of the proliferation signature and a huge amount of further genes. We ranked a total of 71 MCL patients according to their proliferation signature values and separated them according to the median. We define two groups – "s" for small and "b" for big proliferation signature with big difference in the survival time. Patients with a high proliferation signature value live shorter on average, than patients with a low proliferation signature value.

Bild vergrößern

To each single chromosome of the CGH data exploratory data analysis was applied, correspondence analysis [see Additional file 1] and principal component analysis (Figure 2). Both methods are useful for exploring information and structures in data in order to get a first and unbiased impression. Principal components analysis reduces multidimensional data sets to lower dimensions for analysis. Correspondence analysis works similarly, but scales the data, such that both rows and columns can be visualized in one plot. Results show a strong correlation for four bands of chromosome 9, 9p24, 9p23, 9p22 and 9p21 and above median ("s") or below median patient survival ("b").

Bild vergrößern

In the correspondence analysis plot [see Additional file 1], the four bands mentioned before attract most patients of the subgroup "b" and the 1st factor axis separates almost completely the two groups. Bands 9q33 and 9q34, are located relatively far away from the remaining ones. In Figure 2 the second principal component groups almost all the "b" – patients near the four bands 9p24, 9p23, 9p22, 9p21 with vectors of similar length and similar direction. The vectors of 9q33 and 9q34 include along their lengths almost all "s" samples. These results indicate that these six bands of chromosome 9 correlate with good and bad survival between patients. The principal component 1 is an interesting main component, carrying 51% of the variance, but non-trivial to link to a known phenotype (we investigated different possibilities including sex differences, cancer sub-types, patient accrual and correlation with different gene signatures).

Further exploratory data analysis was performed to merge the survival time and the CGH data by the Cox regression hazard model. A univariate Cox regression hazard model was performed on all available bands of the CGH data of all 71 patients. The mentioned four bands of chromosome 9 delivered amongst others the most significant results. The resulting bands are "9p24", "9p23", "9p22", "9p21", "9q31" and "9q32". These comprise the first four bands found on chromosome 9 by the analyses before.

Exploratory analysis pointed to differences between longer and shorter living MCL patients, but rather than forming two distinct subgroups, the patients constitute a coherent continuum. Therefore, the results of the exploratory analysis above were not additionally confirmed by classification tools. However, the differences in gene expression above and below the median of survival correlate well with different gene signatures identified before (proliferation signature) as well as with the new ones described in our study (non-proliferative signatures, see below). To improve survival predictions we further searched with univariate Cox regression hazard analysis for highly significant genes, which correlate strongly with the overall survival time. The cox regression was applied to all data points. However, the first 50 MCL samples served as training set for classification by gene signatures and the remaining data (21 patients) for validation. The idea was here to have a large training data set, but still keep a third of the available data for validation.

A four gene predictor with the genes CDC2, ASPM, tubulin-α and CENP-F reported in [6] could not be tested, as after reannotation by GEPAT [8], mapping of CENP-F seemed uncertain. Predictors with 4, 5 or 6 genes delivered not the same predictive power as the proliferation signature [6] (data not shown). The prediction power was calculated from the correct classification and misclassification for patients over or below the median of survival for 69 patients (the two patients with the median value were excluded).

However, we identified a set of seven genes delivering similar good prognosis separation. It includes (i) the well known key cell cycle kinase CDC2 [21, 22], (ii) the "cell division cycle 20 homolog" (CDC20) required for anaphase and chromosome separation [23] and (iii) the salvage pathway gene HPRT1 (hypoxanthine phosphoribosyltransferase 1), three genes from the 20 genes proliferation signature of Rosenwald [6]. We get improved prediction power including four additional genes (Table 1): (i) centromere protein E (CENPE), a kinesin-like motor protein; it accumulates during G2 phase of cell cycle for chromosome movement or spindle elongation [24]. (ii) BIRC5 (baculoviral IAP repeat-containing 5 gene), an inhibitor of apoptosis (IAP gene family) is expressed in most tumours and in lymphoma [25], participates in the spindle checkpoint and associates with AURKB [26]. (iii) ASPM (abnormal spindle homolog) is essential for normal mitotic spindle function [27]. (iv) Insulin-like growth factor 2 mRNA binding protein 3 (IGF2BP3), is found in the nucleolus, is over-expressed in human tumours and represses IGF2 during late development [28‐30].

The seven genes were used to calculate a multivariate Cox regression hazard model and with its coefficients, a gene expression based survival estimator separated all 71 patients into two subgroups. Two patients had exactly the median of survival and were excluded in this comparison, 56 agreed with the classification according to the gene signatures, 13 did not. Compared to the proliferation [6] signature's ability to distinguish patients with good and bad survival prognosis (Figure 3), the seven gene predictor does it similarly well (Figure 4). The correlation between this classification and the "s" and "b" groups of the proliferation signature is overall about 0.62 and in our validation set (patients 51 – 71) it is 0.81.

Bild vergrößern

A correspondence analysis of the 3.000 genes with the highest variance showed clear clustering of patients with good or bad prognosis, respectively (Figure 1). Using proliferation signature [6] (Figure 3), samples show a little overlap, but are again separated clearly.

Bild vergrößern

Taken together, these results show that the seven gene predictor is able to distinguish patient prognosis as well as the complete proliferation signature, but with less effort.

We found a dense regulatory network of interacting genes correlated with prognosis. Applying a moderate t-test, the well known cell division cycle 2 gene (CDC2/CDK1) for G1 to S and G2 to M transition [31, 32] shows the most significant difference between the longer living "s" and the shorter living "b" patients (Table 2). Furthermore, its interaction partners according to HPRD database [33] show a significant up or down regulation comparing good and bad surviving patients (Figure 5), e.g. WEE1 and CDC25. Moreover, aurora kinases A, B [34] and BUB1 kinase activating the spindle checkpoint [35], are differently regulated between shorter and longer living patients. However, there are further genes involved in this network of directly interacting genes differently regulated in good or bad prognosis patients (Figure 5; Figure 6) such as (i) "proliferating cell nuclear antigen" (PCNA), a cofactor of DNA polymerase delta, helps to increase the processivity of leading strand synthesis during DNA replication in group "b". Because of its ability to interact with multiple partners, it is involved in Okazaki fragment processing, DNA repair, translation, DNA synthesis, DNA methylation, chromatin remodelling and cell cycle regulation [36]. (ii) E2F transcription factor 1 (E2F1), this protein can mediate both cell proliferation and p53-dependent/independent apoptosis [37]. It is lower expressed in group "s". (iii) Nucleolin is an abundant multifunctional phosphoprotein of proliferating and cancerous cells [38‐41] and highly expressed in "b".

Bild vergrößern

Interaction partners of CCND1 are also significantly differently expressed (Figure 7): CCND1 and CDK4 are assumed to be involved in cell cycle progression of MCL, MYC is suspected of increasing MCL's proliferation rate. FOS, JUN and MYBL2 are partly known to play a role in cancer, but not explicitly in MCL. FOS ("v-fos FBJ murine osteosarcoma viral oncogene homolog") and JUN ("jun oncogene") are weakly downregulated in "b". Other interaction partners such as MYC ("V-myc myelocytomatosis viral oncogene homolog (avian)"), MYBL2 ("V-myb myeloblastosis viral oncogene homolog (avian)-like 2"), CDK4 ("Cyclin-dependent kinase 4") and CDK6 show higher gene expression values in bad prognosis patients below the median of survival.

Bild vergrößern

Moreover, there are some genes with similar significance and expression difference, associated with other functions (Table 3). Most of them are associated with DNA metabolism. Three of them, "suppressor of cytokine signaling 1" (SOCS1), "tubulin, alpha 1b" (TUBA1B), and "CCAAT/enhancer binding protein (C/EBP), beta" (CEBPB) are mentioned here. CEBPB, is a transcription factor. It plays an important role in immune and inflammatory responses [42]. Additionally it can stimulate the expression of the collagen type I gene. TUBA1B encodes for an important part of the microtubules. SOCS1 is a member of cytokine-inducible inhibitors of signaling [43] and inhibits protein kinase activity.

We applied the Wilcoxon rank sum test on the CGH data and compared the patients with good "s" and bad prognosis "b" (over and below median of survival). The null hypothesis corresponds to no differences between the two entities. The resulting p-values for every band of chromosome 9 are compared in Figure 8. They show strongly the significance of the first four bands 9p24, 9p23, 9p22 and 9p21. On these bands are MCL related genes such as "cyclin-dependent kinase inhibitor 2B" (CDKN2B) and "cyclin-dependent kinase inhibitor 2A" (CDKN2A). TP53 mutations are associated with the blastoid variant of MCL and with a worse prognosis. The bands 9q33 and 9q34 are less significant. To visualize this result more clearly we plot the densities of the p-values [see Additional file 2]. A peak in the density indicates significant bands of the Wilcoxon test.

Bild vergrößern

The Wilcoxon rank sum test showed similar results for chromosome 7. Here, the bands 7p21, 7p15, 7p14 are potentially related to the classification of "s" and "b" patients. Now the log p-values and their densities are plotted against the bands in Figure 9, the density plot of p-values for chromosome 7 is also shown [see Additional file 3]. The explorative analyses of chromosome 7 could not show such a clear relation as in chromosome 9.

Bild vergrößern

Specific gene expression differences in patients with good or bad prognosis are well supported by the CGH data of chromosome 9. We checked the location of the signature genes as we wondered if they were on chromosome 7 or 9, however this was not the case. Also the genes of the gene network in Figure 6 are located elsewhere. No result mentioned before could explain the relationship between the subgroups and the subgroup-separating CGH data of chromosome 9. We thus investigated the gene expression data of these bands. Again a moderate t-test was applied to rank genes differentially expressed between "s" and "b". The top five are listed in Table 4, e.g. the "Heat Shock 70 kDa protein 5" and a catalytic subunit of "Protein Phosphatase 6". Several of their functions implicate them to be critical in cancer development. Their genomic position revealed a quite remarkable clustering of these genes [see Additional file 4]. Three of the genes seem to be located very closely to each other. The "heat shock 70 kDa protein 5" (HSPA5), also referred to as 'immunoglobulin heavy chain-binding protein' (BiP) targets misfolded proteins for degradation, and has an anti-apoptotic property. It is induced in a wide variety of cancer cells and cancer biopsy tissues and contributes to tumor growth and confers drug resistance to cancer cells [44]. The PPP6C gene encodes for a catalytic subunit of the Ser/Thr phosphatases, the "protein phosphatase 6 catalytic subunit" [45]. The pre-B-cell leukemia transcription factor 3 (PBX3) shows extensive homology to PBX1, a human homeobox gene involved in t(1;19) translocation in acute pre-B-cell leukemia. But in contrast to PBX1 the expression of PBX3 is not restricted to particular states of differentiation or development [46]. It is also known that if HoxB8, a homeobox gene identified as a cause of leukemia, binds to the Pbx cofactors it blocks differentiation in certain cell types [47]. "Prostaglandin-endoperoxide synthase 1" (PTGS1) is the key enzyme in prostaglandin biosynthesis, and is also known to play a role in the human colon cancer [48, 49]. The expression of the alternative splice variants is differentially regulated by cytokines and growth factors [50‐52]. Very little is known about "quiescin Q6-like 1" (QSCN6L1), except its major role in regulating the sensitization of neuroblastoma cells for IFN-gamma-induced apoptosis [53]. A similar clear clustering as on chromosome 9 could not be detected on chromosome 7.

Bild vergrößern