Comparison of genome profiles for identification of distinct subgroups of diffuse large B-cell lymphoma

Diffuse large B-cell lymphoma (DLBCL) is the most common type of non-Hodgkin lymphoma and is known to include pathophysiologically heterogeneous groups.^1-4  DLBCL is also known to be clinically heterogeneous because patients with DLBCL show markedly different clinical courses.⁵  This has drawn attention to the importance of identifying subgroups in heterogeneous DLBCL.

Gene-expression profiling conducted by Alizadeh et al⁶  identified 2 molecularly distinct forms of DLBCL with gene expression patterns indicative of the different stages of B-cell differentiation, that is, activated B cell-like (ABC) and germinal center B-cell-like (GCB) types.^6,7  The ABC group expresses genes characteristic of activated B cells and plasma cells, whereas the GCB group maintains the gene expression program of normal germinal center B cells.^6-8  Those authors also reported that the overall survival of the ABC group was significantly worse than that of the GCB group.

We reported the identification of 3 phenotypically distinct subgroups of DLBCL, CD5⁺, CD5^-CD10⁺, and CD5^-CD10^- DLBCLs.⁹  The CD5⁺ group, found to account for approximately 10% of all DLBCL, has the CD5⁺C10^+-CD19⁺CD20^-CD21^-CD23^-cyclinD1^- phenotype and is characterized by poorer prognosis, frequent extranodal sites, poorer performance status, and higher lactase dehydrogenase levels compared with CD5^- DLBCL.¹⁰  The CD5^-CD10⁺ group shows less-frequent BCL2 protein expression than the other groups, indicating the presence of a definite relationship with normal germinal center cells that usually lack BCL2 expression. Finally, the CD5^-CD10^- group is the most common group and has a higher incidence of BCL6 gene rearrangement than the other 2 groups, although the difference is insignificant.⁹ 

Despite each subgroup of DLBCL being molecularly or phenotypically distinct, the genetic characteristics and their relationship have not been sufficiently studied. Here, we made use of the array-based comparative genomic hybridization (array CGH) to identify genomic imbalances characteristic of the distinct subgroups of DLBCL.^11,12  Gene-expression profiling was also used to clarify the relationship between the ABC/GCB and CD5⁺/CD5^-CD10⁺ subgroups.

Lymph node samples and clinical data were obtained from 99 patients with the protocol approved by the Institutional Review Board of Aichi Cancer Center (36 cases of CD5⁺, 19 cases of CD5^-CD10⁺, and 44 cases of CD5^-CD10^-). Three CD5⁺ cases possessed CD10⁺. Informed consent was provided according to the Declaration of Helsinki. It should be noted that we especially collected a total of 36 cases of CD5⁺ DLBCL, which represent a higher ratio than that found in the standard population, for evaluation of the genetic status of this disease entity. We analyzed genomic imbalances for the 99 cases of DLBCL by use of array CGH. Forty-six of these cases (22 cases of CD5⁺, 7 cases of CD5^-CD10⁺, and 17 cases of CD5^-CD10^-) were subsequently subjected to gene-expression profiling. In addition, none of the patients had a previous history of malignant lymphoma. All the DNA and RNA samples were obtained from tumors at the time of diagnosis before any treatment was administered, while all the patients with DLBCL received adequate doses of a cyclophosphamide, Adriamycin, vincristine, and prednisone (CHOP)-like regimen after diagnosis. All patients' data regarding age, stage, performance status, lactate dehydrogenase (LDH) level, number of extranodal sites, and International Prognostic Index (IPI) score at diagnosis are shown in Table 1.

Total RNA was isolated from each specimen by means of cesium chloride centrifugation. Cyanine 5 (Cy5)- or Cy3-labeled complementary RNA (cRNA) was generated from total RNA by using a Low RNA Input Linear Amplification Kit (Agilent Technologies, Palo Alto, CA). Probes consisted of a mixture of an experimental Cy5-labeled cRNA and control Cy3-labeled cRNA. The latter was prepared from a pool of total RNA from 10 hyperplasia lymph node samples. The microarray glass slide consisted of an Agilent oligonucleotide array custom-made for the Cancer Institute of the Japanese Foundation for Cancer Research, on which a total of 21 619 genes were spotted (Agilent Technologies). Probes were hybridized overnight on the glass slides with an In Situ Hybridization Kit Plus (Agilent Technologies) according to the manufacturer's protocol. Fluorescent images of hybridized microarrays were obtained with an Agilent scanner G2565AA (Agilent Technologies), which were then analyzed with Feature Extraction software (Agilent Technologies) to obtain the ratios of the fluorescence of the experimental Cy5-labeled samples to that of the Cy3-labeled control. All nonflagged fluorescence ratios were log-transformed (base 2) and centered by subtracting the median observed value of each gene determined by cluster analysis. The hierarchical clustering algorithm was applied to DLBCL cases according to the expression level of these genes with the aid of Cluster and TreeView software (Eisen Lab, http://rana.lbl.gov/EisenSoftware.htm).¹³  Of the 100 genes specified by Rosenwald et al⁸  for clustering ABC and GCB groups, 67 of which were available to use for clustering analysis. These 67 genes are BARD1, PIK3CG, LRMP, Hs.1098, BCL6, HDAC1, MYBL2, MME (CD10), STAG3, LMO2, APS, Hs.151051, ADPRT, ITPKB, REL, FLJ20094, Hs.211563, MEF2B, CD44, Hs.75765, IL6, PTPN2, PTPN12, BMI1, Hs.128003, BACH2, HIVEP1, CFLAR, APAF1, RYK, EDG1, KIAA0874, Hs.153649, MADH4, PTPN1, Hs.93213, DCTD, Hs.193857, IL16, SP140, SH3BP5, IRF4 (MUM1), TLK1, KCNA3, TCL1A, PAK1, Hs.188, CXCR4, SLA, CCND2, TGFBR2, ETV6, SPAP1, PM5, PDIR, IGHM, CD22, Hs.296938, Hs.1565, Hs.83126, MAPKAPK3, RUNX1, Hs.55947, S100A4, TFAP4, IRF2, and OPA1. We performed clustering analysis with published microarray data by these 67 genes. DLBCL gene expression profile data generated by the Lymphochip microarrays were obtained from supplemental data of the article by Rosenwald et al.⁸  We confirmed that the 274 DLBCL of the Lymphochip microarray data set could be divided into the ABC and the GCB and the Type 3 with these 67 genes. Distributions of tumors with aforementioned 67 genes were nearly identical to those with the 100 genes that were described by Rosenwald et al.⁸ 

Array CGH was performed for DLBCL cases by previously described methods using custom-made glass slide of Aichi Cancer Center (ACC) array slide version 4.0. The array consisted of 2304 BAC (Bacterial artificial chromosome) and PAC (P-1 derived artificial chromosome) clones (BAC/PAC clones), covering the whole human genome with roughly 1.3 Mb (megabase) of resolution. BAC clones were derived from RP11 and RP13 libraries, and PAC clones were derived from RP1, RP3, RP4, and RP5 libraries. BAC/PAC clones were subjected to degenerate oligonucleotide-primed polymerase chain reaction (PCR). The resulting DNA samples were robotically spotted by an inkjet technique (NGK, Nagoya, Japan) in duplicate onto CodeLink activated slides (Amersham Biosciences, Piscataway, NJ). BAC/PAC clones used were selected based on information from the National Center for Biotechnology Information (NIBC; http://www.ncbi.nlm.nih.gov/) and Ensembl Genome Data Resources (http://www.ensembl.org/). These clones were obtained from the BACPAC Resource Center at the Children's Hospital (Oakland Research Institute, Oakland, CA). DNA preparation, labeling, array fabrication, and hybridization were performed as described previously.^11,12 

For the array, 10 simultaneous hybridizations of healthy male versus normal male were performed to define the normal variation for the log₂ ratio. A total of 91 clones with less than 10% of the mean fluorescence intensity of all the clones, with the most extreme average test over reference ratio deviations from 1.0 and with the largest SD in this set of normal controls was excluded from further analyses. Thus, we analyzed a total of 2213 clones (covered 2988 Mb, 1.3 Mb of resolution) for further analysis. Of the 2213 clones, 2158 (covered 2834 Mb) were from chromosome 1p telomere to 22q telomere; 55 of 2213 clones were from chromosome X.

Because greater than 96% of the measured fluorescence log₂ ratio values of each spot (2 × 2191 clones) ranged from +0.2 to -0.2, the thresholds for the log₂ ratio of gains and losses were set at the log₂ ratio of +0.2 and -0.2, respectively. Regions of low-level gain were defined as log₂ ratio +0.2 to +1.0, those suggested of containing a heterozygous loss/deletion as log₂ ratio -1.0 to -0.2, those showing high-level gain as log₂ ratio greater than +1.0, and those suggested of containing a homozygous losses/deletion as log₂ ratio less than -1.0. We defined region of gain or loss as (1) continuously ordered 3 clones showing gain or loss or as (2) single clones showing recurrent high copy number gain (log₂ ratio >+1.0) or homozygous loss (log₂ ratio <-1.0).^11,12 

Regions of high-level gain and regions of homozygous loss/deletion were also easily detected, as were regions showing low-level gain and those of heterozygous loss/deletion.

Statistical analysis of overall survival of distinct subgroups of DLBCL was conducted by log-rank test. P less than .05 was taken to show a significant difference.

Statistical analysis for array CGH. To analyze genomic regions for statistically significant differences between the 2 patient groups (eg, ABC and GCB), a data set was constructed by defining genomic alterations as copy number gains for log₂ ratio thresholds of +0.2 or greater, and as copy number losses for thresholds of -0.2 or less. Clones showing a gain (log₂ ratio ≥ +0.2) were inputted as “1” versus no-gain clones (log₂ ratio <+0.2) as “0” on an Excel (Microsoft, Redman, WA) template for each case. Similarly, loss clones (log₂ ratio ≤ -0.2) were inputted as “1” versus no-loss clones (log₂ ratio >-0.2) as “0” on another Excel template for each case. Data analyses were then carried out for the following purposes: (1) comparison between the 2 groups (eg, ABC and GCB) of frequencies of gain or loss for each single clone, and (2) comparison of overall survival between cases showing gain or loss of a single clone and cases without either gain or loss. Fisher exact test for probability was used for the former comparison, and a log-rank test for comparing survival curves for the 2 groups was used for the latter.

Statistical analysis for gene-expression profiling. The Mann-Whitney U test was performed for detecting significant differences in expression levels of p16^INK4a between the ABC and GCB groups. All the statistical analyses were conducted with the STATA version 8 statistical package (StataCorp, College Station, TX).

The variety of DLBCLs can be characterized and classified by gene-expression profiling and cell-surface phenotyping. Clinically, ABC DLBCL behaves more aggressively than GCB DLBCL. The survival of CD5⁺ DLBCL cases is shorter, and CD10⁺ DLBCL is relatively indolent (Figure 1).

To determine whether DLBCL with CD5 and/or CD10 markers are related to the ABC and GCB subgroups, we subjected a total of 46 DLBCL cases (22 cases of CD5⁺, 7 cases of CD5^-CD10⁺, and 17 cases of CD5^-CD10^-) to gene-expression profiling. The results showed that the 46 cases could be clearly assigned to either the ABC or GCB groups (Figure 2) and, of special importance, that the CD5⁺ and CD5^-CD10⁺ phenotypes were closely related to the ABC and GCB subgroups, respectively (Table 2). Of the 22 CD5⁺ DLBCL cases, 19 showed the ABC signature, whereas only 3 were characterized by the GCB signature (2 cases of CD5⁺CD10⁺ and 1 case of CD5⁺CD10^-) (P = .001). In sharp contrast, all 7 cases of CD5^-CD10⁺ DLBCL showed the GCB signature (P = .003). CD5^-CD10^- DLBCL cases showed mixed results, expressing either the ABC (9 cases) or GCB (8 cases) signature (P = .533), indicating that it is a heterogeneous entity.

Because it was previously demonstrated that ABC and GCB DLBCLs are molecularly distinct subgroups,^6-8  we initially compared the genomic profiles of these subgroups. Two typical individual genomic profiles of 2 CD5^-CD10^- cases, 1 with an ABC signature and the other with a GCB signature, are shown in Figure 3.

Frequent genomic imbalances (copy number changes) of the ABC group (≥ 6 cases) were gain of chromosome 3, 8q21-q26, 11q21-q25, 16p11-p13, 16q22-q24, 18, 19q13 and X, and loss of 2p11 (Igk locus), 6q12-q27, 8p22-p23, 9p21, and 17p. Frequent genomic imbalances of the GCB group (≥ 4 cases) were gain of 1q22-q32, 2p14-p24, 5p12-p15, 5q15-q31, 6p12-p25, 7, 8q22-q26, 9q33-q34, 11q, 12, 13q31-q33, 16p11-p13, 18q21-q23, 19p, 19q13, 21q, and X and loss of 1p36, 2p11, 3p14, 4p12-p13, 4q33-q34, 6q14-q16, 8p22-p23, 9p21, 13q12-q22, 17p12, and 18q22-q23. Here, frequent genomic gains and losses were defined as greater than 20% for either group.

The ABC group was genomically characterized by more frequent gains of 3p23-q28, 18q11.2-q23, and 19q13.41-q13.43 and loss of 6q22.31-q24.1 and 9p21.3. The GCB group was genomically characterized by more frequent gains of 1q21.1-q23.3, 1q31.1-q42.13, 2p15-p16.1, 7q22.1-q36.2, and 12q13.1-q14 (Fisher exact test, P < .05). Ideogram of the genomic imbalance of ABC and GCB DLBCLs are presented in Figure 4A-B, and the genome-wide frequency representing the genomic imbalance of ABC and GCB DLBCL are presented in Figure 5A.

We found that the genomic imbalance patterns of ABC and GCB DLBCLs are distinctly different. For instance, gain of chromosome 3q23-q28 was observed in 25% to 36% of the ABC group but not in the GCB group (0%), whereas gain of 7q22-q36 was observed in 50% to 61% of the GCB group and far less (< 5%) in the ABC group.

It should be noted that the expression of CD5 had no effect on the genomic imbalance observed in the ABC and GCB groups. The genomic imbalance detected here in the ABC group reflects the dominance of CD5-classified cases: 67% (19 of 28 cases) of the ABC group was of the CD5⁺ type, but the frequency and region of the genomic imbalance of CD5⁺ and CD5^- within the ABC group were similar. There were no significant differences in either region or frequency of the genomic imbalance between CD5⁺ and CD5^- within the ABC group (data not shown). A typical example was loss of 9p21. This loss was specific to the ABC group where, within the ABC group, 13 (68%) of the 19 CD5⁺ DLBCL cases and 6 (66%) of the 9 CD5^- DLBCL cases showed this loss (P = 0.999).

We examined the frequency of gain and loss regions for the CD5⁺ (36 cases), CD5^-CD10⁺ (19 cases), and CD5^-CD10^- (44 cases) groups. Frequent genomic imbalances (≥ 8 cases) in the CD5⁺ group were gain of chromosome 3, 6p22-p25, 7p22-q31, 8q24, 11q22-q25, 12, 16p13-q21, 18, 19, and X and loss of 1p36, 2p11, 6q14-q27, 8p23, 9p21, 15q13-q14, and 17p11-p13. Although gain of 1q21-q32, 7p22-q36, and 12 were characteristic of the GCB group, these gains were also found in 20% or less of the CD5⁺ group.

Frequent genomic imbalances (≥ 4 cases) in the CD5^-CD10⁺ group were gain of 1q, 2p13-p25, 6p21-p25, 7, 8q22-q24, 9q33-q34, 12, 13q31-q33, 15q, 16p13, 19q13.3-q13.4, and X and loss of 1p36, 1p22, 2p11, 3p14, 4p, 6q13-q27, 9p21, and 13q14-q21.

Comparison of the CD5⁺ and CD5^-CD10⁺ subgroups showed that CD5⁺ DLBCL had more frequent gains at chromosome 3 and loss of 9p21 compared with CD5^-CD10⁺ DLBCL, whereas CD5^-CD10⁺ DLBCL showed more frequent gains of 7q22-q36, 12q13-q14, and 17p13 compared with CD5⁺ DLBCL (Table 3). The genome-wide frequency of the genomic imbalance of CD5⁺ and CD5^-CD10⁺ DLBCLs are shown in Figure 5B. Of special importance is that these characteristic genomic profiles of CD5⁺ and CD5^-CD10⁺ DLBCLs are quite similar to those of ABC and GCB DLBCLs, respectively (Figure 5A-B). There were no significant differences in either region or frequency of genomic imbalance between either the CD5⁺ and ABC groups or the CD5^-CD10⁺ and GCB groups (Table 3).

Frequent genomic imbalance of the CD5^-CD10^- group (≥ 9 cases) was gain of 1q21-q42, 3q21-q29, 5p13-p15, 6p21-p25, 7, 11q22, 12q13-q14, 18, 19q13.1-q13.4, and X and loss of 2p11, 3p14.2, 6q12-q27, 9p21.3, 15q14-q15, and 17p11-p13. The genomic imbalance of CD5^-CD10^- DLBCL reflected a mixed type because all of these gains and losses could be observed in both the CD5⁺ and CD5^-CD10⁺ groups. Furthermore, the frequency and region of the genomic imbalance of CD5^-CD10^- DLBCL was very similar to that of ABC plus GCB, as shown in Figure 5. There were no significant differences in either region or frequency of the genomic imbalance between the CD5^-CD10⁺ and ABC plus GCB groups (data not shown). These findings correlate well with the results of gene-expression profiling that showed CD5^-CD10^- DLBCL was evenly distributed in the ABC and GCB groups.

Finally, we tried to identify prognostic variables detected by the array CGH and found that loss of 9p21 had a deleterious effect on patient survival: 37 cases with loss of 9p21 showed significantly poorer survival than the 59 cases without this loss (log-rank test, P = 0.001) (Figure 6A).

Loss of 9p21 was significantly more frequently detected in the ABC group (19 cases) than in the GCB group (5 cases) (Fisher exact test, P = .014). The survival of ABC cases with loss of 9p21 was significantly inferior to that of the ABC cases without such a loss (log-rank test, P = .013), whereas loss at the corresponding region in the GCB group did not affect survival. Similarly, the survival of CD5⁺ cases with loss of 9p21 was significantly poorer than cases without this loss (log-rank test, P = .004). Thus we were able to show that loss of 9p21 (p16^INK4a locus) marks the most aggressive form of DLBCL. Among the CD5^-CD10^- DLBCL cases, loss of 9p21 tended to have a negative effect on survival, although this did not reach statistical significance (log-rank test, P = .068).

As shown in Figure 6B, the minimum common region of 9p loss was located within 2.2 Mb at 9p21.3. Evidence suggesting homozygous losses of 9p21 was found in 6 cases (defined as log₂ ratio <-1.0; 3 cases each of CD5⁺ and CD5^-CD10^-), whereas all GCB or CD5^-CD10⁺ cases failed to show any signs, suggesting homozygous loss at 9p21.3. Thirteen of the 37 cases with loss of 9p21 showed loss at a restricted position of the genome encompassing a single BAC, RP11-149I2, which contains the p16^INK4a tumor suppressor. The expression level of p16^INK4a for the ABC group was significantly lower than that of the GCB group (Figure 6C) (Mann-Whitney U test, P = .001). These results agree well with the finding of a higher frequency of 9p21.3 loss in ABC DLBCL cases.

Several researchers have reported genomic alterations in DLBCL detected by means of conventional CGH or array CGH.^14-18  However, only a few comparative genome analyses of DLBCL subtypes have been conducted.^12,19  In the study presented here, our array CGH enabled us to detect distinct differences in the genomic imbalance patterns of the ABC and GCB subgroups. ABC DLBCL is genomically characterized by gain of 3q, 18q, and 19q and loss of 6q and 9p21, whereas GCB DLBCL is genomically characterized by gain of 1q, 2p, 7q, and 12q. These results thus provide evidence that the ABC and GCB groups are genetically distinct to each other, suggesting that ABC and GCB DLBCL develop tumors via distinct genomic pathways.

Four groups have published reports on the genomic imbalance of DLBCL transformed from follicular lymphoma (FL).^18,20-22  According to their reports, DLBCL transformed from FL is characterized by a genomic imbalance consisting of gain of 2p, 7p, 12p, and 12q and loss of 4q and 13q. Because the characteristic genomic imbalance of DLBCL transformed from FL is similar to that of the GCB group, DLBCL transformed from FL and GCB DLBCL may share certain steps in their genomic aberration program through the development of lymphomagenesis.

We previously reported that CD5⁺ and CD10⁺ DLBCL constitute clinically relevant subtypes. In the study presented here, we report for the first time that CD5⁺ DLBCL is characterized by ABC expression and unique genomic patterns. Recently, Katzenberger et al²³  conducted a cytogenetic and loss of heterozygosity study of de novo CD5⁺ DLBCL and speculated that CD5⁺ DLBCL was likely to originate from the same progenitor cells as B-chronic lymphocytic leukemia (CLL) because the former showed frequent deletions of the D13S25 locus as well as of the p16^INK4a tumor suppressor.²³  However, CD5⁺ DLBCL appears to be different both in terms of its origin and pathway from CLL and mantle cell lymphoma (MCL), which also express CD5. CLL and MCL are both characterized by loss of 6q, 9p21, 11q22-q23, and 13q14-q21 (seen in 30%-50% of cases),^24-27  whereas less than 10% of CD5⁺ DLBCL cases examined by us showed loss of 11q22-q23, and 13q14-q21. Loss of only p16^INK4a appears to be a common characteristic.

We were also able to demonstrate that CD10⁺ DLBCL is characterized by GCB expression and unique genomic patterns. We previously reported that CD10⁺ DLBCL might originate from germinal center progenitor cells because cells with a normal follicular center possess CD5^- and CD10⁺ immunophenotypes and rarely express BCL2. Huang et al²⁸  used gene-expression profiling to demonstrate that CD10⁺ DLBCL is characterized by a GCB signature, as is also evident from our results.^28,29 

We noted that CD5^-CD10^- DLBCL exhibited a mixed genomic imbalance pattern with respect to 2 subtypes (CD5⁺ and CD5^-CD10⁺). This correlates well with the results of gene-expression profiling, which showed that CD5^-CD10^- DLBCL was evenly distributed within the ABC or GCB groups, suggesting that CD5^-CD10^- DLBCL is a genetically heterogeneous entity.

Loss of 9p21 (p16^INK4a) may mark the most aggressive cases. Of special interest in our series was that ABC and CD5⁺ DLBCL cases with loss of 9p21 showed poorer outcomes than cases without this loss. Loss of 9p21 may therefore represent a unique feature, reflecting the most aggressive form of DLBCL. Deletion of 9p21.3 (p16^INK4a locus) is frequently found in aggressive lymphoma and acute lymphoblastic leukemia and less often in low-grade lymphoma.^30-35  CD5⁺ DLBCL is also closely associated with many aggressive clinical features or parameters, and loss of 9p21 in conjunction with inactivation of p16^INK4a may well be a feature of CD5⁺ DLBCL. Indeed, frequent deletion of 9p21 in CD5⁺ DLBCL has been reported by us and another group.^12,23 

To summarize, we were able to demonstrate that ABC and GCB DLBCLs are distinct in terms of gene expression and genomic imbalance. Most of the CD5⁺ and CD5^-CD10⁺ DLBCLs are included in the ABC and GCB groups, respectively. Furthermore, when searching for genomic imbalances that affect patient prognosis, we found that loss of 9p21 (p16^INK4a locus) marks the most aggressive form of DLBCL. As demonstrated with DLBCL, the combined use of gene-expression profiling and array CGH may facilitate better understanding of heterogeneous tumors in general.

We thank Mss H. Suzuki and Y. Kasugai for their outstanding technical assistance. We also thank Dr Ryuzo Ohno, president of the Aichi Cancer Center, for his support.