Despite the established nature of the cell of origin classification a comparative analysis of gene expression across multiple data sets classified with the same algorithm has not been published. To allow such an analysis we have developed a robust classifier based on 20 genes, described in the original Wright classifier implementation 
, and represented on multiple platform types. The use of a small number of classifier genes is an important feature since primary classifier genes show intrinsic bias and need to be disregarded for downstream analyses. Since the inception of the cell of origin classification, variations have been described that employ larger numbers of classifier genes 
. However, to our knowledge, there has been no direct comparison of whether large numbers of classifier genes produce consistent improvement in classification. Here we have directly addressed this question and using a range of different classifier gene choices we find no consistent benefit of using larger numbers of classifier genes. Indeed in several instances classifiers relying on greater numbers of classifier genes generate ABC- and GCB-DLBCL classes with less significant survival separation.
An issue of primary importance in the evaluation of a test is the performance against a “gold standard” and the choice of metric that is used to assess performance. The most significant clinical feature of the cell of origin classification is the ability to separate DLBCL into two major subgroups ABC and GCB with different survival 
. In assessing the performance of classifiers we therefore used survival separation as the metric. The “gold standard” can be seen either as the LPS classifier described by Wright et al. 
, or the classes assigned by subsequent extension of this classifier to include more classifier genes 
. While the LPS classifier effectively separates multiple data sets from fresh frozen material into classes with significant survival separation, classification choices made by machine-learning tools using 20 classifier genes generate more significant differences in survival separation than those made by LPS or the published class assignments for 4/5 data sets (considering the CHOP and R-CHOP components of GSE10846 separately), the original data set of Monti et al. being the exception 
. Furthermore our selected classifier can distinguish prognostic groups in an FFPE derived data set generated on an Illumina platform 
, while the LPS classifier was much less effective in this data set. Another FFPE derived data set included in this analysis, GSE31312, was generated on Affymetrix HG-U133 Plus 2.0 GeneChips using FFPE derived samples 
. This was also readily classified by DAC and at the level of segregation of gene expression was amongst the most concordant with GSE10846.
An important observation emerging from this study is the existence of an extended molecular gray zone, representing cases whose classification is sensitive to the type of classifier implementation used. A substantial group of cases in each data set was equivalently classified by most classifier implementations and thus had a consistent class. In contrast the differences in outcome separation observed were attributable to cases that had more marginal expression of classifier genes and moved been class in a fashion dependent on classifier implementation. While a “molecular gray zone” was inherent in the cell of origin classification, and in its original form encompassed the Type-III or unclassifiable cases 
, when considering the choices made by different classifier implementations the extent of this molecular gray zone was greater. The concept of cases that do not fall neatly into one or other diagnostic group is familiar and indeed is encompassed in distinct categories of the WHO lymphoma classification 
. It is therefore no surprise that gene expression profiling is similarly subject to ambiguity, and that choice of weighting for individual genes results in differential classification of some cases. Such ambiguity may be resolved in future by using more discrete variables, for example the presence or absence of particular pathway mutations. Nonetheless gene expression based classification schemes are likely to continue to provide additional important information, since they can assess the combined impact of multiple molecular abnormalities acting within a tumour cell population to drive a predominant phenotype. An analysis such as that performed in this work can provide the basis on which to select between individual classifier implementations. In making this selection a metric needed to be chosen. We opted for the difference in overall survival between the principal cell of origin classes assigned by each classifier, while ensuring that this did not come at the expense of greater inclusion into the Type-III/unclassified category. We used the consistency of class-associated gene expression as a supporting feature. As illustrated in this work it is possible to improve on classification choices using different implementations, while remaining within an existing classification paradigm. While such choices have limited impact in the research setting, if classification based on gene expression profile is in future linked to treatment choice then classifier implementation will become a significant consideration. It could be argued that a systematic evaluation of classifier implementations is advisable in such circumstances, where sufficient publically available data sets are available.
The development of meta-profiles representing the most consistent differentially expressed genes between ABC- and GCB-DLBCL is significant since these gene lists are uniquely informed by the consistency of differential gene expression between multiple data sets. Indeed limited numbers of genes were detected as ABC- or GCB-associated in all data sets, and these genes are likely to be enriched for core regulators. Amongst the transcription factors most consistently linked to the ABC-subset is BATF
, an NFkB target gene that has recently been identified as a key partner of IRF4 
. This is of particular interest since ABC-DLBCL and myeloma show non-oncogenic addiction to IRF4 function 
. IRF4 depends on partner transcription factors to occupy DNA, the canonical partners being ETS-factors SPIB or PU.1 in B-cells; in contrast BATF plays a dominant role in T-cells 
. IRF4 is itself an NFkB target gene, which is a principle mechanism proposed for its expression in ABC-DLBCL 
expression in ABC-DLBCL may be explained in a similar fashion since it features on multiple NFkB signature lists, including the MSigDB signature V$NFKB_Q6_01. Indeed BATF
is induced transiently in activated B-cells during CD40 driven plasma cell differentiation 
. Amplification of chr19q as well as translocations can deregulate SPIB
, leading to functional addiction to this transcription factor 
. It will therefore be interesting to assess the relative contributions of SPIB and BATF to IRF4 function in ABC-DLBCL.
Another feature of the ABC-DLBCL meta-profile was the fact that three genes, ZBTB32
show a more consistent class-association than any of the primary classifier genes. ZBTB32
encodes a transcription factor, also known as Repressor of GATA (ROG), with functions in T-cell activation and differentiation 
. The role of this gene in B-cell differentiation is somewhat enigmatic, but ZBTB32
has recently been identified as a repressor of transcription rapidly induced during murine B-cell differentiation, which can co-operate with BLIMP1 in silencing CIITA
. The latter is essential for MHC class-II expression, and shows a weak class-association with GCB-DLBCL. Expression of ZBTB32 and repression of CIITA
may thus contribute to immune-evasion in ABC-DLBCL 
(also known as Kv1.3) encodes a potassium channel, which is expressed in effector memory T-cells and memory B-cells 
. KCNA3 has been identified as a potential target for therapy in autoimmune disease 
, and several strategies for inhibition have been developed in this context 
. The strong and consistent expression of KCNA3
identifies this as an intriguing candidate for a novel therapeutic approach in ABC-DLBCL.
Enrichments of gene signatures derived from previous studies 
, across meta-profiles and across individual data sets provides a resource for the identification of class/signature/gene associations of relevance to lymphoma pathogenesis and further confirmed the validity of classifications. A striking feature was the ability to identify gene sets associated with known regions of chromosomal deregulation in DLBCL. Several studies have assessed the contribution of copy number alterations to deregulated gene expression 
, with a notable recent paper using integrated analysis to assess the contribution of multiple regions of copy number change to deregulation of cell-cycle checkpoints 
. Our comparative analysis has allowed a distinct approach identifying over-represented chromosomal cytobands by the consistency of gene expression across multiple data sets. This approach emphasised the particular importance of chr3 and chr18 to ABC-DLBCL, and also identified genes in chromosomal regions as potential drivers of pathogenesis. An example of this was evident in chr18q21. Amplification of chr18q is common in ABC-DLBCL and associated with aberrant expression of BCL2
. Identifying differential expression of BCL2
is complicated by the fact that BCL2
translocations are a common feature of GCB-DLBCL, but nonetheless BCL2
was ABC class-associated. However, the most consistently differentially expressed gene from chr18q21 in ABC-DLBCL was TCF4
. This gene encodes a transcription factor also known as E2-2, which can drive SPIB
expression and co-operate with SPIB to regulate the transcriptional program of plasmacytoid dendritic cells 
. Notably several dendritic cell signatures are enriched in the ABC-DLBCL meta-profile including Dendritic_cell_CD123pos_blood, CD123 being a surface marker of plasmacytoid dendritic cells, ranked 6th
of all signatures in the ABC meta-profile (12.38% overlap, FDR corrected p-value
8.95E-18). Thus in addition to the deregulation of BCL2, amplification of chr18q21 is likely to contribute to the consistent expression of TCF4
, and hence to establishing the transcriptional network of ABC-DLBCL.
In conclusion, the generation of the robust classifier algorithm, DAC, provides a tool with which to consistently classify DLBCL cases regardless of microarray platform type. It has potential applications in the research and clinical setting, since it is designed, and is currently being used, to allow real-time assessment of individual incident cases. Currently real-time classification of DLBCL cases into molecular classes does not affect primary clinical management decisions, but in future this may change. The analysis we have performed highlights the issues surrounding the effect that classifier choice can have on class assignment, and argues for a robust analysis of classifier algorithm in such settings. The development of this classifier has allowed the generation of a useful resource in which the consistency of class-associated gene expression provides a method for identifying associations of relevance to disease biology, and in particular highlights transcription factors operating in ABC-DLBCL.