With advances in high-density DNA microarray technology, it has become possible to screen large numbers of genes to see whether or not they are active under various conditions. This is gene-expression profiling, and there has been an expectation that it will revolutionize cancer diagnosis ( Box 1)1,2. The thinking is that tumour behaviour is dictated by the expression of thousands of genes, and that micro-array analysis should allow that behaviour and the clinical consequences to be predicted. This rationale is sound enough, but until now it has not been substantiated by experiment.
On page 503 of this issue3, Alizadeh et al. deliver such substantiation. The particular cancer they have looked at is diffuse large B-cell lymphoma (DLBCL), a disease that takes in a clinically and morphologically varied group of tumours that affect the lymph system and blood. The authors carried out gene-expression profiling with a ‘Lymphochip’, a microarray carrying 18,000 clones of complementary DNA designed to monitor genes involved in normal and abnormal lymphocyte development.
Using clustering analysis, Alizadeh et al. could separate DLBCL into two categories, which had marked differences in overall survival of the patients concerned. The gene-expression signatures of these subgroups corresponded to distinct stages in the differentiation of B cells, the type of lymphocyte that makes antibodies.
Diffuse large B-cell lymphoma is the most common subtype of non-Hodgkin's lymphoma. With current treatments, long-term survival can be achieved in only 40% of patients. There are no reliable indicators — morphological, clinical, immunohistochemical or genetic — that can be used to recognize subclasses of DLBCL and point to a differential therapeutic approach to patients4.
Expression profiling has already shown its usefulness in identifying genes with high or low expression levels in specific cell types under defined conditions — for instance, when being stimulated with growth factors or treated with drugs, or when the cell's degree of attachment to the extracellular matrix varies (this last characteristic may determine tumour spread)2,5,6,7. More recently, reports on tumour classification have also begun to emerge. Acute leukaemias can effectively be divided into the lymphoblastic and myeloblastic forms by expression profiling8. But in these studies, no multigene-expression signature was found that correlated with a new leukaemia subgroup, or with clinical outcome, in the relatively small group of tumours examined.
In the case of Alizadeh and colleagues' analysis3 of DLBCL, the situation is different. Hierarchical clustering of the gene-expression data divided DLBCLs into two groups: one had the signature of B cells from the germinal centres (the B-cell factories in lymph nodes); the other had the signature of activated B cells. The outlook for patients who had tumours with the activated-B-like signature was much worse — 16 out of 21 died, compared with 6 out of 19 patients with the germinal-centre B-cell signature. Importantly, the predictive value was independent of the standard clinical parameters of prognosis, the International Prognostic Indicator.
That is far from the end of the story, of course. As the authors point out, most of the patients in the ‘favourable prognosis’ group that die do so within the first two years of diagnosis, whereas some of the patients in the ‘poor prognosis’ group were still alive after five years. The question is whether there is a ‘hidden signature’ that, if found, would enable early identification of these subgroups. For the moment, we just cannot say. Testing of more tumours, and using larger or different DNA microarrays, might be needed to resolve this question. In addition, some prognostic indicators might escape detection by expression profiling as they are qualitative, rather than quantitative. That is, genetic (allelic) differences might mean that some genes escape expression screening. They could encode proteins with a different activity or stability that affects tumour progression or response to treatment. In this respect, monitoring of single-nucleotide polymorphisms (SNPs — individual differences at a single base pair that mark a particular genetic variation in the population) would constitute an appealing complementary approach to screening9.
When more expression signatures of larger tumour sets become available, it will become clear how this approach will improve monitoring of the stages in which tumours grow and spread, and therefore prognosis. The expectations are high. Furthermore, the better definition of patient groups, made possible by expression profiling, is of obvious importance for assessing the efficacy of various treatments. Patients will benefit directly from the tailoring of therapies to specific subclasses of tumours.
Finally, gene-expression profiling can be used to identify the genes and pathways that really matter for the tumorigenic process, thereby revealing new targets for therapy. The studies discussed here, and others, show that it is indeed feasible to identify such pathways. For example, high expression of the homeobox gene HOXA9, previously identified as a potent oncogene, was correlated with the failure to induce remission in a small group of patients suffering from acute myeloblastic leukaemia8. The DLBCL study revealed the increased expression of a series of genes involved in the inhibition of apoptosis (programmed cell death); this, too, might be a therapeutic avenue to explore.
A word of caution. Prognoses based on gene-expression signatures, and — in the future — SNP analysis, are likely to have their limitations. These methods will probably provide quite reliable predictions of the patient responses to initial therapies. The question then is how much ‘noise’ will remain as a result of tumour heterogeneity and genetic instability. Both are hallmarks of most malignancies and can lead to recurrent disease, as cancer specialists and their patients know only too well. Time will tell.
For the moment we are witnessing spectacular developments in tumour diagnosis. These developments are going hand in hand with new treatments targeted to defined regulatory pathways that are frequently deregulated in cancer. So although caution is in order, so too is optimism.
Alizadeh, A. A. et al. J. Clin. Immunol. 18, 373– 379 (1998).
Perou, C. M. et al. Proc. Natl Acad. Sci. USA 96, 9212– 9217 (1999).
Alizadeh, A. A. et al. Nature 403, 503–511 (2000).
Harris, N. L. et al. J. Clin. Oncol. 17, 3835– 3849 (1999).
Iyer, V. R. et al. Science 283, 83–87 (1999).
Marton, M. J. et al. Nature Med. 4, 1293– 1301 (1998).
Alon, U. et al. Proc. Natl Acad. Sci. USA 96, 6745– 6750 (1999).
Golub, T. R. et al. Science 286, 531–537 (1999).
Cargill, M. et al. Nature Genet. 22, 231– 238 (1999).
About this article
Current Opinion in Urology (2019)
Partial maximum correlation information: A new feature selection method for microarray data classification
Gene selection and disease prediction from gene expression data using a two-stage hetero-associative memory
Progress in Artificial Intelligence (2019)
Current Oncology Reports (2018)