Acute myeloid leukemia: leukemia stem cells write a prognostic signature

In a recent interesting article, analysis of gene expression between phenotypically defined acute myeloid leukemia (AML) leukemia stem cells (LSCs) and more mature leukemia progenitor cells is used to generate a differentially expressed gene signature for LSCs. Through clever bioinformatic weighting analysis, the authors describe a method to convert this signature into a single score for any given sample and then test the prognostic utility of this 'LSC score' in publicly available gene expression profiles from bulk AML samples. They demonstrate that a high LSC score is associated with poor prognosis in AML patients and further demonstrate that the score is independent of known prognostic factors, including age, karyotype and mutation of the FLT3 or NPM1 genes. These findings are important and directly relate transcriptional dysregulation in AML LSCs with the outcome in patient samples, thus reinforcing the belief that these cellular populations are crucial for the initial propagation and subsequent relapse and resistance of leukemia.

To date, cytogenetic abnormalities have provided most prognostic information, although up to 50% of patients have a normal karyotype [2]. In addition, over the past decade the detection of specifi c molecular mutations, including mutations in the NPM1 and FLT-3 genes, have provided additional prognostic information to predict relapse and overall survival. Global gene expression profi les of bulk leukemia cells have provided further data to improve classifi cation and prognosis [3,4], and have identifi ed overexpression of specifi c genes, such as BAALC and ERG, to be associated with poor survival. Th ese studies have aided our understanding of the biology of AML as a disease characterized by transcriptional dysregulation [5].
Our understanding of the organization of malignancy and the basis of therapeutic failure have recently been informed by the demonstration that certain malignancies are arranged in hierarchies and have a cancer stem or initiating cell at their apices, which propagates and maintains the tumor [6]. Th e cancer stem cell hypothesis is controversial in some solid organ malignancies but is well accepted in leukemias, and AML was the fi rst malignancy in which a cancer stem cell was demonstrated [7,8]. Th e malignant hierarchy in AML is similar to normal hematopoietic ontogeny, with leukemia stem cells (LSCs) in most AML patients being positive for the surface marker CD34, as are normal hematopoietic stem cells, and demonstrating variable expression of the CD38 and CD90 markers [7][8][9][10]. However, although it is thought that the LSC is responsible for disease relapse and secondary resistance, our understanding of this cellular compartment is poor and it is unlikely that signifi cant therapeutic advances towards reducing disease relapse will be achieved until this ignorance is reversed. Specifi cally, it is not yet known whether the gene expression patterns in bulk tumor cells are refl ective of the transcriptional dysregulation in their initiating or stem cells; this knowledge may inform our understanding of their biology and suggest therapeutic targets.
Th is issue is addressed in the interesting recent paper from Gentles and colleagues [11]. Th e authors compare gene expression profi les in LSCs and more diff erentiated 'leukemia progenitor cell' subpopulations from AML

Abstract
In a recent interesting article, analysis of gene expression between phenotypically defi ned acute myeloid leukemia (AML) leukemia stem cells (LSCs) and more mature leukemia progenitor cells is used to generate a diff erentially expressed gene signature for LSCs. Through clever bioinformatic weighting analysis, the authors describe a method to convert this signature into a single score for any given sample and then test the prognostic utility of this 'LSC score' in publicly available gene expression profi les from bulk AML samples. They demonstrate that a high LSC score is associated with poor prognosis in AML patients and further demonstrate that the score is independent of known prognostic factors, including age, karyotype and mutation of the FLT3 or NPM1 genes. These fi ndings are important and directly relate transcriptional dysregulation in AML LSCs with the outcome in patient samples, thus reinforcing the belief that these cellular populations are crucial for the initial propagation and subsequent relapse and resistance of leukemia.
patients, as defi ned by surface phenotype during normal ontogeny (CD34+/CD38-/CD90-and CD34+/CD38+, respec tively). Th eir dataset combines their own analysis of seven paired samples with a published dataset of eight paired gene expression profi les from LSCs that have been functionally validated in immunodefi cient mice [12]. Th e authors found 52 genes to be diff erentially expressed, with the 21 genes down-regulated in the LSC fraction functionally enriched for genes involved in cell cycle, proliferation and diff erentiation, in keeping with the quiescent, diff erentiation-blocked phenotype predicted for LSCs. Th e 31 up-regulated genes were subsequently used to generate an 'LSC signature' , where each gene was given a mathematical weighting relative to its predictive value within the dataset. Th e sum of these weightings could then be calculated for a given sample to obtain a single value or 'LSC score' , and this score further correlated with clinical outcome and known prognostic factors. To address its prognostic value in AML patients, the 'LSC score' was recalculated using the same 31 genes in another dataset of gene expression profi les obtained from bulk blasts of 163 patients with normal karyotype AML. Th e authors demonstrated in this 'test' dataset that a high LSC score was associated with poorer overall survival, both as a continuous variable or a dichotomized value. Importantly, they validated these fi ndings in the normal karyotype patients in another three large publicly available datasets of gene expression profi les across more diverse AML cases, and demonstrated similar fi ndings in patients with cytogenetic abnormalities. Moreover, the independent prognostic value of the LSC score was shown by multivariate Cox regression analysis to be independent of known prognostic factors, including age, karyotype and mutation of the FLT3 or NPM1 genes.
Th ese data are important fi ndings and directly relate the genes dysregulated in LSCs with the outcome in patient samples, thus reinforcing the belief that these cellular populations are crucial for the propagation of leukemia. It is currently not practicable to predict a priori prognosis in patients using global gene expression patterns of even bulk tumor cells, although it may be possible that elements of this signature could be used in quantitative PCR assays in the future, if the elements were further biologically validated and the assays standard ized. Although the prognostic implications of this study are obvious, the biological implications are less so. Th e authors speculate that those leukemias with a high LSC score are likely to have an increased frequency of LSCs within the bulk tumor. Th is interpretation is reasonable and would be in keeping with two previous reports, one of which demonstrated that a high frequency of phenotypic LSCs at diagnosis of AML was predictive of a high minimal residual disease and poor overall survival [13]. Th e other, functional study demonstrated that the effi ciency of disease transfer into immunocompro mised mice (a function of LSC activity) was also predictive of the clinical outcome of the patient [14]. Unfortunately, no correlation of the frequency of the LSCs, as determined by surface phenotype, with the LSC score was provided, as this frequency could also provide a surrogate prognostic marker. In addition, this study may also provide indirect evidence of the target cell for transformation to generate LSCs in AML. Th e LSC score was also derived for the same 31-transcript gene set in normal hematopoietic stem cells and progenitor populations. Of note, the average LSC score for phenotypic AML LSCs was most comparable to the multipotent progenitor and granulocyte monocyte progenitor popula tions. Th ese same compartments have recently been demonstrated to be expanded and to contain functional LSCs in the majority of AML cases [10]. Finally, of the 52 genes diff erentially expressed, only a few candidate genes have been previously implicated in leukemogenesis, including HLF [15], SETBP1 [16] and MEF2C [17]. Further evaluation of the other members of the gene set may yet provide even more information on the biology of the LSC in AML.