Identification of Gene Expression Signatures Accurately Predicting Cytogenetic Subtypes in Pediatric Acute Myeloid Leukemia.

Pediatric acute myeloid leukemia (AML) is a heterogeneous disease, which is classified according to the WHO classification, based on morphology, immunophenotyping and non-random genetic aberrations. AML is hypothesized to arise from two different types of genetic aberrations, i.e. type-I (proliferation enhancing) mutations and type-II (differentiation impairing) mutations. To detect genetic aberrations multiple techniques such as conventional karyotyping, FISH and RT-PCR are being used. In addition to conventional karyotyping, the latter two techniques revealed a higher frequency of aberrations. Still, failures or false negative results should be taken into account. Recent studies have focused on the potential of gene expression profiling (GEP) to classify acute leukemias. To study the clinical value of classification by GEP, we first used a double-loop cross validation (CV) to avoid over-fitting of GEP data and, subsequently, addressed whether the identified GEP was suitable to classify pediatric AML cases in a second independent group of cases. Affymetrix Human Genome U133 plus 2.0 microarrays were used to generate gene expression profiles of 257 children with AML, with high blast counts, if necessary, after enrichment (~80% or more) and good quality RNA. Probe set intensities were normalized using the variance-stabilizing normalization (VSN) implemented in R (version 2.2.0). The patient group was divided into a test cohort (n=170) and an independent validation cohort (n=87). The test cohort was used to construct the classifier using two levels of CV:

• the minimum number of predictive genes was estimated using a 10-fold CV on random subsets of about 113 (~2/3 of total) patients;

• the accuracy of the obtained classifier is estimated on the remaining 57 (~1/3) patients. Candidate genes to represent the GEP in the classifier were those genes that discriminated AML subtypes according to an empirical Bayes linear regression model (Bioconductor package: Limma).

To construct a reliable classifier it was sufficient to use 75 probe sets, representing the top 15 discriminating probe sets for MLL-gene rearranged AML, t(8;21), inv(16), t(15;17) and t(7;12). These subtypes represented ~50% of the included patients. The remaining patients either had normal cytogenetics, random aberrations or no data available (cytogenetic failure). Due to the heterogeneity of these remaining groups discriminative probe sets were not found. This classifier could reliably predict the 5 subtypes with a median accuracy of 93%. Validation of the classifier on the independent cohort confirmed that the sensitivity and accuracy was more than 99%. No gene expression signatures could be found for the molecular aberrations NPM1, CEBPa, MLL-PTD, FLT3, C-KIT, RAS or PTPN1, possibly due to the small number of cases. However, specific gene expression signatures were found for FLT3-ITD within the subset of cases with t(15;17) or normal cytogenetics. Importantly, a high expression of HOXB-cluster related genes was found in cases with FLT3-ITD and normal cytogenetics. In conclusion, GEP can correctly predict several important cytogenetic subtypes of pediatric AML, including cases that are currently classified using different cytogenetic techniques and cases with failed cytogenetic analysis. Prospective studies are needed to validate the use of GEP in the classification of pediatric AML, especially to provide information on its utility in clinical practice. Increasing numbers in rare subtypes may result in the discovery of genes discriminative for them, and may foster GEP as a new diagnostic tool.