Are Microarray-Derived Gene Expression Signatures Useful as a Routine Assay for Prognostic Evaluation of Acute Myeloid Leukemia Samples with Low Leukemic Blast Content ?.

Abstract 1658

Prognostic evaluation is a critical step in newly diagnosed pts with AML. With samples containing a high leukemic blast load, microarray-based gene expression profiling (GEP) and class prediction analyses have demonstrated their ability to assign AML samples to one of the three well characterized favorable-risk AML subtypes, ie, APL, CBFA-AMLs, CBFB-AMLs, with high accuracy and low error rates. One major issue regarding the use of microarray-based GEP in a routine prognostic workflow, concerns its ability to perform accurate class prediction analyses with samples containing as low as 20% blasts. So far, the minimum leukemic cell load that would still lead to correct class assignment has not been extensively studied. Another critical issue for a routine use of microarray-based GEP concerns samples that do not fulfill all quality control criteria along the process. The accuracy of class prediction analyses using data derived from poor quality AML samples also deserves further evaluations.

In this study, using the Illumina BeadChip technology, GEP was first assessed for its ability to correctly classify APLs, CBFA-AMLs, CBFB-AMLs within a training set of 72 samples containing at least 60% of leukemic cells (20 APL samples from 14 pts, 12 CBFA-AML samples - n=6 pts, 11 CBFB-AML samples - n=10 pts). This training set also included normal bone marrow samples as well as a more heterogeneous group of 29 NK-AML samples (28 pts). The classifiers derived from this supervised analysis were then evaluated for their ability to assign to the correct class 93 test samples with either low leukemic blast load (as low as 5% of blasts – n=79 - 22 samples that originally contained less than 60% of leukemic cells, 57 samples with high blast load that were artificially diluted within a pool of normal bone marrows at 50 and 75%, ie, final leukemic blast load after dilution ranging from 10 to 50%), poor quality control criteria (n=10), or with atypical characteristics (n=4). The Class Prediction Module of ArrayMiner 5.3.3 software, which is a supervised method based on a proprietary optimization technique called grouping genetic algorithms (GGA), was used to build the classifiers associated with each class of the Training Set and assess the predictive capacity of these classifiers on Test Set samples.

Using the Training Set and the GGA-base supervised method, the best model fitness was achieved with classifiers including 14 markers per class. With these classifiers, all Training Set samples were assigned to the correct class. Considering Test Set samples, all APL and CBFA-AML samples were assigned to the correct class, including the ones with the lowest leukemic blast load, i.e., 7% for APLs and 15% for CBFA-AMLs. One of the 23 CBFB-AML Test Set samples, which contained 5% of blasts was misclassified. All other CBFB-AML test samples were correctly classified, even those contained as low as 10% of leukemic cells. All but one of the NK-AML test samples were assigned to the correct class, even though some contained as low as 19% of leukemic cells. Concerning the 10 poor quality control samples characterized by either low RIN values and degraded total RNAs, low amounts of labeled cRNAs or both low RIN value and low labeled cRNA, all were assigned to the expected class. All atypical AMLs were assigned, as expected, to the NK-AML group.

In conclusion, favorable cytogenetic risk AMLs with low leukemic cell load or poor quality control criteria can be correctly assign to the appropriate group using GEP. Transposability of this Illumina technology-based model and its efficiency in term of predictive capacity, will be presented using an independent series of 245 AML samples processed at a different institution, on Affymetrix GeneChips.